In this study, a total of 53 Enterococcus isolates isolated from pigeons were subjected to antimicrobial susceptibility testing. Out of the total 53 samples, 15 were collected from the tracheal region and 38 from the cloacal region. According to MALDI-TOF identification, the Enterococcus isolates consisted of 13 Enterococcus faecium, 21 Enterococcus faecalis, 7 Enterococcus durans, 1 Enterococcus mundtii, 4 Enterococcus gallinarum, 2 Enterococcus hirae and 5 Enterococcus columbae (Supplementary Table 1). More than half of the isolates (50.9%) originated from the Dél-Alföld region of Hungary. The majority were recovered from homing pigeons (39.6%), with most birds classified as juvenile pigeons (as defined by <6 weeks of age) accounted for 39.6% of the sampled population and 58.5% of owners maintained medium-sized flocks (51–100 birds). The regional distribution of the isolates is presented in Figure 1. Descriptive data on sample origin, age distribution, flock size by region, and the proportion of MDR isolates are summarized in Supplementary Tables S2–S4.

Following MIC determination, a comprehensive frequency table was generated (Table 1), which includes both MIC₅₀ and MIC₉₀ values for each antibiotic tested. This table provides a quantitative overview of the antimicrobial susceptibility distribution across the entire isolates population and facilitates interpretation by aligning the values with CLSI breakpoints and EUCAST epidemiological cut-off values (ECOFFs).

MIC₉₀ values indicated relatively low resistance to imipenem (8 μg/mL). While neomycin also showed a low MIC90 (64 μg/mL), the absence of an established clinical breakpoint for Enterococcus spp. limits interpretation. In contrast, a more variable susceptibility pattern was observed for other agents, where at least 50% of isolates were susceptible to amoxicillin, amoxicillin-clavulanate, doxycycline, potentiated sulfonamide, and vancomycin, based on MIC₅₀ values.

Importantly, comparison with ECOFFs revealed that 90% or more of the isolates were wild-type for neomycin, while imipenem also showed a wild-type profile in over 50% of cases. These findings not only corroborate the results from categorical resistance classifications but also provide a finer resolution of MIC trends within the population.

Table 1 was included to offer a detailed antimicrobial profile, identify potential therapeutic options, and highlight the antibiotics that may still be effective against commensal isolates. The inclusion of this table also allows comparison with future surveillance datasets.

Detailed MIC data at the isolate level are made available in the Supplementary material (file to enhance) transparency and reproducibility.

Figure 2 presents the distribution of Enterococcus isolates classified as resistant or susceptible to each antimicrobial agent based on established CLSI breakpoints. This figure was included to visually summarize resistance prevalence across the tested antibiotic panel and to identify key agents with high or low effectiveness within the population.

Pearson correlation analysis was employed to explore potential associations between resistances to different antibiotics, offering insights into co-selection patterns and shared resistance mechanisms. The corresponding heatmap (Figure 3) was selected for its ability to visually represent the strength and direction of these relationships. Notably, strong positive correlations were detected between amoxicillin and amoxicillin-clavulanate (0.64), doxycycline and potentiated sulfonamides (0.63), and florfenicol and tylosin (0.54).

Based on phenotypic profiles, 62.3% of isolates met the criteria for MDR, and 3.8% classified as XDR, according to established definitions. No PDR isolates were observed.

Cluster analysis was conducted to identify natural groupings of isolates based on their resistance profiles, helping to uncover patterns not immediately evident through individual antibiotic comparisons. The resulting K-means clustering identified three distinct clusters (Figure 4), each characterized by dominant resistance phenotypes that suggest differing selection pressures or genetic backgrounds. Cluster 1 was primarily defined by high resistance to doxycycline, suggesting potential overuse or cross-resistance with other tetracyclines. Cluster 2 was dominated by enrofloxacin resistance, consistent with the widespread use of fluoroquinolones in intensive poultry production. Cluster 3 was most strongly associated with amoxicillin-clavulanate resistance.

The inclusion of this analysis contributes to the study’s findings by demonstrating heterogeneity within the sample set, reinforcing the notion that resistance dynamics may differ between production systems or bacterial lineages.

