Purpose/Objective: To develop an explainable artificial intelligence (XAI) model to predict the complete/partial response of oligometastatic brain lesions from melanoma (MBMs) after two months from radiosurgical or fractionated stereotactic treatments.
Material/Methods: 50 consecutive patients with a total 102 MBMs treated with CyberKnife® SRS/SRT from December 2012 to December 2018 were selected within the Rabbit study. Local response was evaluated on brain MRI performed during follow-up using response evaluation criteria in solid tumors (RECIST) and revised using response criteria for brain metastases from the response assessment in neuro-oncology criteria (RANO) group. For each treated lesion, the following variables were considered: number of treated lesions, the LDH pre-treatment level, previous treatments (surgery or whole brain RT), BRAF status, time from primary diagnosis to brain lesions discovery (Time_diag), lesion dimensions and site, SBRT dose and fractionation, type of concomitant treatment with ipilimumab (IPI). Six different supervised machine learning models, including logistic regression (LR), k-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), decision tree (DT) and light gradient boosting (LGBM), were trained and internally validated with 5-fold cross validation. Model performances were assessed using receiving operator curves (ROC) and area under the curve (AUC). An explainable approach based on SHapley Additive exPlanations (SHAP) method was deployed to generate individual explanations of the model decisions. Lastly, a decision tree was created to better understand and interpret the complex relationships between the covariates and the target variable.
Results: 23.5% of lesions had a complete or partial response and were used as ground truth for the ML models. The best performing model was the LGBM, reporting a mean AUCs in the 5-fold cross validation of 0.87 (0.85-0.91). SHAP analysis strongly associated complete/partial response with four variables, namely in order of importance: the BRAF status, the total dose, the type of treatment with ipilimumab, and the lesion dimension. In particular, the DT classified patients with BRAF “wild-type” status having only 10.7% probability to reach complete/partial response. Among these lesions, those treated with doses < 24 Gy had 0% probability of complete/partial response. On the other side, lesions with a BRAF “mutated” status reported a probability of complete/partial response of 39.1%. Among these lesions, those treated with concomitant ipilimumab reached 100% probability of complete/partial response.
Conclusion: Machine-learning models enables a reliable prediction of the treatment response of oligometastatic lesions receiving SBRT. This approach may assist radiation oncologists to tailor more individualized treatment plans in this clinical setting. An external validation of the results is mandatory.
Publication Date: 2026-06-19