Statistical and machine learning methods to estimate the impact of antimicrobial resistance on patient outcomes

Thesis

Antimicrobial resistance (AMR) presents a complex, universal threat to healthcare whose estimated impact varies widely by population, microbiology, and analytical methodology. Beset by both inter-study heterogeneity and inconsistent availability of relevant evidence, it is challenging for policy-makers and clinicians to make concerted decisions that balance regional, national and international interests. Accordingly, the World Health Organization’s and the UK’s AMR action plans prioritise the use of electronic health record (EHR) data to measure treatment efficacy in light of local patterns of pathogens and AMR. Whilst EHR data are becoming increasingly available, only a few studies have leveraged them to explore clinical and policy-related repercussions of current definitions of AMR for specific pathogens, and of dysregulated systemic infections like sepsis.

In this thesis, I first analysed patient-level clinical, antibiotic and outcome data to investigate the impact of shifting in-vitro definitions of AMR, focusing on bloodstream infections caused by the most implicated pathogens in high-income countries, Escherichia coli (E. coli) and Enterobacterales. I then examined how unsupervised machine learning approaches for clustering can identify phenotypes of these patients with distinct responses to infection and treatment. Broadening the patient population to those with sepsis, I assessed the validity of phenotypes discovered with different variables in dissimilar healthcare ecosystems. In these analyses, a recurring theme was that whilst studies of the baseline condition and treatment of patients can provide useful insights, they do not account for changes in treatment after baseline. However, naïvely estimating the impact of post-baseline changes to treatment is often done, assuming the absence of time-dependent confounding, including the influence of past treatment history and time-varying covariates on subsequent treatment decisions, and ultimately on the outcome of interest. To address this, I employed marginal structural models to study the impact of differential delays to active antibiotic treatment in E. coli and Enterobacterales bloodstream infections.

Overall, the thesis highlights the challenges of estimating the impact of AMR using routinely-collected EHR data. Further work through inter-institutional and international collaboration will accelerate our ability to study, validate, and sensibly act upon AMR in the coordinated, empirically-justified manner that it demands.

Authors: Yoon, CH

• DOI: 10.5287/ora-jbvdxwep0
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: antimicrobial resistance; machine learning; electronic health records; bacteraemia; causal methods; epidemiological modeling
• Subjects: Medical records; Machine learning; Epidemiology