Single-channel EEG sleep stage classification using synchrosqueezed transform and sequential representation learning

Accurate classification of sleep stages is essential for both sleep research and clinical practice, yet remains challenging due to the complexity of EEG signals and feature overlap between stages. This study introduces a novel three-stage approach using single-channel EEG signals from the Sleep-EDF dataset. We focused on the Fpz-Cz channel, applying the SynchroSqueezed Transform (SST) to convert signals into time-frequency representations (TFRs). To address the complexity of classifying long, 3000-sample signals, we first divided each 30-second epoch into smaller, 256-sample segments. In Stage 1, a Convolutional Neural Network (CNN) was trained on these segments to initialize the model's weights. In Stage 2, contrastive learning further refined the encoder, improving feature separability between sleep stages. This two-step pretraining established a robust foundation for feature extraction. In Stage 3, we returned to the full 3000-sample signals and used the pretrained encoder in a time-distributed manner, extracting features from 256-sample windows. This enabled the encoder to process each segment, with the extracted features then passed to two stacked Gated Recurrent Units (GRUs) to capture long-term dependencies for final classification. This model achieved an accuracy of 83%, highlighting the efficacy of our approach for sleep stage classification.