{"days":[
{"date":"2025-10-12","headline":"First edition of The Public Health AI Handbook published; The history of AI, from the 1956 Dartmouth Conference to the present; MYCIN, early surveillance, and IBM Watson case studies added to the history chapter; Supervised, unsupervised, and reinforcement learning for practitioners; The Iron Law of Machine Learning and the surveillance pyramid; Outbreak detection, from traditional surveillance to BlueDot and BioSense; Compartmental models from SIR through SEIRD, and time series forecasting","entries":[{"category":"milestone","title":"First edition of The Public Health AI Handbook published","change":"The first edition of The Public Health AI Handbook is organized in five parts with 16 chapters and four appendices, and includes a Welcome page, a Preface, and a References page.","why":"Establishes the handbook's first public edition and its overall structure.","chapters":[["Welcome","/"],["Preface","/preface.html"],["References","/references.html"]],"sources":[]},{"category":"content","title":"The history of AI, from the 1956 Dartmouth Conference to the present","change":"AI in Context: A Brief History traces the birth of AI at the 1956 Dartmouth Conference, the Turing Test, and the major historical eras through the present, citing Alan Turing's 1950 paper introducing the Turing Test.","why":"Gives the handbook's opening chapter a complete historical narrative grounded in a primary source.","chapters":[["AI in Healthcare: A Brief History","/foundations/history.html"]],"sources":[{"label":"Turing, 1950","url":"https://doi.org/10.1093/mind/LIX.236.433","type":"paper"}]},{"category":"evidence","title":"MYCIN, early surveillance, and IBM Watson case studies added to the history chapter","change":"AI in Context: A Brief History adds evaluation data for MYCIN (matching infectious disease specialists in a 1979 study) and the HealthMap surveillance system (2006), plus a discussion of IBM Watson's Jeopardy! win and its cancer-treatment recommendations that were later found to be unsafe and incorrect, framed as a demonstration-to-deployment gap.","why":"Gives readers documented historical examples of both a successful and a failed clinical AI deployment.","chapters":[["AI in Healthcare: A Brief History","/foundations/history.html"]],"sources":[{"label":"1979 MYCIN evaluation study","url":"https://doi.org/10.1016/0010-4809(79)90015-4","type":"paper"},{"label":"HealthMap, 2006","url":"https://doi.org/10.1197/jamia.M2544","type":"paper"},{"label":"IBM Watson Jeopardy! results, 2012","url":"https://doi.org/10.1147/JRD.2012.2184356","type":"paper"},{"label":"STAT News, IBM Watson Oncology report, 2018","url":"https://www.statnews.com/2018/07/25/ibm-watson-recommended-unsafe-incorrect-treatments/","type":"news"}]},{"category":"content","title":"Supervised, unsupervised, and reinforcement learning for practitioners","change":"Just Enough AI to Be Dangerous introduces the three fundamental machine learning paradigms, supervised, unsupervised, and reinforcement learning, together with core algorithms and the concepts practitioners need, and recommends further reading for deeper technical study.","why":"Gives readers a foundational, non-technical grounding in machine learning concepts before the chapter's applied AI content.","chapters":[["Machine Learning Fundamentals","/foundations/ai-basics.html"]],"sources":[]},{"category":"content","title":"The Iron Law of Machine Learning and the surveillance pyramid","change":"The Data Problem sets out the Iron Law of Machine Learning, the surveillance pyramid problem, and the other reasons public health data is uniquely challenging for AI, citing Andrew Ng's writing on data-centric AI.","why":"Gives readers a foundational framework for assessing data quality before applying AI to public health problems.","chapters":[["The Data Problem in Public Health AI","/foundations/data.html"]],"sources":[{"label":"Ng, The Batch, data-centric AI","url":"https://www.deeplearning.ai/the-batch/issue-107/","type":"other"}]},{"category":"content","title":"Outbreak detection, from traditional surveillance to BlueDot and BioSense","change":"Disease Surveillance and Outbreak Detection sets traditional surveillance systems beside AI-enhanced ones, among them BlueDot's AI-based outbreak alerts and CDC's BioSense syndromic surveillance platform.","why":"Gives readers a foundation on how surveillance systems detect disease outbreaks, from traditional methods to AI-enhanced tools.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"BlueDot","url":"https://bluedot.global/","type":"other"},{"label":"CDC, BioSense Platform","url":"https://www.cdc.gov/nssp/biosense/index.html","type":"policy"},{"label":"CDC, National Syndromic Surveillance Program","url":"https://www.cdc.gov/nssp/overview.html","type":"policy"}]},{"category":"content","title":"Compartmental models from SIR through SEIRD, and time series forecasting","change":"Epidemic Forecasting works through compartmental models from SIR through SEIRD, R0 estimation, and time series methods including ARIMA, drawing on a review of spatial epidemic models and a textbook on modeling infectious diseases.","why":"Gives readers a technical foundation for building and interpreting epidemic forecasting models.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Delamater et al., 2019","url":"https://wwwnc.cdc.gov/eid/article/25/1/17-1901_article","type":"paper"},{"label":"Keeling and Rohani, 2008","url":"https://press.princeton.edu/books/hardcover/9780691116174/modeling-infectious-diseases-in-humans-and-animals","type":"other"}]}]},
{"date":"2025-10-13","headline":"Genomic surveillance and pathogen analysis: tracking SARS-CoV-2 and Omicron; Large language models in public health: how they work and where they fail; Clinical decision support from MYCIN to machine learning; Evaluating AI systems: technical performance and clinical validation; Racial bias in a risk-prediction algorithm, and the privacy of health data; New chapter: Deployment, Monitoring, and Maintenance; Choosing a development environment and the modern AI software stack; A hospital readmission prediction project, end to end; The pace of AI progress, and the divide between high- and low-resource settings; Regulatory frameworks for AI in healthcare, and where oversight gaps remain; Glossary expanded to 500 or more terms; 15 AI deployments in public health, from BlueDot to Google Flu Trends; Code Repository Guide and Further Reading appendices expanded","entries":[{"category":"content","title":"Genomic surveillance and pathogen analysis: tracking SARS-CoV-2 and Omicron","change":"Genomic Surveillance and Pathogen Analysis covers the central dogma, key concepts in genomic epidemiology, and pathogen tracking during COVID-19, citing the earliest public posting of the SARS-CoV-2 genome, the GISAID sequence database, and one of the first United States Omicron outbreak investigations.","why":"Gives readers a technical foundation in genomic surveillance grounded in the COVID-19 pandemic response.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Virological.org, SARS-CoV-2 genome posting, 2020","url":"https://virological.org/t/novel-2019-coronavirus-genome/319","type":"other"},{"label":"GISAID","url":"https://www.gisaid.org/","type":"dataset"},{"label":"Tegomoh et al., 2021, MMWR","url":"https://www.cdc.gov/mmwr/volumes/70/wr/mm705152e3.htm","type":"paper"}]},{"category":"content","title":"Large language models in public health: how they work and where they fail","change":"Large Language Models in Public Health explains how large language models work, from tokenization through the 2022 release of ChatGPT and the 2023 release of GPT-4, and states their risks: unreliable outputs, bias, and privacy concerns.","why":"Gives readers a technical and risk-aware grounding in large language models before they are asked to evaluate LLM-based tools.","chapters":[],"sources":[{"label":"OpenAI, ChatGPT release, 2022","url":"https://openai.com/blog/chatgpt","type":"announcement"},{"label":"Reuters, ChatGPT user growth, 2023","url":"https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/","type":"news"},{"label":"OpenAI, GPT-4 announcement, 2023","url":"https://openai.com/research/gpt-4","type":"announcement"}]},{"category":"content","title":"Clinical decision support from MYCIN to machine learning","change":"Diagnostic and Clinical Decision Support traces the evolution from rule-based systems such as MYCIN to machine learning, citing validation research on Epic's sepsis prediction model.","why":"Gives readers a historical and evidentiary basis for evaluating modern AI diagnostic tools.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"JAMA Internal Medicine, Epic sepsis model validation","url":"https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307","type":"paper"}]},{"category":"content","title":"Evaluating AI systems: technical performance and clinical validation","change":"Evaluating AI Systems treats technical performance alongside clinical validation, citing a review of deep learning studies in medical imaging, Van Calster et al. on clinical utility assessment, and Gianfrancesco et al. on algorithmic bias.","why":"Gives readers a structured framework for judging whether an AI system is ready for public health use.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Radiology, deep learning review, 2020","url":"https://pubs.rsna.org/doi/10.1148/radiol.2020191432","type":"paper"},{"label":"Van Calster et al., 2019, BMJ","url":"https://www.bmj.com/content/364/bmj.l230","type":"paper"},{"label":"Gianfrancesco et al., 2018, Nature Communications","url":"https://www.nature.com/articles/s41467-018-06170-0","type":"paper"}]},{"category":"content","title":"Racial bias in a risk-prediction algorithm, and the privacy of health data","change":"Ethics, Bias, and Equity introduces the Obermeyer case on racial bias in a healthcare risk-prediction algorithm; Privacy, Security, and Governance examines the Cambridge Analytica case and major health data breaches.","why":"Gives readers grounded starting points for evaluating fairness, bias, and privacy risk in public health AI systems.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"],["Privacy, Security, and Governance for Health AI","/implementation/privacy.html"]],"sources":[]},{"category":"content","title":"New chapter: Deployment, Monitoring, and Maintenance","change":"A new chapter, Deployment, Monitoring, and Maintenance, covers MLOps principles for public health AI systems, citing Gartner's finding that only 53% of AI projects reach production and Sculley et al.'s account of technical debt in machine learning systems.","why":"Fills a gap between building an AI model and successfully operating it in a real public health setting.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[{"label":"Gartner, 2021","url":"https://www.gartner.com/en/newsroom/press-releases/2021-10-11-gartner-says-only-53-percent-of-projects-make-it-from-artificial-intelligence-prototype-to-production","type":"report"},{"label":"Reddy et al., 2020, Nature Medicine","url":"https://www.nature.com/articles/s41591-019-0740-8","type":"paper"},{"label":"Sculley et al., 2015, NIPS","url":"https://papers.nips.cc/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html","type":"paper"}]},{"category":"content","title":"Choosing a development environment and the modern AI software stack","change":"Your AI Toolkit guides readers through choosing development environments and the modern AI software stack, with a curated list of documentation and tutorials for pandas and scikit-learn.","why":"Gives readers a practical starting point for setting up their own AI development environment.","chapters":[["Your AI Toolkit for Public Health","/practical/toolkit.html"]],"sources":[]},{"category":"content","title":"A hospital readmission prediction project, end to end","change":"Building Your First Project walks readers through a hospital readmission prediction project end to end, including model interpretation, common pitfalls such as data leakage, and three dataset options for the exercise, citing the readmission research literature and CMS's Hospital Readmissions Reduction Program.","why":"Gives readers a complete, reproducible project to apply the handbook's concepts to a realistic public health prediction task.","chapters":[["Building Your First Public Health AI Project","/practical/first-project.html"]],"sources":[{"label":"Jencks et al., 2009, NEJM","url":"https://www.nejm.org/doi/full/10.1056/NEJMsa0803563","type":"paper"},{"label":"CMS, Hospital Readmissions Reduction Program","url":"https://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/AcuteInpatientPPS/Readmissions-Reduction-Program","type":"policy"},{"label":"Kansagara et al., 2011, Annals of Internal Medicine","url":"https://www.acpjournals.org/doi/10.7326/0003-4819-155-8-201110180-00008","type":"paper"}]},{"category":"content","title":"The pace of AI progress, and the divide between high- and low-resource settings","change":"Emerging Technologies assesses the pace of AI progress, citing Sevilla et al. on the growth of AI compute. Global Health and Equity examines the AI divide between high- and low-resource settings, citing AI-assisted TB screening from chest X-rays and smartphone-based microscopy.","why":"Planning for public health AI depends both on how fast the technology advances and on whether tools such as AI-assisted TB screening from chest X-rays serve low-resource settings as well as high-resource ones.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"],["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Sevilla et al., 2022, preprint","url":"https://arxiv.org/abs/2202.05924","type":"preprint"},{"label":"Murphy et al., 2020, PLOS Medicine","url":"https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1003198","type":"paper"},{"label":"Khan et al., 2020, Nature Medicine","url":"https://www.nature.com/articles/s41591-020-0907-8","type":"paper"}]},{"category":"content","title":"Regulatory frameworks for AI in healthcare, and where oversight gaps remain","change":"Policy and Governance examines regulatory frameworks for AI in healthcare, citing Gerke et al. and Finlayson et al. in Nature Medicine and the FDA's 510(k) database, which covers about 90% of AI device clearances.","why":"Gives readers a regulatory map for AI medical devices and where oversight gaps remain.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Gerke et al., 2020, Nature Medicine","url":"https://www.nature.com/articles/s41591-020-0842-9","type":"paper"},{"label":"Finlayson et al., 2021, Nature Medicine","url":"https://www.nature.com/articles/s41591-021-01270-5","type":"paper"},{"label":"FDA, 510(k) database","url":"https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm","type":"policy"}]},{"category":"content","title":"Glossary expanded to 500 or more terms","change":"The Glossary is substantially expanded to cover 500 or more AI, machine learning, and public health terms, organized alphabetically.","why":"Gives readers a single reference for terminology used throughout the handbook.","chapters":[["Glossary","/appendices/glossary.html"]],"sources":[]},{"category":"content","title":"15 AI deployments in public health, from BlueDot to Google Flu Trends","change":"The Case Study Library documents 15 real-world case studies spanning surveillance, diagnostics, treatment, child welfare, drug discovery, and rural health, including BlueDot's early COVID-19 detection, the rise and fall of Google Flu Trends, and the Allegheny Family Screening Tool.","why":"Gives readers concrete, documented examples of AI deployments across the full range of outcomes, from success to failure.","chapters":[["Case Study Library - Full Cases","/appendices/case-studies.html"]],"sources":[{"label":"Bogoch et al., 2020, Journal of Travel Medicine","url":"https://academic.oup.com/jtm/article/27/2/taaa008/5709322","type":"paper"},{"label":"Freifeld et al., 2008, PLOS Medicine","url":"https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.0050151","type":"paper"},{"label":"Lang et al., 2023, Lancet Digital Health","url":"https://www.thelancet.com/journals/landig/article/PIIS2589-7500(23)00003-0/fulltext","type":"paper"}]},{"category":"content","title":"Code Repository Guide and Further Reading appendices expanded","change":"The Code Repository Guide and Further Reading appendices are substantially expanded, with setup instructions for the companion code repository and a curated list of resources on foundational machine learning, healthcare AI, public health, and ethics.","why":"Gives readers practical setup guidance and a curated path for further study.","chapters":[["Code Repository Guide","/appendices/code-repository.html"],["Further Reading","/appendices/resources.html"]],"sources":[]}]},
{"date":"2025-10-14","headline":"Case Study Library appendix split into overview and full cases; Check Your Understanding quizzes added across chapters; New appendix: AI Morgue failure post-mortems; Part-ending summary sections added to all five parts","entries":[{"category":"content","title":"Case Study Library appendix split into overview and full cases","change":"Adds a new overview section to the Case Study Library appendix, summarizing all 15 case studies by category (surveillance, diagnostics, forecasting, and more) with links to the full write-ups.","why":"Gives readers a quick-reference index before they read the full case study library.","chapters":[["Case Study Library - Overview","/appendices/case-study-overview.html"]],"sources":[]},{"category":"content","title":"Check Your Understanding quizzes added across chapters","change":"Adds multiple-choice 'Check Your Understanding' quiz sections with explanations to chapters across all five parts, covering topics from data quality to surveillance, genomics, ethics, privacy, deployment, and toolkits.","why":"Lets readers test their grasp of each chapter's key concepts before moving on.","chapters":[["AI in Healthcare: A Brief History","/foundations/history.html"],["Machine Learning Fundamentals","/foundations/ai-basics.html"],["The Data Problem in Public Health AI","/foundations/data.html"],["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"],["Epidemic Forecasting with AI","/applications/forecasting.html"],["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"],["Privacy, Security, and Governance for Health AI","/implementation/privacy.html"],["Your AI Toolkit for Public Health","/practical/toolkit.html"]],"sources":[]},{"category":"content","title":"New appendix: AI Morgue failure post-mortems","change":"Adds an appendix, 'The AI Morgue,' a set of post-mortem case studies of failed health AI projects such as IBM Watson for Oncology, then adds a References section citing STAT News, Wall Street Journal, and IEEE Spectrum reporting on the failures.","why":"Complements the handbook's case studies of AI successes with documented, sourced examples of what went wrong.","chapters":[["The AI Morgue: Failure Post-Mortems","/appendices/failures.html"]],"sources":[{"label":"Ross and Swetlitz, STAT News, 2017","url":"https://www.statnews.com/2017/07/25/ibm-watson-cancer-unsafe/","type":"news"},{"label":"Hernandez and Greenwald, Wall Street Journal, 2018","url":"https://www.wsj.com/articles/ibm-bet-billions-that-watson-could-improve-cancer-treatment-it-hasnt-worked-1533961147","type":"news"},{"label":"Strickland, IEEE Spectrum, 2019","url":"https://spectrum.ieee.org/how-ibm-watson-overpromised-and-underdelivered-on-ai-health-care","type":"news"}]},{"category":"content","title":"Part-ending summary sections added to all five parts","change":"Adds 'What You Should Now Know' recap sections with skill checklists at the end of Parts I through V, closing with a handbook conclusion after Part V.","why":"Gives readers a checkpoint of skills and concepts before moving to the next part.","chapters":[["The Data Problem in Public Health AI","/foundations/data.html"],["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"],["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"],["Building Your First Public Health AI Project","/practical/first-project.html"],["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[]}]},
{"date":"2025-10-15","headline":"New chapter: AI, Misinformation, and the Infodemic; New chapter: practical guide to using LLMs in public health; New chapter: AI-assisted coding and development tools; New chapter: AI-driven behavioral interventions","entries":[{"category":"content","title":"New chapter: AI, Misinformation, and the Infodemic","change":"Adds a new chapter on how generative AI has reshaped public health misinformation, with worked examples.","why":"Gives readers dedicated coverage of AI-driven misinformation and infodemic response, previously absent from the handbook.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[]},{"category":"content","title":"New chapter: practical guide to using LLMs in public health","change":"Adds a new chapter on using large language models in public health practice, covering privacy and compliance requirements, model selection, and safe-use guidance.","why":"Extends the handbook's LLM coverage from theory into hands-on practical guidance.","chapters":[],"sources":[]},{"category":"content","title":"New chapter: AI-assisted coding and development tools","change":"Adds a new chapter on AI coding assistants such as GitHub Copilot for public health data work, covering tool selection, capabilities, and limitations, citing studies on developer productivity and security risks in AI-generated code.","why":"Gives practitioners evidence-based guidance on where AI coding tools help and where they introduce risk.","chapters":[["AI-Assisted Coding for Public Health Analysis","/practical/ai-coding.html"]],"sources":[{"label":"Kalliamvakou et al., GitHub, 2023","url":"https://github.blog/2023-06-27-the-economic-impact-of-the-ai-powered-developer-lifecycle-and-lessons-from-github-copilot/","type":"announcement"},{"label":"Prather et al., 2023","url":"https://doi.org/10.1145/3545945.3569823","type":"paper"},{"label":"Pearce et al., 2022","url":"https://doi.org/10.1109/MSP.2021.3117259","type":"paper"}]},{"category":"content","title":"New chapter: AI-driven behavioral interventions","change":"Adds a new chapter on applying AI and natural language processing to behavioral health interventions, citing Thaler and Sunstein's Nudge alongside digital-health studies on AI-based interventions and mobile sensing.","why":"Extends the handbook into AI applications for behavior change, an area not previously covered.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[{"label":"Thaler and Sunstein, Nudge, 2008","url":"https://yalebooks.yale.edu/book/9780143115267/nudge/","type":"other"},{"label":"Direito et al., 2023","url":"https://doi.org/10.2196/43479","type":"paper"},{"label":"Torous et al., 2016","url":"https://doi.org/10.1002/wps.20344","type":"paper"}]}]},
{"date":"2025-10-16","headline":"Large language models in public health: theory and practice; Epidemic detection versus forecasting","entries":[{"category":"revision","title":"Large language models in public health: theory and practice","change":"Large Language Models in Public Health: Theory and Practice covers model releases, practical use, safety considerations, and provider terms of service.","why":"Gives readers a reference for the capabilities, risks, and terms governing LLM use in public health.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Google DeepMind, Gemini 2.0 Flash, 2024","url":"https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/","type":"announcement"},{"label":"Google, Terms of Service","url":"https://policies.google.com/terms","type":"other"},{"label":"OpenAI, ChatGPT Enterprise","url":"https://openai.com/chatgpt/enterprise","type":"announcement"}]},{"category":"revision","title":"Epidemic detection versus forecasting","change":"Epidemic Forecasting distinguishes detection from prediction, illustrated with the 1987 'Great Storm' forecasting failure and a chaos-theory butterfly-effect analogy.","why":"Clarifies for readers why forecasting an epidemic is harder than simply detecting an ongoing outbreak.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Wikipedia, Butterfly effect","url":"https://en.wikipedia.org/wiki/Butterfly_effect","type":"other"},{"label":"BBC News, 1987","url":"https://www.bbc.co.uk/news/uk-41534451","type":"news"},{"label":"Wikipedia, Great Storm of 1987","url":"https://en.wikipedia.org/wiki/Great_Storm_of_1987","type":"other"}]}]},
{"date":"2025-10-17","headline":"Version 1.0: first complete edition of all 20 chapters","entries":[{"category":"milestone","title":"Version 1.0: first complete edition of all 20 chapters","change":"Marks version 1.0, the first complete edition of the handbook: all 20 chapters across five parts, with appendices and supporting code examples.","why":"Five days after the first chapters went online, every planned chapter had full content.","chapters":[],"sources":[]}]},
{"date":"2025-10-18","headline":"Discussion questions and a development timeline for the history chapter; Transformer architecture diagram in the LLM theory chapter; Diagrams for model behavior and epidemic models","entries":[{"category":"content","title":"Discussion questions and a development timeline for the history chapter","change":"AI in Context: A Brief History adds end-of-chapter discussion questions on historical AI failures, deployment challenges, and equity implications, plus a visual timeline of AI development eras from the 1950s to the 2020s.","why":"Gives readers reflection prompts and a chronological visual reference for the chapter's historical narrative.","chapters":[["AI in Healthcare: A Brief History","/foundations/history.html"]],"sources":[]},{"category":"content","title":"Transformer architecture diagram in the LLM theory chapter","change":"Large Language Models in Public Health: Theory and Practice adds a diagram of the transformer encoder-decoder architecture showing multi-head self-attention, feed-forward layers, and positional encoding, credited to Jay Alammar's 'The Illustrated Transformer'.","why":"Gives readers a visual reference for the attention-based architecture underlying modern language models.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[]},{"category":"content","title":"Diagrams for model behavior and epidemic models","change":"Adds diagrams to four foundational and evaluation chapters illustrating model overfitting and underfitting, SIR and SEIR epidemic models, model validation, and surveillance data pyramids.","why":"Gives readers visual references for core AI evaluation and epidemic modeling concepts introduced in the text.","chapters":[["Machine Learning Fundamentals","/foundations/ai-basics.html"],["The Data Problem in Public Health AI","/foundations/data.html"],["Epidemic Forecasting with AI","/applications/forecasting.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[]}]},
{"date":"2025-10-19","headline":"Obermeyer healthcare algorithm bias case study; Cambridge Analytica case study in the privacy chapter; AI development timeline image in the history chapter; WHO and Nextstrain variant naming conventions","entries":[{"category":"content","title":"Obermeyer healthcare algorithm bias case study","change":"Ethics, Bias, and Equity adds a case study describing the 2019 research led by Ziad Obermeyer that uncovered bias embedded in a widely used commercial healthcare risk-prediction algorithm, cited as Obermeyer et al., 2019.","why":"Gives readers a detailed, concrete example of algorithmic bias in a widely deployed healthcare tool.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[]},{"category":"developments","title":"Cambridge Analytica case study in the privacy chapter","change":"Privacy, Security, and Governance adds a case study on the 2018 Cambridge Analytica disclosure, describing how the firm harvested personal data from 87 million Facebook users, citing the Guardian and New York Times reporting that broke the story.","why":"Illustrates public health privacy risks with a documented, large-scale data harvesting case.","chapters":[["Privacy, Security, and Governance for Health AI","/implementation/privacy.html"]],"sources":[{"label":"The Guardian, Cambridge Analytica Files, 2018","url":"https://www.theguardian.com/news/2018/mar/17/cambridge-analytica-facebook-influence-us-election","type":"news"}]},{"category":"content","title":"AI development timeline image in the history chapter","change":"AI in Context: A Brief History adds a timeline graphic spanning AI eras from the 1950s to today, with a caption linking to an AJOT explainer on AI history.","why":"Gives readers a visual chronology to accompany the chapter's historical narrative.","chapters":[["AI in Healthcare: A Brief History","/foundations/history.html"]],"sources":[{"label":"AJOT, Breaking Down Artificial Intelligence","url":"https://www.ajot.com/insights/full/ai-breaking-down-artificial-intelligence-ai","type":"other"}]},{"category":"policy","title":"WHO and Nextstrain variant naming conventions","change":"Genomic Surveillance and Pathogen Analysis adds detail on the WHO's Greek-alphabet variant naming system for public communication and on Nextstrain's clade naming convention.","why":"Explains the parallel naming systems readers encounter when tracking pathogen variants.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"WHO, SARS-CoV-2 variant tracking","url":"https://www.who.int/activities/tracking-SARS-CoV-2-variants","type":"policy"},{"label":"Nextstrain, clade naming reference","url":"https://docs.nextstrain.org/projects/ncov/en/latest/reference/naming_clades.html","type":"other"}]}]},
{"date":"2025-10-23","headline":"Evaluation chapter expanded with foundation model, MLOps, and regulatory content","entries":[{"category":"content","title":"Foundation model and LLM evaluation section added to the evaluation chapter","change":"Evaluating AI Systems adds a new section on evaluating foundation models and large language models, covering medical LLM benchmarks, factual accuracy and safety metrics, retrieval-augmented generation evaluation, MLOps drift monitoring, and regulatory frameworks including FDA Software as a Medical Device classification and the EU AI Act.","why":"Gives readers current, 2025-relevant methods for evaluating medical LLMs, monitoring deployed models, and meeting regulatory expectations for AI-based medical software.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"FDA, Digital Health Center of Excellence","url":"https://www.fda.gov/medical-devices/digital-health-center-excellence","type":"policy"},{"label":"IMDRF, SaMD risk categorization framework","url":"http://www.imdrf.org/docs/imdrf/final/technical/imdrf-tech-140918-samd-framework-risk-categorization-141013.pdf","type":"other"},{"label":"TRIPOD-AI reporting guideline","url":"https://www.tripod-statement.org/","type":"other"},{"label":"CONSORT-AI reporting guideline","url":"http://www.consort-ai.org/","type":"other"}]}]},
{"date":"2025-10-24","headline":"2025 LLM release timeline updated in the LLM theory chapter","entries":[{"category":"developments","title":"2025 LLM release timeline updated in the LLM theory chapter","change":"Large Language Models in Public Health: Theory and Practice extends its 2025 model release timeline with entries from OpenAI, Anthropic, Google, and xAI, including GPT-5, Claude 4, Gemini 2.5, and Grok 4, extending coverage through October 2025.","why":"Keeps the chapter's survey of available models current as developers released new reasoning models throughout 2025.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"OpenAI, introducing o3 and o4-mini, 2025","url":"https://openai.com/index/introducing-o3-and-o4-mini/","type":"announcement"},{"label":"Anthropic, Claude Opus 4.1 announcement, 2025","url":"https://www.anthropic.com/news","type":"announcement"},{"label":"xAI, Grok 4 Fast announcement, 2025","url":"https://x.ai/news/grok-4-fast","type":"announcement"}]}]},
{"date":"2025-10-25","headline":"Limitations added to the behavioral AI chapter's wearable data discussion","entries":[{"category":"revision","title":"Limitations added to the behavioral AI wearables discussion","change":"AI-Driven Behavioral Interventions adds a note on the limits of physical activity tracking (step counts, sleep patterns, heart rate variability) and a Limitations section covering data quality, privacy, false positives and negatives, and generalizability across populations.","why":"Balances the chapter's discussion of wearable data with caveats about its reliability and privacy risks.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[]}]},
{"date":"2025-10-26","headline":"New AI Safety chapter: protecting patients and populations; Interactive quiz added to the AI Safety chapter","entries":[{"category":"milestone","title":"New AI Safety chapter: protecting patients and populations","change":"Adds a new chapter, AI Safety: Protecting Patients and Populations, covering safety frameworks (IEC 62304, ISO 14971), FDA, EU, and UK regulatory safety guidance, hazard analysis, and the Epic Sepsis Model failure as a case study.","why":"Introduces a dedicated framework for evaluating AI safety, separate from performance, ethics, and privacy, prompted by documented safety failures like Epic's sepsis prediction model.","chapters":[["AI Safety in Healthcare: Protecting Patients and Populations","/implementation/safety.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"],["Privacy, Security, and Governance for Health AI","/implementation/privacy.html"],["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"],["Building Your First Public Health AI Project","/practical/first-project.html"],["AI-Assisted Coding for Public Health Analysis","/practical/ai-coding.html"],["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Sendak et al., 2020, Nature Medicine","url":"https://www.nature.com/articles/s41591-019-0750-6","type":"paper"},{"label":"Leveson and Turner, 1993, An Investigation of the Therac-25 Accidents","url":"https://ieeexplore.ieee.org/document/274940","type":"paper"},{"label":"Amodei et al., 2016, Concrete Problems in AI Safety, preprint","url":"https://arxiv.org/abs/1606.06565","type":"preprint"},{"label":"FDA, Software as a Medical Device: Clinical Evaluation","url":"https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-evaluation-software","type":"policy"},{"label":"European Union, Medical Device Regulation 2017/745","url":"https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32017R0745","type":"policy"},{"label":"IEC 62304:2006+AMD1:2015, medical device software lifecycle standard","url":"https://www.iso.org/standard/38421.html","type":"other"}]},{"category":"content","title":"Interactive quiz added to the AI Safety chapter","change":"AI Safety: Protecting Patients and Populations adds a six-question, click-to-reveal quiz covering safety validation, autonomous AI requirements, incident response, FMEA versus traditional evaluation, FDA versus EU MDR requirements, and safety interlocks for a chemotherapy dosing AI.","why":"Lets readers test their understanding of the chapter's safety concepts.","chapters":[["AI Safety in Healthcare: Protecting Patients and Populations","/implementation/safety.html"]],"sources":[]}]},
{"date":"2025-10-28","headline":"Uganda mTRAC case study and LLM usage policy template added","entries":[{"category":"content","title":"Uganda mTRAC case study and LLM usage policy template added","change":"Global Health and Equity adds a case study on Uganda's mTRAC SMS-based disease surveillance system with AI anomaly detection, plus a practical action framework for implementing AI in resource-limited settings; Large Language Models in Public Health adds a sample LLM usage policy template for public health organizations.","why":"Gives readers a resource-limited-setting example and a ready-to-adapt policy template for governing LLM use.","chapters":[["AI in Global Health and Equity","/future/global-health.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Meng et al., 2022, Health Policy and Planning","url":"https://academic.oup.com/heapol/article/37/2/208/6419736","type":"paper"}]}]},
