COMBINING AUDIO FEATURES AND LIGHTGBM FOR CLASSIFYING COVID-19 PATIENTS THROUGH COUGH SOUNDS

Abstract

Since the outbreak of the COVID-19 pandemic, mass testing has become essential to reduce the spread of the virus. The primary diagnostic tool to detect COVID-19 is a reverse-transcription polymerase chain reaction (RT-PCR) test from the respiratory specimens of the suspected patient, which is invasive and a resourcedependent technique (such as healthcare workers or medical equipment). Because the implementation of the above technique has not satisfied the factors of velocity and costs, researchers around the world have been looking for other characteristics to classify COVID patients. The cough sound feature is a solution feasible for this problem. Some related research has been proposed to classify covid-19 patients through audio recordings based on various machine learning algorithms such as Decision Trees, SVM, RNN. In this study, the proposed solution is combining audio features and LightGBM for classifying covid-19 patients through cough sounds. The selected machine learning algorithm is LightGBM which is a gradient-boosting framework with the following advantages: faster training speed and higher efficiency, lower memory usage, and better accuracy. It is being widely used in many winning solutions of machine learning competitions. Some data features that are important in speech and audio processing are Mel-Frequency Cepstral Coefficients, Chroma-based, Spectral Roll-off, and Mel-Spectrogram which are extracted from cough sounds and used to train the LightGBM model. In this thesis, a web application was built to support COVID-19 classification via audio recordings. The cough sounds will be recorded on the front-end by the user and transmitted to the backend through the client-server model. At the back-end side, the data features will be extracted from cough sounds and processed with the power of the LightGBM model. Then, the prediction result will be responded to the frontend. The evaluating results show the accuracy is 0.82 and f1-scores is 0.76 when using the LightGBM model. The research results show the superiority of the LightGBM algorithm on an imbalanced dataset that contains 23,553 samples.