🌍𝗠𝗲𝗱𝗶𝗰𝗮𝗹 𝗔𝗜 𝘀𝗵𝗼𝘂𝗹𝗱 𝗯𝗲 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲𝗱 𝗶𝗻 𝘁𝗵𝗲 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀, 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝗶𝗲𝘀 𝗮𝗻𝗱 𝗵𝗲𝗮𝗹𝘁𝗵 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝘄𝗵𝗲𝗿𝗲 𝗶𝘁 𝘄𝗶𝗹𝗹 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗯𝗲 𝘂𝘀𝗲𝗱. Delighted to announce the launch of #MAGIC (𝗠𝗲𝗱𝗶𝗰𝗮𝗹 𝗔𝗜 𝗚𝗹𝗼𝗯𝗮𝗹 𝗜𝗻𝗶𝘁𝗶𝗮𝘁𝗶𝘃𝗲 𝗮𝗻𝗱 𝗖𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻) introduced in our new @NaturePortfolio @NatureMedicine correspondence. 📜MAGIC paper: nature.com/articles/s41591-0… 🔗MAGIC website: magic-ai.health/ MAGIC brings together clinicians, AI researchers, policymakers, industry stakeholders and local experts to address a critical question: does strong benchmark performance translate into safe, useful support across diverse healthcare settings? Our initial priorities are: ✅ A global health LLM evaluation benchmark addressing multilingual coverage, regional disease burden and sociocultural adaptability. ✅ A global deployment reporting standard clarifying intended use, target users, healthcare settings, and performance and safety across populations. Thank you to our colleagues #worldwide for helping launch the network. 𝗢𝘂𝗿 𝗴𝗼𝗮𝗹 𝗶𝘀 𝘁𝗼 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱 𝘄𝗵𝗲𝗿𝗲 𝗺𝗲𝗱𝗶𝗰𝗮𝗹 𝗔𝗜 𝘄𝗼𝗿𝗸𝘀, 𝗳𝗼𝗿 𝘄𝗵𝗼𝗺, 𝗮𝗻𝗱 𝘂𝗻𝗱𝗲𝗿 𝘄𝗵𝗮𝘁 𝗰𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻𝘀. #MAGIC #MedicalAI #AI #LLM #GlobalHealth #ResponsibleAI #HealthEquity #DukeNUS @dukenus

Honored to be 𝗿𝗮𝗻𝗸𝗲𝗱 𝟮𝟭𝟬𝘁𝗵 𝗴𝗹𝗼𝗯𝗮𝗹𝗹𝘆 (𝗮𝗺𝗼𝗻𝗴 𝘁𝗵𝗲 𝘁𝗼𝗽 𝟬.𝟬𝟱% 𝘄𝗼𝗿𝗹𝗱𝘄𝗶𝗱𝗲) in the ScholarGPS 2026 Highly Ranked Scholars – Lifetime rankings (Machine Learning specialty). This recognition is a reflection of the collective effort behind our work. My sincere thanks to every member of the 𝗗𝗶𝗴𝗶𝘁𝗮𝗹 𝗠𝗲𝗱𝗶𝗰𝗶𝗻𝗲 𝗟𝗮𝗯 and to our collaborators around the world. #MachineLearning #MedicalInformatics #AI #ResponsibleAI #TrustworthyAI #DigitalHealth #ResearchImpact #DukeNUS

𝗘𝘁𝗵𝗶𝗰𝘀 𝗶𝗻 𝗵𝗲𝗮𝗹𝘁𝗵 𝗔𝗜 𝗰𝗮𝗻𝗻𝗼𝘁 𝗯𝗲 𝗮 𝗳𝗶𝗻𝗮𝗹 𝗰𝗵𝗲𝗰𝗸𝗯𝗼𝘅 𝗯𝗲𝗳𝗼𝗿𝗲 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁. Sharing our new article in @Patterns_CP: "Ethical principles and considerations in health AI." Our central argument: ethical principles should guide the entire #AI #lifecycle, from development and validation to deployment and ongoing monitoring, and must be considered together, not in isolation. We highlight three practical points, among many others: ✅ 𝗘𝘁𝗵𝗶𝗰𝗮𝗹 𝗽𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲𝘀 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁. Efforts to improve equity can create trade-offs with predictive performance, requiring careful assessment of safety and alternative safeguards. ✅ 𝗧𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆 𝗻𝗲𝗲𝗱𝘀 𝗮 𝗽𝘂𝗿𝗽𝗼𝘀𝗲. Clinicians need evidence about local performance, uncertainty and limitations, not just model explanations, to decide when to follow, question or override AI. ✅ 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀. Clinical priorities and available resources shape which safeguards are feasible, particularly in low-resource settings. This is not another checklist. It is an approach to understanding how ethical principles interact and translating them into decisions at the model, workflow and health-system levels. Thank you to our international team for this collaboration. 𝗢𝘂𝗿 𝗴𝗼𝗮𝗹 𝗶𝘀 𝘁𝗼 𝗺𝗮𝗸𝗲 𝗲𝘁𝗵𝗶𝗰𝘀 𝗮𝗻 𝗼𝗻𝗴𝗼𝗶𝗻𝗴 𝗽𝗮𝗿𝘁 𝗼𝗳 𝗵𝗼𝘄 𝗵𝗲𝗮𝗹𝘁𝗵 𝗔𝗜 𝗶𝘀 𝗯𝘂𝗶𝗹𝘁, 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲𝗱 𝗮𝗻𝗱 𝘂𝘀𝗲𝗱, 𝗻𝗼𝘁 𝗮𝗻 𝗮𝗳𝘁𝗲𝗿𝘁𝗵𝗼𝘂𝗴𝗵𝘁. 📖 Read open-access paper: doi.org/10.1016/j.patter.202… #AI #Ethics #ResponsibleAI #HealthAI #AIEthics #TrustworthyAI #DigitalHealth #DukeNUS @CellPressNews