{"date":"2025-10-29","headline":"Check Your Understanding quizzes added to two chapters; AI tool selection and first-project failure modes","entries":[{"category":"content","title":"Check Your Understanding quizzes added to two chapters","change":"Just Enough AI to Be Dangerous and Evaluating AI Systems each add a Check Your Understanding quiz section with multiple-choice questions and explained answers.","why":"Lets readers test their grasp of key concepts before moving on.","chapters":[["Machine Learning Fundamentals","/foundations/ai-basics.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[]},{"category":"content","title":"AI tool selection and first-project failure modes","change":"Your AI Toolkit and Building Your First Project cover practical choices: the toolkit chapter frames tool selection philosophy for 2025, and the first-project chapter walks through common project failure modes, including scope creep, drawn from an analysis of abandoned projects.","why":"Sets realistic expectations before readers begin the chapter's hands-on work.","chapters":[["Your AI Toolkit for Public Health","/practical/toolkit.html"],["Building Your First Public Health AI Project","/practical/first-project.html"]],"sources":[]}]},
{"date":"2025-10-30","headline":"NHS COVID-19 app and OPTUM bias case studies added to the failures appendix; New appendices: AI Vendor Evaluation Checklist and AI Governance Policy Template; New appendix: Career Guide for Public Health AI Practitioners; Explainability and interpretability section added to the evaluation chapter; Emerging AI architectures section added to the LLM chapter","entries":[{"category":"content","title":"NHS COVID-19 app and OPTUM bias case studies added to the failures appendix","change":"The AI Morgue: Failure Post-Mortems adds case studies on the UK NHS COVID-19 contact-tracing app, downloaded by only 28% of the population against a 60% effectiveness target before being abandoned, and on racial bias in the OPTUM/UnitedHealth health-risk algorithm, each with root-cause analysis and a prevention checklist.","why":"Gives readers documented, citable failures to learn from before deploying similar digital health tools.","chapters":[["The AI Morgue: Failure Post-Mortems","/appendices/failures.html"]],"sources":[{"label":"Rowe et al., 2021, PLOS ONE","url":"https://doi.org/10.1371/journal.pone.0248250","type":"paper"},{"label":"The Guardian, NHS contact-tracing app cost, 2020","url":"https://www.theguardian.com/world/2020/sep/29/uk-nhs-covid-19-contact-tracing-app-has-cost-35m-so-far","type":"news"},{"label":"Parikh et al., 2019, JAMA","url":"https://doi.org/10.1001/jama.2019.18058","type":"paper"},{"label":"Vyas et al., 2020, NEJM","url":"https://doi.org/10.1056/NEJMms2004740","type":"paper"},{"label":"Rajkomar et al., 2018, Annals of Internal Medicine","url":"https://doi.org/10.7326/M18-1990","type":"paper"},{"label":"Ledford, Nature, 2019","url":"https://www.nature.com/articles/d41586-019-03228-6","type":"news"}]},{"category":"content","title":"New appendices: AI Vendor Evaluation Checklist and AI Governance Policy Template","change":"Two new appendices give readers practical tools: a six-domain framework for evaluating commercial AI vendors before purchase or deployment, and a customizable AI governance policy template for health departments and healthcare organizations.","why":"Gives readers ready-to-use tools for vendor evaluation and policy-writing instead of only conceptual guidance.","chapters":[["AI Vendor Evaluation Checklist","/appendices/vendor-evaluation.html"],["AI Governance Policy Template","/appendices/policy-template.html"]],"sources":[{"label":"AHRQ, Health IT Safety Toolkit","url":"https://www.ahrq.gov/patient-safety/resources/hitsafety/index.html","type":"policy"}]},{"category":"content","title":"New appendix: Career Guide for Public Health AI Practitioners","change":"A new appendix maps career paths in public health AI, covering major career tracks, required skills by experience level, and education pathways from certificates through PhD programs.","why":"Gives readers considering a career shift a structured map of roles, skills, and training options.","chapters":[["Career Guide: Pathways in Public Health AI","/appendices/career-guide.html"]],"sources":[]},{"category":"content","title":"Explainability and interpretability section added to the evaluation chapter","change":"Evaluating AI Systems adds a section on explainability methods including SHAP, LIME, and attention mechanisms, noting a study finding that 72% of clinicians would not trust or act on predictions from AI systems they could not understand, and reviewing FDA and EU AI Act perspectives on explainability requirements.","why":"Gives readers practical methods for making AI predictions interpretable to clinicians and regulators, not just accurate.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Wachter et al., 2019, NeurIPS, preprint","url":"https://arxiv.org/abs/1711.00399","type":"preprint"}]},{"category":"content","title":"Emerging AI architectures section added to the LLM chapter","change":"Large Language Models in Public Health: Theory and Practice adds a section on AI agents, vision-language models, and small language models, covering the ReAct planning-and-tool-use framework, medical image analysis safety considerations, and locally run efficient models.","why":"Extends the chapter beyond text-only chatbots to the multimodal and agentic systems increasingly used in 2025.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Yao et al., 2023, ICLR, preprint","url":"https://arxiv.org/abs/2210.03629","type":"preprint"},{"label":"Liu et al., 2023, NeurIPS, preprint","url":"https://arxiv.org/abs/2304.08485","type":"preprint"},{"label":"Touvron et al., 2023, Meta AI, preprint","url":"https://arxiv.org/abs/2307.09288","type":"preprint"}]}]},
{"date":"2025-10-31","headline":"License changed to CC BY-NC-ND 4.0","entries":[{"category":"content","title":"License changed to CC BY-NC-ND 4.0","change":"Welcome updates the handbook's license from MIT to Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).","why":"Clarifies that commercial use and derivative works require permission, while educational and non-commercial use remains free with attribution.","chapters":[["Welcome","/"]],"sources":[{"label":"Creative Commons, CC BY-NC-ND 4.0","url":"https://creativecommons.org/licenses/by-nc-nd/4.0/","type":"other"}]}]},
{"date":"2025-11-19","headline":"LLM release timeline expanded through late 2025 in the emerging technologies chapter","entries":[{"category":"revision","title":"LLM release timeline expanded through late 2025 in the emerging technologies chapter","change":"Emerging Technologies expands its language-model release timeline from 2020 through November 2025, covering GPT, Claude, Gemini, DeepSeek, Grok, and Llama model generations and each developer's reported benchmark results.","why":"Keeps the chapter's account of frontier model progress current as new models are released.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[]}]},
{"date":"2025-11-24","headline":"Global health chapter gets a TL;DR on the AI research and access gap in lower-income countries","entries":[{"category":"content","title":"Global health chapter TL;DR on the AI equity gap in low- and middle-income countries","change":"Global Health and Equity adds a chapter summary stating that 90% of AI health research comes from high-income countries even though 80% of the world's population lives in low- and middle-income countries, that only 3% of AI health datasets include data from Africa, and that fewer than 5% of FDA-approved AI medical devices are validated in low- and middle-income populations.","why":"Gives readers the chapter's core equity argument and supporting figures in one place.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[]}]},
{"date":"2025-11-25","headline":"LLM timeline updated with September through November 2025 model releases","entries":[{"category":"revision","title":"LLM timeline updated with September through November 2025 model releases","change":"The AI-assisted coding, emerging technologies, and LLM theory and practice chapters update their model timelines and current-frontier rankings with GPT-5.1, Grok 4.1, Gemini 3 Pro, and Claude Opus 4.5, each developer's most recent release as of late November 2025.","why":"Keeps the handbook's account of which models lead on coding and reasoning benchmarks current.","chapters":[["AI-Assisted Coding for Public Health Analysis","/practical/ai-coding.html"],["Emerging AI Technologies for Public Health","/future/emerging.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"OpenAI, GPT-5.1 announcement, 2025","url":"https://openai.com/index/gpt-5-1/","type":"announcement"},{"label":"xAI, Grok 4.1 announcement, 2025","url":"https://x.ai/news","type":"announcement"},{"label":"Google DeepMind, Gemini 3 Pro announcement, 2025","url":"https://blog.google/technology/google-deepmind/","type":"announcement"},{"label":"Anthropic, Claude Opus 4.5 announcement, 2025","url":"https://www.anthropic.com/claude/opus","type":"announcement"}]}]},
{"date":"2025-11-28","headline":"Skeptical assessment of quantum computing, digital twins, and AI drug discovery added","entries":[{"category":"revision","title":"Quantum computing, digital twins, and AI drug discovery get a critical look","change":"Emerging Technologies adds a skeptical assessment of quantum computing (current systems have only 100 to 1,000 error-prone qubits, with practical public health applications likely decades away), digital twins (mostly unvalidated), and AI-accelerated drug discovery, noting that AI can speed computational steps but cannot replace clinical trials.","why":"Separates genuinely useful emerging technologies from hype for readers deciding where to invest attention.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[]}]},
{"date":"2025-11-29","headline":"License changed to CC BY 4.0","entries":[{"category":"content","title":"License changed to CC BY 4.0","change":"Welcome and References & How to Cite update the handbook's license from CC BY-NC-ND 4.0 to Creative Commons Attribution 4.0 International (CC BY 4.0), permitting commercial use, adaptation, and remixing with attribution.","why":"Removes the non-commercial and no-derivatives restrictions, making the material easier for others to reuse and extend.","chapters":[["Welcome","/"],["References","/references.html"]],"sources":[{"label":"Creative Commons, CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/","type":"other"}]}]},
{"date":"2025-12-11","headline":"New executive summary","entries":[{"category":"content","title":"New executive summary","change":"Adds an executive summary covering key findings on AI in disease surveillance, genomic surveillance, and forecasting, stakeholder-specific recommendations, an evidence quality assessment, and priority investment areas.","why":"Gives readers a single, scannable summary of the handbook's evidence and recommendations.","chapters":[["Executive Summary","/executive-summary.html"]],"sources":[]}]},
{"date":"2025-12-12","headline":"GPT-5.2 added to model comparisons; New wastewater surveillance section in disease surveillance chapter; New model documentation standards section in ethics chapter; New climate-health AI section in emerging technologies; Evidence maturity framework added to three chapters; DHIS2 coverage and vendor evaluation checklist in global health","entries":[{"category":"developments","title":"GPT-5.2 added to model comparisons","change":"Adds OpenAI's GPT-5.2 (released December 11, 2025) to Emerging Technologies and the LLM chapter's model comparisons, noting its 400K token context window and developer-reported reductions of 38% in errors and about 30% in incorrect responses versus GPT-5.1, plus its three-tier pricing.","why":"OpenAI released a new frontier model, and the chapter's model comparisons and pricing needed to reflect it.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"OpenAI, GPT-5.2 announcement, 2025","url":"https://openai.com/index/introducing-gpt-5-2/","type":"announcement"}]},{"category":"content","title":"New wastewater surveillance section in disease surveillance chapter","change":"Adds a section on wastewater surveillance to Disease Surveillance and Outbreak Detection, covering CDC's National Wastewater Surveillance System, AI applications in normalization, forecasting, and variant tracking, limitations, and integration with clinical surveillance.","why":"Wastewater signals catch infections regardless of testing behavior, filling a gap that clinical and digital surveillance miss.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"CDC, National Wastewater Surveillance System","url":"https://www.cdc.gov/nwss/index.html","type":"policy"},{"label":"WastewaterSCAN","url":"https://wastewaterscan.org/","type":"other"},{"label":"Biobot Analytics","url":"https://biobot.io/","type":"other"}]},{"category":"content","title":"New model documentation standards section in ethics chapter","change":"Adds a section to Ethics, Bias, and Equity on Model Cards (Mitchell et al., 2019) and Datasheets for Datasets (Gebru et al., 2021), with templates, a sepsis-model example, and the case for transparency in public health AI.","why":"Gives health departments a standard way to document and evaluate a model's training data and limitations before deployment.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"Hugging Face, Model Cards documentation","url":"https://huggingface.co/docs/hub/model-cards","type":"other"}]},{"category":"content","title":"New climate-health AI section in emerging technologies","change":"Adds a Climate-Health AI section to Emerging Technologies covering heat wave prediction and response, vector-borne disease modeling, air quality and respiratory health, and the challenges of applying AI to environmental health threats.","why":"Extends the handbook's coverage into climate change as an emerging driver of public health risk.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"EPA, AirNow","url":"https://www.airnow.gov/","type":"policy"},{"label":"WHO, Climate Change and Health","url":"https://www.who.int/health-topics/climate-change","type":"policy"},{"label":"CDC, Climate and Health","url":"https://www.cdc.gov/climate-health/","type":"policy"},{"label":"Lancet Countdown on Health and Climate Change","url":"https://www.lancetcountdown.org/","type":"other"}]},{"category":"content","title":"Evidence maturity framework added to three chapters","change":"Adds a 'What We Know vs. What Remains Uncertain' framework to Emerging Technologies, the LLM chapter, and AI-Driven Behavioral Interventions, sorting claims into demonstrated, theoretical, and beyond-current-capabilities tiers.","why":"Helps readers separate what AI has shown it can do from what remains speculative.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[]},{"category":"content","title":"DHIS2 coverage and vendor evaluation checklist in global health","change":"Adds a section on DHIS2, the health information system used in over 80 countries, to Global Health and Equity, along with a one-page AI vendor evaluation checklist for resource-limited settings covering must-have requirements, red flags, and key vendor questions.","why":"Gives health departments with limited technical capacity a practical tool for assessing AI vendors.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"DHIS2","url":"https://dhis2.org/","type":"other"}]}]},
{"date":"2025-12-14","headline":"GPT-5.2 references updated in coding and LLM chapters","entries":[{"category":"revision","title":"GPT-5.2 references updated in coding and LLM chapters","change":"Updates ChatGPT and Cursor descriptions in AI-Assisted Coding and Development Tools to GPT-5.2 and updates the LLM chapter's model version note to December 2025, listing GPT-5.2, Claude Opus 4.5, Gemini 3 Pro, and Grok 4.1 with their release dates.","why":"Keeps the model comparisons current as newer model versions replace the ones previously described.","chapters":[["AI-Assisted Coding for Public Health Analysis","/practical/ai-coding.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[]}]},
{"date":"2025-12-16","headline":"New chapter on substance use AI; One Health section; adoption statistics added","entries":[{"category":"content","title":"New chapter on AI for substance use and overdose prevention","change":"Adds a new chapter covering EHR-based opioid risk screening, PDMP-integrated risk prediction, wastewater-based drug surveillance, digital therapeutics for addiction, and the 42 CFR Part 2 privacy framework governing substance use records.","why":"Gives public health practitioners an evidence-based look at where AI helps and where privacy and stigma limit its use in addiction response.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"],["Glossary","/appendices/glossary.html"],["Quick Reference: All Chapter Summaries (TL;DRs)","/appendices/tldr-compilation.html"]],"sources":[{"label":"Afshar et al., 2025, Nature Medicine","url":"https://www.nature.com/articles/s41591-025-03603-z","type":"paper"},{"label":"Lo-Ciganic et al., 2023","url":"https://pubmed.ncbi.nlm.nih.gov/37001323/","type":"paper"},{"label":"42 CFR Part 2","url":"https://www.ecfr.gov/current/title-42/chapter-I/subchapter-A/part-2","type":"policy"},{"label":"NIDA, HEAL Opioid Use Disorder and Overdose Strategic Plan","url":"https://nida.nih.gov/publications/2022-2026-nida-strategic-plan/heal-opioid-use-disorder-overdose-strategic-plan/nida-heal-opioid-use-disorder-overdose-strategic-plan-fy-2025","type":"policy"}]},{"category":"content","title":"New One Health section in the global health chapter","change":"Adds a section to Global Health and Equity on AI across the human-animal-environment interface, covering the NOHF-Zoonoses framework released by CDC, USDA, and DOI in January 2025, spillover risk prediction, and a case study of H5N1 in U.S. dairy cattle.","why":"The added section shows how AI can integrate zoonotic disease surveillance across the human, animal, and environmental health sectors that traditionally operate separately.","chapters":[["AI in Global Health and Equity","/future/global-health.html"],["Glossary","/appendices/glossary.html"]],"sources":[{"label":"CDC, USDA, and DOI, National One Health Framework to Address Zoonotic Diseases, 2025","url":"https://www.cdc.gov/one-health/media/pdfs/2025/01/354391-A-NOHF-ZOONOSES-508_FINAL.pdf","type":"policy"},{"label":"CDC, One Health","url":"https://www.cdc.gov/one-health/index.html","type":"policy"}]},{"category":"developments","title":"AI adoption statistics added to the welcome page","change":"Adds a section to the Welcome page citing the 2024 NACCHO Public Health Informatics Profile, showing 5% of local health departments use AI tools and 84% have no implementation plans, alongside CDC's reported deployment of 55 AI solutions and an estimated $3.7 million in labor cost savings.","why":"Gives readers a baseline picture of how far local health departments lag behind federal AI adoption.","chapters":[["Welcome","/"]],"sources":[{"label":"NACCHO, 2024 Public Health Informatics Profile","url":"https://www.naccho.org/programs/public-health-infrastructure/performance-improvement/informatics","type":"other"},{"label":"CDC, AI solutions deployment","url":"https://www.cdc.gov/data-modernization/php/ai/cdcs-vision-for-use-of-artificial-intelligence-in-public-health.html","type":"policy"}]}]},
{"date":"2025-12-21","headline":"Frequently Asked Questions appendix","entries":[{"category":"content","title":"Frequently Asked Questions appendix","change":"Adds a new Frequently Asked Questions appendix with 16 questions and answers covering getting started, evaluation and validation, red flags, data and technical issues, ethics, and careers, cross-referenced to the relevant chapters.","why":"Gives readers a single quick-reference page for common questions instead of searching individual chapters.","chapters":[["Frequently Asked Questions","/appendices/faq.html"]],"sources":[]}]},
{"date":"2025-12-22","headline":"Substance use AI: the crisis and application areas","entries":[{"category":"content","title":"Substance use AI: the crisis and application areas","change":"AI for Substance Use and Overdose Prevention describes the scale of the crisis, noting that overdose deaths have more than tripled since 2000 and surpassed peak annual HIV/AIDS deaths, and outlining the chapter's scope across prevention, identification, treatment, and surveillance.","why":"Connects the scale of the overdose crisis with AI applications in prevention, identification, treatment, and surveillance.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"sources":[]}]},
{"date":"2025-12-26","headline":"BEACON: an open-access alternative to BlueDot for outbreak detection","entries":[{"category":"developments","title":"BEACON open-access outbreak detection system","change":"Disease Surveillance and Outbreak Detection adds a section on BEACON, Boston University's freely accessible outbreak intelligence system, describing its AI web-scanning and PandemIQ language-model pipeline, human expert review step, and its 2025 output of over 1,000 outbreak reports covering more than 130 outbreaks, contrasted with BlueDot's proprietary model.","why":"Gives readers a free alternative to BlueDot's proprietary surveillance system, illustrating a different access model for outbreak intelligence.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"BEACON (Boston University), 2025","url":"https://beaconbio.org/","type":"other"}]},{"category":"developments","title":"McKinsey 2025 State of AI survey: the adoption-value gap","change":"Evaluating AI Systems for Healthcare adds findings from McKinsey's 2025 State of AI survey of 1,993 participants across 105 countries: 88% of organizations report using AI, but only 6% qualify as high performers with measurable earnings impact, and only 7% have fully scaled AI. The chapter lists five recurring barriers, including data quality, financial justification, and governance uncertainty.","why":"Gives readers survey evidence for the chapter's point that deployment success requires data infrastructure, training, and governance, not just adoption.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"McKinsey, The State of AI, 2025","url":"https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai","type":"report"}]},{"category":"evidence","title":"Google Population Dynamics Foundation Model for health mapping","change":"AI in Global Health and Equity adds Google's Population Dynamics Foundation Model (PDFM), which predicts local health and demographic patterns from satellite imagery, search trends, and geographic data via graph neural networks, with applications in measles vulnerability mapping, dengue forecasting, and respiratory disease tracking. The model is freely available on GitHub with a one-minute Colab notebook.","why":"Gives readers a free tool for health and demographic mapping in areas that lack complete data, including low-connectivity settings.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Google Research, Population Dynamics Foundation Model","url":"https://github.com/google-research/population-dynamics","type":"dataset"}]}]},
{"date":"2025-12-27","headline":"ML fundamentals expanded; WhatsApp health agents and edge AI techniques added","entries":[{"category":"content","title":"AI health agents on messaging platforms","change":"Emerging AI Technologies adds a section on AI health agents delivered via WhatsApp, covering the Business API architecture and applications in vaccination reminders, chronic disease management, maternal health, and outbreak response, plus a table of implementation considerations covering API access, data privacy, language support, escalation, and consent.","why":"Messaging platforms are described as the largest untapped channel for public health outreach, especially where WhatsApp serves as primary internet access.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Meta, WhatsApp Business Platform","url":"https://business.whatsapp.com/","type":"other"},{"label":"Meta, WhatsApp Cloud API","url":"https://developers.facebook.com/docs/whatsapp/cloud-api","type":"other"}]},{"category":"evidence","title":"Edge AI techniques: knowledge distillation and federated learning","change":"AI in Global Health and Equity adds a section on edge AI techniques: knowledge distillation, where a smaller student model captures 70 to 90% of a larger teacher model's performance, and federated learning, which trains models across distributed devices without centralizing patient data. A cited 2024 systematic review finds federated models can match centralized accuracy with stronger privacy protection.","why":"Edge AI techniques let AI run on local devices in settings with unreliable connectivity, without depending on the cloud.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"npj Digital Medicine, federated learning study","url":"https://www.nature.com/articles/s41746-021-00489-2","type":"paper"},{"label":"Systematic review on federated learning in healthcare, 2024","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10897620/","type":"paper"},{"label":"Keras, Knowledge Distillation example","url":"https://keras.io/examples/vision/knowledge_distillation/","type":"other"}]},{"category":"content","title":"Machine learning fundamentals: core concepts, calibration, and vendor questions","change":"Machine Learning Fundamentals adds a How Machine Learning Actually Works section covering models, loss functions, and gradient descent with a terminology map for epidemiologists, a calibration section on reliability diagrams and Brier score that warns against optimizing for AUC alone, and a checklist of questions to ask vendors and data scientists.","why":"Gives readers the working knowledge to evaluate AI systems, communicate with technical teams, and question vendor claims.","chapters":[["Machine Learning Fundamentals","/foundations/ai-basics.html"]],"sources":[{"label":"Hosmer and Lemeshow, Applied Logistic Regression","url":"https://doi.org/10.1002/0471722146","type":"paper"},{"label":"Mandrekar, 2010","url":"https://doi.org/10.1016/j.jtho.2010.05.018","type":"paper"},{"label":"Steyerberg et al., 2010","url":"https://doi.org/10.1097/EDE.0b013e3181c30fb2","type":"paper"},{"label":"Chawla et al., 2002","url":"https://doi.org/10.1613/jair.953","type":"paper"},{"label":"Stone, 1974","url":"https://doi.org/10.1111/j.2517-6161.1974.tb00994.x","type":"paper"},{"label":"van Rijsbergen, 1979","url":"https://www.dcs.gla.ac.uk/Keith/Preface.html","type":"other"}]}]},
{"date":"2025-12-28","headline":"WHO value chain governance model for AI in health; Open-source LMM governance and epistemic authority erosion; LMM ethics and the environmental cost of AI systems","entries":[{"category":"policy","title":"WHO value chain governance model for AI in health","change":"AI Policy and Governance adds a section on WHO's 2024 guidance for large multi-modal models, describing a three-phase value chain governance model spanning foundation model development, adaptation for healthcare, and clinical deployment, and noting the EU's proposed AI Liability Directive as a liability framework for the same value chain.","why":"Gives readers a governance framework for assigning responsibility across developers, adapters, and clinical deployers of AI models in health.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"WHO, Ethics and governance of artificial intelligence for health: guidance on large multi-modal models, 2024","url":"https://www.who.int/publications/i/item/9789240084759","type":"policy"},{"label":"European Parliament, briefing on the EU AI Liability Directive, 2024","url":"https://www.europarl.europa.eu/RegData/etudes/BRIE/2023/739342/EPRS_BRI(2023)739342_EN.pdf","type":"policy"}]},{"category":"policy","title":"Open-source LMM governance and epistemic authority erosion","change":"LLMs in Public Health adds a section on the governance benefits and considerations of open-source and open-weight models (Meta Llama, Mistral, DeepSeek), drawing on WHO's 2024 guidance, and a new discussion of epistemic authority erosion as a risk distinct from models generating false information.","why":"Applies WHO's 2024 guidance to open-source models and names a distinct risk to expertise and trust beyond models generating false information.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"WHO, Ethics and governance of artificial intelligence for health: guidance on large multi-modal models, 2024","url":"https://www.who.int/publications/i/item/9789240084759","type":"policy"}]},{"category":"policy","title":"LMM ethics and the environmental cost of AI systems","change":"Ethics, Bias, and Equity adds LMM-specific ethical considerations and a new section on the environmental sustainability of AI, covering the carbon and water cost of AI systems, ethical implications for public health, and mitigation strategies. It cites the WHO's six consensus principles for AI in health, published in 2021 and updated in 2024 with guidance for large multi-modal models.","why":"Extends the chapter's ethics framework to LMM-specific risks and to AI's material and environmental costs, not just its outputs.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"WHO, AI ethics and governance guidance for large multi-modal models, 2024","url":"https://www.who.int/news/item/18-01-2024-who-releases-ai-ethics-and-governance-guidance-for-large-multi-modal-models","type":"policy"}]}]},
{"date":"2025-12-29","headline":"HHS AI Use Case Inventory and CDC's ChatCDC tool","entries":[{"category":"policy","title":"HHS AI Use Case Inventory and CDC's ChatCDC tool","change":"LLMs in Public Health adds a Federal Government Adoption note describing CDC's ChatCDC, an internal Azure OpenAI-powered tool deployed in 2024, and HHS's FY2024 inventory of 271 AI use cases (66% growth from FY2023) with at least 24 chatbot implementations. Further Reading adds an entry linking to the HHS AI Use Case Inventory, updated annually under Executive Order 13960.","why":"Gives readers a primary source for the scale of federal AI adoption in health agencies, updated annually.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["Further Reading","/appendices/resources.html"]],"sources":[{"label":"HHS, AI Use Case Inventory, FY2024","url":"https://www.healthit.gov/hhs-ai-usecases","type":"policy"}]}]},
{"date":"2025-12-30","headline":"Evidence-based misinformation debunking strategies","entries":[{"category":"evidence","title":"Evidence-based misinformation debunking strategies","change":"AI, Misinformation, and the Infodemic expands its section on evidence-based debunking, citing research showing that repeating a false claim to correct it can backfire (Swire et al., 2017) and that the messenger delivering health information matters as much as the message itself (Freimuth et al., 2014).","why":"Gives readers research-backed guidance for correcting health misinformation without unintentionally reinforcing it.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[{"label":"Lewandowsky et al., 2012","url":"https://journals.sagepub.com/doi/10.1177/1529100612451018","type":"paper"},{"label":"Swire et al., 2017","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC7462781/","type":"paper"},{"label":"Freimuth et al., 2014","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC3943629/","type":"paper"},{"label":"Lakoff, 2018","url":"https://www.futurehindsight.com/blog/george-lakoffs-truth-sandwich","type":"other"},{"label":"Lewandowsky et al., 2020","url":"https://digitalcommons.unl.edu/scholcom/245/","type":"other"}]}]},
{"date":"2025-12-31","headline":"Foundational techniques behind large language models","entries":[{"category":"revision","title":"Foundational techniques behind large language models","change":"LLMs in Public Health describes large language models as bringing natural language interfaces to complex computation, and names three foundational techniques behind current systems: ReAct (Yao et al., 2023), visual instruction tuning (Liu et al., 2023), and Llama 2 (Touvron et al., 2023).","why":"Grounds the chapter's coverage of LLMs in the specific techniques and models that made them practical.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[]}]},
{"date":"2026-01-01","headline":"PDF edition of the handbook","entries":[{"category":"milestone","title":"PDF edition of the handbook","change":"Enables a formatted PDF edition of the handbook, with bookmarks, headers, and footers, for archival and citation.","why":"Gives readers a citable, archival format of the handbook alongside the website.","chapters":[],"sources":[]}]},
{"date":"2026-01-04","headline":"Expanded tool comparisons in the AI toolkit chapter; New definitions in the evaluation chapter; Executive Order 14110 rescission; December 2025 executive order on federal AI preemption","entries":[{"category":"revision","title":"Expanded tool comparisons in the AI toolkit chapter","change":"Your AI Toolkit for Public Health adds comparisons of virtual environments, Polars versus pandas, XGBoost, PyTorch, TensorFlow, Seaborn, Plotly, Weights and Biases, DVC, and Docker for public health data science workflows.","why":"Gives readers a broader, more current set of tool options for building public health AI projects.","chapters":[["Your AI Toolkit for Public Health","/practical/toolkit.html"]],"sources":[]},{"category":"revision","title":"New definitions in the evaluation chapter","change":"Evaluating AI Systems for Healthcare adds definitions distinguishing precision-recall curves from ROC curves for imbalanced datasets, and explains retrieval-augmented generation and software as a medical device.","why":"Clarifies technical evaluation concepts referenced elsewhere in the handbook.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[]},{"category":"policy","title":"Executive Order 14110 rescission","change":"AI Governance Policy Template notes that Executive Order 14110, which mandated safety testing for high-risk AI systems, was rescinded by Executive Order 14179 in January 2025, shifting federal policy toward deregulation, and clarifies that FDA device regulations and state laws remain in effect.","why":"Keeps the policy guidance current with the change in federal AI policy.","chapters":[["AI Governance Policy Template","/appendices/policy-template.html"]],"sources":[{"label":"White House, Executive Order 14179, 2025","url":"https://www.govinfo.gov/content/pkg/FR-2025-01-31/pdf/2025-02172.pdf","type":"policy"}]},{"category":"policy","title":"December 2025 executive order on federal AI preemption","change":"AI Governance Policy Template adds a Federal Preemption Efforts section on the December 2025 Executive Order on National AI Policy Framework, which directs DOJ challenges to state AI laws, ties broadband funding to repealing them, and has the FTC address state AI output laws, while noting its legal status is unclear absent Congressional action.","why":"Explains a new federal action that creates legal uncertainty for the state AI laws the chapter describes.","chapters":[["AI Governance Policy Template","/appendices/policy-template.html"]],"sources":[{"label":"White House, Executive Order on National AI Policy Framework, 2025","url":"https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/","type":"policy"},{"label":"NPR, 2025","url":"https://www.npr.org/2025/12/11/nx-s1-5638562/trump-ai-david-sacks-executive-order","type":"news"},{"label":"Goodwin Law, 2025","url":"https://www.goodwinlaw.com/en/insights/publications/2025/12/alerts-otherindustries-trumps-ai-preemption-executive-order","type":"other"}]}]},
