Radiomics Models to Predict Tumor Response and Pneumonitis in Non-Small Cell Lung Cancer Patients Treated with Immunotherapy

Background: Checkpoint inhibitor-associated pneumonitis (CIP) after immunotherapy has become a challenging issue in non-small cell lung cancer (NSCLC) patients. This study leverages artificial intelligence (AI) algorithms to analyze radiomic features, aiming to predict the occurrence of CIP, as well as tumor response. Methods: This study analyzed data from 159 stage III-IV NSCLC patients undergoing immunotherapy. The patients were categorized into pneumonitis and non-pneumonitis groups, and 3D radiomic features from both tumors and surrounding regions were extracted using LIFEx software. To address scanner-associated variations, a linear mixed-effect radiomics harmonization model was applied. A random forest algorithm was then used to develop models predicting CIP occurrence and tumor responses based on the pre-treatment CT radiomics. The accuracy was evaluated using the area under the curve (AUC). Results: A total of 159 patients were analyzed, of which only 31 experienced CIP. Most had grade 1 (17/31, 54.8%) or 2 (12/31, 38.7%) pneumonitis; only two (6.5%) patients had grade 3. Patients who developed pneumonitis were more likely to be male (64.5% vs. 38.3%, p = 0.014), had less adenocarcinoma histology (54.8% vs. 78.9%, p = 0.032), and exhibited a higher tumor mutational burden (57.1% vs. 24.5%, p = 0.047). Radiomics analysis reported predictability for CIP with an AUC of 0.60 (95% CI 0.55–0.66). The five-year overall and progression-free survival rates were 24.7% (95% CI 15.2–35.5%) and 9.7% (95% CI 4.4–17.4%), respectively. The radiomics features also exhibited AUCs of 0.63 (95% CI 0.59–0.67) in irRECIST and 0.66 (95% CI 0.61–0.70) in RECIST 1.1 in terms of tumor responses to immunotherapy. Conclusions: This study provides insights into the potential role of radiomic models in predicting CIP and tumor responses from pre-treatment CT images of NSCLC patients treated with immunotherapy.