Short-term peri- and intra-tumoral CT radiomics to predict immunotherapy response in advanced non-small cell lung cancer.

Background: Predicting response to immunotherapy is crucial for advanced non-small cell lung cancer (NSCLC) treatment planning, but effective predictive markers for immunotherapy efficacy are still lacking. This study aimed to develop an explainable machine learning model for predicting immunotherapy responses in advanced NSCLC patients. Methods: A total of 245 advanced NSCLC patients from two centers who received immunotherapy were retrospectively enrolled. For each primary tumor, three regions of interest were analyzed, namely, the intratumoral region (ITR), peritumoral region (PTR), and combined intratumoral and PTR (IPTR). Pre-radiomics features and delta-radiomics features reflecting the rate of change between radiomics features before and after treatment were extracted. Models for predicting immunotherapy responses were established via the extreme gradient boosting (XGBoost) classifier and assessed in terms of discrimination, calibration, and clinical utility. The SHapley Additive exPlanations (SHAP) tool was employed to explore the interpretability of the model. Kaplan-Meier (KM) analysis of progression-free survival (PFS) was conducted to evaluate the prognostic value of the prediction models. Results: The delta-radiomics models of ITR and IPTR demonstrated optimal performance in predicting immunotherapy response, significantly improving the area under the curve (AUC) to 0.85 and 0.83 in the internal validation cohort and 0.84 and 0.86 in the external validation cohort. SHAP revealed a strong relationship between the delta-radiomics feature values and the model-predicted probabilities. KM curves indicated that the high-risk groups identified by the delta-radiomics models had significantly worse PFS than did the low-risk groups across all cohorts. Conclusions: The results demonstrated that a model based on multiple time points outperformed one based on a single time point. The delta-radiomics model has been proved a noninvasive approach for assessing the response of advanced NSCLC patients to immunotherapy and facilitates individualized treatment decision making.