Wastewater Surveillance of Antimicrobial Resistance in Human Populations: A Systematic Review

Wastewater-based surveillance of antimicrobial resistance (AMR) may facilitate convenient monitoring of population-level AMR prevalence without the healthcare-associated bias and data collection restrictions inherent to clinically oriented systems. However, differences in study design and methodology likely contribute to differences in outcomes and interpretation, limiting reproducibility, reliability and meta-analysis. We therefore systematically reviewed studies using wastewater for AMR surveillance in human populations to identify optimal practices to detect wastewater-human AMR correlations. We evaluated 7,063 records and 174 full-text methods in a two-stage screen; 20 studies were included. Risk of bias assessment divided studies into high-risk (n=3), low-risk (n=3) and unclear-risk (n=14). Most studies detected wastewater-human AMR correlations (n=15) but only six studies identified statistically significant associations, most via culture-independent approaches (n=5). Genomic approaches also facilitated higher-resolution AMR monitoring whereas culture-based studies primarily undertook observational comparisons of specific organisms and phenotypic AMR profiles. Studies identifying wastewater-human AMR correlations were consistently associated with sampling wastewater influent irrespective of other methodological approaches. For longitudinal studies, a timeframe of >=6 months was similarly associated. Most influent studies identifying wastewater-human AMR correlations used composite (n=5) or flow-proportional wastewater sampling methods (n=4); however, grab sampling was commonest overall (n=6) and generally appeared similarly effective.Wastewater-based surveillance of AMR in human populations appears relatively robust, with most included studies reporting a correlation despite high diversity in study design and methodology. Our review supports sampling of wastewater influent using composite sampling (at a minimum) as a standard. Impacts of other methodological approaches are less clear; however, a minimum timeframe of six months for longitudinal studies, and increased sampling coverage for culture-independent studies to enable adequate biostatistical analyses appear sensible. As this relatively new field grows, more studies with clear wastewater-based population-level AMR surveillance aims are needed to better determine the impact of confounding features and validate comprehensive “best practice” protocols.


Introduction
Antimicrobial resistance (AMR) is a significant threat to global health 1 and the treatment of infectious diseases, and is compounded by diverse drivers 2,3 facilitating emergence and spread. AMR surveillance is therefore critical to understanding trends and monitoring interventions, as prioritised in the World Health Organisation's global AMR action plan 4 .
Large networks dedicated to sharing AMR data have been established to meet this need, including the European Antimicrobial Resistance Surveillance Network (EARS-Net) and the Global Antimicrobial Resistance Surveillance System (GLASS). However, current surveillance methods are limited by the reliance on patient-level sampling, which is often affected by selection bias towards healthcare-associated settings 5 ; this does not reliably capture AMR prevalence in commensal organisms and in the community, thought to silently constitute most of the true AMR burden [6][7][8] . Additionally, reliance on routine clinical microbiology results often restricts data collection to a limited subset of culturable species, and focuses predominantly on susceptibility phenotypes with limited genotyping. This lack of genotyping hampers the surveillance of high-risk AMR-associated clones and the horizontal transfer of AMR determinants 9 .
Wastewater-based AMR surveillance has the potential to avoid biases in current surveillance methods by simultaneously sampling both healthcare-and community-associated populations 10 . The approach has already been successful in illicit drug monitoring 11 and pathogen (particularly enterovirus) surveillance 12,13 but its application to AMR surveillance is relatively new 7 . Recent wastewater AMR studies have investigated seasonal/geographic AMR distributions 14 , quantified global abundances of AMR genes 6 and identified correlations between wastewater and clinical AMR surveillance data 15 . However, differences in methodology and study design among wastewater AMR studies likely contribute to differences in outcomes/interpretations. For example, the relative impact of wastewater sampling approaches on study outcomes are not well understood, including the effects of grab sampling (i.e. taking single samples at one timepoint) in collecting homogenous solids/unrepresentative samples 16 , or sampling in the presence of unrepresentative, contaminating AMR-associated point sources. Difficulties in standardising AMR testing across traditional surveillance networks 9 could be, for example, circumvented by using metagenomic sequencing to universally probe wastewater resistomes 17,18 . However, despite increasing research in this area, there has not been an attempt to review the data available and assess the remaining knowledge gaps.
We systematically reviewed studies using wastewater for AMR surveillance in human populations, seeking to identify practices that could optimise identification of the correlation of wastewater-human AMR for surveillance purposes. This review examines both study design and methodology to identify the relative potential impact of these factors on study outcomes, and to highlight any limitations and recommendations for future research.

