A Novel Approach to Detection of COVID-19 and Other Respiratory Diseases Using Autoencoder and LSTM

Innumerable approaches of deep learning-based COVID-19 detection systems have been suggested by researchers in the recent past, due to their ability to process high-dimensional, complex data, leading to more accurate prediction of the COVID-19 infected patients. There is a visible dominance of Convolutional Neural Network (CNN) based models analysing chest images like X-rays and Computed Tomography (CT) scans for prediction, while the utilization of audio data for the same is less prevalent. Considering the respiratory system is one of the primary means by which the SARS-CoV-2 virus spreads, respiratory sounds are a potential biomarker for determining the presence of COVID-19. In this paper, we propose a novel approach for the detection of COVID-19 from amidst a dataset comprising of respiratory sound samples of healthy, COVID-19, and other lung diseases which are often misinterpreted as COVID-19. The approach employs an autoencoder for anomaly detection and a Long Short-Term Memory (LSTM) network for the detection of COVID-19 from amongst other lung diseases. The first stage of the model comprises an encoder-decoder-based autoencoder model with baseline reconstruction error, trained in an unsupervised environment, to reconstruct “healthy” audio signals. An LSTM based multi-class classifier is proposed for the second stage to classify the infected samples into the five classes: COVID-19, Bronchiolitis, COPD, Pneumonia and URTI. The experimental results demonstrate the efficacy of our proposed approach in detecting COVID-19 from a 5-class test set of audio samples of patients suffering from respiratory disease, with an accuracy of 98.7%, and an AUC of 1.