BOLD complexity characterizes glioblastoma survival via voxel-wise and localized sample entropy.

PURPOSE: Glioblastoma (GBM) is the most prevalent and lethal primary brain tumor. Non-invasive presurgical biomarkers are urgently needed to predict patients’ overall survival (OS). Here we demonstrated a nuanced prognostic tool using sample entropy to assess Blood-Oxygen-Level-Dependent (BOLD) complexity and predict survival outcome, which is computationally efficient, reproducible, robust to noise, and readily transferable across cohorts. METHODS: Resting-state fMRI from 205 treatment-naïve GBM patients and 1148 cognitively stable healthy controls were evaluated. Sample entropy (SampEn), a complexity metric, was evaluated in relation to OS at four levels: whole brain voxel-wise, 15 resting state networks (RSNs), a 64-feature autoencoded latent space, and complexity dynamics along contrast-enhancing (CE) boundary. RESULTS: GBM patients showed a significant reduction in global SampEn versus controls (p < 0.001). Among RSNs, medial temporal lobe (MTL) and basal ganglia (BGA) SampEn correlated inversely with OS (R² = 0.033 and 0.034; p = 0.008 and 0.006). The latent-space-dependent Cox risk score stratifies patients into high and low survival populations (p < 0.001). The number of SampEn peaks at the CE boundary also correlated negatively with OS (R² = 0.020, p = 0.037). CONCLUSIONS: Voxel-wise SampEn revealed widespread loss of BOLD complexity in GBM. It identifies influences at RSNs and tumor-edge, characterizing survival. Latent space analysis revealed whole-brain SampEn characteristics, which provide a compact, data-driven biomarker that augments conventional Cox modelling and stratifies the patient survival. These findings show fMRI-derived SampEn measures are efficient and robust for risk stratification and mechanistic insight in glioblastoma.