Materials and Methods
For this systematic review, we adopted the "Population Intervention Comparator Outcome" (PICO) framework using the following domains: Wastewater, antimicrobial resistance, bacteria and public health surveillance/methods. The search string was developed through iterative preliminary searches in consultation with a librarian experienced with systematic reviews; subject headings/operators were adapted for each bibliographic database (full search strings are presented in Supplementary dataset 1). Searches were conducted on 01/02/2019 in: Ovid MEDLINE (1946-present), Ovid EMBASE (1974-present), Ovid Global Health (1973-present), Ovid CAB Abstracts (1973-present), Scopus and Web of Science Core Collection. Searches were updated on 21/10/2019 using identical search strings. Search results were limited to the English language and de-duplicated.
Study titles/abstracts were screened (Fig.S1) to determine if the study: (i) was primary research, (ii) collected human-associated wastewater, (iii) reported AMR prevalence, and (iv) compared a wastewater dataset against another non-wastewater dataset of relevance.
Studies performing self-defined wastewater-based surveillance of AMR (i.e. studies explicitly using wastewater for population-level AMR surveillance) were purposely included using parallel criteria. If it was unclear whether a study met all criteria based on title and abstract alone, the study was passed onto the next stage. For studies passing the initial screen, fulltext methods were reviewed and studies were included if they: (i) analysed wastewater samples from a wastewater treatment works (WwTW) and represented a human population in their non-wastewater dataset, or (ii) conducted a self-defined wastewater-based AMR surveillance study as part of the study aim.
All studies were screened by a single reviewer (KKC); two other reviewers (LB, NSi) independently re-screened a random 10% subset of retrieved records to estimate Cohen's Kappa score 19 as a measure of inter-reviewer reliability beyond chance. A score above 0.75 was interpreted as representing excellent agreement beyond chance 20,21 . Consensus was reached by discussion in the case of disagreements.
For included records, data were extracted by a single reviewer (KKC) and checked by a second (LB) using a pre-tested data extraction form piloted on five random included records (Supplementary dataset 2), including: general characteristics (publication year, geographic location), study design (study aim, sampling strategy, sampling point, sampling methods, AMR detection methods, human component, sample sizes) and outcomes (results of wastewater-human comparison with respect to AMR). Statistical methods or modelling approaches were also recorded if used, and a summary of relevant findings and additional points of interest as judged by the reviewer were documented.
Included studies underwent a risk of bias assessment by two independent reviewers (KKC, LB). Risk of bias was assessed using a qualitative, modified approach based on the Cochrane risk of bias tool 22 addressing five bias domains (see below); this focused on systematic methodological differences as reported outcomes were highly diverse. Selection bias was defined as differences introduced across sampled sites using different sampling methods.
Performance bias was defined as differences introduced across comparison groups using different AMR detection approaches. Attrition bias was defined as differences introduced by incomplete outcome data (e.g. incomplete data entry across comparison groups or longitudinally). Reporting bias was defined where outcomes may have been measured but not reported/disproportionately reported. Lastly, other bias was defined as confounding due to the likely presence of AMR/AMR-influencing sources which were not considered and may have affected outcomes. Studies at "high-risk" of bias were those with inconsistencies in sampling, AMR detection methods, measurement and reporting of outcomes; "low-risk" studies broadly maintained consistency across comparators. If information present was insufficient to assess risk of bias, the classification "unclear" was assigned. Discrepancies were resolved by discussion, and an overall qualitative measure (high, low, unclear) was assigned to each study.

