Multi-Method Approaches for Sleep EEG Analysis And Sleep Stage Classification

Multi-Method Approaches for Sleep EEG Analysis And Sleep Stage Classification

Sleep plays a fundamental role in human well-being, and understanding its intricate effects remains a crucial research area. Sleep electroencephalogram (EEG) signal analysis offers a promising direction for uncovering hidden singularities in sleep data. This thesis introduces innovative approaches for untangling sleep stage characteristics from EEG data.

Studied and inspired by the matching pursuit (MP) method, this research firstly developed a multitapers and convolution (MT&C) method that can decompose EEG data based on a dictionary. The MT&C method leverages controlled wavelets to compute spectral estimation, offering a robust basis for sleep EEG analysis, visual guidance, and stage scoring. By adhering to the Rechtshaffen and Kales sleep scoring manual (R&K rules) and the American Association of Sleep Medicine standards (AASM), both the MP and MT&C methods demonstrate an improved classification accuracy. Experimental results on healthy subjects demonstrated an accuracy of 79.4% and 87.6% for the MP and MT&C, respectively. While the MP and MT&C methods differ in definition, they complement each other and contribute to the advancements of sleep EEG analysis.

This thesis further examines the identification and classification of sleep spindles using a new spindles across multiple channels (SAMC) method. The SAMC implements multitapers and convolution to extract the spectral density estimation across multiple EEG channels, providing a comprehensive understanding of the behaviours and characteristics of the sleep spindles across the scalp. The SAMC method performs better than existing approaches, showcasing its potential to accurately identify and categorise sleep spindles.

Lastly, this study employed an advanced time-frequency analysis and incorporated a powerful deep learning model. The proposed method achieves significant performance improvements by employing the MT&C for initial feature extraction and utilising advanced techniques of visual geometric group, squeeze-and-excitation blocks, and scaled exponential linear units with batch normalisation. Across three diverse ii databases, the average accuracy and precision of 87% demonstrated the potential of these techniques in enhancing sleep stage classification.

Overall, this thesis contributes to the field of sleep research by introducing novel multimethod approaches for sleep EEG analysis, sleep stage classification, and spindle identification. The findings highlight the potential for improving understanding and possible diagnosis of sleep-related phenomena, offering new insights into sleep quality and its impact on human health and well-being.
Rationale: This research stems from the critical importance of understanding sleep's effects on human well-being. The focus on EEG signal analysis arises from the potential to uncover hidden aspects of sleep data, contributing to improved diagnosis and comprehension of sleep-related phenomena.
Contributions:
• Introduction of novel multi-method approaches for sleep EEG analysis, sleep stage classification, and spindle identification.
• Development of the MP and MT&C methods, enhancing sleep EEG analysis and classification accuracy.
• Innovation of the SAMC method, outperforming existing approaches in sleep spindle identification.
• Integration of advanced time-frequency analysis and deep learning techniques, significantly improving sleep stage classification across diverse databases. The findings of this thesis offer new insights into sleep quality and its profound impact on human health and well-being.

File reproduced in accordance with the copyright policy of the publisher/author/creator.