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

Osteochondritis dissecans (OCD) is a common developmental orthopedic condition in Thoroughbred racehorses and although arthroscopic surgery is widely used for treatment, its long-term effects on race performance remain unclear. This retrospective study evaluated the effect of OCD surgery on race performance, compared the predictive power of pedigree and management variables and applied interpretable machine learning methods for forecasting race performance outcomes. Data were collected from 75 Thoroughbreds that underwent OCD surgery between 2015 and 2017 and 257 maternal siblings without recorded OCD surgery (controls). Variables included biometric, pedigree, surgical and race performance measures. We also derived additional predictors, including earnings per start and sale-price ratios and quantified racing performance using the field-adjusted percentile metric 'race_pts_avg'. We trained gradient-boosting models (XGBoost and CatBoost) and evaluated predictive performance using R² scores and Shapley Additive Explanations (SHAP). Models that included derived features consistently outperformed those without and the best model (CatBoost) achieved R² = 0.7983. Variables related to surgical history, including age at surgery, showed limited predictive value and lesion severity did not rank among the dominant predictors. In contrast, pedigree, particularly the group-mean encoded family identifier 'h_family', ranked highest. These results indicate that OCD surgery does not significantly impair long-term race performance, while pedigree is the strongest predictor in this cohort. Although this observational study is limited by its regional scope and by unmeasured factors that may influence both surgery status and performance, the findings provide objective context for breeding, treatment and sales decisions involving Thoroughbreds with a history of OCD surgery.