Deep subdomain adaptation subject-specific sleep staging framework with iterative self-training.

Area of Science:

• Computational Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate sleep staging is crucial for diagnosing sleep disorders and assessing sleep quality.
• Deep neural networks struggle with subject-specific sleep staging due to individual variations (age, gender, ethnicity).
• Domain shift in unlabeled target data presents a significant challenge for generalization in sleep classification.

Purpose of the Study:

• To develop a domain adaptation framework (DDAST) to address the domain shift problem in subject-specific sleep staging.
• To improve the generalization capability of deep learning models for sleep classification across different subjects.
• To enhance the accuracy and reliability of automated sleep staging.

Main Methods:

• Proposed a novel discrepancy-based learning framework (DDAST) for sleep staging.
• Implemented adaptive domain-specific batch normalization to merge source and target domain statistics, especially for small target datasets.
• Combined self-training with discrepancy-based unsupervised learning using target domain pseudo-labels for cross-subject alignment.

Main Results:

• Achieved high accuracy in cross-validation experiments: 89.7% on MASS-SS3 and 84.3% on ISRUC-S3.
• Outperformed existing baseline methods in subject-specific sleep staging.
• Ablation studies confirmed the effectiveness of individual framework modules.
• Feature representation visualization demonstrated improved source and target domain alignment.

Conclusions:

• The DDAST framework effectively addresses the domain shift problem in subject-specific sleep staging.
• This approach shows significant promise for clinical applications in sleep disorder diagnosis and monitoring.
• The developed method enhances the performance of deep learning models in personalized sleep analysis.