{"date":"2026-01-08","headline":"OpenAI for Healthcare launches; scrutiny of its benchmark and study claims","entries":[{"category":"developments","title":"OpenAI for Healthcare launch","change":"LLMs in Public Health adds a section on OpenAI for Healthcare, launched January 8, 2026, describing its evidence retrieval features, institutional policy integration, and early adopters including Boston Children's Hospital, Cedars-Sinai, and UCSF.","why":"Covers a new enterprise AI product targeting clinical, research, and administrative workflows in health systems.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"OpenAI, OpenAI for Healthcare announcement, 2026","url":"https://openai.com/index/openai-for-healthcare/","type":"announcement"}]},{"category":"evidence","title":"HealthBench conflict-of-interest critique","change":"LLMs in Public Health notes that HealthBench, the benchmark OpenAI cites to show model superiority, was created by OpenAI researchers, and cites independent critique that strong benchmark performance may not translate into diagnostic accuracy, workflow efficiency, or patient safety.","why":"Flags a potential conflict of interest in a benchmark used to support vendor performance claims.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"OpenAI, HealthBench, 2025, preprint","url":"https://arxiv.org/abs/2505.08775","type":"preprint"},{"label":"Independent critique of HealthBench, PMC, 2025","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC12547120/","type":"paper"}]},{"category":"evidence","title":"GDPval benchmark results scrutinized","change":"LLMs in Public Health notes that OpenAI's claim that GPT-5.2 performs better than human baselines across every role does not match the underlying GDPval results, which show a 40.6% win-or-tie rate against expert human work, and that only 4 of 44 evaluated occupations were healthcare-related.","why":"Tests a vendor performance claim against the primary benchmark data and its limited healthcare coverage.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"OpenAI, GDPval results, 2025, preprint","url":"https://arxiv.org/abs/2510.04374","type":"preprint"}]},{"category":"evidence","title":"Penda Health AI safety net study in Kenya","change":"LLMs in Public Health cites a preprint reporting a 16% reduction in diagnostic errors and 13% reduction in treatment errors across 39,849 patient visits in Kenya, and notes that two patient deaths occurred when AI alerts were ignored and over 35% of critical safety warnings initially went unheeded.","why":"Surfaces safety details from the underlying study that a vendor's promotional description left out.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Penda Health study, Kenya, 2025, preprint","url":"https://arxiv.org/html/2507.16947v1","type":"preprint"},{"label":"STAT News, 2025","url":"https://www.statnews.com/2025/10/01/penda-health-open-ai-safety-net-study-kenya-artificial-intelligence/","type":"news"}]},{"category":"developments","title":"Ambient clinical documentation vendor landscape","change":"LLMs in Public Health surveys ambient scribe vendors, including Microsoft DAX Copilot (deployed in over 150 hospitals, integrated into Epic), Abridge (proprietary models, reported gains in workflow ease and reduced cognitive load), and AWS HealthScribe.","why":"Gives readers a comparison point for evaluating ambient documentation products, with vendor-reported figures labeled as such.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Healthcare IT News, 2025","url":"https://www.healthcareitnews.com/news/nuance-ai-copilot-now-fully-embedded-epic-ehr","type":"news"},{"label":"Microsoft, DAX Copilot product page","url":"https://www.nuance.com/healthcare/dragon-ambient-experience.html","type":"other"},{"label":"JAMIA study, via PMC","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11843214/","type":"paper"},{"label":"Abridge, Mayo Clinic Proceedings summary, 2025","url":"https://www.abridge.com/blog/kumc-research-studies","type":"announcement"},{"label":"AWS, HealthScribe product page","url":"https://aws.amazon.com/healthscribe/","type":"announcement"}]},{"category":"developments","title":"OpenAI-linked healthcare AI startups and a vendor clarification","change":"LLMs in Public Health reports that Ambience Healthcare (backed by the OpenAI Startup Fund) reported a Cleveland Clinic rollout with burnout decreasing from 51.9% to 38.8%, that EliseAI uses OpenAI APIs for front-desk automation, and clarifies that Abridge, despite appearing in OpenAI's announcement, uses proprietary models rather than OpenAI's.","why":"Helps readers distinguish vendors' actual technology relationships from marketing associations with OpenAI.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Ambience Healthcare, Series C announcement, 2025","url":"https://www.ambiencehealthcare.com/blog/ambience-healthcare-announces-243-million-series-c-to-scale-its-ai-platform-for-health-systems","type":"announcement"},{"label":"OpenAI, EliseAI announcement, 2025","url":"https://openai.com/index/eliseai-minna-song/","type":"announcement"},{"label":"Contrary Research, Abridge company profile, 2025","url":"https://research.contrary.com/company/abridge","type":"other"}]},{"category":"policy","title":"FDA regulatory shift on AI clinical decision support","change":"LLMs in Public Health notes that ambient AI documentation tools are generally not FDA-regulated when used only for documentation, references the FDA's January 2025 draft guidance on AI-enabled devices, and reports that in January 2026 the FDA announced it will ease regulation of clinical decision support software meeting certain criteria.","why":"Explains the regulatory boundary for ambient documentation tools and a coming shift in FDA oversight.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"FDA, January 2025 draft guidance on AI-enabled devices","url":"https://www.greenlight.guru/blog/fda-guidance-ai-enabled-devices","type":"other"},{"label":"STAT News, 2026","url":"https://www.statnews.com/2026/01/06/fda-pulls-back-oversight-ai-enabled-devices-wearables/","type":"news"}]}]},
{"date":"2026-01-11","headline":"Anthropic launches Claude for Healthcare; independent benchmarks scrutinize its claims","entries":[{"category":"developments","title":"Anthropic Claude for Healthcare and Life Sciences launch","change":"LLMs in Public Health adds a section on Claude for Healthcare and an expanded Claude for Life Sciences, announced January 11, 2026, describing healthcare connectors (CMS coverage database, ICD-10, NPI registry, PubMed), life sciences connectors for clinical trial and drug discovery data, agent skills for FHIR development and prior authorization review, and HIPAA-ready compliance via Claude for Enterprise.","why":"Covers a new enterprise AI product aimed at clinical, payer, and life sciences research workflows.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Anthropic, Claude for Healthcare announcement, 2026","url":"https://www.anthropic.com/news/healthcare-life-sciences","type":"announcement"},{"label":"Anthropic, Claude for Life Sciences, 2026","url":"https://www.anthropic.com/solutions/life-sciences","type":"announcement"}]},{"category":"evidence","title":"Independent benchmarks for Anthropic's healthcare claims","change":"LLMs in Public Health cites MedAgentBench, a Stanford benchmark testing LLM agents on 300 clinical EHR tasks, on which Claude 3.5 Sonnet v2 achieved a 69.67% success rate (best among tested models but limited to Stanford Hospital records), and SpatialBench, a spatial biology benchmark on which base model accuracy ranged from 20% to 38% across model families.","why":"Gives readers independent, third-party performance data alongside the vendor's own claims.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"NEJM AI, MedAgentBench, August 2025","url":"https://ai.nejm.org/doi/full/10.1056/AIdbp2500144","type":"paper"},{"label":"Stanford ML Group, MedAgentBench dataset","url":"https://github.com/stanfordmlgroup/MedAgentBench","type":"dataset"},{"label":"LatchBio, SpatialBench dataset","url":"https://github.com/latchbio/spatialbench","type":"dataset"}]}]},
{"date":"2026-01-14","headline":"Three types of AI-driven behavioral interventions; Emerging AI Technologies chapter expanded; Digital divide framework and India retinopathy screening case study; Career guide market data and salary tracks; New glossary definitions for bias, fairness, and surveillance; Regulatory framework comparison table added to policy chapter; Memorization risk in healthcare foundation models; Glossary definitions for differential privacy and Google's MedGemma","entries":[{"category":"revision","title":"Three types of AI-driven behavioral interventions","change":"AI-Driven Behavioral Interventions covers three intervention types (chatbot-delivered CBT tools, reinforcement learning for message timing, and NLP-based sentiment analysis) with reported effectiveness ranges, including 10 to 20% better outcomes than wait-list controls for chatbot programs and 15 to 30 percent adherence gains from reinforcement learning timing.","why":"Compares intervention types and their reported effectiveness.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[]},{"category":"revision","title":"Emerging AI Technologies chapter expanded","change":"Emerging AI Technologies adds a Technology Readiness Levels framework, a timeline of predictions about AI limits that were later disproven by newer models, and a section on multimodal foundation models processing text, images, and audio together, including scientific models such as BioGPT and MedPaLM 2.","why":"Gives readers a framework for judging how mature an emerging AI capability actually is.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[]},{"category":"content","title":"Digital divide framework and India retinopathy screening case study","change":"AI in Global Health and Equity adds a four-layer digital divide framework (connectivity, computing, data, and human capacity, each with figures such as 2.9 billion people lacking internet access), edge AI techniques for low-resource deployment, and a case study reporting diabetic retinopathy screening in India at 95% sensitivity and $1.50 per screening across more than 300,000 patients.","why":"Gives readers concrete figures and a real deployment example for AI in resource-constrained health settings.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[]},{"category":"content","title":"Career guide market data and salary tracks","change":"Career Guide adds market context (36% projected job growth for healthcare data scientists, a $7.4 billion CDC Data Modernization Initiative, and a global health AI market projected at $188 billion by 2030) and a table of four career tracks and emerging roles with salary ranges.","why":"Gives readers concrete market and compensation data for planning a career in public health AI.","chapters":[["Career Guide: Pathways in Public Health AI","/appendices/career-guide.html"]],"sources":[]},{"category":"content","title":"New glossary definitions for bias, fairness, and surveillance","change":"Glossary adds definitions for bias, fairness, surveillance, evaluation metrics, and federated learning, each cross-referenced to the chapter with fuller discussion.","why":"Gives readers quick, linked definitions for terms used throughout the handbook.","chapters":[["Glossary","/appendices/glossary.html"]],"sources":[]},{"category":"content","title":"Regulatory framework comparison table added to policy chapter","change":"AI Policy and Governance adds a comparison table of FDA, EU AI Act, UK MHRA, and IMDRF regulatory approaches, details the EU AI Act's four risk categories, and outlines the WHO's three-phase value chain governance approach.","why":"Gives readers a side-by-side reference for how major regulators classify and govern health AI.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[]},{"category":"evidence","title":"Memorization risk in healthcare foundation models","change":"Privacy, Security, and Governance cites Tonekaboni et al. (2025), reporting that healthcare foundation models trained on electronic health records show memorization patterns similar to other models, with patients who have rare conditions especially vulnerable.","why":"Adds evidence that patient-level memorization risk extends to healthcare-specific foundation models.","chapters":[["Privacy, Security, and Governance for Health AI","/implementation/privacy.html"]],"sources":[{"label":"Tonekaboni et al., 2025, preprint","url":"https://arxiv.org/abs/2510.12950","type":"preprint"}]},{"category":"content","title":"Glossary definitions for differential privacy and Google's MedGemma","change":"Glossary adds definitions for differential privacy as a defense against memorization in healthcare foundation models, and for Google's Health AI Developer Foundations program and its MedGemma model family, noting MedGemma 1.5's support for CT, MRI, and histopathology interpretation.","why":"Gives readers definitions for privacy techniques and named open medical AI models referenced elsewhere in the handbook.","chapters":[["Glossary","/appendices/glossary.html"]],"sources":[]}]},
{"date":"2026-01-15","headline":"Zenodo DOI registered for the handbook; Underfitting and overfitting diagram added; System prompt governance: physician persona effects on LLM outputs","entries":[{"category":"milestone","title":"Zenodo DOI registered for the handbook","change":"The handbook is archived on Zenodo with a persistent DOI (10.5281/zenodo.18263442), added to the Welcome page and References & How to Cite.","why":"Gives the handbook a permanent, citable identifier.","chapters":[["Welcome","/"],["References","/references.html"],["AI Governance Policy Template","/appendices/policy-template.html"],["Case Study Library - Overview","/appendices/case-study-overview.html"]],"sources":[{"label":"Public Health AI Handbook archive, Zenodo, 2025","url":"https://zenodo.org/records/18263442","type":"dataset"}]},{"category":"content","title":"Underfitting and overfitting diagram added","change":"Machine Learning Fundamentals adds a figure illustrating underfitting, a well-fit model, and overfitting, sourced from Wikimedia Commons.","why":"Gives readers a visual reference for a core machine learning concept.","chapters":[["Machine Learning Fundamentals","/foundations/ai-basics.html"]],"sources":[]},{"category":"evidence","title":"System prompt governance: physician persona effects on LLM outputs","change":"LLMs in Public Health adds a System Prompt Governance section explaining that system prompts are hidden instructions set at the deployment layer, and cites a Mount Sinai study that tested 20 LLMs across 5 million clinical decisions and found that assigning different physician personas changed model behavior.","why":"Shows practitioners that decisions made by whoever configures an LLM's system prompt, not just the model itself, can shape clinical outputs.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Klang et al., 2026, preprint","url":"https://doi.org/10.1101/2026.01.01.26343302","type":"preprint"}]}]},
{"date":"2026-01-17","headline":"Open-weight AI, global health LLMs, and clinical AI adoption data","entries":[{"category":"content","title":"MomConnect case study: South Africa's LLM-enhanced maternal health platform","change":"Case Studies adds Case Study 16: MomConnect, South Africa's national maternal health platform, reporting 5 million registered users since its 2014 launch, 288,051 monthly active users in 2024, integration with 95% of public health facilities, and 40,000 to 60,000 health questions handled monthly through its LLM-enhanced chatbot.","why":"Gives readers a long-running, national-scale example of AI-assisted maternal health messaging.","chapters":[["Case Study Library - Full Cases","/appendices/case-studies.html"]],"sources":[{"label":"Reach Digital Health, MomConnect 10-year milestone, 2024","url":"https://reachdigitalhealth.medium.com/momconnect-celebrates-10-years-of-transforming-maternal-and-child-health-in-south-africa-b64103c890e1","type":"post"},{"label":"IDInsight, AI chatbot improvements, 2024","url":"https://www.idinsight.org/article/using-ai-to-improve-maternal-health-chatbots-in-south-africa/","type":"other"},{"label":"Hypertext, Google AI integration, 2025","url":"https://htxt.co.za/2025/08/google-leverages-its-ai-to-assist-local-platform-momconnect/","type":"other"},{"label":"Coleman et al., 2017, BMC Pregnancy and Childbirth","url":"https://bmcpregnancychildbirth.biomedcentral.com/articles/10.1186/s12884-017-1337-z","type":"paper"}]},{"category":"evidence","title":"Open-weight model benchmarks: DeepSeek-R1 versus proprietary models","change":"Emerging AI Technologies reports that DeepSeek-R1 scored 0.92 on the US Medical Licensing Examination versus 0.95 for OpenAI's o1, matched o1's 82.0% accuracy on 300 ophthalmology cases at roughly 15 times lower inference cost, and notes that adversarial system prompts caused both open-weight and closed-source models to produce misleading health information.","why":"Gives readers cost and performance evidence for open-weight clinical AI models alongside their safety tradeoffs.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Sandmann et al., Nature Medicine, 2025","url":"https://doi.org/10.1038/s41591-025-03727-2","type":"paper"},{"label":"medRxiv, 2025, preprint","url":"https://www.medrxiv.org/content/10.1101/2025.02.10.25322041v1","type":"preprint"},{"label":"Yang et al., JAMA Network Open, 2024","url":"https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2826474","type":"paper"},{"label":"Ong et al., Nature Health, 2026","url":"https://www.nature.com/articles/s44324-025-00017-2","type":"paper"}]},{"category":"policy","title":"Workforce and employment implications of AI in global health","change":"AI in Global Health and Equity adds a section on AI's impact on healthcare employment, covering exposure by income level, short-term versus long-term effects, jobs at risk, wage dynamics in telemedicine, and workforce preparation, citing the ILO's 2025 analysis of AI exposure in employment.","why":"Extends the chapter's equity focus from AI access to AI's effects on the health workforce.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"ILO, Generative AI and Jobs, 2025","url":"https://www.ilo.org/publications/generative-ai-and-jobs-2025-update","type":"policy"}]},{"category":"evidence","title":"DeepDR-LLM: AI-language model hybrid for diabetic retinopathy care in China","change":"AI in Global Health and Equity describes DeepDR-LLM, a system combining retinal image analysis with a language model fine-tuned on 371,763 real-world management recommendations, reporting a prospective study of 769 patients over 12 weeks that found ophthalmology referral rates of 77.78% versus 58.44% and primary care diagnostic accuracy rising from 81.0% to 92.3% with AI support.","why":"Gives readers outcome data, not just accuracy metrics, for an AI system already integrated into primary care.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Li et al., Nature Medicine, 2024","url":"https://doi.org/10.1038/s41591-024-03139-8","type":"paper"}]},{"category":"evidence","title":"LLMs in global public health: open-weight deployment and validation concerns","change":"AI in Global Health and Equity adds a section on large language models in global public health, noting that DeepSeek has been deployed in over 300 healthcare facilities in China since January 2025 for clinical decision support at lower cost than proprietary models, and that researchers have raised concerns about deploying LLMs for surveillance or policy use without prospective validation.","why":"Weighs the cost advantage of open-weight LLMs for resource-limited health systems against the risk of deploying them without adequate validation.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Chen et al., Journal of Medical Systems, 2025","url":"https://link.springer.com/article/10.1007/s10916-025-02181-4","type":"paper"},{"label":"Wong et al., JAMA, 2025","url":"https://pubmed.ncbi.nlm.nih.gov/40560255/","type":"paper"},{"label":"Ong et al., Nature Health, 2026","url":"https://www.nature.com/articles/s44360-025-00024-7","type":"paper"}]},{"category":"evidence","title":"AfriMed-QA: benchmark for LLMs on African healthcare contexts","change":"AI in Global Health and Equity adds AfriMed-QA, described as the first large-scale, pan-African, multi-specialty medical question-answering dataset for evaluating LLM performance in African healthcare contexts, publicly available for health departments and researchers to test models before deployment.","why":"Gives readers a way to check LLM performance on a population the largest benchmarks often exclude.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Olatunji et al., ACL 2025, preprint","url":"https://arxiv.org/abs/2411.15640","type":"preprint"},{"label":"AfriMed-QA dataset","url":"https://afrimedqa.com/","type":"dataset"}]},{"category":"policy","title":"Global governance initiatives for public health AI","change":"AI in Global Health and Equity describes the WHO, ITU, and WIPO Global Initiative on AI for Health (GI-AI4H), launched in July 2023 to coordinate responsible AI development, alongside an ongoing Lancet Global Health Commission on AI's economic and health impacts and Gates Foundation funding for AI equity and language inclusivity.","why":"Gives readers the institutions and funding shaping global public health AI governance.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"WHO, ITU, and WIPO, Global Initiative on AI for Health, 2023","url":"https://www.who.int/initiatives/global-initiative-on-ai-for-health","type":"policy"},{"label":"Lancet Global Health Commission on AI","url":"https://www.thelancet.com/commissions/global-health","type":"report"},{"label":"Gates Foundation, 2024-2026","url":"https://www.gatesfoundation.org/","type":"other"}]},{"category":"developments","title":"Modern clinical decision support tools: adoption among physicians","change":"Diagnostic and Clinical Decision Support adds a section on LLM-based clinical decision support tools, reporting that OpenEvidence, launched in 2024, reached 45% daily use among U.S. physicians by late 2025, handles over 18 million consultations monthly, and is integrated at UCSF as a SMART on FHIR app within Epic.","why":"Gives readers vendor-reported adoption figures for a fast-growing category of clinical AI tools, labeled as such.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"OpenEvidence product site","url":"https://www.openevidence.com/","type":"other"},{"label":"OpenEvidence, funding announcement, December 2025","url":"https://www.openevidence.com/announcements/openevidence-the-fastest-growing-application-for-physicians-in-history-announces-dollar210-million-round-at-dollar35-billion-valuation","type":"announcement"},{"label":"Axios, December 2025","url":"https://www.axios.com/pro/all-deals/2025/12/19/clinical-decision-support-openevidence-doximity","type":"news"},{"label":"UCSF Open Proposals, 2025","url":"https://open-proposals.ucsf.edu/Artificial-Intelligence-Machine-Learning-Demo/2025/proposal/19434","type":"other"},{"label":"MedCity News, September 2025","url":"https://medcitynews.com/2025/09/clinical-decision-support-healthcare-ai/","type":"news"}]},{"category":"evidence","title":"Independent evidence of a clinical AI benchmark-reality gap","change":"Diagnostic and Clinical Decision Support and Evaluating AI Systems for Healthcare cite the Stanford-Harvard ARISE State of Clinical AI Report 2026, which reviewed more than 500 clinical AI studies and found that benchmark performance does not predict real-world accuracy.","why":"Warns readers that strong benchmark scores for clinical AI tools may not hold up in real-world use.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Stanford-Harvard ARISE, State of Clinical AI Report, 2026","url":"https://www.arise-ai.org/report","type":"report"},{"label":"Stanford Medicine News, 2026","url":"https://medicine.stanford.edu/news/current-news/standard-news/clinical-ai-has-boomed.html","type":"news"}]},{"category":"developments","title":"OpenAI's consumer health AI usage and ChatGPT for Health","change":"LLMs in Public Health adds a section reporting that 40 million people globally use ChatGPT daily for health questions and about 230 million people ask health questions weekly, citing OpenAI's January 2026 report; Diagnostic and Clinical Decision Support notes OpenAI's January 2026 announcement of ChatGPT for Health as a health education tool without clinical validation or regulatory oversight.","why":"Shows the scale of consumer reliance on general-purpose AI for health information outside clinical settings.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"OpenAI, AI as a Healthcare Ally report, January 2026","url":"https://cdn.openai.com/pdf/2cb29276-68cd-4ec6-a5f4-c01c5e7a36e9/OpenAI-AI-as-a-Healthcare-Ally-Jan-2026.pdf","type":"announcement"}]}]},
{"date":"2026-01-22","headline":"VIM framework for AI value disclosure; OpenEvidence adoption and AMA physician survey","entries":[{"category":"evidence","title":"Values In the Model (VIM) framework for AI value disclosure","change":"Ethics, Bias, and Equity adds the RAISE consortium's VIM framework (Goldberg et al., NEJM AI, 2026), which calls for AI systems to disclose their allocative values, fairness trade-offs, threshold decisions, and edge case handling; Further Reading adds the RAISE Consortium as a resource.","why":"Gives readers a framework for documenting the value judgments embedded in AI systems, beyond standard model cards and fairness audits.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"],["Further Reading","/appendices/resources.html"]],"sources":[{"label":"Goldberg et al., NEJM AI, 2026","url":"https://doi.org/10.1056/AIp2501266","type":"paper"},{"label":"Harvard Bioethics, RAISE symposium","url":"https://bioethics.hms.harvard.edu/events/2nd-international-symposium-responsible-ai-social-ethical-healthcare-raise","type":"other"}]},{"category":"developments","title":"OpenEvidence adoption figures and AMA physician survey on AI tools","change":"Diagnostic and Clinical Decision Support now cites OpenEvidence's developer-reported figures of over 40% daily use among U.S. physicians and over 8.5 million monthly consultations, and an AMA survey finding 47% of physicians rank oversight as their top AI regulatory priority and 87% cite liability protection as critical for adoption.","why":"Adoption figures and physicians' stated priorities for oversight and liability protection are primary evidence on how far physicians use and trust AI tools in clinical decisions.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"AMA, February 2025","url":"https://www.ama-assn.org/practice-management/digital-health/2-3-physicians-are-using-health-ai-78-2023","type":"report"}]}]},
{"date":"2026-01-25","headline":"Deferral as a safety mechanism for healthcare AI","entries":[{"category":"evidence","title":"Deferral as a safety mechanism for healthcare AI","change":"AI Safety in Healthcare adds a section on deferral as a safety mechanism, arguing systems should recognize their limits and defer to human judgment rather than answer with false confidence, and adds under-triage, omission, and unsafe non-deferral rate metrics. The Glossary and Evaluating AI Systems for Healthcare add matching coverage citing Azad et al., Nature Medicine, 2026.","why":"Gives readers a framework and named metrics for when an AI system should decline to answer rather than respond with false confidence.","chapters":[["AI Safety in Healthcare: Protecting Patients and Populations","/implementation/safety.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["Glossary","/appendices/glossary.html"]],"sources":[{"label":"Azad et al., Nature Medicine, 2026","url":"https://doi.org/10.1038/s41591-025-04198-1","type":"paper"}]}]},
{"date":"2026-01-26","headline":"FDA AI/ML medical device authorization count","entries":[{"category":"policy","title":"FDA AI/ML medical device authorization count","change":"AI in Healthcare: A Brief History adds that over 1,300 AI/ML-enabled medical devices had received FDA authorization by late 2025, with about 77% in radiology, alongside the existing example of Google's diabetic retinopathy screening deployed in India and Thailand.","why":"Grounds the chapter's radiology-dominant FDA authorization claim in the FDA's own device list.","chapters":[["AI in Healthcare: A Brief History","/foundations/history.html"]],"sources":[{"label":"FDA AI-Enabled Medical Devices","url":"https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices","type":"policy"}]}]},
{"date":"2026-01-28","headline":"RAG evidence review, FUTURE-AI framework, and TRIPOD-LLM reporting standard added","entries":[{"category":"evidence","title":"Evidence review of retrieval-augmented generation in healthcare","change":"LLMs in Public Health adds a section on retrieval-augmented generation (RAG), citing the first review of RAG in healthcare (Ng et al., 2025, NEJM AI), which found RAG reduces ungrounded answers by 30% to 50% versus a base LLM for clinical Q&A, though retrieval quality remains a bottleneck.","why":"Gives readers a primary evidence base for RAG's benefits and limitations across clinical Q&A, literature synthesis, guidelines navigation, and patient education.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Ng et al., 2025, NEJM AI","url":"https://doi.org/10.1056/AIra2400336","type":"paper"}]},{"category":"evidence","title":"FUTURE-AI international consensus framework for trustworthy clinical AI","change":"Ethics, Bias, and Equity adds the FUTURE-AI framework (Lekadir et al., 2025, BMJ), developed by 117 experts across 50 countries through a Delphi process, covering Fairness, Universality, Traceability, Usability, Robustness, and Explainability with 30 best practices across the AI development lifecycle.","why":"Gives readers an internationally developed consensus framework for building and auditing trustworthy clinical AI.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"Lekadir et al., 2025, BMJ","url":"https://doi.org/10.1136/bmj-2024-081554","type":"paper"}]},{"category":"evidence","title":"TRIPOD-LLM reporting guidelines and the LLM reporting crisis","change":"Evaluating AI Systems for Healthcare adds TRIPOD-LLM (Gallifant et al., 2025, Nature Medicine), the first LLM-specific extension of TRIPOD reporting standards, and cites a systematic review (Huo et al., 2025, JAMA Network Open) of 519 LLM health studies finding 99.3% used closed-source models without identifying which version.","why":"Gives readers a reporting standard for LLM studies and evidence of how rarely current studies specify which model version they used.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Gallifant et al., 2025, Nature Medicine","url":"https://doi.org/10.1038/s41591-024-03425-5","type":"paper"},{"label":"Huo et al., 2025, JAMA Network Open","url":"https://doi.org/10.1001/jamanetworkopen.2024.60928","type":"paper"}]}]},
{"date":"2026-01-30","headline":"FDA loosens Clinical Decision Support software oversight for AI diagnostic tools","entries":[{"category":"policy","title":"FDA loosens Clinical Decision Support software oversight (January 2026)","change":"AI Policy and Governance adds the FDA's January 2026 Clinical Decision Support guidance change: a single recommendation is now exempt from medical device regulation if it is the only clinically appropriate option (a term FDA left undefined), time-critical decisions no longer automatically trigger exclusion from the exemption, and the SaMD Clinical Evaluation guidance was withdrawn on January 7, 2026.","why":"Alerts readers that generative AI tools offering diagnostic suggestions may now reach clinical and public health settings without FDA review under the updated exemption criteria.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Epstein Becker Green analysis on FDA CDS guidance, 2026","url":"https://www.ebglaw.com/sp_resources-blogpost-stranger-things-new-fda-updates-flip-expectations-for-the-med-tech-industry-in-2026-upside-down","type":"other"}]}]},
{"date":"2026-02-02","headline":"Kim et al. systematic review finds only 6% of AI imaging studies externally validated","entries":[{"category":"evidence","title":"Kim et al. systematic review on external validation gaps in AI imaging studies","change":"Evaluating AI Systems for Healthcare cites Kim et al. (Korean Journal of Radiology, March 2019), a systematic review of 516 AI medical imaging studies that found only 6% (31 of 516) performed external validation on data from a different institution.","why":"Readers judging an imaging model have the primary source for how rarely AI imaging studies test performance on data from a different institution.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Kim et al., 2019, Korean Journal of Radiology","url":"https://pubmed.ncbi.nlm.nih.gov/30799571/","type":"paper"}]}]},
{"date":"2026-02-03","headline":"Cost-effectiveness analysis framework added to the deployment chapter","entries":[{"category":"evidence","title":"Cost-effectiveness analysis framework added to the deployment chapter","change":"AI Deployment: Why Most Prototypes Fail adds a Total Cost of Ownership framework, decision templates for AI versus epidemiologist staffing, and the CHEERS-AI evaluation checklist (Bullement et al., 2024, Value in Health), alongside a review finding AI can improve diagnostic accuracy and reduce costs but that economic benefits may be overstated (El Arab and Al Moosa, 2025, npj Digital Medicine).","why":"Gives health department directors a cost framework and reporting standard for deciding whether an AI investment is worth it.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[{"label":"El Arab and Al Moosa, 2025, npj Digital Medicine","url":"https://pubmed.ncbi.nlm.nih.gov/40858882/","type":"paper"},{"label":"Bullement et al., 2024, Value in Health","url":"https://pubmed.ncbi.nlm.nih.gov/38795956/","type":"paper"},{"label":"CHEERS-AI checklist, EQUATOR Network","url":"https://www.equator-network.org/reporting-guidelines/consolidated-health-economic-evaluation-reporting-standards-for-interventions-that-use-artificial-intelligence-cheers-ai/","type":"other"}]},{"category":"revision","title":"CheXNet generalization gap example added to evaluation chapter","change":"Evaluating AI Systems for Healthcare adds a concrete example of the generalization gap: CheXNet-style models scored an AUC of 0.93 internally but dropped to 0.82 at external hospitals (Zech et al., 2018, PLOS Medicine).","why":"Gives readers a concrete number illustrating why internal performance can overstate real-world accuracy.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[]},{"category":"content","title":"Data quality audit tactical guide added to the data chapter","change":"The Data Problem in Public Health AI adds a tactical guide for detecting data drift, facility-specific artifacts, and reporting-delay anomalies, with thresholds for when to monitor versus halt and retrain (mild drift of 1 to 2 features: monitor; severe drift of more than 5 features: halt and retrain) and rules for imputation versus exclusion of missing data.","why":"Gives practitioners concrete thresholds and decision rules for handling data problems that are conceptual elsewhere in the chapter.","chapters":[["The Data Problem in Public Health AI","/foundations/data.html"]],"sources":[]},{"category":"content","title":"Traditional reinforcement-learning agents and LLM-based agents","change":"Machine Learning Fundamentals distinguishes traditional reinforcement-learning agents from newer LLM-based agents that plan, use tools, and iterate autonomously.","why":"Distinguishes two approaches to AI agents and their capabilities.","chapters":[["Machine Learning Fundamentals","/foundations/ai-basics.html"],["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[]},{"category":"revision","title":"Forecast accuracy figures by time horizon added to forecasting chapter","change":"Epidemic Forecasting with AI adds median error figures by horizon: 10 to 15% for one-week forecasts and about 50% for four-week forecasts, noting error grows rapidly with forecast horizon.","why":"Gives readers a concrete sense of how quickly epidemic forecast accuracy degrades over time.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[]},{"category":"content","title":"Organizational readiness and change management framework added to deployment chapter","change":"AI Deployment: Why Most Prototypes Fail adds guidance on resource-constrained settings and organizational change management: a five-dimension readiness assessment, five resistance patterns with matching interventions, and a discontinuation-decision framework with pause triggers to plan before go-live.","why":"Gives health departments a structured way to assess readiness and plan for resistance before committing to an AI deployment.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[]}]},
