Predictive Modeling of Pneumonia-Related Mortality in Long-Term Immunosuppressive Therapy Patients: A Machine Learning Approach with Interpretability

Abstract

Abstract Background The use of corticosteroids and immunosuppressive agents has become a cornerstone in the management of inflammatory and autoimmune diseases, but it comes with an increased risk of infections, particularly pneumonia. Machine learning (ML) and SHAP (SHapley Additive exPlanations) offer a promising approach to predict pneumonia-related mortality in patients on long-term immunosuppressive therapy. Methods Data from a retrospective cohort analysis of pneumonia patients undergoing glucocorticoid treatment were collected from six academic hospitals in China. Features such as demographics, clinical symptoms, disease severity, laboratory data, and treatment information were analyzed. Feature selection was performed, and three survival analysis models (Cox regression, Random Survival Forest, Fast Survival Support Vector Machine) were developed. Model interpretability was enhanced using SHAP. Results Among 716 patients, 74.02% survived, and 25.97% died within 90 days. Dyspnea, ventilation support, and certain laboratory values were associated with higher mortality. Eight predictors (Platelet, Albumin, Aspartate Aminotransferase, PH, Glucose, Blood Urea Nitrogen, Oxygenation index, Persistent lymphocytopenia) were identified for model development. The Random Survival Forest model outperformed others, showing a C-index of 0.754 and a Time-dependent AUC of 0.795. SHAP analysis revealed the impact of these predictors on patient outcomes. Conclusion Machine learning, coupled with SHAP analysis, identifies key predictors and enhances prediction accuracy for pneumonia-related mortality in patients on long-term immunosuppressive therapy. This approach facilitates risk stratification and informed clinical decision-making, potentially improving patient outcomes.