Sleep Stage Classification from Physiological Signals

5-class sleep stage classification (Wake, N1, N2, N3, REM) using deep learning on the PhysioNet/Computing in Cardiology Challenge 2018 dataset.

Dataset

• Source: PhysioNet Challenge 2018 — Massachusetts General Hospital Sleep Lab
• Subjects: 1,985 polysomnographic recordings
• Signals: 13 channels (EEG, EOG, EMG, ECG, SaO₂) sampled at 200 Hz
• Labels: 30-second epochs annotated per AASM guidelines

Models Implemented

Model	Accuracy	N3 AUC	Wake AUC	Notes
FFNN (baseline)	56%	0.95	0.86	64 hidden units, ReLU
LSTM	57%	0.96	0.91	64→32 units, dropout 0.3
Transformer	50%	0.95	0.77	Attention-based temporal modeling

Key Features

• Power spectral density extraction (delta, theta, alpha, sigma bands)
• Sleep spindle detection (12–14 Hz sigma band)
• 3-fold cross-validation on training set
• ROC curves and confusion matrices for all models

Project Structure

├── Sleep.ipynb              # FFNN baseline
├── Sleep_with_LSTM.ipynb    # LSTM implementation
├── SleepTransformer.ipynb   # Transformer implementation
└── README.md

Results Discussion

All models achieved similar overall accuracy (~50-57%), but with different class-wise performance:

• N3 (deep sleep): All models perform well (AUC 0.95-0.96) due to distinctive delta waves
• N1 (light sleep): Hardest to classify — frequently confused with Wake and N2
• REM: Poor recall across all models, likely due to class imbalance and reliance on EOG signals

The LSTM showed improved Wake detection (AUC 0.91) by capturing temporal dependencies, while the Transformer struggled with the limited training data.

Tools & Libraries

• PyTorch / Keras
• NumPy, Pandas, Scikit-learn
• MNE (EEG processing)
• Matplotlib, Seaborn

References

Ghassemi MM, et al. "You Snooze, You Win: The PhysioNet/Computing in Cardiology Challenge 2018." Computing in Cardiology Conference (CinC), 2018.