Prediction of clinical mastitis antimicrobial treatment outcome using machine learning

• Grant number: 23/00286-3
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: September 01, 2023
• End date: August 31, 2024
• Field of knowledge: Agronomical Sciences - Veterinary Medicine
• Principal Investigator: Marcos Veiga dos Santos
• Grantee: Breno Luis Nery Garcia
• Supervisor: Diego Borin Nobrega
• Host Institution: Faculdade de Medicina Veterinária e Zootecnia (FMVZ). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Institution abroad: University of Calgary, Canada
• Associated to the scholarship: 21/09134-6 - Use of rapid on-farm test for decision on clinical mastitis treatment and use of microbiological culture as a selection criteria for selective dry cows therapy, BP.DR

Abstract

The increase in antimicrobial resistance (AMR) is a concern worldwide, and has been associated with the overuse of antimicrobials in animal production. In dairy cattle, clinical mastitis (CM) is the main reason for antimicrobial use. Thus, antimicrobial stewardship practices in dairy cattle production will necessarily consider therapy of CM. The ability to predict the risk of clinical and bacteriological cure (BC) will significantly affect the selection of treatment protocols for CM, which could lead to reduced use of antimicrobials on farms. This proposal aims to test and evaluate the ability of machine learning (ML) algorithms to estimate the risk of clinical and BC of cows with mild or moderate mastitis treated with antimicrobials. Retrospective data from approximately 10,000 CM cases from 2,200 Brazilian dairy herds will be used. From each CM case, detailed information on producer-observed clinical cure, intramammary treatment and causative agent will be obtained and included in models. Additionally, we will test a set of cow-level variables to determine their importance and impact on model performance. The following algorithms will be tested: stochastic gradient descent, decision trees, support vector machine, random forests, and artificial neural networks. Accuracy will be measured using cross-validation and receiver operating characteristic curves. Data will be divided randomly into training and test sets, and cross-validation will be performed for model development. For each algorithm, we will describe the sensitivity, specificity, and predictive values. Following selection of final models, we will test the performance of algorithms to predict the risk of cure in another database for which data on bacteriological cure is available. The database comprises of data from 400 CM cases evaluated for bacteriological cure post-treatment (14±3 and 21±3 days after treatment). Results obtained will support antimicrobial decision making on farms, and promote judicious use of antimicrobials. (AU)