Therapeutic hybrid intelligence with neural and knowledge-based expert reasoning for SRS: an AI model for breast cancer brain metastases

Abstract

Background Prescription dose selection for breast cancer brain metastases treated with stereotactic radiosurgery remains largely guided by tumor size, anatomical constraints, and institutional practice rather than individualized tumor-specific estimates of local failure. We developed THINKERS-Breast, a mixture-of-experts artificial intelligence framework for personalized dose evaluation after Gamma Knife radiosurgery. Methods We performed a retrospective single-center tumor-level study of breast cancer brain metastases treated with Gamma Knife radiosurgery. Variables available before or at treatment were used to train a mixture-of-experts neural network with discrete-time survival modeling. Margin dose was incorporated as a queryable input to enable repeated candidate dose evaluation. Internal validation used grouped 5-fold cross-validation and a grouped holdout test split by patient. Performance was assessed using area under the receiver operating characteristic curve (AUC) for 12-month local failure, mean absolute error (MAE) for expected time to local failure, Brier score, and calibration metrics. Results The cohort included 3,098 tumors from 504 patients. In grouped cross-validation, THINKERS-Breast achieved raw mean AUC >0.807 and calibrated mean AUC of 0.864 for 12-month local failure. Raw and calibrated Brier scores were <0.14 and <0.16, respectively, with calibration intercepts ranging from -0.31 to +0.27. In the grouped holdout set, AUC was 0.781 (95% CI, 0.704–0.857), and MAE for expected time to local failure was 1.55 months (95% CI, 0.78–3.42). Conclusions THINKERS-Breast provides an internally validated framework for tumor-specific local failure prediction and dose-policy evaluation after Gamma Knife radiosurgery for breast cancer brain metastases. External validation is required before clinical deployment.