Deep learning-based survival prediction model for adult diffuse low-grade glioma: a multi-cohort validation study.

Abstract

Background: Accurate prognostication in adult diffuse low-grade glioma (DLGG) patients remains challenging due to the complex interplay of clinical and molecular factors. This study aimed to develop and validate a deep learning-based model for predicting survival in DLGG patients.

Methods: We analyzed 1,079 DLGG patients across three cohorts: training (n = 836), internal validation (n = 210), and external validation (n = 33). A deep learning model (DeepSurv) was developed incorporating seven clinicopathological variables. Model performance was assessed using C-index and integrated Brier scores (IBS). Feature importance was evaluated through permutation importance analysis and SHAP values.

Results: The cohorts demonstrated comparable baseline characteristics except for resection extent (P < 0.001). The model achieved robust performance with C-indices of 0.81, 0.76, and 0.87 in the training, internal validation, and external validation cohorts, respectively. Low IBS values (0.03-0.04) confirmed strong predictive accuracy across all cohorts. Age emerged as the strongest prognostic factor, showing non-linear effects particularly pronounced in IDH-wildtype tumors after age 50. IDH mutation status was the second most influential factor, while radiation therapy alone and tumor size showed limited prognostic value.

Conclusion: Our deep learning model demonstrates reliable prognostic capabilities for DLGG patients, with age and IDH status as key determinants of survival. The model has been implemented as a web-based platform ( https://seerlggs-f4nze5jr7iuu9k9uuaemjf.streamlit.app ) for clinical use, offering personalized survival predictions. These findings contribute to more precise prognostication and may aid in treatment strategy optimization for DLGG patients.