Machine learning-based prediction of cerebral artery territorial infarcts in acute ischemic stroke using non-contrast CT
Yang Guo, S Liu, Sun H, Xiukun Jin, Yan Zhang, Jian Sun, Zhiqun Wang, Lvming Zhang
Background Ischemic stroke is a leading global cause of death and disability. Accurate, rapid infarct territory localization is critical for timely treatment, but ultra-early ischemic changes on non-contrast CT (NCCT) are subtle and highly physician-dependent. Methods This retrospective study enrolled acute ischemic stroke patients with single-territory infarction. Clinical data and NCCT images were collected. Quantitative imaging features (HU values, infarct volume, spatial distribution) were extracted from vascular perfusion territories. A strict leak-proof pipeline was adopted: dataset stratification, missing value imputation, scaling, and Synthetic Minority Oversampling Technique (SMOTE) were applied only to the training set, full feature retention (VIF < 5), and Logistic Regression (LR) and XGBoost models were constructed. Standardized labeling was performed. Anterior cerebral artery (ACA), middle cerebral artery (MCA), and posterior cerebral artery (PCA) territories were defined by two neuroradiologists using diffusion-weighted imaging (DWI) (gold standard, Kappa analysis). Multi-territory infarcts were excluded. TotalSegmentator segmented parenchymal perfusion territories. DWI infarct ROIs were registered to NCCT. Model performance was assessed by Area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision and calibration (Brier score, slope), with DeLong-Bonferroni tests. Robust validation used 5-fold repeated stratified cross-validation (10 repetitions). Results A total of 118 eligible patients were enrolled (median age: 72.0 years; IQR: 65.0–79.0; 85.6% male). All patients received NCCT within 24 h of stroke onset, with an average NIHSS score of 6.05 ± 3.26. Inter-rater Kappa for infarct labeling reached 0.89 (excellent agreement). On the test set, XGBoost yielded AUCs of 0.92 (ACA), 0.87 (MCA), and 0.95 (PCA), with a cross-validated mean AUC of 0.94 ± 0.03. Mild overfitting was noted due to a performance drop from training to test data. LR produced AUCs ranging 0.78–0.86 with balanced sensitivity. For left/right/bilateral infarct classification, XGBoost achieved a higher mean AUC (0.90 vs. 0.82, p < 0.01). Both models exhibited good calibration. Conclusion XGBoost demonstrates excellent diagnostic performance for infarct territory localization, whereas LR offers stable sensitivity. Strict protocols and robust validation support its clinical auxiliary value. Future studies will incorporate multi-center external validation and multi-territory infarcts to improve generalizability.