Predictive Analytics for Intensive Care Admission Following Surgery

Abstract

Deciding which surgical patients require postoperative intensive care remains a judgment call made with incomplete information, and legacy scoring systems such as APACHE II, POSSUM, and the Surgical Apgar Score were derived from modest cohorts using a small set of fixed-weight variables. This review examines the growing body of machine learning and deep learning approaches developed to predict postoperative ICU admission, with particular attention to the harder and more clinically consequential problem of unplanned rather than planned transfer. Tree-based ensembles trained on combined preoperative and intraoperative variables consistently outperform classical logistic regression and legacy risk indices, with disease-specific models in oncologic, spinal, and cardiac surgery populations showing further gains from tailored feature sets, while deep learning and multimodal architectures are beginning to integrate time-series and narrative EHR data for real-time deterioration prediction. Despite this progress, nearly every model reviewed remains single-center, retrospective, and externally unvalidated, and calibration, class imbalance, and clinician trust in model interpretability remain unresolved barriers to routine deployment. We conclude that prospective, multicenter validation paired with implementation studies measuring the effect of model-guided triage on actual bed utilization and outcomes is the necessary next step before these tools can move from promising retrospective analyses to routine perioperative practice.