Deep learning based sleep stage classification studies

Deep learning based sleep stage classification studies

sleep stage classification plays a vital role in diagnosing and treating sleep and mental related diseases. The utilization of traditional visual scoring in sleep research is limited due to the intricate nature of the procedure and the substantial time commitment. Most of the previous sleep-scoring investigations have focused on the process of feature extraction. These methods are laborious and can only attain limited precision. Hence, this thesis aims to develop innovative and effective deep-learning based algorithms and models for automatically detecting and classifying the distinctive features associated with each sleep stage. The first proposed method involves integrating convolutional neural networks (CNNs) with gated recurrent units (GRUs) by leveraging several bio-signals, such as electroencephalography (EEG), electromyogram (EMG), and electrooculogram (EOG). The proposed method is able to extract concealed characteristics from multi-channel bio-signal data and effectively distinguish sleep stages. It was tested and evaluated using two databases, namely the Sleep Heart Health Study (SHHS) database and the University College Dublin Sleep Apnoea (UCDDB) database. The primary contribution of this study is in its ability to enhance classification performance compared to previous methods. It can learn features from diverse datasets while maintaining the model architecture. The second approach proposed integrates a CNN and a Long Short-Term Memory (LSTM) model, by employing a Mel-frequency cepstral coefficient (MFCC) feature. The findings of the second study suggest that the MFCC feature possesses the ability to preserve pertinent information about each sleep stage, hence reducing the duration required for retrieving the raw data and enhancing the computational efficiency of the model. The third method presented in this thesis proposed the integration of CNN, GRUs, EfficientNet, and a modified ResNet-50. This method employed the capabilities of EfficientNet in effectively managing the dimensions of the network, including breadth, depth, and resolution. This approach has demonstrated promising outcomes in the assessment of two distinct databases the SHHS database and the Sleep-EDF-78 database. The findings of this study indicate that utilizing numerous consecutive EEG epochs as input data sequences is an effective approach for auto sleep classification.

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