Sleep Stage Classification With Multi-Modal Fusion and Denoising Diffusion Model.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Sleep stage classification is vital for assessing sleep quality and preventing sleep disorders.
• Current algorithms struggle with physiological signal noise and suboptimal multi-modal feature fusion.
• Existing methods often concatenate features without accounting for their interdependencies.

Purpose of the Study:

• To propose Diff-SleepNet, an efficient framework for sleep stage classification using multi-modal input.
• To address challenges of noise in physiological signals and feature correlation in sleep data.
• To improve the accuracy and robustness of automated sleep analysis.

Main Methods:

• A diffusion model with peak signal-to-noise ratio (PNSR) loss is employed for adaptive noise filtering.
• Multi-modal signals are transformed into multi-view spectra for feature extraction using a transformer-based backbone.
• A multi-scale attention module is utilized for robust fusion of extracted features.

Main Results:

• The Diff-SleepNet framework demonstrated effectiveness in classifying sleep stages across three datasets (SHHS, Sleep-EDF-SC, Sleep-EDF-X).
• Experimental results indicate superior performance compared to peer methods.
• The proposed approach successfully handles noise and integrates multi-modal information.

Conclusions:

• Diff-SleepNet offers an effective solution for noise reduction and feature fusion in sleep stage classification.
• The framework shows significant advantages over existing methods, enhancing sleep analysis capabilities.
• This work contributes to more accurate and reliable sleep quality assessment and disorder prevention.