Background/Objectives: The Japanese Veterinary Antimicrobial Resistance Monitoring System (JVARM) conducts longitudinal monitoring of antimicrobial resistance (AMR) in indicator bacteria from food-producing animals. For
Escherichia coli from healthy pigs, slaughterhouse-based sampling has been conducted for approximately a decade, yielding a substantial accumulation of MIC
[...] Read more.
Background/Objectives: The Japanese Veterinary Antimicrobial Resistance Monitoring System (JVARM) conducts longitudinal monitoring of antimicrobial resistance (AMR) in indicator bacteria from food-producing animals. For
Escherichia coli from healthy pigs, slaughterhouse-based sampling has been conducted for approximately a decade, yielding a substantial accumulation of MIC data. While JVARM reporting has traditionally focused on annual resistance proportions by drug, the availability of long-term data enables investigation of cross-drug relationships, including MIC similarity and co-resistance patterns. This study aimed to (i) identify the co-resistance structure among antimicrobial agents using MIC- and phenotype-based similarity measures and (ii) identify drug resistances most strongly associated with multidrug resistance (MDR).
Methods: We analyzed broth microdilution MIC data obtained annually for
E. coli isolates from healthy pigs in the JVARM program in Japan between 2012 and 2023. Antimicrobial resistance was classified from MIC results and annual resistance prevalence was calculated for each antimicrobial. For the co-resistance and MDR analyses, isolate-level data were pooled across the full study period. To identify co-resistance structure, we performed hierarchical clustering using (i) correlation-based similarity of MIC profiles and (ii) Jaccard similarity of binary resistance profiles (resistant/susceptible classification). Multidrug resistance (MDR; ≥3 antimicrobial classes) was further modeled using XGBoost with each drug resistance as a predictive feature, and feature contributions were evaluated using gain, permutation importance, and SHAP values. We also examined how SHAP-based attributions varied when the outcome definition was set to ≥1-, ≥2-, or ≥3-class resistance.
Results: Within the study period, resistance remained highest for tetracycline and moderate for streptomycin, ampicillin, sulfamethoxazole–trimethoprim, and chloramphenicol, whereas resistance to other agents was low. MIC-based correlation analysis revealed coordinated variation among ampicillin, sulfamethoxazole–trimethoprim, streptomycin, chloramphenicol, and tetracycline. Separately, Jaccard similarity of binary resistance profiles identified two closely positioned co-resistance groupings (Ampicillin/Streptomycin/Tetracycline and chloramphenicol/sulfamethoxazole–trimethoprim). Ampicillin was identified as the medoid in both MIC-based and resistance-profile similarity spaces, with streptomycin also positioned near the center in both structures. In the XGBoost model for MDR (≥3 classes), ampicillin resistance was consistently the highest-contributing feature when evaluated by gain, permutation importance, and SHAP. When we examined how SHAP-based attributions varied across outcome definitions (≥1-, ≥2-, and ≥3-class resistance), feature importance largely followed resistance prevalence at ≥1–≥2 classes (tetracycline highest) but shifted at ≥3 classes to ampicillin as the top feature.
Conclusions: Both MIC-based and phenotype-based analyses revealed co-resistance structures. Under the MDR definition used in this study, explainable machine-learning analyses showed that ampicillin resistance emerged as a leading resistance feature associated with MDR. Because these findings are associative rather than causal, further work will be needed to clarify mechanisms. These findings have important implications for antimicrobial resistance control in the Japanese pig sector, indicating that stewardship strategies may need to be tailored according to antimicrobial class and underlying co-resistance structure.
Full article