135 Imaging Patterns Predict Patient Survival and Molecular Subtype in Glioblastoma Using Machine Learning Techniques

Abstract INTRODUCTION: Several studies have examined correlates between imaging features of neoplasm and patient survival or tumor genetic composition; however, few have generated predictive models robust enough to enter clinical practice. In this study, we use advanced pattern analysis and machine learning to identify a combination of imaging features on initial magnetic resonance (MR) images to predict overall survival and molecular subtype in patients with glioblastoma (GB). METHODS: We performed a retrospective followed by a prospective cohort study of GB patients. Imaging features were extracted from structural, diffusion, and perfusion MR images at time of diagnosis. A machine-learning algorithm was used to examine multiple features simultaneously to determine which set of features was most predictive of survival. The model was tested prospectively in a separate cohort of patients. In a subset of patients for which genetic data were obtained, machine learning was used to classify the likelihood of molecular subtype affiliation based on imaging. Tenfold cross-validation was performed. RESULTS: The accuracy of the model in predicting survival was 77% in the retrospective study (n = 105) and 79% in the prospective study (n = 29). Constellations of imaging markers related to infiltration and diffusion of tumor cells into edema, microvascularity, and blood-brain barrier compromise were predictive of shortened survival. A separate model was generated to predict molecular subtype. The accuracy of individual subtype predictions was 85% for classical (n = 20), 84% for mesenchymal (n = 28), 88% for neural (n = 29), and 86% for proneural (n = 22). CONCLUSION: Unlike prior studies, we analyzed the entirety of imaging data in an integrative fashion, leveraging the power of pattern analysis and machine learning to predict survival and molecular subtype with high accuracy and reproducibility in GB. Our noninvasive model utilizes multiparametric imaging obtained routinely for GB patients, making it readily translatable to the clinic.