Predictors of epistaxis incidence in Thoroughbred racehorses in the United States

Abstract

Epistaxis (EP), or bleeding from the nose, may negatively impact racehorse health and performance. To test the hypothesis that a combination of quantifiable factors could predict susceptibility to an EP event, a database of Thoroughbred racehorses experiencing EP unrelated to physical trauma for 2024 in the United States was collated. Data for all horses experiencing EP (n = 202) and all non-EP horses from the same races (n = 1286) were included for analysis. Twenty-five potential risk parameters were extracted from veterinary exams, race data, and environmental factors. The correlations (r) of these risk factors with EP incidence were used to choose predictors for the parsimonious models. Logistic regression modeling was performed using Python due to the binomial distribution of the outcome variable of interest (EP). Weighted logistic regression models with the top 5 (Model 1) and 10 (Model 2) correlated risk factors were compared for model performance (Table 1). Within model 1, EP diagnosis had a positive relationship (P < 0.0001) in the fitted model, whereas off odds had a negative correlation (P = 0.04). Model 1 achieved a high average accuracy score of 84%. For predicting non-EP, precision, recall, and F₁-score of model 1 were all 90%. Predictive EP capability of model 1 was moderate, with precision, recall, and F₁-score values all being 48%. The receiver operating characteristic area under the curve (ROC-AUC) score for this model was 69%, indicating significant utility of model 1 in predicting EP. Within model 2, which had the top 10 strongest correlations, previous EP was the most significant predictor (P < 0.0001), similar to model 1. Precision, recall, and F₁-score for predicting non-EP were all 90% for model 2, as well. For predicting EP, these scores were 46%, 48%, and 47%, respectively. Model 2 had similar model performance statistics as model 1, with an accuracy score of 83% and a ROC-AUC score of 69%. Our analysis favored model 1 to predict EP, due to its parsimonious nature. Further evaluation with advanced models, such as XGBoost and Random Forests, is recommended to gain a more comprehensive understanding of the factors influencing epistaxis.