A Deep Learning Approach for Automatic Sleep Stage Segmentation via Single Channel EEG Signals

Abstract

In this paper we develop and evaluates Convolutional Neural Network (CNN) models for automatic sleep stage segmentation using single-channel EEG signals from the Sleep-EDF dataset. We explored three architectures: CNN-CNN, CNN-CNN-Conditional Random Field (CRF), and CNN-Long Short-Term Memory (LSTM). The CNN-CNN-CRF model, incorporating a CRF layer for sequence labeling, demonstrated the highest performance with an accuracy of 89% and an F1 score of 82%, outperforming CNN-CNN (accuracy 87%, F1 score 81%) and CNN-LSTM (accuracy 71%, F1 score 76%). This approach surpasses existing state-of-the-art models by effectively capturing temporal dependencies between sleep epochs and ensuring consistent sequence labeling. The findings suggest that integrating CRFs with CNNs enhances classification performance, providing a robust solution for automated sleep stage segmentation. These results highlight the potential of deep learning in improving sleep analysis, with implications for sleep quality assessment and clinical applications.