DeltaSHAP: Explaining Prediction Evolutions in Online Patient Monitoring with Shapley Values

Explains changes in patient risk, not just individual predictions: DeltaSHAP focuses on why a patient's risk score is increasing or decreasing over time, rather than explaining isolated predictions. This addresses a critical clinical need—doctors need to understand what's driving deterioration to intervene quickly, not just know the current risk level.

Provides real-time insights for time-sensitive decisions: The algorithm delivers explanations fast enough for online monitoring systems (33% faster than existing methods), which is essential in clinical settings where delays in understanding risk changes could affect patient outcomes.

Shows both what changed and by how much: Unlike some explanation methods, DeltaSHAP indicates both the direction (which factors increased vs. decreased risk) and magnitude of each factor's contribution, giving clinicians actionable information about which specific changes matter most.

Uses only realistic feature combinations: The method bases explanations on actual combinations of patient measurements that were observed, rather than hypothetical scenarios. This makes the explanations more trustworthy and practical for clinical decision-making.

Demonstrates measurable improvements in explanation quality: On the MIMIC-III patient dataset, DeltaSHAP achieved 62% better explanation quality compared to state-of-the-art methods, suggesting it more accurately identifies which factors truly drive changes in patient risk predictions.