Development of a recurrence risk prediction model for intracranial atherosclerotic stroke using HR-VWI combined with radiomics.

Abstract

Objective: Based on high-resolution vessel wall imaging (HR-VWI) and radiomics, this study constructed a predictive model for stroke recurrence in patients with symptomatic intracranial atherosclerotic stenosis (sICAS), and evaluated the performance of the model.

Materials and methods: This study retrospectively included 90 patients with sICAS who were treated in a tertiary hospital in China from January 2020 to December 2024. No a priori power calculation was performed; the sample size was based on the availability of complete HR-VWI and follow-up data during the study period. The imaging features of plaques were measured in T1WI and enhanced T1WI images of all patients, and the regions of interest of plaques were delineated and radiomics features were extracted. All the samples were randomly divided into training set (n = 63) and validation set (n = 27) at a ratio of 7 : 3. To address class imbalance, we employed the Synthetic Minority Over-sampling Technique (SMOTE) within the training set during model development. After feature screening, radiomic score (rad-score) was constructed. Based on the conventional imaging features, radiomics features and clinical related information of vascular plaques, the clinical prediction model, radiomics prediction model and the combined model of the two were established in turn. Receiver operating characteristic (ROC), area under the curve (AUC), calibration curve and decision curve analysis (DCA) were used to comprehensively evaluate the predictive efficacy of various models. At the same time, a nomogram was constructed to visualize the model, providing an intuitive reference for clinical practice.

Results: During the follow-up period, a total of 26 patients (28.9%) had stroke recurrence. Patients in the recurrence group had higher levels of low-density lipoprotein and total cholesterol, lower plaque load, and higher degree of plaque enhancement, and the differences were statistically significant (P = 0.004), and plaque enhancement grade was an independent risk factor (OR = 11.023, P = 0.001). Finally, a total of 14 radiomics features were selected to construct the rad-score scoring system. In the training set and validation set, the AUC corresponding to the combined prediction model reached 0.947 and 0.925, respectively, and its prediction efficiency was significantly better than that of a single clinical model and an independent radiomics model. The nomogram showed good calibration and high clinical net benefit.

Conclusion: The combined model of HR-VWI radiomics features and traditional imaging indicators has excellent performance in the prediction of stroke recurrence in sICAS patients. However, due to the limited sample size (n = 90, with only 26 recurrent events) and the high-dimensional radiomic feature space (n = 1920), Our findings are at a high risk of overfitting. Despite the use of rigorous regularization techniques (LASSO) and internal cross-validation, the results should be interpreted as preliminary and hypothesis-generating. These findings are from a single-center study and urgently require external validation in larger independent cohorts before clinical translation.