[Paper Review] Adversarial Domain Adaptation with Self-Training for EEG-based Sleep Stage Classification.

Emadeldeen Eldele, Mohamed Ragab+5 more|arXiv (Cornell University)|Jul 9, 2021

Domain Adaptation and Few-Shot Learning15 references4 citations

TL;DR

This paper proposes an adversarial domain adaptation framework with self-training for EEG-based sleep stage classification, using unshared attention mechanisms to preserve domain-specific features and a dual-classifier self-training strategy with pseudo-labeling to align fine-grained class distributions across domains. The method achieves state-of-the-art performance across six cross-domain sleep staging scenarios.

ABSTRACT

Sleep staging is of great importance in the diagnosis and treatment of sleep disorders. Recently, numerous data driven deep learning models have been proposed for automatic sleep staging. They mainly rely on the assumption that training and testing data are drawn from the same distribution which may not hold in real-world scenarios. Unsupervised domain adaption (UDA) has been recently developed to handle this domain shift problem. However, previous UDA methods applied for sleep staging has two main limitations. First, they rely on a totally shared model for the domain alignment, which may lose the domain-specific information during feature extraction. Second, they only align the source and target distributions globally without considering the class information in the target domain, which hinders the classification performance of the model. In this work, we propose a novel adversarial learning framework to tackle the domain shift problem in the unlabeled target domain. First, we develop unshared attention mechanisms to preserve the domain-specific features in the source and target domains. Second, we design a self-training strategy to align the fine-grained class distributions for the source and target domains via target domain pseudo labels. We also propose dual distinct classifiers to increase the robustness and quality of the pseudo labels. The experimental results on six cross-domain scenarios validate the efficacy of our proposed framework for sleep staging and its advantage over state-of-the-art UDA methods.

Motivation & Objective

• To address domain shift in EEG-based sleep staging when training and testing data come from different distributions.
• To overcome the limitations of existing unsupervised domain adaptation (UDA) methods that use shared models and global distribution alignment.
• To preserve domain-specific features by employing unshared attention mechanisms in source and target domains.
• To improve classification performance by aligning class-level distributions in the target domain using pseudo-labels.
• To enhance pseudo-label quality through dual distinct classifiers in a self-training framework.

Proposed method

• Introduces unshared attention mechanisms that learn domain-specific representations in the source and target domains separately, preserving unique feature patterns.
• Employs a self-training strategy where the model generates pseudo-labels for the unlabeled target domain data.
• Uses two distinct classifiers to predict pseudo-labels, increasing robustness and reliability of the pseudo-labels.
• Applies adversarial training to align the feature distributions between source and target domains at the class level.
• Optimizes the model end-to-end using a combination of classification loss, adversarial loss, and consistency loss on pseudo-labels.
• Aligns fine-grained class distributions by encouraging the model to produce consistent predictions across the source and target domains.

Experimental results

Research questions

• RQ1 Can unshared attention mechanisms improve domain-specific feature learning in cross-domain EEG-based sleep staging?
• RQ2 Does self-training with pseudo-labels enhance classification performance when domain shift exists?
• RQ3 Can dual classifiers improve the quality and robustness of pseudo-labels in unsupervised domain adaptation?
• RQ4 How effective is the proposed method in aligning class-level distributions across domains compared to global distribution alignment?
• RQ5 Does the framework outperform state-of-the-art UDA methods in multiple cross-domain sleep staging scenarios?

Key findings

• The proposed method achieves state-of-the-art performance across six cross-domain EEG sleep staging scenarios.
• The use of unshared attention mechanisms leads to better preservation of domain-specific features, improving model generalization.
• The self-training strategy with dual classifiers significantly enhances the quality of pseudo-labels, reducing noisy predictions.
• Fine-grained class distribution alignment results in improved classification accuracy compared to global distribution alignment.
• The model demonstrates strong robustness and consistency across diverse subject and device domains in sleep staging.
• Quantitative results show consistent performance gains over existing SOTA UDA methods in all six evaluated cross-domain settings.