Skip to content

S4 — AI and Machine Learning in Clinical Medicine

Full session title: AI and Machine Learning in Clinical Medicine: Bridging or Separating Model Intelligence and Human Expertise? Papers: 27 (pp. 12–416) — the largest session at PSB 2026, roughly half the meeting's peer-reviewed volume Session introduction: Nateghi Haredasht F. et al., pp. 1–11 · PDF · PMID 41758129

Organizers:

  • Fateme Nateghi Haredasht — Stanford Center for Biomedical Informatics Research
  • Joseph D. Romano — Biostatistics, Epidemiology and Informatics, University of Pennsylvania
  • Brett K. Beaulieu-Jones — Computational Biomedicine and Biomedical Data Science, University of Chicago
  • Dokyoon Kim — Biostatistics, Epidemiology and Informatics, University of Pennsylvania
  • Alexander Ioannidis — Biomedical Data Science, Stanford; Biomolecular Engineering, UC Santa Cruz
  • Geoffrey H. Tison — Cardiology, Center for Biosignal Research, UCSF
  • Roxana Daneshjou — Biomedical Data Science, Stanford
  • Jonathan H. Chen — Medicine and Center for Biomedical Informatics Research, Stanford

Reading this session

The framing question in the title — bridging or separating model intelligence and human expertise — is answered mostly in the separating direction. An unusual share of these papers are evaluations, benchmarks, guardrails, and negative results rather than model announcements. That is the notable thing about this session and the reason it is worth the corpus's attention: it is a large, peer-reviewed sample of what happens when clinical LLM claims are tested carefully.

Four sub-clusters organize the 27 papers:

Foundation-model evaluation and benchmarks (the important cluster). Zhang X. et al., Automated Chest X-ray Report Generation Remains Unsolved, is a multi-institution negative result and the single most citable paper here for anyone arguing that benchmark scores overstate clinical readiness. Alongside it: Johri et al.'s PanEndoAtlas / PanEndoFM — 420k endoscopy images from 30 datasets, 13 countries, 111 GI diseases, with a GI foundation model benchmarked against EndoFM-LV, EndoSSL, ViT-B/16 and ResNet-50; Pal et al.'s ReXVQA chest X-ray VQA benchmark; Sambara et al.'s 3DReasonKnee for grounded reasoning in medical vision-language models; and Banerjee et al.'s ReXecution framework, which names the gap between what a medical AI intends and what it executes.

LLM agents and agentic evaluation. Eric Chen et al.'s MedAgentBench v2 is the direct link to the corpus's agentic-ai thread — an agent benchmark, revised, which is a healthier signal than a first-version leaderboard.

LLM-as-evaluator and safety infrastructure. Grolleau et al. (MedFactEval / MedAgentBrief), Wu et al. (concordance with specialist eConsult responses), McCoy et al. (clinical consultation templates), and W. Chen et al. (retrieval-augmented guardrails for AI-drafted patient portal messages, with an error taxonomy). Collectively: how do you check an LLM's clinical output at scale without a physician reading every one.

EHR foundation models and clinical NLP. Burkhart et al.'s Quantifying surprise in clinical care uses FM-derived event informativeness across a whole hospitalization context — the closest thing here to a "what is the model actually representing" paper. The NER/NLP papers (Dey, Prioleau, Hwang, F. Chen, Pugh, Zolensky) are a solid but more conventional block.

Papers

All PDFs at https://psb.stanford.edu/psb-online/proceedings/psb26/<file>.

