Predicting overall survival in diffuse glioma from the presurgical connectome

Abstract

Abstract Diffuse gliomas are incurable brain tumors, yet there is significant heterogeneity in patient survival. Advanced computational techniques such as radiomics show potential for presurgical prediction of survival and other outcomes from neuroimaging. However, these techniques ignore non-lesioned brain features that could be essential for improving prediction accuracy. Gray matter covariance network (connectome) features were retrospectively identified from the T1-weighted MRIs of 305 adult patients diagnosed with diffuse glioma. These features were entered into a Cox proportional hazards model to predict overall survival with 10-folds cross-validation. The mean time-dependent area under the curve (AUC) of the connectome model was compared with the mean AUCs of clinical and radiomic models using a pairwise t-test with Bonferroni correction. One clinical model included only features that are known presurgery (clinical) and another included an advantaged set of features that are not typically known presurgery (clinical +). The median survival time for all patients was 134.2 months. The connectome model (AUC 0.88 ± 0.01) demonstrated superior performance (P < 0.001, corrected) compared to the clinical (AUC 0.61 ± 0.02), clinical + (AUC 0.79 ± 0.01) and radiomic models (AUC 0.75 ± 0.02). These findings indicate that the connectome is a feasible and reliable early biomarker for predicting survival in patients with diffuse glioma. Connectome and other whole-brain models could be valuable tools for precision medicine by informing patient risk stratification and treatment decision-making.