Results
Summary of included studies 7,063 de-duplicated records were retrieved, 174 full-text methods were reviewed, and 20 studies were included (Fig.1). Inter-rater reliability (on n=701 studies) was supported (Cohen kappa score=0.76) 20,21 ; screening conflicts were observed for 27 studies. 9/20 studies were self-defined wastewater AMR surveillance studies; 11 studies performed wastewater AMR surveillance indirectly by investigating: AMR transmission into wastewater environments (n=6); AMR transmission routes to humans (n=3); and epidemiological links between wastewater and human bacterial isolates (n=2). Study summaries are presented in Tables 1 (aims, overview and overall risk of bias) and 2 (methodology).
Three studies were judged at high-risk of bias, three at low-risk and 14 had an unclear-risk ( Studies assigned an unclear risk of bias did not provide sufficient information to evaluate bias, such as partial description of sewer inputs or absent descriptions of sampling methods. Amongst included studies, samples were obtained from 67 countries, although most (48/67) were represented as part of a single global study 6 (Fig.2 . Three studies observed no relationship between wastewater and human AMR, sampling either influent via flow-proportional sampling, effluent with unreported methods, or drainage ditches via grab sampling. The remaining two studies were inconclusive and sampled influent and/or effluent with unreported methods (Fig.3A).

Longitudinal sampling and study timeframes
Thirteen studies conducted longitudinal sampling with total sampling timeframes ranging from 3-14 months (median: 12 months, IQR: 6-12). 10/13 studies reported sampling intervals, ranging from two weeks to six months (median: one month, IQR: 1-1.5). The remaining studies did not conduct longitudinal sampling (n=6) or were unclear in their approach (n=1).
Eight longitudinal studies identified wastewater-human AMR correlations, with the median timeframe and sampling interval range being 12 and 0.5-6 months respectively. The remaining five reported no correlation or were inconclusive; the median timeframe and sampling interval range in these studies were 11 and 0.5-2 months respectively. All snapshot studies and one study with unclear timeframe also identified wastewater-human AMR correlations (Fig.3B).
Nine studies performed direct wastewater-human AMR comparisons where the human population sampled directly contributed to the WwTW sampled. Six studies performed indirect comparisons where the human population sampled did not contribute to the WwTW sampled. One study undertook both evaluations, and four were unclear in this respect. For direct studies, 5/9 reported wastewater-human AMR correlations; however, all six indirect studies reported wastewater-human AMR correlations.
Most culture-based studies did not perform any statistical analyses (10/11) with the exception of one study 34 which modelled the relationship between resistance rates (defined in the study as the proportion of resistant isolates) in wastewater (WwTW influent and hospital effluent) and clinical isolates (hospital and primary care). For this study, a significant correlation was found between mean resistance rates in: (i) hospital wastewater and hospital clinical isolates, and (ii) WwTW influent and primary care urine isolates. This study also observed mean resistance rates were lower in influent compared to urine isolates (twofold difference in five antibiotics) whereas similar rates were found between hospital wastewater and hospital clinical isolates. An exception was cefadroxil resistance which was more prevalent in influent than in primary care isolates. Despite the cefadroxil difference, rates of ESBL-positive isolates were significantly higher in clinical isolates across settings.
Observational wastewater-human AMR relationships were reported for eight studies comparing wastewater isolates to: clinical isolates 16,26,29,31,35,36 (n=6), national AMR data 24 (n=1) and faecal isolates from healthy carriers 27 (n=1). Among clinical isolate studies, the most common observation was the sharing of drug resistance to specific antimicrobial classes 16,26 , multi-drug resistance 35 , specific resistance profiles 26,36 , ESBL genes 36 and drugresistant species 29 , between wastewater and clinical isolates. Two clinical isolate-based studies also observed a correlation between wastewater isolates cultured and study country outbreaks and/or endemic circulation of resistant clones 26,31 . The study investigating faecal isolates from healthy carriers 27 reported correlation of the top four resistance phenotypes (ciprofloxacin, tetracycline, erythromycin, fosfomycin) between wastewater and faecal isolates, but with higher prevalence of resistance to these four antibiotics, multi-drug resistance and vancomycin-resistant enterococci in wastewater. Lastly, a comparison with national resistance data showed increasing resistance in wastewater isolates reflected increases in resistance reported for clinical blood isolates 24 .
The remaining two culture-based studies reported no wastewater-human correlation, or an inconclusive outcome. Both these studies investigated gentamicin resistance but in different species (Enterococcus spp. and E. coli). The study 32 focussed on Enterococcus spp. found high clonal diversity and low genetic relatedness amongst cultured strains using PFGE, contrasting with the observation that AMR phenotype at the strain level matched national levels. The E. coli-focussed study 23 found no evidence of strain similarity using PFGE and observed no association in gentamicin resistance gene prevalence between clinical and wastewater isolates.

