Clinical AI &
Medical Education

The clinical AI landscape has shifted dramatically in 2025–2026. Adam Rodman's landmark study (Harvard/BIDMC) shows LLMs scoring 10/10 on the r-IDEA clinical reasoning framework — surpassing the 9/10 average for attending physicians. Yet this headline result coexists with a troubling paradox: clinicians augmented with AI decision support sometimes underperform AI alone. Automation bias causes clinicians to delegate reasoning to the system, so when the AI errs on a difficult case, augmented clinicians fail to catch it. Simply providing AI predictions is insufficient — systems must actively engage clinician reasoning. This gap between capability and safe deployment is the central open problem in the domain.

On the education side, AI is reshaping how medicine is taught. Stanford's Fall 2025 medical curriculum now integrates AI literacy throughout, not as an elective. Faculty surveys reveal the scale of the challenge: 91% agree AI will help clinical practice, but only 12% feel "very familiar" with AI capabilities. Adaptive tutoring and virtual patient simulation — grounded in cognitive architectures (ACT-R for illness scripts, IRT for ability estimation, FSRS for spaced repetition, ZPD for scaffolding) — offer a principled approach to closing this gap at scale. Logical Neural Networks (LNNs), demonstrated at AMIA 2025, combine the interpretability of first-order logic rules with learnable weights, outperforming random forests and gradient boosting on clinical diagnosis tasks while remaining auditable by domain experts.

These problems connect directly to Ashish's research portfolio. Cardio-Sahayak is a CVD clinical decision support system calibrated for South Asian populations, where standard Western biomarker thresholds (BMI, HbA1c, lipid profiles) systematically misperform. Virtual Tumor Board is a multi-agent simulation of the multi-disciplinary tumor board process. Onco-Shikshak is Ashish's flagship medical education system: a 6-agent AI tumor board simulation integrating ACT-R, IRT, FSRS, Bloom's Taxonomy, ZPD, and Deliberate Practice to model learner cognition at multiple levels simultaneously. Project B11 feeds directly into Cardio-Sahayak; project I13 extends Onco-Shikshak; project A10 targets the hardest open problem in both systems.

The neuro-symbolic direction — led by collaborator Ishaan Varior (Direction D3) — targets what no existing system achieves: a dual-process clinical reasoning architecture where System 1 (LLM pattern matching) generates rapid hypotheses and System 2 (formally verified symbolic rules) analytically validates them, with metacognitive monitoring of when to escalate. No current clinical AI system has formal safety proofs that safety-critical rules — "never give immunotherapy to a patient with active autoimmune disease" — cannot be overridden by the neural component. That is what A10 attempts to build.