Predicting Antimicrobial Resistance Using Machine Learning on Electronic Health Records: A Comparative Study of Ensemble Models

Antimicrobial resistance (AMR), the ability of microbes to survive exposure to drugs intended to eliminate them, is a critical global health concern exacerbated by the overuse and misuse of antibiotics. In this study, we leverage machine learning techniques to predict AMR and evaluate the performance of several advanced supervised algorithms. Using the Antibiotic Resistance Microbiology Dataset (ARMD), a detailed electronic health record (EHR) dataset containing rich clinical, demographic, microbiological, and treatment data from Stanford Healthcare [1, 2].Our approach involves robust data preprocessing to predict the likelihood that a patient’s bacterial iso- late responds to a specific antibiotic as either resistant or susceptible, based on clinical characteristics, microbiological findings, treatment history, and demographic information. We compare the perfor- mance of state-of-the-art machine learning models, including XGBoost, LightGBM, Random Forest, and HistGradientBoostingClassifier, in building reliable predictive models of antibiotic susceptibility. By benchmarking these models on a large real-world dataset, this research identifies effective pre- dictive strategies that can support antimicrobial stewardship, enhance clinical decision-making, and contribute to addressing the growing challenge of AMR.