Parameter optimization for respiratory sound classification

Abstract

Respiratory conditions are one of the most common illness we have encountered. There are different forms of respiratory conditions, some of which proves to be potentially severe or fatal such as pneumonia and Pulmonary Edema. For such conditions, it is important to have access to early detection to reduce the risks of mortality. However, current methods of diagnosis, requires the patient to seek medical help at hospitals or clinics. In recent years, machine learning techniques have been researched to detect fluid accumulation in the lungs. These techniques could potentially be used to help patients with conditions such as Pulmonary Edema by acting as a form of early detection. However, as research in lung water detection is limited and new, more tests are required to investigate the accuracy of this technique. Thus, to further improve the machine learning technique on detecting lung water, the parameters employed will be varied to study the effects on the algorithm. In this study, the author had tested the segmentation window length of the data, the overlap window for the Mel-Frequency Cepstral Coefficient (MFCC) as well as three classification schemes. It was observed that the algorithm was not susceptible to the segmentation window length as well as the overlap window parameters. However, the majority classification scheme was able to yield better results as it has a higher sensitivity and specificity compared to the two other schemes tested.