Machine Learning-Based Prognosis Prediction in Glioblastoma Multiforme Patients by Integrating Clinical Data with Multimodal Radiomics

Objectives: Glioblastoma multiforme (GBM) is considered the most aggressive primary brain tumor, which often exhibits tumor heterogeneity. Hypoxia is a key aspect of intratumoral heterogeneity that contributes to poor prognosis in GBM. In this study, we aimed to develop machine learning (ML) models using radiomics and clinical features for the prediction of one-year survival for GBM. Methods: Data from 35 patients in the ACRIN 6684 trial, including fluoromisonidazole (FMISO)-positron emission tomography (PET), magnetic resonance (MR) (T1, T2, and fluid-attenuated inversion recovery (FLAIR)) images, and clinical information, were retrieved from The Cancer Imaging Archive (TCIA). Three ML algorithms, namely, support vector machine (SVM), random forest (RF), and linear regression (LR), were utilized to analyze selected features. Receiver-operating characteristic (ROC) curves were utilized to evaluate the predictive performance of the models. Several statistical analyses, namely, the permutation test, the permutation importance of selected features, Fisher's exact test, and the unpaired t-test, were performed to analyze the models and features. Results: FMISO achieved the best performance in radiomics models, with an area under the curve (AUC) of 0.870. The clinical data model achieved the best performance of all models, with an AUC of 0.921, outperforming the combined all sequential forward selection (SFS) model (AUC: 0.862). Female sex (p = 0.030) and younger age (p = 0.0043) were significantly associated with better prognosis. Conclusions: Our proposed models have the potential to predict the one-year survival of GBM and facilitate personalized therapy. Future studies with a larger sample size are needed to confirm the generalizability of the models.