Co-Founder & Chief Medical Officer Omniscient Neurotechnology Sydney , New South Wales, Australia
Introduction: Clinical trials for novel glioblastoma (GBM) therapies have a high failure rate, and many of these failures may be attributed to methodological factors. GBM trials need to utilize more informative covariates, which may overcome methodological limitations and lead to better hypothesis testing. To that end, we explore the effects of connectomics on survival in GBM.
Methods: Presurgical T1 and diffusion tractography images, along with demographics, and molecular markers were obtained for 289 patients from the UPenn-GBM dataset. Several radiomic and structural connectivity features were extracted from the scans. These were then used as feature sets in several machine learning models to model survival using regression and extreme gradient boosted trees. We also performed power analysis to ascertain the number of subjects which would be required to demonstrate differences in prospective trials.
Results: Models based on structural connectivity generally performed better than models relying on other factors. For example, the model utilizing PageRank centrality showed superior performance (training R2=0.84, test R2=0.08) than models trained on demographic features and molecular markers like MGMT and IDH-1 (training R2=0.11, test R2=0), radiomic features (training R2=0.46, test R2=0), or fractional anisotropy profiles of white matter tract bundles (training R2=0.57, test R2=0.10). The low test performance was suggestive of overfitting due to having an underpowered sample. Our power analysis confirmed this, suggesting that as many as 2000 patents may be necessary to predict overall survival. Finally, we trained several models predicting survival by stratifying outcomes into three survival categories, and demonstrated that using the structural connectome may be a suitable marker to predict stratified outcomes to traverse the logistical barriers in conducting sufficiently powered studies (test AUC = 0.7).
Conclusion : This stunning finding suggests that previous GBM trials which did not control for structural connectivity were likely underpowered around 10 fold. The implications for future trial design are important.
