Global surveillance of antimicrobial resistance in food animals using priority drugs maps

The study addresses how to prioritize and geographically target surveillance and interventions against antimicrobial resistance (AMR) in food animals across LMICs, where systematic surveillance is sparse. With most antimicrobials globally used in animals, rising AMR threatens animal health and potentially human health. Existing efforts in high-income countries inform policy, but LMICs rely on point prevalence surveys (PPS) that typically use summary metrics and do not disaggregate by antimicrobial–bacterium combinations. The research aims to produce fine-scale maps of resistance prevalence for key drugs in Escherichia coli and nontyphoidal Salmonella in food animals, and to predict which antimicrobials are most likely to exceed critical resistance thresholds in the near future at 10 × 10 km resolution. This enables tailored surveillance and policy actions sensitive to geographic heterogeneity and co-resistance patterns.

The authors situate their work within global AMR surveillance efforts. Systematic surveillance in high-income countries has informed policies (e.g., restrictions on ceftiofur, fluoroquinolone bans) and human AMR burden has been estimated for 88 pathogen–drug pairs. However, LMIC animal AMR mapping has largely used PPS and aggregate indicators (e.g., fraction of drugs with resistance >50%), limiting targeted action by drug class. The WHO Medically Important Antimicrobials (MIA) list offers priorities but lacks geographic granularity. Prior global analyses (e.g., Van Boeckel & Pires 2019) identified hotspots across multiple pathogens but did not provide drug-specific maps for individual bacteria in animals. The paper builds on these by focusing on E. coli and nontyphoidal Salmonella and seven widely used, medically important antimicrobial classes, integrating co-resistance patterns and environmental/anthropogenic covariates.

Data: 1,088 point prevalence surveys (PPS) from LMICs (2000–2019) on AMR in commensal E. coli and nontyphoidal Salmonella from healthy food animals (cattle, pigs, poultry, sheep, horse, goat). Samples included meat, live animal swabs, food products, slaughter swabs, and feces. AST methods largely followed CLSI/EUCAST; breakpoint differences were adjusted per published methods. Drugs modeled: tetracycline (TET), ampicillin (AMP), sulfamethoxazole–trimethoprim (SXT), chloramphenicol (CHL), ciprofloxacin (CIP), gentamicin (GEN), cefotaxime (CTX). For mapping each drug’s prevalence, all PPS reporting that drug were used. For predicting priority antimicrobials, complete seven-drug profiles were required; 806 PPS with at least four drugs were included. Missing drug prevalences were imputed using multivariate imputation by chained equations (MICE), imputing 21% of 9,877 estimates while propagating uncertainty. Geospatial modeling: Prevalence of resistance for each drug was mapped at 10 × 10 km resolution using an ensemble (Gaussian process stacked generalization). Child models: boosted regression trees, LASSO logistic regression, and feed-forward neural networks, trained on environmental and anthropogenic covariates (e.g., national antimicrobial use in 2013/2020; disaggregated drug-specific AMU in 2020 by animal biomass; animal density; temperature; precipitation; pesticide application). Child predictions were stacked via Gaussian process regression fitted with INLA to capture residual spatial autocorrelation. Summary AMR metrics N10/N25/N50 were computed as the number of drugs (of 7) with resistance exceeding 10%, 25%, or 50% per pixel; hotspots defined using thresholds (e.g., N50 ≥ 3). Temporal trends: Logistic regressions estimated 2000–2019 temporal trends per drug (with one outlier removed for TET and AMP based on Box–Tidwell). These trends informed projections to estimate time to reach thresholds. Priority antimicrobials: Defined as the drug with the highest probability that its resistance prevalence will exceed a critical threshold (10%, 25%, or 50%) in the near future for each pixel. Procedure (illustrated for 50%): binarize each pixel’s 2015 resistance profile (1 if ≥50%, 0 otherwise). For drugs with 0, predict whether they will exceed 50% using LASSO logistic regression incorporating (i) co-resistance structure via the proportion of PPS supporting each subsequent resistance profile permutation and the current N50; and (ii) risk factors: relative and total AMU, animal density, climate variables, and other covariates. Repeat over all 64 permutations per target drug and apply to all pixels. Select the drug with the highest exceedance probability as the priority antimicrobial. Estimate time to threshold using fitted temporal trend models from the pixel’s current prevalence to 50%. Validation and uncertainty: Four-fold spatial cross-validation produced AUCs for exceedance prediction per drug. Uncertainty in prevalence maps came from posterior predictive variance. Priority antimicrobial uncertainty combined (a) 15 Monte Carlo runs of imputed datasets and (b) variance in Bayesian posterior predictions; uncertainty per pixel was the fraction of runs yielding a different priority drug than the final map.

