Researchers in South Korea have built an automated system that reads laboratory steroid reports and uses machine learning to determine whether a racehorse is a stallion, a mare or a gelding, with performance that clearly beats the traditional single-threshold approach. The study, published in Frontiers in Veterinary Science, analyzed 291 post-race urine samples collected at Seoul Racecourse in 2025 and showed that a Random Forest classifier, fed with internally normalized gas chromatography–tandem mass spectrometry data, could classify sex cohorts with a five-fold cross-validation accuracy of 90.7 ± 2.3 percent. On a stratified held-out test set of 45 samples the model reached 95.3 percent accuracy, an F1-score of 0.95 and an area under the ROC curve of 0.98, outperforming Logistic Regression, Decision Tree, Support Vector Machine and Gradient Boosting models evaluated on the same data.

The motivation behind the work lies in the regulatory framework of horse racing. Anti-doping authorities, including the International Federation of Horseracing Authorities, prohibit the illicit use of anabolic androgenic steroids to enhance performance, and veterinary surveillance relies on sensitive mass spectrometry methods to monitor controlled substances. Accurate classification of stallions, mares and geldings is essential because sex-specific physiological baselines differ dramatically: under normal conditions, circulating concentrations of specific anabolic androgens in mares and geldings are expected to be very low or undetectable. If the sex of a horse is misclassified, an investigation into possible doping could be compromised, since what looks like an abnormal steroid level for a gelding may be entirely normal for a stallion.

In South Korea, the Korea Racing Authority records the sex of each registered racehorse at the time of initial competition registration, providing a reference classification for routine doping control. However, gelding may occur during a horse’s racing career, and records do not always reflect the most current status. The authors argue that automated classification based on biochemical steroid profiling offers an important complementary verification tool. The problem with the current approach is that equine steroid decisions typically rest on single-marker thresholds, such as testosterone alone, or on manual review of individual chromatographic profiles by experienced analysts. Equine steroid metabolism is a complex, nonlinear network rather than a single pathway, so sex and gonadal status are reflected in multivariate patterns across androgens, estrogens and downstream metabolites, not in one hormone.

This limitation is particularly acute for geldings, which have very low concentrations of gonadal steroids after castration. Distinguishing geldings from stallions and mares can be so challenging that classification accuracy approaches 50 percent in some analyses using threshold-based rules. Single-marker thresholds may also be affected by instrumental drift, batch-to-batch variation and biological differences between animals, and they fail to capture relationships between upstream precursors such as dehydroepiandrosterone (DHEA) and downstream metabolites. Beyond accuracy, the manual workflow itself creates a bottleneck: analysts must extract quantitative results from static PDF reports before classification, which increases processing time and cost and may introduce errors such as peak misidentification or transcription mistakes.

The new Steroid Detection System addresses both problems in a single pipeline. It begins with automated extraction of data from PDF laboratory reports generated by TraceFinder software. Each report contains sample metadata, a total ion chromatogram, a threshold table for target steroid compounds and individual selected reaction monitoring chromatogram panels with compound identity, expected and measured retention times, and peak area/height values. Approximately 87 percent of reports are machine-readable and processed through direct text-layer extraction using the pdfplumber library, while the remaining 13 percent are scanned image-based documents handled by an optical character recognition fallback pathway. Post-extraction validation filters anomalous values, flags missing entries and logs them to a per-sample audit trail, and the pipeline achieved 100 percent report processability across all 291 training samples.

After extraction, every peak area value is normalized against the deuterated internal standard d3-testosterone, producing dimensionless ratios that correct systematic batch-to-batch variation in instrument response, extraction recovery and derivatization efficiency. Ratios exceeding 50 are capped at that maximum to prevent extreme optical-character-recognition-derived outliers from dominating the feature space. The samples themselves were collected as part of routine post-race regulatory doping control, and any samples flagged as positive for illicit or therapeutic substances in the original workflow were excluded, so that the metabolic profiles used for training and validation were not confounded by exogenous substance administration. Sex labels were assigned manually from official racing registration records.

