How much is enough? Optimising sampling frames for genomic surveillance of Escherichia coli and Klebsiella spp. bloodstream infections – a retrospective study
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Optimising sampling frames for genomic surveillance of E. coli and Klebsiella may support interventions to mitigate bloodstream infections (BSIs), but approaches to estimating sample size and how these relate to bacterial population diversity at multiple genetic levels (strain/plasmid/antimicrobial resistance genes [ARGs]) are lacking.
Methods
Using systematically collected, regionally-stratified, whole genome sequencing data from a 6-month genomic survey of English E. coli / Klebsiella BSI isolates (n=1,939; NEKSUS), we used Bayesian approaches to estimate the sample sizes required to capture genetic diversity at strain-(MLST, fastBAPS clusters), plasmid- and ARG-levels, using two diversity measures: ‘coverage’ (i.e. proportion of the bacterial population/plasmids/ARGs represented by features observed in the sample), and the number/proportion of unique features observed at varying surveillance sample sizes. Our method is available as a user-friendly R Shiny application, EpiSENTRY ( https://01a0c3dd-90d2-e597-58a5-51391a468387.share.connect.posit.cloud/ ).
Findings
Randomly sampling 1,400 E. coli / Klebsiella BSI isolates each achieved ≥80% coverage of bacterial lineages and ARGs (i.e. capture features constituting ≥80% of the population) at 95% certainty, but captured lower proportions of unique features (28% MLSTs, 59% fastBAPS clusters, 19% plasmid subcommunities, and 57% ARGs for E. coli; 49%, 87%, 50% and 83% for Klebsiella , respectively). Sample sizes required for 80% coverage of E. coli plasmid populations were higher than MLSTs (1,928[95% CrI:1,837–2,025] vs 1,351[1,230–1,472]), while for Klebsiella , estimates were lower for plasmids (1,115[1,052–1,181] vs 1,422[1,331–1,515] for MLSTs). Lower sample sizes adequately captured 80% coverage of fastBAPS clusters and ARGs (649[575 – 727] and 27[23–30], respectively, for E. coli ; 113[84–145] and 83[72-96] for Klebsiella ), due to feature-specific frequency distributions. Klebsiella BSIs were more diverse than E. coli , with sampling 10.9% vs 3.2% of total annual BSIs in England needed to reach 80% MLST coverage. Plasmid populations and Klebsiella MLSTs were region-specific, emphasising the need for regionally-representative sampling.
Interpretation
Bayesian analysis facilitates estimates of context-specific genomic surveillance sample sizes. Cost-effective targets for coverage require further optimisation.
Funding
National Institute for Health and Social Care Research (NIHR) Health Protection Research Unit in Healthcare Associated Infections and Antimicrobial Resistance (NIHR207397); UKHSA and NIHR Oxford Biomedical Research Centre (BRC); UKHSA PhD Funding Competition; Robertson Fellowship; cloud compute infrastructure for this work was donated by Oracle Corporation Infrastructure.
Research in context
Evidence before this study
A MEDLINE (Ovid) search for genomic surveillance sample size estimation (“sample size estimation/calculation” AND “surveillance”; no restrictions on publication type, date or language) retrieved 62 records (22 included). Most studies (68% [15/22]) used calculations based on frequentist formulae, which require a user-specified arbitrary minimum frequency threshold for the feature/disease under surveillance. Surveillance outcomes were mostly prevalence/incidence (59% [13/22]) and presence/absence (54% [12/22]) of diseases/genomic features, and all studies considered diseases/features individually. Re-analyses of COVID-19 variant surveillance in high-income settings suggested that one third of the sequencing effort would have sufficed. Several publications highlight estimates should be tailored to specific surveillance goals. No studies extended surveillance sample size calculations to capturing population diversity of multiple genomic features simultaneously (e.g. plasmids, antimicrobial resistance [AMR] genes [ARGs]), although these may be important for bacterial genomic surveillance.
Bayesian estimation is a powerful method of statistical inference where prior beliefs about the distribution of a population parameter (e.g.: the frequency of a certain genomic feature) can be updated to a more realistic posterior distribution using the likelihood of the available data, thus incorporating uncertainty from potentially noisy data.
Added value of this study
We applied Bayesian analysis to large-scale genomic data from a 6-month nationally-representative survey of English bloodstream infection-associated E. coli and Klebsiella to estimate the posterior frequency distribution, with uncertainty, of four genomic features (multi-locus sequence type [MLST], fastBAPS cluster, plasmid subcommunities, ARGs). We used estimates to inform meaningful minimum feature frequency thresholds for conventional sample size calculations, by relating frequency to proportion of population diversity captured (i.e.: coverage – the proportion of the population represented by the features detected at least once in the sample), as well as number/proportion of unique features captured. This method overcomes limitations of ecological estimators when scaling to large populations by using all survey data to inform estimates, while placing more weight on epidemiologically important common features. We have made our method available as a user-friendly R Shiny application, EpiSENTRY ( https://01a0c3dd-90d2-e597-58a5-51391a468387.share.connect.posit.cloud/ ), allowing users to easily estimate sample sizes from their own data.
Implications of all the available evidence
Randomly sampling 1,400 E. coli and Klebsiella bloodstream (BSI) isolates, each, in England sufficiently captured strains and ARGs representing 80% of the estimated total bacterial population (80% coverage) at 95% certainty. While this captured the most common BSI-associated bacterial genomic features, >51% of unique MLSTs, >50% of plasmid subcommunities, >13% of fastBAPS clusters, and >17% of ARGs may remain undetected even at high coverages because there are so many unique variants at all genomic levels. Target coverage thresholds may need to be adjusted, or alternative targeted sequencing methods used (e.g. selection based on phenotype) for rare features of critical public health importance (e.g. carbapenemases). Klebsiella BSIs were more diverse than E. coli , relative to incidence, requiring 10.9% vs 3.2% of BSIs to be sequenced to capture comparable diversity. Plasmid populations and Klebsiella MLSTs were also region-specific, emphasising the need for regionally-representative sampling. Sample size estimates should be context-specific to best support national, regional, or local surveillance, and potentially inform infection prevention and control (IPC) efforts. Cost-effectiveness analyses are required to optimise sampling strategies.
