Estimating Progression-Free Survival in Patients with Primary High-Grade Glioma Using Machine Learning

Background/Objectives: High-grade gliomas are the most common primary malignant brain tumors in adults. These neoplasms remain predominantly incurable due to the genetic diversity within each tumor, leading to varied responses to specific drug therapies. With the advent of new targeted and immune therapies, which have demonstrated promising outcomes in clinical trials, there is a growing need for image-based techniques to enable early prediction of treatment response. This study aimed to evaluate the potential of radiomics and artificial intelligence implementation in predicting progression-free survival (PFS) in patients with highest-grade glioma (CNS WHO 4) undergoing a standard treatment plan. Methods: In this retrospective study, prediction models were developed in a cohort of 51 patients with pathologically confirmed highest-grade glioma (CNS WHO 4) from the authors' institution and the repository of the Cancer Imaging Archive (TCIA). Only patients with confirmed recurrence after complete tumor resection with adjuvant radiotherapy and chemotherapy with temozolomide were included. For each patient, 109 radiomic features of the tumor were obtained from a preoperative magnetic resonance imaging (MRI) examination. Four clinical features were added manually-sex, weight, age at the time of diagnosis, and the lobe of the brain where the tumor was located. The data label was the time to recurrence, which was determined based on follow-up MRI scans. Artificial intelligence algorithms were built to predict PFS in the training set (n = 75%) and then validate it in the test set (n = 25%). The performance of each model in both the training and test datasets was assessed using mean absolute percentage error (MAPE). Results: In the test set, the random forest model showed the highest predictive performance with 1-MAPE = 92.27% and a C-index of 0.9544. The decision tree, gradient booster, and artificial neural network models showed slightly lower effectiveness with 1-MAPE of 88.31%, 80.21%, and 91.29%, respectively. Conclusions: Four of the six models built gave satisfactory results. These results show that artificial intelligence models combined with radiomic features could be useful for predicting the progression-free survival of high-grade glioma patients. This could be beneficial for risk stratification of patients, enhancing the potential for personalized treatment plans and improving overall survival. Further investigation is necessary with an expanded sample size and external multicenter validation.