Computed Tomographic Angiography-Based Radiomics Predicted Stroke Recurrence in Patients With Symptomatic...

The aim of the current study was to investigate the predictive values of computed tomographic angiography-derived radiomics features (RFs) for stroke recurrence in symptomatic intracranial atherosclerotic stenosis.

This was a substudy of a cohort study recruiting patients with ischemic stroke or transient ischemic attack with symptomatic intracranial atherosclerotic stenosis (50%-99% stenosis) confirmed by computed tomographic angiography from 2 centers. Baseline clinical variables were collected. RFs were extracted from computed tomographic angiography source images from the plaque region alone and the plaque and lumen region, respectively. All patients received guideline-recommended medical treatment. The primary outcome was same-territory ischemic stroke recurrence within 1 year. Four machine learning algorithms were used to develop and test predictive models for the primary outcome.

Overall, 203 patients (median age, 60 years) were randomly split into training (n=121, 60%) and testing (n=82, 40%) sets. No significant difference was observed in clinical variables between patients with and without primary outcome. Of 1046 RFs extracted, 4 from the plaque region and 5 from the plaque and lumen region were selected. Among all machine learning algorithms, models based on the plaque and lumen region exhibited better discrimination of the primary outcome than plaque region RFs-based model in the training (area under the curve 0.86-0.87 versus 0.79-0.80) and testing set (area under the curve 0.74-0.81 versus 0.58-0.62), consistent with internal validation result with repeated cross-validation. A similar trend was observed in performance metrics, calibration analysis, and reclassification indices.

In patients with symptomatic intracranial atherosclerotic stenosis, computed tomographic angiography-derived RFs of the plaque and lumen region had better predictive value for stroke recurrence than RFs of the plaque region alone, offering a promising tool for stroke risk stratification.