Recipient Organization
UNIVERSITY OF CALIFORNIA, DAVIS
410 MRAK HALL
DAVIS,CA 95616-8671
Performing Department
(N/A)
Non Technical Summary
Recent findings from our on metritis spontaneous cure (MSC) using deep artificial neural networks (ANN) correctly identified 80% of the cows that experienced a spontaneous cure and 76% of all cows that cured or failed to cure from metritis. Albeit these results were highly encouraging, an external validation, a comprehensive economic analysis, and an assessment of other already commercially available technologies, such as cow genomics and cow sensor behavioral data, would be critical to optimize predictive models and solidify the evidence for the implementation of selective therapy for metritis and reduce AMU in dairy cows in the various dairy scenarios. Furthermore, metritis is a polymicrobial disease in which cows failing to cure metritis have an increased abundance of genes associated with LPS synthesis (galE) and quorum sensing signaling (pgi)that might contribute to microbes interactions associated with metritis cure failure.Therefore, there is a critical need to externally validate our MSC deep ANN predictive model's cost-effectiveness and evaluate if the integration of microbial and host genomics and cow sensor behavioral data can improve MSC predictability and cost-effectiveness. Our long-term goal is to develop highly accurate artificial intelligence-driven modelsthat predict metritis cure in dairy cows in diverse scenarios and reduce antimicrobial use without compromising animal welfare and the cost-effectiveness of dairy farms. Our central hypothesis is that the integration of host and microbial genomics and cow's sensor behavioral data will improve MSC predictability metritis and allow the development of selective therapy that is safe and cost-effective in a variety of scenarios that represent the dairy landscape in the US and beyond.
Animal Health Component
60%
Research Effort Categories
Basic
20%
Applied
60%
Developmental
20%
Goals / Objectives
Our long-term goal is to develop a highly accurate artificial intelligence-driven model that predicts metritis spontaneous cure (MSC) in dairy cows in diverse scenarios and reduces antimicrobial use (AMU) without compromising animal welfare and cost-effectiveness. Our central hypothesis is that the integration of host and microbial genomics and cow's sensor behavioral data will improve MSC predictability and judicious AMU in a cost-effective and safe manner.The scenarios depicted will represent the broad tapestry of the dairy landscape in the US. The rationale for the proposed research is that metritis is a highly prevalent and costly polymicrobial disease that is a top AMU driver. Identifying cows that can experience MSC and cows that benefit from treatment is a pivotal paradigm shift needed to improve antimicrobial stewardship and reduce the risk of antimicrobial resistance (AMR) dissemination.A recent collaborative effort from 11,733 dairy cows from 16 different farms in 4 different United States regions highlighted that current metritis incidence ranged from 20.5% to 34.9%, with an average cost of U$513 per case and annual losses estimated at over $1 billion dollars. Failure to cure metritis harbors a significant detrimental impact on health, reproduction, and culling. Regarding antimicrobial stewardship, metritis portrays a contentious scenario in which antimicrobial use (AMU) narrowly improves cure rates from 55-62% to 75-78% compared to spontaneous cure4-6 This situation has led to the investigation of predictive models to identify cows thatcan cure metritis effectively to reduce AMU. Recent research from our team indicated that the integration of cows and environmental factors using machine learning (ML) models predicted metritis cure with high sensitivity (Se=81-85%) and moderate positive predictive value (PPV=75-78%). However, specificity (Sp=26-39%) and negative predictive value (NPV=33-50%) were low. A follow-up study focusing only on metritis spontaneous cure (MSC) using deep artificial neural networks (ANN) correctly identified 80% of the cows that experienced a spontaneous cure and 76% of all cows that cured or failed to cure from metritis. Albeit these results were encouraging, an external validation, a comprehensive economic analysis, and an assessment of other commercially available technologies such as cow genomics10 and cow sensor behavioral dataare critical to optimize predictive models and solidify the evidence for the implementation of a judicious reduction of AMU for metritis in various scenarios that represent the spectrum of dairy farms in the US. Furthermore, metritis is a polymicrobial disease with Bacteroides, Porphyromonas, and Fusobacterium appearing to behave synergistically in cows developing the diseaseand failing to cure. Our data shows that cows failing