Therapeutic hybrid intelligence with neural and knowledge-based expert reasoning for SRS (THINKERS): a...

Brain metastases from gastrointestinal (GI) malignancies are heterogeneous, while stereotactic radiosurgery dose selection remains guided by lesion size, location, prior treatment, and physician judgment rather than individualized estimates of local failure and toxicity. We developed THINKERS-GI, a three-expert mixture-of-experts framework for outcome prediction and dose-policy evaluation after Gamma Knife radiosurgery (GKRS).

We performed a retrospective single-center study of 300 patients with 1,233 GI brain metastases. Pretreatment variables were used to train a three-expert mixture-of-experts neural network with discrete-time competing-risk survival modeling. Margin dose was incorporated as a queryable input for candidate-dose evaluation. Internal validation used grouped 5-fold cross-validation and a closed grouped holdout set. Performance was assessed using time-dependent AUC, C-index, Brier score, calibration, and MAE.

In grouped cross-validation, time-dependent AUC ranged from 0.808 to 0.894, with a C-index of 0.889 ± 0.034. The calibrated 12-month Brier score was 0.019 ± 0.013, and MAE for time to local failure was 1.8 ± 0.7 months. In the grouped holdout set, 12-month AUC was 0.806 (95% CI, 0.758-0.851), C-index was 0.781 (95% CI, 0.742-0.818), calibrated Brier score was 0.101, and MAE was 1.27 months (95% CI, 1.2-1.9). Dose sweeping selected a dose within ±2 Gy of the delivered dose in 91.8% of lesions and increased predicted 12-month local control from 82.3% to 84.4%, while reducing predicted adverse radiation effect risk from 7.9% to 7.6%.

THINKERS-GI provides an internally validated framework for individualized local failure prediction, toxicity estimation, and dose-policy evaluation after GKRS for GI brain metastases.

Not applicable.