Using electronic health records to improve management of E. coli bloodstream infections

The aim of this thesis is to use linked electronic health records of hospital admissions, and microbiology and haematology results in order to inform management of infections in Oxfordshire, and more generally worldwide. Escherichia coli is one of the leading bacterial pathogens causing bloodstream infections; not only is overall incidence rising, but rising resistance to commonly used empiric antibiotics is also a major concern. Despite playing such a significant role in the burden of infections, the epidemiology of E. coli bloodstream infections is still unclear, particularly when considering unselected populations. I showed that the rise in E. coli bloodstream infections in Oxfordshire is driven by truly community-associated cases, that is, cases identified in the first 48 hours of a hospital admission, or outside of a hospital admission, who had not been admitted to the hospital in the past year. Interestingly, the rate of increase in incidence was faster the further away the previous hospital admission had been. However, rising incidence did not seem to be driven by increasing numbers of patients with evidence of a previous urinary tract infection. The number of co-amoxiclav resistant infections was rising significantly faster than the number of co-amoxiclav susceptible infections and the highest number of resistant infections in 2016 was seen in community-associated cases. However, considering 30-day mortality and various biomarkers of infection, there was no evidence for changes in the severity of infections over time. Higher co-amoxiclav use in primary care was associated with higher rates of co-amoxiclav resistant E. coli urinary tract infections in the subsequent year, providing the evidence needed in order to support the aim of lowering inappropriate use of broad-spectrum antibiotics in the community. Co-amoxiclav susceptibility is particularly challenging to define with traditional laboratory testing methods, with different methodologies leading to different results both in terms of minimum inhibitory concentrations, but particularly when considering a dichotomised susceptible/resistant phenotype. Recent advances in whole genome sequencing technology and analysis tools, as well as decreases in costs, increase the potential utility of predicting phenotype from sequencing-derived genotype, particularly for challenging bacteria-drug combinations. I found that machine learning algorithms – statistical methods which learn from patterns within the data without being programmed explicitly – could predict co-amoxiclav resistance where information is extracted from assembled sequences, either through extracting information about genetic features from mapping onto resistance databases or by considering presence and absence of DNA ‘words’ (k-mers). Crucially, feature selection and expert knowledge are not required when constructing these matrices of genetic features, making these algorithms particularly appealing when considering the constant accumulation of new genetic resistance mechanisms. Finally, by comparing the proportions of resistant bloodstream infections and infections in other less invasive sites, I found that the latter could be used as a surveillance tool for antimicrobial resistance in low- and middle-income countries, since these sites are easier to sample from and cheaper to carry out antimicrobial susceptibility testing on.