Abstract
An enhanced ear-EEG method could be pivotal in remote monitoring settings, allowing for continuous, non-invasive sleep quality assessment in elderly patients with conditions like dementia or sleep apnea.
Introduction
Polysomnography (PSG) remains the "gold standard" for sleep analysis, yet its clinical application is hindered by the logistical burden of multiple sensors and the "first-night effect," where unfamiliar lab environments bias results. Wearable "hearables" using ear-EEG offer a discreet alternative for long-term monitoring. However, ear-EEG signals differ significantly from traditional scalp-EEG due to electrode distance from cortical activity. Furthermore, while large scalp-EEG datasets exist, ear-EEG data—particularly for older populations prone to sleep disorders—is scarce. This study investigates a transfer learning approach to bridge this gap.
Methods
The study involved 17 older adults (ages 65–83). Sleep data were recorded simultaneously using standard PSG and a custom ear-EEG sensor consisting of a viscoelastic earplug with a flexible electrode.
The researchers employed a feature-based transfer learning model:
- Base Model: The YASA (Yet Another Sleep Analyzer) open-source tool, pre-trained on 31,000 hours of scalp-EEG data using LightGBM.
- Transfer Learning: Instead of full retraining, "continued training" was used, appending new decision trees to the ensemble to capture ear-EEG nuances while retaining broad scalp-EEG knowledge.
- Interpretability: SHAP (SHapley Additive exPlanations) analysis was used to identify influential features, such as delta or beta power, in the model's decision process.
Datasets
The YASA model was pre-trained on seven datasets comprising 3,163 full-night PSG recordings. The study dataset included 13 older adults (mean age 71.8) after excluding those with high electrode impedance.
Key Insights
- N3 Detection Breakthrough: The most significant improvement was in N3 (deep sleep) detection. The baseline scalp-trained model achieved only 48.9% sensitivity on ear-EEG, which rose to 74.1% after fine-tuning.
- Feature Re-prioritization: SHAP analysis showed that fine-tuning shifted the model's focus. For instance, "fast delta power" became the primary feature for N3 detection post-fine-tuning.
- Geriatric Specificity: Older adults exhibit unique sleep architectures. This study highlights the necessity of adapting general models for reliable geriatric care.
Key Results
Fine-tuning the ear-EEG model increased overall classification accuracy from 70.1% to 73.7%. The average Cohen’s kappa score reached 0.639, signifying "substantial" agreement with human experts.
Despite these gains, ear-EEG still trailed scalp-EEG (which achieved 77.3% accuracy), reflecting the inherent signal attenuation within the ear canal.
Limitations and Future Work
The study's small sample size (n=13) limits definitive clinical conclusions. Electrode impedance remains a technical hurdle, requiring improved sensor application protocols. Future research will target larger cohorts with diverse conditions like Alzheimer's or Parkinson's to validate home-based clinical monitoring.
Summary updated