Machine Learning-Based Prediction of Metastasis-Level Response to Combined SBRT and Immunotherapy
Abstract
Purpose
Although immune checkpoint blockade (ICB) benefits only a subset of patients, combining ICB with stereotactic body radiation therapy (SBRT) may enhance therapeutic response. Building on a validated CT-based radiomics score (RS) that predicts patient-level response to SBRT plus pembrolizumab (SBRT+P), this retrospective study developed a machine-learning (ML)-based approach to predict metastasis-level response by identifying lesions susceptible to SBRT-induced T-cell infiltration, with the goal of enabling more informed treatment decisions for patients with multiple metastases
Methods
This retrospective study included 68 patients with 140 metastases, who received site-specific SBRT followed by pembrolizumab within 7 days. Metastatic response was assessed at first follow-up (3 months) using RECIST 1.1 criteria for soft-tissue metastases and MD Anderson response criteria for bone metastases. For each metastasis, 85 features were extracted, including 41 intratumoral radiomics features, 41 peritumoral radiomic features, and 3 lesion-location indicators. An ML-based prediction framework incorporating backward feature selection and shuffle–split validation was developed to predict metastasis-level response. Kernel SHAP was employed to quantify the contribution of selected features to model predictions, enabling interpretation of feature–response relationships while accounting for nonlinear feature interactions. Model performance was compared with a radiomics score–based logistic regression model (RS-Logistic)
Results
Forty-five of 140 metastases (32%) responded to SBRT+P. The ML model outperformed the RS-Logistic (AUC = 0.77 vs. 0.68). Five features were selected, including intratumoral and peritumoral intensity and textural features, and one lesion-location indicator. Our preliminary results indicate that dense soft-tissue lesions characterized by large, internally coherent homogeneous regions, sharp local-scale intensity transitions, and irregular peritumoral boundaries are more likely to respond to SBRT+P
Conclusion
The proposed ML model integrates radiomic features of metastases and their peritumoral microenvironment with lesion location to predict lesion-level response to combined SBRT-immunotherapy. Further studies are warranted to validate the predictive value of these biomarkers in larger, external cohorts