2076 - FLAIR Connectomics Support Prognostic Glioblastoma Architecture through Neural Survival Modeling
Presenter(s)
V. Gopalakrishnan1, V. Parekh2,3, R. Agyekum1, Y. Tarui4, K. J. Redmond1, and L. R. Kleinberg1; 1Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 2Diagnostic and Interventional Imaging, UTHealth Houston, Houston, TX, 3Department of Neurology and Neurosurgery, Johns Hopkins Hospital, Baltimore, MD, 4Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD
Purpose/Objective(s): We evaluated whether pre- and early postoperative FLAIR-derived tumor-region connectomic features in glioblastoma predict overall survival (OS), add prognostic value beyond molecular variables, and improve prediction when integrated with molecular/clinical features in a DeepSurv multimodal neural network. We hypothesized that connectomic signatures capture biologically meaningful infiltrative tumor architecture and provide complementary prognostic signal.
Materials/Methods: In a retrospective cohort (46 patients; 44 in adjusted Cox models), 105 connectomic features were extracted from preop and postop FLAIR-based graph metrics. We built compact connectomic survival signatures (preop and postop) and evaluated discrimination (C-index, 12-month AUC) and Cox associations. We also trained DeepSurv models comparing connectomics-only, clinical/molecular baselines, and multimodal fusion (preop, postop, and combined pre+post feature sets) using cross-validated out-of-fold performance. Molecular/pathway association analyses were also assessed with multiplicity-aware interpretation.
Results: Both preop and postop connectomic signatures showed clinically meaningful OS signal. The preop signature provided the strongest overall discrimination (CV C-index 0.618; CV 12-month AUC 0.636) and was significant in unadjusted Cox analysis (HR 1.14, p = 0.024), with similar direction after adjustment. The postop signature also remained prognostic and was significant in adjusted analysis (HR 1.18, p = 0.020), supporting persistent survival-relevant information in the postoperative FLAIR pattern despite expected post-surgical confounding. DeepSurv showed complementary postoperative signal, with multimodal fusion improving out-of-fold discrimination and calibration relative to both postoperative connectomics-only and clinical/molecular baseline models (including C-index gains of +0.077 and +0.061), whereas preop fusion provided no clear advantage over simpler preop models. While gene/pathway associations did not remain significant after correction, nominal BRCA1/PTEN-related signals support prioritizing pathway-informed validation analyses rather than broad feature-level screening.
Conclusion: FLAIR connectomics demonstrates reproducible prognostic signal for OS in glioblastoma in both preoperative and postoperative settings, with preop features showing stronger discrimination and postop features retaining adjusted prognostic value. These findings support that FLAIR network topology reflects infiltrative organization not fully captured by standard molecular variables. The DeepSurv results suggest that postoperative multimodal fusion may be a useful direction for integrating imaging topology with tumor biology and clinical context, but larger cohorts and external validation are needed to stabilize nonlinear models and better define connectomic-genomic relationships.