Improving Risk Prediction of Methicillin-Resistant Staphylococcus aureus Using Machine Learning Methods With Network...

Area of Science:

• Computational epidemiology
• Health informatics
• Machine learning in healthcare

Background:

• Healthcare-associated infections (HAIs) from multidrug-resistant organisms (MDROs) like MRSA and CDI pose a significant burden.
• Current MDRO screening is resource-intensive, necessitating innovative approaches.

Purpose of the Study:

• To develop automated tools for predicting MDRO colonization or infection risk using electronic health record (EHR) data.
• To enhance infection control strategies and guide empiric antibiotic coverage decisions.

Main Methods:

• Retrospective development of a machine learning model using EHR data to detect MRSA colonization/infection.
• Inclusion of clinical, nonclinical, and network-based features derived from patient data.
• Exploration of heterogeneous models for specific patient subpopulations to optimize performance.

Main Results:

• Penalized logistic regression demonstrated superior performance, with an 11% AUC improvement using polynomial feature transformation.
• Key predictors of MDRO risk included antibiotic use, surgery, device use, dialysis, comorbidities, and network features.
• Network features provided the most significant performance improvement, increasing model accuracy by at least 15%.

Conclusions:

• Machine learning effectively predicts MRSA risk using EHR data, integrating clinical and nonclinical factors.
• Network features are highly predictive, substantially improving upon existing methods for MDRO risk assessment.
• Heterogeneous models tailored to patient subpopulations enhance predictive accuracy for infection control.