Techniques for Cough Sound Analysis

Abstract

Coughing is a common symptom of pulmonary ailment and serves as a valuable measure when quantifying pulmonary health. This dissertation contains the results of the research and development of a set of techniques to enable researchers to investigate pulmonary health through cough sounds. A variety of signal processing and machine learning approaches are included, each with various performance and usability tradeoffs. We propose a novel algorithm that makes use of the best of the traditional signal processing approaches, combined with the recent advances in deep learning to provide new cough detection and classification results previously unattainable, especially when considered in the context of model runtime performance. We detail the construction a classifier for tuberculosis coughs, and develop a new tool to deal with bifurcated datasets we dub a discriminative adversarial network.