| Summary: | Obstructive Sleep Apnea (OSA) refers to when an individual constantly experience pause or shallow breathing during their sleep. The current gold standard for OSA analysis is to use Polysomnography (PSG) system where individuals are required to sleep overnight in hospitals. During the test, individuals will be wired to multiple sensing systems to collect data which will be manually processed by sleep technologists to determine if an individual suffers from sleep disorder. However, this process is time consuming and challenging as huge amount of data need to be analysed. The aim of this project is to develop a machine learning based algorithm that can detect presence of OSA. It will be developed using MATLAB based on the snoring data. The algorithm includes three stages namely, Feature Extraction, Feature Selection and Classification. Features such as Mel Frequency Cepstral Coefficient (MFCC), Formant frequency and Kurtosis were extracted from the audio data. Followed by feature selection using Fisher’s Ratio to compute and select the more discriminative features. Lastly, the selected features will be used to perform classification. In this project, the classifier used is Support Vector Machine (SVM). This project achieved an accuracy, sensitivity and specificity of 96.7%, 100% and 93.3% respectively for the testing data. This result is predicted using the first training model which has an accuracy of up to 100% when all selected features are considered. In this project, Additive White Gaussian Noise (AWGN) was added to the original audio data and subsequently included as part of testing data. The purpose is to determine the effect on accuracy of prediction results when noise is added. This project has an accuracy of up to 96.0% when SNR is more than 50dB. Hence, it indicates that the model is robust to noise when SNR of more than 50dB is added as signal becomes clearer.
|
|---|