Classification of cough recordings for Covid-19 detection: Development of Machine Learning and Deep Learning models incorporating Concept Drift Adaptation

Abstract

This thesis examines the effectiveness of COVID-19 detection from cough recordings using machine and deep learning techniques in an attempt to reduce the cost and time required to diagnose the patient. In addition, the application of methods to adapt to concept drifts due to the ever-changing characteristics of the virus is examined, with the aim of maintaining the performance of the developed models. To this end, the use of different machine learning (random forests, multi-layer perceptron) and deep learning (convolutional neural networks) methods, as well as transfer learning approaches through the exploitation of pre-trained models are explored. The development and evaluation of the models is based on the use of the Coswara dataset. To address the unbalanced nature of the dataset, techniques for synthetic data generation (SMOTE), cost-sensitive learning and classification threshold optimization are exploited. The highest performance based on the AUROC metric (80.21%) is achieved by a convolutional neural network architecture that uses the pre- trained VGG-16 as the base model. To adapt to the concept drift, the last dense layers of the model are retrained using an appropriate normalization method, which leads to an improvement of the model's performance with respect to the AUROC metric by up to 5%.