Adaptive automatic sleep stage classification under covariate shift.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Medicine

Background:

• Current automatic sleep stage classification (ASSC) methods using polysomnographic (PSG) signals struggle with inter-subject variability, limiting their reliability.
• This variability poses a challenge for applying models trained on one group of subjects to new, unseen individuals.

Purpose of the Study:

• To propose a novel adaptive sleep scoring method robust to inter-subject variability.
• To enhance the reliability of automatic sleep stage classification (ASSC) for diverse subjects.

Main Methods:

• Applied Maximum Overlap Discrete Wavelet Transform (MODWT) to extract features from EEG, EOG, and EMG signals.
• Utilized Minimum-Redundancy Maximum-Relevance (mRMR) for significant feature selection.
• Employed Importance Weighted Kernel Logistic Regression (IWKLR) for domain adaptation in classification.

Main Results:

• The proposed unsupervised domain adaptation method demonstrated robustness to inter-subject differences.
• Leave-One-Out Cross-Validation (LOOCV) confirmed the method's state-of-the-art performance in ASSC.
• The covariate shift model effectively addressed variations in sleep feature distributions between training and testing phases.

Conclusions:

• The novel adaptive sleep scoring method significantly improves ASSC reliability by addressing inter-subject variability.
• Unsupervised domain adaptation offers a promising approach for developing more generalizable sleep analysis tools.
• The combination of MODWT, mRMR, and IWKLR provides an effective framework for robust automatic sleep staging.