Comparison of deep transfer learning algorithms and transferability measures for wearable sleep staging
Samuel H Waters1, Gari D Clifford2
1Department of Bioengineering, Georgia Institute of Technology, Atlanta, United States. swaters36@gatech.edu.
Biomedical Engineering Online
|September 12, 2022
Summary
Transfer learning significantly improves electroencephalogram (EEG)-based sleep staging using wearable sensors. Retraining neural network head layers and using transferability measures are key to enhancing performance for remote medical data acquisition.
Area of Science:
- Biomedical Engineering
- Machine Learning
- Signal Processing
Background:
- Wearable sensors offer a promising alternative to in-hospital monitoring for medical data acquisition.
- A significant challenge in developing wearable sensor technology is the limited availability of relevant datasets.
- Transfer learning, utilizing in-hospital recordings, can enhance the performance of wearable sensor systems.
Purpose of the Study:
- To evaluate the effectiveness of various supervised transfer learning algorithms for electroencephalogram (EEG)-based sleep staging.
- To assess the utility of transferability measures in predicting the performance of transfer learning models on unseen medical data.
- To investigate optimal transfer learning strategies for sleep staging using single-channel EEG data from commercial wearable systems.
Main Methods:
- Two neural network architectures (bespoke and open-source) were pre-trained on six large clinical polysomnogram (PSG) datasets.
- Models were subsequently re-trained on a target dataset of 75 full-night, single-channel EEG recordings from 24 subjects.
- Several transferability measures were evaluated for their correlation with model accuracy on the target dataset.
Main Results:
- Transfer learning improved performance on the sleep staging task, with re-training the head layers of the neural networks being most effective in 63.9% of cases.
- Transferability measures demonstrated significant correlations with predictive accuracy, indicating their utility in assessing model performance.
- The study successfully applied transfer learning to enhance sleep staging using limited in-home EEG data.
Conclusions:
- Re-training the head layers of neural networks is a highly effective strategy for boosting performance in EEG-based sleep staging via transfer learning.
- Transferability measures serve as valuable indicators for predicting the success of transfer learning applications in medical signal analysis.
