Explainable AI models for predicting venous thromboembolism following revascularization therapy in ischemic stroke patients: A retrospective cohort study.

Deep vein thrombosis (DVT) is a serious complication in acute ischemic stroke (AIS) patients undergoing revascularization therapy, but prediction tools remain limited. To develop and validate machine learning models for predicting DVT risk in AIS patients after revascularization therapy, and to enhance clinical decision-making through model interpretability. A retrospective cohort study was conducted using data from the Shenzhen Stroke Database, including AIS patients who underwent endovascular thrombectomy and/or thrombolytic therapy. Various machine learning models, including random forest (RF), support vector machine, gradient boosting machine, decision tree, and Gaussian naive Bayes, were trained and validated using a 70:30 train-validation split. The synthetic minority over-sampling technique was applied to address class imbalance. Among 362 AIS patients undergoing revascularization therapy, DVT incidence was 8.84%. The RF model achieved the highest prediction accuracy with an area under the curve of 0.87. Key predictors included D-dimer levels, aspirin use, National Institutes of Health Stroke Scale score during hospitalization, international normalized ratio, and anti-infective treatments. SHapley Additive exPlanations analysis enhanced model interpretability, providing clear insights into individual predictor contributions. The RF model significantly improved DVT risk prediction in AIS patients post-revascularization, offering a more accurate and interpretable tool for clinical practice.