to cure metritis have an increased abundance of genes associated with LPS synthesis (galE) and quorum sensing signaling (pgi)14 that might contribute to metritis cure failure. Besides the microbial contributions, validation of genome-enabled prediction showed a 10.8% reduction in metritis incidence between best and worst groups, and heritability for tested models improved from 0.12 to 0.17, but its incorporation in predictive models to select cows MSC has not been investigated. Microbiome-related, behavioral, and genome data represent unique complex layers of information that can help explain and predict metritis cure but have not been critically evaluated. One of the most significant challenges of integrating all these layers of information to predict metritis cure is the analytical feasibility of statistical models with many variables and assumptions needed for multiple comparison corrections. A robust approach to circumvent the analytical feasibility issues is to use MLand deep ANN models. The lack of restrictive parametric assumptions for data analysis and the development of predictive tools makes artificial intelligence-driven models an excellent candidate for improving the prediction of classic statistical models, as shown previously in studies of cows with metritis, as we previously reported9, and cow performance, disease, and microbial genomics. Deep ANN uses a layered structure of ML algorithms and can automate the variable extraction process, unlike standard ML techniques that require manual variable selection. Deep ANN is an ML inspired by biological neural networks, on which each ANN encompasses nodes (analogous to cell bodies) that communicate with other nodes via connections (similar to axons and dendrites). Therefore, there is a critical need to externally validate our MSC deep ANN predictive model accuracy and cost-effectiveness and evaluate if the integration of microbial and host genomics and cow sensor behavioral data can improve the MSC predictability and cost-effectiveness. Our objectives are:I.2. Specific objectivesObjective 1: To create a database of cows' phenotype, genotype, behavioral data, and environmental factors and collect uterine swab samples of Holstein metritic cows.Sub-objective 1. A.: To diagnose metritis and randomly allocate cows to NO deep ANN model9 for MSC (Control, n=400) or deep ANN for MSC (MSC, n=400). MSC cows will be randomly allocated to MSC-treated (n=200) and MSC-untreated (n=200).Sub-objective 1.B.: To collect uterine swabs at metritis diagnosis to characterize uterine microbial genomics biomarkers for MSC.Sub-objective 1.C.: To collect data for cows' genotype, metritis risk factors, milk yield, culling, reproductive performance, behavioral sensor data (e.g., lying, activity, rumination, and walking), and environmental factors (e.g., season, type of facility, type of bedding, etc.).Objective 2: To characterize uterine microbial genomics and relationships with host genomicsSub-objective 2.A.: Perform deep shotgun sequencing of uterine samples to characterize the microbiome, resistome, and virulome linked to MSC.Sub-objective 2.B.: Detect functional relationships and candidate causal effects of host SNPs, microbes, virulence factors, and antimicrobial genes with MSC.Sub-objective 2.C.: Identify and select top microbial genomic factors (e.g., microbial species, antimicrobial resistance genes) and host genomics (e.g., SNPs) associated with MSC.Objective 3: Compare the model performance and cost-effectiveness of 1) deep ANN predictive model (ANN); 2) ANN plus genomics (ANN-GEN); 3) ANN plus microbial markers (ANN-MIC), 4) ANN plus cow behavior sensor data (ANN-BEH), and 5) full model with deep ANN, genomics, microbial markers, and cow behavior sensor data (ANN-FULL).Sub-objective 3.A.: Compare Se, Sp, PPV, NPC, accuracy and other predictive performance metrics f ANN, ANN-GEN, ANN-MIC, ANN-BEH, and ANN-FULL.Sub-objective 3.B.: Develop a comprehensive economic model with inputs from the study and sensitivity analysis to evaluate the cost-effectiveness of the predictive models compared.Objective 4: Develop and deliver a broad education and extension program on best management practices for antimicrobial stewardship for metritis in dairy herds.Sub-Objective 4.A.: Develop best management practices based on return on investment and safety according to farm scenarios that are supported by our research findings.Sub-Objective 4.B.: Present the current strategies for improving antimicrobial stewardship for metritis of this study through extension events, website, apps, and curated social media.Sub-Objective 4.C.: Evaluate changes in the understanding and subsequent use of predictive models for MSC and its impact for AMU over time in response to the research findings using producer focus groups and data collected from individual dairies and DHIA.
