Multimodal DeepSurv model with SHAP-based interpretation for predicting local failure in patients with brain
Long Jin1, Qifan Zhao2, Ying Xiao1
1Department of Radiation Oncology, Shaanxi Provincial People's Hospital, Xi'an, China.
Objective:
Local Progression-Free Survival (LPFS) is an important clinical endpoint following hypofractionated stereotactic radiotherapy (HSRT) for brain metastases (BMs). However, accurate prediction of local failure risk remains challenging because conventional clinical factors do not fully capture tumor heterogeneity. Therefore, we aimed to develop and validate an explainable radiomics-based deep learning survival model to predict LPFS after HSRT and identify patients at elevated risk of local failure.
Methods:
This retrospective study included 100 patients with BMs treated with HSRT between 2019 and 2023 as training dataset, while utilized medical images and clinical features from 40 patients from publicly available dataset PROTEAS project. Radiomic features were extracted from the contrast-enhancing BM (T1-weighted contrast-enhanced MRI) and edema (T2-FLAIR MRI). A baseline Cox proportional hazards model incorporating clinical variables was established for comparison. Three DeepSurv models were developed using different combinations of clinical, BM-derived, and edema-derived radiomic features. Model performance was evaluated using the concordance index (C-index). Kaplan-Meier analysis was performed for risk stratification, and Shapley Additive Explanations (SHAP) were used to identify the most influential prognostic features.
Results:
The best predictive performance was achieved by the DeepSurv model integrating clinical, BM-derived, and edema-derived radiomic features, outperforming the clinical-only Cox model (C-index: 0.61 vs. 0.72). The model effectively stratified patients into distinct prognostic groups with significantly different local failure-free survival outcomes (P = 0.002). SHAP analysis identified edema Gray Level Co-occurrence Matrix (GLCM) Cluster Prominence as the most important predictor of local failure, followed by edema Large Area High Gray Level Emphasis (LAHGLE), tumor GLCM Cluster Prominence, and tumor LAHGLE. Patients with high edema GLCM Cluster Prominence exhibited significantly worse local failure-free survival than those with low values (P = 0.0155).
Conclusion:
A combination of clinical, tumor-derived, and edema-derived radiomic features predicted Local Progression-Free Survival more accurately than clinical variables alone. Peritumoral edema features contributed more strongly to local failure prediction than tumor features, highlighting the prognostic importance of the tumor microenvironment. Patients received targeted therapy with low-risk radiomic profiles benefit from HSRT for local control.
