A classification model for concurrent posterior circulation infarction in vertebrobasilar dolichoectasia using machine learning.

BACKGROUND

Vertebrobasilar dolichoectasia (VBD) is a cerebrovascular arteriopathy, and posterior circulation infarction (PCI) is frequently observed in this population. Currently, there are no widely available tools to classify the concurrent presence of PCI in patients with VBD. This study aimed to develop a machine learning-based classification model to identify concurrent PCI in this population using cross-sectional data.

METHODS

Clinical data were retrospectively collected from Tianjin Fourth Central Hospital. The cohort was randomly divided into a 70% training set and a 30% validation set. Eight candidate models were trained after Boruta feature selection. Model discrimination, calibration, and clinical utility were assessed using the receiver operating characteristic curve, calibration curve, and decision curve analysis. The best performing model was interpreted via SHAP-based variable importance and contribution analysis, and web calculator was subsequently deployed.

RESULTS

Our study included 380 participants, with median age 67 years, 262 (68.95%) male and 118 (31.05%) female, included 154 (40.53%) had PCI. Basilar artery diameter (BAD), basilar artery length (BAL), basilar artery midline distance (BAMD), systolic blood pressure, platelet, triglyceride, and uric acid were identified as factors associated with concurrent PCI. The Random Forest model (AUC: 0.846, 95% CI: 0.789-0.895) and the XGBoost model (AUC: 0.884, 95% CI: 0.842-0.925) exhibited the best discrimination, calibration, and clinical utility. SHAP summary of both optimal models indicated that BAD and BAMD were the dominant contributors to the classification of concurrent PCI.

CONCLUSION

We developed and internally validated a machine learning model that integrates clinical and imaging variables to classify concurrent PCI status in VBD. As this is a cross-sectional classification model, external validation and prospective studies are strictly required before clinical implementation. A freely accessible web calculator is provided as a preliminary tool to assist in identifying individuals with concurrent PCI and supporting clinical decision-making.