NIMG-13. ROBUSTNESS OF PROGNOSTIC STRATIFICATION IN DE NOVO GLIOBLASTOMA PATIENTS ACROSS 22 GEOGRAPHICALLY DISTINCT INSTITUTIONS: INSIGHTS FROM THE RESPOND CONSORTIUM

Abstract

Abstract PURPOSE Glioblastoma is the most prevalent primary malignant brain tumor in adults, with a median overall survival (OS) of approximately 15 months and only limited advancements in prognostication and survival prediction. This study aims to evaluate an AI-based prognostic stratification model for OS prediction trained on the ReSPOND consortium data and to validate its performance on an independent dataset. METHODS The AI model was trained on a cohort of 2,293 glioblastoma patients from 22 institutions across three continents. For validation, an independent cohort of 78 treatment-naïve patients was used from three institutions. Preoperative structural MRI scans were utilized for feature extraction. Automated segmentation defined three tumor sub-compartments: enhancing, necrotic, and peritumoral T2-FLAIR abnormality. The AI predictor incorporated variables such as patient age, normalized tumor sub-compartment volume, spatial distribution characteristics, and morphologic descriptors. The overall survival predictor index provided a continuous value between 0 and 1 for patient stratification that higher values indicating longer predicted OS. Generalizability was assessed using Leave-One-Cohort-Out-Cross-Validation (LOCOCV) for training data, and the model was subsequently applied to the validation cohort. RESULTS Survival analysis demonstrated a concordance index of 0.64 for LOCOCV training data and 0.59 for the independent validation data, indicating effective prognostic stratification of patients. CONCLUSION Multi-parametric AI assisted image analysis extracts prognostic biomarkers, which correlate with OS in glioblastoma patients. The generalizability of this method was validated using the extensive centralized glioblastoma imaging dataset registry from the ReSPOND consortium and an independent dataset, demonstrating its generalizability across diverse patient populations and acquisition settings. This model holds promise for robust prognostic stratification and prediction in de novo glioblastoma patients.