Abstract
OBJECTIVE: This study aims to systematically compare traditional statistical methods and machine learning approaches for survival prediction in adult-type diffuse glioma (ADG), evaluating their performance, interpretability, and clinical applicability. METHODS: Using two public datasets and one private retrospective cohort, four models (Cox Proportional Hazards, Random Survival Forest, Neural Multi Task Logistic Regression, and DeepSurv) were developed and validated. Missing values were handled in three different ways and subjected to sensitivity analysis. RESULTS: In both the internal test cohorts and two external validation cohorts, the neural network-based deep learning model outperformed the random survival forest and Cox proportional hazards model, with the DeepSurv model demonstrating the most robust performance. Key prognostic factors included age, molecular pathology, chemotherapy, extent of resection, extent of disease, and radiotherapy. Sensitivity analysis confirmed model stability. The model has been packaged and made publicly available. CONCLUSION: For retrospective real-world ADG patient data with high heterogeneity and partial missingness, the DeepSurv model demonstrates superior survival prediction performance and stability compared to conventional approaches.