Data Augmentation for Audio Based Machine Learning Classifying Brachycephalic Obstructive Airway Syndrome (BOAS) in Dogs

Abstract

Breathing problems of varying degree are common amongst dog breeds with shorter snouts also called brachycephalic dogs. The process of classifying each case consists of a veterinarian visit where tests are preformed to assess the severity on a scale from zero to three. In this master thesis, we aim to simplify this procedure by machine learning and will be working with two hypothesis. Hypothesis I is a continuation of the master thesis Brachycephalic Obstruction Airway Syndrome (BOAS) classification in dogs based on respiratory noise analysis using machine learning by Moa Mårtensson. Here we augmented the audio files to generate a larger data set and extracted multiple features. The features include MFCC, ZCR and RMS that are fed to a LSTM network. The second hypothesis aims to classify BOAS(-) and (+), this hypothesis uses frequency data enhanced with SMOTE and a CNN. We show that it is possible to classify BOAS using machine learning, but that more data is required in order to confidently diagnose BOAS. We can conclude that hypothesis II using data collected from the Littmann device shows the best result on unseen audio files. There is a possibility to further develop this into a tool for both veterinarians and dog owners. This thesis is a collaboration between Chalmers University of Technology and the Swedish University of Agricultural Sciences in Uppsala.