Overall Survival Predictions of GBM Patients Using Radiomics: An Explainable AI Approach Using SHAP

Glioblastoma multiforme (GBM) is a WHO grade IV tumor and its heterogeneity pushes oncologists to focus on more personalized treatments for individual patients. This challenge is aided by radiomics, which involves the extraction of valuable features from Magnetic Resonance images, the most common non invasive method to identify Glioblastoma multiforme. Analyzing radiomic features through machine learning has been a robust way to predict overall survival (OS) days of GBM patients which significantly affects the personalized treatments and wellbeing.Despite the promise of machine learning for predicting GBM patient overall survival (OS), the limited interpretability of complex models hinders their clinical adoption and the translation of valuable insights gained from radiomic features into actionable treatment plans. Therefore, this study is focused on designing a highly accurate Random Forest (RF) model to predict the OS days of GBM patients from radiomic features and utilize SHapely Additive exPlanation (SHAP) values to provide a comprehensive interpritable analysis of the model, enabling clinicians to understand the role of radiomic features in patient survival prediction. The designed RF model achieved a validation accuracy of 62.5% surpassing previous work done on similar studies. The interpritable analysis revealed critical insights for personalized treatment, including the identification of outlier patients with unique radiomic features, risk inflection points for key features that may guide treatment decisions and pairwise interactions of radiomic features that may be novel biomarkers that affect the OS of GBM patients. This study's findings pave the way for the development of more clinically-usable machine learning models for personalized GBM treatment planning, ultimately improving patient outcomes.