Interpretability of SHAP Explanations for Artificial Intelligence Sepsis Alerting Under Bedside Monitor Alarm Fatigue Filtering: A Healthcare Reliability Evaluation

Sepsis alerting models increasingly rely on Shapley additive explanations (SHAP) to justify individual predictions at the bedside, yet the reliability of those explanations has not been characterised under the alarm-fatigue filtering that governs which monitor signals reach the model. Power curves, minimum detectable effect curves, coefficient intervals, a scatter fit, a kernel density estimate and a receiver operating characteristic are reported to characterise the estimator's operating characteristics. The framework reframes SHAP reliability as a property of the alerting pipeline rather than of the attribution algorithm alone, and specifies the conditions under which bedside explanations can be treated as stable evidence for clinical review.