Explainable Machine Learning Model Predicts Melanoma Brain Metastases Response to SBRT

An explainable machine learning model can reliably predict whether oligometastatic brain lesions from melanoma will achieve a complete or partial response two months after stereotactic radiosurgery or fractionated stereotactic radiotherapy. In a study published in the Journal of Neuro-Oncology, researchers developed and evaluated artificial intelligence tools to forecast local treatment outcomes for patients with melanoma brain metastases treated with CyberKnife stereotactic body radiation therapy. The findings indicate that combining computational learning algorithms with post-hoc explanation methods identifies key biological and treatment factors that govern therapeutic success.

The retrospective investigation, selected within the Rabbit study framework, analyzed fifty consecutive patients presenting with a total of 102 melanoma brain metastases treated between December 2012 and December 2018. The primary clinical objective was to predict complete or partial tumor regression at two months following radiosurgical or fractionated stereotactic intervention. Tumor response assessments were performed via follow-up brain magnetic resonance imaging using the Response Evaluation Criteria in Solid Tumors and revised according to the response criteria for brain metastases defined by the Response Assessment in Neuro-Oncology group.

For every treated lesion, the research team recorded multiple baseline clinical, anatomical, and therapeutic variables. These candidate features included the total number of treated lesions, pre-treatment lactate dehydrogenase levels, prior therapeutic interventions such as surgical resection or whole-brain radiation therapy, and the discovery interval between the primary melanoma diagnosis and brain lesion detection. The investigators also documented lesion dimensions, anatomical location, stereotactic body radiation therapy dose and fractionation schedule, BRAF genetic mutation status, and the type of concomitant systemic treatment received with ipilimumab.

To evaluate predictive capabilities, the researchers trained and internally validated six distinct supervised machine learning algorithms using five-fold cross-validation. The evaluated architectures comprised logistic regression, k-nearest neighbors, support vector machines, random forests, decision trees, and light gradient boosting machines. Algorithmic discrimination and diagnostic efficacy were quantified using receiver operating characteristic curves and area under the curve metrics. A decision tree model was also constructed to interpret the complex predictive interactions linking clinical covariates directly to tumor response.

Across the entire cohort, twenty-three point five percent of the treated brain lesions demonstrated a complete or partial response, which served as the reference standard ground truth for algorithmic optimization. The light gradient boosting machine emerged as the superior model, achieving a mean cross-validated area under the curve of 0.87, with a reported performance range across validation folds of 0.85 to 0.91. These results demonstrated the highest comparative accuracy among the evaluated supervised architectures in predicting post-radiosurgical outcomes.

To interpret the internal logic of the models, the investigators applied the SHapley Additive exPlanations method to generate individualized explanatory profiles for predictive outputs. This analytical framework revealed that complete or partial response was strongly governed by four primary covariates, listed in descending order of relative feature importance: tumor BRAF mutation status, total administered radiation dose, the specific protocol of concomitant ipilimumab administration, and baseline lesion dimensions.

The decision tree analysis detailed specific outcome stratifications based on these identified determinants. Lesions characterized by a wild-type BRAF status exhibited an overall response probability of only 10.7 percent, and those within this wild-type subset receiving radiation doses below 24 Gray recorded a zero percent probability of complete or partial response. Conversely, lesions harboring a mutated BRAF status demonstrated an overall response probability of 39.1 percent, and patients with BRAF-mutated lesions who received concomitant ipilimumab therapy achieved a one hundred percent probability of achieving a complete or partial response.

The investigators concluded that explainable machine learning workflows provide reliable predictive assessments of treatment response in oligometastatic melanoma lesions treated with stereotactic body radiation therapy. While these algorithmic tools offer substantial promise to assist radiation oncologists in designing tailored, individualized treatment regimens for this patient population, the authors emphasized that formal external validation of the findings across independent cohorts remains mandatory before clinical adoption.

Subject of Research: Machine learning prediction of melanoma brain metastases response to stereotactic radiotherapy

Article Title: Explainable machine learning predictive models for treatment response after SBRT in melanoma brain metastases

Article References: Cilla, S., Franco, R. D., Borzillo, V., Scipilliti, E., Mottareale, R., Serra, M., Buonanno, F., Macchia, G., Deodato, F., Romano, C., Mercogliano, S., Ametrano, G., Vitale, M. G., Ascierto, P., Pezzulla, D., & Ravo, V. (2026). Explainable machine learning predictive models for treatment response after SBRT in melanoma brain metastases. Journal of Neuro-Oncology, 180(2), Article 42. https://doi.org/10.1007/s11060-026-05836-5

Image Credits: AI Generated

DOI: 10.1007/s11060-026-05836-5

Keywords: melanoma brain metastases, stereotactic body radiation therapy, CyberKnife, machine learning, explainable artificial intelligence, SHAP analysis, light gradient boosting machine, BRAF mutation status, ipilimumab, oligometastatic disease, Journal of Neuro-Oncology, RANO criteria, RECIST, radiation dose, decision tree, neuro-oncology