Prospective biopsy-controlled validation of an AI model for predicting glioblastoma infiltration: Results from the SupraGlio trial

BACKGROUND: Glioblastoma recurrence is driven by diffuse microscopic infiltration beyond the contrast-enhancing tumour margin. GlioMap is an open-access AI model predicting voxelwise infiltration and recurrence risk from multiparametric MRI. This prospective study aimed to validate GlioMap's biological accuracy and prognostic relevance through histopathological assessment, transcriptomic profiling, and survival analysis within the SupraGlio trial (NCT05735171). METHODS: Patients with newly diagnosed glioblastoma underwent neuronavigated biopsies targeting AI-predicted high-risk (HRoR) and low-risk of recurrence (LRoR) regions beyond the contrast-enhancing tumour. Histopathological infiltration served as the ground truth, and transcriptomic profiling characterised each region's molecular phenotype. Model performance was evaluated using accuracy and area under the receiver operating characteristic (ROC) curve (AUC). Survival analyses assessed the prognostic value of postoperative HRoR volume. RESULTS: Fifty-eight biopsies from 27 patients were analysed. GlioMap achieved 0.81 accuracy (95% CI, 0.71-0.91) and 0.84 AUC (95% CI, 0.73-0.93) for histologically confirmed infiltration. Transcriptomic analysis of 48 samples from 16 patients revealed progressive upregulation of invasion- and angiogenesis-related genes (CD44, CHI3L1, STAT3, VEGFA) and downregulation of neuronal markers (MBP, GABRA1) from LRoR to HRoR regions and tumour core, confirming a neural-to-mesenchymal gradient. Postoperative HRoR volume >1.6 cm³ predicted shorter overall survival (P = .04) and progression-free survival (P = .008). CONCLUSIONS: To our knowledge, this study provides the first prospective, biopsy-controlled, molecular validation of an AI model for mapping glioblastoma infiltration. By accurately identifying histologically and transcriptionally infiltrated regions, GlioMap offers a biologically grounded imaging biomarker that could guide extended resection and personalised radiotherapy planning, potentially improving tumour control and patient outcomes.