A Bayesian Semi-Mechanistic Framework for Modelling Antimicrobial Resistance
Abstract
Antimicrobial resistance is recognised as a global health threat but trends in the prevalence of resistance are often obscured by heterogeneity between data sources. Acinetobacter baumannii is a globally significant cause of antibiotic-resistant healthcare-associated infections. Data on A. baumannii resistance in Asia are available from 18 data sources, each using different methods to identify infections and resistance. Mechanistic mathematical models attempt to explain trends in resistance with biologically meaningful parameters but do not account for data heterogeneity. Purely statistical models can account for heterogeneity but have limited utility in predicting responses to interventions. Here we use a semi-mechanistic model assuming that trends in prevalence follow logistic growth dynamics, described by two parameters representing the coefficient for resistance acquisition and the loss of resistance. Using a Bayesian framework to account for uncertainty and missing data, we assume that these two parameters could vary by country, time and data source, and that neighbouring countries could exhibit spatial autocorrelation. We compared six models with different assumptions for these parameters, along with a stacked ensemble of these models. For four out of seven pathogen-antibiotic combinations considered, the stacked model performed best. For the other pathogen-antibiotic combinations, models that accounted for heterogeneity between data sources always performed better than models that did not. This framework improves existing approaches by offering a more biologically plausible structure that can quantify the effects of covariates on resistance dynamics and effectively handle heterogeneous data sources. This methodology provides a potentially valuable foundation for better understanding AMR resistance dynamics and evaluating public health interventions.
Author Summary
Antimicrobial resistance occurs when disease-causing microorganisms become unaffected by the drugs designed to destroy them. This is a growing public health threat since these medicines become ineffective and can make infections difficult to treat. Understanding resistance trends across countries and over time is challenging because available surveillance data is sparse and comes from many different sources that can vary in terms of how the data is collected, the types of patients being treated, and the type of facility the data is being collected. We developed a way to model resistance trends by accounting for differences between data sources, something existing models do not consider, while maintaining the biological nature of how resistance growth is modelled. We applied our method, consisting of 6 different models, using data on a pathogen in Asia from 2000 to 2022. Our results show that accounting for variability between data sources improves our ability to predict resistance trends. Creating models that can more accurately estimate resistance patterns are essential for public health planning and evaluating health interventions. Our framework provides a potentially valuable way to better understand resistance trends and can be used as a foundation to model and assess different public health interventions.
Related articles
Related articles are currently not available for this article.