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

BACKGROUND

Prescription dose selection for intracranial meningiomas treated with stereotactic radiosurgery remains guided by tumor volume, anatomical constraints, proximity to critical structures, and institutional practice rather than individualized estimates of long-term local progression. We developed THINKERS-Meningioma, a mixture-of-experts artificial intelligence framework for personalized dose evaluation after stereotactic radiosurgery.

METHODS

We performed a retrospective single-center study of 800 patients with intracranial meningiomas treated with stereotactic 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 stratified 5-fold cross-validation and a stratified holdout test split. Performance was assessed using area under the receiver operating characteristic curve (AUC) for 5-year local tumor progression, mean absolute error (MAE) for expected time to local progression, Brier score, and calibration metrics.

RESULTS

In stratified cross-validation, THINKERS-Meningioma achieved a raw mean AUC of 0.854 ± 0.077 and calibrated mean AUC of 0.861 ± 0.088 for 5-year local tumor progression. Raw and calibrated Brier scores were 0.0021 ± 0.0020 and 0.0016 ± 0.0024, respectively. The cross-validation MAE was 1.67 ± 0.20 months. In the stratified holdout set, raw AUC was 0.841 and calibrated AUC was 0.861 (95% CI, 0.829-0.915). MAE for expected time to local progression was 1.71 months (95% CI, 1.00-2.01).

CONCLUSIONS

THINKERS-Meningioma provides an internally validated framework for individualized 5-year local progression prediction and dose-policy evaluation after stereotactic radiosurgery for intracranial meningiomas.

CLINICAL TRIAL NUMBER: Not applicable.