{"date":"2026-02-05","headline":"Lost-in-the-middle limitation of long-context LLMs (Liu et al., 2024)","entries":[{"category":"evidence","title":"Lost-in-the-middle limitation of long-context LLMs","change":"LLMs in Public Health adds a section on the lost-in-the-middle problem: model performance drops by more than 30% when relevant information sits in the middle of a long context rather than at the beginning or end (Liu et al., 2024), with a table of risks for literature review, surveillance reports, patient histories, and protocol analysis.","why":"Warns readers that large advertised context windows do not guarantee even attention across a document, which matters for anything read end-to-end by an LLM.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Liu et al., 2024, TACL","url":"https://aclanthology.org/2024.tacl-1.9/","type":"paper"}]}]},
{"date":"2026-02-07","headline":"Course syllabus template for teaching AI in public health; Executive summary updated with data quality and cost-of-ownership findings","entries":[{"category":"content","title":"Course syllabus template for teaching AI in public health","change":"Course Syllabus adds a 12-week (and 6-week intensive) graduate syllabus template for AI in Public Health Practice, mapping each week to a handbook chapter and adding supplementary readings including Wiens et al., 2019 (Nature Medicine) and Reich et al., 2019 (PLOS Computational Biology).","why":"Gives instructors a ready-to-adapt course structure for teaching AI evaluation to MPH, DrPH, and health department audiences.","chapters":[["Course Syllabus Template","/appendices/course-syllabus.html"]],"sources":[{"label":"Wiens et al., 2019, Nature Medicine","url":"https://www.nature.com/articles/s41591-019-0548-6","type":"paper"},{"label":"Mooney and Pejaver, 2018, Annual Review of Public Health","url":"https://www.annualreviews.org/content/journals/10.1146/annurev-publhealth-040617-014208","type":"paper"},{"label":"Reich et al., 2019, PLOS Computational Biology","url":"https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007486","type":"paper"},{"label":"Shaw et al., 2019, JMIR","url":"https://www.jmir.org/2019/7/e13659","type":"paper"}]},{"category":"revision","title":"Executive summary updated with data quality and cost-of-ownership findings","change":"Executive Summary adds that data quality is public health AI's binding constraint, citing statistical drift-detection methods most health departments lack capacity to run, and that budget overruns of 30 to 40% are common when total cost of ownership is underestimated, with hidden costs such as data pipeline optimization often reaching 63% of total expenses.","why":"Brings the executive summary in line with the data quality and cost-effectiveness guidance already added to the chapters.","chapters":[["Executive Summary","/executive-summary.html"]],"sources":[]}]},
{"date":"2026-02-09","headline":"Self-referential chapter openers replaced with direct statements across many chapters","entries":[{"category":"revision","title":"Self-referential chapter openers replaced with direct statements","change":"Removes self-referential openers such as This chapter examines... across thirteen chapters and appendices, replacing them with direct statements, for example opening AI Failures Post-Mortems with a direct statement that healthcare AI failures cost billions and can harm patients.","why":"Makes chapter openings state their content directly instead of describing what the chapter will do.","chapters":[["The AI Morgue: Failure Post-Mortems","/appendices/failures.html"],["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"],["The Data Problem in Public Health AI","/foundations/data.html"],["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"],["AI in Global Health and Equity","/future/global-health.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"],["AI Safety in Healthcare: Protecting Patients and Populations","/implementation/safety.html"]],"sources":[]}]},
{"date":"2026-02-10","headline":"Mechanistic interpretability case study for RL-based Medicaid care coordination; Constrained fairness optimization cuts disparities in Medicaid RL system","entries":[{"category":"evidence","title":"Mechanistic interpretability case study for RL-based Medicaid care coordination","change":"Evaluating AI Systems for Healthcare adds a case study auditing a SARSA reinforcement learning system for Medicaid care coordination across three states (250,000 decisions, 45,000 beneficiaries), combining attention analysis, SHAP, and sparse autoencoders to identify seven interpretable reasoning motifs linking social factors, such as housing instability, to clinical outcomes (Basu et al., 2026).","why":"Gives readers a concrete method for explaining sequential AI decisions in population health management, not just single predictions.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Basu et al., 2026, BMJ Health & Care Informatics","url":"https://doi.org/10.1136/bmjhci-2025-101935","type":"paper"},{"label":"Basu et al., 2025, JMIR","url":"https://doi.org/10.2196/74264","type":"paper"}]},{"category":"evidence","title":"Constrained fairness optimization cuts disparities in Medicaid RL system","change":"Ethics, Bias, and Equity adds that constrained fairness optimization on a Medicaid care coordination reinforcement learning system cut race-group disparity by 37% and sex-group disparity by 28%, with baseline harm of 1.03% to 1.33% tightened to 1.10% to 1.20% across groups and only a 0.2% decrease in overall policy value (Basu et al., 2026).","why":"Shows that fairness constraints can reduce demographic disparities without a meaningful loss in policy effectiveness.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"Basu et al., 2026, BMJ Health & Care Informatics","url":"https://doi.org/10.1136/bmjhci-2025-101935","type":"paper"}]}]},
{"date":"2026-02-11","headline":"Glossary terms added for conformal prediction, mechanistic interpretability, and SARSA","entries":[{"category":"content","title":"Glossary terms added for conformal prediction, mechanistic interpretability, and SARSA","change":"Glossary adds three terms: conformal prediction (a distribution-free method for calibrating prediction thresholds with coverage guarantees, useful for safety envelopes that flag high-risk cases for review), mechanistic interpretability (methods that expose a model's internal reasoning rather than just its output), and SARSA (a conservative, on-policy reinforcement learning algorithm suited to safety-critical healthcare).","why":"Defines the technical terms introduced by the recent Medicaid reinforcement learning case study for readers encountering them for the first time.","chapters":[["Glossary","/appendices/glossary.html"]],"sources":[]}]},
{"date":"2026-02-12","headline":"HAIRA: a maturity framework for healthcare AI governance; AI Safety and Risk reading list in Further Reading","entries":[{"category":"evidence","title":"HAIRA: a maturity framework for healthcare AI governance","change":"AI Deployment: Why Most Prototypes Fail adds a Governance Maturity Assessment section describing the Healthcare AI Governance Readiness Assessment (HAIRA), a five-level maturity model across seven governance domains, drawn from a systematic review of 35 healthcare AI governance frameworks.","why":"A systematic review found most healthcare AI governance frameworks assume resources few public health agencies and community health systems have, so the chapter adds a maturity model scaled to available resources.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"],["AI Policy and Governance in Healthcare","/future/policy.html"],["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"Hussein et al., 2026","url":"https://doi.org/10.1038/s41746-026-02418-7","type":"paper"},{"label":"Bedoya et al., 2022","url":"https://doi.org/10.1093/jamia/ocac078","type":"paper"},{"label":"Kim et al., 2024","url":"https://doi.org/10.1371/journal.pdig.0000390","type":"paper"}]},{"category":"content","title":"AI Safety and Risk reading list in Further Reading","change":"Further Reading adds an AI Safety and Risk section recommending Dan Hendrycks's Introduction to AI Safety, Ethics, and Society and the Center for AI Safety, and AI Safety in Healthcare cites the Hendrycks book for its frameworks on catastrophic risk and safety engineering.","why":"Gives readers curated entry points into AI safety frameworks and evaluation methodology relevant to public health AI deployment.","chapters":[["Further Reading","/appendices/resources.html"],["AI Safety in Healthcare: Protecting Patients and Populations","/implementation/safety.html"]],"sources":[{"label":"Hendrycks, Introduction to AI Safety, Ethics, and Society, 2024","url":"https://doi.org/10.1201/9781003530336","type":"other"},{"label":"Center for AI Safety, Statement on AI Risk","url":"https://www.safe.ai/statement-on-ai-risk","type":"announcement"}]}]},
{"date":"2026-02-13","headline":"CHART statement for evaluating conversational AI; mTRAC automated outbreak detection example; Medicaid care coordination case study updated","entries":[{"category":"evidence","title":"CHART statement for evaluating conversational AI","change":"Evaluating AI Systems for Healthcare adds the CHART statement, a 12-item, 39-subitem checklist for evaluating chatbot and conversational AI, as a complement to TRIPOD-LLM.","why":"Gives readers a checklist specific to chatbot and conversational AI, an evaluation gap general model-reporting guidelines do not cover.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Huo et al., 2025","url":"https://doi.org/10.1001/jamanetworkopen.2025.30220","type":"paper"}]},{"category":"evidence","title":"mTRAC automated outbreak detection example","change":"AI in Global Health and Equity adds mTRAC, an SMS-based reporting system with automated anomaly detection used in Uganda, as an example of automated outbreak detection.","why":"Gives readers a low-resource example of automated outbreak detection already deployed in a national health system.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Cummins, 2013","url":"https://onlinelibrary.wiley.com/doi/abs/10.1111/1759-5436.12017","type":"paper"},{"label":"UNICEF Uganda, mTRAC","url":"https://www.unicef.org/uganda/what-we-do/mtrac","type":"other"}]},{"category":"revision","title":"Medicaid care coordination case study updated","change":"Ethics, Bias, and Equity and Evaluating AI Systems for Healthcare revise the Medicaid care coordination case study: a SARSA reinforcement learning system across three states (250,000 decisions, 45,000 beneficiaries) reduced acute care events by 12 percentage points (NNT 8.3) while narrowing the race and ethnicity equalized odds disparity from 8.9% to 5.6% and the gender disparity from 5.3% to 3.8%.","why":"Refines the case study's outcome and fairness figures for a mixed-methods reinforcement learning deployment already covered in both chapters.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[]}]},
{"date":"2026-02-19","headline":"Federal CDC database disruption documented in the surveillance chapter","entries":[{"category":"evidence","title":"CDC surveillance database disruption, 2025–2026","change":"Disease Surveillance and Outbreak Detection adds a Federal Surveillance Infrastructure Disruption section: a peer-reviewed audit found 38 of 82 CDC databases updated at least monthly had stopped updating, about 87% of them tracking vaccination coverage, while ILINet, FluView, and BioSense remain operational.","why":"Warns AI practitioners that surveillance models calibrated against paused federal baselines lose their reference points and need dependency audits.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Jacobs et al., 2026","url":"https://doi.org/10.7326/ANNALS-25-04022","type":"paper"}]}]},
{"date":"2026-02-20","headline":"West Coast Health Alliance, measles wastewater validation, and LMIC forecasting added","entries":[{"category":"developments","title":"Aidoc's expanded multi-condition CT triage clearance","change":"Diagnostic and Clinical Decision Support adds Aidoc's January 2026 FDA 510(k) clearance (K252970) for multi-condition abdominal CT triage covering 14 findings in one workflow, noting the vendor-reported 97% sensitivity and 98% specificity await independent validation and that Aidoc's \"foundation model\" label has no FDA regulatory definition.","why":"Gives readers the regulatory scope and vendor-reported accuracy of an expanded imaging triage clearance, flagging marketing language without a regulatory basis.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Aidoc","url":"https://www.aidoc.com/","type":"announcement"}]},{"category":"policy","title":"CDC FluSight shifts to hospital admissions target","change":"Epidemic Forecasting with AI updates its description of CDC's FluSight challenge: beginning with the 2023 to 2024 season, CDC shifted the primary forecasting target from outpatient influenza-like illness percentage to weekly laboratory-confirmed influenza hospital admissions, and the chapter adds a section on adapting forecasting methods for data-scarce settings.","why":"Keeps the chapter's description of the CDC's flagship forecasting challenge current with its 2023 target change and highlights that most validated forecasting infrastructure serves high-income countries only.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"CDC, FluSight challenge","url":"https://www.cdc.gov/flu-forecasting/index.html","type":"policy"}]},{"category":"policy","title":"Biosecurity Modernization and Innovation Act of 2026","change":"Genomic Surveillance and Pathogen Analysis adds the proposed Biosecurity Modernization and Innovation Act of 2026 (Cotton-Klobuchar), which would require gene synthesis providers to screen orders and customers, establish a NIST biosecurity governance sandbox, and mandate a government-wide assessment of federal biosecurity authority gaps; the chapter notes the bill is not yet enacted.","why":"Gives readers the specific provisions of pending federal biosecurity legislation and its industry and government backers before any vote.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Cotton, U.S. Senate, 2026","url":"https://www.cotton.senate.gov/news/press-releases/cotton-klobuchar-introduce-bill-to-establish-federal-biotech-security-framework","type":"policy"}]},{"category":"policy","title":"West Coast Health Alliance as a CDC surveillance alternative","change":"Disease Surveillance and Outbreak Detection adds the West Coast Health Alliance (California, Oregon, Washington, and Hawaii; launched September 2025), which coordinates regional public health guidelines and immunization recommendations independent of the CDC, and cites an editorial accompanying the CDC database disruption audit.","why":"Shows readers a concrete state-level response as fragmented alternatives emerge to fill gaps left by paused federal surveillance databases.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"West Coast Health Alliance announcement, 2025","url":"https://www.gov.ca.gov/2025/09/03/california-oregon-and-washington-to-launch-new-west-coast-health-alliance-to-uphold-scientific-integrity-in-public-health-as-trump-destroys-cdcs-credibility/","type":"policy"},{"label":"Marrazzo, 2026","url":"https://doi.org/10.7326/ANNALS-25-05527","type":"paper"}]},{"category":"evidence","title":"Wastewater validation of measles detection, Mesa County","change":"Disease Surveillance and Outbreak Detection adds a measles wastewater surveillance validation case: during a 2025 Mesa County, Colorado outbreak, wastewater detected wild-type measles virus about four days before clinical case counts rose, and seven confirmed cases were ultimately linked to the outbreak.","why":"Demonstrates wastewater surveillance detecting an outbreak ahead of clinical case reporting, evidence relevant amid disruptions to federal case-based surveillance databases.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Jensen et al., MMWR, 2026","url":"https://www.cdc.gov/mmwr/volumes/75/wr/mm7502a2.htm","type":"paper"},{"label":"CDC, National Wastewater Surveillance System, measles data","url":"https://www.cdc.gov/nwss/rv/measles.html","type":"policy"}]},{"category":"evidence","title":"International and LMIC epidemic forecasting","change":"Epidemic Forecasting with AI adds an International and LMIC Forecasting section: the European Forecast Hub found ensembles outperform individual models across 32 countries (Sherratt et al., 2023); modeling found only 1.4% of African region SARS-CoV-2 infections were captured by reported data (Cabore et al., 2022); and a dengue challenge found skill was lowest in high-incidence seasons (Johansson et al., 2019).","why":"Extends the forecasting chapter beyond high-income, well-resourced settings to the surveillance gaps and adapted methods that low- and middle-income countries require.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"European COVID-19 Forecast Hub","url":"https://covid19forecasthub.eu/","type":"other"},{"label":"Sherratt et al., 2023","url":"https://doi.org/10.7554/eLife.81916","type":"paper"},{"label":"Cabore et al., 2022","url":"https://doi.org/10.1016/S2214-109X(22)00233-9","type":"paper"},{"label":"Johansson et al., 2019","url":"https://doi.org/10.1073/pnas.1909865116","type":"paper"}]}]},
{"date":"2026-02-23","headline":"Performance gaps in surveillance and AI early-warning systems; Randomized trial: LLM medical assistants do not improve lay decisions","entries":[{"category":"evidence","title":"Performance gaps in surveillance and AI early-warning systems","change":"Disease Surveillance and Outbreak Detection and related chapters add evidence that event-based surveillance systems (HealthMap, WHO EIOS) detected only 9.2% of influenza outbreaks within a two-week window across 24 countries (Ganser et al., 2022), and that a systematic review of 67 AI-based early warning studies found no pooled sensitivity benchmark across systems (Villanueva-Miranda et al., 2025).","why":"Quantifies how far event-based and AI early-warning surveillance still lag real-time outbreak detection, especially in lower-income countries.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"],["Epidemic Forecasting with AI","/applications/forecasting.html"],["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Ganser et al., 2022","url":"https://doi.org/10.2196/36211","type":"paper"},{"label":"Villanueva-Miranda et al., 2025","url":"https://doi.org/10.3389/fpubh.2025.1609615","type":"paper"}]},{"category":"evidence","title":"Randomized trial: LLM medical assistants do not improve lay decisions","change":"Evaluating AI Systems for Healthcare adds a preregistered randomized trial of 1,298 UK adults testing GPT-4o, Llama 3, and Command R+ as medical assistants: despite strong model-only performance, participants using LLM assistance did not show reliable improvement over controls using their usual methods such as internet search, with human-LLM interaction, not model knowledge, driving the gap.","why":"Shows that benchmark-style model performance did not predict real-world usefulness for lay users, a gap standard testing methods failed to catch.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Bean et al., 2026, Nature Medicine","url":"https://doi.org/10.1038/s41591-025-04074-y","type":"paper"}]}]},
{"date":"2026-02-24","headline":"OpenAI's ChatGPT for Health; ARGO surveillance after Google Flu Trends","entries":[{"category":"developments","title":"OpenAI's ChatGPT for Health","change":"Diagnostic and Clinical Decision Support adds OpenAI's January 2026 announcement of ChatGPT for Health, positioned as a health education tool that answers questions using GPT-4-level reasoning without clinical validation or regulatory oversight.","why":"Gives readers the developer's own framing of a new consumer health AI product and its explicit lack of clinical validation.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"OpenAI, January 2026","url":"https://openai.com/index/introducing-chatgpt-health/","type":"announcement"}]},{"category":"content","title":"ARGO surveillance after Google Flu Trends","change":"Disease Surveillance and Outbreak Detection adds ARGO, the AutoRegression-with-Google-search-data method developed after Google Flu Trends' failure (Yang et al., 2015).","why":"Explains how ARGO combines search data with autoregression for disease surveillance.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"],["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Lo-Ciganic et al., 2022","url":"https://pubmed.ncbi.nlm.nih.gov/35623798/","type":"paper"},{"label":"Yang et al., 2015","url":"https://www.pnas.org/doi/abs/10.1073/pnas.1515373112","type":"paper"}]}]},
{"date":"2026-02-25","headline":"CDC and HHS federal AI adoption in public health","entries":[{"category":"policy","title":"CDC and HHS federal AI adoption in public health","change":"LLMs in Public Health adds a Federal Government Adoption section: CDC deployed ChatCDC in 2024 as the first agency-wide generative AI chatbot, reporting $3.7 million in labor cost savings and 55 AI solutions in production, while HHS reported 271 AI use cases in FY2024, a 66% increase from FY2023.","why":"Gives readers concrete federal benchmarks for generative AI adoption scale and reported cost savings inside public health agencies.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"CDC, Public Health Data Strategy, September 2025","url":"https://www.cdc.gov/public-health-data-strategy/php/story/using-ai-to-improve-public-health-efficiency.html","type":"policy"},{"label":"National Academy of Medicine, AI in Health Care: Hope, Hype, Promise, Peril, 2022","url":"https://doi.org/10.17226/27111","type":"report"},{"label":"National Academy of Medicine, Generative AI in Health and Medicine, 2025","url":"https://doi.org/10.17226/28907","type":"report"},{"label":"National Academy of Medicine, AI Code of Conduct for Health and Medicine, 2025","url":"https://www.nationalacademies.org/publications/29087","type":"report"}]}]},
{"date":"2026-02-27","headline":"PandemIQ Llama open-sourced with expanded outbreak-detection performance metrics","entries":[{"category":"evidence","title":"PandemIQ Llama: open-source outbreak detection model","change":"Disease Surveillance and Outbreak Detection details PandemIQ Llama, an open-source language model built on Meta's Llama-3.1-8B and trained on 5.8 billion tokens across 31 languages, now freely available on GitHub and Hugging Face; BEACON has published over 1,300 disease reports covering nearly 600 outbreaks in 195 countries, integrated into WHO's EIOS system.","why":"Gives readers the technical specifications and open license of a purpose-built outbreak-detection model now reproducible and auditable outside its developers.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"PandemIQ Llama (GitHub)","url":"https://github.com/noc-lab/PandemIQ-Llama","type":"dataset"},{"label":"Yang et al., 2026","url":"https://par.nsf.gov/servlets/purl/10657311","type":"paper"},{"label":"Bhadelia et al., 2025","url":"https://doi.org/10.1093/infdis/jiaf642","type":"paper"}]}]},
{"date":"2026-03-01","headline":"ChatGPT Health: independent triage accuracy evaluation","entries":[{"category":"evidence","title":"ChatGPT Health: independent triage accuracy evaluation","change":"Adds a case example to the Diagnostic and Clinical Decision Support chapter on ChatGPT Health, describing an independent Mount Sinai evaluation across 960 clinician-written scenarios in 21 domains that found a 52% under-triage rate for emergency cases, an inverted-U accuracy pattern, and anchoring bias when symptoms were minimized (OR 11.7).","why":"OpenAI launched ChatGPT Health as a consumer health tool; this is the first independent evaluation of its clinical triage safety.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Ramaswamy et al., Nature Medicine, 2026","url":"https://doi.org/10.1038/s41591-026-04297-7","type":"paper"}]}]},
{"date":"2026-03-06","headline":"Vital Strategies AI readiness framework for public health systems","entries":[{"category":"developments","title":"Vital Strategies AI readiness framework for public health systems","change":"Adds a Further Reading entry for Vital Strategies' Foundations and Futures report, a 63-country AI readiness assessment across governance, data systems, connectivity, human capacity, and financing, including its AI Use-Case Prerequisites Matrix.","why":"Gives public health leaders and funders a framework for assessing AI adoption readiness, particularly in LMICs.","chapters":[["Further Reading","/appendices/resources.html"]],"sources":[{"label":"Vital Strategies, Foundations and Futures report, 2026","url":"https://www.vitalstrategies.org/resources/foundations-and-futures-public-health-in-the-ai-era/","type":"report"}]},{"category":"content","title":"Case study: open-source NLP for clinical note surveillance in Brazil","change":"Adds a case study to the AI in Global Health and Equity chapter describing Vital Strategies' deployment of open-source NLP within Brazil's SUS health system to extract surveillance signals, including underreported gender-based violence, from 39 million unstructured clinical notes.","why":"Documents a national-scale deployment of open-source NLP for surveillance, relevant to LMIC health systems seeking similar capability without licensing costs.","chapters":[["AI in Global Health and Equity","/future/global-health.html"],["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Vital Strategies, Foundations and Futures report, 2026","url":"https://www.vitalstrategies.org/resources/foundations-and-futures-public-health-in-the-ai-era/","type":"report"},{"label":"WHO and IndiaAI, AI Health Casebook, 2026","url":"https://impact.indiaai.gov.in/events/casebook-on-ai-health-use-cases-across-the-global-south","type":"report"}]}]},
{"date":"2026-03-08","headline":"WHO EIOS 2.0 upgrade and CEPI-BEACON zoonotic risk integration","entries":[{"category":"developments","title":"CEPI-BEACON-UC Davis integration for zoonotic spillover risk","change":"Adds detail to the Disease Surveillance and Outbreak Detection chapter on CEPI's partnership with UC Davis and BEACON, funded at up to $1 million, to connect the VISTA zoonotic virus risk-ranking platform (formerly SpillOver) with BEACON's outbreak surveillance pipeline.","why":"Links upstream zoonotic spillover risk prediction with downstream outbreak surveillance, rather than operating the two as separate systems.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"CEPI, new global collaboration announcement, 2025","url":"https://cepi.net/new-global-collaboration-uses-experts-and-ai-spot-next-pandemic","type":"announcement"}]},{"category":"policy","title":"WHO EIOS 2.0: AI-powered epidemic intelligence upgrade","change":"Adds detail to the Disease Surveillance and Outbreak Detection chapter on WHO's October 2025 EIOS 2.0 upgrade, its first major revision since 2017, including AI-powered signal detection from radio and other open sources, a multilingual interface, and adoption by over 110 Member States and about 30 organizations.","why":"WHO upgraded its open-source epidemic intelligence platform, widening the range of public signals it can monitor for outbreak detection.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"WHO, EIOS version 2.0 upgrade announcement, 2025","url":"https://www.who.int/news/item/13-10-2025-who-upgrades-its-public-health-intelligence-system-to-boost-global-health-security","type":"policy"}]}]},
{"date":"2026-03-15","headline":"WHO digital health strategy extension; AI evaluation reporting standards added","entries":[{"category":"policy","title":"WHO Global Strategy on Digital Health extended to 2027","change":"Adds detail to the AI in Global Health and Equity chapter on the World Health Assembly's May 2025 extension of the Global Strategy on Digital Health through 2027, noting 129 countries have adopted national digital health strategies and a 2028-2033 follow-up framework is in development.","why":"Gives readers the policy framework within which AI-specific digital health initiatives operate.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"WHO, World Health Assembly, May 2025","url":"https://www.who.int/news/item/23-05-2025-world-health-assembly-endorses-extension-of-the-global-digital-health-strategy-to-2027","type":"policy"}]},{"category":"evidence","title":"Reporting standards for AI clinical evaluation: SPIRIT-AI, DECIDE-AI, CONSORT-EHEALTH","change":"Adds three reporting standards to the Evaluating AI Systems for Healthcare chapter: SPIRIT-AI for AI clinical trial protocols (Cruz Rivera et al., 2020), DECIDE-AI for early-stage AI decision support evaluation (Vasey et al., 2022), and CONSORT-EHEALTH for digital health intervention reporting (Eysenbach, 2011).","why":"These established checklists set what investigators should report when designing and evaluating AI clinical studies, from trial protocols to early-stage decision support.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Cruz Rivera et al., 2020","url":"https://doi.org/10.1136/bmj.m3210","type":"paper"},{"label":"Vasey et al., 2022","url":"https://doi.org/10.1136/bmj-2022-070904","type":"paper"},{"label":"Eysenbach, 2011","url":"https://doi.org/10.2196/jmir.1923","type":"paper"}]}]},
{"date":"2026-03-18","headline":"Anthropic user survey, Microsoft Copilot health data, METR coding trial","entries":[{"category":"developments","title":"Anthropic global AI user survey: trust and equity findings","change":"Adds findings from Anthropic's survey of 80,508 AI users across 159 countries to two chapters: Global Health notes that respondents in lower-income regions expressed more optimism about AI's potential than those in high-income regions, while Ethics catalogs the top concerns eroding trust, led by unreliability (26.7%) and job displacement (22.3%).","why":"Shows that distrust of AI is not uniform globally and names the specific concerns public health AI deployments must address.","chapters":[["AI in Global Health and Equity","/future/global-health.html"],["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"Anthropic, 2025","url":"https://www.anthropic.com/features/81k-interviews","type":"announcement"}]},{"category":"developments","title":"Microsoft Research analysis of Copilot health conversations","change":"Adds findings to the LLMs in Public Health chapter from Microsoft Research's analysis of over 500,000 de-identified Copilot health conversations from January 2026, finding nearly 1 in 5 involve personal symptom assessment and that emotional wellbeing queries rise at night and on mobile devices.","why":"Gives an independent, platform-scale complement to OpenAI's self-reported consumer health AI usage data, though both remain company-disclosed metrics.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Tolmachev, Costa-Gomes, and Sounderajah, Microsoft Research, 2026","url":"https://www.microsoft.com/en-us/research/publication/how-people-use-copilot-for-health/","type":"announcement"}]},{"category":"evidence","title":"METR randomized trial on AI coding productivity","change":"Adds a caveat to the AI-Assisted Coding for Public Health chapter: a METR 2025 randomized controlled trial found experienced open-source developers took 19% longer on tasks using AI tools despite predicting a 20% speedup, contrasting with a previously cited 55% faster result from less experienced developers on simpler tasks.","why":"Qualifies earlier AI coding productivity claims by showing effects depend heavily on developer experience and task complexity.","chapters":[["AI-Assisted Coding for Public Health Analysis","/practical/ai-coding.html"]],"sources":[{"label":"METR, 2025","url":"https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/","type":"report"}]}]},
{"date":"2026-03-28","headline":"Physics-informed neural forecasting: Fujita and Akutsu 2025; Case study: rapid genomic containment of H5N1 in the Cayman Islands; Wastewater genomic surveillance: variant tracking with Freyja and NWSS; NarxCare feature replication finds low precision: Wang et al. 2026; CDC's first AI strategy and agentic tool guidance; Federal preemption of state AI laws: December 2025 executive order; Toolkit addition: uv, a Rust-based Python package manager","entries":[{"category":"evidence","title":"Physics-informed neural forecasting: Fujita and Akutsu 2025","change":"Adds a third epidemic forecasting approach to the Epidemic Forecasting with AI chapter: Fujita and Akutsu's neural network constrained by an SIR module, which produced more stable multi-region forecasts than baseline deep learning models while preserving interpretable epidemiologic parameters.","why":"Gives readers a structure-constrained alternative to purely data-driven forecasting models.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Fujita and Akutsu, PLOS One, 2025","url":"https://doi.org/10.1371/journal.pone.0331611","type":"paper"}]},{"category":"content","title":"Case study: rapid genomic containment of H5N1 in the Cayman Islands","change":"Adds a case study to the Genomic Surveillance chapter describing the Cayman Islands' November 2025 rapid whole genome sequencing of A(H5N1) clade 2.3.4.4b within hours, enabling containment with no human cases, contrasted with the persisting US H5N1 spread across migratory flyways and dairy cattle.","why":"Illustrates how in-country genomic sequencing capacity can be the difference between containment and a wider outbreak.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"PAHO, March 2026","url":"https://www.paho.org/en/news/18-3-2026-early-detection-and-genomic-surveillance-key-containing-ah5n1-cayman-islands","type":"policy"},{"label":"USGS, 2026","url":"https://www.usgs.gov/centers/nwhc/science/avian-influenza","type":"policy"}]},{"category":"content","title":"Wastewater genomic surveillance: variant tracking with Freyja and NWSS","change":"Adds a new section to the Genomic Surveillance chapter on wastewater-based epidemiology, describing how the Freyja tool deconvolves sequencing mixtures to estimate circulating lineage proportions (Karthikeyan et al., Nature, 2022) and pointing to the CDC's National Wastewater Surveillance System dashboard.","why":"Gives readers a population-level surveillance method that captures cases without relying on clinical testing.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Karthikeyan et al., Nature, 2022","url":"https://www.nature.com/articles/s41586-022-05049-6","type":"paper"},{"label":"Freyja (open-source tool)","url":"https://github.com/andersen-lab/Freyja","type":"dataset"},{"label":"CDC, National Wastewater Surveillance System dashboard","url":"https://www.cdc.gov/nwss/rv/COVID19-nationaltrend.html","type":"policy"}]},{"category":"evidence","title":"NarxCare feature replication finds low precision: Wang et al. 2026","change":"Adds a study to the Substance Use and Overdose Prevention chapter: Wang et al. replicated NarxCare-aligned features across PDMP and claims data and found precision from 0.01 to 0.32, far below the vendor-reported benchmark of 0.75, and call commercial tool opacity a barrier to independent bias auditing.","why":"Independently tests a widely used prescription-monitoring risk score and finds its published accuracy claim unsupported.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"sources":[{"label":"Wang et al., npj Digital Medicine, 2026","url":"https://doi.org/10.1038/s41746-026-02491-y","type":"paper"}]},{"category":"policy","title":"CDC's first AI strategy and agentic tool guidance","change":"Adds detail to the Disease Surveillance and Outbreak Detection chapter on CDC's first AI strategy (FY 2026 to 2030) and separate guidance for state, tribal, local, and territorial partners on agentic research tools, noting the guidance limits these tools to narrow synthesis tasks with expert validation.","why":"CDC officials reported that about 10% of the agency's roughly 100 AI use cases were already agentic, prompting formal guardrails for their use.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"CDC, AI strategy, 2026","url":"https://www.cdc.gov/ai/strategy/index.html","type":"policy"},{"label":"CDC, agentic research guidance for STLT partners, 2026","url":"https://www.cdc.gov/ai/resources/considerations-for-agentic-research-in-public-health.html","type":"policy"},{"label":"Alder, FedScoop, March 2026","url":"https://fedscoop.com/cdc-releases-ai-strategy-guidance-with-eye-toward-agentic-uses/","type":"news"}]},{"category":"policy","title":"Federal preemption of state AI laws: December 2025 executive order","change":"Adds a section to the AI Policy and Governance chapter on the December 11, 2025 executive order directing federal preemption of state AI regulations, including a DOJ litigation task force, a Commerce Department report due March 2026, and funding conditions, and notes healthcare-specific patient safety laws were absent from its stated exceptions.","why":"Flags an unresolved gap in patient safety protections as federal preemption of state AI laws proceeds.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Hooper Lundy, analysis, 2025","url":"https://hooperlundy.com/the-trump-executive-order-on-ai-regulation-key-takeaways-for-the-health-care-industry/","type":"other"},{"label":"Ropes and Gray, analysis, 2026","url":"https://ropesgray.com/en/insights/alerts/2026/03/examining-the-landscape-and-limitations-of-the-federal-push-to-override-state-ai-regulation","type":"other"}]},{"category":"content","title":"Toolkit addition: uv, a Rust-based Python package manager","change":"Adds uv, a Rust-based Python package manager from Astral, to the Your AI Toolkit for Public Health chapter as a faster alternative to pip, venv, and pyenv for environment and package management in reproducible analysis pipelines.","why":"Gives practitioners a faster tool option for building reproducible public health analysis environments.","chapters":[["Your AI Toolkit for Public Health","/practical/toolkit.html"]],"sources":[{"label":"uv (Astral)","url":"https://github.com/astral-sh/uv","type":"dataset"}]}]},
