A Hybrid Explainable Boosting Machine and SHAP Framework for Early Prediction of Sepsis-Induced Hypotension in ICU Settings

Abstract

Sepsis-induced hypotension kills a substantial portion of ICU patients, yet clinicians frequently miss its early signs amid the noise of continuous monitoring. Standard early warning tools treat hemodynamic variables as static snapshots and cannot capture the nonlinear trajectories that precede decompensation. This paper describes a clinical decision-support system built around an Explainable Boosting Machine (EBM) coupled with SHapley Additive exPlanations (SHAP), trained on 4,212 adult sepsis admissions from a U.S. academic medical center (2019–2023). The objective was two-hour-ahead prediction of hypotension onset. The final model reached an AUROC of 0.94 (95% CI: 0.92–0.96), sensitivity 0.89, and specificity 0.91—outperforming logistic regression, random forest, XGBoost, and a deep neural network without sacrificing transparency. Heart rate variability, lactate acceleration, and MAP second-derivative were the three leading predictors. A prospective bedside simulation with 15 intensivists showed a 17-percentage-point accuracy gain (71% → 88%), 40% faster decisions, and substantially higher confidence. A Shapley interaction analysis also uncovered a clinically novel synergy: concurrent rising lactate and suppressed HRV inflated hypotension risk 3.2-fold relative to either signal alone. The framework is fully deployable (<50 ms inference), HIPAA-compliant, and designed to meet pending FDA Class II guidance on explainable clinical decision support.