Attention-Based Adaptive Sampling for Continuous EMG Data Streams

This paper presents an online attention-based adaptive sampling approach for EMG data streams. Our sampling strategy is based on dynamically tuning the duty cycle of an EMG monitoring and recognition system. A response model was developed to adjust the EMG system's sampling rate depending on the signal pattern. The response model was implemented by two sampling rate states. We report a case study of an eyeglasses diet monitoring system that implements the adaptive sampling strategy to monitor the Temporalis muscle activity. We show that the adaptive sampling approach can reduce energy consumption in a free-living study dataset with ten participants. Compared to a static uniform sampling, our approach yields an energy saving on 70%, while recognition performance remained above 80%.