Machine Learning-based Overall Survival Prediction in GBM Patients Using MRI Radiomics

Glioblastoma multiforme (GBM) is regarded as the most prevalent primary tumor of the central nervous system in the brain. However, due to the lack of information, it is too hard to understand the underlying progression patterns and the prognosis of patients. In this study, we evaluated the overall survival predictive (time to event analysis) power of radiomic features extracted from MRI, along with the help of feature selection (FS) and machine learning (ML) algorithms. The MR images of 119 patients and their overall survival status were obtained. The data were randomly split into 70% and 30%, indicating training and testing datasets, respectively. Twelve preprocessing methods (e.g., bin discretization, Laplacian of Gaussian, and wavelet transform), 5 FS (e.g., Boruta, Cindex, Random Survival Forest), and 7 ML (e.g., Glmnet, CoxBoost) algorithms were recruited to form a total of 420 models. The models with C-index as FS method showed more decent results than others. The highest-achieving model (C-index = 0.72) was the combination of LOG sigma 1 mm as the preprocessing, C-index as the feature selector, and Coxph as the ML algorithm. Our findings represent the power of radiomic features to be utilized in the overall survival prediction of patients with glioblastoma and their prognostication in general.