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

Motion artefacts remain a major barrier to reliable wearable sleep behavior analysis in real-world environments. Although recent advances in flexible and textile-based sensing technologies have enabled unobtrusive monitoring of respiratory and upper-airway activities during sleep, their performance often degrades substantially under unconstrained body movement conditions. Here, we present a motion-conditioned multimodal sensing framework for robust sleep behavior classification under motion-rich scenarios. A wearable smart throat band integrating textile strain sensors and inertial sensing modules was developed to synchronously capture physiological micro-vibrations and motion dynamics from the neck region. A motion-conditioned deep learning architecture, termed MoCo Net, was further introduced, where inertial motion embeddings modulate strain feature extraction through feature-wise linear modulation(FiLM), enabling adaptive representation learning under different motion regimes. The proposed framework was evaluated on a dataset collected from 24 participants, comprising 93.6 hours of wearable recordings and 33,702 annotated signal segments across four sleep-related behaviors. Under fully unconstrained motion conditions, the proposed method achieved an overall classification accuracy of 94.3%, substantially outperforming strain-only sensing, which achieved 84.3% accuracy under the same conditions. Furthermore, leave-one-subject-out (LOSO) evaluation achieved an average zero-shot accuracy of 82.1%, which further improved to 91.0% after introducing only 15 samples per behavior for few-shot adaptation. These results demonstrate the importance of explicitly modeling motion context for robust wearable physiological behavior sensing in real-world sleep environments.