Updates in Predictive and Generative AI in Medicine: Uses and Pitfalls
Recorded session from the SGIM 2026 Annual Meeting.
Artificial intelligence is being integrated into medicine at a rapid pace, but its applications fall into two distinct domains: predictive AI and generative AI. Distinguishing between the two matters — success in generative AI (like drafting an HPI from a conversation) is often mistakenly assumed to signal progress in predictive AI (like the accuracy of inpatient falls-risk scores), when the two are largely unrelated. For phenomena shaped by human behavior in particular, predictive accuracy remains a fundamental barrier.
This session highlights influential papers published over the past 12–24 months, critically analyzing their relevance to general internal medicine practice, teaching, and system-level integration. The first part focuses on predictive AI models used for risk stratification and clinical outcome prediction, such as sepsis mortality and falls prevention — examining both their potential benefits and their methodological limitations, including overfitting, lack of prospective validation, overreliance on AUC as a sole accuracy metric, and the inherent uncertainty in predicting future outcomes. Attendees leave equipped to critically appraise predictive AI papers and spot bias or overstated performance claims.
The second part addresses generative AI, reviewing studies evaluating large language models for tasks central to clinicians and educators — clinical reasoning, teaching, coaching, and assessment. The session weighs both promise and risk: generative AI can improve efficiency and provide educational scaffolding for learners, but unchecked errors can cause patient and learner harm, and outputs may reflect cognitive biases or inconsistency even with careful prompting. Presenters share cases where generative AI has been implemented successfully, while emphasizing the need for careful human oversight.
Throughout, presenters use a critical appraisal framework so attendees leave with practical tools for interpreting new AI literature responsibly — not just an update on cutting-edge studies, but the means to weigh the uses and pitfalls of predictive and generative AI in their own clinical and educational work.
Presenters
Amiran Baduashvili, MD
Verity Schaye, MD, MHPE
Karlen Ulubabyan, MD
Justin Choi, MD
Alice J. Tang, MD, MHPE
