Robust Sleep Behavior Monitoring Under Motion Artefacts via Motion-Conditioned Learning

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning for Healthcare

Background:

• Motion artifacts significantly hinder accurate wearable sleep behavior analysis in real-world settings.
• Flexible and textile-based sensors struggle with performance degradation during unconstrained body movements.
• Robust sleep monitoring requires addressing challenges posed by motion-rich environments.

Purpose of the Study:

• To develop a motion-conditioned multimodal sensing framework for reliable sleep behavior classification.
• To enhance the robustness of wearable physiological sensing under significant motion.
• To enable adaptive representation learning for diverse motion regimes in sleep analysis.

Main Methods:

• Developed a wearable smart throat band integrating textile strain sensors and inertial sensing modules.
• Introduced MoCo Net, a motion-conditioned deep learning architecture using feature-wise linear modulation (FiLM).
• Synchronously captured neck micro-vibrations and motion dynamics for multimodal analysis.

Main Results:

• Achieved 94.3% classification accuracy under fully unconstrained motion, outperforming strain-only sensing (84.3%).
• Demonstrated strong zero-shot performance (82.1% accuracy) with leave-one-subject-out evaluation.
• Showcased significant improvement to 91.0% accuracy with minimal few-shot adaptation (15 samples/behavior).

Conclusions:

• Explicitly modeling motion context is crucial for robust wearable physiological behavior sensing.
• The proposed framework significantly enhances sleep behavior analysis accuracy in motion-rich scenarios.
• This approach paves the way for more reliable real-world sleep monitoring solutions.