Project Methods
Objective 1.: To create a database of cows' phenotypes, genotype, behavioral data, and environmental factors and collect uterine swab samples of Holstein metritic cows. Cows diagnosed with metritis will be randomly allocated to NO deep ANN prediction9 for MSC (Control, n=400), or deep prediction ANN for MSC (MSC, n=400). Cows allocated to MSC will be randomly re-allocated to MSC-treated (n=200) and MSC-untreated (n=200). The proposed approach will allow us to externally validate the deep ANN for MSC benefits when compared to the current dairy industry standard (MSC-untreated vs. Control) and assess if predicted benefits would withstand if cows predicted to have MSC are treated (MSC-untreated vs. MSC-treated). Drs. Lima and Pereira will use their relationship with the dairy industry to secure farmers who fulfill the inclusion criteria. Dr. Lima will lead the sample and data collection for the study.Farms, housing, diets, and management: The farms queried to be part of the study are located in San Joaquin Valley in California and milk between 2,000 and 6,000 Holstein cows. The farms have genotyped cows and activity-automated sensor devices. Information including lactation number, sire identification, calving date, dystocia, stillbirth, twins, vulvar-vaginal laceration, retained fetal membranes, calf gender, disease postpartum, milk production, first service pregnancy per AI, daily and test-day milk production, days open at 300 DIM culling/death will be collected from the farms' DairyComp 305 database (Valley Agricultural Software, Tulare, CA). Cow sensor data (lying, resting, activity, rumination, eating, and walking) will also be collected from specific software used on each farm. Environmental-related factors such as season, temperature-humidity index, housing, bedding, and stocking rate will be obtained from all farms. Data will be consolidated into an Excel spreadsheet for further analysis.Metritis monitoring and diagnosis. Metritis diagnosis will be performed after morning milking, based on vaginal discharge (VD) and rectal temperature (RT) that will be evaluated on 4, 7, and 10 DIM and as requested by farm employers. Vaginal discharge will be retrieved using the Metricheck device and scored using a modified 0 to 5 scale: 0 = no secretion material retrieved, 1 = clear mucus, 2 = flecks of pus in the vaginal discharge, 3 = <50% pus in the vaginal discharge, 4 = >50% pus in the vaginal discharge, 5 = watery, fetid, red-brownish vaginal discharge49. Rectal temperature will be assessed using a digital thermometer, and RT ≥39.5°C will be determined to have a fever. Cows with fetid, watery, red-brownish VD (score = 5) will be diagnosed as having metritis34. Metritis type will be classified as metritis with or without fever.Randomization and treatment allocation. A total of 800 cows (approximately 200 from each farm) with metritis without an exclusion criterion will be blocked by parity (primiparous vs. multiparous), and metritis type (with or without fever) at enrollment and within each randomly allocated to NO deep ANN prediction9 for MSC (Control, n=400), or deep ANN prediction for MSC (MSC, n=400). Then the cows allocated to MSC will be randomly re-allocated to MSC-treated (n=200) and MSC-untreated (n=200). Treated cows will receive a subcutaneous administration of 6.6 mg of ceftiofur crystalline-free acid /kg of body weight administered at diagnosis of metritis and 72 hours later (Excede®, Zoetis, Madison, NJ). Enrolled cows will not be eligible to receive non-steroidal anti-inflammatories. Still, they may receive supportive therapies such as oral propylene glycol and intravenous administration of fluids such as sodium chloride, dextrose, and Ringer's lactate, among others, following farms' treatment protocols. These therapies can be administered any time after enrollment, following farm protocols and the veterinarian's discretion.Health evaluations, clinical cure definition, escape