Detection of COVID-19 based on cough sound and accompanying symptom using LightGBM algorithm

Abstract

Coronavirus disease 19 (COVID-19) is an infectious disease whose diagnosis is carried out using antigen-antibody tests and reverse transcription polymerase chain reaction (RT-PCR). Apart from these two methods, several alternative early detection methods using machine learning have been developed. However, it still has limitations in accessibility, is invasive, and its implementation involves many parties, which could potentially even increase the risk of spreading COVID-19. Therefore, this research aims to develop an alternative early detection method that is non-invasive by utilizing the LightGBM algorithm to detect COVID-19 based on the results of feature extraction from cough sounds and accompanying symptoms that can be identified independently. This research uses cough sound samples and symptom data from the Coswara dataset, and cough sound’s features were extracted using the log mel-spectrogram, mel frequency cepstrum coefficient (MFCC), chroma, zero crossing rate (ZCR), and root mean square (RMS) methods. Next, the cough sound features are combined with symptom data to train the LightGBM. The model trained using cough sound features and patient symptoms obtained the best performance with 95.61% accuracy, 93.33% area under curve (AUC), 88.74% sensitivity, 97.91% specificity, 93.17% positive prediction value (PPV), and 96.33% negative prediction value (NPV). It can be concluded that the trained model has excellent classification capabilities based on the AUC values obtained.