AdaptiveEdge: Adaptive Model Update for Motor-Intent Decoding With Knowledge Distillation and Efficient EMG Sensor System

Abstract

Recent advancements in electromyogram (EMG)-based gesture decoding have enabled the development of active rehabilitation devices and enhanced human-machine interaction capabilities. While production-grade EMG sensors offer improved signal-to-noise ratios, their technical complexity necessitate innovative solutions to address inherent limitations. Additionally, EMG-based motor-intent decoders are prone to performance degradation due to factors such as fatigue, electrode shifts, and varying acquisition conditions. To address these challenges, we propose a low-cost EMG sensor grid alongside an advanced decoding strategy named AdaptiveEdge. This adaptive model update strategy integrates offline training with real-time on-device parameter updates, facilitating seamless adaptation to diverse EMG disturbance scenarios. Our comprehensive experiments demonstrated significant accuracy improvements: AdaptiveEdge yielded 10.18% higher accuracy (88.66%) when both offline and on-device update were utilized compared to 78.48% without offline training. Furthermore, AdaptiveEdge not only enhances decoding accuracy but also optimizes memory usage and energy consumption, making it particularly suitable for on-device applications such as neuroprosthetics. These advancements collectively pave the way for more effective and practical EMG-based devices, thereby improving human-machine interaction capabilities. The code associated with this study can be accessed here: https://github.com/deremustapha/AdpativeEdge