Improving empiric antibiotic prescribing by applying a Bayesian decision theory approach to phenotypic and genomic resistance data.

Lead Research Organisation: University of Bristol
Department Name: Cellular and Molecular Medicine


The emergence of antimicrobial resistance (AMR) is one of the greatest challenges currently facing modern medicine (O'Neill, 2014). AMR is partly driven by inappropriate antimicrobial (AM) use, hence, one element of the UK's AMR strategy it to focus on reducing inappropriate AM prescribing (UK five year national plan, 2019). Although in-roads have been made by various national and local AM stewardship initiatives, the majority of AMs that are prescribed are initiated empirically for presumed bacterial infection in the absence of a microbiological culture with associated AM sensitivities. The national stewardship guideline (Start Smart then Focus, 2015) requires a review of AM use at 48 to 72 hours, when the AMs can be stopped, continued, switched to an oral option or changed, dependent on microbiological results and clinical response.

The initial empiric AM choices for the management of bacterial infections are informed by an understanding of local phenotypic AMR patterns from blood culture isolates, and other sample types which are collected by all acute NHS trusts. While this local phenotypic resistance data is wildly available, it is currently an underused resource due to a lack of understanding of circulating AMR mechanisms, the interplay between mechanisms and how this data can be used to guide prescribing.

Amongst the most common AM decisions hospital clinicians make are "which oral antibiotic is optimal following successful treatment with this intravenous (iv) antibiotic?" and "what is my second line choice of antibiotic given failure of initial treatment?" Currently these decisions are informed by the local resistance rates to antibiotic X, rather than the local resistance rate of antibiotic X given the presumed resistance/sensitivity to antibiotic Y.
We intend to apply a Bayesian decision theory approach to our existing local phenotypic resistance data, and to verify and improve this model using existing and prospectively collected regional genomic data to inform decision making regarding antibiotic switches. We will initially focus on interplay of resistance mechanisms between commonly used iv antibiotics (piperacillin/tazobactam and third generation cephalosporins) with aminoglycoside antibiotics and three commonly use oral antibiotics (cotrimoxazole, ciprofloxacin, and coamoxiclav). We have selected these antibiotic combinations to address the most pressing day-to-day clinical concerns.
We will also determine the minimum size of dataset required for robust results, and therefore the extent to which this approach can be applied to smaller groups of patients that are heavily exposed to antibiotics such as Haematology and bone marrow transplant (BMT) patients, were antibiotic selection can be problematic. By fully utilising our understanding of local AMR patterns we can maximise the chance of successful antimicrobial treatment and minimise the inappropriate use of antibiotics. Once developed this approach could be rapidly applied to other regions of the UK (and internationally), using existing phenotypic data with or without the support of a local genomic surveillance program. As resistance genes spread over time the importance of understanding the co-dependencies of these genes to the management of patients will inevitably increase. This project will be a useful and timely addition to our understanding of this important problem.


Technical Summary

The aim of this project is to provide a Baysian predictive tool that can be used by clinicians to decide on the best empiric choice of antimicrobial (AM) in a variety of clinical scenarios. The basis of this model will be phenotypic antimicrobial resistance (AMR), patient outcome, and whole genome sequence data for bloodstream infections. Severn pathology lab and Winpath LIMS serves 3 NHS trusts generating about 2000 Gram-negative isolates per year from blood cultures with associated AMR and outcome data. We will collect up to 1000 of these isolates. Using the data extracted from LIMS, clinical, biochemical and microbiological data can be associated with each sample's unique identifier. These isolates will be sent in batches for WGS to the MicrobesNG facility in University of Birmingham.
We will receive monthly data report on blood cultures sent from hospitals sharing Winpath LIMS. We will use the data within the R statistics environment to apply a Bayesian predictive modelling approach with Bayesian network decision theory to our local phenotypic AMR data. We will derive predictive probabilities for resistance and outcomes by assessing the extent of correlations at multiple levels: a) to derive estimates of co-resistance between AMs; b) to model correlations with metadata; c) to understand temporal effects on the model.
Following collection of the WGS data and completion of the initial model we will compare our predictive model to our genomic data and integrate these results into our final model. This will be piloted at University Hospitals Bristol NHS Foundation Trust to assess and improve usability in a real-world context..

Planned Impact

The aim of this project is to improve empiric antibiotic prescribing, by optimising our use of phenotypic and genotypic resistance data.

When a pathogen can be isolated from a patient's blood cultures, or other significant sample types, AM therapy can be directed based on the susceptibility testing of that isolate. Unfortunately, less than 7% of blood cultures taken result in a clinically significant isolate being identified and in patients with sepsis only a third will have positive blood cultures. The majority of AM decisions are therefore empiric, taken before the isolate sensitivities are available or in the absence of a specific pathogen being identified. Even a small improvement in empiric prescribing will have a significant benefit to the UK population.

The scale of the problem
An estimated 147,000 people are admitted to UK hospitals each year with sepsis including 10,000 children1. Data from the USA suggests that 50% of all hospital patients will be taking an AM with 25% prescribed two or more at any given time. Point prevalence studies suggest about a third of patients in our hospital will be on AM treatment. The majority of these antibiotics have been prescribed empirically.

Direct Patient Benefits
Earlier treatment with the correct AM therapy has been demonstrated to reduced morbidity and mortality in sepsis. An incorrect empiric switch of AM therapy will deprive a patient of the benefit. Further empiric changes in AM therapy may be required, leading to longer hospital stays, increased risk of side effects and increased risk of C. difficile infection. The aim of this project is to minimise the probability of an incorrect empiric AM switch.

Economic Benefits
The estimated direct hospital costs of sepsis are £830 million per year with estimated indirect costs of £6.9 billion per year1 while infectious diseases account for 7% of UK deaths and cost an estimated £30 billion per year2. Improved empiric antibiotic selection will reduce both the direct and indirect costs associated with bacterial infection as described above.
The tools we generate will have economic value, and follow on projects could lead to a fully developed Clinical Decision Support Application.

Protection of the Global Antibiotic Resources
Pipeline new AM agents are limited and will be finite, hence every effort needs to be made to preserve this important resource. Any AM use will promote AMR. This is likely to be mediated through the exposure of a patient's endogenous flora to AM agents and the consequent evolutionary pressure on the patient's microbiome. Correct AM use drives resistance, but with an associated patient benefit, while the wrong AM selection will drive resistance with no associated benefit. The UK national AMR strategy 2019-2024 requires a 1% reduction in AM use, year on year. The successful completion of this project will make a significant contribution to that target by reducing inappropriate prescribing.

The secondary outputs of the project include the comprehensive, regional, WGS data set of invasive gram-negative isolates. This will be an important resource for both existing projects on AMR in the region and for planned future work on virulence factors by the Avison group. This data could be of great benefit in a wide range of other key research areas, in particular the understanding of gram-negative sepsis, and working towards the national target of reducing gram-negative sepsis rates by 50%.

1 .


10 25 50