Internal validation of THINKERS-PA for outcome prediction after pituitary adenoma radiosurgery

Outcome prediction after Gamma Knife radiosurgery (GKRS) for pituitary adenomas remains guided by tumor anatomy, functional status, prior treatment, optic-apparatus proximity, and physician experience rather than individualized estimates of tumor response, biochemical remission, and endocrine morbidity. We developed THINKERS-PA, a mixture-of-experts (MoE) framework for outcome prediction and prescription-dose simulation after GKRS.

We performed a retrospective study of 125 patients with pituitary adenomas treated with GKRS. Variables available at treatment planning were used to train a three-expert MoE neural network integrating clinical and endocrine history, tumor anatomy and imaging, and radiosurgical dosimetry. Discrete-time survival heads estimated tumor progression, biochemical remission in functioning adenomas, and new hypopituitarism. Secondary endpoints include a favorable composite outcome, visual deterioration, and need for additional treatment. Validation used repeated nested 5-fold cross-validation. A dose-sweeping module evaluated treatment doses.

The cohort included 72 functioning and 53 nonfunctioning adenomas. Tumor progression occurred in 14.4%, biochemical remission in 59.7% of functioning adenomas, and new hypopituitarism in 25.0% of patients at risk. Mean out-of-fold time-dependent AUCs were 0.82 for progression, 0.84 for biochemical remission, and 0.81 for hypopituitarism. Corresponding integrated Brier scores were 0.083, 0.143, and 0.119; calibration slopes were 0.89, 0.91, and 0.87; and censoring-aware mean absolute errors were 10.7, 8.6, and 9.8 months. Dose-sweeping estimates were within ± 1 Gy of the delivered dose in 71% of patients.

THINKERS-PA provides an internally validated framework for individualized prediction of tumor, endocrine, and treatment-related outcomes after GKRS. External multicenter validation is required before clinical implementation.