Our paper on ethical considerations and principles for medical AI was just published Led by @nliulab w/ @YilinNing and a global team of AI researchers cell.com/patterns/fulltext/S…

Glad to be included again in the 2026 World's Top 2% Scientists list (Stanford & Elsevier), and heartened to be ranked 122 among 22,436 researchers in Medical Informatics, my primary research field. While this recognition is encouraging, citations are not the purpose of our work. What matters most is conducting meaningful research, addressing real healthcare needs, and contributing to the communities we serve. Look forward to continuing our work on trustworthy and responsible AI for healthcare, with clinical value, patient well-being and community benefit at its heart. #MedicalInformatics #ResponsibleAI #TrustworthyAI #AI #LLM #DigitalHealth #ResearchImpact #DukeNUS

🌍The global #AI divide in healthcare is widening, and privacy laws are part of why. Sharing our new article in @NaturePortfolio @npjDigitalMed. The core problem: clinical AI is trained mostly on data from high-resource settings, while low-resource settings, which carry the greatest disease burden, remain underrepresented. Pooling global data could be a fix to this. But privacy regulations, data sovereignty concerns, and the risk of "digital colonialism" mean that data often cannot move across borders. It's a deadlock: local performance needs diverse data, but privacy concerns keep that data locked in place. Privacy-enhancing technologies (#PETs) such as de-identification, synthetic data, federated learning, and cryptographic methods, could offer a way through, shifting the paradigm from data-sharing to knowledge-sharing. But not all PETs are equally deployable. In resource-constrained settings, techniques like statistics-based federated learning and transfer learning are more feasible than computationally expensive approaches like homomorphic encryption. The rise of #LLMs adds new urgency, and new risks, to this picture. Real progress requires more than good algorithms. We outline four pathways: 1) data governance built on sovereignty rather than extraction, 2) harmonized model-sharing guidelines, 3) accessible/lightweight infrastructure, and 4) trust-based partnerships. 📜 Read paper here: nature.com/articles/s41746-0… #GlobalHealth #DigitalHealth #AI #HealthEquity #PrivacyEnhancingTechnologies #FederatedLearning #HealthAI #DataGovernance #LMIC #ResponsibleAI #DukeNUS

[🌍𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 & 𝗣𝗼𝗹𝗶𝗰𝘆 𝗖𝗼𝗿𝗻𝗲𝗿] A healthcare AI model can be accurate when tested and still be unreliable when used. Why? Because clinical reliability is not a fixed property of a model. It depends on which patients encounter it, where and how it is used, and whether its performance remains acceptable as clinical practice changes. A new Review in @CellCellPress, "Avoiding common failures in AI for health and medicine," examines reliability across both predictive and generative AI. The authors highlight three recurring and often interconnected failures. 1️⃣ 𝗘𝗿𝗿𝗼𝗻𝗲𝗼𝘂𝘀 𝗼𝘂𝘁𝗽𝘂𝘁𝘀 𝗺𝗮𝘆 𝗯𝗲 𝗱𝗶𝗳𝗳𝗶𝗰𝘂𝗹𝘁 𝘁𝗼 𝗱𝗲𝘁𝗲𝗰𝘁 Predictive models can learn from incomplete records, inconsistent labels, or proxies that do not accurately represent the underlying clinical condition. Generative AI can produce fluent and authoritative responses that are clinically incorrect, making unreliable outputs particularly difficult to recognize. 2️⃣ 𝗔𝘃𝗲𝗿𝗮𝗴𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗰𝗮𝗻 𝗰𝗼𝗻𝗰𝗲𝗮𝗹 𝘂𝗻𝗲𝗾𝘂𝗮𝗹 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 A system may perform differently across patient populations, hospitals, scanners, input formats, or clinical settings for reasons unrelated to genuine clinical differences. Strong overall results may therefore coexist with clinically important failures for particular populations or contexts. 3️⃣ 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗰𝗮𝗻 𝗱𝗲𝗴𝗿𝗮𝗱𝗲 𝗮𝗳𝘁𝗲𝗿 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 Changes in patient populations, workflows, documentation practices, coding systems, or data collection can make a previously validated model less reliable. These changes may be gradual, difficult to observe, and entangled with the model's own effects on clinical decisions and recorded outcomes. The paper also explains why familiar technical fixes are often insufficient. Data cleaning, model refinement, retrieval systems, recalibration, and retraining can help, but they frequently depend on assumptions that don't hold in dynamic clinical environments. Addressing one known failure may also leave other failures undetected or introduce new ones. The authors therefore frame reliability as a lifecycle responsibility. Reliable healthcare AI is not achieved by optimizing a model once. It requires continuous coordination across technical evaluation, clinical oversight, monitoring, and institutional governance. Olawale Salaudeen, et al. Avoiding common failures in AI for health and medicine. Cell. 2026; 189: 5893-5904. doi.org/10.1016/j.cell.2026.… #AIGovernance #HealthAI #ResponsibleAI #TrustworthyAI #ClinicalAI #AISafety #GenerativeAI #DigitalHealth #HealthcareInnovation #DukeNUS [𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 & 𝗣𝗼𝗹𝗶𝗰𝘆 𝗖𝗼𝗿𝗻𝗲𝗿] is a regular series in which we will share and reflect on relevant research, frameworks, and policy developments shaping trustworthy and ethical AI in healthcare.