Title First author Pages PDF PMID
Inference Gap in Domain Expertise and Machine Intelligence in Named Entity Recognition: Creation of and Insights from a Substance Use-related Dataset Sumon Kanti Dey 12–26 dey.pdf 41758130
ColonCrafter: A Depth Estimation Model for Colonoscopy Videos Using Diffusion Priors Romain Hardy 27–41 hardy.pdf 41758131
A Clinician-Guided Framework for Endoscopic AI: Developing PanEndoAtlas and Benchmarking Foundation Models Across the Full GI Spectrum Shreya Johri 42–56 johri.pdf 41758132
Higher-order Interaction Matters: Modeling Epidemics via Dynamic Hypergraph Neural Networks Songyuan Liu 57–70 liu_s.pdf 41758133
Scoring Physician Risk Communication in Prostate Cancer Using Large Language Models Guillermo Lopez-Garcia 71–84 lopez-garcia.pdf 41758134
SeizureFormer: A Multi-Scale Transformer for Seizure Risk Forecasting from RNS-Derived Biomarkers Tianning Feng 85–98 feng.pdf 41758135
3DReasonKnee: Advancing Grounded Reasoning in Medical Vision Language Models Sraavya Sambara 99–113 sambara.pdf 41758136
Abstention and Threshold Identification for Uncertainty Management in Clinical Decision Tools — Pediatric Autism Classifiers Aiden Ko 114–129 ko.pdf 41758137
Evaluation of Large Language Models as Emergency Department Revisit Predictors Emma Chen 130–143 chen_emma.pdf 41758138
Speaker Role Identification in Clinical Conversations Andrew Zolensky 144–157 zolensky.pdf 41758139
Learning Causally Predictable Outcomes from Psychiatric Longitudinal Data Eric V. Strobl 158–172 strobl.pdf 41758140
Quantifying surprise in clinical care: Detecting highly informative events in EHRs with foundation models Michael C. Burkhart 173–188 burkhart.pdf 41758141
Retrieval Augmented Guardrails for AI Drafted Patient Portal Messages: Error Taxonomy Construction and Large-Scale Evaluation Wenyuan Chen 189–204 chen_w.pdf 41758142
Leveraging Large Language Models for Adverse Drug Event Detection: Token vs Span-Based NER Howard Prioleau 205–218 prioleau.pdf 41758143
Towards Automated Analysis of Gaze Behavior from Consumer VR Devices for Neurological Diagnosis Lio Schmitz 219–235 schmitz.pdf 41758144
Automated Chest X-ray Report Generation Remains Unsolved Xiaoman Zhang 236–250 zhang_x.pdf 41758145
ReXVQA: A Large-scale Visual Question Answering Benchmark for Generalist Chest X-ray Understanding Ankit Pal 251–264 pal.pdf 41758146
Detecting PTSD in Clinical Interviews: A Comparative Analysis of NLP Methods and LLMs Feng Chen 265–279 chen_f.pdf 41758147
ED-Explain: Personalized Video Instructions for Patients Discharged from the Emergency Department Luyang Luo 280–293 luo.pdf 41758148
The Intention-Execution Disconnect in Medical AI: The ReXecution Framework Oishi Banerjee 294–308 banerjee.pdf 41758149
Leveraging Generative AI for Interpretable Clinical Decision Making Through Causal Graphs Mehmet Eren Ahsen 309–323 ahsen.pdf 41758150
Leveraging LLMs to Derive Multiple Sclerosis Progression Assessments from Clinical Notes Sy Hwang 324–337 hwang.pdf 41758151
WATCH-SS: Detecting Cognitive Impairment from Spontaneous Speech Sydney Pugh 338–353 pugh.pdf 41758152
MedAgentBench v2: Improving Medical LLM Agent Design Eric Chen 354–371 chen_eric.pdf 41758153
Automated Evaluation of LLM Response Concordance with Human Specialist Responses on Physician-to-Physician eConsult Cases David JH Wu 372–387 wu_d.pdf 41758154
MedFactEval and MedAgentBrief: A Framework and Workflow for Generating and Evaluating Factual Clinical Summaries François Grolleau 388–399 grolleau.pdf 41758155
Asking the Right Questions: Benchmarking LLMs in the Development of Clinical Consultation Templates Liam G. McCoy 400–416 mccoy.pdf 41758156

Connections to the corpus

Sources