Compact machine learning model for perioperative stroke prediction prior to surgery: A retrospective cohort study

Perioperative stroke significantly impacts postoperative outcomes. Current risk stratification methods for perioperative stroke prediction lack accuracy and practicality. We aimed to develop a machine learning (ML) model that improves both accuracy and usability for predicting perioperative strokes. Data from 36,502 patients at Seoul National University Hospital (SNUH) were utilized to develop and internally validate. An external validation was conducted using data from 404 patients at Boramae Medical Center (BMC). Perioperative stroke was defined as a brain infarction of ischemic etiology, occurring within 30 days post-surgery. We developed ML-based prediction models comprising preoperative features and compared them with cardiovascular scores. Additionally, we developed a compact model utilizing the 10 most significant features of the best-performing model. The CatBoost-based prediction model showed the best discriminatory power for high-risk patients and outperformed cardiovascular scores in the external validation set (area under the receiver operating characteristics curve [AUC], 0.867 [95% CI: 0.830-0.896]; revised cardiac index score, 0.528 [95% CI: 0.497-0.575; p < 0.000]; CHA₂DS₂VASc score, 0.706 [95% CI: 0.659-0.748; p < 0.000]. The compactmodel also improved performance (AUC of 0.875 [95% CI: 0.860-0.952]). Our ML-based perioperative stroke prediction model improves accuracy and clinical usability.