GUARD-LGG: a glioma upfront AI-driven radiotherapy response decipher for lower-grade glioma

Radiotherapy (RT) remains a cornerstone in managing lower-grade glioma (LGG), but its response varies widely. Reliable pre-treatment prediction of RT response could facilitate patient management but remains unavailable. Here we present GUARD-LGG, a machine learning model designed to predict early RT response in LGG before treatment. GUARD-LGG was trained on 963 patients from a major center, tested on 187 patients from seven independent institutions, and 328 patients prospectively enrolled in a registered clinical trial. GUARD-LGG achieved AUCs of 0.946 (95% CI: 0.938–0.962), 0.929 (95% CI: 0.894–0.965), 0.882 (95% CI: 0.830–0.934), and 0.913 (95% CI: 0.878–0.949) in the training, internal validation, external, and prospective test sets, respectively. Model interpretation highlighted residual tumor, age, T2-weighted texture roughness and T1-weighted contrast intensity changes as major predictors. Radiogenomic analyses suggested potential associations between predicted unfavorable RT response and enrichment of RT resistance-related pathways. Consistently, patients with predicted unfavorable RT response exhibited shorter overall survival. In conclusion, GUARD-LGG enables individualized prediction of early RT response before therapy, offering a tool for pre-treatment response stratification in LGG.