Truace
Subscribe

Model predicts which melanoma and kidney cancer brain tumors will return after radiosurgery

Study title: Pre-treatment machine learning prediction of local failure and time to local failure after Gamma Knife radiosurgery for radioresistant brain metastases

Researchers built a random survival forest, a machine learning model, to predict which melanoma and renal cell carcinoma brain metastases would fail local control after Gamma Knife radiosurgery, and when. The model used only variables known before treatment. On a holdout test set of 179 lesions it reached a C-index of 0.842.

Read the study abstract

Background Local control after Gamma Knife radiosurgery (GKRS) for radioresistant brain metastases remains variable. Melanoma and renal cell carcinoma brain metastases are relatively radioresistant, yet pre-treatment prediction of local failure and its timing remains limited. We developed and internally validated a survival machine learning model to predict lesion-specific local failure and time to local failure using only pre-GKRS variables. Methods We performed a retrospective single-center study of melanoma a…

Truvace record2 min read1 sourceEvidence: Moderate · 60/100 — Observational study · n = 893 · 1 independent sourcePeer-reviewedTRV-2026-2430 · cite or check it
Pre-treatment machine learning prediction of local failure and time to local failure after Gamma Knife radiosurgery for radioresistant brain metastases"Medical Laboratory" by ben.dracup, CC BY 2.0.

In brief

Local control after this radiosurgery is variable for these relatively radioresistant tumors, and pre-treatment prediction of failure and its timing has been limited. The authors say the model could support individualized risk stratification and follow-up planning. The results come from internal validation at one center.

Main points

  1. The retrospective single-center cohort included 893 lesions; 228 (25.5%) developed local failure.
  2. The model was a Random Survival Forest trained on baseline demographic, clinical and tumor-level variables available before treatment.
  3. Local failure was modeled as a right-censored survival outcome, with each lesion as the prediction unit.
  4. Data were split at the patient level to minimize information leakage.
  5. On the 179-lesion holdout set, dynamic AUC was 0.904 at 6 months, 0.890 at 12 months and 0.851 at 24 months.
  6. Brier scores were 0.130, 0.153 and 0.152 at those time points.
  7. Patient-grouped 5-fold cross-validation gave a mean C-index of 0.882.
  8. The strongest predictors were age at treatment, pre-treatment KPS and tumor volume.

The gain

In a single-center study of 893 brain metastases, a random survival forest using only pre-treatment data predicted local failure after Gamma Knife radiosurgery with a C-index of 0.842 on a 179-lesion holdout set.

The full story

The study focused on melanoma and renal cell carcinoma brain metastases treated with Gamma Knife radiosurgery, tumor types considered relatively radioresistant. Of 893 lesions, 228 (25.5%) developed local failure.

A Random Survival Forest was trained on demographic, clinical and tumor-level variables available before treatment. Each lesion was a prediction unit, while splits were made by patient to limit information leakage. Performance was judged with the concordance index, cumulative/dynamic AUC and Brier score under patient-grouped 5-fold cross-validation.

On the 179-lesion holdout set the C-index was 0.842 and the dynamic AUC fell from 0.904 at 6 months to 0.851 at 24 months. Cross-validation gave a mean C-index of 0.882. Age at treatment, pre-treatment KPS and tumor volume were the strongest predictors.

Sources

  1. Peer-reviewedJournal of Neuro-Oncology2026-11-02

The debate