Depth-Camera-Based Under-Blanket Sleep Posture Classification Using Anatomical Landmark-Guided Deep Learning Model.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Sleep Medicine Technology

Background:

• Sleep disorders affect millions globally, necessitating advanced monitoring solutions.
• Current sleep posture monitoring technologies face challenges with occlusions like blankets.
• Deep learning offers potential for non-invasive and accurate sleep analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for improved under-blanket sleep posture classification.
• To investigate the impact of anatomical landmark features on classification accuracy.
• To assess the robustness of the model against varying blanket conditions.

Main Methods:

• Utilized a system with integrated visible light and depth cameras for data acquisition.
• Trained deep learning models (ResNet-34, EfficientNet B4, ECA-Net50) on depth images.
• Incorporated anatomical landmark coordinates derived from visible light pose estimation.
• Compared model performance with and without landmark features across four blanket conditions.

Main Results:

• ECA-Net50 demonstrated superior sleep posture classification performance.
• Integrating anatomical landmark features boosted ECA-Net50's F1 score from 87.4% to 92.2%.
• Landmark-guided models exhibited reduced sensitivity to blanket-induced interference.

Conclusions:

• The proposed deep learning approach with anatomical landmarks significantly enhances under-blanket sleep posture classification.
• This method shows promise for more reliable sleep disorder monitoring systems.
• Leveraging anatomical features improves model resilience to common real-world monitoring challenges.