CNN-BASED MODEL COMPARISON FOR COVID-19 DETECTION USING COUGH SOUND

• Main Author: Lucky Mahendra, Dimas
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76671
• Institution: Institut Teknologi Bandung

Sars-CoV-2 virus or widely knwon as COVID-19 was classified as a pandemic throughout all over the world in 2019-2022. This phenomenon motivated a lot of researchers to study a machine learning implementation for detecting this virus by using cough sound. The difficulties of the researches regarding this topic is the limited amount of COVID-19 positive patient data available to be used for machine learning compared to the COVID-19 negative patient data, and also there are many variations of model architecture that could be used. Some researches shows that CNN-based model results in a higher performance score when compared to shallow learning-based model, but the combination between the preprocessing data method and the model used could be studied further to achieve the most optimal result. On this research, an experiment was conducted by using two types of imbalance data handling method, which are Random Undersampling and data augmentation. The experiment also used two types of feature, which are Mel Frequency Cepstral Coefficient (MFCC) and Log-Mel Spectrogram, and also three types of model architecture, which are CNN, ResNet-50, and VGGish. For the baseline model, a combination of Random Undersampling, MFCC, and CNN model was used. The experiment is the continued by using a variation of the total positive training data after the preprocessing phase is done with the amount of 1000-1500 data. The best experiment result is achieved by using a combination of White Noise Injection as the imbalance data handling method, MFCC for the feature, and CNN model architecture. This variation results in 83% specificity score, 44% sensitivity score, and 79% F1-Score. This variation gives an improvement of 0.248 specificity score, 0.0178 sensitivity score, and 0.02 F1-Score when compared with the baseline model. In conclusion, this combination is the best combination to classify COVID-19 by using cough sound in this research.