Antimicrobial resistance surveillance: can we estimate resistance in bloodstream infections from other types of specimen?

Background: Antimicrobial resistance (AMR) is a major global health threat. Standard approaches to AMR surveillance through susceptibility testing of isolates from blood cultures are difficult in low- and middle-income countries (LMIC), where lack of laboratory capacity prevents routine patient-level antimicrobial susceptibility testing, and systematic testing of invasive specimens may not be feasible. Other specimen types could provide an alternative but effective approach to surveillance, but the relationship between resistance prevalence in these and bloodstream infections has not been systematically evaluated. Methods: We used data from Oxfordshire, UK, 1998-2018, to investigate associations between resistance rates in Escherichia coli and Staphylococcus aureus isolates from blood and other specimens, comparing proportions resistant in each calendar year using time series cross-correlations, for multiple antibiotics. We also compared the proportion of resistant isolates from blood versus other specimens across drug-years, overall and across four arbitrary resistance categories (<5%, 5-10%, 10-20%, >20%). We repeated analysis across four high-income and 12 middle-income countries, and in three hospitals/programmes in LMICs. Findings: 8102 E. coli bloodstream infections, 322087 E. coli urinary tract infections, 6952 S. aureus bloodstream infections and 112074 S. aureus non-sterile site cultures were included from Oxfordshire. Resistance trends over time in isolates from blood versus other specimens were strongly correlated (maximum cross-correlation 0.51-0.99 with strongest associations between proportions in the same year for 18/27 pathogen-drug combinations). Resistance prevalence was broadly congruent across drug-years for each species, particularly allowing for uncertainty in estimation. 207/312 (66%) species-drug-years had resistance prevalence in other specimen types within +/-5% of that blood isolates, and 276/312 (88%) within +/-10%. 215/312 (69%) species-drug-years were in the same resistance category for blood and other specimen types; 305 (98%) were the same or adjacent resistance categorisation. Results were similar across multiple countries in high- and middle-income settings, and the three LMIC hospitals/programmes. Interpretation: Resistance in bloodstream and other less invasive infections are strongly related, suggesting the latter could be a surveillance tool for AMR in LMICs. These infection sites are easier to sample and cheaper to obtain the necessary numbers of susceptibility tests, thus providing more cost-effective evidence for decisions including empiric antibiotic recommendations.


Introduction 73
Antimicrobial resistance (AMR) is among the top ten global health threats, 1 and is 74 particularly acute in low-to middle-income countries (LMICs). [2][3][4] Surveillance is a key tool 75 to combat rising AMR, as highlighted by the World Health Organization (WHO) 5 and 76 formalised in its Global Antimicrobial Resistance Surveillance System (GLASS), 6,7 77 particularly in LMICs where lack of laboratory capacity prevents routine patient-level 78 antimicrobial susceptibility testing. 8 The lack of surveillance data was identified as a key 79 contributor to global AMR, 9 since the broad-spectrum empiric treatment approach that is 80 pragmatic in hospitals with limited laboratory facilities 2,10 generally results in 81 overtreatment. 11 Strengthening surveillance capacity in LMICs is therefore a major focus of 82 international initiatives. To date, these programmes have generally focused on improving 83 capacity for blood culture surveillance due to their high mortality; however, this requires 84 substantial capital investment and sustainable funding for LMICs, where baseline laboratory 85 functioning is poor. 12 Blood cultures are comparatively costly, require trained staff due to the 86 invasive nature of blood sampling, and can have slow turnaround times, reducing their ability 87 to inform individual clinical management. Additionally, the low positivity rate means that 88 very high sample throughput is needed to confidently estimate resistance prevalence. 89 Methods). We estimated yearly resistance prevalence for all antimicrobials with 133 susceptibilities for >65% samples in that year (mostly >80%). Testing used manual disk 134 diffusion before 2013, and thereafter automated testing with the BD Phoenix TM Automated 135

