Digital Structuring of Equine Upper Airway Endoscopic Findings for Machine Learning-Based Non-Destructive Diagnostics

Endoscopy is a non-destructive diagnostic method widely used to evaluate equine upper airway disorders; however, interpretation remains subjective and depends on clinician experience. This study aimed to standardize endoscopic findings and establish a reproducible data pipeline for machine learning-based analysis within a non-destructive testing framework.

A total of 109 equine endoscopic examinations were retrospectively analyzed. The following pathological features were assessed using an expert-based scoring system: edema, hyperemia, upper tracheal mucus accumulation (0–3), lymphoid tissue hyperplasia (0–4), laryngeal paralysis (1–4), dorsal displacement of the soft palate, and dorsal pharyngeal collapse. Edema and hyperemia were the most frequent findings (75.2% and 69.7%, respectively). Tracheal mucus accumulation showed predominance of mild grades, while lymphoid tissue hyperplasia demonstrated a broad distribution across severity levels. Laryngeal paralysis (7.3%), dorsal displacement of the soft palate (9.2%), and dorsal pharyngeal collapse (5.5%) were less frequent.

The dataset is structured for supervised learning with clearly defined target variables. The primary task is the classification of tracheal mucus severity, implemented as both a binary (low [0–1] vs high [2–3]) and a multi-class (0–3) problem. Additional pathological features are included as auxiliary labels to enable multi-label analysis.

To ensure methodological rigor, data splitting is performed at the case level, and model evaluation is defined using cross-validation with accuracy, sensitivity, specificity, and F1-score. This study provides a justified and reproducible framework for transforming clinical endoscopic data into machine learning-ready inputs, supporting the development of objective diagnostic tools in equine respiratory medicine.