A tool to enhance antimicrobial stewardship using similarity networks to identify antimicrobial resistance patterns across farms

Cecilia Aguilar‐Vega, Caterina Scoglio, María J. Clavijo, Rebecca Robbins, Locke A. Karriker, Xin Liu, Beatriz Martínez‐LópezScientific 2023 · DOI

TL;DR

Researchers have developed a network-based visualization tool that maps antimicrobial resistance (AMR) patterns across swine farms. By treating farms as nodes in a "similarity network," the system identifies clusters of farms with nearly identical resistance profiles, allowing for targeted biosecurity interventions and more precise antibiotic stewardship.

What Problem Does This Paper Address?

Antimicrobial resistance (AMR) is a "One Health" crisis, impacting humans, animals, and the environment. In the intensive swine industry, high animal density creates a perfect storm for the emergence and spread of resistant bacteria.

The challenge for veterinarians and production managers is that AMR data is often viewed in isolation—one farm or one pathogen at a time. This "siloed" view makes it difficult to see broader trends or identify "hub" farms that might be inadvertently spreading resistant strains through animal movements. The authors sought to create a holistic tool that translates complex laboratory data into actionable maps of resistance across entire production systems.

Methods

The researchers employed Similarity Networks, a technique more common in social science or genomics than in veterinary medicine.

  1. Data Binarization: Phenotypic resistance data (Minimum Inhibitory Concentrations) for seven key swine pathogens were converted into a binary format: 0 for susceptible and 1 for resistant/intermediate.
  2. Monoplex Networks: For each pathogen, they created a network where each node is a farm. An "edge" (connection) was drawn between two farms if their resistance patterns were statistically similar, calculated using Euclidean distance.
  3. Thresholding: To prevent the network from becoming a "hairball" of too many connections, they developed a thresholding criterion that retained only the most significant similarities while keeping the network's core structure intact.
  4. Community Detection: They used the Louvain algorithm to group farms into "communities" that share nearly identical AMR profiles.
  5. Centrality Measures: They calculated metrics like betweenness centrality to identify "hub" farms—those that act as bridges between different resistance communities and are therefore critical points for intervention.
  6. Multiplex Networks: Finally, they layered these individual pathogen networks into a multilayer (multiplex) network to see if resistance patterns for one bug (like E. coli) overlapped with others (like Salmonella).

Datasets

The study utilized a robust dataset from a large swine production system in the United States:

  • Scope: 710 initial animal holdings, narrowed down to 479 farms for the final multiplex analysis.
  • Pathogens: Seven major bacterial threats, including E. coli, Salmonella spp., and Streptococcus suis.
  • Antimicrobials: Up to 17 different antibiotics across 10 classes (e.g., penicillins, tetracyclines, macrolides).
  • Timeline: Data spanned from late 2018 through the end of 2021.

Key Insights and Ideas

  • The Density of Resistance: Pathogens like Actinobacillus suis and Bordetella bronchiseptica formed highly dense networks. This means their resistance patterns are remarkably uniform across the entire system, suggesting these bugs might have "innate" resistance or that a single resistant clone has already become dominant.
  • The "Hub" Farm Strategy: By identifying farms with high betweenness centrality, managers can find the "connectors." If a new resistance pattern emerges, these are the farms where biosecurity and movement restrictions will have the highest impact on stopping the spread.
  • Pathogen Independence: One of the most surprising findings was the low overlap in the multiplex network. A farm’s resistance profile for E. coli was a poor predictor of its profile for Salmonella. This suggests that resistance is often pathogen-specific rather than a general "farm-level" characteristic, requiring tailored stewardship for different diseases.

Key Results

  • Resistance Leaders: E. coli and Salmonella spp. isolates showed the highest overall AMR levels, while A. suis and G. parasuis remained the most susceptible to common treatments.
  • Community Profiles: The tool successfully identified distinct sub-groups. For example, in the S. suis network, "Community 25" was 99% susceptible to clindamycin, while almost every other community in the system showed over 80% resistance.
  • Network Morphology: The researchers observed that "denser" networks (where resistance is widespread and uniform) actually had fewer "hubs," whereas less dense networks (like E. coli) had clear, high-centrality nodes that are easier to target for clinical management.

Limitations and Future Work

The authors acknowledge several hurdles:

  • Clinical Bias: The data came from clinical isolates (sick pigs) rather than random sampling, which might overrepresent resistance levels.
  • Missing Data: Not every farm tested for every pathogen or every antibiotic. While they used statistical imputation to fill gaps, complete datasets would yield more precise maps.
  • The "Missing Link": The study lacked metadata on geographical location and animal movement "flows." The next logical step is to overlay these AMR similarity networks onto actual physical transport maps to see if trucks and pig movements are the primary drivers of these resistance clusters.

By integrating these network tools into routine veterinary surveillance, the industry can move from reactive treatment to proactive, data-driven stewardship.

Summary updated

Abstract

Antimicrobial resistance (AMR) is one of the major challenges of the century and should be addressed with a One Health approach. This study aimed to develop a tool that can provide a better understanding of AMR patterns and improve management practices in swine production systems to reduce its spread between farms. We generated similarity networks based on the phenotypic AMR pattern for each farm with information on important bacterial pathogens for swine farming based on the Euclidean distance. We included seven pathogens: Actinobacillus suis, Bordetella bronchiseptica, Escherichia coli, Glaesserella parasuis, Pasteurella multocida, Salmonella spp., and Streptococcus suis; and up to seventeen antibiotics from ten classes. A threshold criterion was developed to reduce the density of the networks and generate communities based on their AMR profiles. A total of 479 farms were included in the study although not all bacteria information was available on each farm. We observed significant differences in the morphology, number of nodes and characteristics of pathogen networks, as well as in the number of communities and susceptibility profiles of the pathogens to different antimicrobial drugs. The methodology presented here could be a useful tool to improve health management, biosecurity measures and prioritize interventions to reduce AMR spread in swine farming.
BibTeX
@inproceedings{aguilarvega2023tool,
  author = {Cecilia Aguilar‐Vega and Caterina Scoglio and María J. Clavijo and Rebecca Robbins and Locke A. Karriker and Xin Liu and Beatriz Martínez‐López},
  title = {{A tool to enhance antimicrobial stewardship using similarity networks to identify antimicrobial resistance patterns across farms}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-29980-4},
}
Figure