This project presents an automated approach for sleep stage classification using electroencephalogram (EEG) signals from the PhysioNet Sleep-EDF Expanded dataset. Manually scoring these signals is time-consuming and subject to variability, motivating the use of machine learning for automation. In this work, classical models and a 1D Convolutional Neural Network are used to classify 30-second EEG epochs into five sleep stages: Wake, N1, N2, N3, and REM. EEG signals were preprocessed and segmented, with handcrafted features extracted for classical models, while the 1D-CNN operated directly on raw EEG signals. Performance was evaluated using accuracy, balanced accuracy, and macro F1-score. Results show that the 1D-CNN outperforms classical methods, particularly in minority classes such as REM, while N1 remains the most challenging stage due to its transitional nature.
Figure 1. 30-second Epoch Segment - REM
- Christian Bammann – Contributor of all data preprocessing, SVM/RF/CNN model development, experimentation, evalutation, and report/presentation preparation.
| File | Description |
|---|---|
figures |
Folder containing visualization figures |
outputs |
Folder containing metrics and plots for each model |
training |
Folder containing all Python training notebooks |
README.md |
Project Overview |
report.pdf |
IEEE-style Report |
requirements.txt |
List of required dependencies |
Run the notebooks in the 'training/' directory:
classical_eeg.ipynbcnn_eeg.ipynb
| Model | Acc | Balanced Acc | Macro F1 |
|---|---|---|---|
| SVM | 86.9% | 79.6% | 72.1% |
| RF | 92.2% | 72.0% | 74.4% |
| 1D-CNN | 93.0% | 83.3% | 79.9% |
| Model | Wake | N1 | N2 | N3 | REM |
|---|---|---|---|---|---|
| SVM | 94.0% | 28.6% | 84.4% | 80.6% | 52.9% |
| RF | 95.6% | 28.0% | 84.5% | 82.9% | 66.3% |
| 1D-CNN | 98.7% | 37.4% | 88.0% | 86.4% | 77.7% |
Link to PhysioNet Sleep-EDF Expanded Dataset: Sleep-EDF Dataset
This work presents an automated approach for sleep stage classification using both classical machine learning models and a 1D convolutional neural network. The 1D-CNN achieved the best overall performance, with notable improvements in macro F1-score and balanced accuracy, demonstrating its ability to better handle class imbalances and capture complex EEG signal patterns. While the model shows strong performance, particularly in the Wake and N2 stages, the N1 stage remains a key challenge to classify due to its transitional and ambiguous nature. These results emphasize that performance evaluation should prioritize class-balanced metrics and clinically meaningful outcomes rather than overall accuracy alone. From a clinical standpoint, accurate classification of sleep stages such as N3 and REM are essential for assessing sleep quality and supporting diagnosis of sleep disorders. The improvements demonstrated in this work highlight the potential of deep learning methods for scalable and consistent sleep analysis