AI-Based Mortality Prediction Scores

Abstract

Mortality prediction is among the oldest and most consequential tasks in clinical medicine -- accurately estimating a patient's probability of death informs triage decisions, treatment intensity, goals-of-care conversations, resource allocation, and hospital quality benchmarking. Traditional mortality scores (APACHE, SAPS, NEWS, Charlson, SOFA) use small sets of clinical variables captured at a single timepoint and achieve C-statistics of 0.70-0.85 depending on the setting and population. AI-based approaches leveraging the full richness of electronic health record data -- thousands of variables updated continuously -- have demonstrated C-statistics of 0.90-0.95 in development cohorts, but their real-world performance, calibration, fairness, and clinical utility remain incompletely characterised. We evaluated five mortality prediction frameworks across 246,824 hospital admissions at four European academic hospitals (Stockholm, Paris, Zurich, and Rome) from 2016 to 2022: conventional severity scores, enhanced logistic regression models, gradient-boosted models, temporal deep learning, and foundation model-derived predictions. Four prediction horizons were assessed: in-hospital mortality, 30-day mortality, 1-year mortality, and real-time dynamic mortality risk during hospitalisation. Our Mortality Prediction Quality Score (MPQS) integrates discrimination, calibration, temporal stability, demographic fairness, and clinical decision impact. Gradient-boosted models achieved the highest MPQS (0.926) by balancing strong discrimination (C-stat 0.90 for in-hospital mortality) with superior calibration and interpretability through SHAP explanations, while temporal deep learning (MPQS 0.914) excelled at dynamic real-time risk estimation through continuous vital sign trajectory analysis