aceMachine learning prediction of local control after Gamma Knife radiosurgery to post-resection cavities from brain metastases: a proof-of-concept study
Abstract: Background 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. Methods 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 40…
"Brain - CT scan - Metastatic pulmonary adenocarcinoma Case 239 (7603361920)" by Yale Rosen from USA is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.
Researchers retrospectively evaluated 401 post-resection cavities from brain metastases treated with Gamma Knife radiosurgery at a single institution from 2014 to 2024. They trained a gradient boosting classifier on eight routine treatment-time features to predict local control, estimating performance with five-fold stratified cross-validation.
By the July 2026 publication date, the model had shown improved discrimination over chance with ROC-AUC 0.735 and PR-AUC 0.802 and acceptable calibration, suggesting potential to inform surveillance. What remains uncertain is whether performance holds outside this single-center retrospective cohort, as authors noted need for external validation and prospective evaluation.
- Retrospective cohort of 401 post-resection cavities treated with GKRS at a single institution from 2014-2024 with local control as primary endpoint.
- Model used eight routine features: age, sex, pre-treatment Karnofsky Performance Status, primary tumor category, single vs multiple metastases, lobe/structure, eloquence, and cavity volume.
- Performance estimated with five-fold stratified cross-validation with out-of-fold predictions; prevalence baseline ROC-AUC 0.494 versus gradient boosting ROC-AUC 0.735.
- At prespecified threshold 0.50, gradient boosting accuracy 0.701 with sensitivity 0.783 and specificity 0.566; feedforward neural network performed worse at ROC-AUC 0.672.
A gradient boosting classifier trained on eight routine treatment-time features improved prediction of local control after Gamma Knife radiosurgery to post-resection cavities from brain metastases, achieving ROC-AUC 0.735 compared to chance-level baseline.
The rundown
The study analyzed 401 post-resection cavities treated with Gamma Knife radiosurgery between 2014 and 2024 at one center, using five-fold stratified cross-validation and out-of-fold predictions to estimate performance against a prevalence-only baseline.
Eight routinely available treatment-time variables were used as inputs, and the gradient boosting model outperformed both the baseline (ROC-AUC 0.494) and a feedforward neural network (ROC-AUC 0.672, PR-AUC 0.768, Brier 0.219) with reported operating characteristics at threshold 0.50.
Sources
- Peer-reviewedJournal of Neuro-Oncology2026-07-10
How should this claim be treated?
The debate