Prediction of immune checkpoint inhibitor-related pneumonitis in lung cancer: development and validation of multiple machine learning models.

objectivesThis study aims to develop and validate combined radiomics machine learning (ML) models and clinical semantic feature models to accurately predict immune checkpoint inhibitor-related pneumonitis (CIP) risk in lung cancer patients receiving immunotherapy. Additionally, it seeks to create multiple risk assessment tools for early CIP detection, enhancing patient management and outcomes.

methodsFrom August 2020 to September 2024, candidate predictors were obtained from 210 patients receiving immunotherapy for lung cancer. These predictors included clinical semantic features, blood markers, and imaging histology features. The outcomes for CIP were derived from electronic medical records as well as imaging assessments. Five machine learning algorithms were utilized to develop models and compare their predictive performance.

resultsA total of four potential predictors related to CIP were identified and used to construct a clinical semantic model. The radiomics model constructed based on the support vector machine (SVM) algorithm showed better agreement between the train and test cohorts. The combined model (Area under the curve [AUC], 0.933 and 0.909, respectively) had the best predictive performance. The combined model greatly improved the predictive performance of the CIP. Decision curve analysis (DCA) showed improvement of the combined model in cohorts. Nomograms and Shapley additive explanation of the risk prediction model used to visualize and interpret early predictions of CIP.

conclusionThe combined model constructed by integrating CT radiomics features and clinical semantic features demonstrated the optimal predictive performance. The combined model for early prediction and screening of CIP has the potential for broader clinical applications.