Network graph analysis was employed to examine co-resistance patterns among antimicrobials, offering a visual representation of how frequently resistance co-occurred across isolates (Figure 5). This approach helps to identify potential co-selection mechanisms or linked resistance determinants that may be transmitted together.

Edge thickness in the graph reflects the frequency of co-resistance between two antibiotics, providing an intuitive map of resistance clustering. Tylosin emerged as a central node, showing the highest number of co-resistances, particularly with florfenicol (n = 33), enrofloxacin (n = 25), and doxycycline (n = 23). These frequent pairings suggest shared resistance pathways or common usage in poultry production that may drive joint selection.

Conversely, imipenem- and vancomycin-resistant isolates appeared as peripheral or isolated nodes, indicating that resistance to these agents is rare and does not typically co-occur with resistance to other compounds.

The decision tree, which consisted of four hierarchical levels, effectively distinguished MDR from non-MDR isolates. This supervised machine learning method enables the stepwise stratification of isolates based on their resistance profiles, helping to pinpoint variables most predictive of MDR.

The tree grew to a maximum depth of four, balancing model interpretability and classification accuracy. Among the 15 tested antibiotics, resistance to florfenicol, potentiated sulfonamide, enrofloxacin, and amoxicillin emerged as the most discriminative features, sequentially splitting the dataset into MDR and non-MDR subsets. However, a 10-fold cross-validation yielded an overall accuracy of 75%, precision of 72%, and recall of 78%, suggesting reasonable internal consistency despite the sample size limitation.

This finding suggests that resistance to these compounds — particularly when occurring in combination — strongly correlates with broader resistance phenotypes. The model’s predictive structure offers practical insight into field diagnostics: identifying resistance to just a few of these antibiotics may serve as an early warning indicator for MDR emergence.

To complement the empirical findings and assess the expected baseline of MDR, a Monte Carlo simulation was performed using 10,000 iterations (Figure 6). This stochastic modeling technique enabled the estimation of MDR prevalence under conditions of random resistance distribution, while preserving the real dataset’s size and antibiotic count.

The simulation yielded an average MDR rate of 59.39%, with a normal distribution curve reflecting consistent simulation behavior. The 95% confidence interval ranged from 50.0 to 69.0%, which provides a robust benchmark against which the observed MDR rate (62.3%) can be compared.

Figure 6 thus serves as a statistical validation of the MDR rates, reinforcing the conclusion that strategic interventions are needed to mitigate further selection and spread.

We also compared our findings with publicly available human antimicrobial resistance data to assess potential overlaps and One Health relevance (Figure 7). For aminopenicillins, the resistance profile of our isolates resembled that of E. faecalis from human clinical settings, which typically shows low resistance levels. This similarity may indicate shared susceptibility patterns between poultry- and human-associated isolates of this species.

However, a stark contrast was observed with E. faecium, where nearly all human isolates (98.5%) displayed aminopenicillin resistance, highlighting the species-specific nature of resistance development. This divergence suggests that E. faecium poses a higher therapeutic challenge in clinical practice and may reflect distinct evolutionary or usage pressures in human versus veterinary contexts.

Notably, none of the poultry-derived isolates showed resistance to aminoglycosides, a finding that diverges from certain human clinical isolates, where aminoglycoside resistance is more prevalent.

Vancomycin resistance, a major concern for both animal and human health, was identified in 15.1% of our poultry isolates. When species-level human data were considered, only 1.0% of E. faecalis isolates were resistant, whereas resistance was significantly higher (43.6%) among human E. faecium isolates.

