Machine learning-based models for predicting 30-day cerebral infarction after transient ischemic attack: a retrospective study

Transient ischemic attack (TIA) is an important warning sign of impending stroke. This study aimed to develop models for predicting secondary cerebral infarction within 30 days after TIA in patients with diffusion-weighted imaging (DWI)-negative TIA using least absolute shrinkage and selection operator (LASSO) regression and machine-learning algorithms, and to compare their predictive performance. We retrospectively enrolled 270 patients with DWI-negative TIA treated between ​​January 2023 and July 2025. Patients were stratified by outcome and randomly divided into training ( n = 189) and test ( n = 81) sets at a 7:3 ratio. All predictor selection, feature standardization, and hyperparameter tuning were performed exclusively within the training set using stratified cross-validation, and the test set was locked until final evaluation. LASSO, applied within the training set, retained seven predictors: age, sex, histories of hypertension and diabetes, TIA duration ≥ 1 h, FVH, and an incomplete circle of Willis. Support vector machine (SVM), random forest (RF), and XGBoost models were developed, with logistic regression as the reference. Discrimination was assessed by the area under the receiver operating characteristic curve (AUC) with percentile-bootstrap 95% confidence intervals (CIs), accuracy, precision, sensitivity, specificity, and F1 score; calibration (Brier score, calibration intercept and slope), precision–recall curves, decision-curve analysis, and confusion matrices were also evaluated, and AUCs were compared using DeLong tests. Secondary cerebral infarction within 30 days occurred in 86 of 270 patients, and LASSO retained seven predictors. In the test set, RF attained the numerically highest AUC (0.846, 95% CI 0.751–0.929) and the lowest Brier score (0.144); Brier scores ranged from 0.144 to 0.173 and average precision from 0.71 to 0.74 across the four models. DeLong testing showed no significant differences among the four models in the test set (all P > 0.05). At sensitivity-prioritized thresholds, the models detected 69–96% of test-set events. Repeating feature selection within the training data alone retained the same core predictors and left the test-set AUCs essentially unchanged. Against the approximated ABCD2 clinical score (test AUC 0.804, 95% CI 0.690–0.900), the imaging-augmented models showed comparable overall discrimination (DeLong P = 0.38–0.67). In this modest, single-center cohort of patients, all four models showed useful but only moderate discrimination for secondary cerebral infarction within 30 days, and calibration and sensitivity were imperfect. No machine-learning model significantly outperformed logistic regression in discrimination; this reflects a lack of demonstrated superiority rather than evidence of equivalence. The models may serve as adjuncts to clinical assessment rather than substitutes for it, and require external validation and calibration before clinical use.