Epoch-level and Sequence-level Multi-Head Self-Attention-based Sleep Stage Classification

Inthis paper, we propose a Transformer-based sleep stage classification method by effectively combining epoch-level and sequence-level multi-head self-attention with multiple outputs to aid the training. Most of the existing automated sleep staging methods utilized convolutional neural networks (CNNs) for feature extraction and recurrent neural networks (RNNs) or Transformer for capturing global dependencies. However, there are disadvantages when utilizing CNNs and RNNs (e.g., large parameter sizes, information loss). Therefore, Transformer-only architecture with epoch-level multi-head self-attention for capturing intra-epoch dependencies and sequence-level multihead self-attention for capturing inter-epoch dependencies in 30-second PSG signals is utilized. Furthermore, a sequence of learnable class tokens is deployed to store summary of each sleep epoch. As a result, three different outputs are produced, in which two are used as references during the backpropagation in training. Numerical simulation results on the Sleep-EDFX78 dataset confirm that the proposed method show improvements compared to the existing Transformer-based methods.