Estimating overall survival of glioblastoma patients using clinical variables, tumor size, and location

Accurate prognosis of glioblastoma is crucial for better-informed treatment decisions, potentially leading to improved disease management. We investigated whether clinical variables, tumor size, and location, can serve as prognostic factors.

A retrospective, multicenter study enrolled 1318 adult patients with histopathologically confirmed glioblastoma undergoing first-time surgery, with survival censored for 188 patients. Pre-operative brain MRIs were used to compute tumor size and derive advanced radiological features describing tumor location, later refined by expert-based opinion. Post-operative MRIs were used to measure the enhancing residual tumor volume. The prognostic quality of all variables, measurements, and features was assessed as inputs of three survival regression models (CoxPH, Random Survival Forests, DeepSurv) to predict overall survival, under five timepoints of patient treatment: onset presentation, assessment by multidisciplinary board, intervention planning, post-intervention evaluation, and chemoradiotherapy planning. Model evaluation was performed with the C-index, Brier Score over Time, and Integrated Brier Score.

Multivariable Cox analysis identified most clinical variables and tumor size as strong predictors of patient survival, with varying hazard ratios across timepoints. DeepSurv was consistently the top performing model under all possible inputs and at all timepoints, yielding mean test C-index scores ranging from 61.71% to 70.29%, and mean Integrated Brier Scores ranging from 8.57% to 7.63%.

Clinical variables, tumor size, and location carry prognostic value for the overall survival of patients with glioblastoma. The best predictive performance was observed under a Deep Survival model using all variables at the stage of chemoradiotherapy planning.