HR-VWI Radiomics for Predicting 1-Year Adverse Events in Symptomatic ICAS: A Comparison of Plaque-Only and Integrated Plaque-Vessel Models.

RATIONALE AND OBJECTIVES: To compare an integrated "Plaque-Vessel" high-resolution vessel wall imaging (HR-VWI) radiomics model against a "Plaque-only" model for predicting 1-year adverse clinical events in symptomatic intracranial atherosclerotic stenosis (ICAS). MATERIALS AND METHODS: This retrospective study enrolled 272 symptomatic ICAS patients. A temporal split was applied: patients from 2022 (n = 133) constituted the training and validation cohorts, while those from 2023 (n = 139) served as an independent temporal test cohort. Radiomic features were extracted from T1-weighted and contrast-enhanced T1-weighted HR-VWI. Two signatures were constructed: a "Plaque" model (plaque ROI only) and a "Plaque-Vessel" model (incorporating plaque, vessel wall, and lumen). The primary endpoint was a composite of 1-year adverse clinical events. Performance was assessed using the area under the curve (AUC) and decision curve analysis (DCA), comparing against a qualitative imaging model (comprising intraplaque hemorrhage, plaque enhancement, and positive remodeling), the Essen Stroke Risk Score (ESRS), and NIHSS. RESULTS: The Plaque-Vessel model yielded the highest predictive performance, achieving an AUC of 0.800 (95% CI: 0.724-0.863) in the temporal test set. This performance showed a trend toward improvement compared to the Plaque-only model (AUC = 0.721, P = .27) and significantly outperformed the qualitative plaque features model (AUC = 0.631, P = .03), ESRS (AUC = 0.554, P < .01), and NIHSS (AUC = 0.543, P < .01). Calibration curves showed good agreement. DCA demonstrated that the Plaque-Vessel signature provided greater net clinical benefit than the Plaque-only model or clinical scores across the majority of threshold probabilities. CONCLUSION: The integrated Plaque-Vessel radiomics model significantly outperforms conventional clinical scores and routine visual plaque assessment for predicting 1-year adverse events. However, given the inconclusive statistical difference between the integrated and plaque-only models, these findings remain exploratory. The model requires further large-scale validation before it can be considered ready to guide routine clinical decisions.