Background and Purpose: The purpose of this study was to explore radiomics features and construct models to predict symptomatic hemorrhage in cerebral cavernous malformation (CCM) lesions. Methods: A total of 86 CCM patients (153 unruptured CCM lesions) from the TOUCH cohort were enrolled, along with 25 CCM patients (50 unruptured CCM lesions) from CRESS cohort for external validation. According to 3-year prospective follow-up, the lesions were classified into hemorrhagic group or stable group. Baseline clinical characteristics were filtered using univariate and multivariate analysis. Radiomics features were extracted from magnetic resonance imaging (MRI) sequences (T2-weighted and T1-weighted images) before follow-up, and radiomics variables were screened by Least Absolute Shrinkage and Selection Operator (LASSO) regression. Models were constructed using machine learning- Gaussian Naive Bayes (GNB) algorithm. Model performance was evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Validation was performed through internal validation, external validation and cross-validation. Results: A total of 9 statistically significant variables were identified. Three predictive models were constructed: a clinical model, a radiomics model and an ensemble model (integrating clinical and radiomics characteristics). These models demonstrated robust predictive performance The AUC values of training dataset were 0.743, 0.760 and 0.826, respectively; the AUC values of internal validation dataset were 0.839, 0.736 and 0.922, respectively; the AUC values of external validation dataset were 0.690, 0.783 and 0.717, respectively. Conclusions: This study demonstrated the machine learning-derived ensemble model showed superior predictive performance, with enhanced sensitivity and model efficacy.
Li et al. (Thu,) studied this question.