WaveCrossNet: interpretable cross-modal automatic sleep staging for OSA patients

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Obstructive sleep apnea (OSA) presents challenges for automatic sleep staging due to fragmented sleep and frequent stage transitions.
• Existing methods struggle with non-stationary OSA signals, relying on fixed denoising and standard positional encoding, leading to module stacking and poor interpretability.

Purpose of the Study:

• To develop an enhanced Transformer-based model, WaveCrossNet, for accurate and interpretable automatic sleep staging in obstructive sleep apnea patients.
• To address limitations in feature extraction, multimodal correlation modeling, and interpretability of current sleep staging techniques for OSA.

Main Methods:

• Proposed WaveCrossNet, an end-to-end Transformer model integrating adaptive wavelet thresholding and a dual-channel CNN for temporal modeling.
• Implemented energy-adaptive subband thresholding for selective denoising and preservation of transient events.
• Utilized position-aware attention with dynamic cross-modal fusion to model complementary EEG and EOG roles across sleep stages.
• Incorporated hard example mining loss to address class imbalance in minority sleep stages.

Main Results:

• WaveCrossNet achieved 84.9% accuracy on a dataset of 200 SHHS OSA records with subject-independent splits.
• The model demonstrated superior performance compared to state-of-the-art methods on both healthy and OSA datasets.
• Attention weight visualization linked feature correlations to specific physiological events, enhancing model interpretability.

Conclusions:

• WaveCrossNet offers an interpretable, end-to-end solution for accurate sleep staging in obstructive sleep apnea.
• The integrated approach of adaptive denoising, representation learning, and temporal modeling effectively handles non-stationary OSA signals.
• The model's ability to jointly optimize multiple aspects of sleep staging improves performance and interpretability for OSA patients.