Multicenter validation of a scalable, interpretable, multitask prediction model for multiple clinical outcomes.

Abstract

Predicting multiple postoperative complications remains challenging in perioperative care. Current approaches often address complications individually, limiting the potential for integrated risk assessment. We developed and externally validated a scalable, interpretable, tree-based multitask learning model to predict three critical postoperative outcomes-acute kidney injury (AKI), postoperative respiratory failure (PRF), and in-hospital mortality-using 16 preoperative features generally available in electronic health records. Our model achieved AUROCs of 0.805, 0.789, and 0.863 for AKI; 0.886, 0.925, and 0.911 for PRF; and 0.907, 0.913, and 0.849 for mortality in the derivation cohort and external validation cohorts A and B, respectively (all p < 0.001, except for AKI in derivation and PRF in cohort B). We also elucidated the contribution of each input variable to predictions among different outcomes. Our findings highlight the potential of multitask learning to streamline preoperative risk assessment and present a scalable, interpretable, and generalizable framework for improving perioperative care.