Predicting Burn Injury Readmissions using Machine Learning Methods

Overview

certain readmissions are unavoidable, many can be prevented, potentially easing the financial burden on families and government healthcare syste ms. Despite the unique risks and costs associated with burn injuries, few studies have investigated related readmissions. Our research addresses this gap by evaluating the predictive accuracy of five widely used machine learning methods in forecasting readmission rates related to burn injuries through extensive simulation studies with various factors such as sample size, numb er of features, data dependency structures, and readmission rate prevalence. Our findings reveal that Support Vector Machine and Random Forest methods demonstrate the highest accuracy when the data signal is strong, and logistic regression models could perform competently in scenarios with weak data signals. We apply these methods to the Nationwide Readmissions Database (NRD) and employ a random under - sampling strategy due to the significant data imbalance and a low readmission prevalence. Overall, the Random Forest method emerged as the most effective predict ing method. We further identify five key factors influencing readmission likelihood: longer initial hospital stays and the absence of emergency department services at discharge increase readmission rates while being located in a major metropolitan area, initial weekend admissions, and primary insurance coverage through Medicaid or private providers are associated with reduced readmission rates. These insights offer valuable guidance for physicians and policymakers in formulating strategies and guidelines to mitigate burn-related readmissions.