Transfer Learning for Clinical Sleep Pose Detection Using a Single 2D IR Camera.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Science

Background:

• Sleep quality significantly impacts health and well-being.
• Accurate, large-scale sleep quality quantification is needed for diagnosis and epidemiology.
• Body posture is a key indicator of sleep quality.

Purpose of the Study:

• To develop and implement a non-contact sleep monitoring system analyzing body posture and movement.
• To quantify sleep poses using supervised machine learning on infrared camera data.
• To evaluate the system's performance against manual scoring and polysomnography.

Main Methods:

• A non-contact, vision-based infrared camera system was designed.
• Supervised machine learning with transfer learning (ResNet-152) was applied to infrared data.
• The system detected four predefined poses and an empty bed state during overnight sleep.

Main Results:

• The ResNet-152 model outperformed de novo CNN and other pre-trained networks.
• The system achieved superior sleep pose estimation compared to other video-based methods.
• Performance exceeded the clinical standard using polysomnography position sensors.

Conclusions:

• Infrared video capture combined with deep learning AI can accurately quantify sleep poses and transitions in realistic conditions.
• This non-contact approach offers superior pose estimation compared to current clinical methods.
• The technology enables scalable diagnosis and epidemiology of poor sleep.