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

Area of Science:

• Medical Informatics
• Health Services Research
• Machine Learning in Healthcare

Background:

• Accurate prediction of hospital readmissions after pneumonia is crucial for effective patient intervention.
• Existing models often exclude continuous variables, limiting the capture of complex patient data.
• Electronic Medical Record (EMR) data offers potential for enhanced predictive modeling.

Purpose of the Study:

• To evaluate advanced machine learning (ML) models, including continuous variables, for predicting 30-day readmission after pneumonia.
• To compare the performance of ML models against traditional logistic regression.
• To identify novel predictors of readmission.

Main Methods:

• Utilized EMR data from adult patients with initial pneumonia admissions (April 2018 - February 2024).
• Employed eXtreme Gradient Boosting (XGBoost) and deep neural networks for readmission prediction.
• Compared ML model performance (AUROC, AUPRC) with logistic regression.

Main Results:

• The study included 2,752 patients; the 30-day readmission rate was 9.9%.
• ML models showed comparable performance (AUROC 0.62-0.64) to logistic regression (AUROC 0.63).
• Drug abuse and Blood Urea Nitrogen (BUN) values emerged as significant, previously understated predictors.

Conclusions:

• Advanced ML models incorporating continuous variables did not outperform logistic regression in this cohort.
• Identification of understated predictors like drug abuse and BUN offers new avenues for risk stratification.
• Future research should integrate social determinants of health data for improved predictive accuracy.