IMG-10. GUARD: A glioma unified AI-driven radiotherapy decision system

Abstract This study aimed to develop and validate a machine learning-based system that non-invasively predicts radiotherapy (RT) sensitivity in patients with low-grade glioma (LGG) prior to treatment, providing a tool for personalized RT recommendations. Our study collected longitudinal MRI data from six institutions at four time points: preoperative, postoperative, pre-RT, and post-RT. A retrospective training cohort of 963 patients (11,556 MRIs) was selected from 13,690 reviewed LGG cases at Beijing Tiantan Hospital. An external test set of 147 patients (1,764 MRIs), screened from a total of 4,715 LGG cases, was collected from five independent centers: Shandong Provincial Hospital; Harbin Medical University Cancer Hospital; Sixth Medical Centre, General Hospital of the People’s Liberation Army; Affiliated Hospital of Hebei University; and China-Japan Union Hospital of Jilin University. Additionally, a prospective test cohort of 328 patients (3,936 MRIs) was enrolled from a registered clinical trial (NCT06454097). The GUARD model was trained on radiomic features extracted from preoperative scans using manually segmented tumor regions together with important clinical variables. Its performance was comprehensively assessed across all datasets, achieving AUCs of 0.948, 0.911, 0.892, and 0.915 in the training, internal validation, external, and prospective datasets, respectively. SHAP analysis revealed that residual tumor, patient age, and radiomics features such as T2-weighted texture roughness and T1-weighted contrast enhancement changes were important predictors of RT sensitivity. The model-derived Radscore stratified patients into biologically distinct groups in radiogenomics analysis. Specifically, tumors in the low Radscore group, associated with low RT sensitivity, was characterized by significant enrichment in radio-resistant pathways, including oxidative phosphorylation, fatty acid metabolism, hypoxia, and DNA repair. Additionally, patients in this group exhibited significantly worse overall survival. These findings suggest that the GUARD model offers accurate and interpretable predictions of RT sensitivity and may serve as a clinically valuable tool to guide personalized RT recommendations for LGG patients.