Recipient Organization
CORNELL UNIVERSITY
(N/A)
ITHACA,NY 14853
Performing Department
(N/A)
Non Technical Summary
Every time that we use antibiotics in animals or people, we may contribute to the rise of antimicrobial-resistant bacteria. These bacteria can survive antibiotic treatment, threatening human and animal health. National surveillance systems are essential tools for detecting and responding to resistant bacteria. Surveillance can also show us when antibiotic use policies are successful in reducing resistance. However, variation in laboratory processes across time and institutions makes it difficult to assess trends in resistance over many years. In addition, we have limited tools for analyzing multidrug resistance.This seed grant supports the development of new analytic tools for antimicrobial resistance surveillance. We will examine Escherichia coli in cattle and the restrictions on cephalosporin use. Resistant E. coli can be transmitted from cattle to humans through beef products, direct contact, or the environment. First, we will fill in gaps that arise from varied laboratory processes. Random forests, a machine learning method, will predict missing resistance data (i.e., minimum inhibitory concentrations, MIC). Then, statistical models can identify previously hidden trends in resistance data over time. We will also investigate the impact of uncertainty in resistance testing on our conclusions. Association mining, another machine learning tool, will analyze MICs and genomic sequencing data to reveal changes in multidrug resistance. Overall, our research will improve antimicrobial resistance surveillance and help identify successful resistance mitigation strategies.
Animal Health Component
50%
Research Effort Categories
Basic
50%
Applied
50%
Developmental
(N/A)
Goals / Objectives
Our overall long-term goal is to leverage big data to understand the relationship between antimicrobial resistance (AMR) in livestock, AMR in humans, and antimicrobial use (AMU) regulations. This will enable the creation and continuation of effective AM stewardship policies. To achieve this goal, we must first address barriers to transforming AMR surveillance data into actionable insights. Current AMR analysis methods obscure increases or decreases in resistance because data simplification is required to avoid the MIC and multidrug resistance (MDR) analysis barriers described above. Whenever data is reduced to a simpler format with less variability, statistical power is lost. We hypothesize that novel analytic techniques, which do not require data simplification or transformation, will reveal AMR trends that would be missed with the current AMR analysis standards. Thus, these new methods will provide stronger evidence to support antimicrobial stewardship policies. We will test our hypothesis on cattle-associated Escherichia coli AMR data from NARMS, NCBI Pathogen Detection, and the National Animal Health Laboratory Network (NAHLN). Our proposed work will enable surveillance programs to more effectively detect AMR trends and evaluate the merits of AMR mitigation policies.Objective 1: Quantify the impact of AMU policies on AMR by comparing MIC distributions.Objective 1.1: Impute missing MIC data.Objective 1.2: Create antimicrobialclass resistance indicators.Objective 1.3: Model MIC trends over time with Cox PH regression.Objective 1.4: Compare survival analysis models to the Mann-Kendall test for trend in binary resistance prevalence, which is the NARMS standard.Objective 1.5: Evaluate the impact of MIC uncertainty with sensitivity analyses.Objective 2: Characterize and compare MDR patterns across the food chain by applying machine learning tools to phenotypic and genotypic AMR indicators.Objective 2.1: Jointly analyze phenotypic and genotypic AMR indicators to characterize MDR.Objective 2.2: Compare MDR at each step in the food chain.Objective 2.3: Compare the MDR patterns identified with association mining to a standard MDR analytic approach.
Project Methods
The proposed analytic methods are innovative and novel for AMR studies. While machine learning has been used to generate MICs in some bacterial species from genomic data, our machine learning MIC imputation method will be the first to predict narrower MIC intervals from other phenotypic data to compensate for AST protocol changes. Association mining is a novel approach to MDR analysis. Our previous work demonstrated that association mining is adaptable for MDR surveillance. Here, we propose the next step--developing hierarchical association rules at the AM drug and class levels from two data sources, phenotypic AST results and genomic sequences, to understand MDR dynamics in the food supply chain.Data: We will use publicly available phenotypic and genotypic surveillance data, downloaded from NARMS and NCBI. Cattle-associated E. coli was selected as the model organism because it has consistent genotype-phenotype associations, and NARMS and NAHLN test it against a wide range of AMs.Obj. 1: Quantify the impact of AMU policies by comparing MIC distributions.Obj. 1.1: Impute missing MIC data. For AMs with inconsistent concentrations used for susceptibility testing, missing MIC data will be imputed with random forests, predicting an MIC of interest from other AMs and covariates (e.g., year). The model outcome is the MIC of an AM of interest. The model inputs are MICs from other AMs and covariates. The random forest classifier will then be used to predict MICs where AMs and concentrations tested changed.Obj. 1.2: Create AM class resistance indicators. For AMs that are not consistently tested, we will create ordinal AMR indicators for each AM class or sub-class by aligning the MIC distributions for each AM within the class at the median or mode and summing the distributions.Obj. 1.3: Model MIC