Machine Learning Prediction of Local Failure in GI Brain Metastases on the Day of Gamma Knife Radiosurgery.

BACKGROUND: Time to local failure after stereotactic radiosurgery (SRS), including Gamma Knife radiosurgery (GKRS), for gastrointestinal (GI) brain metastases is difficult to predict and may vary substantially across tumors and patients. We developed and internally evaluated a survival machine learning framework to estimate tumor-specific time to local failure using pre-SRS features. METHODS: We performed a retrospective study of GI brain metastases treated with GKRS. Local failure was modeled as a time-to-event outcome under right censoring. A Random Survival Forest survival model was trained using treatment-time demographic, clinical, tumor, histologic, and radiosurgical variables restricted to information available before or at GKRS. To avoid within-patient leakage, validation was performed using patient-grouped cross-validation. Model performance was assessed using the concordance index (C-index) and IBS. Among tumors with observed local failure, temporal prediction accuracy was additionally evaluated using mean absolute error (MAE) and coefficient of determination (R²). RESULTS: The cohort included 90 patients and 218 treated tumors from 2014 to 2024. Median age at treatment was 68.0 years, median pre-treatment Karnofsky Performance Status was 60, and 91.3% of tumors occurred in patients with multiple metastases. Fifty-seven tumors experienced local failure. The model achieved a mean grouped cross-validated C-index of 0.79 and a mean IBS of 0.0031 (95% CI, 0.0014-0.0042). Among tumors with local failure, conditional MAE was 1.11 months and R² was 0.26. CONCLUSIONS: Pre-SRS survival machine learning is feasible for predicting time to local failure in GI brain metastases treated with GKRS and supports further validation in larger multicenter cohorts.