Materials and methods
Microbiology System, Beckton Dickinson. Resistance was as defined by the laboratory when 136 the test was done, using the same breakpoints regardless of specimen type, following 137 European Committee on Antimicrobial Susceptibility testing (EUCAST) recommendations 138 for bloodstream infections in each year. 20 We did not de-duplicate isolates by patient, 139 reflecting data that would be easily available from routine laboratory systems. 140 141 For each pathogen-drug combination, we first estimated the proportion resistant in blood 142 versus other samples in each calendar year (with 95% confidence intervals (CI)) using Lin's 143 concordance correlation coefficient (CCC) for comparisons. 21 We used time-series cross-144 correlation functions to identify the time differences at which the correlation between 145 resistance prevalence was strongest. As perfect agreement is unrealistic, we also considered 146 agreement within what might be considered clinically acceptable error (±5% and ±10% 147 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted October 14, 2020. . difference in prevalence), and agreement within arbitrary categories which may nevertheless 148 be clinically helpful, specifically: <5% resistance (most would readily prescribe an 149 antibiotic), 5-10% (most would prescribe for mild infections), 10-20% (unclear) and >20% 150 (many would not prescribe if other options were available). We used logistic random-effects 151 meta-analysis to estimate the overall difference between resistance prevalence in the two 152 sample types, treating every calendar year as an independent study, assessing heterogeneity 153 across years using the I 2 statistic. 22 We used meta-regression to estimate the effect of calendar 154 year on the proportion of resistant bloodstream isolates (with its standard error), assuming the 155 proportion of resistant isolates from other infection sites was known (given their greater 156 numbers). We also used meta-regression to directly estimate the association between log odds 157 of resistance in bloodstream infections (outcome, with its standard error) and the log odds of 158 resistance in other infection sites (explanatory variable, fixed). (Full details in the 159 Supplementary Material.) 160 161 One possibility is that resistance rates differ between blood and other infection sites because 162 of differences in the underlying populations being sampled. A secondary analysis therefore 163 considered matched pairs of isolates from the same patient, selecting the closest culture from 164 a different site up to 3 days before or 2 days after the blood culture, 15 assessing concordance 165 using McNemar's test. We also considered whether susceptibility of a previous E. coli UTI 166 could predict susceptibility of a subsequent E. coli BSI, by selecting the temporally closest 167 urine culture taken between 3-90 days before the blood culture. 15,23 168

169
We also analysed pathogen-drug combinations at the country level (pooling multiple 170 hospitals/regions) using the ATLAS dataset, comprising 633,820 isolates from 73 countries 171 between 2004-2017. 18 We considered all years where at least 30 isolates were tested for a 172 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted October 14, 2020. .
given drug in LMICs (95% CI width around prevalence always <37%); and 100 isolates for 173 HIC (width <20%) given greater numbers from HIC and strong inverse association between 174 numbers tested and resistance prevalence in HIC, suggesting preferential testing of resistant 175 isolates. Finally, we also analysed microbiology data from Angkor Hospital for Children 176  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . rates in the same year most strongly correlated (maximum cross-correlation at lag 0). For 198 these four antibiotics, there was no evidence of strong variation in the relationship between 199 resistance rates over time (Supplementary Table 1 Table 1). Overall resistance 207 prevalence in BSI was highly correlated with resistance prevalence in UTI (CCC=0.93 (95% 208 CI 0.91-0.95) ( Table 1). Agreement was also relatively high across our four pre-defined 209 resistance categories, with 83/132 (63%) drug-years in the same resistance category and 210 45/132 (34%) in adjacent categories (Figure 2, right-hand panel). Although numbers were 211 smaller, broadly similar results were seen for Klebsiella spp. (Figure 2 CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . proxy for blood culture surveillance. Agreement was poorer for gentamicin and tetracycline 229 resistance, although rates were low (<15%). As might be expected, agreement was slightly 230 better between resistance prevalence in blood and other sterile site cultures than blood and 231 non-sterile site cultures (Supplementary Figure 7). As agreement was reasonable in both, 232 and non-sterile site cultures would be more likely to be taken in LMICs, we focussed on this 233 group of infection sites. Overall, resistance prevalence was more similar in bloodstream and 234 other non-sterile infection sites in S. aureus than E. coli (Supplementary Table 1 Table 1). Overall resistance prevalence in blood was highly 243 correlated with resistance prevalence in non-sterile isolates (CCC=0.97 (0.95-0.97)) ( Table  244 1). Overall agreement between resistance in concurrent blood and non-sterile cultures within 245 a patient was extremely high (96-100%) for S. aureus, with ciprofloxacin, erythromycin and 246 oxacillin having <5% of resistant blood cultures with susceptible non-sterile cultures and 247 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . Table 5). Again, although numbers were smaller, broadly similar results were seen for S. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. .

Discussion 271
Here, we show that AMR prevalence from infections in sites other than blood could 272 potentially be used to infer AMR in bloodstream infections, which could reduce the initial 273 training staff to carry out antimicrobial susceptibility testing on blood cultures, arguing that if 294 this can be achieved, other types of culture will become possible too. However, setting up 295 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . such systems is very expensive, both in terms of infrastructural and running costs, and it will 296 be many years until sufficient data on blood cultures have been accumulated for surveillance 297 purposes, particularly given initial low uptake and low rates of positivity in these samples. 298 While longitudinal surveillance data is available at a limited number of research sites, these 299 programmes cover a limited geographic area and may not be representative of other locations 300 in a country. Incorporating blood culture testing into public hospitals may be further biased 301 by differences in ability to pay. 2 Collecting specimens from urine/skin surface/genital sites is 302 cheaper and less invasive than sampling blood, and therefore substantially easier to do on 303 large numbers from different communities, as well as having higher positivity rates, 304 increasing their utility for AMR surveillance. This is true even in HICs, with clearly narrower 305 confidence intervals from larger numbers of non-blood Oxfordshire isolates. Even allowing 306 for the greater manual processing needed for non-blood cultures, for example, obtaining 100 307 E. coli isolates would only require testing 361 urine samples assuming 37% are culture-308 positive and 75% of these are E. coli, but 7693 blood cultures, assuming 13% are positive and 309 10% of these are E. coli (estimates based on Oxfordshire data [1998][1999][2000][2001][2002][2003][2004][2005][2006][2007][2008][2009][2010][2011][2012][2013][2014][2015][2016]. Whilst blood 310 cultures may be perceived as having greater diagnostic utility, this may be limited practically, 311 e.g. by long turnaround times. 312

