Fig 1.
Representative raw EEG and EMG epochs with corresponding spectrograms for each sleep stage.
The first two columns display time-series plots of raw EEG and EMG signals for each sleep stage, respectively. The last two columns show spectrograms derived from the same EEG and EMG data, with horizontal black lines indicating stage-specific dominant frequency bands: REM exhibits theta‑band prominence, Wake shows strong EMG activity across higher frequencies, and NREM is characterized by low‑frequency dominance.
Fig 2.
Neural network architectures for sleep stage classification.
Table 1.
Performance Comparison of 1D-CNN, AccuSleep, and DeepSleepNet Using Raw, Z-Scored, and Mixture Z-Scored Data for Within-Subject and Cross-Subject Validation.
Fig 3.
Overall performance comparison of neural network models for sleep stage classification.
One-way ANOVA was conducted to compare the performance of the three networks under the same condition (comparisons limited to adjacent bars; *p < 0.05, **p < 0.01, ***p < 0.001). Each bar represents the mean performance with standard deviation.
Fig 4.
Detailed performance metrics for sleep stage classification across different models and conditions.
One-way ANOVA was conducted to compare the performance of the three networks under the same condition (*p < 0.05, **p < 0.01, ***p < 0.001). Each bar represents the mean performance with standard deviation.
Fig 5.
F1-scores of sleep stage classification for each model with cross-subject data across different normalization conditions.
One-way ANOVA was conducted to compare the performance of the three networks under the same condition (*p < 0.05, **p < 0.01, ***p < 0.001). Each bar represents the mean performance with standard deviation.
Fig 6.
Grad-CAM based saliency maps for the 2D-CNN model.
The first column shows spectrograms of representative EEG epochs for each sleep stage. Next two columns display saliency maps of the first and second convolutional layers, and the last column presents the saliency map of the last convolutional layer, illustrating stage-specific frequency preference.
Fig 7.
First-layer convolutional filters of DeepSleepNet and their peak frequency distribution.
For both narrow and wide convolutional branches, the first 16 filters out of 64 are visualized. The rightmost column shows histogram plots summarize the peak frequencies extracted from all 64 filters, shown separately for narrow and wide convolutional branches. Peak frequency for each filter was defined as the frequency component with the highest amplitude within that filter’s learned weights.
Table 2.
F1-scores reported in the literature for 1D-CNN and CNN + biLSTM models applied to rodent and human sleep datasets, alongside results from the present study.