Comparative Evaluation of Time Series Forecasting Approaches for Facility-Level Antibiotic Resistance Outcomes in the Veterans Health Administration

Antibiotic resistance is a critical public health threat, particularly in hospital settings where vulnerable populations face heightened risks of infection and adverse outcomes. Forecasting resistance trends at the facility level is essential for guiding local antibiotic stewardship, informing infection control strategies, and improving patient safety. In this study, we conducted a comparative evaluation of time series forecasting methods to predict facility-level antibiotic resistance trends in hospital-onset infections within the Veterans Health Administration system. Using data from January 2007 to March 2022 across 30 high-volume facilities, representing approximately half of all admissions and patient days, we assessed the performance of autoregressive integrated moving average (ARIMA), vector autoregression (VAR), and long short-term memory (LSTM) models. Our results showed that LSTM models consistently outperformed ARIMA and VAR, with accuracy notably enhanced by incorporating antibiotic use covariates. These findings underscore the value of integrating antimicrobial utilization data and deep learning approaches to improve resistance forecasting and support targeted stewardship efforts.