PATH-69. AI-based prognostic risk stratification of adult-type diffuse gliomas using H&E-stained slides

Abstract BACKGROUND Adult-type diffuse gliomas (ADG) are aggressive brain tumors with highly variable clinical outcomes. Reliable prognostic assessment from clinical routine H&E-stained whole slide images (WSIs) contributes to clinical decision-making and patient management optimization, but remains challenging due to tumor heterogeneity. We present an artificial intelligence (AI) approach for overall survival risk stratification directly from H&E-stained WSIs offering a promising approach to uncover novel prognostic markers. DATA AND METHOD We utilized 1,533 WSIs (735 patients) from the multi-institutional TCGA-GBM and TCGA-LGG collections, under a 10-fold cross-validation configuration with patient-level partitioning. After comprehensive WSI curation, we systematically benchmarked eight pathology-specific foundation models (FMs) and an AI model pretrained on ImageNet, along seven multiple instance learning (MIL) methods, within a weakly supervised framework to predict patient-level risk scores. These scores were used to further stratify patients in high- vs. low-risk groups, and evaluate performance according to subtype (oligodendroglioma, astrocytoma, glioblastoma). RESULTS The UNI FM model with MambaMIL achieved the optimal C-index of 0.76 (95% confidence interval (CI): [0.734, 0.784]). Stratifying patients into high- vs. low-risk groups (p<0.005, log-rank) yielded a highly favorable hazard ratio (HR) of 0.33 (95%CI:[0.26,0.43]). Subtype-specific analysis in astrocytoma, IDH-mutant and glioblastoma, IDH-wildtype showed HR of 0.27(95%CI:[0.12,0.64] and 0.92 (95%CI:[0.76,1.12]), respectively, demonstrating subtype-dependent performance. Key insights include: (i) pathology-specific FMs substantially outperform ImageNet-based model; (ii) advanced spatially informed MILs contribute marginal performance improvements; (iii) no single MIL method consistently improves performance across all FMs, indicating that the optimal MIL choice may be FM-dependent. CONCLUSION Our results, indicating strong concordance between predicted risk and actual outcomes, support the integration of FMs with MIL methods for WSIs-based risk stratification across ADG patients. Our findings underscore the potential of computational pathology complementing existing neuropathological diagnostic assessment towards enhancing patient prognostic stratification, improving clinical decision-making in glioma care, and advancing personalized neuro-oncology.