Intra- and Inter-epoch Temporal Context Network with Pyramidal Feature Hierarchy for Automatic Sleep Scoring on Raw Electrograms

Abstract

This study proposes a deep learning architecture that learns intra- and inter-epoch temporal contexts and channel-based model ensemble for automatic sleep scoring. To classify the sleep stage from half-minute PSG data, sleep experts investigate sleep-related events and frequency characteristics in the epoch and consider the transition rules between the successive epochs. Inspired this idea, in the first chapter, Intra- and Inter-epoch Temporal Context Network (IITNet-v1) is proposed. IITNet-v1 extracts representative features at a sub-epoch level by a modified ResNet-50 and then captures intra- and inter-epoch temporal context from the series of feature sequence vi Bi-LSTM. The proposed architecture was evaluated on two public datasets (Sleep-EDF and MASS) as the number of input epochs, the sequence length (L), increased from one to ten. As a result, the performance dramatically increased until the sequence length was four and converged around at ten. In the second chapter, the experiments for improving sleep scoring performance were conducted on the same dataset under the conditions of the sequence length ten obtained in the first chapter. For the enhancement, attention layer and pyramidal feature hierarchy were introduced to IITNet-v1 in order to consider entire hidden states and multi-scale feature maps, respectively. In addition, various experiments were carried out to investigate the network configuration. Consequently, two anchored VGG-13-based IITNet-v1 (called IITNet-v2) surpassed IITNet-v1. In the last chapter, the channel-based model ensemble was performed in order to consider EEG, EOG, and EMG together. Two well-known model ensemble methods (majority voting and softmax probability averaging) were used in this study. The softmax probability averaging method considering two-channel EEG and single-channel EOG performed better than proposed single models. Finally, IITNet-v2 and its ensemble model outperformed the existing algorithms.