Shared Multidrug Resistance Patterns in Chicken-Associated Escherichia coli Identified by Association Rule Mining

Abstract

Using multiple antimicrobials in food animals may incubate genetically-linked multidrug-resistance (MDR) in enteric bacteria, which can contaminate meat at slaughter. The U.S. National Antimicrobial Resistance Monitoring System tested 21,243 chicken-associated Escherichia coli between 2004 and 2012 for resistance to 15 antimicrobials, resulting in >32,000 possible MDR patterns. We analyzed MDR patterns in this dataset with association rule mining, also called market-basket analysis. The association rules were pruned with four quality measures resulting in a <1% false-discovery rate. MDR rules were more stable across consecutive years than between slaughter and retail. Rules were decomposed into networks with antimicrobials as nodes and rules as edges. A strong subnetwork of beta-lactam resistance existed in each year and the beta-lactam resistances also had strong associations with sulfisoxazole, gentamicin, streptomycin and tetracycline resistances. The association rules concur with previously identified E. coli resistance patterns but provide significant flexibility for studying MDR in large datasets.

Keywords: Escherichia coli; antimicrobial resistance; association rule mining; foodborne bacteria; machine learning; multidrug resistance.

Figure 1

Frequent itemset generation from the Table 3 dataset using the Apriori algorithm. Each candidate itemset is represented by a black circle. The itemsets with a red slash are eliminated after examining the top row of itemsets and comparing the support of each itemset to the minimum support of 0.4. [B] has a support of 0.33 and therefore all of its supersets must have a support less than or equal to 0.33. All remaining itemsets in the second row are frequent with a support >0.4. Itemsets with a blue slash are eliminated after calculating the support for three-item itemsets.

Figure 2

Number of association rules and distribution of rule quality measures. Five rule quality measures were calculated: confidence (A), lift (B), phi (C), ralambrodrainy (D), and support (E). Frequency polygons for each year-source dataset are shown R: retail; S: slaughter. The number of rules before (all rules) and after (best rules) pruning with confidence >0.75, support >0.01, lift >2, and absolute value of phi >0.5 are shown in (F). Ralambrodrainy was not used for pruning rules because a very small cut-off (>0.005) resulted in too few rules for analysis.

Figure 3

Rule overlap and cumulative rule stability. Rule overlap is the proportion of rules shared between two datasets (the number of rules that are in both datasets divided by the total number of rules within the datasets). When calculated for consecutive years, rule overlap is plotted on the earlier year. Cumulative rule stability (A) averages the proportion of rules shared across all previous years. (B) is rule overlap between slaughter and retail isolates for a given year and (C) is rule overlap between consecutive years, calculated separately for slaughter and retail isolates.

Figure 4

Decomposed rule graphs for Escherichia coli isolated from chicken carcasses at slaughter. The best-rules identified in each year (confidence >0.75, support >0.01, lift >2, phi >0.5) were decomposed into nodes (antimicrobials) and edges connecting the antecedents to the consequent. Nodes are colored based on antimicrobial class (bright green = beta-lactams; purple = aminoglycosides; yellow = sulfonamides; red = tetracycline; blue = fluoroquinolones; dark green = phenicols; pink = macrolides). Node size is proportional to node degree (number of other connected nodes). Edge thickness is proportional to the number of rules involving each pair of antimicrobials and edge darkness is proportional to the average phi for those rules.

Figure 5

Decomposed rule graphs for Escherichia coli isolated from chicken retail meat. The best-rules identified in each year (confidence >0.75, support >0.01, lift >2, phi >0.5) were decomposed into nodes (antimicrobials) and edges connecting the antecedents to the consequent. Nodes are colored based on antimicrobial class (bright green = beta-lactams; purple = aminoglycosides; yellow = sulfonamides; red = tetracycline; blue = fluoroquinolones; dark green = phenicols). Node size is proportional to node degree (number of other connected nodes). Edge thickness is proportional to the number of rules involving each pair of antimicrobials and edge darkness is proportional to the average phi for those rules.

Figure 6

Distribution of association rule quality measures under the null hypothesis (H₀) of no associations. The percent of rules under the null hypothesis (H₀) that exceed a given quality measure cut-off was calculated for each year-source dataset and at 20 different cut-off values for confidence (A), phi (B), and lift (C). Boxes are the interquartile ranges among the year-source datasets; solid line is the median, whiskers extend up to 1.5 times the interquartile range and any outliers are marked with points.