Abstract
Brachycephalic obstructive airway syndrome (BOAS) is a highly prevalent respiratory disease affecting popular short-faced dog breeds such as Pugs and French bulldogs. BOAS causes significant morbidity, leading to poor exercise tolerance, sleep disorders and a shortened lifespan. Despite its severity, the disease is commonly missed by owners or disregarded by veterinary practitioners. A key clinical sign of BOAS is stertor, a low-frequency snoring sound. In recent years, a functional grading scheme has been introduced to semi-objectively grade BOAS based on the presence of stertor and other abnormal signs. However, correctly grading stertor requires significant experience and adding an objective component would aid accuracy and repeatability. This study proposes a recurrent neural network model to automatically detect and grade stertor in laryngeal electronic stethoscope recordings. The model is developed using a novel dataset of 665 labelled recordings taken from 341 dogs with diverse BOAS clinical signs. Evaluated via nested cross validation, the neural network predicts the presence of clinically significant BOAS with an area under the receiving operating characteristic of 0.85, an operating sensitivity of 71% and a specificity of 86%. The algorithm could enable widespread screening for BOAS to be conducted by both owners and veterinarians, improving treatment and breeding decisions.