Multimodal Sleep Stage and Sleep Apnea Classification Using Vision Transformer: A Multitask Explainable Learning...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Sleep Medicine

Background:

• Sleep is vital for health, requiring accurate staging and disorder detection for quality assessment.
• Current methods often use single signals and treat sleep stages and disorders separately.
• There's a need for multimodal, multilabel frameworks for comprehensive sleep analysis.

Purpose of the Study:

• To develop a novel 1D-Vision Transformer for simultaneous classification of sleep stages and sleep disorders.
• To leverage correlations between sleep disorders and specific sleep stage patterns.
• To utilize multimodal, multilabel sensory data for improved sleep analysis.

Main Methods:

• Proposed a 1D-Vision Transformer model for integrated sleep stage and disorder classification.
• Trained and tested the model using multimodal-multilabel data: photoplethysmogram, respiratory flow, and respiratory effort signals.
• Analyzed encoder attention weights to understand feature influence on model predictions.

Main Results:

• Achieved 78% accuracy (Cohen's Kappa 0.66) for five-stage sleep classification.
• Achieved 74% accuracy (Cohen's Kappa 0.58) for sleep apnea classification.
• Identified respiratory patterns (troughs, peaks) as significant contributors to classification.

Conclusions:

• The 1D-Vision Transformer effectively performs simultaneous sleep stage and disorder classification.
• The model demonstrates strong performance in identifying sleep stages and sleep apnea.
• Feature analysis highlights the importance of respiratory signals in sleep disorder detection.