Predictive Value of Machine Learning Models for Functional Recovery Following Endovascular Thrombectomy in Patients with Acute Vertebrobasilar Artery Occlusion

Introduction: This study aimed to develop machine learning prediction models to identify high-risk populations with Vertebrobasilar Artery Occlusion (VBAO) who are likely to experience poor prognosis after Endovascular Thrombectomy (EVT), thereby informing clinical decision-making. materials and methods: A retrospective cohort of 204 patients with acute VBAO undergoing EVT at the Department of Neurology, Affiliated Hospital of Xuzhou Medical University (January 2020–October 2024) was analyzed. Multimodal data, including medical history, clinical symptoms, laboratory parameters, hemodynamic indices, and procedural details, were collected. Patients were stratified into favorable (modified Rankin Scale mRS 0–3) and poor prognosis (mRS 4–6) groups based on 90-day outcomes. Seven machine learning models—logistic regression, random forest, gradient boosting, XGBoost, LightGBM, CatBoost, and AdaBoost—were constructed and evaluated using five-fold cross-validation. Model performance metrics and predictor importance were analyzed. Methods: A retrospective cohort of 204 patients with acute VBAO undergoing EVT at the Department of Neurology, Affiliated Hospital of Xuzhou Medical University (January 2020– October 2024) was analyzed. Multimodal data, including medical history, clinical symptoms, laboratory parameters, hemodynamic indices, and procedural details, were collected. For missing data, Multiple Imputation (MI) was performed to reduce information bias. Patients were stratified into favorable (modified Rankin Scale mRS ≤2) and poor prognosis (mRS >2) groups based on 90-day outcomes. Seven machine learning models-logistic regression, random forest, gradient boosting, XGBoost, LightGBM, CatBoost, and AdaBoost-were constructed and evaluated using five-fold cross-validation. Model performance metrics and predictor importance were analyzed. results: Among 204 patients, the random forest model demonstrated the highest performance (AUC = 0.9648, 95% CI: 0.9356–0.9941), significantly outperforming traditional LR (AUC = 0.76, p lt; 0.01). Key predictors included hypertension (HTN), type 2 diabetes mellitus (T2DM), heart ratxmmmme (HR), National Institutes of Health Stroke Scale (NIHSS) score, procedural duration, blood glucose (BG), and albumin. SHapley Additive exPlanations (SHAP) analysis identified BG as the most influential predictor, contributing 31.2% to the model. Decision curve analysis (DCA) confirmed clinical utility at a threshold probability of 0.1, supporting its application in guiding preventive or therapeutic interventions. Results: Among 204 patients, the random forest model demonstrated the highest performance (AUC = 0.9648, 95% CI: 0.9356-0.9941), significantly outperforming traditional LR (AUC = 0.76, p <0.01). Key predictors included Hypertension (HTN), Type 2 Diabetes Mellitus (T2DM), Heart Rate (HR), National Institutes of Health Stroke Scale (NIHSS) score, operation time, Blood Glucose (BG), and albumin. SHapley Additive exPlanations (SHAP) analysis identified BG as the most influential predictor, accounting for 31.2% of the model's predictive power. Decision Curve Analysis (DCA) confirmed clinical utility at a threshold probability of 0.1, supporting its application in guiding preventive or therapeutic interventions. Discussion: The random forest model’s superiority lies in its ability to capture non-linear variable interactions, overcoming LR's limitations. BG and systemic inflammatory indices (SIRI, NLR) highlight the role of metabolic-inflammatory crosstalk in prognosis. The model integrates easily accessible data (avoiding over-reliance on imaging) and offers SHAP-based interpretability, but is limited by single-center design, small sample size, and lack of external validationissues to address in future multicenter studies. conclusion: This machine learning model provides a precise decision-support tool for preoperative risk stratification in VBAO patients, offering significant clinical value for optimizing post-EVT neurological outcomes. Conclusion: This machine learning model provides a precise decision-support tool for preoperative risk stratification in VBAO patients, offering significant clinical value for optimizing post- EVT neurological outcomes.