Evaluating Predictive Performance of Machine Learning Algorithms That Integrate Routine Clinical Variables With Imaging-Derived Information in Stroke Recurrence Risk

Aims/Background: Stroke recurrence remains a significant challenge in post-stroke management, with traditional prediction models often showing limited accuracy. This study aims to compare the performance of multiple machine learning (ML) algorithms that integrate routine clinical variables with imaging-derived features in predicting stroke recurrence risk, and to identify the optimal predictive model. Methods: This retrospective cohort study enrolled 350 patients with ischemic stroke who were admitted to The Fifth People’s Hospital of Jinan between January 2018 and December 2021. Patients were divided into three groups based on the time of first stroke onset: Group A (n = 110), Group B (n = 120), and Group C (n = 120). Routine clinical variables (age, gender, hypertension, and diabetes) and imaging features (infarct size and location) were collected. Four ML-based algorithms—logistic regression, random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost)—were used to construct predictive models. The predictive performance of these models was evaluated by area under the curve (AUC), sensitivity, specificity, and accuracy. Results: The XGBoost model showed the superior predictive performance, achieving the highest AUC of 0.86, followed by the random forest model (0.82), support vector machine model (0.78), and logistic regression model (0.75). The most influential predictors for stroke recurrence were found to be infarct size, history of hypertension, and fasting blood glucose levels. Conclusion: ML-based algorithms that integrate routine clinical variables with imaging-derived data can predict stroke recurrence risk effectively, with the XGBoost model demonstrating superior predictive performance, which may further support more individualized clinical decision-making.