{"date":"2026-04-02","headline":"Ambient scribe reduces documentation time: Ma et al. 2025; PandemicLLM: language models for epidemic forecasting; PathGen: federated pathogen genomics platform previewed; EU AI Act: phased high-risk compliance deadlines; Toolkit addition: Ruff, a Python linter and formatter","entries":[{"category":"evidence","title":"Ambient scribe reduces documentation time: Ma et al. 2025","change":"Adds a study to the Diagnostic and Clinical Decision Support chapter: a prospective quality improvement study across 45 physicians in 8 ambulatory specialties (17,428 encounters) found ambient scribe use in 55% of encounters cut daily EHR time by 19.95 minutes, including 5.17 fewer after-hours minutes.","why":"Gives a real-world, multi-specialty measurement of ambient scribe impact on clinician documentation burden.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Ma et al., JAMIA, 2025","url":"https://doi.org/10.1093/jamia/ocae304","type":"paper"}]},{"category":"evidence","title":"PandemicLLM: language models for epidemic forecasting","change":"Adds a fourth forecasting approach to the Epidemic Forecasting with AI chapter: PandemicLLM feeds epidemiologic, policy, spatial, and genomic context into large language models, reporting better 1- to 3-week COVID-19 forecasts than CDC Forecast Hub baselines in retrospective US experiments, though evidence is limited to one disease setting.","why":"Gives readers a context-rich alternative to case-count-only forecasting models, with an explicit caveat on its limited evaluation.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Du et al., Nature Computational Science, 2025","url":"https://doi.org/10.1038/s43588-025-00798-6","type":"paper"}]},{"category":"developments","title":"PathGen: federated pathogen genomics platform previewed","change":"Adds detail to the Disease Surveillance and Outbreak Detection chapter on PathGen, previewed by Duke-NUS's Asia Pathogen Genomics Initiative in December 2025, a federated 'sovereign-by-design' platform combining pathogen genomics with clinical, population, climate, and mosquito-habitat data, with pilots scheduled from early 2026 through 2027.","why":"Adds a data-sovereign alternative to centralized commercial AI surveillance platforms like BlueDot.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Duke-NUS, December 2025","url":"https://www.duke-nus.edu.sg/newshub/media-releases/pathgen","type":"announcement"}]},{"category":"policy","title":"EU AI Act: phased high-risk compliance deadlines","change":"Adds detail to the AI Policy and Governance chapter on the EU AI Act's phased deadlines: most high-risk obligations apply from August 2, 2026, but systems that are safety components of already-regulated products, including medical devices under Article 6(1), have until August 2, 2027.","why":"Helps public health teams deploying AI in regulated products identify which EU compliance deadline applies to their system.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"EU AI Act, Article 113","url":"https://artificialintelligenceact.eu/article/113/","type":"policy"}]},{"category":"content","title":"Toolkit addition: Ruff, a Python linter and formatter","change":"Adds a Ruff callout to the Your AI Toolkit for Public Health chapter describing the tool as a fast linter and formatter that replaces Flake8, isort, and Black behind one CLI, with setup guidance for pip and uv.","why":"Reduces tool sprawl for public health projects juggling notebooks, scripts, and CI checks.","chapters":[["Your AI Toolkit for Public Health","/practical/toolkit.html"]],"sources":[{"label":"Ruff (Astral)","url":"https://docs.astral.sh/ruff/","type":"other"}]}]},
{"date":"2026-04-13","headline":"Shadow AI: health systems govern only the visible tip; AI chatbot sycophancy risk: Cheng et al. 2026","entries":[{"category":"evidence","title":"Shadow AI: health systems govern only the visible tip","change":"Adds a section to the AI Policy and Governance chapter on Ötleş and colleagues' argument that restricting enterprise AI tools without addressing consumer-grade 'shadow AI' on personal devices drives use underground rather than reducing risk, and links to the handbook's governance policy template.","why":"Argues that banning or restricting official AI tools does not eliminate risk, only visibility into it.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Ötleş et al., NEJM AI, 2026","url":"https://doi.org/10.1056/AIe2600236","type":"paper"}]},{"category":"evidence","title":"AI chatbot sycophancy risk: Cheng et al. 2026","change":"Adds detail to the AI-Driven Behavioral Interventions chapter on a 2026 Science study across 11 leading AI models that found they affirmed users' actions 49% more often than humans, including in scenarios involving deception or illegality, reducing willingness to take responsibility after even a single interaction.","why":"Warns that behavioral health and coaching tools must resist inappropriate affirmation, not just detect crises.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[{"label":"Cheng et al., Science, 2026","url":"https://doi.org/10.1126/science.aec8352","type":"paper"}]}]},
{"date":"2026-04-22","headline":"Consumer wearables for outbreak detection: the DETECT study; FDA guidance on Predetermined Change Control Plans","entries":[{"category":"evidence","title":"Consumer wearables for outbreak detection: the DETECT study","change":"Disease Surveillance covers consumer wearable surveillance: the DETECT study found Oura Ring signals from 63,153 participants detected COVID-19 2.75 days before testing at 82% sensitivity (Mason et al., 2022), and a follow-up analysis found that refinements to baseline correction improved detection accuracy (Phipps et al., 2024).","why":"Gives readers evidence on a passive, population-scale signal that can precede laboratory confirmation, while noting it remains a research tool, not an operational system.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Mason et al., Scientific Reports, 2022","url":"https://doi.org/10.1038/s41598-022-07314-0","type":"paper"},{"label":"Phipps et al., NPJ Digital Medicine, 2024","url":"https://doi.org/10.1038/s41746-024-01129-1","type":"paper"}]},{"category":"revision","title":"FDA guidance on Predetermined Change Control Plans","change":"Evaluating AI Systems for Healthcare adds FDA's December 2024 final guidance allowing Predetermined Change Control Plans for pre-specified AI-enabled device updates without a new marketing submission.","why":"Describes the regulatory path for pre-specified updates to AI-enabled devices.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[]}]},
{"date":"2026-05-06","headline":"AI governance framework comparison for health departments; HTI-1 algorithm transparency in vendor due diligence; MMWR report on LLM-assisted outbreak hypothesis generation","entries":[{"category":"content","title":"AI governance framework comparison for health departments","change":"AI Policy and Governance adds a table comparing WHO 2021 ethics guidance, WHO 2024 large multimodal model guidance, NIST AI RMF 1.0, OECD AI Principles, UNESCO's Recommendation on the Ethics of AI, and the EU AI Act by what each contributes and its main limitation, with a matching FAQ entry.","why":"Health departments face a recurring practical question of which framework answers which governance problem, not whether to use one.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"OECD AI Principles","url":"https://www.oecd.org/en/topics/ai-principles.html","type":"policy"},{"label":"UNESCO Recommendation on the Ethics of AI","url":"https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence","type":"policy"}]},{"category":"policy","title":"HTI-1 algorithm transparency in vendor due diligence","change":"Vendor Evaluation Checklist adds a section on the ONC HTI-1 final rule's algorithm transparency requirements for certified health IT, with due-diligence questions on certification status and fairness, validity, and safety documentation.","why":"Gives procurement teams a primary-source basis for algorithm transparency requirements when evaluating vendors.","chapters":[["AI Vendor Evaluation Checklist","/appendices/vendor-evaluation.html"]],"sources":[{"label":"ONC, HTI-1 Final Rule","url":"https://www.healthit.gov/regulations/hti-rules/hti-1-final-rule/","type":"policy"}]},{"category":"evidence","title":"MMWR report on LLM-assisted outbreak hypothesis generation","change":"Disease Surveillance and Outbreak Detection adds a section on an Illinois county fair Salmonella Agbeni investigation, where a health department used ChatGPT 4.0 to generate additional exposure hypotheses after interviews pointed only to fair-tent beer and environmental samples were no longer available; the outbreak conclusion (contaminated ice) still rested on standard epidemiologic evidence.","why":"Gives a documented boundary case for using LLMs in field epidemiology without displacing the standard evidentiary hierarchy.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Houser et al., 2026, MMWR","url":"https://doi.org/10.15585/mmwr.mm7507a1","type":"paper"}]}]},
{"date":"2026-05-21","headline":"AI-generated forecasting models benchmarked against the CDC ensemble; Transfer-learning early warning system deployed across 50 states; British eConsult study on digital access equity","entries":[{"category":"evidence","title":"AI-generated forecasting models benchmarked against the CDC ensemble","change":"Epidemic Forecasting with AI adds a 2026 Nature study in which an LLM and tree-search system (Empirical Research Assistance) generated 14 COVID-19 hospitalization forecasting models that outperformed the CDC ensemble and all individual benchmark models.","why":"Evaluation pipelines need to audit algorithmically generated forecasting models to the same calibration and external-validity standards as human-built ones.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Aygun et al., 2026","url":"https://doi.org/10.1038/s41586-026-10658-6","type":"paper"}]},{"category":"evidence","title":"Transfer-learning early warning system deployed across 50 states","change":"Disease Surveillance and Outbreak Detection adds a 2026 Nature Communications study of a transfer-learning early warning system deployed across all 50 states in the 2024-2025 respiratory season, combining search trends with epidemiological data to detect 98.0% of outbreak onsets with an average 5-week lead time.","why":"The comparison with Google Flu Trends shows search data becomes useful for surveillance only when paired with epidemiological data, anomaly detection, and prospective deployment.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Garrido-Garcia et al., 2026","url":"https://doi.org/10.1038/s41467-026-72655-7","type":"paper"}]},{"category":"evidence","title":"British eConsult study on digital access equity","change":"AI in Global Health and Equity adds an observational analysis of 43,657,891 online consultations across England, Scotland, and Wales from 2019 to 2023, finding peak eConsult coverage of 43.7% of English, 11.6% of Scottish, and 50.5% of Welsh practices, higher in more affluent areas.","why":"Digital front doors need standardized recording, equity monitoring, and escalation audits before they count as access improvements.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Kerr et al., 2025","url":"https://doi.org/10.1136/bmjdhai-2025-000032","type":"paper"}]}]},
{"date":"2026-05-22","headline":"CDC and HHS AI adoption figures","entries":[{"category":"revision","title":"CDC and HHS AI adoption figures","change":"LLMs in Public Health updates federal AI adoption figures: CDC reports a staff chatbot with over $3.7 million in estimated labor cost savings, 103 AI solutions as of December 31, 2025, and an AI Accelerator program that includes TowerScout, Clinical Narratives, and NewsScape. HHS reports 271 AI use cases in FY2024, up 66% from FY2023.","why":"Keeps the chapter's account of federal AI deployment current as CDC and HHS reported updated figures.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[]}]},
{"date":"2026-05-26","headline":"CDC AI Strategy and GenAI guidance for public health agencies; Systematic review on generative AI and health misinformation; Federated learning review for cross-border disease surveillance; CDC data modernization milestones as AI prerequisites","entries":[{"category":"policy","title":"CDC AI Strategy and GenAI guidance for public health agencies","change":"LLMs in Public Health and AI Policy and Governance add CDC's FY 2026-2030 AI Strategy and companion GenAI considerations, which frame GenAI adoption for state, tribal, local, and territorial agencies as a governance decision: define permitted uses, protect sensitive data, require human review, and document use before scaling.","why":"Public health agencies need concrete task-fit and governance guardrails as they adopt generative AI tools.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"CDC, March 2026","url":"https://www.cdc.gov/ai/resources/considerations-for-genai-in-public-health.html","type":"policy"},{"label":"CDC, April 2026","url":"https://www.cdc.gov/public-health-data-strategy/php/about/phds-milestones.html","type":"policy"}]},{"category":"evidence","title":"Systematic review on generative AI and health misinformation","change":"AI, Misinformation, and the Infodemic adds a 2026 BMC Public Health systematic review finding that generative AI increases the volume, speed, and perceived credibility of health disinformation, while detection and labeling mitigations remain early-stage and under-tested.","why":"Users often cannot distinguish AI-generated from human-authored health misinformation, and mitigation tools are not yet reliable.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[{"label":"Saeidnia et al., 2026","url":"https://doi.org/10.1186/s12889-025-26148-9","type":"paper"}]},{"category":"evidence","title":"Federated learning review for cross-border disease surveillance","change":"Disease Surveillance and Outbreak Detection and AI in Global Health and Equity add a 2026 Nature Medicine review of federated learning for infectious disease surveillance, which keeps data local across sources but remains limited in adoption, depending on shared protocols, governance, and local analytic capacity.","why":"Outbreak data that is confidential or cannot be pooled can still be analyzed across sources if it stays local under a federated approach.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"],["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Khurana et al., 2026","url":"https://doi.org/10.1038/s41591-026-04351-4","type":"paper"}]},{"category":"content","title":"CDC data modernization milestones as AI prerequisites","change":"The Data Problem in Public Health AI adds a section on CDC's 2026 Public Health Data Strategy milestones: at least 86% of U.S. emergency departments sending syndromic data, automated bed-capacity feeds from at least 17 state and territorial jurisdictions, and electronic case reporting processing for at least 70% of ELC-funded health departments connected to AIMS.","why":"Models built on delayed or inconsistent data feeds inherit those defects, so data infrastructure determines whether AI can be evaluated at all.","chapters":[["The Data Problem in Public Health AI","/foundations/data.html"]],"sources":[]}]},
{"date":"2026-05-27","headline":"NIST AI 800-4 post-deployment monitoring guidance","entries":[{"category":"policy","title":"NIST AI 800-4 post-deployment monitoring guidance","change":"AI Deployment: Why Most Prototypes Fail adds a reference to NIST's Center for AI Standards and Innovation (CAISI) catalog of post-deployment monitoring challenges and monitoring categories in NIST AI 800-4.","why":"Gives deployment teams a primary-source catalog of post-deployment monitoring challenges and categories.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[{"label":"Rao et al., 2026, NIST AI 800-4","url":"https://doi.org/10.6028/NIST.AI.800-4","type":"report"}]}]},
{"date":"2026-05-29","headline":"NIST AI RMF critical infrastructure profile concept note","entries":[{"category":"policy","title":"NIST AI RMF critical infrastructure profile concept note","change":"AI Policy and Governance and AI Deployment: Why Most Prototypes Fail add NIST's April 7, 2026 draft concept note for a Trustworthy AI in Critical Infrastructure profile, extending the NIST AI RMF's govern, map, measure, and manage functions toward critical infrastructure use.","why":"Gives readers the primary source for NIST's newest critical-infrastructure extension of the AI RMF.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"],["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[{"label":"NIST, 2026, draft concept note","url":"https://www.nist.gov/system/files/documents/2026/04/08/Draft%20Concept%20Note_%20Development%20of%20the%20NIST%20AI%20RMF%20Trustworthy%20Use%20of%20AI%20in%20Critical%20Infrastructure%20Profile.pdf","type":"policy"}]}]},
{"date":"2026-06-01","headline":"FDA final guidance on Clinical Decision Support software; Federal surveillance database pause count; Multi-agent AI scientist systems in Nature; Licensure framework for autonomous clinical AI","entries":[{"category":"policy","title":"FDA final guidance on Clinical Decision Support software","change":"The Executive Summary now cites FDA's updated final guidance on Clinical Decision Support software, issued January 6, 2026 and re-issued January 29, 2026, clarifying which CDS functions are excluded from the device definition under section 520(o)(1)(E).","why":"Distinguishes CDS functions excluded from the device definition from those subject to device oversight.","chapters":[["Executive Summary","/executive-summary.html"]],"sources":[{"label":"FDA, Clinical Decision Support Software guidance, 2026","url":"https://www.fda.gov/media/109618/download","type":"policy"},{"label":"FDA, Clinical Decision Support Software FAQs, 2026","url":"https://www.fda.gov/medical-devices/software-medical-device-samd/clinical-decision-support-software-frequently-asked-questions-faqs","type":"policy"}]},{"category":"evidence","title":"Federal surveillance database pause count","change":"Disease Surveillance and Outbreak Detection now cites a count of federal data disruptions: 46% of CDC databases updated at least monthly were paused with no explanation, and 87% of paused databases were vaccination-related, as of October 28, 2025.","why":"Quantifies the surveillance feed disruptions that the chapter says AI models calibrated on those feeds need to audit for.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Annals of Internal Medicine, 2025","url":"https://doi.org/10.7326/ANNALS-25-04022","type":"paper"}]},{"category":"evidence","title":"Multi-agent AI scientist systems in Nature","change":"Disease Surveillance and Outbreak Detection notes that Nature studies on multi-agent AI scientist systems show them generating hypotheses and proposing experiments, but describes this as disciplined hypothesis expansion under human accountability, not autonomous decision-making.","why":"Clarifies that this new multi-agent research literature does not extend to autonomous surveillance decisions.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Gottweis et al., 2026","url":"https://doi.org/10.1038/s41586-026-10644-y","type":"paper"},{"label":"Ghareeb et al., 2026","url":"https://doi.org/10.1038/s41586-026-10652-y","type":"paper"}]},{"category":"evidence","title":"Licensure framework for autonomous clinical AI","change":"AI Policy and Governance adds a note on a proposed licensure-based approach for autonomous clinical AI, urging agencies to require a defined scope of practice, clinician escalation boundaries, supervised deployment evidence, outcome monitoring, and authority to restrict or revoke use after errors.","why":"A JAMA Perspective framed AI systems that make care decisions without clinician review as a practice-authorization question for agencies considering AI as a workforce intervention.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Bergman et al., 2026","url":"https://doi.org/10.1001/jama.2026.5483","type":"paper"}]}]},
{"date":"2026-06-03","headline":"Expanded FDA Clinical Decision Support guidance detail","entries":[{"category":"revision","title":"Expanded FDA Clinical Decision Support guidance detail","change":"AI Policy and Governance expands its January 2026 Clinical Decision Support Software Guidance section with a table covering the non-device CDS threshold, recommendation type, transparency requirement, and device examples such as risk scores and image analysis.","why":"Gives readers FDA's specific statutory criteria and device examples in place of a secondary law-firm analysis of the guidance.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[]}]},
{"date":"2026-06-10","headline":"Anthropic Claude for Healthcare and Life Sciences","entries":[{"category":"developments","title":"Anthropic Claude for Healthcare and Life Sciences","change":"LLMs in Public Health describes Anthropic's January 2026 Claude for Healthcare product, with HIPAA-ready connectors for CMS, ICD-10, NPI, and PubMed, and a separate Claude for Life Sciences expansion adding Medidata and ClinicalTrials.gov connectors.","why":"Anthropic launched distinct healthcare and life-sciences product lines with different connectors and skills, a distinction the chapter's technology survey now tracks.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Anthropic, January 2026","url":"https://www.anthropic.com/news/healthcare-life-sciences","type":"announcement"},{"label":"Claude for Life Sciences","url":"https://claude.com/solutions/life-sciences","type":"announcement"}]}]},
{"date":"2026-06-23","headline":"Nature Medicine study: general-purpose LLMs outperform specialized clinical AI tools","entries":[{"category":"evidence","title":"Head-to-head evaluation of clinical AI tools","change":"Evaluating AI Systems for Healthcare adds a Nature Medicine study comparing OpenEvidence and UpToDate Expert AI against general-purpose large language models on medical benchmarks and physician queries, finding general-purpose LLMs outperformed both specialized tools.","why":"The study did not evaluate downstream patient outcomes, so it supports further implementation and safety testing rather than routine clinical deployment.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Vishwanath et al., 2026","url":"https://doi.org/10.1038/s41591-026-04431-5","type":"paper"}]}]},
{"date":"2026-06-25","headline":"NACCHO360 AI for Public Health conference track added to Further Reading","entries":[{"category":"content","title":"NACCHO360 AI for Public Health track","change":"Further Reading adds the NACCHO360 conference's 2026 AI for Public Health track as practitioner-oriented programming for local health departments.","why":"Gives readers a practitioner-oriented conference track focused on AI in local public health practice.","chapters":[["Further Reading","/appendices/resources.html"]],"sources":[{"label":"NACCHO360, 2026 AI for Public Health track","url":"https://www.naccho360.org/naccho360/schedule/ai-for-public-health-track","type":"other"}]}]},
{"date":"2026-06-29","headline":"DeepSeek's low-cost deployment across 90 Chinese tertiary hospitals","entries":[{"category":"evidence","title":"DeepSeek deployment in Chinese hospitals","change":"Emerging AI Technologies adds a 2025 Journal of Medical Systems study documenting DeepSeek deployment across 90 Chinese tertiary hospitals, illustrating how efficient model architectures enable local deployment without ongoing API costs, though clinical impact data remains unreported.","why":"Gives readers a cost and deployment data point for resource-limited settings evaluating open-weight models.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Chen et al., 2025","url":"https://doi.org/10.1007/s10916-025-02181-4","type":"paper"}]}]},
{"date":"2026-07-11","headline":"TAKE IT DOWN Act enforcement; multilingual AFP surveillance chatbot in Ethiopia","entries":[{"category":"policy","title":"TAKE IT DOWN Act and FTC enforcement","change":"AI, Misinformation, and the Infodemic adds the TAKE IT DOWN Act (Public Law 119-12), which addresses nonconsensual intimate imagery including AI-generated digital forgeries and requires covered platforms to remove reported content within 48 hours, and notes FTC enforcement of the platform requirements began May 19, 2026.","why":"No broad federal law governs AI-generated misinformation, so the chapter tracks this narrower law as the closest federal action.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[{"label":"Congress, 2025","url":"https://www.govinfo.gov/content/pkg/PLAW-119publ12/pdf/PLAW-119publ12.pdf","type":"policy"},{"label":"FTC, 2026","url":"https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-begins-enforcing-take-it-down-act","type":"policy"}]},{"category":"evidence","title":"Multilingual chatbot for AFP surveillance in Ethiopia","change":"Disease Surveillance and Outbreak Detection adds a 2026 study of AFP Assistant, a retrieval-augmented chatbot trained on 468 expert-validated question-and-answer pairs, supporting Amharic, Afaan Oromo, and English for acute flaccid paralysis surveillance in Ethiopia, reporting 0.90 accuracy in in-house testing without yet establishing field effectiveness.","why":"Gives readers a documented approach to multilingual surveillance interfaces in low-resource language settings.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Biru et al., 2026","url":"https://doi.org/10.1038/s41746-026-02944-4","type":"paper"}]}]},
{"date":"2026-07-13","headline":"MuILI multilingual influenza dataset; Lao PDR immunization registry evaluation","entries":[{"category":"evidence","title":"Multilingual influenza surveillance dataset (MuILI)","change":"Disease Surveillance and Outbreak Detection adds a 2026 study introducing MuILI, a dataset of 4,284 tweets annotated for influenza-like illness across five European languages, comparing multilingual language models and cross-lingual transfer approaches across France, Italy, Spain, and Germany, though it does not establish performance in low-resource languages.","why":"Gives readers a benchmark for multilingual social-media surveillance classifiers beyond English-language tools.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Mishra et al., 2026","url":"https://doi.org/10.1038/s44482-026-00020-y","type":"paper"}]},{"category":"evidence","title":"Lao PDR electronic immunization registry evaluation","change":"AI in Global Health and Equity adds a mixed-methods evaluation of Lao PDR's DHIS2 Tracker-based electronic immunization registry, finding high user acceptance but vaccine type missing for 15.66% of events, vaccination date missing for 14.91%, and only 15.26% of vaccinations recorded within three days.","why":"Shows that workforce capacity, electricity, connectivity, governance, and financing remain prerequisites before registry data can support forecasting or coverage analysis.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Patel et al., 2026","url":"https://doi.org/10.1038/s44482-026-00018-6","type":"paper"}]}]},
{"date":"2026-07-17","headline":"AMR prediction generalization benchmark; Quantitative tuberculosis resistance prediction; Resistome surveillance needs genetic and functional context; Wastewater AMR surveillance and forecasting baselines; AMRnet described as rule-based infrastructure, not AI; One Health AMR surveillance and limits on transmission inference","entries":[{"category":"evidence","title":"AMR prediction generalization benchmark","change":"Genomic Surveillance and Pathogen Analysis adds a benchmark spanning 78 species-antibiotic datasets showing machine-learning resistance prediction methods perform best for closely related strains, while the rule-based ResFinder system handles divergent genomes better, with performance varying substantially for several beta-lactams.","why":"Random train-test partitions can overstate operational accuracy, so the chapter frames generalization across geography, time, and genetic relatedness as the key evaluation question.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Hu et al., 2024","url":"https://doi.org/10.1093/bib/bbae206","type":"paper"}]},{"category":"evidence","title":"Quantitative tuberculosis resistance prediction","change":"Genomic Surveillance and Pathogen Analysis adds a 2026 study of convolutional neural networks predicting minimum inhibitory concentrations for eight Mycobacterium tuberculosis drugs from gene sequences, with 89% of predictions within one twofold dilution, and higher predicted rifampicin MICs associated with unfavorable outcomes in 373 South African patients.","why":"Quantitative MIC prediction can shorten the time to an initial resistance estimate for tuberculosis treatment decisions.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Kulkarni et al., 2026","url":"https://doi.org/10.1038/s41467-026-72225-x","type":"paper"}]},{"category":"evidence","title":"Resistome surveillance needs genetic and functional context","change":"Genomic Surveillance and Pathogen Analysis adds a 2026 review finding that host attribution and mobile-element resolution remain major analytical challenges for resistome surveillance, even as detection of resistance genes within microbial communities has improved.","why":"Detecting a resistance gene does not by itself establish which organism carries it or whether it is expressed, a distinction surveillance teams need before acting on the data.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Chaudhry et al., 2026","url":"https://doi.org/10.1038/s41467-026-75659-5","type":"paper"}]},{"category":"evidence","title":"Wastewater AMR surveillance and forecasting baselines","change":"Genomic Surveillance and Pathogen Analysis adds a U.S. study mapping 11 resistance genes across 163 wastewater treatment plants in 40 states, and an English hospital-group study where XGBoost forecasts of resistance prevalence outperformed a naive baseline mainly when prevalence changed substantially.","why":"Establishes an explicit baseline for judging whether a forecasting model adds value over simply carrying forward the prior year's resistance level.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Kim et al., 2026","url":"https://doi.org/10.1038/s41467-026-71195-4","type":"paper"},{"label":"Vihta et al., 2024","url":"https://doi.org/10.1038/s43856-024-00606-8","type":"paper"}]},{"category":"evidence","title":"AMRnet described as rule-based infrastructure, not AI","change":"Genomic Surveillance and Pathogen Analysis adds AMRnet, a public genome-derived resistance dashboard and API, describing its indicators as calculated from known markers and imported rules rather than learned by an AI model.","why":"Distinguishes surveillance infrastructure that applies known rules from AI systems that learn patterns from data.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"AMRnet, genomic AMR surveillance platform","url":"https://www.amrnet.org/","type":"other"},{"label":"Cerdeira et al., 2026","url":"https://doi.org/10.1093/nar/gkaf1101","type":"paper"}]},{"category":"revision","title":"One Health AMR surveillance and limits on transmission inference","change":"AI in Global Health and Equity cautions that detecting the same resistance marker in more than one sector does not by itself establish cross-species transmission; Genomic Surveillance and Pathogen Analysis is the chapter for resistome analysis, wastewater AMR surveillance, and AMRnet.","why":"Distinguishes population-level AMR surveillance from patient-level susceptibility and cross-species transmission claims.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[]}]},
{"date":"2026-07-22","headline":"Evidence matched to five claim types, from technical performance to lifecycle safety","entries":[{"category":"revision","title":"Evidence matched to five claim types, from technical performance to lifecycle safety","change":"Evaluating AI Systems for Healthcare matches evidence to five claim types: technical performance, transportability, operational utility, population impact, and lifecycle safety.","why":"A single evidence ladder does not fit every deployment question, so the chapter now asks readers to match the study design and appraisal method to the specific claim and deployment risk.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["Quick Reference: All Chapter Summaries (TL;DRs)","/appendices/tldr-compilation.html"]],"sources":[]},{"category":"evidence","title":"External validation gap in medical-imaging AI studies","change":"Evaluating AI Systems for Healthcare and the Chapter Summaries TL;DR cite a review of 516 medical-imaging AI studies in which only 31 used external validation, alongside the previously cited finding that a widely implemented sepsis model had 33% sensitivity and 12% positive predictive value on external evaluation.","why":"Shows that most published AI validation stops short of the external testing needed to judge transportability to new settings.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["Quick Reference: All Chapter Summaries (TL;DRs)","/appendices/tldr-compilation.html"]],"sources":[{"label":"Kim et al., 2019","url":"https://doi.org/10.3348/kjr.2019.0025","type":"paper"}]},{"category":"policy","title":"GRADE certainty-of-evidence framing","change":"Evaluating AI Systems for Healthcare cites the CDC ACIP GRADE Handbook to clarify that GRADE assigns certainty to a body of evidence for a specific outcome, not to a journal, vendor, individual paper, or generic validation stage.","why":"Gives readers a primary methodological source for judging evidence certainty instead of where or by whom a study was published.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"CDC ACIP GRADE Handbook, 2024","url":"https://www.cdc.gov/acip-grade-handbook/hcp/chapter-7-grade-criteria-determining-certainty-of-evidence/index.html","type":"policy"}]},{"category":"evidence","title":"Reporting standards for locked-model evaluation","change":"Evaluating AI Systems for Healthcare cites TRIPOD+AI for transparent reporting and PROBAST+AI for quality and risk-of-bias appraisal of locked-model evaluations that assess calibration, threshold-specific errors, missing data, and subgroup performance.","why":"Gives readers current reporting and appraisal standards for the technical-performance claim.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Collins et al., 2024","url":"https://doi.org/10.1136/bmj-2023-078378","type":"paper"},{"label":"Moons et al., 2025","url":"https://doi.org/10.1136/bmj-2024-082505","type":"paper"}]},{"category":"evidence","title":"Operational utility standards for surveillance systems","change":"Evaluating AI Systems for Healthcare cites CDC surveillance-system attributes guidelines and, for live decision support, DECIDE-AI reporting items to judge whether a system works in public health operations.","why":"Gives readers standards for prospective evaluation of data pipelines, timeliness, workflow, and failure modes before operational deployment.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"CDC, 2001","url":"https://www.cdc.gov/mmwr/preview/mmwrhtml/rr5013a1.htm","type":"policy"},{"label":"Vasey et al., 2022","url":"https://doi.org/10.1038/s41591-022-01772-9","type":"paper"}]},{"category":"evidence","title":"CONSORT-AI for population impact claims","change":"Evaluating AI Systems for Healthcare cites CONSORT-AI reporting standards for randomized AI intervention trials that provide comparative evidence on decisions, outcomes, harms, equity, and resource use.","why":"Gives readers a reporting standard for the strongest form of evidence on whether using a system actually improves outcomes.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Liu et al., 2020","url":"https://doi.org/10.1038/s41591-020-1034-x","type":"paper"}]},{"category":"evidence","title":"NICE framework for lifecycle safety monitoring","change":"Evaluating AI Systems for Healthcare cites the NICE Evidence Standards Framework for Digital Health Technologies, which separates design, value, performance, deployment, and post-deployment evidence for ongoing monitoring of drift, adverse effects, and unintended consequences.","why":"Gives readers a framework for judging whether benefit persists after deployment rather than treating validation as a one-time event.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"NICE, Evidence Standards Framework for Digital Health Technologies","url":"https://www.nice.org.uk/what-nice-does/digital-health/evidence-standards-framework-esf-for-digital-health-technologies","type":"other"}]}]},
