Abstract
MGMT promoter methylation is a critical predictive biomarker in high-grade gliomas (HGG), but its assessment currently relies on invasive tissue sampling. We aimed to develop an explainable, modality-adaptive, and calibrated radiomics model for non-invasive prediction of MGMT promoter methylation using multi-center MRI data.
Pre-operative MRI from the UCSF-PDGM and UPENN-GBM cohorts was analyzed using radiomics extracted from intratumoral and peritumoral regions. Conventional (T1, T2, FLAIR) and advanced (DWI/ADC, ASL) MRI sequences were included. A novel modality-adaptive framework was implemented, allowing the model to automatically ignore advanced modalities when unavailable. After feature ranking and redundancy reduction, six machine-learning classifiers were optimized and Platt-calibrated. Model performance was evaluated on a held-out test set using ROC metrics, calibration assessment, and Decision Curve Analysis (DCA). Feature contributions were interpreted using SHAP.
The top-performing LightGBM model, trained on the 500 most important radiomic features, achieved an AUC of 0.67, recall of 0.90, and accuracy of 0.72 on the independent test set. The model demonstrated strong sensitivity for identifying methylated tumors, minimizing false-negative predictions. Calibration improved clinical net benefit across a range of threshold probabilities on DCA. Feature attribution analysis revealed balanced contributions from conventional and advanced MRI modalities, with texture and intensity-based descriptors being most influential. Notably, low-frequency FLAIR wavelet intensity features were associated with unmethylated tumors.
This explainable, modality-adaptive radiomics model identified biologically consistent MGMT-related imaging patterns and demonstrated decision-analytic value for clinical risk stratification. The framework supports real-world applicability in heterogeneous imaging environments. Future work should focus on external validation and integration with clinical and molecular biomarkers to further enhance predictive performance.