Therapeutic hybrid intelligence with neural and knowledge-based expert reasoning for SRS (THINKERS): an AI model for lung cancer brain metastases.

Abstract

Prescription dose selection for lung brain metastases treated with stereotactic radiosurgery (SRS) remains largely guided by generalized practice patterns rather than tumor-specific modeling of local failure dynamics. We developed a Therapeutic Hybrid Intelligence with Neural and Knowledge-based Expert Reasoning for SRS (THINKERS), an artificial intelligence framework for personalized dose evaluation in lung brain metastases. We performed a retrospective single-center study of lung brain metastases treated with Gamma Knife radiosurgery. Only variables available at or before treatment were included. The final model used a mixture-of-experts (MoE) deep neural network with discrete-time survival modeling. Margin dose was incorporated as an explicit input variable, allowing repeated evaluation across candidate dose levels for tumor-specific dose recommendation. Internal validation consisted of grouped 5-fold cross-validation and a grouped holdout test split by patient. The final analytic cohort included 767 patients with 3,728 treated lung brain metastases. In grouped cross-validation, the MoE model achieved a mean Area Under the Curve (AUC) of 0.876 for 12-month local failure and a mean absolute error (MAE) of 0.99 months. In the grouped holdout test set, the model achieved an AUC of 0.863 (95% CI, 0.776-0.942) and an MAE of 1.26 months (95% CI, 0.44-1.49). Probabilistic performance was favorable, with a Brier score of 0.061, calibration intercept of 0.18, and calibration slope of 0.87. THINKERS-Lung provides an internally validated framework for tumor-specific SRS dose evaluation in lung brain metastases and supports the feasibility of AI-guided personalized radiosurgical decision support. Not applicable.