Machine learning prediction of local control after Gamma Knife radiosurgery to post-resection cavities from brain...

Large symptomatic brain metastases require initial surgical resection. However, local control (LC) after Gamma Knife radiosurgery (GKRS) to resection cavities remains variable. Quantitative risk stratification using routinely available treatment-time variables could inform surveillance and multidisciplinary decision-making.

We performed a retrospective study of post-resection cavities treated with GKRS at a single institution (2014-2024). The primary endpoint was LC. The cohort comprised of 401 post-resection cavities. A gradient boosting classifier was trained using eight routine treatment-time features: age, sex, pre-treatment Karnofsky Performance Status, primary tumor category, single vs. multiple metastases, lobe/structure, eloquence, and cavity volume. Performance was estimated using five-fold stratified cross-validation with out-of-fold predictions and compared with a prevalence-only baseline. Discrimination and calibration were assessed using the receiver operating characteristic area under the curve (ROC-AUC), and Brier score; operating characteristics were reported at a prespecified probability threshold of 0.50.

We compared the performance of two different AI models for predicting LC. The prevalence baseline demonstrated chance-level discrimination (ROC-AUC 0.494). The gradient boosting model improved performance with ROC-AUC 0.735 and PR-AUC 0.802 with acceptable calibration (Brier 0.208). At threshold 0.50, accuracy was 0.701 with sensitivity 0.783 and specificity 0.566. A feedforward neural network trained on the same features performed worse (ROC-AUC 0.672; PR-AUC 0.768; Brier 0.219).

A machine learning model using routine treatment-time variables can meaningfully stratify LC after GKRS to post-resection cavities. The gradient boosting model showed the best performance supporting further external validation and prospective evaluation.

Not applicable.