Deep-Learning Model for Personalized Prediction of Positive MRSA Culture Results Using Patient’s Time-Series Electronic Health Records

Abstract Methicillin-resistant Staphylococcus aureus (MRSA) is a common bacterial cause of morbidity and mortality. Our deep-learning model (PyTorch_EHR) processes time-series structured electronic health record (EHR) data, including previous cultures and antimicrobial exposures, to predict the lab result of MRSA culture positivity over the next two weeks. After training and evaluation on data from 8,164 MRSA and 22,563 non-MRSA patient events from Memorial Hermann Hospital System, Houston, Texas, the PyTorch_EHR outperformed traditional machine learning methods logistic regression and light GBM (Area Under the Curve of Receiver Operating Curve [AUC]PyTorch_EHR=91.12%, AUCLR=85.91%, AUCLGBM=89.11%). External validation using the MIMIC-IV dataset of 393,713 patient events from a tertiary care center in Boston, Massachusetts, confirmed PyTorch_EHR’s accuracy (AUCPyTorch_EHR=85.50%, AUCLR=83.24%, AUCLGBM=82.48%). The model maintained its accuracy across most subgroup analyses based on infection type. The cumulative incidence curves based on our model successfully high-, medium-, and low-risk patients. This study demonstrates the potential of deep-learning models to predict the presence of MRSA-positive cultures to optimize MRSA antimicrobial therapy.