Physics-Informed Transfer Learning to Enhance Sleep Staging.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• At-home sleep staging via wearable sensors offers a cost-effective, less disruptive alternative to in-hospital polysomnography, especially for long-term monitoring.
• Machine learning (ML) for wearable sleep staging faces challenges due to limited data and discrepancies between wearable and in-hospital electroencephalography (EEG) from differing electrode placements.

Purpose of the Study:

• To enhance transfer learning performance for wearable sleep staging by generating synthetic EEG data.
• To address the paucity of wearable sensor data by leveraging electrophysiological models.

Main Methods:

• Utilized Low-Resolution Electromagnetic Tomography Analysis (LORETA) to map real in-hospital EEG data to brain current estimates.
• Generated synthetic wearable EEG by estimating signals from these current estimates at various head locations.
• Augmented existing datasets with this synthetic data to train ML models.

Main Results:

• Statistically significant improvement in wearable sleep staging accuracy, increasing from 80.8% to 81.3% on average.
• Performance gains were dependent on transfer learning parameters and specific data sources used.

Conclusions:

• Synthesized data using physical electrophysiological models can enhance ML performance for sleep staging tasks.
• This approach offers a novel transfer learning method, particularly valuable for scarce EEG data or non-standard electrode configurations.