The Enterococcus isolates (n = 53) analyzed in this study were recovered from a total of 660 oropharyngeal and cloacal swab samples collected in 2022 with the assistance of veterinary practitioners attending healthy domestic pigeons (Columba livia domestica), maintained in private lofts for sport, for meat or ornamental purposes. This represents a recovery rate of 8.0%. The term “domestic” here refers to birds bred and kept under human care, as opposed to feral or wild-living pigeons. Sampling was performed on a random basis, with the specific aim of maximizing geographic coverage across Hungary to enhance the representativeness of the collected isolates. Oropharyngeal and cloacal swab samples were collected during routine diagnostic examinations. The samples were obtained from pigeon populations located in Észak-Magyarország (n = 3), Észak-Alföld (n = 2), Dél-Alföld (n = 8), Közép-Magyarország (n = 5), Közép-Dunántúl (n = 1), Dél-Dunántúl (n = 1), and Nyugat-Dunántúl (n = 2). The isolates were obtained as pure cultures derived from the collected oropharyngeal and cloacal samples. Sampling was performed using Amies-type (carbon-free, aluminum shaft) swabs (Biolab Zrt., Budapest, Hungary). We spread the samples on m-Enterococcus modified agar (Merck KGaA, Darmstadt, Germany) for the identification of the Enterococcus genus, following the manufacturer’s specifications, which indicate that only Enterococcus species exhibit growth on this medium, with colony colors ranging from light pink to red. The species identification of the strains was determined using MALDI-TOF mass spectrometry (Flextra-LAB Kft., Budapest, Hungary) and Biotyper software version 12.0 (Bruker Daltonics GmbH, Bremen, Germany, 2024) (42). The isolates were stored at −80 °C using the Microbank™ cryopreservation system (Pro-Lab Diagnostics, Richmond Hill, Canada).

Human antimicrobial resistance data were provided by the National Public Health Center of Hungary. Resistance to ampicillin was used as the comparator in human cases, while amoxicillin was used for veterinary isolates. For third-generation cephalosporins, ceftriaxone was selected for comparison. Aminoglycoside resistance data were aggregated across gentamicin, tobramycin, and amikacin; additional specific data for neomycin were also available. Fluoroquinolone resistance was assessed as a group in human isolates, whereas enrofloxacin was evaluated separately for veterinary samples. Both national and region-specific human data were shared in Excel format with official authorization from the Chief Medical Officer and included percentage resistance rates.

Phenotypic resistance was evaluated by determining MIC values according to the guidelines of the Clinical Laboratory Standards Institute (CLSI) (43). Breakpoints were interpreted using CLSI standards [35], and compared with epidemiological cut-off values (ECOFFs) published by the European Committee on Antimicrobial Susceptibility Testing (EUCAST) (44). The choice to apply CLSI breakpoints was based on their widespread usage in veterinary antimicrobial susceptibility testing and the broader availability of CLSI interpretive criteria for the antimicrobials tested. While EUCAST clinical breakpoints are increasingly adopted in European human healthcare settings, CLSI remains a practical and accepted standard in animal studies. EUCAST ECOFFs were included to complement the clinical interpretation with an epidemiological assessment, distinguishing wild-type from non-wild-type isolates and providing a broader perspective on emerging resistance trends. This dual approach allows for both clinically and epidemiologically relevant interpretations, and has been applied in previous veterinary AMR surveillance studies (45, 46).

In the case of neomycin, no CLSI or EUCAST breakpoint is available for Enterococcus spp., thus we referred to a meat-product–based threshold of 1,024 μg/mL reported in the literature (47). Similarly, for tylosin, an 8 μg/mL breakpoint was adopted from published sources (48).

The bacterial isolates stored at −80 °C were revived 1 day before testing by suspending them in 3 mL of cation-adjusted Mueller-Hinton broth (CAMHB; Biolab Zrt, Budapest, Hungary), followed by incubation at 37 °C for 18–24 h. MIC testing was performed using 96-well microtiter plates (VWR International, LLC., Debrecen, Hungary). Except for the first column, each well was filled with 90 μL of CAMHB. Stock solutions of 1,024 μg/mL of each tested antimicrobial agent (Merck KGaA, Darmstadt, Germany) were prepared according to CLSI protocols (49).

Amoxicillin and amoxicillin-clavulanate were dissolved in phosphate buffer (2:1 ratio; pH 7.2, 0.01 mol/L), while imipenem was dissolved in phosphate buffer at pH 6. Doxycycline, neomycin, tylosin, and vancomycin were dissolved in distilled water. For potentiated sulfonamides (trimethoprim-sulfamethoxazole, 1:19 ratio), sulfamethoxazole was dissolved in hot water with a few drops of 2.5 mol/L NaOH, and trimethoprim was dissolved in 0.05 mol/L HCl-containing distilled water. Enrofloxacin was dissolved in distilled water with a few drops of 1 mol/L NaOH, and florfenicol was dissolved in 95% ethanol diluted with distilled water. All antimicrobial agents used in this study were obtained from Sigma-Aldrich (Merck KGaA, Darmstadt, Germany).

