Abstract
Background/Objectives: Immune checkpoint inhibitors (ICIs) have been extensively used for the treatment of non-small cell lung cancer patients in recent years, providing a significant survival benefit. However, a major drawback of ICI-related immunotherapy is the risk of developing post-surgical pneumonitis. Methods: In this study, we propose a deep learning-embedded, multi-modality prediction approach to assess whether patients will develop ICI-pneumonitis after receiving ICI-based immunotherapy. This approach utilizes multi-modal data, including clinical data and pre-treatment lung screening computed tomography (CT) images. We extracted three types of features: (1) deep learning features from CT scans using a pre-trained vision transformer; (2) radiomic features from CT scans using pre-defined radiomic algorithms; (3) clinical features from patients' electronic health records. We then compared ten machine learning algorithms for prediction based on these extracted features. Results: Our experiments demonstrated that using all three types of features leads to the best prediction result, with a prediction accuracy rate of 0.823 and an area under the receiver operating characteristic curve of 0.895. Conclusion: Multimodal approaches can result in superior prediction results compared to single modality approaches. This study demonstrates the feasibility of developing machine learning algorithms to accurately predict ICI-pneumonitis and contributes to the early identification of patients who are at a higher risk of developing pneumonitis.