The Random Forest classifier, built with 300 decision trees and balanced class weighting to compensate for the unequal class distribution of 122 males, 102 females and 67 geldings, delivered per-class cross-validation accuracies of 98.4 percent for males, 87.3 percent for females and 82.1 percent for geldings. The gelding figure represents a critical advance over the rule-based baseline of roughly 50 percent, which the authors attribute to the model’s ability to learn nonlinear androgen metabolite interaction patterns. In the head-to-head comparison on the held-out test set, Gradient Boosting ranked second with 93.1 percent accuracy and an AUC of 0.97, while Logistic Regression scored lowest among the machine learning models at 81.4 percent accuracy and an AUC of 0.85, confirming that the gender-class boundaries are nonlinear in the normalized feature space. An ablation analysis in which the model was retrained using only the top five features reduced overall cross-validation accuracy from 90.7 ± 2.3 percent to 87.2 ± 2.8 percent, confirming that the lower-ranked features provide complementary discriminatory information.

Feature-importance analysis of the trained Random Forest identified 5α-androstane-3β,17α-diol as the strongest single predictor with an importance of 0.212, followed closely by testosterone at 0.209 and nandrolone at 0.191. Together these three biomarkers accounted for 61.2 percent of the total classification signal, with the remaining contribution distributed across 5(10)-estrene-3β,17α-diol at 0.107, boldenone at 0.066 and lower-ranked features each below 0.05. The authors interpret this distribution as a quantitative formalization of the conditional, multi-marker reasoning that experts have historically applied qualitatively: the same testosterone ratio can correspond to different phenotypes depending on the accompanying 5α-metabolite levels, because gonadal steroidogenesis proceeds through cascading enzymatic conversions and the diagnostic information about gonadal activity is distributed across multiple downstream products rather than concentrated in a single terminal hormone. This explains, they argue, why the largest gains over the threshold approach appeared in the male and gelding classes, whose profiles differ not in any one compound but in the coordinated pattern across the androgen and metabolite groups.

The system also demonstrated utility beyond routine classification when applied to an anomalous specimen. A registered gelding, Sample IMG_9892, showed an internal-standard-normalized testosterone ratio of 2.577 and a nandrolone ratio of 2.691, along with a substantial 5α-androstane-3β,17α-diol ratio of 1.072, all substantially exceeding the androgen levels expected under normal physiological conditions for a castrated male. The Random Forest predicted this sample as male with 83.6 percent confidence, and that prediction was consistent with an independent chemical analysis performed in the wet laboratory using GC–MS/MS, which also indicated elevated testosterone and nandrolone and classified the sample as male. From a regulatory perspective, the authors note, a gelding presenting such a hormonal profile would indicate either illicit exogenous administration of anabolic androgenic steroids or incomplete castration resulting in residual testicular function, and the system flagged the profile automatically without manual expert review.

The study has clear limitations of scope. It was conducted on a single population of 291 samples from one racecourse in one year, the gelding class was the smallest at 67 samples and the least accurately classified, and the held-out test set contained only 45 samples, so the reported test-set metrics carry uncertainty. The system also accepts pre-tabulated CSV and Excel input files and returns predictions through a web-based interface with a confidence score and probability distribution, allowing non-technical laboratory staff to submit files without programming expertise. The authors conclude that internal-standard-normalized multivariate steroid profiling combined with Random Forest classification is a powerful, automated method for sex cohort determination of equines in regulatory anti-doping control, one that reduces manual data-processing steps while formalizing expert interpretation into an objective, scalable model for large-scale surveillance.

Subject of Research: Machine learning classification of equine sex cohorts from GC–MS/MS urinary steroid profiles for anti-doping control

Article Title: AI-driven, qualitative GC–MS/MS-based endogenous steroid profiling with automated classification for sex cohort determination

Article References: Khan, T., Jang, I., Kim, H. C., Yu, J., & Kwak, Y. B. (2026). AI-driven, qualitative GC–MS/MS-based endogenous steroid profiling with automated classification for sex cohort determination. Frontiers in Veterinary Science, 13, Article 1939926. https://doi.org/10.3389/fvets.2026.1939926

Image Credits: AI Generated

DOI: 10.3389/fvets.2026.1939926

Keywords: horse racing, anti-doping, GC-MS/MS, steroid profiling, Random Forest, machine learning, geldings, testosterone, urine analysis, Seoul Racecourse, Korea Racing Authority, automated classification