Large Language Models (LLMs) excel at medical question-answering, yet progress toward medical reasoning models is limited by the lack of scalable, trustworthy reasoning supervision. We propose tapping into the so far underused resource of medical consensus guidelines. These map numerical clinical measurements to disease scores and can optionally be curated for site-specific exceptions. Automatically generating gold-standard reasoning chains from consensus rules enables both large-scale supervised fine-tuning and automatic evaluation. Applying SOFA, SAPS-II, and Sepsis-3 guidelines to MIMIC-III, we show that small fine-tuned models significantly outperform larger one-shot prompted models in derivation correctness and F1 score, achieving near-perfect deductive inference and reliable exception handling. Trustworthy step-by-step reasoning does not reliably emerge from one-shot prompting, even when models receive the consensus definition and an example. Performance across all models declines when forecasting future clinical variables, indicating that, on the evaluated structured ICU tasks, the central challenge is not rule application, but the orthogonal problem of generalizing into the future for prognostic reasoning.
翻译:暂无翻译