ICML2026 Medical LLM AI paper notes paper summaries Medical Imaging Alignment/RLHF Adversarial Robustness Domain Adaptation Reasoning

🩺 Medical LLM¶

🧪 ICML2026 · 4 paper notes

📌 Same area in other venues: 📷 CVPR2026 (1) · 🔬 ICLR2026 (20) · 💬 ACL2026 (47) · 🤖 AAAI2026 (12) · 🧠 NeurIPS2025 (17) · 🧪 ICML2025 (4)

🔥 Top topics: Medical Imaging ×2

A Machine-Learned Comorbidity Index: Traditional comorbidity scores (Charlson, Elixhauser) are linear rules with weights manually calibrated for mortality, performing poorly on other clinical outcomes. This paper utilizes neural networks to compress ICD codes from an admission into a scalar score, trained by maximizing the normalized HSIC (kernel dependence) between this score and multiple clinical outcomes. This ensures the single score provides consistent severity ranking across mortality, readmission, length of stay, and ICU admission. The dependence metrics on MIMIC-III/IV significantly exceed those of traditional indices and various machine learning baselines.
ClinTutor-R1: Advancing Scalable and Robust One-to-Many Alignment in Clinical Socratic Education: This paper proposes ClinTutor-R1, the first vision-language agent for one-to-many alignment in clinical Socratic education. By constructing the 48k ClinTeach dialogue dataset via the ClinEdu multi-agent simulator, and utilizing explicit Theory of Mind (ToM) reasoning alongside three-axis rubric reinforcement learning, the model maintains stable teaching quality even when scaled to 10 students, outperforming baselines by 20% and reaching GPT-4o performance levels.
Exploring Accurate and Transparent Domain Adaptation in Predictive Healthcare via Concept-Grounded Orthogonal Inference: ExtraCare utilizes a "dictionary metric-induced orthogonal decomposition" to decouple Electronic Health Record (EHR) patient representations into "cross-domain invariant label information" and "domain-specific covariate residuals." It surpasses existing domain adaptation baselines on two real-world EHR datasets while mapping each latent variable back to specific ICD medical concepts via sparse dimension ablation. This informs clinicians exactly what was "preserved" and "discarded" during the adaptation process.
MedCase-Structured: A Text-to-FHIR Dataset for Benchmarking Diagnostic Reasoning in Clinically Realistic EHR Settings: The authors propose a "multi-stage LLM + terminology grounding + repair loop" pipeline to convert free-text medical cases into HL7 FHIR R4 standard bundles. Using this, they construct the MedCase-Structured dataset (1,408 cases, 82.5% success rate) from MedCaseReasoning. Experiments demonstrate that diagnostic accuracy for GPT-5.4, Gemini-3.1-Pro, and Claude-Opus-4.6 consistently drops by 4–23% when using structured FHIR inputs compared to raw text.