A deep transfer learning approach for wearable sleep stage classification with photoplethysmography

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Wrist-worn wearable photoplethysmography (PPG) offers potential for unobtrusive home sleep monitoring and sleep disorder screening.
• Training deep neural networks for sleep stage classification is challenging due to limited large-scale sleep studies with PPG and gold-standard polysomnography annotation.

Purpose of the Study:

• To develop and evaluate a deep recurrent neural network model for 4-class sleep stage classification using PPG data.
• To investigate the effectiveness of transfer learning strategies in adapting a model trained on electrocardiogram (ECG) data to a PPG dataset.

Main Methods:

• A deep recurrent neural network was initially trained on a large ECG dataset (292 participants) for 4-class sleep stage classification.
• Three transfer learning variations were applied to adapt the trained model to a smaller PPG dataset (60 participants).
• Performance was evaluated using Cohen's kappa and accuracy, comparing transfer learning strategies against baseline models.

Main Results:

• The domain and decision combined transfer learning strategy yielded the best results, achieving a Cohen's kappa of 0.65 ± 0.11 and accuracy of 76.36% ± 7.57%.
• This performance significantly outperformed models trained solely on PPG or ECG data.
• The achieved accuracy for PPG-based 4-class sleep stage classification is unprecedented in existing literature.

Conclusions:

• Transfer learning is a viable and effective method for developing reliable sleep stage classification models for new sensor technologies like PPG.
• This approach brings unobtrusive home sleep stage monitoring closer to clinical application.
• Further research should validate the approach in patient populations with sleep disorders such as insomnia and sleep apnea.