MRI-based radiomics model for predicting symptomatic hemorrhage in cerebral cavernous malformations: a two-center study

Abstract

Predicting hemorrhage in cerebral cavernous malformations (CCM) supports clinical decisions by preventing the risk of fatal or disabling outcomes. This study developed models to predict symptomatic hemorrhage using radiomics features. A total of 153 non-hemorrhagic lesions from the TOUCH cohort (NCT03467295) were analyzed for model development and 54 lesions from the CRESS cohort (NCT04076449) for external validation, with lesions classified as hemorrhagic or stable based on 3-year follow-up. Clinical variables were screened via univariate analysis, and radiomics features from T1- and T2-weighted MRI were selected through LASSO regression. Multivariate logistic regression was used to build three models: clinical, radiomics, and an integrated model combining both. The integrated model demonstrated superior predictive performance, achieving AUCs of 0.860 (training), 0.976 (internal validation), and 0.980 (external validation). Decision curve and calibration analyses confirmed its clinical utility and robustness. This integrated model effectively stratifies hemorrhage risk in this study and shows potential as a supporting tool to inform clinical decision-making in CCM, pending further prospective validation.