Predictive Analytics for Intraoperative Complications: Enhancing Perioperative Safety with AI

Abstract

Intraoperative problems significantly impact patient safety and surgical outcomes, with early detection of such problems being important to improve perioperative care. This article explores algorithms of machine learning to predict intraoperative complications using preoperative and intraoperative data from a large retrospective cohort of 121,898 adult surgical procedures at a single academic medical center between 2012 and 2016. To model the prediction of outcomes such as acute kidney injury (AKI), delirium, deep vein thrombosis (DVT), pulmonary embolism (PE), and pneumonia, five models were trained: logistic regression, support vector machine, random forest, gradient boosting tree (GBT), and deep neural network (DNN). Combination datasets performed better than preoperative or intraoperative data alone. The highest AUROC was 0.91 (GBT; pneumonia), 0.85 (aKI; GBT), 0.88 (DVT;  GBT), 0.76 (PE; DNN), and 0.999 (delirium; GBT) (Table 2). Including missing data variables yielded significant performance gain in all categories. SHapley Additive exPlanations (SHAP) discovered significant, patient-specific risk factors in a clinically relevant manner, thus enhancing interpretability. These findings demonstrate the potential for AI-driven predictive analytics to provide physicians with interpretable, real-time decision support, reduce complication rates and enhance perioperative safety overall.