MultiDiseasePred: An Interpretable Multi-Disease Risk Prediction Framework and Its Application for Early Risk Stratification in the Emergency Department

Abstract At emergency department triage, clinicians must assess multiple overlapping acute conditions simultaneously, yet many existing risk tools are developed for individual conditions and provide outputs that are difficult to compare across diseases with substantially different frequencies. We developed and validated MultiDiseasePred, an interpretable multi-task learning framework for simultaneous risk estimation across multiple non-mutually exclusive clinical targets, and demonstrated its application for early multi-condition risk profiling in emergency department triage. MultiDiseasePred estimated risks for seven acute conditions (sepsis, ARDS, COPD/asthma exacerbation, acute coronary syndrome, viral pneumonia, pulmonary embolism, acute kidney injury) and two disposition endpoints (hospitalization; a critical outcome) in a single inference pass. A per-target gate weighted a shared expert pool differently for each target, allowing the model to learn shared clinical and physiological signals across related conditions while preserving target-specific prediction pathways. To reduce feature acquisition burden while maintaining model performance, a parsimonious feature selection procedure identified 12 variables available at triage from routinely available candidate predictors, without requiring laboratory results, imaging, or downstream clinical documentation. Calibrated probabilities were summarized using prevalence-adjusted lift measures to enable comparison across targets with different baseline prevalence, and each risk estimate was accompanied by target-specific feature attributions indicating which triage findings contributed to the prediction. Across 2,203,437 encounters from three health systems, the model achieved a macro-AUROC of 0.900 (95% CI, 0.897–0.903) for the seven acute conditions and 0.852 (0.849–0.856) for the two disposition endpoints in internal validation, outperforming NEWS and generally matching or exceeding single-task machine-learning baselines. In external validation without local retraining or recalibration, acute-condition macro-AUROC was 0.793 at Beth Israel Deaconess Medical Center and 0.866 at Stanford. These findings support MultiDiseasePred as a generalizable framework for interpretable, early multi-condition risk profiling and motivate prospective evaluation in emergency care workflows.