Direct qPCR-based approach (n=2)
Two studies directly detected AMR genes using qPCR of 229 and eight resistance genes respectively. Only the study 15 targeting 229 resistance genes performed statistical testing but both reported a relationship between wastewater AMR and national AMR data.
The study targeting 229 resistance genes compared a longitudinal survey of 13 European WwTWs across seven countries with contemporaneous surveillance data from EARS-Net.
The relative abundance of influent AMR genes clustered significantly based on high versus low national antibiotic consumption, with AMR gene distributions mirroring EARS-Netdescribed north-to-south and west-to-east geographic gradients. A higher relative abundance of most AMR gene classes was observed in high antibiotic consumption countries except for tetracycline and macrolide-lincosamide-streptogramin B resistance genes. The study also performed counts of antibiotic-resistant culturable bacteria from samples but showed no significant correlation between counts and AMR gene quantification.
The second qPCR-based study reported relationships between wastewater samples in Tunisia and Spain, and respective national AMR surveillance data. Intra-country comparisons showed the most commonly recovered beta-lactamase gene (blaTEM) and quinolone resistance genes qnrS and qnrA in wastewater were also most commonly reported in clinical surveillance. This was not however true for CTX-M-9 group ESBL genes, where a higher wastewater prevalence was observed. Inter-country comparisons highlighted higher AMR gene detection in samples from Spain coinciding with higher antibiotic use, and high mecA prevalence thought to be potentially consistent with high-pig farming densities.
Biostatistical evaluations of sequencing data were performed in all (6/7) but one study 37 .
Four studies reported a significant relationship between wastewater and human AMR using WGS (n=3) and metagenomics (n=1). Of these, two WGS studies compared wastewater isolates (E. coli and E. faecium) with same-species bacteraemia isolates using core genome phylogenies. In the E. faecium study 38 , wastewater and clinical isolates were phylogenetically interspersed, indicating mixing, with divergence events implying recent local emergence and dissemination. Each WwTW sampled also contained genetically diverse populations but these remained comparable to diversity in national bloodstream infection isolates. Network analysis showed geographic clustering of WwTWs and bloodstream isolates, and linked bloodstream isolates to 9/20 WwTWs (three without hospital sewer input). The E. coli study 39  Another WGS study 28 reporting a wastewater-human AMR association compared wastewater with community-acquired UTI ESBL-E. coli isolates from residents in the WwTW catchment, identifying significantly higher single-and multi-drug resistance in urine isolates compared to those from wastewater even when adjusting for the clinically common ST131.
Both community and wastewater samples shared blaCTX-M-15 as the most commonly identified ESBL gene and shared similar prevalence of other bla genes (blaOXA-1, blaTEM-1B) except for blaCTX-M-14 which was more prevalent in wastewater.
The final study 6 reporting a wastewater-human AMR correlation conducted a large-scale cross-sectional global wastewater survey and compared resistomes to multiple epidemiological variables. This study showed total AMR gene abundances from a given site/country correlated with sanitation and general health metrics, with up to 89% of observed resistome variation significantly explained regardless of within-sample AMR gene diversity. A high human development index (i.e. generally high-income countries) was linked to significantly lower abundance of AMR genes in wastewater metagenomes.
The remaining 3/7 sequencing-based studies reported no relationship 33 between wastewater and human AMR or inconclusive results 33,40 . One inconclusive study conducted a metagenomics-based longitudinal surveillance study 40 comparing resistomes generated from drainage ditch wastewater and contemporaneous local surveillance data consisting of household morbidity and healthcare usage, including diarrhoea/fever cases and clinical faecal culture. Significant increases/decreases in resistome read abundance were not associated with increases in reported diarrhoea or clinical faecal isolates. However, nonsignificant increases in pathogen read abundances appeared to coincide with increased reported illness/clinic visits, possibly reflecting issues with methodological sensitivity (i.e. impact of sequencing depth) to detect significant changes.
The second inconclusive study was a metagenomic and functional metagenomic longitudinal evaluation of wastewater and human faecal samples collected from individuals within the WwTW catchment 33 . Characterisation of resistomes, including AMR gene network analysis, showed significantly higher phylogenetic species diversity and abundance of AMR proteins in wastewater (WwTW influent, street-access wastewater) compared to faecal samples.
Faecal samples were significantly enriched for drug efflux mechanisms whereas wastewater samples contained more aminoglycoside acetyltransferase, class D β-lactamase, and dihydrofolate reductase genes. However, extensive sharing of sulphonamide-resistance conferring AMR genes was detected, indicating a possible resistance-specific relationship.
Lastly, a longitudinal study 37 reported low to no relationship between wastewater and human AMR based on its comparison of clinical and WwTW effluent carbapenem-resistant P. aeruginosa using PFGE, multilocus sequence typing (MLST), WGS to evaluate acquired carbapenemase genes, and culture-based AST. PFGE pulsotypes showed very limited overlap between clinical and wastewater samples (n=4). However, WGS showed dissemination of the blaVIM-2 carbapenemase gene in clinical and wastewater samples.

