Comparison of predicted survival curves and personalized prognosis among cox regression and machine learning approaches in glioblastoma

Abstract

Background: Glioblastoma is an aggressive primary brain tumor with a poor prognosis. At present, time-to-event machine learning (ML) approaches have been used for prognostication in neuro-oncology. The present study aimed to compare the predictive performances among Cox hazard regression, parametric survival regression, and time-to-event ML algorithms. In addition, the secondary objective was to deploy a personalized survival curve for each patient’s condition. Methods: A retrospective cohort study was conducted on glioblastoma patients admitted between December 2007 and June 2021 in a tertiary center in Southern Thailand. Various clinical, radiological, and therapeutic characteristics were collected, and variables related to prognosis were analyzed using a backward stepwise technique. Therefore, a 70:30 data split was performed for the training model and testing performances among the Cox hazard regression, parametric survival models, and various time-to-event ML approaches. Time-to-event performance metrics were used for predicting main outcomes such as Harrell’s concordance index (C-index) and root mean square error (RMSE). Results: There were 208 glioblastoma patients in this cohort, and three variables were used for developing the predictive model using various time-to-event approaches. The multilayer perceptron had the highest value of Harrell’s C-index, which was 0.659 [95% confidence interval (CI): 0.657–0.661], while Cox regression had a C-index of 0.648 (95% CI: 0.642–0.653). The random survival forest model had the lowest RMSE of 0.980 (95% CI: 0.979–0.981) for the estimated number of patients at risk over time, while Cox regression had RMSE of 1.006 (95% CI: 1.005–1.007). The personalized prognosis by the Kaplan-Meier curves could demonstrate the prognosis of the patients in each condition for the recommendation for personal treatment. Conclusions: Time-to-event survival approaches were designed to show the personalized survival curves in each condition for a physician to make a personal treatment recommendation. Therefore, choosing patients with a favorable prognosis would lead to cost-effectiveness management for high-cost standard treatment.