Temporal Convolutional Network-based Multi-View Sleep Staging

In this paper, we propose a novel approach for multi-view sleep staging using Temporal Convolutional Networks (TCNs). Our method effectively processes both time and time-frequency domain representations by incorporating intra and inter-epoch level layers of TCNs. Unlike previous methods that heavily relied on a single input type, such as time-domain signals or time-frequency representations, our approach tackles the challenge of simultaneously utilizing both domains compared to using only one as input. While many existing automated sleep staging methods employ recurrent neural networks (RNNs) to capture sequential information at intra and inter-epoch levels, there are potential limitations associated with RNNs, such as gradient vanishing or exploding, and computational complexity. Therefore, we adopt a TCN-based architecture to effectively capture intra and inter-epoch dependencies in 30-second EEG signals. Moreover, we propose a novel gradient blending method that considers both validation and training losses at recent and current timesteps. This method is designed to facilitate training and achieve enhanced performance by emphasizing on the recent trend of the training process. On numerical simulation results, conducted on the SleepEDF-20 and DRM-Sub dataset demonstrate that our proposed method outperforms the existing automated sleep staging methods.