Explainable Machine Learning Model for Post-SRS Overall Survival Estimation in Renal Cell Carcinoma Brain Metastases Patients
Semiha Ozgul, Melis Parabas, Peng Li, Zuhair Majeed, Yusuf Acikgoz, Sherise Desiree Ferguson, Eric Jonasch, Subha Perni, Thomas H Beckham, Jing Li, Dima Suki, Merve Hasanov, Elshad HasanovAbstract
Background
Brain metastases (BM) from renal cell carcinoma (RCC) significantly affect survival, and disease-specific prognostic tools for patients undergoing stereotactic radiosurgery (SRS) remain limited. We aimed to identify prognostic factors and develop an interpretable machine learning (ML) model for individualized risk stratification following SRS.
Methods
We retrospectively analyzed 169 patients with RCC BM treated with SRS between November 2015 and May 2021 in this retrospective cohort study. Univariate Cox regression were performed. ML analyses used a comprehensive survival modeling framework integrating multiple algorithms and ensemble strategies, with final model selection based on nested cross-validation (CV) performance and interpretation using Shapley Additive Explanations (SHAP). Performance was evaluated using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (AUC), Brier score, and calibration metrics; a web-based calculator was developed.
Results
Median overall survival (OS) was 1.73 years (95% confidence interval (CI), 1.26–2.60). In the test cohort, the selected model elastic net–regularized Cox regression (α = 0.2) achieved a C-index of 0.68 with 95% CI: 0.60-0.76 and a 1-year AUC of 0.74 with 95% CI: 0.61-0.86. Using the training cohort median risk score, patients were stratified into low- and high-risk groups with significantly different OS in both cohorts. SHAP analysis identified pre-intervention Karnofsky Performance Status , BM locations, age at intervention, pre-intervention tumor volume, and number of BM as influential predictors.
Conclusions
We developed an interpretable ML survival model that enables individualized OS prediction and clinically meaningful risk stratification after SRS while reflecting established prognostic factors.