A convolutional transformer for EEG sleep stage classification

Abstract

Sleep is one of the most critical physiological processes in life, with profound implications for human health and functioning. Sleep stage classification is a fundamental task in sleep research, aiding in the understanding of the various stages of sleep and playing a crucial role in medical and health monitoring. Traditional Convolutional Neural Networks (CNNs) have limitations in handling long-range dependencies, focusing only on information within specific windows. For example, missing transitions between Rapid Eye Movement (REM) sleep stages and Non-Rapid Eye Movement (NREM) sleep stages. RNN may not effectively capture local features. For instance, RNNs might not effectively capture short awakenings or microarousals during sleep, which are crucial for accurate sleep stage classification. Subsequently, one limitation of transformers is their inefficiency in handling sequential data with long-range dependencies. To address these issues, we propose the Conformer model, which combines convolutional neural networks and transformers to efficiently capture both local and global dependencies in an electroencephalogram sequence. Conformer models employ various parameter reduction strategies, making the model more lightweight and suitable for embedded devices and low-resource environments. To validate our approach, we conducted comprehensive experiments using a widely collected single-channel EEG dataset and evaluated the model's performance. When compared to traditional methods and other deep learning approaches, our model demonstrates a significant advantage in capturing subtle features and patterns associated with different sleep stages.