A Digital Twin Framework for Glioblastoma Survival Prediction Using Explainable Machine Learning and Multimodal MRI Radiomic Features

Abstract

Glioblastoma multiforme (GBM)is one of the most common and lethal forms of brain tumor, with a median overall survival time of being only 12-15 months.GBM is highly heterogeneous meaning that there is considerable variation in tumor characteristics, behavior, and treatment outcomes across patients. This leads to differences in survival times and treatment outcomes among individual patients. Accurate survival stratification at diagnosis is therefore critical to personalizing therapeutic strategies. This paper presents a Digital Twin (DT) framework for GBM survival prediction that integrates multimodal radiomic biomarkers extracted from the BraTS2020 dataset with an XGBoost classifier and SHAP. Sixteen features spanning volumetric, intensity, and GLCM texture domains are extracted from FLAIR, T1-contrast, and segmentation MRI modalities, all constrained to tumor voxels. Class imbalance was addressed using the SMOTE applied exclusively to training data. The XGBoost DT achieved a hold-out accuracy of 63.83%, an AUC-ROC of 0.6725 (95% CI: 0.507–0.812), and an F1 score of 0.738. A complementary Cox Proportional Hazards model yielded a test-set concordance index of 0.6481, supporting the classifier’s findings under a time-to-event framing. Seven clinically grounded scenarios were simulated, and a Monte Carlo sensitivity analysis confirmed that scenario-level survival probability estimates are robust to plausible variation in the simulated perturbations. This framework provides a foundation for patient-specific prognostic simulation and supports the integration of AI-driven digital twins into neuro-oncological clinical workflows.