An interpretable machine-learning model for predicting the need for diagnostic cerebral angiography in acute ischemic stroke patients in imaging-limited primary hospitals

In resource-limited primary hospitals, timely imaging may be unavailable for patients with acute ischemic stroke, complicating decisions on whether immediate Diagnostic Cerebral Angiography (DSA) is warranted. This study aimed to develop an interpretable machine-learning model to predict the likelihood of immediate angiography in AIS patients without imaging-based assessment. This single-center retrospective cohort study used de-identified AIS data from the Department of Neurology, General Hospital of the Western Theater Command, between June 2021 and December 2022. Of 1,026 screened patients, 605 were included and randomly split into a training set ( n = 425) and a validation set ( n = 180; 7:3). The outcome was whether DSA was performed without prior imaging evaluation. Thirty-two candidate predictors were collected using the first available measurements within 24 h after admission. Feature selection combined LASSO and Boruta with expert review, yielding seven predictors (TG, SBP, COPD, Hcy, age, MAP, and HbA1c). Seven algorithms (LR, DT, SVM, RF, XGBoost, LightGBM, and ANN) were trained; SMOTE was applied to the training set to address class imbalance, with nested cross-validation and hyperparameter optimization. Discrimination, calibration, and decision curve analysis were used for performance and clinical utility assessment. Model interpretability was evaluated using SHAP. The proportions of patients undergoing DSA were 19.05% in the training cohort and 18.89% in the validation cohort. All models showed good performance in the training set: LR achieved the highest AUC (0.974), XGBoost performed best on several classification metrics, and RF yielded the lowest Brier score (0.007). In the validation set, discrimination decreased for both RF and XGBoost with comparable performance; however, RF demonstrated better calibration and higher net benefit within threshold probabilities of 0.10–0.40 and 0.60–0.90, and was therefore selected as the final model. SHAP analysis identified TG as the most influential predictor, followed by SBP, COPD, Hcy, age, MAP, and HbA1c. An interpretable random forest model incorporating seven routinely available clinical indicators can predict the likelihood of immediate DSA in AIS patients without imaging support, thereby assisting primary care physicians in referral decision-making, offering practical decision assistance in resource-limited settings. External validation and prospective studies are needed to confirm robustness and generalizability.