trends over time with Cox PH regression. We will conduct two primary analyses: (i) MIC trends from 2013 - 2018 using NARMS cecal, ground beef, and human illness isolates, and (ii) the effect of the 2012 cephalosporin restriction on MICs using NARMS ground beef and human illness isolates from 2005 - 2018. Each AM will be modeled separately with the MIC or the ordinal AMR indicator as the outcome. Year of sample collection and isolate source (i.e., cecal vs meat vs human) will be the primary explanatory variables. We will compare year coefficients between 2005-2011 and 2012-2018 to quantify the effect of the 2012 cephalosporin restrictions.Obj. 1.4: Compare survival analysis models to the Mann-Kendall test for trend in binary resistance prevalence, which is the NARMS standard. We will tabulate the number of AMs with an unsignificant Mann-Kendall trend test but with a significant effect (P < 0.05) for year in the Cox PH model.Obj. 1.5: Evaluate the impact of MIC uncertainty with sensitivity analyses. We will build Cox PH models under three scenarios, comparing the results to the original model to quantify the impact of MIC censoring: (i) the actual MIC occurs at the midpoint of the MIC interval, (ii) the actual MIC occurs at the reported MIC, (iii) the actual MIC occurs at the lower limit of the MIC interval. We will also evaluate the effect of MIC measurement error with a sensitivity analysis, sampling the MIC from a uniform distribution of the reported MIC +/- one dilution.Obj. 1 expected results. We expect that the MIC distribution analysis (i.e., Cox PH model) will be more sensitive to changes in AMR than prior gold-standards of AMR trend analysis (e.g., Mann-Kendall tests) because the MIC provides a more detailed view of AMR than resistant/susceptible categorization. When analyzing the FDA's cephalosporin restrictions, we expect to find a larger change in cephalosporin and beta-lactam resistance (e.g., smaller MICs or slower rate of MIC increase) than in other resistances due to strong cross-resistance among beta-lactams.Obj. 2: Characterize and compare MDR patterns across the food chain by applying machine learning tools to phenotypic and genotypic AMR indicators.Obj. 2.1: Jointly analyze phenotypic and genotypic AMR indicators to characterize MDR. We will generate hierarchical association sets, with phenotypic resistance and resistance genes clustered by AM class, to allow comparisons across phenotypic and genotypic data. For example, the hierarchical association set for both the phenotypic pattern [ampicillin, gentamicin, sulfisoxazole] and the genotypic pattern [blaCMY-2, aadA2, sul1] is [beta-lactam, aminoglycoside, folate-pathway inhibitor]. We will compare the results of the joint phenotypic-genotypic analysis to association sets generated from (i) phenotypic data only (i.e., reported MIC values or resistant/susceptible), and (ii) AMRFinder gene data only (i.e., gene presence or absence).Obj. 2.2: Compare MDR at each step in the food chain. The association sets will be compared across the datasets - sick cattle (NAHLN data), slaughterhouses (NARMS cecal data), retail ground beef (NARMS meat data), and sick humans (NARMS human data) - to quantify MDR variation across the food chain. We will quantify the association between specific MDR patterns and each source by creating association rules of the form "MDR → Source", similar to explanatory and outcome variables in regression models. The year will be added to the rule to describe the changes in MDR patterns over time.Obj. 2.3: Compare the MDR patterns identified with association mining to a standard MDR analytic approach. We will tabulate the most common MDR phenotypes and genotypes (a standard approach) and calculate the proportion that are identified as association sets. Next, we will compare the MDR pattern and source relationship quantified by association rules (e.g., "MDR → Source, Year") to logistic regression models (a standard approach) with common MDR phenotypes as outcomes and year and source as explanatory variables.Obj. 2 expected results. We also expect the association sets and rules to provide substantially more detail on MDR trends than phenotype tabulation and logistic regression, while still capturing the most common patterns. Obj. 2 will create a powerful, flexible toolbox for MDR analysis by overcoming two fundamental challenges in AMR analysis: i) massive numbers of potential resistance patterns to analyze, and ii) distinct surveillance datasets for AST and genome sequencing.Efforts: Results will be published in peer-reviewed journals and work-in-progress shared at national conferences, such as the Conference of Research Workers in Animal Diseases and the American Association of Bovine Practitioners Conference. At the latter, we will discuss our findings on AMU restrictions with AM prescribers. Regular communication with stakeholders (e.g., California Antimicrobial Use and Stewardship veterinarians, NARMS/NAHLN and diagnostic lab researchers) will guide our interpretation of results and plans for future work.Evaluation: We will evaluate our progress and success by comparing milestone completion dates to our proposed timetable. The use of specific products (e.g., reproducible R code, peer-reviewed publications) by the target audience will be measured through user-access data and citations. We will qualitatively assess the impact on the work on AMR surveillance through our regular discussions with stakeholders.TimetableStartEndObj. 1.11/248/24Obj. 1.25/248/24Obj. 1.39/244/25Obj. 1.41/254/25Obj. 1.55/258/25Prepare Obj. 1.1/1.2 abstract for conference5/248/24Prepare and publish manuscript for Obj. 11/258/25Obj. 2.11/2412/24Obj. 2.29/248/25Obj. 2.35/2512/25Prepare Obj 2.1/2.2 abstract for conference1/254/25Prepare and publish manuscript for Obj. 25/2512/25