In a The New York Times @nytimes op-ed, geriatrician Dr. Rachael Bedard shares a striking confession: #AI is making her a better doctor, but she worries it's making her trainees worse. As Dr. Adam Rodman, physician and AI researcher at Harvard Medical School, put it in the piece: "the cognitive offloading can cause never-skilling." The New York Times article: lnkd.in/g_K7wQKc Glad to see our research aligning with what's being observed at a major international venue. Read more about this concept in our @NatureMedicine article "AI-induced never-skilling in medical education" (nature.com/articles/s41591-0…). #NeverSkilling #AIinMedicine #AI #MedicalEducation #ClinicalReasoning #DukeNUS

📢New in @NatureHumBehav: "How AI companions could deepen social inequality," with @qiyang_zhang & @renwenzhang. AI companions feel more patient & available than any human relationship - so who really benefits? nature.com/articles/s41562-0… #AI #inequality #SMU #NTU #DukeNUS

AI companions are often seen as tools for connection and support. In this Comment, @nliulab and colleagues discuss how they could also exacerbate existing social inequalities. [7/10] #AICompanions nature.com/articles/s41562-0…

Retrieval-Augmented Generation (#RAG) in medicine: promising, but still early-stage. #LLMs are transforming clinical practice, but hallucinations, static knowledge, and bias remain real risks. Our @CellRepMed paper shows where the field stands. cell.com/cell-reports-medici… #DukeNUS

Attending International Conference on Machine Learning (#ICML) at Seoul. Happy to connect and talk about #AI and #ML for #healthcare and #medicine.

Skin cancer #AI models can exhibit #gender #bias. Our work introduces #LesionAttn, which focuses on clinician-annotated lesions while down-weighting background signals. @npjDigitalMed @NaturePortfolio nature.com/articles/s41746-0… #Dermatology #AIFairness #MedicalAI #DigitalHealth

We've written about AI-induced physician deskilling that has already surfaced thelancet.com/journals/lance… @tberzin AI-induced never-skilling among newly trained doctors, while not yet proven, is a serious concern that needs to be addressed @NatureMedicine @nliulab nature.com/articles/s41591-0…

Our new @NatureMedicine Perspective argues that uncritical use of #AI in early medical training may weaken independent clinical reasoning, and proposes a framework for integrating AI while protecting core competencies. nature.com/articles/s41591-0… @NaturePortfolio @dukenus #DukeNUS

𝗖𝗮𝗻 𝗔𝗜 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗼𝗻𝘀 𝗿𝗲𝗮𝗹𝗹𝘆 𝗿𝗲𝗽𝗹𝗮𝗰𝗲 𝗵𝘂𝗺𝗮𝗻 𝗿𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝘀𝗵𝗶𝗽𝘀? We argue that #AI companionship is not just a technical innovation but a deep relational shift. papers.ssrn.com/sol3/papers.… #AICompanions #GenAI #DigitalIntimacy #AIethics #DukeNUS

Ambient AI is no longer just documenting care. It’s starting to shape it. In our latest in @npjDigitalMed article, we examine how ambient AI scribes are moving into complex care settings, and why that changes everything. nature.com/articles/s41746-0…

📜Innovating global regulatory frameworks for generative AI in medical devices is an urgent priority. @NaturePortfolio @npjDigitalMed Read paper: doi.org/10.1038/s41746-026-0… #GenerativeAI #LLM #AIinHealthcare #DigitalHealth #MedTech #HealthAI #AIRegulation #ResponsibleAI #DukeNUS

By analyzing 19,123 #NLP-related studies using topic modeling, we mapped the research landscape on #LLMs vs non-LLMs, and compared task distributions and application contexts across these two technical paradigms. @NaturePortfolio npj Health Systems nature.com/articles/s44401-0…