A retrospective investigation assessing computational algorithms for secondary stroke risk has found that machine-learning models achieve moderate discrimination in predicting thirty-day cerebral infarction following a transient ischemic attack, without demonstrating superiority over traditional logistic regression. While the analyzed algorithms exhibited useful predictive capacity, their overall discriminative accuracy remained moderate, and their diagnostic sensitivity and statistical calibration were imperfect across the cohort.
Transient ischemic attack is an important warning sign of impending stroke. The study retrospectively enrolled 270 patients with diffusion-weighted imaging-negative transient ischemic attack treated between January 2023 and July 2025. Across this full cohort of 270 individuals, secondary cerebral infarction occurred within thirty days in 86 patients.
The 270 enrolled patients were stratified according to their thirty-day clinical outcome and randomly divided into a training set of 189 patients and a testing set of 81 patients, establishing a seven-to-three allocation ratio. The test set was locked until final evaluation. In accordance with the study protocol, all predictor selection procedures, feature standardization steps, and hyperparameter tuning were conducted exclusively within the training cohort using stratified cross-validation routines.
Predictor selection within the training set was executed using least absolute shrinkage and selection operator, or LASSO, regression. This screening process narrowed the initial feature pool down to seven retained clinical and imaging predictors: patient age, biological sex, documented medical histories of hypertension and diabetes, transient ischemic attack symptom duration lasting at least one hour, FVH, and the structural presence of an incomplete circle of Willis. Repeating feature selection within the training data alone retained the same core predictors.
Using the seven predictors selected by LASSO, the researchers trained three distinct machine-learning models: a support vector machine, a random forest algorithm, and an extreme gradient boosting model, known as XGBoost. In parallel, a standard multivariable logistic regression model was constructed using the same predictors to serve as the established reference standard. The performance of each model was comprehensively evaluated through discrimination, measured by the area under the receiver operating characteristic curve with percentile-bootstrap 95 percent confidence intervals, overall accuracy, precision, sensitivity, specificity, and F1 score. The research team also evaluated model calibration through Brier scores, calibration slopes, and calibration intercepts, alongside decision-curve analysis, precision-recall curves, and confusion matrices.
In the locked test set of 81 patients, the random forest algorithm achieved the numerically highest discrimination, yielding an area under the receiver operating characteristic curve of 0.846 with a percentile-bootstrap 95 percent confidence interval spanning from 0.751 to 0.929. The random forest model also attained the lowest Brier score at 0.144. Across all four models, Brier scores ranged from 0.144 to 0.173, while average precision values ranged between 0.71 and 0.74. Formal statistical comparisons using DeLong tests demonstrated that the differences in the area under the receiver operating characteristic curve among the random forest, support vector machine, XGBoost, and logistic regression models were not statistically significant, with all resulting P values exceeding 0.05.
When the researchers applied classification thresholds prioritizing diagnostic sensitivity, the models successfully identified between 69 percent and 96 percent of secondary cerebral infarction events within the test set. The authors also benchmarked their imaging-augmented models against an approximated ABCD2 clinical score, which yielded a test-set area under the receiver operating characteristic curve of 0.804 with a 95 percent confidence interval from 0.690 to 0.900. DeLong test comparisons showed that the imaging-augmented algorithms achieved overall discrimination comparable to the approximated ABCD2 score, with P values ranging from 0.38 to 0.67. Furthermore, re-running feature selection solely within the training partition left the test-set discrimination values essentially unchanged.
In evaluating the broader clinical implications of their findings, the authors stressed that their retrospective cohort was derived from a single medical center and comprised a modest sample size. They underscored that calibration and sensitivity remained imperfect across all models, and that the absence of a statistically significant performance advantage for machine learning over logistic regression indicates a lack of demonstrated superiority rather than definitive 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.
Subject of Research: Prediction of 30-day cerebral infarction following diffusion-weighted imaging-negative transient ischemic attack using machine learning
Article Title: Machine learning-based models for predicting 30-day cerebral infarction after transient ischemic attack: a retrospective study
Article References: Wu, J., Zhang, D., Liu, X., Sun, K., Jin, J., & Wang, Z. (2026). Machine learning-based models for predicting 30-day cerebral infarction after transient ischemic attack: a retrospective study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-03004-9
Image Credits: AI Generated
DOI: 10.1186/s12880-026-03004-9
Keywords: transient ischemic attack, cerebral infarction, stroke prediction, machine learning, diffusion-weighted imaging, random forest, logistic regression, LASSO regression, Machine, learning-based, models, predicting
News Source: Teresa Odom. (November 10, 2026). Machine Learning Shows Moderate Discrimination Predicting 30-Day Cerebral Infarction After TIA. Scienmag.



