Open Collections

Automatic sleep stage classification using multitaper spectral estimation, wavelet transform and neural network

University of British Columbia

Laboratory Polysomnography (PSG) is conducted using American Academy of Sleep Medicine (AASM) guidelines to diagnose sleep disorders, which has several limitations. In particular, the manual scoring process used for annotation of sleep stages is highly subjective, inefficient, and very time consuming. Therefore, the main objective of the present research is to automate the sleep stage classification process. The designing process of the automated system developed in the present thesis is based on five experiments, and utilizes a professionally annotated PSG data base of 994 patients and Feed Forward Neural Networks (FNN). It has been evident from the initial experimental results that the Composite Multiscale Sample Entropy (CMSE) feature does not have the potential to distinguish sleep stages as it only produced an overall accuracy of 61%. The present work uses the new frequency band information feature that has been specified in the AASM manual. This has in a much improved accuracy of 71% for the sleep stage classification. The frequency band information extracted with Multitaper algorithm has produced an accuracy of 76.8% (~77%) indicating that the Multitaper algorithm produces more accurate spectral estimations than the previous Fast Fourier Transform (FFT) algorithm. The experimental results obtained from the hybrid feature vector have an overall accuracy of 80.1% for sleep stage classification, which is a 19% improvement over that of the initial experiment. The experimental results also conclude that hybrid feature vectors that are for sleep stage classification are more effective than individual features. Finally, the model prediction analysis conducted using different threshold values has produced a significant overall accuracy of 96.8% (~97%) for the sleep stage classifier. The present work verifies that the finalized model has a better potential in sleep stage classification than that in previous studies and has improved control over the prediction accuracy levels permitted at the model output.

Text

2024-03-31

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/77455

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/