To compare the predictive accuracy of deep neural network and ensemble learning models for 3-month functional outcome in anterior circulation acute ischemic stroke (AIS)

Collateral circulation critically influences infarct dynamics and recovery in AIS. Artificial intelligence (AI) allows automated, data-driven evaluation of collateral flow, offering potential for individualized prognostication and treatment optimization.

This retrospective cross-sectional analytic study utilized a curated dataset of 240 consecutive patients with anterior circulation AIS. Each sample included structured clinical features and neuroimaging features as input variables, with 3-month modified Rankin Scale (mRS) scores as output labels. The dataset was randomly split into training, validation, and testing cohorts. Seven supervised models—Decision Tree, Random Forest, Bagging, AdaBoost, Gradient Boosting, XGBoost, and Light Gradient Boosting Machine (LGBM)—were developed to classify outcomes (favorable mRS ≤2 vs. poor mRS >2). Model performance was compared across metrics, including accuracy, F1 score, recall, and ROC-AUC.

Of 240 patients (age 59 ± 11 years; 71% male), median NIHSS was 6 (IQR 3–10) and median Clot Burden Score 9 (IQR 7–10). NIHSS correlated with mRS (r = 0.48, p < 0.001) and CBS (r = 0.26, p < 0.001).LGBM achieved the highest performance (accuracy 81%, F1 = 0.88, ROC-AUC = 0.68), followed by XGBoost (accuracy 80%, F1 = 0.87) and Gradient Boosting (accuracy 79%, F1 = 0.87). Decision Tree yielded the lowest accuracy (73%, F1 = 0.83). Ensemble models consistently outperformed single classifiers.

Deep learning models integrating CTA-derived collaterals and clinical data accurately predict 3-month outcomes after anterior-circulation AIS.