{"date":"2026-07-28","headline":"Statistical rigor caveat for MedQA benchmark comparisons","entries":[{"category":"evidence","title":"Statistical rigor caveat for MedQA benchmark comparisons","change":"Evaluating AI Systems for Healthcare adds a caveat that MedQA benchmark figures are bare point estimates with no uncertainty attached, citing a 2024 statistical framework for language model evaluations that recommends paired comparisons when models are scored on the same question set.","why":"A small percentage-point gap between benchmarks can reflect question-set noise rather than a real capability difference; a companion handbook carries the full statistical treatment.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Miller, 2024, preprint","url":"https://arxiv.org/abs/2411.00640","type":"preprint"}]}]},
{"date":"2026-07-30","headline":"Dermatology AI equity study; KFF poll on patient use of health AI","entries":[{"category":"evidence","title":"Dermatology AI skin-tone labeling evidence","change":"Diagnostic and Clinical Decision Support renames its dermatology case study to Image-Based Equity and cites a 2025 study finding that no assessed skin-tone labeling method generalized reliably across in-person assessment, clinical photographs, and dermoscopy images.","why":"Clarifies that Fitzpatrick skin type is not a validated proxy for skin tone, race, or a clinical image's optical properties, so programs need a justified, reported measurement method.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Weir et al., 2025","url":"https://doi.org/10.1038/s41746-025-02245-2","type":"paper"}]},{"category":"content","title":"Pathogen Genomics for Public Health course","change":"Genomic Surveillance and Pathogen Analysis adds a Courses and Training section linking to open-access, self-paced courses on genomic epidemiology and sequencing strategies, developed under the U.S. Pathogen Genomics Centers of Excellence and supported by the CDC Office of Advanced Molecular Detection.","why":"Gives practitioners a free training pathway into genomic epidemiology and outbreak-response sequencing.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[{"label":"Pathogen Genomics for Public Health course, Harvard Medical School","url":"https://learn.hms.harvard.edu/pathogen-genomics-public-health","type":"other"}]},{"category":"content","title":"Genomic surveillance FAQ","change":"Genomic Surveillance and Pathogen Analysis adds an FAQ section covering how quickly the first SARS-CoV-2 genome was shared, the difference between GISAID and Nextstrain, how SNP distances support transmission inference, and what quality metrics matter for genomic surveillance.","why":"Answers common practitioner questions about genomic surveillance tools and terminology in one place.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"sources":[]},{"category":"evidence","title":"KFF poll on patient use of AI for health information","change":"Diagnostic and Clinical Decision Support cites a March 2026 nationally representative KFF poll in which 65% of users cited a need for immediate information as a reason for using AI for health information, and 18% cited lacking a regular clinician or an available appointment.","why":"Frames patient chatbot use as a possible signal of unmet access and communication needs, not only a model-safety question, and calls for evaluating whether deployments hand off to human care.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"KFF, 2026","url":"https://www.kff.org/public-opinion/kff-tracking-poll-on-health-information-and-trust-use-of-ai-for-health-information-and-advice/","type":"report"}]}]},
{"date":"2026-08-06","headline":"Dengue risk-mapping study added to emerging technologies","entries":[{"category":"evidence","title":"Dengue risk-mapping study added to emerging technologies","change":"Adds a Colombia dengue risk-mapping study to Emerging AI Technologies: a spatiotemporal U-Net++ model combining satellite, census, climatic, and socioeconomic data reached a test mIoU of 0.66 nationally (0.74 in validation), a generalization gap the chapter states explicitly; the work remains research-stage without operational validation.","why":"The chapter previously covered dengue only as temporal forecasting, with no geospatial risk-stratification evidence.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Velandia et al., 2026","url":"https://doi.org/10.1038/s41598-026-63542-8","type":"paper"}]}]},
{"date":"2026-08-09","headline":"New chapters on LLMs, deployment, evaluation, and global health","entries":[{"category":"milestone","title":"New chapters on LLMs, deployment, evaluation, and global health","change":"Adds focused chapters covering topics such as LLM foundations and selection, MLOps, deployment operations, performance metrics, explainability, validation, and global-health implementation, equity, and data governance.","why":"Provides dedicated reference chapters for model selection, deployment, evaluation, equity, and data governance.","chapters":[["AI in Global Health and Equity","/future/global-health.html"],["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"],["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"],["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"],["MLOps for Public Health AI","/implementation/mlops.html"],["Performance Metrics for Public Health AI","/implementation/performance-metrics.html"],["Large Language Model Foundations for Public Health","/future/llm-foundations.html"],["Global Health AI Data Governance","/future/global-data-governance.html"]],"sources":[]},{"category":"content","title":"New chapter: Geospatial AI and Spatial Epidemiology","change":"Adds a new chapter on geospatial AI, covering an operational example (CDC's TowerScout computer-vision tool for locating cooling towers in Legionnaires' disease investigations) and validation approaches designed for spatial data.","why":"Health outcomes can cluster for many reasons besides transmission, including environmental conditions, healthcare access, and how boundaries are drawn.","chapters":[["Geospatial AI and Spatial Epidemiology","/applications/geospatial.html"]],"sources":[{"label":"WHO, GIS Centre for Health","url":"https://www.who.int/data/GIS","type":"policy"},{"label":"WHO Regional Office for Europe, 2026","url":"https://www.who.int/europe/publications/i/item/WHO-EURO-2026-11407-51179-77989","type":"policy"},{"label":"CDC, AI's Role in Stopping Legionnaires' Disease, 2026","url":"https://www.cdc.gov/ai/ai-success-stories/ais-role-in-stopping-legionnaires-disease.html","type":"policy"},{"label":"CDC, Identifying Cooling Towers, 2026","url":"https://www.cdc.gov/investigate-legionella/php/public-health-strategy/identifying-cooling-towers.html","type":"policy"},{"label":"Iyer et al., 2025","url":"https://doi.org/10.1007/s40572-025-00497-4","type":"paper"}]},{"category":"content","title":"New chapter: Emerging AI Architectures for Public Health","change":"Adds a new chapter on multimodal, agentic, and domain-specific AI architectures for public health, covering the shift from chatbots to agents and multimodal systems and the evidence base for retrieval-augmented generation (RAG) in healthcare.","why":"Keeps pace with AI capabilities beyond single-turn chat that general LLM guidance does not cover.","chapters":[["Emerging AI Architectures for Public Health","/future/ai-architectures.html"]],"sources":[]},{"category":"content","title":"New chapter: AI in Public Health Emergency Operations","change":"Adds a new chapter on using AI during public health emergency response, built around an \"Emergency AI Control Loop\" that defines authority and decision rights, sets an approved information boundary, and requires traceable outputs.","why":"AI can speed up finding patterns in an emergency, but the chapter says command decisions are not simply an information-retrieval problem.","chapters":[["AI in Public Health Emergency Operations","/implementation/emergency-operations.html"]],"sources":[]},{"category":"evidence","title":"Ensemble forecasting guidance calibrated with new evaluation evidence","change":"Updates Epidemic Forecasting with AI, its TL;DR, and the Executive Summary to note that ensembles are a strong default but not an unconditional rule, citing CDC's 2024-2025 FluSight evaluation and a disease-agnostic learned-weight ensemble study that requires an evidence gate before weighting models (Murph et al., 2026, Nature Communications).","why":"Recent operational evidence shows ensembles do not always outperform, so the chapter now states when the default should be tested rather than assumed.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"],["Quick Reference: All Chapter Summaries (TL;DRs)","/appendices/tldr-compilation.html"],["Executive Summary","/executive-summary.html"]],"sources":[{"label":"CDC, 2024-2025 FluSight Evaluation","url":"https://www.cdc.gov/flu-forecasting/evaluation/2024-2025-report.html","type":"report"},{"label":"Murph et al., 2026, Nature Communications","url":"https://doi.org/10.1038/s41467-026-70937-8","type":"paper"}]}]},
{"date":"2026-08-11","headline":"Medical-countermeasure authority separated from AI threat detection","entries":[{"category":"policy","title":"Medical-countermeasure authority section added to emergency operations","change":"Adds a section to AI in Public Health Emergency Operations distinguishing AI-supported threat detection from the authorization of medical countermeasures, citing FDA's countermeasure framework, WHO's 2026 national public health agency framework, and FDA's post-dispensing monitoring guidance; a cited learning-health-system model is noted as a framework, not evidence.","why":"AI can inform the common operating picture, but the chapter states it cannot authorize a countermeasure or establish that a product fits a population.","chapters":[["AI in Public Health Emergency Operations","/implementation/emergency-operations.html"]],"sources":[{"label":"FDA, What Are Medical Countermeasures, 2024","url":"https://www.fda.gov/emergency-preparedness-and-response/about-mcms/what-are-medical-countermeasures","type":"policy"},{"label":"WHO, National Public Health Agency Framework, 2026","url":"https://www.who.int/publications/i/item/B09726","type":"policy"},{"label":"FDA, Medical Countermeasure Monitoring and Assessment, accessed August 2026","url":"https://www.fda.gov/emergency-preparedness-and-response/about-mcms/medical-countermeasure-monitoring-and-assessment","type":"policy"},{"label":"Ankolekar et al., 2024","url":"https://doi.org/10.1016/j.csbj.2024.05.014","type":"paper"}]}]},
{"date":"2026-08-12","headline":"FAQ adds a surveillance and evaluation section","entries":[{"category":"content","title":"FAQ adds a surveillance and evaluation section","change":"The FAQ chapter adds a Surveillance and Evaluation section with new questions, including what role AI can play in disease surveillance and how a health department should evaluate an AI tool before deployment.","why":"Gives readers direct answers to common surveillance and evaluation questions in one place.","chapters":[["Frequently Asked Questions","/appendices/faq.html"]],"sources":[]}]},
{"date":"2026-08-16","headline":"TB REACH India study on AI-assisted chest radiography yield; FAQ sections added to nearly every chapter","entries":[{"category":"evidence","title":"TB REACH India study on AI-assisted chest radiography yield","change":"The clinical AI chapter now cites a TB REACH active case-finding program in Nagpur, India, in which 8,303 chest radiographs were read by both a radiologist and AI software (qXR); the program reported about a 15.8% increase in overall tuberculosis yield attributable to qXR alone.","why":"Reports the additional tuberculosis yield attributable to AI-assisted chest radiography in an active case-finding program.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Vijayan et al., 2023, PLOS Digital Health","url":"https://doi.org/10.1371/journal.pdig.0000404","type":"paper"}]},{"category":"content","title":"FAQ sections added to nearly every chapter","change":"Most chapters across the foundations, applications, implementation, future, and practical parts, plus the welcome page, preface, executive summary, and several appendices, add a Frequently Asked Questions section answering chapter-specific questions, including a surveillance-chapter answer clarifying that BlueDot alerted clients to the Wuhan pneumonia cluster the same day WHO was informed, not nine days earlier.","why":"Readers get quick, sourced answers to common questions, including the timing of BlueDot's COVID-19 alert, widely repeated as nine days before WHO was informed.","chapters":[["Welcome","/"],["Preface","/preface.html"],["Executive Summary","/executive-summary.html"],["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"],["Privacy, Security, and Governance for Health AI","/implementation/privacy.html"],["AI Policy and Governance in Healthcare","/future/policy.html"],["AI in Healthcare: A Brief History","/foundations/history.html"],["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[]}]},
{"date":"2026-08-18","headline":"MLOps chapter adds regression testing guidance for agentic AI systems","entries":[{"category":"content","title":"MLOps chapter adds regression testing guidance for agentic systems","change":"The MLOps chapter adds a new section on regression testing for generative and agentic public health tools: freezing the full configured system (model, prompts, tools, retrieval, permissions), replaying held-out failure cases, running repeated trials because stochastic systems can behave inconsistently, and calibrating automated judges against blinded human experts.","why":"Gives implementers a concrete method for catching regressions in agentic tools that a single successful demo would miss.","chapters":[["MLOps for Public Health AI","/implementation/mlops.html"]],"sources":[{"label":"NIST AI 800-2, initial public draft, 2026","url":"https://doi.org/10.6028/NIST.AI.800-2.ipd","type":"policy"},{"label":"Yao et al., 2025","url":"https://proceedings.iclr.cc/paper_files/paper/2025/hash/1b126cc38b8638e07bef37e7b2bb72bf-Abstract-Conference.html","type":"paper"},{"label":"Chen et al., 2024","url":"https://doi.org/10.18653/v1/2024.emnlp-main.474","type":"paper"}]}]},
{"date":"2026-08-19","headline":"Output-format guidance distinguishes formatting from schema validation","entries":[{"category":"revision","title":"Output-format guidance distinguishes formatting from schema validation","change":"The LLM-selection chapter's guidance on requesting structured output now distinguishes format instructions, which make a response easier to inspect, from actual schema enforcement, and recommends provider-supported constrained generation, explicit schema validation, rejecting unknown fields, and tested retry behavior for production workflows.","why":"Prevents readers from treating a well-formatted response as proof that its structure was actually validated.","chapters":[["Selecting and Using Large Language Models","/future/llm-selection-use.html"]],"sources":[]},{"category":"evidence","title":"Error rates depend on the model, task, prompt, and scoring method","change":"LLMs in Public Health now cites research showing reported rates of unsupported output are evaluation-specific and do not transfer across models, tasks, prompts, and scoring methods, alongside the earlier finding that Alkaissi and McFarlane documented false citations in ChatGPT case exercises.","why":"Supports the chapter's caution against citing a single error-rate figure as if it generalized across tools and tasks.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Ji et al., 2023","url":"https://doi.org/10.1145/3571730","type":"paper"},{"label":"Schmidtova et al., 2025","url":"https://aclanthology.org/2025.inlg-main.4/","type":"paper"}]}]},
{"date":"2026-08-21","headline":"LLM chapter tightens PHI and Business Associate Agreement guidance; Ethics chapter clarifies what a bias audit examines","entries":[{"category":"revision","title":"LLM chapter tightens PHI and Business Associate Agreement guidance","change":"The LLM-in-public-health chapter sets the rule for protected health information: covered entities and business associates must not disclose PHI to an LLM vendor unless HIPAA permits it and safeguards are in place, a Business Associate Agreement is required when the vendor acts as a business associate, and consumer accounts should not receive PHI.","why":"Gives readers a clearer, more specific rule for when PHI can and cannot go to an LLM vendor.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[]},{"category":"revision","title":"Ethics chapter clarifies what a bias audit examines","change":"The ethics chapter defines a bias audit as an examination of whether an AI system produces systematically different performance, access, or outcomes across populations and whether the evidence for deployment is adequate, citing Obermeyer et al., who found a widely deployed risk-prediction algorithm disadvantaged Black patients by using spending as a proxy for need.","why":"Sharpens the working definition practitioners use when they run their own bias audits.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[]}]},
{"date":"2026-08-28","headline":"Swedish registry study added on hip-fracture prediction and PPV limits","entries":[{"category":"evidence","title":"Swedish registry hip-fracture prediction model added","change":"The clinical AI chapter adds a nationwide Swedish cohort of 3,542,647 adults with 142,327 incident hip fractures, where a DeepSurv model (FRACTURE-ML) reached a holdout AUC of 0.89 at 1 year; at 80% sensitivity and a 2.3% five-year incidence, precision was about 7%, with no external validation or implementation trial reported.","why":"Extends the chapter's PPV-at-low-incidence lesson to a non-imaging, registry-based prediction model.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Axelsson et al., 2026","url":"https://doi.org/10.1371/journal.pmed.1005190","type":"paper"}]}]},
{"date":"2026-08-30","headline":"AI regulation and organizational governance; Emerging-technologies chapter adds evidence-tiered readiness framework; Global health AI: evaluating the complete service","entries":[{"category":"content","title":"AI regulation and organizational governance","change":"The AI-policy chapter distinguishes regulation, which sets legal duties, from organizational governance, which sets who may approve, monitor, and stop an AI system's use, and compares FDA, EU, and international regulation.","why":"Gives readers a structured way to separate legal classification from an organization's own governance decisions.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"EU AI Act, 2024","url":"https://eur-lex.europa.eu/legal-content/EN/ALL/?uri=CELEX%3A32024R1689","type":"policy"},{"label":"NIST AI Risk Management Framework","url":"https://airc.nist.gov/airmf-resources/airmf/5-sec-core/","type":"policy"}]},{"category":"content","title":"Emerging-technologies chapter adds evidence-tiered readiness framework","change":"The emerging-technologies chapter classifies claims about foundation models, federated learning, edge AI, digital twins, and quantum computing as demonstrated, theoretical, or beyond current capabilities, citing federated-learning validation guidance, on-device model guidance, and CDC's outbreak-investigation framework.","why":"Gives readers a consistent readiness test instead of treating each new architecture as a single wave of progress.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Rieke et al., 2020","url":"https://doi.org/10.1038/s41746-020-00323-1","type":"paper"},{"label":"LiteRT guidance","url":"https://developers.google.com/edge/litert/conversion/tensorflow/build/overview","type":"other"},{"label":"CDC, outbreak investigation steps","url":"https://www.cdc.gov/foodborne-outbreaks/outbreak-basics/investigation-steps.html","type":"policy"}]},{"category":"content","title":"Global health AI: evaluating the complete service","change":"The global health AI chapter directs readers to evaluate the complete workflow, population, infrastructure, and governance at the intended site rather than a transferred model alone, citing WHO guidance that digital interventions are not substitutes for functioning health systems.","why":"Connects model performance to whether the full service works locally.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"World Health Organization","url":"https://www.who.int/publications/i/item/9789241550505","type":"policy"}]}]},
{"date":"2026-09-01","headline":"McBain et al. on measuring mental-health conversations in AI chat logs","entries":[{"category":"evidence","title":"McBain et al. study on measuring mental-health conversations in chat logs","change":"AI-Driven Behavioral Interventions cites a JAMA Network Open analysis of 620,699 WildChat ChatGPT conversations: an LLM classifier flagged 0.21% as mental-health related under a conservative definition versus 4.90% under an expansive one, a measurement finding rather than a population prevalence estimate.","why":"Crisis-detection rates depend on how a mental-health conversation is defined, so readers should not treat WildChat percentages as a US prevalence rate.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[{"label":"McBain et al., 2026","url":"https://doi.org/10.1001/jamanetworkopen.2026.30635","type":"paper"}]}]},
{"date":"2026-09-02","headline":"Abulibdeh et al. census of FDA-authorized AI/ML devices; Tao et al. RCT of the PreA pre-visit patient-facing LLM; Li et al. RCT of the P&P Care community health chatbot; Chen and Bento on school-level measles transmission risk; Chen et al. systematic review of LLM triage accuracy; Kamulegeya et al. systematic review of AI models for TB treatment failure","entries":[{"category":"evidence","title":"Abulibdeh et al. census of FDA-authorized AI/ML devices","change":"Evaluating AI Systems for Healthcare cites a PLOS Digital Health census of all 1,357 FDA-authorized AI/ML devices through December 2025: only 34 (2.5%) had a registered prospective trial, 12 (0.9%) posted results, and 3 (0.2%) were evaluated on patient-centered outcomes.","why":"Gives the evaluation chapter a primary source bounding how many FDA-authorized devices have prospective, patient-outcome evidence.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Abulibdeh et al., 2026","url":"https://doi.org/10.1371/journal.pdig.0001597","type":"paper"}]},{"category":"evidence","title":"Tao et al. RCT of the PreA pre-visit patient-facing LLM","change":"Diagnostic and Clinical Decision Support adds a multicenter RCT of PreA, a patient-facing LLM used before specialist visits at two Chinese tertiary centers: across 2,069 patients, PreA-only shortened consultation time from 4.41 to 3.14 minutes and raised coordination and communication scores versus no preassessment.","why":"Gives the chapter a workflow-level trial of a patient-facing pre-visit LLM, distinct from evidence on diagnosis accuracy or clinical outcomes.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Tao et al., 2026","url":"https://doi.org/10.1038/s41591-025-04176-7","type":"paper"}]},{"category":"evidence","title":"Li et al. RCT of the P&P Care community health chatbot","change":"Diagnostic and Clinical Decision Support adds a 2,113-person community RCT across 11 Chinese provinces comparing the P&P Care chatbot with versus without preparatory e-learning: health-awareness scores were higher with e-learning (2.95 vs 2.34), with smaller gains among rural participants.","why":"Broadens the chapter's patient-facing AI evidence with a differently designed community trial focused on health awareness rather than diagnosis or care-seeking.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Li et al., 2026","url":"https://doi.org/10.1038/s44360-025-00021-w","type":"paper"}]},{"category":"evidence","title":"Chen and Bento on school-level measles transmission risk","change":"Epidemic Forecasting with AI adds a Nature Medicine gravity-kernel analysis of US school-entry MMR records for more than 50,000 schools, 13,000 districts, and 3,000 counties (2013 to 2025): school-level measles Rᵥ crossed 1 in 2022 to 2023 while county-level averages stayed below threshold, as susceptibility rose from about 5% to 10% after the pandemic.","why":"Gives the spatial-scale point in the forecasting chapter a real-data example instead of a small toy mobility matrix.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Chen and Bento, 2026","url":"https://doi.org/10.1038/s41591-026-04561-w","type":"paper"}]},{"category":"evidence","title":"Chen et al. systematic review of LLM triage accuracy","change":"AI in Public Health Emergency Operations adds an npj Digital Medicine systematic review of 50 studies (25 LLMs) that found pooled triage accuracy of LLMs versus clinicians of 1.01 (95% CI 0.94 to 1.09, I² 94%), with only 3 prospective studies, all triage.","why":"Gives the emergency-operations chapter a pooled accuracy estimate for consumer-facing LLM triage against clinician judgment, most of it from retrospective studies.","chapters":[["AI in Public Health Emergency Operations","/implementation/emergency-operations.html"]],"sources":[{"label":"Chen et al., 2026","url":"https://doi.org/10.1038/s41746-026-02409-8","type":"paper"}]},{"category":"evidence","title":"Kamulegeya et al. systematic review of AI models for TB treatment failure","change":"Evaluating AI Systems for Healthcare adds a PLOS One PRISMA/PROBAST synthesis of 34 studies on AI/ML models for tuberculosis treatment failure: a 19-study pool (100,790 participants) reached AUC 0.836, but only 8 of 34 had external validation and 1 of 34 was low risk of bias overall.","why":"Shows why a pooled AUC should not be read as a tuberculosis-program endpoint.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Kamulegeya et al., 2026","url":"https://doi.org/10.1371/journal.pone.0355124","type":"paper"}]}]},
{"date":"2026-09-04","headline":"Surveillance readiness in LMICs and the Caribbean, a Lassa fever model comparison","entries":[{"category":"policy","title":"Trial-start latency and population access to new therapies","change":"AI Policy and Governance adds a subsection distinguishing drug trial-start latency from device classification. It cites a New York Times opinion arguing that trial-start latency, not discovery science, is the binding bottleneck for AI-designed therapies, and notes that HHS launched Operation TrialBlazer in June 2026 as a clinical-trials reform effort.","why":"Separates the policy question of how quickly trials can start from the SaMD and CDS device-classification analysis already in the chapter.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Teslo, New York Times, September 2026","url":"https://www.nytimes.com/2026/09/04/opinion/clinical-trials-drugs-science.html","type":"news"},{"label":"HHS, Operation TrialBlazer announcement, 2026","url":"https://www.hhs.gov/press-room/hhs-launches-clinical-trials-reform-initiative.html","type":"policy"}]},{"category":"evidence","title":"Francis and Reddy narrative review of LMIC surveillance integration","change":"Disease Surveillance and Outbreak Detection adds a SANRA-structured narrative review of digital, AI, genomic, and vaccine-intelligence pathways toward integrated surveillance in low- and middle-income countries, noting that only about 14% of the AI outbreak-detection models reviewed had external validation in LMIC settings.","why":"Extends the chapter's equity coverage beyond urban and rural digital access to fragmented indicator systems, connectivity, and workforce gaps that also delay outbreak signals in LMICs.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Francis & Reddy, 2026","url":"https://doi.org/10.3389/fpubh.2026.1843782","type":"paper"}]},{"category":"evidence","title":"Addison et al. survey of AI early-warning readiness in Trinidad and Tobago","change":"Disease Surveillance and Outbreak Detection adds a 2026 survey of County Medical Officers of Health in Trinidad and Tobago (6 of 9 responded): all six prioritized AI early warning for infectious disease and flooding and equity for underserved populations, while budget was the leading barrier and familiarity was only low to moderate.","why":"Adds leader-perception evidence from a Small Island Developing State, distinct from evaluating an implemented AI early-warning system.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Addison et al., 2026","url":"https://doi.org/10.3389/fpubh.2026.1931303","type":"paper"}]},{"category":"evidence","title":"Alasow et al. proposed AI-CHEWS for the Horn of Africa","change":"Emerging AI Technologies adds a Globalization and Health comment proposing an IGAD-anchored five-layer AI Climate-Health Early Warning System (AI-CHEWS) for eight Horn of Africa countries, fusing earth-observation, case-reporting, and vulnerability data with SMS fallback in low-connectivity districts; the authors note fewer than 50 dual-trained public-health and data-science professionals across the region.","why":"The system is a proposed design, not a trained model or evaluated deployment; the authors call for future trials measuring health impact rather than only detection accuracy.","chapters":[["Emerging AI Technologies for Public Health","/future/emerging.html"]],"sources":[{"label":"Alasow et al., 2026","url":"https://doi.org/10.1186/s12992-026-01233-9","type":"paper"}]},{"category":"evidence","title":"Tadema et al. comparison of Random Forest and SARIMA for Lassa fever forecasting","change":"Epidemic Forecasting with AI adds a retrospective comparison of climate-informed Random Forest models to a SARIMA benchmark using weekly Lassa fever surveillance from six Nigerian states (2018 to 2025, n = 6,801 cases): Random Forest outperformed SARIMA in three states and SARIMA performed better in the other three, a 3/3 geographic split rather than a uniform result.","why":"Shows that model performance varies by location, so the chapter favors locally tailored climate-informed surveillance over a single national forecasting approach.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Tadema et al., 2026","url":"https://doi.org/10.1186/s12879-026-14203-8","type":"paper"}]}]},
