A sound based algorithm for detecting obstructive sleep apnoea

• Main Author: Ang, Shu Hui
• Other Authors: Ser Wee
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: http://hdl.handle.net/10356/67046
• Institution: Nanyang Technological University

Obstructive Sleep Apnoea (OSA) is a chronic respiratory disorder which disrupts the patient’s sleep when his upper airway involuntarily collapses during the night, causing the person to stop breathing for typically 10 seconds. Traditional methods to detect OSA involves a full night polysomnography (PSG) which is invasive, complex and costly. Studies have shown multiple characteristics to distinguish patients with OSA, and snore signals acquisition is favourable for this development. This project's objective is to develop a sound-based algorithm for automatic OSA detection and it will include the study of a machine learning technique. This project focuses on the spectral analysis of these snore signals by doing Fast Fourier Transform (FFT). The sound-based apnoea detection algorithm involves extracting multiple features from the snore signals, followed by a feature selection algorithm which will rank the prominence of these features using Fisher's Ratio (FR). Subsequently, the classification algorithm uses the Support Vector Machine (SVM) to distinguish the OSA and normal snorers. The spectral analysis had revealed a disparity between the nature of OSA and normal snores. OSA snores had much higher energy intensities and occurred mainly at higher frequencies as compared to the latter. Additionally, another distinction between the type of snores were their peak frequencies, kurtosis and skewness of the spectrum. With these knowledge and numerical data, SVM was used to form a suitable training data set. Consequently, testing will determine the best decision threshold and conduct a performance evaluation based on their accuracy, sensitivity and specificity. This project showed an average accuracy, specificity and sensitivity of 93.54%, 94.12% and 88.56% respectively for the decision threshold of using the top 2 features ranked by FR simultaneously. In conclusion, this project used the frequency-domain approach to discuss the various sound based algorithm useful in detecting the presence of OSA, which could assist future works of developing a sound-based diagnostic method on a portable device which is less invasive, less complex and less costly.