From Scalp to Ear-EEG: A Generalizable Transfer Learning Model for Automatic Sleep Scoring in Older People

Area of Science:

• Biomedical Engineering
• Neuroscience
• Sleep Medicine

Background:

• Scalp electroencephalogram (EEG) is a standard for sleep monitoring, with abundant data and established models.
• Emerging modalities like ear-EEG lack extensive datasets and pre-trained models for sleep analysis.
• Older adults, a population often experiencing sleep disturbances, are underrepresented in current ear-EEG research.

Purpose of the Study:

• To investigate the effectiveness of transfer learning from scalp EEG models to ear-EEG for sleep analysis.
• To evaluate the performance of fine-tuned models on an older adult cohort.
• To explore feature-based transfer learning using LightGBM for ear-EEG data.

Main Methods:

• Utilized pre-trained scalp EEG models and applied them to ear-EEG data from 17 older adults (aged 65-83).
• Employed LightGBM for transfer learning, comparing direct application with fine-tuning using ear-EEG data.
• Analyzed classification accuracy and feature importance shifts (SHAP values) for sleep stages.

Main Results:

• Initial accuracy of pre-trained models on ear-EEG was 70.1%; fine-tuning improved accuracy to 73.7%.
• Fine-tuning led to statistically significant accuracy improvements (p < 0.05) for 10 out of 13 participants.
• Enhanced average Cohen's kappa score of 0.639 indicated stronger agreement between automated and expert sleep stage classifications.

Conclusions:

• Fine-tuning pre-trained scalp EEG models on ear-EEG data is a viable strategy to enhance sleep classification accuracy.
• This approach shows particular promise for analyzing sleep in older populations using feature-based transfer learning.
• The study highlights the adaptability of transfer learning across different populations and computational methods for ear-EEG analysis.