Abstract:We present a high performance and computation-efficient deep neural network method for automatic sleep stage classification. Using a multi-model integration strategy with multiple signal channels as input, we propose a method that automatically detects signal shedding and eliminates its effects. We designed a compact network, using bottleneck structure, depth-wise separable convolution, and shortcut connection, which dramatically reduces the number of parameters of the neural network model. Our approach has achieved significant improvements in multiple metrics on the Sleep-EDF dataset (accuracy 86.1%, mF1 81.1%, Cohen's Kappa 80.0%), while the number of model parameters is only several tenths of the other methods. We further explored the application of deep neural networks in automatic classification of sleep stage and laid the foundation for automated sleep analysis software to be applied on portable or wearable devices.

Computation-Efficient Multi-Model Deep Neural Network for Sleep Stage Classification.