Bayesian latent class mixture models for antimicrobial resistance data with censoring

Abstract

Antimicrobial resistance (AMR) has been one of the most serious global public health threats in the 21st century. Our research focused on developing statistical methodologies to study the quantitative problems related with AMR through the surveillance program in the United States, the National Antimicrobial Resistance Monitoring System (NARMS). However, the records of the minimum inhibitory concentration (MIC) in the surveillance program are censored data with unknown facts of resistance level. Additionally, AMR usually presents an overlapped mixture of susceptible and resistant isolates, which necessitates the modeling of such a distribution and separation of the two components. To account for the data censorship, we augmented each laboratory measurement with its latent MIC variable that exists within the interval as indicated by the observation. A latent class Gaussian mixture model was assumed to reflect the high likelihoods of therapeutic success and failure, whose mixing weights represent the probabilities of the two clusters. The work series began with the task of monitoring the temporal pattern of AMR in Chapter 2 where a linear regression was constructed on the basis of the latent class mixture model with censorship. In Chapter 3, we aimed to detect similar direction of AMR changes across populations. With a multivariate normal distribution associating different populations, their correlation for the conventionally unnoticed mean shift of the susceptible bacteria was quantified. In Chapter 4, we developed a proactive tool for detection of antibiotic multidrug resistance, motivated by its huge impact to healthcare and economic. Another Bayesian framework that estimates the resistance level jointly for antibiotics of different classes was proposed. The model provides inferences of correlation in the latent MIC and the binary susceptibility, both of which should be considered when assessing multidrug resistance. The models in our work were evaluated to be accurate and robust according to the simulation studies. Applications of our work can provide useful information to the surveillance programs, which further warrants investigation and clinical intervention of antimicrobial use.