{"date":"2026-09-05","headline":"Banerjee and Nayak scenario cost model for autism identification; Silva et al. comparison of CatBoost and GRU dengue forecasts across Brazil; Huang et al. evaluation of free-tier chatbots on HIV/AIDS FAQs; Abdul Rahman and Tun climate-scenario model of global disease burden; Mello and Nagireddy on phased AI for Medicaid work-requirement eligibility; Google AI Overview compared with specialized clinical evidence tools; Dominikowski et al. and Janvier et al. on gray-market performance-enhancing peptides; Praschk et al. audit of LLM vaccine-stance classification; Qing et al. on VAERS/FAERS signals as a vaccine-safety evaluation control; Saleem Al-Ansary et al. XGBoost classifier for foodborne outbreak detection; Mahmood and Topol on tokenising population health journeys; Jiang et al. ensemble model for school-based vision screening; Fuad, Milki, and Aziz on dengue forecasting complexity in Bangladesh; Xiang et al. on symptom-based pathogen prediction for home testing; Haq et al. analysis of E. coli contamination in Bangladesh drinking water; Kim et al. on survey weighting bias in NHANES-based AI models; Stürenburg et al. classification of global health aid by disease burden; Leibovitch et al. stage 1 DECIDE-AI evaluation of the SHAKED CDSS","entries":[{"category":"evidence","title":"Banerjee and Nayak scenario cost model for autism identification","change":"Diagnostic and Clinical Decision Support adds a calibrated U.S. scenario cost-of-illness model for autism identification delayed by specialist evaluation queues, explicitly framed as a projection exercise rather than an observed national savings figure.","why":"Extends the chapter's bottleneck-and-downstream-action rule to non-imaging pathways, where specialist evaluation queues delay identification.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Banerjee and Nayak, 2026","url":"https://doi.org/10.3389/fpubh.2026.1886106","type":"paper"}]},{"category":"evidence","title":"Silva et al. comparison of CatBoost and GRU dengue forecasts across Brazil","change":"Epidemic Forecasting with AI adds a 2026 International Journal of Biometeorology walk-forward comparison of CatBoost and GRU dengue-forecasting models across all 27 Brazilian state capitals, extending the chapter's point that multimodel forecasting patterns from COVID-19 also apply to dengue.","why":"Gives the LMIC forecasting discussion a Brazil-specific model comparison alongside its existing dengue-challenge citations.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Silva et al., 2026","url":"https://doi.org/10.1007/s00484-026-03300-7","type":"paper"}]},{"category":"evidence","title":"Huang et al. evaluation of free-tier chatbots on HIV/AIDS FAQs","change":"LLMs in Public Health adds a peer-reviewed evaluation in which three infectious-disease specialists scored free-tier ChatGPT-3.5 and DeepSeek-R1 (search off, default settings) on 26 common HIV/AIDS questions.","why":"Fills a gap in the chapter's own evidence hierarchy for free-tier chatbot accuracy on population health information seeking.","chapters":[["Large Language Models in Public Health: Theory and Practice","/future/llm-theory-practice.html"]],"sources":[{"label":"Huang et al., 2026","url":"https://doi.org/10.3389/fpubh.2026.1886493","type":"paper"}]},{"category":"evidence","title":"Abdul Rahman and Tun climate-scenario model of global disease burden","change":"Epidemic Forecasting with AI adds a PLOS One climate-health scenario example: a GBD, CMIP6, and ERA5 simulation projecting global disease burden at 1.5, 2.0, and 2.5 degrees Celsius of warming, with a Random Forest R² of 0.89.","why":"The paper's own 'AI-forecast' label describes a scenario simulation, illustrating the chapter's distinction between forecasts and projections in a climate-health example.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Abdul Rahman & Tun, 2026","url":"https://doi.org/10.1371/journal.pone.0354159","type":"paper"}]},{"category":"policy","title":"Mello and Nagireddy on phased AI for Medicaid work-requirement eligibility","change":"AI Policy and Governance adds a subsection on AI for Medicaid eligibility redetermination under the 2025 budget reconciliation law's new community-engagement (work) rules, citing a JAMA Health Forum argument for introducing AI in phases rather than all at once.","why":"Addresses a live implementation question created by HR 1's new Medicaid work-requirement rules for state eligibility redeterminations.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Mello & Nagireddy, 2026","url":"https://doi.org/10.1001/jamahealthforum.2026.4102","type":"paper"}]},{"category":"revision","title":"Google AI Overview compared with specialized clinical evidence tools","change":"Diagnostic and Clinical Decision Support notes that public search answers such as Google AI Overview can match specialized clinical evidence tools on some clinician-query benchmarks, while not substituting for clinical care or institutional decision support.","why":"Extends the chapter's patient-facing safeguards to general-purpose search answers, not only dedicated chatbots or clinical decision-support tools.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[]},{"category":"evidence","title":"Dominikowski et al. and Janvier et al. on gray-market performance-enhancing peptides","change":"AI, Misinformation, and the Infodemic adds a parallel visual-misinformation failure mode: gray-market injectable peptides marketed as 'research compounds' and dosed from online protocols, citing a 2026 Frontiers in Endocrinology review that maps GH-IGF-1-axis performance-enhancing peptides.","why":"Extends the misinformation chapter's visual and product-image failure modes to gray-market peptides sold and dosed outside clinical oversight.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[{"label":"Dominikowski et al., 2026","url":"https://doi.org/10.3389/fendo.2026.1822475","type":"paper"},{"label":"Janvier et al., 2018","url":"https://doi.org/10.1016/j.talanta.2018.06.023","type":"paper"}]},{"category":"evidence","title":"Praschk et al. audit of LLM vaccine-stance classification","change":"AI, Misinformation, and the Infodemic adds a zero-shot audit of two reasoning LLMs classifying vaccine stance on 3,060 rehydrated 2020-21 X posts: o4-mini reached 0.819 accuracy and 0.820 macro-F1, and Gemini 2.5 Flash reached 0.799 accuracy and 0.800 macro-F1.","why":"Gives the misinformation chapter a measured accuracy audit rather than a deployed vaccine-stance classification system.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[{"label":"Praschk et al., 2026","url":"https://doi.org/10.1007/s10916-026-02449-3","type":"paper"}]},{"category":"evidence","title":"Qing et al. on VAERS/FAERS signals as a vaccine-safety evaluation control","change":"Evaluating AI Systems for Healthcare adds an evaluation control: do not treat VAERS/FAERS disproportionality analysis combined with in-sample machine learning as incidence, causality, or a vaccine-safety endpoint. The cited study analyzed 9,446 AVA reports and found 180 depression-related cases.","why":"Guards against reading VAERS/FAERS signals as causal vaccine-safety findings.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Qing et al., 2026","url":"https://doi.org/10.1186/s40249-026-01483-0","type":"paper"}]},{"category":"evidence","title":"Saleem Al-Ansary et al. XGBoost classifier for foodborne outbreak detection","change":"Disease Surveillance and Outbreak Detection adds a Frontiers in Public Health case study that trained an XGBoost outbreak classifier on 61 cleaned Eastern Province foodborne events, illustrating the base-rate problem with real records rather than hypothetical weekly figures; the classifier is described as a proof of concept, not an outbreak early-warning system.","why":"Gives the base-rate-problem discussion a real-world foodborne surveillance example instead of a hypothetical weekly calculation.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Saleem Al-Ansary et al., 2026","url":"https://doi.org/10.3389/fpubh.2026.1883871","type":"paper"}]},{"category":"content","title":"Mahmood and Topol on tokenising population health journeys","change":"Large Language Model Foundations for Public Health adds a new section on tokenising population health journeys, describing how Mahmood and Topol translate recorded health data (codes, measurements, text, images, and time intervals) into sequenced tokens that support population-scale prediction.","why":"Introduces the concept of translating longitudinal health records into token sequences, extending the chapter's tokenization teaching from single words to population health journeys.","chapters":[["Large Language Model Foundations for Public Health","/future/llm-foundations.html"]],"sources":[{"label":"Mahmood & Topol, 2026","url":"https://doi.org/10.1016/S0140-6736(26)01773-3","type":"paper"}]},{"category":"evidence","title":"Jiang et al. ensemble model for school-based vision screening","change":"Diagnostic and Clinical Decision Support adds a nested cross-validated ensemble model trained on 8,035 Guangdong children aged 5 to 10 in a school vision screening program, framed as a different kind of 'early warning' claim than outbreak detection.","why":"Distinguishes school-based screening performance from outbreak early-warning system performance, a distinction the chapter draws repeatedly.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Jiang et al., 2026","url":"https://doi.org/10.3389/fpubh.2026.1910531","type":"paper"}]},{"category":"evidence","title":"Fuad, Milki, and Aziz on dengue forecasting complexity in Bangladesh","change":"Epidemic Forecasting with AI adds a PLOS One panel study across 11 Bangladesh districts illustrating that greater model complexity does not equal greater forecasting skill, alongside a note that a review of 67 AI-based early-warning studies found no pooled sensitivity benchmark across systems.","why":"Reinforces the point that forecasting-model complexity does not guarantee skill, adding a Bangladesh dengue example alongside the chapter's broader early-warning evidence gap.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"sources":[{"label":"Fuad, Milki & Aziz, 2026","url":"https://doi.org/10.1371/journal.pone.0353069","type":"paper"}]},{"category":"evidence","title":"Xiang et al. on symptom-based pathogen prediction for home testing","change":"Disease Surveillance and Outbreak Detection adds a PLOS Digital Health home-testing study asking whether 13 self-reported symptoms can identify the likely pathogen well enough to prioritize multiplex PCR testing, distinguishing syndrome-level surveillance such as ILINet from pathogen-specific detection.","why":"Clarifies that ILINet-style syndromic surveillance identifies a syndrome, not a specific pathogen, a distinction relevant to home-testing triage design.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Xiang et al., 2026","url":"https://doi.org/10.1371/journal.pdig.0001656","type":"paper"}]},{"category":"evidence","title":"Haq et al. analysis of E. coli contamination in Bangladesh drinking water","change":"Disease Surveillance and Outbreak Detection adds a PLOS One analysis of 6,069 Bangladesh MICS 2019 household drinking-water samples, finding E. coli contamination in 39.2% of samples (95% CI 37.4 to 41.2) with geographic clustering by ward.","why":"Gives the surveillance chapter a moderate-prevalence example of a data-verification rule, using household water-quality rather than disease case data.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"Haq, Ador & Nobrega, 2026","url":"https://doi.org/10.1371/journal.pone.0343606","type":"paper"}]},{"category":"evidence","title":"Kim et al. on survey weighting bias in NHANES-based AI models","change":"The Data Problem in Public Health AI adds a Frontiers in Digital Health NHANES methods paper analyzing 6,024 training and 1,926 test adults (2007 to 2014 cycles), giving the chapter's population-inference rule an empirical example based on unweighted versus weight-aware model performance.","why":"Gives the chapter's population-inference rule an empirical NHANES example, following the same rule's earlier theoretical form.","chapters":[["The Data Problem in Public Health AI","/foundations/data.html"]],"sources":[{"label":"Kim, Rogers & Wang, 2026","url":"https://doi.org/10.3389/fdgth.2026.1834715","type":"paper"}]},{"category":"evidence","title":"Stürenburg et al. classification of global health aid by disease burden","change":"AI in Global Health and Equity adds a study using Llama-3.1-70B to classify 3.7 million OECD Creditor Reporting System project descriptions (about USD 332 billion, 157 countries, 2000 to 2022) into 17 Global Burden of Disease categories.","why":"Extends the chapter's aid-versus-burden mismatch discussion beyond workforce capacity to how development aid funding is distributed across disease categories.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Stürenburg et al., 2026","url":"https://doi.org/10.1038/s41467-026-76542-z","type":"paper"}]},{"category":"evidence","title":"Leibovitch et al. stage 1 DECIDE-AI evaluation of the SHAKED CDSS","change":"Evaluating AI Systems for Healthcare adds a Nature Medicine DECIDE-AI stage 1 evaluation of SHAKED, a multi-LLM clinical decision-support system tested in a tertiary emergency department over 4 weeks with 1,138 patients, one wing using the system and one following routine rotations.","why":"Gives the evaluation chapter a DECIDE-AI-framework example of early-stage clinical evaluation for a multi-LLM decision-support system.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Leibovitch et al., 2026","url":"https://doi.org/10.1038/s41591-026-04601-5","type":"paper"}]}]},
{"date":"2026-09-09","headline":"van Genderen and van Sprang review of AI across the travel-health continuum; Ye et al. five-phase lifecycle staging framework for medical AI evaluation","entries":[{"category":"evidence","title":"van Genderen and van Sprang review of AI across the travel-health continuum","change":"Disease Surveillance and Outbreak Detection adds a Journal of Travel Medicine review mapping AI across pre-, peri-, and post-travel care, citing AI-assisted border screening that flagged SARS-CoV-2 outbreaks up to 9 days earlier than conventional surveillance and malaria-risk prediction at AUC 0.98 (sensitivity 100%, specificity 72%).","why":"Adds a narrative review's synthesis of mostly retrospective evidence on travel-related AI screening, not a prospective multicenter implementation trial.","chapters":[["AI in Disease Surveillance and Outbreak Detection","/applications/surveillance.html"]],"sources":[{"label":"van Genderen and van Sprang, 2026","url":"https://doi.org/10.1093/jtm/taag082","type":"paper"}]},{"category":"evidence","title":"Ye et al. five-phase lifecycle staging framework for medical AI evaluation","change":"Evaluating AI Systems for Healthcare adds a pointer to an npj Digital Medicine framework proposing five gated lifecycle phases with fall-back triggers for staging medical AI evaluation, noting that a companion handbook carries the fuller teaching.","why":"Gated lifecycle phases with fall-back triggers give evaluators a proposed, not yet validated, structure for deciding when a medical AI system is ready for its next stage of evaluation.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Ye et al., 2026","url":"https://doi.org/10.1038/s41746-026-03155-7","type":"paper"}]}]},
{"date":"2026-09-11","headline":"GPT-4o narrows a cross-country clinical-vignette performance gap in a new RCT","entries":[{"category":"evidence","title":"Rounding et al. multi-country RCT of GPT-4o on clinical vignettes","change":"AI in Global Health and Equity adds a multi-country randomized trial of GPT-4o assistance on standardized clinical vignettes (249 physicians in Indonesia, Kenya, and the Netherlands): the largest absolute gain was in Kenya (+18 percentage points), versus +10.7 in Indonesia and +7.2 in the Netherlands, narrowing the Kenya-Netherlands performance gap.","why":"Gives the equity discussion a concrete case where AI assistance narrowed rather than widened a cross-country performance gap, evidence from controlled vignettes rather than patient outcomes.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Rounding et al., 2026","url":"https://doi.org/10.1038/s41746-026-03111-5","type":"paper"}]}]},
{"date":"2026-09-15","headline":"Setting-dependent gains in the Rounding et al. vignette trial; Parikh et al. on epidemiology of population AI exposure","entries":[{"category":"revision","title":"Setting-dependent gains in the Rounding et al. vignette trial","change":"AI in Global Health and Equity notes that the Rounding et al. trial's largest vignette gains occurred in Kenya, illustrating that LLM assistance effects may be setting-dependent, while stressing this is still not proof of reduced mortality or safer care.","why":"Draws out an implication of the RCT already cited in the chapter: assistance benefits may vary by setting rather than apply uniformly.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"sources":[]},{"category":"evidence","title":"Parikh et al. on epidemiology of population AI exposure","change":"Evaluating AI Systems for Healthcare adds a Nature Health perspective distinguishing ambient AI exposure (algorithmic curation and institutional AI decisions reaching people without opt-in) from personal AI exposure (direct tool use), arguing epidemiology needs frameworks to study population-level AI health effects. It is a framework perspective with survey illustration, not a causal estimate.","why":"Frames AI as a population-level determinant of health, not only a clinical device, extending the chapter's evaluation scope beyond individual tool use.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Parikh et al., 2026","url":"https://doi.org/10.1038/s44360-026-00188-w","type":"paper"}]}]},
{"date":"2026-09-20","headline":"New evidence-gap review and interpretive lens for patient-facing AI in primary care","entries":[{"category":"evidence","title":"Habibi et al. scoping review of patient-facing AI in primary care","change":"Diagnostic and Clinical Decision Support adds a PRISMA-ScR scoping review of 65 studies (from 3,469 records, search through March 2025) on patient-facing AI in primary health care: 27 tools (42%) were already publicly accessible, but patient safety was evaluated in only 8% of studies and access in 6%.","why":"Gives readers an evidence-gap warning for patient-facing chatbots already publicly available, not proof that any named tool improves outcomes.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Habibi et al., 2026","url":"https://doi.org/10.1186/s12916-026-05223-x","type":"paper"}]},{"category":"evidence","title":"Parikh and Kohli on an interpretive lens for patient-facing AI evaluation","change":"Diagnostic and Clinical Decision Support adds a Lancet Digital Health commentary arguing that patient-facing generative AI can shape expectations and perceived legitimacy before clinical contact, so evaluation should track how answers are framed, how uncertainty is shown, and whether language mix and validation context are disclosed alongside escalation and equity metrics.","why":"Adds a complementary interpretive framework alongside the Habibi scoping review, for judging how patient-facing AI tools are perceived, not only what they measure.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Parikh & Kohli, 2026","url":"https://doi.org/10.1016/j.landig.2026.101072","type":"paper"}]}]},
{"date":"2026-09-22","headline":"Zhou et al. EAGLE model for opportunistic esophageal cancer detection","entries":[{"category":"evidence","title":"Zhou et al. EAGLE model for opportunistic esophageal cancer detection","change":"Diagnostic and Clinical Decision Support adds EAGLE, a two-stage model for esophageal malignancy detection on noncontrast chest CT, externally tested across eight centers in three countries (n = 11,466) at about 98.5% specificity and 90% sensitivity for cancer, with a prospective hospital run reaching about 42% positive predictive value.","why":"The studies show multicenter diagnostic performance and implementation results, not a mortality-outcome trial, and several authors are industry-affiliated with closed code and data.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Zhou et al., 2026","url":"https://doi.org/10.1038/s41591-026-04656-4","type":"paper"}]},{"category":"policy","title":"UK National Commission recommendations for regulating AI in healthcare","change":"AI Policy and Governance adds a new subsection on the UK National Commission's September 2026 recommendations for a future AI-in-healthcare regulatory framework: staged authorisation for new AI models, continuous real-world monitoring, stronger traceability, and proportionate transparency when AI is used in care.","why":"These are independent advisory recommendations, not enacted law, and the chapter keeps its existing MHRA change-programme material alongside this new layer.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"National Commission into the Regulation of AI in Healthcare, September 2026","url":"https://www.gov.uk/government/publications/national-commission-into-the-regulation-of-ai-in-healthcare-recommendations-for-a-future-regulatory-framework/national-commission-into-the-regulation-of-ai-in-healthcare-recommendations-for-a-future-regulatory-framework","type":"policy"}]},{"category":"policy","title":"CDC guidance on agentic 'deep research' tools for public health research","change":"AI Policy and Governance adds CDC's Considerations for Agentic Research in Public Health, which favors well-scoped, information-rich tasks such as policy scans and early synthesis with qualified human verification, while advising against restricted or sensitive data, statistically rigorous analytics, and legal or clinical judgment calls.","why":"Gives readers CDC's own effective-use principles for agentic research tools: informational exploratory guidance, not a vendor accuracy validation or permission to place PHI in external agents.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"CDC, Considerations for Agentic Research in Public Health, 2026","url":"https://www.cdc.gov/ai/resources/considerations-for-agentic-research-in-public-health.html","type":"policy"}]},{"category":"developments","title":"ASPPH AI framework for academic public health workforce and education","change":"AI Policy and Governance and Your AI Toolkit for Public Health add ASPPH's June 2026 Artificial Intelligence Framework, a task-force roadmap across four pillars: Education; Teaching and Learning; Practice and Research; and Policy, Regulatory, and Architectural Frameworks, urging AI literacy as core public health professionalism.","why":"Gives schools and programs of public health a workforce and curriculum roadmap.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"],["Your AI Toolkit for Public Health","/practical/toolkit.html"]],"sources":[{"label":"ASPPH, Artificial Intelligence Framework, June 2026","url":"https://aspph.org/initiatives/ai-for-public-health/ai-framework-report/","type":"report"},{"label":"ASPPH, press release, June 2026","url":"https://aspph.org/aspph-releases-ai-framework-to-guide-responsible-innovation-in-academic-public-health/","type":"other"},{"label":"ASPPH, AI Framework recommendations (PDF), 2026","url":"https://aspph-prod-web-assets.s3.us-east-1.amazonaws.com/AI-Framework-Recommendations.pdf","type":"other"}]},{"category":"evidence","title":"Singletary and Baker on phased AI adoption in public health agencies","change":"AI Deployment: Why Most Prototypes Fail adds a Journal of Public Health Management and Practice article arguing that health departments should adopt AI through a phased approach rather than a single enterprise-wide rollout, paired with the chapter's existing pilot, canary, and stop-criteria teaching.","why":"Adds a short peer-reviewed practice article to a chapter that already teaches phased deployment discipline, without treating it as outcome evidence for any specific vendor.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[{"label":"Singletary & Baker, 2026","url":"https://doi.org/10.1097/PHH.0000000000002486","type":"paper"}]},{"category":"policy","title":"WHO report on ethics review and oversight for AI-related health research","change":"Ethics, Bias, and Equity adds a WHO July 2026 report on ethics review and oversight for AI-related health research, organizing it into three categories (AI used for data science, research conducted with AI tools, and research on AI tools) and noting particular risks for low- and middle-income countries, including data colonialism and ethics dumping.","why":"Addresses ethics gaps when AI enters the research pipeline itself, not only clinical deployment, complementing the WHO AI-for-health ethics principles already taught in the chapter.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"WHO, Artificial intelligence-related health research: ethics review and oversight, 2026","url":"https://www.who.int/publications/i/item/9789240124073","type":"policy"}]},{"category":"evidence","title":"Al-Ajlouni and Njei proposed AI Productivity Index for health systems","change":"Evaluating AI Systems for Healthcare adds a Clinical Medicine opinion proposing an AI Productivity Index, a preliminary composite lens covering operational impact, implementation burden, opportunity costs, downstream utilisation, and equity of gains or burdens, to sit beside safety, efficacy, and formal economic evaluation.","why":"Argues that clinical accuracy alone is a weak proxy for system value, though this is an opinion framework with illustrative domains, not a validated scored instrument.","chapters":[["Evaluating AI Systems for Healthcare","/implementation/evaluation.html"]],"sources":[{"label":"Al-Ajlouni & Njei, Clinical Medicine, 2026","url":"https://doi.org/10.1016/j.clinme.2026.100638","type":"paper"}]}]},
{"date":"2026-09-23","headline":"Tiwari et al. on scaling diabetic retinopathy screening across three countries","entries":[{"category":"evidence","title":"Tiwari et al. Nature Medicine comment on scaling diabetic retinopathy screening","change":"Context-Appropriate Global Health AI adds a pointer to a 2026 Nature Medicine Comment on the Aravind program family's diabetic retinopathy screening expansion across India, Thailand, and Australia to more than a million screenings, framed as implementation lessons rather than a mortality RCT.","why":"Extends the chapter's Aravind case study with a broader multi-country update, while keeping the boundary that this is implementation experience, not outcome-trial evidence.","chapters":[["Context-Appropriate Global Health AI","/future/global-ai-implementation.html"]],"sources":[{"label":"Tiwari et al., 2026","url":"https://doi.org/10.1038/s41591-026-04643-9","type":"paper"}]}]},
{"date":"2026-09-25","headline":"Anthropic reports Claude drafting situation reports in a 2026 Ebola response in DRC","entries":[{"category":"developments","title":"Anthropic reports Claude use in a CEPI-convened Ebola response in eastern DRC","change":"AI in Public Health Emergency Operations adds a vendor-reported case: an Anthropic feature describes a CEPI-convened partnership with WHO AFRO and INRB that used Claude to draft situation reports during the 2026 Bundibugyo ebolavirus response, with named staff verifying facts. The chapter notes it is a company feature story with no peer-reviewed outcome data.","why":"Illustrates the chapter's documentation guidance: AI may draft situation reports while humans verify, and AI does not set objectives or allocate life-safety resources.","chapters":[["AI in Public Health Emergency Operations","/implementation/emergency-operations.html"]],"sources":[{"label":"Anthropic, Ebola response feature, 2026","url":"https://www.anthropic.com/features/ebola-response","type":"announcement"}]}]},
{"date":"2026-09-29","headline":"An HPV vaccine chatbot raised literacy without changing 2-week vaccination decisions","entries":[{"category":"evidence","title":"Kobayashi et al. trial of an HPV vaccine chatbot versus a government leaflet","change":"AI-Driven Behavioral Interventions reports a JAMA Network Open randomized trial of an HPV vaccine chatbot versus a government leaflet among female caregivers in Japan with unvaccinated daughters aged 12-18 (Kobayashi et al., 21 September 2026; immediate mITT n=704; 2-week n=477). Adjusted literacy was 0.30 points higher on a 0-7 scale with the chatbot (95% CI 0.18-0.43 immediately; 0.13-0.47 at 2 weeks). Vaccination decisions at 2 weeks were 40.3% and 39.6% (adjusted OR 0.74, 95% CI 0.41-1.33).","why":"In this trial, conversational AI raised vaccine literacy and left short-term vaccination decisions unchanged.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"],["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"sources":[{"label":"Kobayashi et al., 2026","url":"https://doi.org/10.1001/jamanetworkopen.2026.34696","type":"paper"}]}]},
{"date":"2026-09-30","headline":"Nature editorial and FDA discussion paper on AI devices (comments due 19 Oct 2026; not guidance)","entries":[{"category":"policy","title":"Nature editorial and FDA discussion paper on generative-AI medical devices","change":"AI Policy and Governance cites a September 2026 Nature editorial urging risk-proportionate real-world testing for generative-AI medical products that inform clinical decisions, and FDA's 18 August 2026 discussion paper on generative-AI-enabled devices. Comments under docket FDA-2026-N-7874 are due 19 October 2026. The document seeks input and is not draft or final guidance.","why":"The comment window is open, and the paper is not a clearance standard an agency can apply as guidance.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Nature, 2026","url":"https://doi.org/10.1038/d41586-026-03046-7","type":"paper"},{"label":"FDA, Considerations for the Regulation of Generative AI-Enabled Medical Devices, 18 August 2026","url":"https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request","type":"policy"},{"label":"FDA docket FDA-2026-N-7874","url":"https://www.regulations.gov/docket/FDA-2026-N-7874","type":"policy"}]},{"category":"developments","title":"Google announces Gemini 4 Argon for Fairwind cyber defense","change":"Selecting and Using Large Language Models carries Google's September 30, 2026 announcement of Gemini 4 Argon for phased Fairwind cyber-defense access, with a paid API and AI Ultra rollout still forthcoming. Wiz Scan for Good reports that Argon found a healthcare-software vulnerability exposing hospital PII; the chapter treats that report as a cyber-defense vignette, not clinical accuracy evidence.","why":"Readers choosing a model need the current name, and a vendor cyber report is not clinical performance evidence.","chapters":[["Selecting and Using Large Language Models","/future/llm-selection-use.html"]],"sources":[{"label":"Google, Gemini 4 Argon announcement, September 30, 2026","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","type":"announcement"},{"label":"Google, Fairwind Program","url":"https://blog.google/innovation-and-ai/technology/safety-security/fairwind-program/","type":"announcement"},{"label":"Google DeepMind, Frontier Safety Framework","url":"https://deepmind.google/blog/strengthening-our-frontier-safety-framework/","type":"announcement"}]},{"category":"policy","title":"ARPA-H ADVOCATE awards for agentic cardiovascular clinical AI","change":"AI Policy and Governance notes ARPA-H's September 2026 ADVOCATE awards for agentic clinical AI in heart failure, a supervisory safety agent, and multi-system deployment plans aimed at FDA authorization pathways, including sites that serve counties without a cardiologist. A companion handbook carries the performer teams and the evaluation fence.","why":"A federal funding program for cardiovascular clinical AI is a deployment plan, separate from device clearance.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"ARPA-H, 2026","url":"https://arpa-h.gov/news-and-events/arpa-h-launches-worlds-first-bid-build-fda-authorized-clinical-ai-cardiovascular","type":"policy"},{"label":"ARPA-H, ADVOCATE program","url":"https://arpa-h.gov/explore-funding/programs/advocate","type":"policy"}]}]},
{"date":"2026-10-01","headline":"Bessemer and Bain report healthcare AI returns by use case","entries":[{"category":"developments","title":"Bessemer and Bain report healthcare AI returns by use case","change":"AI Deployment: Why Most Prototypes Fail describes a Bessemer Venture Partners and Bain survey of 226 executives reporting average returns of 3.5 times cost in about 12 months. These are self-reported enterprise returns, with higher multiples in administrative workflows, and do not replace local total-cost-of-ownership analysis.","why":"Procurement decisions need returns measured for the intended workflow and local costs.","chapters":[["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"]],"sources":[{"label":"Guerra and Kraus, 2026","url":"https://www.bvp.com/atlas/the-2026-healthcare-ai-roi-scorecard","type":"report"}]}]},
{"date":"2026-10-05","headline":"Rosenbaum: algorithm disclosure did not change flu-vaccine uptake","entries":[{"category":"evidence","title":"Rosenbaum: algorithm disclosure did not change flu-vaccine uptake","change":"AI-Driven Behavioral Interventions describes three randomized trials with more than 90,000 patients: high-risk messages raised flu-vaccine uptake by 1.1 to 1.4 percentage points over plain reminders, while naming the algorithm did not change uptake. The trials measured vaccination at one US health system, not influenza illness.","why":"The trials inform how health systems explain algorithm-targeted vaccine messages to patients.","chapters":[["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"]],"sources":[{"label":"Rosenbaum et al., 2026","url":"https://doi.org/10.1038/s41562-026-02603-4","type":"paper"}]},{"category":"evidence","title":"SMART-AI screening for substance misuse from inpatient notes","change":"Substance Use and Overdose Prevention describes retrospective SMART-AI screening from inpatient notes, with external validation AUROCs of 0.88 for alcohol misuse and 0.94 for opioid misuse. An external AUROC for non-opioid drug misuse was not reported.","why":"Screening performance depends on the substance endpoint and the population used for validation.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"sources":[{"label":"Afshar et al., 2022","url":"https://doi.org/10.1016/S2589-7500(22)00041-3","type":"paper"}]},{"category":"evidence","title":"External validation of an alcohol-use classifier","change":"Substance Use and Overdose Prevention describes Lin and colleagues' external validation of an alcohol NLP classifier at an independent Midwestern center, with AUROC 0.91 and specificity 0.98 for unhealthy alcohol use in the first 24 hours of notes.","why":"Independent validation helps assess whether a clinical-note classifier transports to another center.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"sources":[{"label":"Lin et al., 2022","url":"https://doi.org/10.1111/add.15730","type":"paper"}]},{"category":"evidence","title":"Pediatric substance-use detection from structured and narrative records","change":"Substance Use and Overdose Prevention describes Ni and colleagues' retrospective pediatric EHR system combining rules, NLP, and deep learning for patients aged 10 to 20. AUROC was at least 0.957 across alcohol, tobacco, marijuana, and opiate labels for lifetime, current, and family use.","why":"Narrative records contain substance-use information that structured fields may miss.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"sources":[{"label":"Ni et al., 2021","url":"https://doi.org/10.1093/jamia/ocab116","type":"paper"}]},{"category":"evidence","title":"Cigarette-cessation prediction in the PATH cohort","change":"Substance Use and Overdose Prevention describes a random-forest model using PATH survey data to predict next-wave cigarette cessation among adult smokers, with balanced accuracy of 72% on held-out data. Prediction discrimination does not establish a treatment effect.","why":"Population survey models address cessation prediction beyond clinical-note screening.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"sources":[{"label":"Issabakhsh et al., 2023","url":"https://doi.org/10.1371/journal.pone.0286883","type":"paper"}]},{"category":"evidence","title":"Wu and colleagues on consumer AI care pathways","change":"Diagnostic and Clinical Decision Support cites a Nature Health perspective arguing that consumer health AI matters through platform integration: who controls symptom interpretation, care routing, payment, and treatment completion.","why":"Platform control shapes care pathways as well as the model performance a consumer encounters.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Wu et al., 2026","url":"https://doi.org/10.1038/s44360-026-00195-x","type":"paper"}]},{"category":"developments","title":"European academies recommend evaluation of AI net benefit","change":"AI Policy and Governance records the EASAC and FEAM recommendation that EU policymakers mandate transparent evaluation of AI net benefit before and after deployment, covering clinical outcomes, patient safety, security, equity, quality of life, and cost-effectiveness.","why":"Net-benefit evaluation connects technical performance to the consequences of deployment.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"Cole-Hunter et al., EASAC and FEAM, 2026","url":"https://doi.org/10.1553/EASAC-FEAM_Report_AI-in-Healthcare_2026","type":"report"}]},{"category":"evidence","title":"LungDiag accuracy on simulated respiratory cases","change":"Diagnostic and Clinical Decision Support describes a randomized comparison among 2,400 adults: LungDiag had model-based accuracy of 70.0% versus 55.4% for mobile web search in the primary compliant-case analysis. Simulated respiratory cases do not establish effects on real-world care-seeking, diagnostic delay, or clinical outcomes.","why":"The comparison informs interpretation of consumer symptom-assessment performance.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Liang et al., 2026","url":"https://doi.org/10.1038/s44360-026-00189-9","type":"paper"}]},{"category":"evidence","title":"Luo and colleagues on ophthalmology pre-consultation histories","change":"Diagnostic and Clinical Decision Support describes a single-center randomized trial in 172 patients: an LLM pre-consultation system exceeded residents' history-completeness scores by an adjusted 17.7 points. The exploratory history-alone diagnostic advantage narrowed with ocular signs; the trial measured history quality, not patient outcomes.","why":"History-taking evaluations require attention to examination findings and the outcomes actually measured.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"sources":[{"label":"Luo et al., 2026","url":"https://doi.org/10.1038/s41746-026-03232-x","type":"paper"}]},{"category":"evidence","title":"Celi and colleagues on reliability, equity, and community governance","change":"Ethics, Bias, and Equity cites Celi and colleagues' framing essay: reliability without equity and community governance is incomplete, and transferring health data and models from the Global South to benefit commercial actors elsewhere raises digital extractivism concerns. This is a framing essay, not new empirical equity evidence.","why":"Community governance and the distribution of benefits belong in assessments of health AI reliability.","chapters":[["Ethics, Bias, and Equity in Healthcare AI","/implementation/ethics.html"]],"sources":[{"label":"Celi et al., 2026","url":"https://doi.org/10.1145/3833880","type":"paper"}]}]},
