Machine learning-based prediction and quantification of OCD surgery and pedigree effects on racehorse performance.

Overview

• This study used machine learning to investigate how arthroscopic surgery for osteochondritis dissecans (OCD) affects the long-term race performance of Thoroughbred horses.
• It also compared the influence of pedigree versus surgery and management factors on predicting race outcomes, finding that pedigree was the strongest predictor and surgery had limited impact.

Background and Objectives

• Osteochondritis dissecans (OCD) is a developmental orthopedic condition frequently seen in Thoroughbred racehorses that can potentially impact their athletic performance.
• Arthroscopic surgery is commonly performed to treat OCD lesions, but there is uncertainty about how this surgery affects a horse’s long-term racing success.
• The study aimed to:
• Evaluate the long-term impact of OCD surgery on race performance.
• Compare the predictive strength of pedigree data and management factors for race results.
• Apply interpretable machine learning techniques to forecast race outcomes and understand contributing variables.

Data and Methods

• Retrospective data were collected for 75 Thoroughbreds that underwent OCD surgery between 2015 and 2017, along with 257 maternal siblings without surgery to serve as controls.
• Variables included:
• Biometric data (e.g., horse measurements)
• Pedigree information (family identifiers and relationships)
• Surgery details (age at surgery, lesion severity)
• Race performance metrics (earnings, starts, percentile rankings)
• Additional derived features included earnings per start and sale-price ratios to provide richer predictors of performance.
• Race performance was quantified using a field-adjusted metric called ‘race_pts_avg’, which normalizes performance relative to competition.
• Two gradient boosting algorithms, XGBoost and CatBoost, were trained and compared for predictive accuracy using R² scores.
• Interpretability was achieved using SHapley Additive exPlanations (SHAP), which quantify how each feature influences the output of the machine learning model.

Main Findings

• Models that incorporated the derived performance features (like earnings per start) consistently performed better than models that did not include them.
• The best-performing model, CatBoost, achieved a high R² value of 0.7983, indicating strong predictive power.
• Variables related directly to OCD surgery, such as age at surgery and lesion severity, showed limited importance in predicting race performance.
• Pedigree information, particularly a feature called ‘h_family’ which encoded maternal family lineage, was the most influential predictor of performance.
• This suggests that genetic lineage outweighs the impact of OCD surgery on long-term racing success in this cohort.

Implications

• The study provides evidence that undergoing arthroscopic surgery for OCD does not significantly reduce a Thoroughbred’s future race performance potential.
• Pedigree remains the most critical factor in forecasting performance, reinforcing the importance of breeding decisions.
• These insights may guide breeders, trainers, veterinarians, and buyers when making treatment, management, and sales decisions for horses with a history of OCD surgery.
• Additionally, the use of interpretable machine learning models allows stakeholders to understand which factors drive performance prediction, increasing trust in the results.

Limitations

• The study was limited geographically and used a relatively small sample size which may affect generalizability.
• As an observational study, there could be unmeasured confounding factors that influence both the likelihood of surgery and race performance outcomes.
• Future research with larger, more diverse populations and prospective designs would be beneficial to confirm these results.