Artifact Removal From Sleep-Disordered EEG by Wavelet Enhanced Independent Component Analysis

Abstract

In the field of sleep research, the quantitative analysis of electroencephalography (EEG) data acquired during sleep offers invaluable insights. However, the presence of artifacts in such data can severely distort analytical outcomes. Therefore, this study aims to develop an innovative artifact detection and rejection tool to enhance the analysis of sleep EEG data from overnight polysomnography (PSG) in patients with sleep disorders. While recent advancements have seen the trend of artifact removal using a hybrid method, these techniques typically require pre-labeled data for training machine learn- ing models, introducing a dependency on prior knowledge. Addressing this limitation, our research introduces a novel unsupervised learning approach, utilizing hierarchical clustering to identify artifactual components within wavelet-enhanced independent component analysis (ICA)-separated data. We present a unique set of features for clustering, including kurtosis, zero-crossing count, skewness, Hjorth parameters, and R ́enyi entropy, tailored to discern artifacts in EEG recordings. Our methodology affords the flexibility of fully automated artifact removal or semi-automated processes involving visual inspection of hierarchical dendrograms. Comparative analyses demonstrate that this new method not only refines EEG signal quality but also surpasses traditional manual cleaning techniques in performance. The findings underscore the potential of hierarchical clustering in the unsupervised learning landscape for artifact detection, heralding a significant step forward in the preprocessing of EEG data for sleep studies.