Development and Validation of a machine learning model for predicting gastrointestinal bleeding after endovascular thrombectomy in acute ischemic stroke: a two-center study

Background We aimed to develop and externally validate an interpretable machine-learning model for predicting overt or endoscopically confirmed gastrointestinal bleeding within 72 hours after EVT. Methods This two-center retrospective study included a development cohort from the First Affiliated Hospital of Guilin Medical University (May 2021 to June 2025; n = 485) and an external validation cohort from Guilin People's Hospital (January 2024 to June 2025; n = 72). The primary outcome was overt or endoscopically confirmed gastrointestinal bleeding occurring within 72 hours after EVT. The intended prediction time was immediately after EVT completion, and all nine final predictors were available at or before that time. Predictors were selected using least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm, and seven supervised algorithms were compared. SHapley Additive exPlanations (SHAP) were used for model interpretation, and a web-based calculator was developed. Results The primary outcome occurred in 142 of 485 patients (29.3%) in the development cohort, and 20 patients (4.1%) experienced severe or fatal GIB. Nine predictors were retained in the primary model: NIHSS, Glasgow Coma Scale, modified Rankin Scale, Alberta Stroke Program Early CT Score (ASPECTS), location of vessel occlusion, the number of EVT passes, modified Thrombolysis in Cerebral Infarction grade, pre-admission anticoagulant therapy, and neutrophil count. The primary support vector machine (SVM) achieved AUROCs of 0.931 (95% CI, 0.902–0.961), 0.916 (0.857–0.974), and 0.865 (0.780–0.951) in the training, internal test, and external validation cohorts, respectively, with Brier scores of 0.094 and 0.136 in the internal test and external validation cohorts. Logistic regression (LR) showed similar discrimination in both validation cohorts. SHAP analysis characterized the contribution of each predictor, and the model was implemented as an online calculator for individual risk estimation. Conclusion This SVM model integrated nine variables available at or before EVT completion and showed good discrimination for estimating the probability of overt or endoscopically confirmed GIB during the subsequent 72 hours. Prospective multicenter validation across geographically diverse regions and practice settings is warranted before clinical use.