Taqdees Khan (Inje University)

Hee Cheol Kim (Inje University)

Jundong Yu (Korea Racing Authority)

Young Beom Kwak (Inje University)

Taqdees Khan (Inje University)

Inhyeok Jang (Korea Racing Authority)

Hee Cheol Kim (Inje University)

Jundong Yu (Korea Racing Authority)

Young Beom Kwak (Inje University)

The proper identification of sex cohorts is essential for the successful implementation of anti-doping within the regulatory framework for horse racing, as anti-doping investigations could be compromised if the sex of a stallion, mare or gelding were misclassified. The single-threshold approach is not very useful for assessing a horse’s hormonal profile because of the nonlinear relationships among testosterone and other hormones and metabolites. In the meantime, manual inspection of GC–MS/MS analytical reports is increasingly becoming an operational hurdle in a high-throughput screening environment. In the present study, an automated Steroid Detection System was developed using PDF-based data extraction and normalization with an internal standard (IS) of deuterated testosterone (d3-testosterone), and a Random Forest (RF) classifier for automated gender prediction in horses. A total of 291 post-race urine samples from Males (n = 122), Females (n = 102), and Geldings (n = 67), collected at the Seoul Racecourse, Korea, were analyzed by GC–MS/MS in multiple reaction monitoring (MRM) mode. IS-normalized ratios of 10 key biomarker compounds were used as input features for classification. The Random Forest classifier (300 decision trees, balanced class weighting) achieved an overall 5-fold stratified cross-validation accuracy of 90.7 ± 2.3%, with per-class accuracies of 98.4% (Male), 87.3% (Female) and 82.1% (Gelding). On a stratified held-out test set (n = 45), the Random Forest achieved an accuracy of 95.3%, F1-score of 0.95, and AUC of 0.98, outperforming Logistic Regression (81.4%, 0.80, 0.85), Decision Tree (84.7%, 0.84, 0.88), SVM (87.2%, 0.87, 0.92), and Gradient Boosting (93.1%, 0.93, 0.97). Feature importance analysis revealed that 5α-androstane-3β,17α-diol (EN3), testosterone (EN1), and nandrolone (EN4) were the most important classifiers, accounting for 61.2% of the classification signal. The system was also shown to be useful for detecting abnormal androgenic profiles, as shown in a case study of a specimen of a masculinized gelding, confirmed by wet-lab GC–MS/MS analysis. The multivariate steroid profiling with IS normalization and Random Forest classification is a powerful, automated tool for gender classification of horses in regulatory anti-doping control.​​