Risk-Calibrated Medical LLM Triage with Physician-Rubric Self-Verification and Patient-Facing Evidence Cards: An Empirical Health Bench Study

Abstract

Medical large language models can answer patient questions fluently while remaining unreliable about urgency, uncertainty, and safe escalation. This study evaluated a risk-controlled post-generation architecture on HealthBench 2025, comprising 5,000 physician-annotated medical conversations, 57,237 weighted rubric criteria, 4,206 physician ideal responses, and 11,112 stored answers across 2,778 four-candidate cases. The architecture combined a word-level triage classifier, fold-local rubric retrieval, a word–character–retrieval ensemble with temperature scaling, deterministic answer reranking, triage-policy self-verification, selective escalation, and patient-facing evidence cards. Five-fold out-of-fold evaluation on 453 physician-labeled emergency-referral cases produced a macro-F1 of 0.645 for the self-verified ensemble versus 0.620 for direct TF-IDF; expected calibration error fell from 0.126 to 0.058, and under-triage fell from 18.5% to 16.1%. At a 0.70 confidence threshold, coverage was 32.0%, selective risk was 22.1%, and urgent-route recall reached 98.1%. On the 2,778 candidate-selection cases, rubric-RAG raised weighted rubric-alignment coverage from 0.801 to 0.845. Self-verification preserved 0.845 alignment while reducing the response-level triage-policy error proxy from 53.8% to 42.3%. The results show that retrieval improves content coverage, whereas a distinct safety check is needed to convert content quality into safer disposition behavior. Confidence-aware abstention and concise evidence cards provide an auditable interface for patient-facing use, while the remaining uncertainty supports clinician oversight rather than autonomous deployment.