Supervised Machine Learning Techniques for Breeding Value Prediction in Horses: An Example Using Gait Visual Scores.
Authors: Bussiman Fernando, Alves Anderson A C, Richter Jennifer, Hidalgo Jorge, Veroneze Renata, Oliveira Tiago
Journal: Animals : an open access journal from MDPI
Summary
# Editorial Summary Gait visual scoring remains fundamental to equine genetic selection programmes, yet the inherent subjectivity of these assessments can constrain breeding progress. Researchers compared traditional genetic evaluation methods with three machine learning approaches (artificial neural networks, random forest regression, and support vector regression) to predict breeding values for five gait traits in Campolina horses, utilising over 5,000 phenotypic records across 107,951 animals spanning 14 generations. Whilst all machine learning models achieved comparable predictive accuracy to conventional multiple-trait statistical models, artificial neural networks showed marginally superior accuracy but introduced higher bias and over-dispersion, particularly when predicting values for younger animals without extensive performance data. The findings suggest machine learning offers a viable complement to established EBV estimation, though practitioners should recognise that these algorithms tend to produce inflated predictions for young stock and may require adjustment factors when used operationally. For breeding programmes seeking to incorporate emerging analytical tools, this research indicates that machine learning warrants cautious integration alongside—rather than replacement of—traditional quantitative genetic approaches, especially when making selection decisions on animals with limited phenotypic histories.
Read the full abstract on PubMed
Practical Takeaways
- •Machine learning offers a viable complementary tool to traditional genetic evaluation methods for predicting breeding values from gait scores, though current implementations may not fully replace established MTM approaches
- •When using machine learning predictions for young horses or animals with limited records, expect reduced reliability and wider prediction intervals compared to traditional methods
- •Subjective visual gait scoring limitations may be partially addressed through machine learning, but the technology works best when integrated with established genetic evaluation frameworks rather than as a standalone replacement
Key Findings
- •Machine learning models (ANN, RFR, SVR) achieved comparable accuracy to traditional multiple-trait model (MTM) for predicting breeding values from visual gait scores
- •Artificial neural networks showed slightly higher accuracy but demonstrated the highest bias and over-dispersion compared to other methods
- •Machine learning predictions were biased and over-dispersed particularly for young animals with limited phenotypic data
- •Five visual gait score traits (dissociation, comfort, style, regularity, development) in Campolina horses were successfully modeled using adjusted phenotypes from over 107,000 pedigreed horses