Interactions between study features
When assessing the distribution of study features relative to the study outcome, wastewater sampling point and AMR evaluation approach, studies identifying wastewaterhuman AMR correlations were consistently associated with sampling influent (Fig.3). All wastewater sampling methods also identified wastewater-human AMR correlations when influent was sampled, with composite sampling associated to studies identifying correlations regardless of AMR evaluation approach (Fig.3A). For longitudinal studies, a timeframe of >=6 months was most closely associated with the identification of a wastewater-human AMR correlation when sampling influent or sludge whilst using cultureor qPCR-based AMR evaluation (Fig.3B). The presence of sewer inputs with distinct wastewater types appeared to cluster with studies identifying a wastewater-human correlation where these sampled influent, irrespective of AMR evaluation approach (Fig.3C).
No human dataset type was exclusively associated with a given outcome/sampling point.
However, clinical isolate studies were common amongst those identifying wastewaterhuman AMR correlations, irrespective of sampling points (although most sampled influent or influent and effluent; Fig.3D). For additional plots see (Figs.S6-8).

Antibiotic prescribing and residue measurements
Five studies investigated the relationship between antibiotic prescribing/measured concentrations and AMR prevalence. National prescribing/usage was associated with qPCR-ARG abundances 30,41 and individual resistance phenotypes 32 , but not metagenomic-ARG abundances 6 . Local clinical prescribing was associated with clinical resistances 36 while primary care prescribing showed no association with qPCR-ARGs. Antimicrobial residue measurements were not associated with either qPCR-15 or metagenomic-ARG 6 abundances.

