A Multi-view Deep Survival Combined Model for Predicting Stroke Recurrence in Symptomatic Intracranial Atherosclerosis.
Authors
Affiliations (4)
Affiliations (4)
- Department of MRI Center, The First Affiliated Hospital of Henan University of Chinese Medicine, Zhengzhou, China (Z.L., T.H., L.Z., X.Z., H.L., Q.X., G.Z., Y.G.).
- The Second Affiliated Hospital of Zhengzhou University, Zhengzhou, China (Q.Z.).
- Department of Radiology Center, The First Affiliated Hospital of Xinxiang Medical University, Xin Xiang, China (L.H.).
- Department of MRI Center, The First Affiliated Hospital of Henan University of Chinese Medicine, Zhengzhou, China (Z.L., T.H., L.Z., X.Z., H.L., Q.X., G.Z., Y.G.). Electronic address: [email protected].
Abstract
Symptomatic intracranial atherosclerotic stenosis (sICAS) is associated with a high risk of stroke recurrence. Current risk stratification approaches based on high-resolution vessel wall imaging (HR-VWI) remain dependent on subjective human assessment, which limits their precision. From June 2020 to December 2024, HR-VWI images were retrospectively collected from 363 patients with sICAS across 2 medical institutions. Among the 363 patients, there were 79 cases of stroke recurrence (21.76%) and 284 cases without recurrence (78.24%). The cohort was divided into a Training/Validation set (n=290) and a Test set (n=73). Using the T1-weighted contrast-enhanced sequence, we developed a Multi-View Deep Survival Combined Model. This model employs a Vision Transformer and radiomics to analyze MR images and uses DeepSurv to extend the modeling capacity of the traditional Cox proportional hazards model. It enables the prediction of stroke recurrence risk related to intracranial culprit plaques. Model performance was evaluated using time-dependent receiver operating characteristic curves and the C-index. Additionally, decision curve analysis and calibration curves were used to comprehensively validate the model's practical value and clinical application prospects. The Combined Model exhibited superior predictive performance, achieving a C-index of 0.872 (95% CI: 0.785-0.958) in the internal validation set and 0.803 (95% CI: 0.711-0.895) in the external test set. This performance significantly outperformed that of clinical models, radiomics models, and standalone deep learning models. The model also demonstrated excellent time-dependent predictive accuracy for 1, 2, and 3-year recurrence (area under the curve: 0.841, 0.870, and 0.802, respectively). Calibration and decision curve analysis confirmed the model's clinical utility. By integrating automated multi-view deep feature learning and DeepSurv-based survival analysis, the Combined Model provides a robust and objective tool for stratifying recurrence risk in patients with sICAS. It outperforms conventional methods and holds significant potential for guiding personalized secondary prevention strategies.