313
One strength of our study is the different datasets used, and the robustness of findings across 314 these, including continuous surveillance of one large region (Oxfordshire), and global 315 analysis (ATLAS) albeit of a selection of infections within each country. The 316 hospital/programme-level LMIC datasets are highly curated, but relatively small. One crucial 317 underpinning assumption is that pathogens causing other infections are representative of 318 pathogens causing BSI in terms of antibiogram, either because the source of infection is 319 commonly from commensal colonising organisms that become pathogenic opportunistically 320 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . or because both types of infection arise from a similar reservoir. Even though this may hold 321 within the populations studied, it is possible, although fairly unlikely, that this may not 322 generalise more widely. Generalisability is supported by the good country-level agreement in 323 ATLAS, even though pooling different regions within countries, which could have 324 potentially differed further between blood and other infection sites. This raises an important 325 limitation of our approach, namely that different people were sampled for blood and other 326 specimens. However, our analysis of samples from the same patients suggests this is unlikely 327 to cause major bias. Another limitation is that we have only considered surveillance of the 328 proportion of resistant infections, not numbers or population-level rates of resistance; 329 however, in terms of empiric treatment recommendations, proportions rather than absolute 330 numbers that are generally of interest. We have not considered surveillance of AMR across 331 the entire antibiogram, i.e. patterns of resistance within one isolate across different 332 antibiotics. We also pooled all samples together: in future work we could consider whether 333 relationships are generalizable between different groups of patients (e.g. by age, 334 nosocomial/community infections). Finally, the thresholds we considered were arbitrary, 335 which is why we considered two different approaches (varying margins of error and varying 336 categorical thresholds). 337 338 One question is how good the agreement between resistance rates in isolates from blood 339 versus other specimens should be for surveillance of resistance in samples from other body-340 sites to be useful -clinically and for understanding overall trends. In our dataset, different 341 resistance rates were observed in different sample types for many pathogen-drug 342 combinations, with some variation over time and less than perfect agreement. However, 343 cross-correlations were generally high, as was agreement within 10% and across categories 344 that could inform empirical antimicrobial therapy. Several biological reasons for the lack of 345 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . perfect agreement are plausible: for example, a large proportion of E. coli BSIs have a 346 urinary focus, but susceptibility and appropriate empiric treatment generally limit progression 347 to BSI, in contrast to resistant, and perhaps sub-optimally treated UTIs. Hence one might 348 expect the proportion of drug-resistant BSIs to be higher due to inadequate empiric treatment 349 of a drug-resistant UTI, but underlying time trends to track each other, as we observed. One 350 could ask whether blood is the best sample for AMR surveillance. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . and unwillingness of patients to have samples taken (often due to cost) have been identified 371 as key factors influencing antibiotic use. 32,33 LMICs face numerous challenges in setting up 372 surveillance systems similar to those in HICs, including lack of coherent governance, budget, 373 technical expertise, information technology systems and co-ordination. Our study shows that 374 body-sites which are easier to sample, cheaper, faster and easier to grow from compared to 375 blood could provide a feasible approach to AMR surveillance, providing evidence for empiric 376 treatment recommendations. They could also be amenable to survey-type approaches which 377 could be particularly valuable as setting up a local laboratory infrastructure for successful and 378 useful antibiogram production is challenging. Using population level surveillance as a 379 substitute for individual patient testing could be considered a pragmatic steppingstone in 380 improving a site's diagnostic capacity for individual patient management. Nevertheless, the 381 urgent threat posed by AMR demands that we not let the perfect be the enemy of the good. 382 383 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . https://doi.org/10.1101/2020 and that any discrepancies from the study as planned (and, if relevant, registered)  is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. .  is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. .

Figure 1 Resistance prevalence in E. coli and S. aureus in blood versus non-blood cultures in 548
Oxfordshire , 1998-2018 549 E. coli S CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. .

576
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. .

Figure 3 Resistance prevalence in E. coli in blood versus non-blood cultures in high-income 577
countries present in the ATLAS dataset, 2004ATLAS dataset, -2016 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. .

Figure 4 Resistance in blood and non-blood cultures (A) E. coli, (B) S. aureus, ATLAS 581
dataset split into high-income countries and middle-income countries 582

E. coli bloodstream infections versus non-blood 583
High-income countries 584 585 Middle-income countries 586 587 S. aureus bloodstream infections versus non-blood 588 High-income countries 589 590 591 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted October 14, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

Middle-income countries
The copyright holder for this preprint this version posted October 14, 2020. . . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

Figure 5 Resistance prevalence in E. coli in blood versus non-blood cultures in three
The copyright holder for this preprint this version posted October 14, 2020. .

621
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted October 14, 2020. . is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.