A 180 μL aliquot from a 512 μg/mL stock solution (diluted 1:1 with CAMHB) was pipetted into the first well, followed by a two-fold serial dilution across columns. After the 10th column, excess volume was discarded to maintain 90 μL per well. The bacterial inoculum was adjusted to 0.5 McFarland standard using a nephelometer (ThermoFisher Scientific, Budapest, Hungary), and 10 μL was added to each well starting from column 11 and proceeding backward (43).

MIC values were read using the Sensititre™ SWIN™ system (ThermoFisher Scientific, Budapest, Hungary) and interpreted with VIZION System Software v3.4 (ThermoFisher Scientific, Budapest, Hungary, 2024). MICs were determined once per isolate, following CLSI guidelines, as is standard practice in large-scale antimicrobial surveillance studies. The reference quality control strain used was E. faecalis ATCC 29212. This strain is recommended by CLSI guidelines for antimicrobial susceptibility testing of Enterococcus spp. and was employed to ensure the accuracy and reliability of MIC determinations throughout the study.

All statistical analyses were performed using R version 4.2.2 in the RStudio environment (50). The Shapiro–Wilk test was used to assess normality. For non-normally distributed data, non-parametric tests were applied. Differences in antimicrobial resistance across categories were evaluated using the Kruskal–Wallis test (51), which does not assume normality and is suitable for comparing medians across multiple groups. Post hoc pairwise comparisons were carried out using Mann–Whitney U tests (52), and t-tests, with Bonferroni correction applied to adjust for inflated p-values due to multiple comparisons (53). The potential risk of increased Type II error (false negatives) associated with the Bonferroni correction was taken into account.

Correlation analysis between antimicrobial agents was visualized via heatmaps using the corrplot (v0.92) and pheatmap (v1.0.12) R packages. Hierarchical clustering was performed using the cluster (v2.1.4) package, visualized with factoextra (v1.0.7) and dendextend (v1.16.0).

To identify patterns of co-resistance, network graph analysis was conducted using the igraph (v1.3.5) and ggraph (v2.1.0) packages. In the network graph, each node represented a specific antimicrobial agent, and edges were drawn based on pairwise co-resistance patterns derived from binary resistance data (resistant = 1, susceptible = 0) across isolates. The thickness of edges corresponded to the number of isolates exhibiting simultaneous resistance to both agents, highlighting the frequency of co-occurrence. Overall, the network graph complements the statistical analyses by revealing structural patterns of resistance interaction, further substantiating the selective pressures shaping the resistance landscape in the studied region.

For MDR isolates prediction, a decision tree model was built using rpart (v4.1.16), evaluated with caret (v6.0.93), and visualized using rpart.plot (v3.1.2). Given the relatively small sample size, the decision tree analysis was used as an exploratory tool to identify patterns rather than to produce definitive predictive metrics. In the decision tree, the input variables were binary resistance profiles (0 or 1) for each antibiotic, and the target variable was the classification of isolates as MDR or non-MDR based on established criteria. The model identified which antibiotics’ resistance status was most predictive of MDR and used sequential splits to stratify isolates accordingly.

Monte Carlo simulations were employed to estimate the prevalence of MDR isolates using 10,000 bootstrap iterations. The boot (v1.3.28), dplyr (v1.1.0), and ggplot2 (v3.4.0) packages were used for resampling, aggregation, statistical summary, and visualization. Each iteration randomly assigned resistance phenotypes across antibiotics within the constraints of the actual data (same number of isolates and antibiotic variables), allowing estimation of an expected MDR distribution under random conditions. This served as a baseline to assess whether the observed MDR prevalence was likely driven by non-random, external factors such as antibiotic usage patterns.

All analyses were conducted using open-source R packages to ensure full transparency and reproducibility. Antibiotic susceptibility profiles (antibiograms) were compiled from laboratory results and visualized using Microsoft Excel (Figures 6, 7).