Transfer Learning for Automatic Sleep Staging Using a Pre-Gelled Electrode Grid.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Novel sensor solutions for home sleep monitoring offer potential for continuous observation and new insights.
• Automatic evaluation of data from new sensors is crucial due to differences from classical polysomnography (PSG).
• Limited datasets from new sensor technologies hinder the training of automatic algorithms.

Purpose of the Study:

• To develop an automatic sleep phase detection method for new sensor technologies with limited training data.
• To circumvent high system-specific training data requirements using pre-training and finetuning.
• To enable automatic sleep phase detection for small test series of novel sensor data.

Main Methods:

• Employed pre-training on large, publicly available polysomnography (PSG) datasets.
• Utilized finetuning on a small dataset (12 nights) from a new sensor technology (pre-gelled electrode grid).
• Captured electroencephalography (EEG), electrooculography (EOG), and electromyography (EMG) data; analysis focused on EEG and EOG.

Main Results:

• Achieved an overall F1 score of 0.81 for automatic sleep phase detection.
• Specific F1 scores: wake 0.84, N1 0.62, N2 0.81, N3 0.87, REM 0.88.
• Considered spatial channel distribution and approximated classical electrode positions.

Conclusions:

• Pre-training and finetuning enable accurate automatic sleep phase detection even with small datasets from new sensors.
• The developed method is effective for analyzing data from novel home sleep monitoring technologies.
• This approach facilitates advancements in sleep medicine through accessible and automated sleep analysis.