Multimodal diffusion on low-rank approximation radiomic phenotypes with deep learning survival model for prediction of survival in post-operative GBM patients.

Abstract

e14005 Background: Glioblastoma (GBM) exhibits substantial biological and imaging heterogeneity, resulting in variable postoperative outcomes that are inadequately captured by conventional prognostic factors. Although multimodal radiomics enables quantitative tumor characterization, its high dimensionality limits clinical robustness and translation. This study aimed to develop a clinically translatable multimodal diffusion model for personalized survival prediction in patients with postoperative GBM. Methods: A retrospective cohort of 450 patients with de novo glioblastoma from the University of Pennsylvania, each with standard preoperative multimodal MRI (T1, T2, FLAIR, and contrast-enhanced T1), was analyzed. From each modality, 354 radiomic features were extracted and subsequently compressed into a stable 11-dimensional latent representation using density-based Isomap (PR-Isomap), with dimensionality determined by an isometry-driven gap statistic. These low-rank radiomic phenotypes were integrated through a multimodal diffusion framework to generate clinically actionable imaging embeddings. Survival risk was modeled using a deep survival model, optimizing the log-hazard function via a negative log-likelihood loss for individualized outcome prediction. Results: The optimized survival modeling framework demonstrated reliable prognostic performance, yielding a binary classification accuracy of 72.7% through an ensemble consensus of machine learning classifiers evaluated under 10-fold cross-validation, alongside a concordance index of 0.60. Using MRI-derived latent embeddings (hidden dimension = 128; learning rate = 1×10⁻⁴; dropout = 0.2), risk stratification based on hazard-score quartiles produced well-separated Kaplan–Meier survival curves, providing clear evidence that the learned imaging representations capture clinically meaningful differences in patient survival trajectories. In the final model, risk stratification based on hazard-score quartiles identified well-balanced high- and low-risk cohorts (n = 113 each), with clear separation between patients below the first quartile (Q1 = 0.0812) and above the third quartile (Q3 = 0.0820), further supporting the robustness of the proposed survival risk grouping. Conclusions: The integration of deep learning–driven multimodal diffusion modeling with radiomic embedding approximation establishes a clinically robust and objective framework for prognostic stratification in GBM. By capturing complementary imaging-derived phenotypes, this approach enables individualized post-operative risk assessment and supports precision neuro-oncology through data-informed optimization of adjuvant therapeutic strategies.