Predicting time to local failure after gamma knife radiosurgery for melanoma brain metastases using survival machine learning.

BACKGROUND: Melanoma brain metastases treated with Gamma Knife radiosurgery show heterogeneous local control trajectories. Accurate prediction of time to local control loss could support risk stratification and follow-up planning using routinely available clinical and treatment variables. METHODS: A retrospective dataset was constructed from a melanoma brain metastasis radiosurgery cohort. The endpoint was time to local control loss with right censoring for lesions maintaining local control at last follow-up. Clinical, demographic, and treatment features were included, focusing on age, sex, race, pre-treatment KPS, systemic therapy, number of metastases, lesion location, eloquence, tumor volume, and margin dose. A Random Survival Forest model was trained using one-hot encoding for categorical variables. Performance was assessed with five-fold stratified cross-validation at the lesion-level using Harrell C index. Feature importance was estimated with permutation importance. RESULTS: A total of 884 lesions were analyzed, including 198 local control loss events. The Random Survival Forest achieved high discrimination with a mean C index of 0.919 ± 0.020 across folds. The most influential predictors were age at treatment, margin dose, pre-treatment KPS, tumor volume, and isodose line. Additional contributions were observed from anatomic location and therapy category. CONCLUSIONS: A Random Survival Forest survival model accurately predicted time to local control loss in melanoma brain metastases using routinely collected variables, with strong discrimination and transparent feature importance. This approach enables individualized risk estimation and time-based predictions that can be integrated into clinical decision support and follow-up strategies.