Discussion
In our review, most (15/20, 75%) studies reported that wastewater AMR reflected AMR in human catchment populations, but the extent to which this relationship could be evaluated was dependant on study design, sampling strategies, and genotypic versus phenotypic testing approaches; testing approach also determined outcome quality.
As anticipated, we found WwTW influent is the most population-representative sampling point; previous studies have described transformation of microbial and AMR composition during treatment 42 . Sampling influent appears to be the single most important feature associated with identifying a human-wastewater AMR correlation (Fig.3, Figs.S6-8).
However, several studies identifying such a correlation analysed treated wastewater, indicating transformed samples may remain useful for wastewater surveillance, potentially dependent on treatment process. However, only 3/8 non-influent studies reported WwTW treatment types, and evaluating the impact of these remains important as specific treatments may favour specific species/AMR determinants 43 .
Composite and 24H flow-proportional sampling may be better than single grab sampling given that wastewater composition changes significantly over short timescales 44 and individual samples may be "flooded" by homogenous solid material 16 . However, several influent studies using grab-like composites with 2-3 sub-samples in quick succession, and non-influent studies using single grab samples, also demonstrated a human-wastewater AMR correlation. Grab sampling, which was the most commonly used method, is convenient and avoids significant autosampler-associated workload and capital costs. Despite potential sample homogeneity, half of all studies identifying a wastewater-human AMR correlation used grab samples in combination with culture-based/phenotypic approaches (Fig.3A). Grab samples may therefore be representative for such analyses although further research is needed to characterise the extent to which a single grab can accurately reflect temporal flux in AMR.
Longitudinal studies demonstrating a human-wastewater AMR correlation typically included longer sampling intervals (six months versus two months for negative/inconclusive studies) and timeframes. For the highest resolution sampling campaigns 26,40 which used two week sampling intervals across different timeframes, one study observed an association between ampicillin-resistant wastewater and contemporaneous clinical isolates over 12 months, whereas the other found changes in metagenomic read abundances were not matched by changes in contemporaneous clinical surveillance over three months. Directly comparing these studies is difficult due to different methodologies, but the findings would support the fact that studies sampling over a timeframe of ≥6 months are more likely to capture associations with human population-level AMR (Fig.3B).
The unexpected finding that human-wastewater AMR correlations were not observed in some cases where the wastewater being sampled received direct inputs from the population being surveyed can likely be explained by study aims and methods. Most of these direct studies aimed to show fine-scale links in small catchment areas via biostatistical analyses whereas most indirect studies compared broad (often national) observations without statistical testing. Only 2/6 indirect studies reported significant relationships compared to 4/9 direct studies. Smaller catchments also limit sample sizes potentially obscuring signals from detection which may have contributed to unclear/negative outcomes 40 . When comparing direct studies, similar sampling methods and test approaches are seen, but amongst studies identifying wastewater-human AMR correlations, more WwTWs and higher numbers of wastewater samples were sampled per study (Fig.S3). This suggests higher sampling coverage may aid capturing fine-scale changes in smaller catchments.
The presence of specific sewer inputs from potentially major sources of AMR in wastewater (e.g. hospitals) was interestingly associated with positive outcomes (Fig.3C), and likely linked to whether the AMR mechanisms studied were hospital-associated/emerging or already widely disseminated in the community. For example, one study 39 sampled WwTWs with and without hospital input, and found the most clinically-prevalent E. coli ESBL gene was also present in all WwTWs -indicating widespread prevalence across settings. Another study 23 focussing on E. coli gentamicin resistance in hospital effluent, WwTW influent and domestic wastewater, found significantly lower prevalence in domestic wastewater than in both influent and hospital effluent -indicating likely nosocomial emergence. Only measuring hospital-associated wastewater in this context would therefore confound community AMR estimates. Importantly, half of all studies did not report sewer inputs at all, making interpretation of findings more difficult; a similar lack of reporting was observed for WwTW PE and flow estimates (Fig.S4).
Culture-dependent and independent test approaches were both associated with the identification of wastewater-human AMR correlations (culture-based studies: 9/11, cultureindependent studies: 6/9). However, for the six studies that identified statistically significant positive correlations, most used culture-independent approaches (n=5). Culture-based approaches appear useful for reporting observational wastewater-human AMR correlations, but studies employing culture-independent methods are relatively species-and mechanism-agnostic, and may be more tractable and statistically robust for public health surveillance purposes. Combination approaches may capture the breadth and detail of genetic relationships between bacterial strains and important AMR genes.
For culture-based studies the evaluation of overlap between wastewater and human isolates is limited to comparisons of phenotypic AMR profiles, which may reflect significant differences in underlying genotypes, particularly for Gram-negative organisms such as Enterobacterales. Most studies 25,28,32,34,35 showed AMR prevalence in wastewater was lower, but common AMR phenotypes were frequently shared, indicating culture/phenotype-based surveillance of wastewater may be useful for low-resolution monitoring of AMR in clinical isolates. Targeted culture-based approaches could also pick up relevant high-risk drugresistant clones (MRSA, VRE, CRE, MDR-Salmonella Typhimurium DT104) 26,27,29,35 . Increasing AMR prevalence in wastewater isolates appeared to mirror national clinical AMR rates in one study 24 , although this study was subject to reporting bias.
Culture-independent approaches could be used for more detailed analyses such as exploratory ordination, phylogenetics, and statistical modelling. As a result, comparisons of relative AMR gene abundance, genomic relatedness and evolution, and resistome variation could be reported. In-depth genomic data may be more useful for higher-resolution AMR monitoring in clinical isolates and human carriage, but at a higher resource cost. In comparisons with prescribing data, targeted qPCR methods appeared to perform better than metagenomics which may reflect sensitivity trade-offs that occur with lower sequencing depth. However, neither approach correlated with wastewater antimicrobial concentrations; this may however be linked to detection limits and sample variability 6,15 .
This review has some limitations. Our highly comprehensive search strategy meant that screening could not be carried out in duplicate; however, the risk of single reviewer bias was mitigated by validating the screening strategy with three reviewers on 10% of records. We excluded non-English publications, potentially missing some relevant studies. 11/20 included studies were not self-defined wastewater surveillance studies, and so the relevance of their findings to the systematic review question was based on reviewer judgement. Studies were highly diverse in reported design/outcomes, complicating standardisation for narrative synthesis. Finally, the premise of our review was based on the assumption that wastewater-human AMR correlations occur, and to identify study approaches which best characterise this phenomenon, but it may be that in some cases there is genuinely no association.

