We aim to provide a realistic assessment of sleep stage classification (SSC) models in real-world label-scarce scenarios. Most of the proposed deep learning-based models assume access to a vast amount of labeled data, which is not usually achievable.
We explored the efficacy of self-supervised learning (SSL) algorithms in scenarios with limited labeled data. We examined existing sleep stage classification models, evaluated their performance under the few-labeled data settings, and explored the efficacy of different SSL algorithms (pretext and contrastive methods) to improve their performance.
We used the Sleep Heart Health Study, where we randomly chose 20 subjects from the patients during the first visit (SHHS-1 dataset). We selected 1 EEG channel, i.e., C4-A1 with a sampling rate of 125 Hz.
The results suggest that the performance of existing SSC models degrades in the few-labels regime. However, self-supervised pretraining with contrastive methods ensures improved performance against supervised training under the same settings. In addition, SSL algorithms improved the models’ capacity to learn temporal information in EEG data. Notably, fine-tuning the SSL-pretrained models with 5 or 10% of labels can achieve very close performance to the supervised training with 100% of labels. In addition, contrastive SSL algorithms are more robust to dataset imbalance, and have better transferability when a domain shift exists.
The code of this evaluation is available at https://github.com/emadeldeen24/eval_ssl_ssc. The code is also generic and can be easily customized for other sleep stage classification models.
Dr. Emadeldeen Eldele, Nanyang Technological University and Centre of Frontier AI Research, A*STAR
Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C., & Li, X. (2023). Self-Supervised Learning for Label Efficient Sleep Stage Classification: A Comprehensive Evaluation. IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, 31, 1333–1342. https://doi.org/10.1109/TNSRE.2023.3245285