{"date":"2026-10-06","headline":"UK government accepts all 44 National Commission recommendations on healthcare AI","entries":[{"category":"policy","title":"UK government accepts all 44 National Commission recommendations on healthcare AI","change":"AI Policy and Governance records UK acceptance of all 44 recommendations on 6 October 2026, including work toward a public medical-device adverse-incident database and AI Airlock Phase 3 on post-market surveillance. Acceptance does not itself change UK law; an implementation roadmap is due in Spring 2027.","why":"The response sets priorities for population-level safety monitoring of AI-enabled medical devices.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"sources":[{"label":"UK government response to the National Commission, October 2026","url":"https://www.gov.uk/government/publications/government-response-to-the-national-commissions-recommendations-on-the-regulation-of-ai-in-healthcare/government-response-to-the-national-commissions-recommendations-on-the-regulation-of-ai-in-healthcare","type":"policy"},{"label":"UK government announcement, October 2026","url":"https://www.gov.uk/government/news/government-backs-recommendations-of-nhs-doctors-led-ai-commission","type":"policy"}]},{"category":"evidence","title":"Google researchers report mixed PDFM gains for cholera and dengue forecasts","change":"Epidemic Forecasting with AI describes a Google-authored preprint reporting that PDFM place embeddings improved eight-week cholera emergence ranking in the DRC but did not help at one to two weeks. One-month dengue forecasts improved overall in Mexico, while the median municipality did slightly worse.","why":"Forecast evaluation must examine lead time and local performance before population-wide use.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"],["AI in Global Health and Equity","/future/global-health.html"]],"sources":[{"label":"Muslim et al., 2026, preprint","url":"https://arxiv.org/abs/2610.05699","type":"preprint"},{"label":"Google Research, 2026","url":"https://research.google/blog/earth-ais-planetary-geospatial-foundation-models-for-global-public-health/","type":"announcement"}]}]},
{"date":"2026-10-07","headline":"OECD Expert-Led Madrid Action Plan for scaling AI in health","entries":[{"category":"policy","title":"OECD Expert-Led Madrid Action Plan for scaling AI in health","change":"AI Policy and Governance describes OECD and Spanish Ministry of Health conference proceedings proposing ten priority actions for scaling AI in health, plus measurement of readiness, use, and impact. EL-MAP is an action-plan proposal, not an adopted guideline or binding standard.","why":"The proposal connects implementation capacity, oversight, infrastructure, and financing to responsible scale.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"summary":"The OECD Expert-Led Madrid Action Plan proposes priorities for scaling AI in health, including measurement of readiness, use, and impact.","section":"/future/policy.html#sec-future-policy-intl-governance","sources":[{"label":"OECD, 2026","url":"https://www.oecd.org/en/publications/proceeding-report-from-madrid-conference-on-the-responsible-scale-of-ai-in-healthcare_cf9439c3-en.html","type":"policy"}]},{"category":"developments","title":"Google reports public SynthID Detector access for content provenance","change":"AI, Misinformation, and the Infodemic describes public image, video, and audio uploads to Google's SynthID Detector for watermarks from participating companies. A negative result cannot authenticate a file because content outside the program and deliberate watermark removal can leave no detectable mark.","why":"Health communicators need to distinguish evidence of content provenance from evidence that a message is authentic.","chapters":[["AI, Health Misinformation, and the Infodemic","/future/misinformation.html"]],"summary":"Google reports public SynthID Detector access for images, video, and audio; a negative watermark result cannot establish authenticity.","section":"/future/misinformation.html#sec-future-misinformation-detection-and-verification-technologies","sources":[{"label":"Google, October 2026","url":"https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/","type":"announcement"},{"label":"OpenAI, May 2026","url":"https://openai.com/index/advancing-content-provenance/","type":"announcement"}]}]},
{"date":"2026-10-08","headline":"Anthropic cyber-defense access for hospitals and public infrastructure","entries":[{"category":"developments","title":"Anthropic cyber-defense access for hospitals and public infrastructure","change":"Selecting and Using Large Language Models describes Defense Access eligibility for regional hospitals and municipal utilities; Anthropic reports offering models and technical support to more than half of U.S. states since June 2026. These are security-team access programs, not evidence of clinical or surveillance accuracy.","why":"Public health security teams can assess access programs for defending the systems their organizations operate.","chapters":[["Selecting and Using Large Language Models","/future/llm-selection-use.html"]],"summary":"Anthropic lists regional hospitals and municipal utilities as eligible for Defense Access and reports offering models and technical support to more than half of U.S. states.","section":"/future/llm-selection-use.html#sec-future-llm-selection-use-choosing-the-right-llm-a-decision-framework","sources":[{"label":"Anthropic, Cyber Verification Program, October 2026","url":"https://www.anthropic.com/news/cyber-verification-program","type":"announcement"},{"label":"Anthropic, Cyber Mission, October 2026","url":"https://www.anthropic.com/news/anthropic-cyber-mission","type":"announcement"}]}]},
{"date":"2026-10-09","headline":"HHS civil enforcement for 42 CFR Part 2 substance-use treatment records; PDMP-based overdose risk scores: limited validation evidence; COVID-19 Forecast Hub: forecast horizons and how error grew with them; Dermatology and chest radiograph models lose accuracy outside their original data; Gates Foundation plans at least US$1 billion for equitable AI over two years; The trivy-action tag compromise and pinning third-party actions to immutable hashes","entries":[{"category":"policy","title":"HHS civil enforcement for 42 CFR Part 2 substance-use treatment records","change":"AI for Substance Use and Overdose Prevention states that the 2024 Part 2 final rule has required compliance since February 16, 2026. From that date the HHS Office for Civil Rights accepts complaints and breach notifications for Part 2 records, with penalties aligned with HIPAA.","why":"Programs that hold substance-use treatment records now face civil enforcement for Part 2 violations, the first time it has been available.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"summary":"Since February 16, 2026, the HHS Office for Civil Rights accepts complaints and breach notifications for Part 2 substance-use treatment records.","section":"/applications/substance-use.html#sec-substance-use-implementation-challenges","sources":[{"label":"HHS Office for Civil Rights, February 2026","url":"https://www.hhs.gov/press-room/hhs-announce-civil-enforcement-program-sud-patient-records.html","type":"policy"}]},{"category":"evidence","title":"PDMP-based overdose risk scores: limited validation evidence","change":"AI for Substance Use and Overdose Prevention reports a cross-sectional study in 19 pharmacies that found fair concurrent validity of the PDMP Narcotic Score against the WHO ASSIST screening test, not against overdose, and a 2023 JAMIA perspective noting that the Overdose Risk Score has never been clinically evaluated.","why":"Agencies and clinicians using these scores have no clinical evaluation showing that they predict overdose.","chapters":[["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"]],"summary":"Published validation of PDMP Narcotic and Overdose Risk Scores is limited; neither has been shown to predict overdose.","section":"/applications/substance-use.html#sec-substance-use-cdc-caution-on-pdmp-risk-scores","sources":[{"label":"Cochran et al., 2021","url":"https://doi.org/10.1016/j.drugalcdep.2021.109067","type":"paper"},{"label":"McElfresh et al., 2023","url":"https://doi.org/10.1093/jamia/ocad110","type":"paper"}]},{"category":"revision","title":"COVID-19 Forecast Hub: forecast horizons and how error grew with them","change":"Epidemic Forecasting with AI describes the COVID-19 Forecast Hub, launched in April 2020: as of September 28, 2021, during Delta, it reported only one-week-ahead case forecasts because of persistent inaccuracies. For the two teams forecasting all 50 states, 20-week error was 2.9 to 4 times the 1-week score.","why":"How quickly forecast error grows tells agencies how far ahead an ensemble forecast can support a decision.","chapters":[["Epidemic Forecasting with AI","/applications/forecasting.html"]],"summary":"As of September 28, 2021, the COVID-19 Forecast Hub reported only one-week-ahead case forecasts; 20-week error reached 2.9 to 4 times the 1-week score.","section":"/applications/forecasting.html#sec-forecasting-case-study-covid-19-forecasting-hub","sources":[]},{"category":"revision","title":"Dermatology and chest radiograph models lose accuracy outside their original data","change":"Global Health AI Equity and Local Capacity reports three dermatology models reaching ROC-AUC 0.56 to 0.67 on a skin-tone-diverse image set against 0.88 to 0.94 on their original test data, worse on Fitzpatrick types V and VI, and chest radiograph models that performed worse externally in three of five comparisons.","why":"Models moved to sites with different populations or equipment need the same external validation before use.","chapters":[["Global Health AI Equity and Local Capacity","/future/global-equity-capacity.html"]],"summary":"Dermatology and chest radiograph models lost accuracy on new populations and sites, so models moved to new settings need external validation.","section":"/future/global-equity-capacity.html#sec-future-global-equity-capacity-examples-of-population-bias","sources":[]},{"category":"developments","title":"Gates Foundation plans at least US$1 billion for equitable AI over two years","change":"Global Health AI Data Governance describes the Gates Foundation's September 2026 plan: roughly 40% for health care, including decision support for frontline health workers and maternal and newborn care tools, and 10% for digital foundations such as datasets in underserved languages. It is a funding plan and reports no evaluation results.","why":"The announcement ties effective tools to data from the populations they serve, with countries and communities deciding how data is managed.","chapters":[["Global Health AI Data Governance","/future/global-data-governance.html"]],"summary":"The Gates Foundation plans at least US$1 billion over two years for equitable AI, about 40% for health care; no results are reported yet.","section":"/future/global-data-governance.html#sec-future-global-data-governance-llms-global-health","sources":[{"label":"Gates Foundation, September 2026","url":"https://www.gatesfoundation.org/ideas/media-center/press-releases/2026/09/goalkeepers-report-equitable-ai","type":"announcement"}]},{"category":"developments","title":"The trivy-action tag compromise and pinning third-party actions to immutable hashes","change":"MLOps for Public Health AI pins each third-party action in its example pipeline to a full SHA hash rather than a version tag. In March 2026 an attacker force-pushed 76 of 77 version tags of aquasecurity/trivy-action to credential-stealing malware for about 12 hours, and the maintainers' advisory directs users to pin actions to full SHA hashes.","why":"A version tag can be moved to different code; a full SHA hash cannot.","chapters":[["MLOps for Public Health AI","/implementation/mlops.html"]],"summary":"After the March 2026 trivy-action tag compromise, the MLOps example pins third-party actions to full SHA hashes, as the maintainers advise.","section":"/implementation/mlops.html#sec-implementation-mlops-mlops-fundamentals","sources":[{"label":"Aqua Security, trivy-action security advisory, 2026","url":"https://github.com/aquasecurity/trivy/security/advisories/GHSA-69fq-xp46-6x23","type":"announcement"}]}]},
{"date":"2026-10-10","headline":"Google's Med-PaLM 2: question-answering performance and clinical limits; EASAC and FEAM recommend clearer Annex III guidance for healthcare AI; Obermeyer et al.: health costs as a biased proxy for health need; Murphy et al.: modeled screening capacity with CAD4TB; Anthropic's AI interviews describe existing Claude.ai users; Sweeney's demographic-uniqueness estimate uses 1990 Census data","entries":[{"category":"revision","title":"Google's Med-PaLM 2: question-answering performance and clinical limits","change":"Google's Med-PaLM 2 study combines medical-question benchmarks, physician review of long-form answers, and adversarial testing. Performance Metrics for Public Health AI explains why these evaluations do not establish clinical competence.","why":"Public health teams need evaluation methods that match the decisions and workflows a model would support.","chapters":[["Performance Metrics for Public Health AI","/implementation/performance-metrics.html"]],"summary":"Google's Med-PaLM 2 study evaluates question answering through benchmarks and physician ratings; these results do not establish clinical competence.","section":"/implementation/performance-metrics.html#sec-implementation-performance-metrics-evaluating-foundation-models-and-large-language-models","sources":[{"label":"Singhal et al., 2025","url":"https://doi.org/10.1038/s41591-024-03423-7","type":"paper"}]},{"category":"revision","title":"EASAC and FEAM recommend clearer Annex III guidance for healthcare AI","change":"EASAC and FEAM recommend clearer guidance on which healthcare AI deployments fall under Annex III of the EU AI Act, to support consistent high-risk classification across Member States. AI Policy and Governance presents this as advice from the academies.","why":"Healthcare AI classification determines which regulatory obligations apply to a particular use.","chapters":[["AI Policy and Governance in Healthcare","/future/policy.html"]],"summary":"EASAC and FEAM recommend clearer Annex III guidance for consistent high-risk classification of healthcare AI across EU Member States.","section":"/future/policy.html#sec-future-policy-regulatory-landscape","sources":[{"label":"Cole-Hunter et al., EASAC and FEAM, 2026","url":"https://doi.org/10.1553/EASAC-FEAM_Report_AI-in-Healthcare_2026","type":"report"}]},{"category":"revision","title":"Obermeyer et al.: health costs as a biased proxy for health need","change":"A widely used algorithm predicted health costs as a proxy for health need and gave Black patients lower risk scores than equally sick White patients. Global Health AI Equity and Local Capacity describes the mechanism reported by Obermeyer and colleagues.","why":"The outcome chosen for prediction can reproduce unequal access to care in resource-allocation decisions.","chapters":[["Global Health AI Equity and Local Capacity","/future/global-equity-capacity.html"]],"summary":"Obermeyer and colleagues found that predicting health costs as a proxy for health need produced racial bias in a widely used algorithm.","section":"/future/global-equity-capacity.html#sec-future-global-equity-capacity-algorithmic-fairness-across-populations","sources":[{"label":"Obermeyer et al., 2019","url":"https://www.science.org/doi/10.1126/science.aax2342","type":"paper"}]},{"category":"revision","title":"Murphy et al.: modeled screening capacity with CAD4TB","change":"Murphy and colleagues evaluated CAD4TB on chest radiographs and modeled daily screening capacity and cost for different software versions. Context-Appropriate Global Health AI describes those estimates as results of one evaluation, not observed deployment counts.","why":"Planning a screening service requires distinguishing modeled capacity from measured performance in its intended workflow.","chapters":[["Context-Appropriate Global Health AI","/future/global-ai-implementation.html"]],"summary":"Murphy and colleagues' CAD4TB screening-capacity and cost estimates come from an efficiency analysis, not observed deployment counts.","section":"/future/global-ai-implementation.html#sec-future-global-ai-implementation-successful-global-health-ai-implementations","sources":[{"label":"Murphy et al., 2020","url":"https://pubmed.ncbi.nlm.nih.gov/32218458/","type":"paper"}]},{"category":"revision","title":"Anthropic's AI interviews describe existing Claude.ai users","change":"Anthropic's interviews describe the concerns and aspirations of existing Claude.ai users across countries. AI in Global Health and Equity explains that the interviewees do not represent whole populations and that the study did not test tools designed for local context.","why":"Public health teams need to know who supplied the evidence before applying findings about trust to other communities.","chapters":[["AI in Global Health and Equity","/future/global-health.html"]],"summary":"Anthropic's interviews describe existing Claude.ai users, not whole populations, and did not test tools designed for local context.","section":"/future/global-health.html#sec-future-global-health-the-paradox-of-need","sources":[{"label":"Anthropic, 2026","url":"https://www.anthropic.com/features/81k-interviews","type":"announcement"}]},{"category":"revision","title":"Sweeney's demographic-uniqueness estimate uses 1990 Census data","change":"Sweeney estimated demographic uniqueness using 1990 Census data and the combination of date of birth, gender, and five-digit ZIP code. Large Language Model Foundations for Public Health uses this historical analysis to explain why removing direct identifiers does not settle re-identification risk.","why":"Privacy decisions depend on which attributes remain available for linking records.","chapters":[["Large Language Model Foundations for Public Health","/future/llm-foundations.html"],["Large Language Model Exercises for Public Health","/future/llm-exercises.html"]],"summary":"Sweeney's demographic-uniqueness estimate uses 1990 Census data and combines date of birth, gender, and five-digit ZIP code.","section":"/future/llm-foundations.html#sec-future-llm-foundations-privacy-security","sources":[{"label":"Sweeney, 2000","url":"https://doi.org/10.1184/R1/6625769.v1","type":"report"}]}]},
{"date":"2026-10-16","headline":"Wastewater sequencing and early detection of SARS-CoV-2 variants of concern; Screening digital mammography accuracy in U.S. community practice; Undocumented SARS-CoV-2 infections in China before the January 2020 travel restrictions","entries":[{"category":"evidence","title":"Wastewater sequencing and early detection of SARS-CoV-2 variants of concern","change":"AI in Genomic Surveillance and Pathogen Analysis reports that in a 295-day San Diego study, wastewater sequencing detected emerging variants of concern up to 14 days earlier than clinical sequencing (Karthikeyan et al., 2022). The chapter places this beside CDC's National Wastewater Surveillance System, which integrated the Freyja variant-deconvolution tool to track SARS-CoV-2 variant composition nationally.","why":"Wastewater sequencing depends less on clinical testing, so it can give health departments an earlier signal of an emerging variant.","chapters":[["AI in Genomic Surveillance and Pathogen Analysis","/applications/genomics.html"]],"summary":"In a 295-day San Diego study, wastewater sequencing detected emerging SARS-CoV-2 variants of concern up to 14 days earlier than clinical sequencing.","section":"/applications/genomics.html#sec-genomics-genomic-data-collection-quality-and-biases","sources":[{"label":"Karthikeyan et al., 2022, Nature","url":"https://doi.org/10.1038/s41586-022-05049-6","type":"paper"}]},{"category":"evidence","title":"Screening digital mammography accuracy in U.S. community practice","change":"AI in Diagnostic and Clinical Decision Support reports that in U.S. community practice, screening digital mammography had 86.9% sensitivity and 88.9% specificity, missing about 13% of cancers present at screening (Lehman et al., 2017). The figures frame the breast cancer screening case, where double reading improves accuracy but doubles cost.","why":"Measured human reading performance in routine screening is the baseline against which AI support for mammography should be judged.","chapters":[["AI in Diagnostic and Clinical Decision Support","/applications/clinical.html"]],"summary":"In U.S. community practice, screening digital mammography had 86.9% sensitivity and 88.9% specificity, missing about 13% of cancers present at screening.","section":"/applications/clinical.html#sec-clinical-ai-in-diagnostic-imaging-the-flagship-application","sources":[{"label":"Lehman et al., 2017, Radiology","url":"https://doi.org/10.1148/radiol.2016161174","type":"paper"}]},{"category":"revision","title":"Undocumented SARS-CoV-2 infections in China before the January 2020 travel restrictions","change":"The Data Problem in Public Health AI illustrates how surveillance selection effects lead to systematic underestimation of disease burden: Li et al. (2020) estimated that 86% of SARS-CoV-2 infections in China were undocumented before the 23 January 2020 travel restrictions.","why":"Models trained on reported case counts inherit this gap and can confuse more testing with more disease.","chapters":[["The Data Problem in Public Health AI","/foundations/data.html"]],"summary":"Li and colleagues estimated that 86% of SARS-CoV-2 infections in China were undocumented before the 23 January 2020 travel restrictions.","section":"/foundations/data.html#sec-data-why-public-health-data-is-uniquely-challenging","sources":[{"label":"Li et al., 2020, Science","url":"https://doi.org/10.1126/science.abb3221","type":"paper"}]}]},
{"date":"2026-10-20","headline":"Course syllabus: grading rubric, learner AI-use policy, and CEPH competency map","entries":[{"category":"content","title":"Course syllabus: grading rubric, learner AI-use policy, and CEPH competency map","change":"Course Syllabus provides a grading rubric, a learner AI-use policy that keeps surveillance and case data out of unapproved tools, an assignment map to CEPH foundational competencies, and a course evaluation plan.","why":"Instructors can adapt the assessment criteria and data-use safeguards to public health AI teaching.","chapters":[["Course Syllabus Template","/appendices/course-syllabus.html"]],"summary":"Course Syllabus provides a grading rubric, a learner AI-use policy, an assignment map to CEPH foundational competencies, and a course evaluation plan.","section":"/appendices/course-syllabus.html#sec-course-ceph-alignment","sources":[]}]}
],"highlights":[{"date":"2026-10-20","title":"Course syllabus: grading rubric, learner AI-use policy, and CEPH competency map","tier":"major","category":"content","chapter":["Course Syllabus Template","/appendices/course-syllabus.html"],"sources":[]},{"date":"2026-10-10","title":"Google's Med-PaLM 2: question-answering performance and clinical limits","tier":"major","category":"revision","chapter":["Performance Metrics for Public Health AI","/implementation/performance-metrics.html"],"sources":[{"label":"Singhal et al., 2025","url":"https://doi.org/10.1038/s41591-024-03423-7","type":"paper"}],"vendor":true},{"date":"2026-10-10","title":"EASAC and FEAM recommend clearer Annex III guidance for healthcare AI","tier":"major","category":"revision","chapter":["AI Policy and Governance in Healthcare","/future/policy.html"],"sources":[{"label":"Cole-Hunter et al., EASAC and FEAM, 2026","url":"https://doi.org/10.1553/EASAC-FEAM_Report_AI-in-Healthcare_2026","type":"report"}],"section":"/future/policy.html#sec-future-policy-regulatory-landscape"},{"date":"2026-10-10","title":"Obermeyer et al.: health costs as a biased proxy for health need","tier":"major","category":"revision","chapter":["Global Health AI Equity and Local Capacity","/future/global-equity-capacity.html"],"sources":[{"label":"Obermeyer et al., 2019","url":"https://www.science.org/doi/10.1126/science.aax2342","type":"paper"}],"section":"/future/global-equity-capacity.html#sec-future-global-equity-capacity-algorithmic-fairness-across-populations"},{"date":"2026-10-09","title":"HHS civil enforcement for 42 CFR Part 2 substance-use treatment records","tier":"major","category":"policy","chapter":["AI for Substance Use and Overdose Prevention","/applications/substance-use.html"],"sources":[{"label":"HHS Office for Civil Rights, February 2026","url":"https://www.hhs.gov/press-room/hhs-announce-civil-enforcement-program-sud-patient-records.html","type":"policy","short_label":"HHS OCR"}]},{"date":"2026-10-08","title":"Anthropic cyber-defense access for hospitals and public infrastructure","tier":"major","category":"developments","chapter":["Selecting and Using Large Language Models","/future/llm-selection-use.html"],"sources":[{"label":"Anthropic, Cyber Verification Program, October 2026","url":"https://www.anthropic.com/news/cyber-verification-program","type":"announcement"},{"label":"Anthropic, Cyber Mission, October 2026","url":"https://www.anthropic.com/news/anthropic-cyber-mission","type":"announcement"}],"section":"/future/llm-selection-use.html#sec-future-llm-selection-use-choosing-the-right-llm-a-decision-framework","vendor":true},{"date":"2026-10-07","title":"OECD Expert-Led Madrid Action Plan for scaling AI in health","tier":"major","category":"policy","chapter":["AI Policy and Governance in Healthcare","/future/policy.html"],"sources":[{"label":"OECD, 2026","url":"https://www.oecd.org/en/publications/proceeding-report-from-madrid-conference-on-the-responsible-scale-of-ai-in-healthcare_cf9439c3-en.html","type":"policy"}],"section":"/future/policy.html#sec-future-policy-intl-governance"},{"date":"2026-10-06","title":"UK government accepts all 44 National Commission recommendations on healthcare AI","tier":"landmark","category":"policy","chapter":["AI Policy and Governance in Healthcare","/future/policy.html"],"sources":[{"label":"UK government response to the National Commission, October 2026","url":"https://www.gov.uk/government/publications/government-response-to-the-national-commissions-recommendations-on-the-regulation-of-ai-in-healthcare/government-response-to-the-national-commissions-recommendations-on-the-regulation-of-ai-in-healthcare","type":"policy"},{"label":"UK government announcement, October 2026","url":"https://www.gov.uk/government/news/government-backs-recommendations-of-nhs-doctors-led-ai-commission","type":"policy"}],"section":"/future/policy.html#sec-uk-national-commission"},{"date":"2026-10-05","title":"Rosenbaum: algorithm disclosure did not change flu-vaccine uptake","tier":"major","category":"evidence","chapter":["AI-Driven Behavioral Interventions in Public Health","/future/behavioral-ai.html"],"sources":[{"label":"Rosenbaum et al., 2026","url":"https://doi.org/10.1038/s41562-026-02603-4","type":"paper"}],"section":"/future/behavioral-ai.html#sec-future-behavioral-ai-application-areas"},{"date":"2026-10-01","title":"Bessemer and Bain report healthcare AI returns by use case","tier":"major","category":"developments","chapter":["AI Deployment in Healthcare: Why Most Prototypes Fail","/implementation/deployment.html"],"sources":[{"label":"Guerra and Kraus, 2026","url":"https://www.bvp.com/atlas/the-2026-healthcare-ai-roi-scorecard","type":"report"}],"section":"/implementation/deployment.html#sec-implementation-deployment-cost-effectiveness"},{"date":"2026-09-30","title":"Nature editorial and FDA discussion paper on generative-AI medical devices","tier":"major","category":"policy","chapter":["AI Policy and Governance in Healthcare","/future/policy.html"],"sources":[{"label":"Nature, 2026","url":"https://doi.org/10.1038/d41586-026-03046-7","type":"paper","short_label":"Nature editorial"},{"label":"FDA, Considerations for the Regulation of Generative AI-Enabled Medical Devices, 18 August 2026","url":"https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request","type":"policy","short_label":"FDA discussion paper"},{"label":"FDA docket FDA-2026-N-7874","url":"https://www.regulations.gov/docket/FDA-2026-N-7874","type":"policy","short_label":"Docket FDA-2026-N-7874"}],"section":"/future/policy.html#sec-cds-guidance-2026"},{"date":"2026-09-25","title":"Anthropic reports Claude use in a CEPI-convened Ebola response in eastern DRC","tier":"landmark","category":"developments","chapter":["AI in Public Health Emergency Operations","/implementation/emergency-operations.html"],"sources":[{"label":"Anthropic, Ebola response feature, 2026","url":"https://www.anthropic.com/features/ebola-response","type":"announcement","short_label":"Read report"}],"section":"/implementation/emergency-operations.html#sec-implementation-emergency-operations-emergency-operations-functions","vendor":true},{"date":"2026-09-22","title":"AI-in-health guidance from the UK National Commission, CDC, and WHO","tier":"landmark","category":"policy","chapter":["AI Policy and Governance in Healthcare","/future/policy.html"],"sources":[{"label":"National Commission into the Regulation of AI in Healthcare, September 2026","url":"https://www.gov.uk/government/publications/national-commission-into-the-regulation-of-ai-in-healthcare-recommendations-for-a-future-regulatory-framework/national-commission-into-the-regulation-of-ai-in-healthcare-recommendations-for-a-future-regulatory-framework","type":"policy","short_label":"UK Commission"},{"label":"CDC, Considerations for Agentic Research in Public Health, 2026","url":"https://www.cdc.gov/ai/resources/considerations-for-agentic-research-in-public-health.html","type":"policy","short_label":"CDC guidance"},{"label":"WHO, Artificial intelligence-related health research: ethics review and oversight, 2026","url":"https://www.who.int/publications/i/item/9789240124073","type":"policy","short_label":"WHO guidance"}],"section":"/future/policy.html#sec-uk-national-commission"},{"date":"2026-09-11","title":"Rounding et al. multi-country RCT of GPT-4o on clinical vignettes","tier":"major","category":"evidence","chapter":["AI in Global Health and Equity","/future/global-health.html"],"sources":[{"label":"Rounding et al., 2026","url":"https://doi.org/10.1038/s41746-026-03111-5","type":"paper","short_label":"Read paper"}],"section":"/future/global-health.html#sec-rounding-gpt4o-trial"}],"months":{"2025-10":"The handbook launched October 12, 2025 with five draft chapters, reached Version 1.0 on October 17 with all 20 chapters complete, then added a new AI Safety chapter on October 26. The month also introduced new chapters on deployment, misinformation, AI-assisted coding, and behavioral interventions, added a 500-plus-term glossary and 15 case studies, and closed with the license changed to CC BY-NC-ND 4.0.","2025-11":"The handbook extended its language-model release timeline through late November 2025, adding GPT-5.1, Grok 4.1, Gemini 3 Pro, and Claude Opus 4.5, and added a skeptical assessment of quantum computing, digital twins, and AI drug discovery to Emerging Technologies. Global Health and Equity gained a TL;DR noting that 90% of AI health research comes from high-income countries, and the license changed from CC BY-NC-ND 4.0 to CC BY 4.0.","2025-12":"December added a new executive summary, a chapter on AI for substance use and overdose prevention, and sections on wastewater surveillance and climate-health AI. The handbook added OpenAI's GPT-5.2 to its model comparisons, covered the Epic sepsis prediction model's positive predictive value of 12%, and added WHO's 2024 governance guidance for large multi-modal models, BEACON's open-access outbreak detection system, and CDC's ChatCDC tool.","2026-01":"January covered the launches of OpenAI for Healthcare and Anthropic's Claude for Healthcare, adding independent scrutiny of each vendor's benchmark claims, including HealthBench's conflict of interest and the Stanford-Harvard ARISE State of Clinical AI Report's benchmark-reality gap. The handbook registered a Zenodo DOI, added evidence including DeepSeek-R1 benchmarks and South Africa's MomConnect case study, and noted that Executive Order 14110 was rescinded and the FDA loosened Clinical Decision Support oversight.","2026-02":"The handbook added Kim et al.'s finding that only 6% of AI imaging studies were externally validated. The month documented a peer-reviewed audit of CDC surveillance database disruptions, added the West Coast Health Alliance as a CDC surveillance alternative, and cited a Nature Medicine trial finding LLM medical assistants did not improve lay decisions.","2026-03":"March added the first independent evaluation of ChatGPT Health's triage accuracy, a Mount Sinai study in Nature Medicine, alongside WHO's EIOS 2.0 upgrade and CEPI's zoonotic spillover risk partnership with BEACON and UC Davis. The month also added CDC's first AI strategy, a case study on rapid H5N1 genomic containment in the Cayman Islands, and findings from Anthropic's global AI user survey and a METR trial on AI coding productivity.","2026-04":"April added PandemicLLM, a fourth epidemic forecasting approach using language models, and a study finding ambient scribes reduced documentation time across 45 physicians. The month added evidence on Shadow AI governance gaps, a Science study identifying chatbot sycophancy risk across 11 AI models, the DETECT study on consumer wearables for outbreak detection, and updated the FDA's AI-enabled device count from over 1,300 to over 1,400.","2026-05":"May added a Nature study benchmarking AI-generated forecasting models against the CDC ensemble and a Nature Communications study of a transfer-learning early warning system deployed across 50 states. The month added CDC's FY2026-2030 AI Strategy, NIST's AI 800-4 monitoring guidance, and a systematic review finding generative AI increases health misinformation.","2026-06":"June added FDA's final guidance on Clinical Decision Support software and a note that 46% of CDC databases were updated at least monthly amid federal surveillance disruptions. The handbook cited a Nature Medicine study finding general-purpose LLMs outperformed specialized clinical AI tools such as OpenEvidence and UpToDate Expert AI, a proposed licensure framework for autonomous clinical AI, and a study documenting DeepSeek's deployment across 90 Chinese tertiary hospitals.","2026-07":"July added guidance matching evidence to five claim types, adding reporting standards including TRIPOD+AI, PROBAST+AI, DECIDE-AI, CONSORT-AI, and the NICE Evidence Standards Framework. The month also added antimicrobial resistance prediction studies, the TAKE IT DOWN Act's protections against nonconsensual intimate imagery, a Weir et al. dermatology AI skin-tone labeling study, and a KFF poll on patient use of AI for health information.","2026-08":"August split four chapters, including Global Health and Equity and LLMs in Public Health, into 18 focused chapters and added new chapters on Geospatial AI and Spatial Epidemiology, Emerging AI Architectures, and AI in Public Health Emergency Operations. The month added a Swedish hip-fracture prediction study and a TB REACH chest radiography study.","2026-09":"September added a census of 1,357 FDA-authorized AI/ML devices and surveillance, forecasting, and clinical studies. It records UK recommendations for regulating AI in healthcare, a WHO ethics-review report for AI-related health research, and a Nature editorial with FDA's discussion paper on generative-AI devices (comments due 19 October 2026; not guidance). Gemini 4 Argon is cyber-defense currency, and an HPV chatbot trial raised literacy while 2-week vaccination decisions were unchanged.","2026-10":"The UK government accepted all 44 National Commission recommendations on healthcare AI, while the OECD Expert-Led Madrid Action Plan proposes priorities for scaling AI safely. Rosenbaum and colleagues found small gains in flu-vaccine uptake from high-risk messages, with no difference from algorithm disclosure. HHS civil enforcement covers Part 2 substance-use records, and Google's Med-PaLM 2 question-answering evaluations do not establish clinical competence. Course Syllabus provides assessment criteria, a learner AI-use policy, a CEPH competency map, and an evaluation plan."},"upkeep":["2025-10","2025-11","2025-12","2026-01","2026-02","2026-03","2026-04","2026-05","2026-06","2026-07","2026-08","2026-09","2026-10"]}