therapy. Daily RT will be assessed in the first five days and on day 12. Vaginal discharge will be evaluated on days 5 and 12, as previously described. The clinical cure will be evaluated on day 12, and it will be characterized as a VD score < 5. The spontaneous cure will be characterized by a cure in the CON cows and treatment failure will be characterized by non-cure in the TRT group. To ensure proper welfare to study cows, escape therapy will be considered for all treatment groups based on the following signs: persistent dehydration, anorexia, depression, systemic shock, or any other clinical signs attributable to metritis with or without elevated rectal temperature. Farm employees will administer escape therapy following standard farm protocols, including alternative antibiotics (e.g., oxytetracycline, penicillin, ampicillin) and anti-inflammatories. Cows that receive escape therapy, die, or are sold because of metritis will be considered non-cured. Body condition will be scored (BCS) at enrollment in all cows based on a 5-point scoring system.Uterine swab collection.As previously described, uterine swab samples will be collected from all cows at metritis diagnosis using a 30" double-guarded sterile culture swab.Genotypic characterization: Interrogation of genomic variation will be made by collecting the information generated from the use of the bovine highly specific Dairy Bovine Clarifide Ultra (Zoetis Services LLC), which contains over 62,000 highly informative SNPs uniformly distributed across the entire genome of dairy cattle. Animals will be nominated, along with pedigree and genotype, to the Council of Dairy Cattle Breeding (CDBC) to obtain CDCB genetic evaluation predictions for metritis and other related wellness traits as previously described10. Individual values for each cow for wellness traits will be collected and used.Objective 2.: To characterize uterine microbial genomics and relationships with host genomics. We will use deep shotgun sequencing to characterize the microbiome, virulome, and resistome of the uterine swab on the day of metritis diagnosis. We will perform deep shotgun sequencing of uterine samples in 600 cows from the 4 farms enrolled in the study (Control=200, MSC-treated=200, and MSC-untreated). The 200 cows from the Control group will be matched pair to equally represent farms, parity, and fever at metritis diagnosis to minimize sources of variability across the groups compared. The same 600 cows (Control=200, MSC-Treated=200; MSC-Untreated=200) will have their genotypical data assessed to detect functional relationships and potential causal effects between host genomics and MSC. Dr. Lima will perform animal selection and data curation, and Dr. Weimer will lead the shotgun sequencing and bioinformatics of the proposed objective.?Objective 3.: Compare the model performance and cost-effectiveness of 1) deep ANN predictive model (ANN); 2) ANN plus genomics (ANN-GEN); 3) ANN plus microbial markers (ANN-MIC), 4) ANN plus cow behavior sensor data (ANN-BEH), and 5) full model with deep ANN, genomics, microbial markers, and cow behavior sensor data (ANN-FULL).Objective 4.: Develop and deliver a broad education and extension program on best management practices for antimicrobial stewardship for metritis in dairy herds. The extension program will be led by Drs. Lima and Pena-Levano, who have extension expertise and access to resources and stakeholders. The working hypothesis is that producers and stakeholders will perceive comprehensive extension programming as highly relevant. Together with the advisory panel (dairy farmers, extension experts, consultants, veterinarians, and researchers), strategies for judicious AMU for metritis will be created and disseminated based on scenarios of our findings and their relationships with health, milk yield, culling, and reproductive traits. The best management practices will be developed based on predictions, inferences, and impact on AMR safety.