Published 2019
| Version v1
Journal article
Restricted
Comprehensive analysis of machine learning models for prediction of sub-clinical mastitis: Deep Learning and Gradient-Boosted Trees outperform other models
Authors/Creators
- 1. School of Animal and Veterinary Sciences, The University of Adelaide, Adelaide, Australia, School of Basic Sciences, University of Qom
- 2. School of Animal and Veterinary Sciences, The University of Adelaide, Adelaide, Australia
- 3. School of Animal and Veterinary Sciences, The University of Adelaide, Adelaide, Australia, Genomics Research Platform, School of Life Sciences, La Trobe University, Melbourne, Victoria, Australia, School of Information Technology and Mathematical Sciences, Division of Information Technology Engineering & Environment, University of South Australia, South Australia, Australia, School of Computer Science, Faculty of Engineering, Computer and Mathematical Sciences, The University of Adelaide, South Australia, Australia, Australian Centre for Antimicrobial Resistance Ecology, School of Animal and Veterinary Sciences, The University of Adelaide, South Australia, Australia
- 4. School of Animal and Veterinary Sciences, The University of Adelaide, Adelaide, Australia, Australian Centre for Antimicrobial Resistance Ecology, School of Animal and Veterinary Sciences, The University of Adelaide, South Australia, Australia, Davies Research Centre, School of Animal and Veterinary Sciences, The University of Adelaide, South Australia, Australia
Description
Sub-clinical bovine mastitis decreases milk quality and production. Moreover, sub-clinical mastitis leads to the use of antibiotics with consequent increased risk of the emergence of antibiotic-resistant bacteria. Therefore, early detection of infected cows is of great importance. The Somatic Cell Count (SCC) day-test used for mastitis surveillance, gives data that fluctuate widely between days, creating questions about its reliability and early prediction power. The recent identification of risk parameters of sub-clinical mastitis based on milking parameters by machine learning models is emerging as a promising new tool to enhance early prediction of mastitis occurrence. To develop the optimal approach for early sub-clinical mastitis prediction, we implemented 2 steps: (1) Finding the best statistical models to accurately link patterns of risk factors to sub-clinical mastitis, and (2) Extending this application from the farms tested to new farms (method generalization). Herein, we applied various machine learning-based prediction systems on a big milking dataset to uncover the best predictive models of sub-clinical mastitis. Data from 364,249 milking instances were collected by an electronic automated in-line monitoring system where milk volume, lactose concentration, electrical conductivity (EC), protein concentration, peak flow and milking time for each sample were measured. To provide a platform for the application of the models developed to other farms, the Z transformation approach was employed. Following this, various prediction systems [Deep Learning (DL), Nave Bayes, Generalized Liner Model, Logistic Regression, Decision Tree, Gradient-Boosted Tree (GBT) and Random Forest] were applied to the non-transformed milking dataset and to a Z-standardized dataset. ROC (Receiver Operating Characteristics Curve), AUC (Area Under The Curve), and high accuracy demonstrated the high sensitivity of GBT and DL in detecting sub-clinical mastitis. GBT was the most accurate model (accuracy of 84.9%) in prediction of sub-clinical bovine mastitis. These data demonstrate how these models could be applied for prediction of sub-clinical mastitis in multiple bovine herds regardless of the size and sampling techniques. 2019 Elsevier Ltd
Files
Restricted
The record is publicly accessible, but files are restricted. Log in to check if you have access.