Interpreting Sepsis Prediction through Positional Explanation

Chaehyeon Kim, Gary E. Weissman, Eric Wong

Keywords: Explainability, Interpretability, Sepsis Prediction

Abstract: Sepsis prediction models achieve high accuracy but go unused at the bedside. Their explanations cannot answer the questions clinicians actually ask: which measurement signaled deterioration, and when? Existing attribution methods score features but conflate a measurement's value with its timing, obscuring trajectories needed for early action. We introduce Position-aware eXplanation (PaX), a framework that decomposes each attribution into a "what" score for the measurement feature and a "when" score for its temporal position. Across Mamba, GPT-2, and MedGemma on PhysioNet and MC-MED, PaX surfaces second-order early-warning indicators that may signal early deterioration, improves alignment with clinical expertise, and exposes positional biases obscured by standard attribution methods.

Submission Number: 158