Machine learning for identifying risk factors of nosocomial infection in cancer patients with immune checkpoint inhibitor-related pneumonia

BACKGROUND: This retrospective study used machine learning to find the risk factors of nosocomial infection in cancer patients with immune checkpoint inhibitor-related pneumonia. METHODS: We analyzed data of 120 patients with immune-related pneumonia from a specialized cancer hospital collected between January 2020 and December 2023. Linear logistic regression and nonlinear support vector machine (SVM) models were used to evaluate the predictive factors for nosocomial infection risk among the patients. RESULTS: We found a nosocomial infection rate of 45.83%, predominantly lower respiratory tract infections, among cancer patients with immune-related pneumonia. Severity and mortality rates for the immune-related pneumonia with nosocomial infection group were significantly higher than those for the non-infected group. Logistic regression analysis showed that immune-related pneumonia was significantly associated with the diagnosis time and with C-reactive protein levels. Nonlinear SVM model SHapley Additive exPlanation graph analysis revealed that diagnosis time, tumor radiotherapy, pulmonary dysfunction, and age were risk factors for nosocomial infections in immune-related pneumonia. CONCLUSIONS: Our results highlight the potential of using machine learning to predict the infection risk of immune-related pneumonia. Future multicenter prospective studies are needed to optimize and improve the models and methods used in this study.