Logo image
Back
A first step towards AI-assisted grading of laryngeal function in Thoroughbred yearlings
Conference proceeding

A first step towards AI-assisted grading of laryngeal function in Thoroughbred yearlings

Lucy Green, L Talas, J Fennell, Josephine Hardwick, Samantha Franklin and Kate Allen
Conference Proceedings. 46th Bain Fallon Memorial Lectures: Sports Medicine, Lameness, Endocrinology, Surgery
46th Bain Fallon Memorial Lectures (Melbourne, VIC, 19/07/2026–23/07/2026)
07/2026
url
PublisherView

Abstract

Introduction Laryngeal function (LF) grading relies on subjective visual interpretation, leading to low interrater agreement among veterinarians. This study investigates the feasibility of using machine learning (ML) to predict LF grades from endoscopic videos, and to assess whether the use of ML could facilitate more objective grading practices. Materials and methods A dataset comprising 4,605 post-sale Thoroughbred laryngoscopy recordings were graded by consensus using the Havemeyer scale. Videos were analysed with an ML pipeline: a convolutional neural network was used to segment laryngeal structures, enabling extraction of quantitative features. These features were used to train an XGBoost classifier to predict LF grade. Results On an unseen portion (20%) of the dataset, the model matched consensus grade 52% of the time for classification by grade, and 82% for “pass/fail” (threshold: > grade III.1). Agreement with consensus was moderate for multiclass (quadratic-weighted Cohen’s κ = 0.55) and substantial for binary predictions (κ = 0.64); comparable to previously reported interveterinarian agreement. Performance may improve with label denoising and increased training data for grades ≥ III.1. Feature distributions showed significant overlap between grades, highlighting the inherent challenges in separation. Relevance to Australian clinical equine practice Poor agreement in grading at Thoroughbred yearling sales has diminished stakeholder confidence in the process. Unlike human graders, given identical parameters, the model always assigns the same grade to a given video; and is able to report quantitative metrics about LF from endoscopy video. Computer-assisted LF grading could support more objective decision making for veterinarians and purchasers at sales.

Details

Metrics

1 Record Views
Logo image