UTILIZATION OF SHALLOW LEARNING FOR COVID-19 CLASSIFICATION BASED ON COUGH SOUNDS

• Main Author: Arif Rahman, Dandy
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/57189
• Institution: Institut Teknologi Bandung

Coronavirus disease 2019 (COVID-19) is an ongoing pandemic. In general, COVID-19 is transmitted through droplets produced when an infected person coughs, sneezes, or exhales. Several researchs have been conducted to use cough sounds as a distinguishing feature between people who is infected COVID-19 and those without. In this research, the author tries to utilize artificial intelligence to classify COVID-19 using cough recordings using shallow learning modeling. This method can serve as a tool to prioritize a person for further diagnosis, namely RTPCR. In this study, the contributions made were to try different feature extractions, unbalanced data handling, and several modeling techniques to classify COVID-19 using cough recordings. The feature extraction techniques tested include Mel Frequency Cepstrum Coefficient (MFCC), Non-negative Matrix Factorization on MFCC (NMF-MFCC), Non-negative Matrix Factorization on spectograms (NMFspectogram), and log mel spectograms. Balanced data handling techniques used are undersampling, oversampling, and Synthetic Minority Over-Sampling Technique (SMOTE). While the modeling techniques tried are K-nearest neighbor (KNN), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost). Based on the evaluation and analysis of the experimental results, the NMF-spectogram tends to produce better performance compared to other feature extraction techniques. Unbalanced data handling techniques tend to produce the same performance, so the undersampling technique is preferable from the point of view of memory utilization and model time. Modeling using XGBoost tends to produce the best performance, although the best results are obtained by the SVM model, but SVM modeling using MFCC-based features is always stuck in overfitting problems, so it cannot generalize the problem well. Based on the experimental results, the best results were obtained using a combination of NMF-Spectrogram features, undersampling methods, and SVM. This combination gives a sensitivity value of 90.9%, specificity 55.6% and AUC-ROC 73.3%.