Explaining Temporal Effects in Sepsis Prediction

Sepsis prediction models remain opaque to clinicians which hinder clinician adoption: without understanding why a patient is flagged as high-risk, accurate predictions may be ignored, delaying critical intervention. Existing explainability methods focus on feature importance and often overlook timing, thus failing to capture the temporal influences inherent in time-series data. We propose Positional Explanation, which separates attributions into feature content and it's position to highlight temporal effects, enabling clinicians to identify early warning indicators and monitor for specific physiological changes at critical time windows before sepsis develops. Applied to GPT-2 and Mamba models finetuned for sepsis prediction on PhysioNet and MC-MED benchmarks, our method achieves higher faithfulness scores and reveals temporal patterns in sepsis progression that existing techniques miss, potentially enabling earlier detection and improved patient outcomes.