Alert Burden, Calibration, and Clinical Utility in Early Sepsis Prediction: A Systematic Review of Explainable Multimodal Models

Abstract

Early sepsis prediction models often report discrimination while providing limited information about calibration, alert volume, or the number of clinically actionable warnings generated in practice. We systematically reviewed explainable multimodal sepsis prediction studies published through 2025. Database searches yielded 1,926 records, of which 74 studies met inclusion criteria. Most models combined vital signs and laboratory data; fewer incorporated clinical text or waveform information. AUROC was reported in 93% of studies, but only 38% reported calibration and 27% quantified alert burden at an operational threshold. Explanations were commonly generated using feature attribution, yet only 16 studies evaluated explanation faithfulness or stability. External validation was performed in 22% of studies and often showed lower performance than internal testing. Studies that reported decision-curve or alert-rate analyses provided a clearer picture of potential clinical utility than discrimination metrics alone. Evidence supports continued development of multimodal sepsis prediction, but deployment claims should be tempered when calibration, alert burden, and external generalisation are unreported. Standard evaluation should include these measures alongside explainability outputs.