Acoustic Signal Analysis and Feature- Based Classification of BOAS

Abstract

This thesis examines a feature-based approach for classifying Brachycephalic Obstructive Airway Syndrome (BOAS) in dogs using acoustic signal analysis and machine learning. Audio recordings of dogs breathing, collected both before and after physical exercise, were preprocessed through normalization, filtering, and data augmentation techniques to enhance signal quality. Features were extracted using the openSMILE toolkit and refined through statistical tests, notably the Mann-Whitney U-test, to identify those most indicative of BOAS severity. Two modeling strategies were employed: separate classifiers for pre- and post-exercise recordings and a hybrid model that incorporates both. The hybrid model, trained using decision tree-based methods including Random Forest and XGBoost, demonstrated superior performance, achieving an AUC of 1.0 and an average prediction confidence of 88.5% when evaluated on an unseen dataset of five dogs. Although more data is needed to ensure the model’s reliability and generalization to unseen data, these findings highlight the potential of a feature-based tool as a practical and accessible option for BOAS classification, thereby improving the health and welfare of brachycephalic dogs. v