Os04.1.a a Normative Human Glioblastoma-Connectome Constructed From Tumor-Connected Brain Regions Improves Survival Rates Estimation

BACKGROUND Glioblastoma (GBM) is an aggressive brain tumor with a poor prognosis. Recent evidence suggests that GBMs interact functionally and structurally with spatially distant brain regions, shaping behavior, survival, and evolution. Here, we studied the white matter connectivity between the GBM and the brain to understand the topological properties of GBM networks and their relationship to survival rates. METHODS Using survival data from two independent cohorts (N=367 and N=496), we obtained streamlines intersecting each tumor based on normative tractography models and computed the connectivity between brain regions. After binarizing the networks, we obtained the size of the giant component, the average degree and clustering coefficient, the number of hubs, and the modularity. Then, we applied principal component (PC) analysis, T-distributed Stochastic Neighbor Embedding (TSNE), and Uniform Manifold Approximation and Projection (UMAP) to obtain a low-dimensional representation of the network topology. These features were incorporated into Cox and logistic models of patient survival, along with sex, age at diagnosis, extent of resection (EOR), and methylation (MGMT). RESULTS We first fitted Cox and logistic models using sex, age, EOR, and MGMT. All covariates except sex showed significant hazard ratios (p<0.001; Wald’s t-test) and contributed meaningfully to the logistic fit. The Cox models using the TSNE and UMAP components as additional features were more likely than the baseline (p<0.0001, p=0.0024, df=2; log-likelihood ratio tests). In the UMAP model, both embedding dimensions had significant hazard ratios (HR=1.037, p=0.003; HR=0.952, p=0.020; two-sided Wald’s t-tests). In the TSNE model, the results were similar but only the second component was associated with a significant hazard ratio (HR=1.003, p=0.051; HR=0.981, p=0.002; two-sided Wald’s t-tests). These findings were replicated in the accompanying logistic regression models, where both components of the TSNE and UMAP embeddings now significantly contributed to the models, which again, outperformed the standard model (p<0.001, df=2; log-likelihood ratio tests). The first two PCs explained over 85% of the variance and showed borderline significant hazard ratios (p=0.053, p=0.080; Wald’s t-test). Their contribution to logistic regression was significant (p<0.01), and both models outperformed baselines (p<0.05, df=2; Wald’s z-test). CONCLUSIONS We explored a network model of human GBMs using neuroimaging tools to improve prognosis. These preliminary findings support the view of GBM as a distributed white matter tumor, suggesting that this paradigm shift could lead to more precise and potentially personalized prognostic models. FUNDING NCN (UMO-2024/53/N/NZ4/03513), EU Horizon 2020 (857533), FNP (MAB PLUS/2019/13), and Minister of Science and Higher Education (MEiN/2023/DIR/3796).