A Novel CNN plus Transformer Network for EEG-based Automatic Sleep Staging

Sleep is vital for human health, underpinning memory consolidation, restoration, and overall well-being. Accurately classifying sleep stages is crucial for diagnosing sleep disorders. Traditional sleep staging relies on experts reading spectrum analysis, which is time-consuming and labor-intensive. Current research methods are either multi-channel with high information dependency or single-channel but with low accuracy. Single-channel detection with high-accuracy results have been relatively lacking. Besides, most existing methods rely on hand-engineered features, which can be tedious and error-prone. In this paper, we propose a novel sleep stage classification system, named ConvTransSleepNet, which combines the strengths of one-dimensional convolutional neural networks (ID-CNN) and Transformer layers. This hybrid architecture retains the inherent properties of traditional CNN models while incorporating the powerful self-attention mechanism of Transformer layers. The ID-CNN component is specifically designed to adapt to one-dimensional input data, such as biomedical signals like EEG. The transformer layers in ConvTransSleepNet allows the model to effectively capture complex temporal relationships and contextual information present in the EEG signals, further improving the sleep stage classification performance. Tested on the Sleep-EDF dataset, ConvTransSleepNet achieved an overall accuracy of 84.6%. Particularly, it demonstrated high accuracy for the wake, N2, N3, and REM stages. In comparison to other state-of-the-art approaches, our model showcased superior performance, with an overall accuracy of 0.84 and a macro-averaged Fl score of 0.79. These findings highlight ConvTransSleepNet's potential as an efficient tool for sleep stage classification.