From unsupervised to semi-supervised adversarial domain adaptation in electroencephalography-based sleep staging
Elisabeth R M Heremans1, Huy Phan2, Pascal Borzée3
1STADIUS Center for Dynamical Systems, Signal Porcessing and Data Analytics - Department of Electrical Engineering (ESAT), KU Leuven, Kasteelpark Arenberg 10, Leuven, 3001, Belgium.
Journal of Neural Engineering
|May 4, 2022
Summary
Adversarial domain adaptation improves automated sleep stage classification from wearable devices, even with limited patient data. This transfer learning approach enhances accuracy and enables personalized sleep monitoring.
Area of Science:
- Biomedical Engineering
- Machine Learning
- Sleep Science
Background:
- Wearable sleep monitoring generates vast datasets, but limited labels hinder automated sleep stage classification.
- Transfer learning and domain adaptation are key for developing data-efficient sleep analysis models.
Purpose of the Study:
- Investigate adversarial domain adaptation for sleep staging using wearable data from patient populations.
- Evaluate the impact of target domain labels and domain mismatch on model performance.
- Assess the personalization capabilities of adversarial domain adaptation for individual patients.
Main Methods:
- Applied adversarial domain adaptation to wearable sleep datasets from diseased patients.
- Examined the influence of pseudo-labels and real target labels.
- Analyzed the effect of domain mismatch between source and target data.
- Implemented unsupervised adversarial domain adaptation for personalization.
Main Results:
- Adversarial domain adaptation significantly improved sleep staging accuracy by 7%-27% compared to non-adapted models.
- Performance gains were further enhanced by incorporating pseudo-labels and real target labels.
- Unsupervised domain adaptation achieved 1%-2% performance improvement for personalized sleep models.
Conclusions:
- Adversarial domain adaptation offers a flexible framework for semi-supervised and unsupervised transfer learning in sleep staging.
- This method is highly valuable for analyzing wearable electroencephalography data in clinical and research settings.