- Weighted mean resistance prevalence across LMIC PPS: E. coli vs. nontyphoidal Salmonella, respectively: TET 59% (n=745) vs. 54% (n=597); AMP 57% (n=779) vs. 46% (n=632); SXT 45% (n=649) vs. 36% (n=501); CHL 35% (n=656) vs. 26% (n=553); CIP 30% (n=796) vs. 26% (n=624); GEN 28% (n=882) vs. 23% (n=650); CTX 33% (n=446) vs. 19% (n=334). - Temporal changes (2000–2019): +12% (TET), +33% (AMP), +19% (SXT), +20% (CHL), +16% (CIP), +11% (GEN), +37% (CTX); significant for all except TET (p=0.061). Significant increases by species: poultry (AMP, CHL, CIP, CTX) and pigs (AMP, SXT, CHL, CIP, GEN, CTX); not significant in cattle. - Hotspots (N50 ≥ 3): E. coli—southern/eastern China, central Asia, northern India, northern Brazil, Chile; nontyphoidal Salmonella—northeastern China. AMP resistance hotspot in E. coli in northern/eastern Brazil; SXT/GEN (E. coli) and CHL/CIP/GEN (Salmonella) hotspots in northeastern China. - Uncertainty: highest for CTX resistance in nontyphoidal Salmonella (mean SD ~19.9%), lowest for TET resistance in nontyphoidal Salmonella (mean SD ~4.5%). - Priority antimicrobials (50% threshold): Africa and South America—78% of locations prioritized TET or AMP; Asia—77% prioritized AMP or SXT (TET already ≥50% in 83% of locations). SXT priority regions included NE India, S and NE China, S Brazil, Turkey, Iran; AMP priority in N/W China, Mongolia, W India; CHL priority in S/E China. GEN and CTX prioritized rarely (0.02%) and scattered. - Time to exceed 50%: for AMP-priority pixels average 1.7 years (animal-biomass weighted); for CIP-priority pixels average 12.4 years. For TET, AMP, CTX priority areas, average times <7 years. - Model performance: AUCs for exceedance predictions ranged 0.880–0.994 across drugs. Co-resistance covariates most influential (ΔAUC 0.224–0.494 when removed). Environmental/anthropogenic covariates added little for TET/AMP (ΔAUC ~0.002 and −0.003) but improved others (ΔAUC 0.109–0.416). Key covariates included AMU, pesticide application, precipitation cycles, and night LST amplitude.

The study delivers fine-scale, drug- and pathogen-specific AMR maps for food animals across LMICs and identifies, per location, the antimicrobial most likely to cross critical resistance thresholds. This disaggregation addresses the limitations of prior PPS-based summaries and aligns animal AMR mapping more closely with human-focused burden assessments. Findings confirm known global hotspots (China, India, Turkey, Iran, Brazil, Chile) and reveal variations by bacterium and drug class, underscoring the need for tailored policies. E. coli generally exhibits higher resistance than nontyphoidal Salmonella. Regions like northeastern China show elevated resistance for multiple classes except tetracyclines and penicillins, which are already high globally, highlighting urgency to preserve remaining classes. Lower AMR in Africa is consistent with lower veterinary AMU. The priority antimicrobial maps indicate that in low-AMR regions, tetracyclines and penicillins are most likely to cross 50%, driven mainly by universal co-resistance patterns and widespread, accessible use of older drugs. In higher-AMR regions, sulfonamides and amphenicols are more likely to exceed 50%, with evidence of co-selection and persistence despite regulations. The predictive framework, supported by high AUCs, integrates co-resistance with environmental and AMU covariates, offering a practical basis for risk-based surveillance targeting and informing interventions to reduce AMU and improve farm biosecurity.

This work compiles 1,088 PPS to create high-resolution global maps of resistance prevalence for seven key antimicrobials in E. coli and nontyphoidal Salmonella in food animals across LMICs, and introduces a computational framework to identify, per pixel, the drug most likely to exceed critical resistance thresholds and the expected time to exceedance. The outputs can guide geographically targeted AMR surveillance and policy, prioritize drug classes for monitoring, and inform strategies to reduce AMU and enhance farm biosecurity in hotspot regions. Future research should expand to additional pathogen–drug combinations, incorporate more robust region-specific temporal trends as PPS data grow, integrate explicit policy and AMU datasets, consider serovar-level resolution for Salmonella, and refine thresholds and timelines under varying intervention scenarios.

Key limitations include: (1) Uncertainty from predictive mapping and imputation of missing resistance data, particularly for cefotaxime (missing in 51% of surveys; high SD of imputed values). (2) Restricted scope to seven drugs and two bacteria due to limited PPS reporting; Salmonella results not serovar-specific. (3) Pooling data across years for spatial mapping due to sparse spatio-temporal coverage, precluding full spatio-temporal modeling. (4) Assumption that resistance trends continue at past rates, not accounting for future AMU policy changes. (5) Lack of a comprehensive, up-to-date database of country-specific veterinary AMU regulations as covariates; policy effects considered only indirectly via AMU estimates (2013, 2020). (6) Choice of resistance thresholds (10%, 25%, 50%) is somewhat subjective; sensitivity analyses were performed. Despite these constraints, uncertainty was quantified via Monte Carlo imputation and Bayesian posterior variance, and model performance remained high in spatial cross-validation.