DSleepNet: Disentanglement Learning for Personal Attribute-Agnostic Three-Stage Sleep Classification Using Wearable...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Non-invasive sleep monitoring is crucial for understanding sleep disorders and related diseases.
• Conventional deep learning models struggle with personal attributes (PAs), limiting generalization.

Purpose of the Study:

• To introduce DSleepNet, a novel deep learning approach for robust sleep stage monitoring.
• To disentangle personal attribute-specific and -agnostic features for improved model generalization.

Main Methods:

• DSleepNet utilizes two probabilistic encoders to separate PA-specific and PA-agnostic features.
• An Independent Excitation mechanism removes correlations within the latent feature space.
• The model operates without requiring target cohort data or PA data during inference.

Main Results:

• DSleepNet's PA-agnostic features improved mean F1 score by up to 8.7% and Cohen's Kappa by 4.7% over baseline CNN.
• Significant reduction in the impact of personal attributes, particularly sleep apnea severity.
• Demonstrated robustness across various personal attribute settings.

Conclusions:

• DSleepNet offers a robust and generalizable solution for non-invasive sleep stage monitoring.
• The disentanglement approach effectively mitigates the influence of personal attributes on model performance.
• This method holds promise for advancing sleep disorder research and clinical applications.