Clustering Sleep EEG Recordings to improve the Automated Sleep Scoring Procedure

Sleep has a critical role in promoting health. The polysomnography (PSG) and the analysis of the sleep recordings are the main tools to evaluate sleep and to identify sleep disorders. Sleep scoring is an essential procedure for sleep analysis. Its automation is gaining importance over years. It reduces the time-consuming tedious repetitive work of the physicians and it improves the reproducibility of the results. The aim of this work is to optimize the performance of an existing deep learning-based automated sleep scoring system. The innovative idea is that, to simplify and to enhance the automatic scoring procedure, the learning architecture need to be trained on sub-groups of PSG recordings. Indeed, we propose to cluster in sub-groups the raw data of the training set and to use the data of each cluster to train independent Neural Networks. Two different approaches are proposed. The Approach1 considers the whole night electroencephalography (EEG) recordings as instance to be clustered, whilst the Approach2 considers night excerpts of 90 seconds to be clustered. In both clustering approaches, all the features are extracted from the raw Fpz-Cz EEG derivation. Multiresolution analysis, multiscale entropy and autoencoding feature extraction techniques are employed to analyze the temporal evolution of the EEG signal. Feature selection, feature ranking and dimensionality reduction techniques are applied on the extracted feature vectors. Spectral Clustering and Mini Batch K-Means algorithms are used for the Approach1 and the Approach2, respectively. The clustering results are validated on two databases of different size, relying on the performance, in terms of overall accuracy, of the automated sleep scoring system. The Approach1 does not seem to be an appropriate solution. It is supposed that the signals a e compared using too large time scale, so they can share parts of recordings really similar or totally different. The Approach2 has proven to be a promising solution, showing a substantial improvement in performance of the automated sleep scoring system. In order to validate the latter approach, further analysis should be carried out by testing the procedure on databases with a larger number of subjects.