Two-Step Deep Learning for Estimating Human Sleep Pose Occluded by Bed Covers.

Area of Science:

• Computer Science
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate sleep posture identification is crucial for understanding sleep health and developing targeted interventions.
• Existing methods often lack the precision to differentiate between subtle variations in sleep positions.
• Deep learning offers potential for advanced image analysis in sleep studies.

Purpose of the Study:

• To propose and evaluate a novel two-step deep learning method for identifying 12 distinct sleep postures.
• To enhance the accuracy of sleep pose estimation compared to traditional and single-stage deep learning approaches.
• To provide a more refined sleep pose identification system for research and clinical applications.

Main Methods:

• A two-step deep learning architecture was developed, utilizing transfer learning with a VGG-19 convolutional neural network (CNN).
• The VGG-19 network initially classifies images into four broad categories: supine, left, right, and prone.
• Subsequent specialized CNNs further refine these classifications into 12 specific sleep pose labels using infrared images from 10 participants.

Main Results:

• The proposed two-step deep learning method achieved an average accuracy of 85.6% in identifying sleep postures.
• This represents a significant improvement over traditional CNNs trained from scratch (74.5% accuracy) and a single-stage VGG-19 approach (78.1% accuracy).
• The method demonstrated enhanced precision in distinguishing between the 12 pre-defined sleep positions.

Conclusions:

• The novel two-step deep learning approach offers a highly accurate and effective method for sleep pose identification.
• This technique provides a valuable tool for objective sleep monitoring and analysis.
• Further research can explore the integration of this method into real-time sleep tracking systems.