Expression of Interest Title: Data Challenge – Are antibiotic breakpoints globally consistent, does it matter if not?
Team 8923, led by Robert Beardmore of University of Exeter, UK, with Emily Wood (University of Exeter), Pablo Catalan (Universidad Carlos III, Spain) and Jon Iredell (Westmead Hospital/University of Sydney, Australia)
Expression of Interest:
Antibiotic resistance is quantified numerically by so-called breakpoints which, ideally, ensure pathogens are cleared by patients when treated at dosages above the breakpoint. Bodies that publish breakpoints determine their values by committee who use data, like distributions of pathogens’ minimal inhibitory concentrations (MIC) to delineate boundaries between putative drug-susceptible (S) and drug-resistant (R) subpopulations. Importantly for patients, human experts in these decision pipelines, EUCAST in Europe and CLSI (and, more recently, USCAST) in the US, make different decisions: CLSI breakpoints are typically higher than EUCAST’s. A key question is this: are these differences warranted, or not? To address this, we would like to examine whether breakpoints are consistent with the available MIC data in ATLAS both within and between continents. We also wish to understand whether the aforementioned expert committees might be supported by AI and Machine Learning that can use microbiology statistics, like MICs, to help determine breakpoints ‘automatically’, or else to estimate uncertainties in prior breakpoint decisions. We plan to explore this by examining EUCAST, CLSI and ATLAS MIC data together and there are several ways to complete our analysis. For instance, we know Enterobacterales species have the same EUCAST breakpoints for each antibiotic, and similar CLSI ones. ATLAS can help us answer, therefore, whether Enterobacterales species have statistically different MIC profiles in practise. If so, this would show breakpoint-setting policies are inconsistent with real-world MIC data. Our hope is that supervised classification algorithms can differentiate rationally between pathogen species and then be used to estimate what optimal breakpoints might be, for example using neural nets to model & extend prior CLSI and EUCAST decisions. As a multi-continental team of clinicians, microbiologists and data scientists, we expect that patients’ antibiotic treatments might, one day, be improved by our work, for example by providing rational tests for breakpoint policy inconsistencies by using the data found in ATLAS. At the very least, our project is a starting point for the use of AI methods in antibiotic breakpoint-setting policy.
Data Contributor: Pfizer Inc.