Interpretable Feature Extraction from Clinical Notes for Sepsis Prediction: Comparing Rule-Based, LLM, and Hybrid Approaches.

Embedding-based approaches integrate clinical notes into sepsis prediction models but produce uninterpretable representations, obscuring which clinical findings drive predictions, and limiting both trust and regulatory acceptance. We compared three frameworks for extracting interpretable features from MIMIC-III clinical notes (217 sepsis cases, 2,170 controls): rule-based (M1), zero-shot LLM with retrieval-augmented generation (M2), and hybrid combining symbolic pre-filtering with LLM quantification (M3), evaluating 19 SOFA, qSOFA, and SIRS-based features using expert validation of 190 predictions. M1 achieved high precision but limited recall, M2 attained broad coverage (recall: 0.95, F1: 0.83) but lower precision (0.74), while M3 demonstrated superior balance (precision: 0.90, recall: 0.94, F1: 0.92), extracting quantitative measurements in JSON. Hybrid architectures combining symbolic constraints with LLMs enable accurate, interpretable extraction, improving precision and maintaining recall with outputs.