GPNet: Granularity-Aware Pyramid Network with Graph Aggregation for Sleep Staging and Face-Emotional Recognition Speed Prediction

Accurate classification of sleep stages is crucial for sleep quality assessment, health monitoring, and disease prevention. To effectively extract significant waveform features and capture the interactive coupling of features at different layers from single-channel electroencephalogram (EEG) signals, this study proposes the Granularity-Aware Pyramid Network with Graph Aggregation for Sleep Staging (GPNet) model. Specifically, the model first extracts fine-grained time-frequency features from multi-resolution input signals using the feature pyramid. Subsequently, an adaptive deep attention mechanism is incorporated into the layer with the highest depth-wise information to explore the correlations between local and global features. Finally, graph convolution is employed to learn the coupling interactions among high-level features across multiple layers. Comparative experiments conducted on the Sleep-EDF-X datasets demonstrate that GPNet exhibits highly competitive performance compared to other models. Additionally, GPNet predicts post-sleep recognition speed of negative emotions, revealing a negative correlation with REM(%) sleep and suggesting that sleep mitigates negative effects.