LEARNING REPRESENTATIONS OF BRAIN-GLIOMA CONNECTIVITY TO UNDERSTAND AND INFORM SURVIVAL PROGNOSIS

Glioblastoma (GBM) is increasingly recognized as a network disease, rather than an isolated histopathological entity, that evolves within and perturbs the brain connectome. We investigated whether patient-specific brain-glioma circuitry contains prognostic information beyond standard clinical factors. We analyzed 867 patients (WHO 2021) by filtering a normative tractogram for tumor-intersecting streamlines. Then, we computed patient-specific tract density maps, defined as the number of streamlines traversing each voxel. These maps were flattened, and the first 100 Principal Components (PCs) were extracted to represent circuit heterogeneity. Although three principal components explained over 50% of the variance in white matter-GBM interactions, nearly 50 components were required to reach 90%, indicating highly heterogeneous spatial patterns, with tumors preferentially involving either anterior or posterior ipsilateral and cross-hemispheric pathways but not both (Fig. left). Survival was assessed using regularized Cox models, comparing baseline clinical covariates against models sequentially incorporating PCs. While connectivity features alone showed significant but rather weak prognostic effects (C-index ≈ 0.65; Fig. center), models integrating connectivity with clinical covariates outperformed clinical baselines (C-index 0.72 vs 0.67; Fig. right). Prognostic performance improved monotonically as more PCs were added, indicating that the survival-related signal is distributed across multiple aspects of the brain-glioma circuitry rather than localized to a single pathway or pattern. Ranking PCs by survival association improved performance at low cut-offs but converged when all components were included (Fig. right). These findings motivate the use of representation learning, particularly deep-learning models, to capture the complex, distributed structure of brain-glioma connectivity for survival prediction.