Abstract 2791: Deep learning-based survival prediction of post-operative GBM patients via multimodal radiomic phenotypes drawn from Kaniadakis vector embedding in latent space.

Abstract Purpose: The aim of this study was to use multimodal imaging biomarkers and AI-guided personalized precision survival prediction in improving outcomes for patients with post-operative glioblastoma (GBM). Method: A cohort of 450 de novo GBM cases from the University of Pennsylvania, each featuring the four standard multimodal MRI sequences (T1, T2, FLAIR, T1-Gadolinium), was analyzed. We extracted 354 radiomic biomarkers per modality (Pyradiomics) and reduced these high-dimensional features to a stable 11-dimensional latent space using Density-Based Isomap (PR-Isomap), with the optimal dimension determined by a gap statistic, which we developed specifically for such projection. These latent biomarkers were then normalized, fused across modalities, and projected into a risk-sensitive feature space via Kappa (Kaniadakis) Vector Embedding. The resulting multimodal features fed into a deep learning-based survival model using the log-hazard ratio, optimized via a survival loss function based on the negative log-likelihood of the Cox proportional hazards model for patient outcome prediction. Results: The final survival-prediction model achieved a 71.7% accuracy in binary classification via a consensus of multiple machine learning classifiers with 10-fold cross-validation and a C-index of 0.5196, indicating a modest but statistically meaningful predictive capability using the imaging embeddings alone. To assess prognostic separation, a Kaplan-Meier survivorship curve was stratified by hazard-score quartiles. Distinct thresholds were established: the first quartile Q1) hazard score was -0.36 and the third quartile (Q3) score was -0.29. The median hazard score of -0.33 separated the cohort into 225 high-hazard patients and 225 low-hazard patients. The resulting stratified Kaplan-Meier curves demonstrated clear separation between these risk groups, definitively confirming the capacity of the MRI-derived embeddings to differentiate patient survival trajectories. Conclusion: The integration of deep learning with multimodal MRI embeddings offers a powerful, objective methodology for GBM prognostication. The resulting quantitative hazard scores successfully differentiate high- and low-risk patient survival trajectories. This multimodal imaging biomarker-driven approach provides personalized risk information for post-operative GBM patients, and supporting precision neuro-oncology by informing the selection of personalized adjuvant therapies. Citation Format: Roy Nasr, Bardia Rodd, . Deep learning-based survival prediction of post-operative GBM patients via multimodal radiomic phenotypes drawn from Kaniadakis vector embedding in latent space [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2791.