Estimating coverage of empiric treatment regimens for childhood bloodstream infection based on routine microbiological data

Abstract

This research paper style thesis comprises six papers, each addressing a different aspect of the selection of empiric antibiotic regimens for the treatment of severe childhood infections, focussing on suspected bloodstream infection. Antibiotics are a means to effectively manage life-threatening bacterial infections, such as bloodstream infections. Recommendations for life-saving empiric antibiotic treatment for bloodstream infection are traditionally based on knowledge of the epidemiology of the targeted infection, and are strongly influenced by knowledge about antibiotic resistance in causative pathogens. The underlying assumption is that the in vitro phenomenon of antimicrobial resistance relates to a poor response to antibiotics in vivo. Bacteria causing bloodstream infection are increasingly found to be resistant to antibiotics and this can vary by region, hospital and patient group. It is therefore necessary to select and review best options for empiric treatment taking into account these trends. Details on the current approaches, data sources and the advantages and limitations of both are discussed in the first part of thesis (chapters 2-5). The methods for selecting optimal empiric treatment from microbiological data, including information on antimicrobial resistance, are poorly defined. It is unclear which approach is most informative clinically and which can still use microbiology data generated as part of routine care and utilized for surveillance. Importantly, empiric regimens must be based on knowledge of the bacteria associated with a specific infection syndrome including their relative frequency as well as their resistance patterns. The probability that a given regimen will cover the next clinically identified episode of the infection in question can then be derived as guidance for regimen selection. In the second part of the thesis, a specific method for constructing a weighted-incidence syndromic combination antibiogram (or WISCA) to estimate coverage is therefore developed and presented. The WISCA is derived from a Bayesian decision tree model, and has the advantages of explicitly combining relative incidence and resistance patterns for a given syndrome as well as accurately reflecting imprecision of coverage estimates. The Bayesian decision tree WISCA is used to investigate coverage of empiric antibiotic regimens at hospital level in Europe, including potential methods for dealing with heterogeneity between centres while still supporting data pooling to improve precision (Chapter 6). A further application is the estimation and comparison of coverage offered by recommended regimens for neonatal sepsis in Asian countries with data pooling at the level of country (Chapter 7). Finally, the potential influence of patient characteristics on selection of antibiotics of last resort (i.e. those with a broad therapeutic spectrum but likely to be strong drivers for the selection of antimicrobial resistance and therefore to be used only when necessary) was investigated (Chapter 8). This demonstrates that certain patients or infection episodes are more likely to be treated with last resort antibiotics than others, and would seem to indicate expected heterogeneity among neonates and children with bloodstream infection. The Bayesian WISCA provides a useful approach to pooling information to guide empiric therapy and could increase confidence in the selection of specific regimens. In presented analyses, it provides evidence for the continued use of narrow-spectrum regimens in certain contexts, and could be further developed to address data pooling and allow the integration of local resistance data with surveillance data for data-based modification of high-level treatment recommendations (Chapter 9). Further work should focus on promoting the uniform reporting of coverage (and WISCA) to enable robust meta-analysis of antimicrobial resistance data and address best methods for dealing with small sample sizes expected at hospital-level and for stratified coverage estimates.