The morphometric features derived as the predictors of acute ischemic stroke recurrence.

BACKGROUND AND

OBJECTIVE

Several approaches have been used to estimate the risk of recurrent ischemic stroke, including clinical risk scores, models based on routine clinical data, and diffusion-weighted imaging radiomics. Most of these methods, however, focus on established clinical factors or features of the acute infarct and do not fully reflect structural changes across the whole brain. It remains uncertain whether morphometric measures derived from T1-weighted imaging can improve the prediction of recurrence within 1 year. We therefore evaluated the predictive value of these measures.

METHODS

A total of 201 subjects with acute ischemic stroke (AIS) were assigned to the recurrence and nonrecurrence groups. Cortical morphometric features were extracted from T1-weighted imaging (T1WI) using FreeSurfer. Three feature selection strategies were evaluated: independent-samples t-tests alone, least absolute shrinkage and selection operator (LASSO) regression alone, and independent-samples t-tests followed by LASSO regression. Eight machine learning models were developed using the selected features, resulting in 24 feature selection model combinations. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, precision, F1 score, calibration analysis, and decision curve analysis. Shapley additive explanations (SHAP) were used to assess feature importance and interpret model predictions.

RESULTS

Among the 24 feature selection model combinations, SVM using features selected by independent-samples t-tests followed by LASSO regression achieved the highest AUC, with an AUC of 0.841 (95% CI, 0.786-0.897), accuracy of 0.771, sensitivity of 0.705, specificity of 0.824, precision of 0.792, and F1 score of 0.738. Calibration analysis showed that this model had the lowest Brier score of 0.162, while decision curve analysis demonstrated a favorable net benefit across a broad range of threshold probabilities. SHAP analysis identified the folding index of the left medial orbital-olfactory sulcus and the mean curvature of the left subcallosal gyrus as important contributors to model predictions.

CONCLUSION

This interpretable and reproducible model is a valuable tool for predicting ischemic stroke recurrence, aiding clinical risk stratification and informing individualized treatment decisions.