Machine learning based models for audio signal classification : Application to COVID-19 Detection

Abstract

In this thesis, we propose novel methods for COVID-19 screening, where cough sound is used to employ various techniques. Our first approach is based on hybridizing Convolutional Neural Network (CNN) with Long-Short Term Memory (LSTM) along with Attention mechanism and spectral data augmentation. The second one is a set of ML approaches that are based on ensemble learning, where Random Forest algorithm was primarily used.The first approach consists of forwarding a raw Low-Level Descriptors (LLD) vector to the classifier, the second and the third are based on feature space compression and dimensionality reduction using Stacked Autoencoders and Locally Linear Embedding(LLE), respectively, to reduce computing complexity and make use of the most discriminant features to perform the classification task. The proposed methods are evaluated on a publicly available dataset called COUGHVID, and the results demonstrate their feasibility and high performance. The contributions of this thesis include the development of novel Deep Learning (DL) and ML-based architectures for COVID-19 screening, which can be used in both clinical and remote settings. These findings highlight the potential of ML and DL-based diagnosis system in improving the speed and accuracy of COVID-19 screening and its potential to assist healthcare providers for decision making, especially during pandemics.