Deciphering Optimal Radar Ensemble for Advancing Sleep Posture Prediction through Multiview Convolutional Neural...

Area of Science:

• Biomedical Engineering
• Sleep Science
• Sensor Technology

Background:

• Accurate sleep posture assessment is vital for sleep quality evaluation and sleep disorder diagnosis.
• Traditional methods face challenges in low-light and obstructed environments.
• Radar technology offers a promising non-invasive solution for monitoring sleep posture.

Purpose of the Study:

• To determine the optimal number and placement of radar sensors for accurate sleep posture estimation.
• To evaluate different deep learning models for sleep posture classification using radar data.
• To identify the most effective radar sensor configurations for practical applications.

Main Methods:

• Utilized radar sensors to monitor 70 participants in nine distinct sleep postures under varying blanket thicknesses.
• Developed a novel Spatial Radio Echo Map (SREM) technique for multi-radar data fusion.
• Employed a Multiview Convolutional Neural Network (MVCNN) framework with various deep feature extractors (ResNet-50, EfficientNet-50, DenseNet-121, etc.) for posture classification.

Main Results:

• DenseNet-121 achieved the highest accuracy (0.534 for coarse, 0.804 for fine-grained classification).
• A single left-located head radar sensor was found to be optimal, yielding 0.809 accuracy.
• Analysis explored trade-offs between sensor quantity and classification accuracy, with configurations using fewer sensors showing promising results.

Conclusions:

• Established a foundation for optimal radar sensor configuration in sleep posture monitoring.
• Demonstrated the effectiveness of radar technology and deep learning for non-invasive sleep posture assessment.
• Highlighted the potential for simplified sensor setups without significant accuracy loss.