Deep Learning for Survival Prediction in Glioblastoma: Time-dependent Model Interpretability Using MRI, Clinical, and Molecular Data.
Authors
Affiliations (9)
Affiliations (9)
- Interdisciplinary Programs in Cancer Biology, Seoul National University Graduate School, Seoul, Republic of Korea.
- Department of Radiology, Seoul National University Hospital, 101 Daehak-ro, Jongno-gu, Seoul 110-744, Republic of Korea.
- Department of Biomedical Sciences, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Department of Radiology, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, South Korea.
- Department of Pathology, Seoul National University Hospital, Seoul, Republic of Korea.
- Department of Neurosurgery, Seoul National University Hospital, Seoul, Republic of Korea.
- Department of Radiology, Seoul National University College of Medicine, Seoul, Korea.
- Department of Radiology, Research Institute of Radiological Science, Center for Clinical Imaging Data Science, Yonsei University College of Medicine, Seoul, Republic of Korea.
- Healthcare AI Research Institute, Seoul National University Hospital, Seoul, Korea.
Abstract
Purpose To develop a multimodal model for survival prediction and time-dependent model interpretability in glioblastoma by integrating preoperative MRI with clinical and molecular variables. Materials and Methods This retrospective multicenter study included two institutional cohorts (February 2007-December 2024) and two public external test sets. A deep learning-based prognostic index (DPI) was generated from preoperative multiparametric MRI using a Vision Transformer. The DPI was integrated with clinical variables (age, sex, Karnofsky performance status [KPS], extent of resection [EOR]), molecular markers (<i>IDH</i> mutation, <i>MGMT</i> promoter methylation), histopathology, and WHO grade using a random survival forest model. Model performance was evaluated using the concordance index (C-index), and time-dependent model interpretability was assessed using Survival SHapley Additive Explanations (SurvSHAP(t)). Associations between DPI and clinical and molecular variables were evaluated using correlation and group-wise statistical tests. Results A total of 1,883 patients (mean age, 57.7 ± 14.8 [SD] years; 983 female) were included. The multimodal model integrating MRI and clinical and molecular variables achieved C-indexes of 0.77, 0.73, and 0.63 for the internal test set and two external test sets, respectively. In comparison, the image-only model achieved C-indexes of 0.73, 0.65, and 0.60 across the same cohorts. SurvSHAP(t) analysis showed that prognostic influence peaked at approximately 12 months for EOR and 24 months for <i>MGMT</i> promoter methylation, whereas <i>IDH</i> mutation and WHO grade increased in importance over time. The imaging-derived DPI consistently ranked among the strongest predictors of survival and showed moderate correlations with age, KPS, <i>IDH</i> mutation status, and WHO grade. Conclusion The multimodal model showed good performance for glioblastoma survival prediction and enabled time-dependent model interpretability, identifying the imaging-derived prognostic index as a complementary biomarker with sustained prognostic importance over time. ©RSNA, 2026.