NEJM AI Grand Rounds

The promise of medical AI is not simply better performance — it is better care grounded in evidence patients and clinicians can trust. Dr. Xiao Liu, a clinician and medical AI researcher affiliated with the University of Birmingham and Microsoft AI, has spent her career asking what rigorous evaluation should look like as new technologies move from publications into practice. She describes reporting guidelines as tools for transparency rather than prescriptions and explains why seeing product development from inside industry changed her understanding of implementation. Looking ahead, Liu imagines a medical profession in which readily available knowledge shifts physicians toward a more interpersonal role: accompanying patients through uncertainty, deterioration, treatment failure, recovery, and life-changing diagnoses. At the same time, she sees AI opening new scientific possibilities by revealing biomarkers, disease phenotypes, and treatment-response patterns that current medical categories may miss. The challenge is to preserve rigor without letting familiar methods prevent medicine from learning something genuinely new.

Transcript.

A patient can deteriorate while the warning signs remain buried in a noisy electronic record. Dr. Suchi Saria, Founder and CEO of Bayesian Health and a faculty member at Johns Hopkins University, has spent her career trying to surface those signals sooner. She explains why a strong retrospective model is only the starting point — and why prospective validation, workflow integration, physician adoption, monitoring, and financial sustainability determine whether clinical AI ever reaches patients. Her work on early sepsis recognition became personal after she lost her nephew to the condition, turning an academic agenda into an urgent effort to prevent failures to rescue. Saria also challenges the current center of gravity in health care AI: tools increasingly prepare for visits, transcribe conversations, and support billing, while comparatively few directly help clinicians care for patients. She envisions a third modality of medicine alongside drugs and devices — rigorous, software-based interventions that identify when to act, what to do, and for whom.

Transcript.

Brandon Rice believes that improving health care sometimes means improving the systems behind it. Through his work at Weave, he focuses on modernizing the regulatory infrastructure that helps bring new medicines to market. He discusses the challenge of organizing scientific knowledge, communicating with regulators, and navigating processes that can span more than a decade. His vision is that AI can reduce administrative burden while preserving the rigor required for patient safety and scientific progress.

Transcript. 

Medical expertise has always been scarce. Dr. Karan Singal believes AI can help change that. Drawing on his work at OpenAI and earlier efforts behind Med‑PaLM, he discusses how clinicians and patients are already using AI to answer questions, support decisions, and navigate care. He argues that the future of health AI is not only about improving model performance, but also about helping people advocate for themselves more effectively. Through HealthBench and ChatGPT for Clinicians, his team is exploring how to make these systems safer, more useful, and more trustworthy. The result is a vision of health care where expertise becomes more accessible without losing sight of clinical responsibility.

Transcript. 

Dr. Travis Zack, Chief Medical Officer of OpenEvidence, takes us behind the scenes of the start and growth of the company, and brings a clinician’s perspective to one of medicine’s hardest questions: how should artificial intelligence support decision-making? In this episode, he emphasizes that reasoning—not just correctness—defines good care, and that evidence must be contextual, accessible, and usable. He explores how physicians use AI to reduce uncertainty, why global constraints challenge the idea of a single “right answer,” and how trust depends on transparent use of medical literature. For clinicians navigating complex decisions, this conversation highlights both the promise and the limits of AI—and the enduring importance of human judgment.

Transcript. 

Doctronic CMO Dr. Byron Crowe describes how administrative complexity can interfere with timely, effective treatment, and how AI may help address those challenges. Crowe discusses Doctronic’s use of autonomous AI to renew prescriptions, arguing that this application can streamline care while maintaining clinical oversight. For physicians, this shift raises important questions about workflow, responsibility, and patient engagement. Crowe emphasizes that the goal is not automation for its own sake, but more reliable and accessible care. As these tools evolve, their impact will depend on how thoughtfully they are integrated into clinical practice.

Transcript.

Dr. Kyunghyun Cho is a leading AI researcher best known for co-authoring a landmark 2014 paper that introduced neural machine translation. In this episode, he discusses his wide-ranging career spanning fundamental AI research, co-founding Prescient Design (acquired by Genentech), and driving applications of AI in health care. For clinicians, Cho’s core message is pragmatic: AI should help health care run better. After years of work at NYU Langone, he reframed AI in medicine from solving rare diagnostic puzzles to improving operational prediction at scale. Cho emphasizes purpose‑built data, careful fine‑tuning, and regulatory accountability. His perspective connects technical rigor with system stewardship—and insists that patient voices must be present in AI governance.

Transcript. 

Clinical AI only helps patients if clinicians and health systems trust it. Seth Hain describes how Epic is building foundation models that respect institutional autonomy, minimize burden, and prioritize safety. He discusses scaling laws in structured medical data, cautious deployment for clinical interventions, and why understanding causality—not just correlation—is essential. This conversation reframes AI not as disruption, but as infrastructure for safer, more reliable care.

Transcript. 

For Dr. Marinka Zitnik, the promise of AI in medicine begins with acknowledging the scale of the problem. Most patients with rare diseases have no approved treatments, and traditional drug development timelines make progress painfully slow. In this conversation, she describes how AI-driven drug repurposing offers a way to work within existing constraints while still opening new therapeutic possibilities.

She also highlights a structural issue that has limited impact: machine learning and biology communities often work in parallel, not together. By building shared benchmarks and collaborative spaces, Marinka argues, researchers can focus models on problems that truly matter for patients.

The episode introduces her definition of AI agents as systems that can take actions and learn from outcomes — a capability she sees as essential for scientific discovery beyond static prediction. Throughout the discussion, Marinka returns to the value of academic freedom: the ability to chase difficult questions that require long time horizons and interdisciplinary thinking.

Transcript.

For Dr. Zak Kohane, this year’s advances in AI weren’t abstract. They were personal, practical, and deeply tied to care. After decades studying clinical data and diagnostic uncertainty, he finds himself building his own EHR, reviewing his child’s imaging with AI, and re-thinking the balance between incidental and missed findings. Across each story is the same insight: clinicians and machines make mistakes for different reasons — and understanding those differences is essential for safe deployment.

In this episode, Zak also highlights where AI is spreading fastest, and why: reimbursement. While dermatology and radiology aren’t broadly using AI for interpretation, revenue-cycle optimization is advancing rapidly. Meanwhile, ambient documentation has exploded — not because it increases accuracy or throughput, but because it improves clinician satisfaction in strained systems.

Yet the most profound theme, he argues, is values. Models already show implicit preferences: some conservative, some aggressive. And unlike human clinicians, no regulatory framework examines how those preferences form. Zak calls for a new form of oversight that centers patients, recognizes bias, and bridges clinical expertise with technical transparency.

Transcript.