Mamba-based deep learning approach for sleep staging on a wireless multimodal wearable system without...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Wearable technology offers a non-intrusive method for collecting physiological data relevant to sleep staging.
• Traditional polysomnography (PSG) requires specialized equipment and a clinical setting, limiting widespread sleep analysis.
• Developing accurate sleep staging algorithms for wearable sensors is crucial for remote and accessible sleep monitoring.

Purpose of the Study:

• To evaluate a Mamba-based deep learning model for sleep staging using data from the ANNE One wearable system.
• To assess the performance of the model across different sleep stage classifications (3, 4, and 5 classes).
• To determine the feasibility of using non-EEG wearable data for clinical sleep analysis.

Main Methods:

• Wearable sensor data (ECG, accelerometry, temperature, PPG) were collected from 357 adults concurrently with PSG.
• A Mamba-based recurrent neural network architecture was trained on the wearable sensor data, using manually scored PSG as ground truth.
• Ensembling techniques were applied to model variants to enhance performance.

Main Results:

• The ensembled Mamba model achieved a 3-class balanced accuracy of 84.02% and a Cohen's kappa of 72.89%.
• For 4-class sleep staging, the model reached a balanced accuracy of 75.30% and an F1 score of 74.10%.
• Performance varied with more granular sleep stage classifications, indicating potential for further model refinement.

Conclusions:

• The Mamba-based deep learning model effectively infers major sleep stages from the ANNE One wearable system.
• This approach provides a viable, non-EEG alternative for sleep staging in adults within a clinical sleep lab setting.
• The findings support the application of this wearable technology and AI model for broader sleep health monitoring.