Machine Learning-Based Radiomic Features for Glioblastoma Overall Survival Prediction

This study aims to prognosticate the overall survival (OS) of patients afflicted with gliomas undergoing treatment. The predictive models are constructed through diverse combinations of feature selection techniques and machine learning algorithms. MRI scans (T1 and T2 FLAIR) obtained from pre-treatment, one week post-treatment, and two months later yielded 112 features. Delta radiomics, derived from pre and post-treatment features are used for feature selection using Mutual information, Chi-squared, and F-test. Six machine learning methods (RF, Logistic Regression, KSVM, MLP, Gradient Boosting, XGBoost) were used to build a classification model predicting Overall Survival (OS). F-test outperformed, yielding the highest AUC, ACC, TPR, and TNR in conjunction with Gradient Boosting.