MS thesis Columbia University

Thesis Title: Developing Gaussian Mixture Models to Forecast Bacterial Antimicrobial Resistance Incidence Across 71 Countries’ Large Routinely Collected Health Data

Study Aims:
Could standardized, global forecasting models of bacterial antimicrobial resistance be developed that detect and forecast deviations across pathogen-antibiotic pairs, countries, age groups, and sexes over time?

Methods:
Following Catalán et al., we modelled log-transformed minimal inhibitory concentrations (MIC) distributions of each pathogen-antibiotic pair as Gaussian mixture models to determine the number of clusters and classify clusters as resistant or susceptible (greatest vs. lowest mean; all other clusters are intermediate). We incorporated the mean of each MIC cluster into a hierarchical Gaussian Process regression framework, well-suited for forecasting trends with quantified uncertainty, to produce out-of-sample forecasts of expected resistance baselines and their uncertainty. These models also forecast the probability of baseline exceedance events, particularly across pathogen-antibiotic pairs, countries, age groups, and sexes over time, anticipating where an outbreak might occur and which populations are most at risk. We estimated 80% simulated power to detect a 5% change in resistance each year (α = 0.05) using Monte Carlo simulation, following standard practices for estimating simulated power in outbreak detection methods. We used two-sided Z-tests (α = 0.05) to compare the forecasted incidence of baseline exceedance events in 2018-2023 with the incidence of baseline exceedance events reported in ATLAS’s 2018-2023 data to see if our models reliably detected and forecast incidence patterns when aligned with current CLSI and EUCAST guidelines, using ATLAS as the gold standard.