Leveraging nonlinear relationships and interactions to improve 30-day pneumonia readmission machine learning models

Background Research is needed to develop more accurate readmission prediction models that identify patients at the highest risk of readmission after their initial pneumonia hospitalization. Improving prediction accuracy will support the implementation of more effective, personalized interventions to lower readmission rates. Published models tend to rely on traditional methods or advanced machine learning models that exclude continuous variables, overlooking opportunities to uncover nonlinear relationships and interactions. In response, we used electronic medical record (EMR) data, including continuous variables such as vitals, alongside more advanced machine learning (ML) models. Methods Using EMR data from a single academic medical center, we identified adults with initial pneumonia admissions between April 2018 and February 2024. We predicted 30-day readmission using eXtreme Gradient Boosting (XGBoost) and deep neural networks, and compared their performance with that of traditional logistic regression. Results We identified 2,752 patients admitted with pneumonia during the study period (mean age = 70.0 years, 49.1% female). The 30-day readmission rate was 9.9%. The average AUROC for our ML models ranged from 0.62 to 0.64, and AUPRC was 0.15 to 0.16, comparable to traditional logistic regression (0.63 and 0.17, respectively). Previously underemphasized predictors included drug abuse and BUN values. Conclusion Using more advanced machine learning models and continuous variables yielded similar performance to logistic regression models. However, we identified previously understated predictors of readmission after pneumonia hospitalization. Future efforts should focus on gathering important data not readily available in EMR, such as social determinants of health, to potentially enhance the models.