Radiomics-based models to predict IDH mutation status and prognosis in gliomas using MRI: a multicenter study.

Introduction: Gliomas are infiltrative primary intracranial tumors with marked biological and clinical heterogeneity. Prognosis varies widely and depends on tumor grade, histopathological characteristics, and molecular alterations. Isocitrate dehydrogenase mutation is a key prognostic biomarker and is associated with improved treatment response and longer overall survival. Radiomics enables the extraction of quantitative features from routinely acquired medical images. This study evaluated radiomics-based machine learning models for noninvasive prediction of isocitrate dehydrogenase mutation status and overall survival in glioma patients. Methods: From T2-weighted MRI scans of 638 gliomas (213 from a local institution (discovery), 425 from a public dataset (validation)), 1,820 radiomics features were extracted. Machine learning models were constructed and trained on the discovery cohort and externally validated to predict isocitrate dehydrogenase mutation status. A radiomics risk score was computed using Lasso regression, and patients were stratified into high- and low-risk groups using the median radiomics risk score for Kaplan-Meier analysis. Cox regression assessed the prognostic value of radiomics risk score along with clinical features (age, sex, WHO grade, isocitrate dehydrogenase mutation status). A nomogram incorporating independent predictors to estimate 1-, 2-, and 3-year overall survival was assessed using the concordance index and calibration curves. Results: Logistic regression and random forest classifier models achieved area under the receiver operating characteristic curve of 0.90 and 0.68 in the discovery and validation cohorts, respectively, for isocitrate dehydrogenase mutation prediction using 12 top radiomic features. High-risk patients showed significantly shorter median overall survival than low-risk patients in both discovery and validation cohorts (21 vs. 30 months, 10 vs. 19.5 months, respectively; P <0.001). Age, radiomics risk score, and isocitrate dehydrogenase mutation status were significant prognostic factors (P <0.05). The nomogram achieved concordance indices of 0.83 and 0.75 in the discovery and validation cohorts, with good calibration. Conclusion: Radiomics from preoperative T2-weighted MRI enabled prediction of isocitrate dehydrogenase mutation status and overall survival in gliomas. The radiomics risk score was an independent prognostic factor and, combined with clinical variables, enabled personalized risk stratification in gliomas. External validation further confirmed the generalizability of the proposed models.