Conclusion
Overall, wastewater-based surveillance of population-level AMR appears relatively robust, with most included studies reporting wastewater-human AMR correlations despite high diversity in study design, methods and metadata. Our review suggests sampling of influent and composite sampling optimise the chance of identifying human-wastewater AMR correlations and are most suitable for wastewater-based AMR surveillance studies. Impacts of timeframe, comparison type and the number of WwTWs sampled were less clear based on the available data, however a minimum timeframe of six months appears sensible since longitudinal changes may be clearer over longer timescales. Biostatistical analyses likely benefit from increased sampling coverage of the catchment area by including more WwTWs and larger numbers of wastewater samples. A culture-independent approach facilitates agnostic detection and statistical synthesis but is more resource-intensive and may be limited in sensitivity by low sequencing depths; culture-based approaches however do enable low-resolution clinical surveillance and targeted detection of specific clones and/or AMR phenotypes. Studies should clearly report sewer inputs, WwTW features (e.g. PE, flow rates, treatment methods), sampling and sample processing methods, sampling intervals and timeframes. As this relatively new field grows, more studies with clear wastewaterbased population-level AMR surveillance aims are needed to better determine individual feature contributions and confirm comprehensive "best practice" protocols.

Figure 2: Geographic distribution of wastewater sampling and test approach of included studies
Centroids of countries sampled by included studies are plotted with colours and shapes according to citation and test approach respectively. World Bank regional coverage by study was as follows: East Asia and Pacific (n=1), Europe and Central Asia (n=13), Latin America and the Caribbean (n=2), Middle East and North Africa (n=7), North America (n=1), South Asia (n=1), Sub-Saharan Africa (n=2). Non-Aarestrup et al. 2019 studies are plotted with jitter around the centroid for the map focussing on Europe.

Figure 3: Interactions between study features in relation to outcome, sampling point and AMR evaluation.
Number of studies with a specific study feature (A=wastewater sampling method, B = longitudinal timeframe, C = reported sewer inputs, D = human dataset used for comparison) plotted against study outcomes, wastewater sampling point and AMR evaluation approach. NAs represent studies without the plotted feature (e.g. snapshot studies for longitudinal timeframe) or did not report feature used. The bar heights represent the number of studies for a given comparison.

Relevant comparison Relevant outcomes Overall risk of bias
explained up to 89% of observed resistome variation between samples regardless of diversity of AMR genes. Countries with higher HDI have significantly lower abundance of AMR genes and the number of passengers flying into a country has no effect on abundance. No significant association between temperature at collection with AMR genes abundance. No significant effect was found when investigating total usage of all antimicrobials on the abundance of AMR genes on a class-level. No significant association was found between the abundance of AMR genes on a class-level and the antimicrobial residue levels measured. Hendriksen et al. 2019

Kenya
To monitor circulating pathogens and AMR genes via metagenomics of urban wastewater with comparison to other concurrent disease surveillance.
Non-limited Resistomes of wastewater compared to surveillance data and laboratoryconfirmed cases of wastewater source populations.
Increases/decreases in read abundances for bacteria/AMR genes were not reflected in the contemporaneous PBIDS data. Nonsignificant observed increase in read abundances of pathogens appeared to coincide with reported illness/visits to clinic.