Using a generative EMG network to enable compact wearable devices for gesture recognition and gait prediction

Stanford scientists have developed a generative learning framework paired with a compact wearable EMG device that extrapolates limited sensor inputs to reconstruct muscle activity equivalent to that captured by high-density sensor arrays. By reducing the required sensor count from 32 channels to 6 without sacrificing prediction accuracy, GenENet enables practical, comfortable wearable devices for applications such as sign language recognition and gait dynamics monitoring.

Electromyography (EMG) is widely used for non-invasive monitoring of muscle activity and is essential for applications ranging from gesture recognition to gait analysis. Capturing the complex interplay between muscle groups and corresponding body movements requires high-density surface EMG instruments, often with 32 to 256 electrodes. However, increasing the electrode count enlarges the device footprint, raises power consumption, and limits comfort and practicality for everyday use. While advances in electrode materials and signal processing algorithms have improved individual sensor performance, these improvements have not addressed the fundamental trade-off between sensing resolution and wearable form factor. Accordingly, methods that can achieve high-density sensing performance from compact, low-channel-count devices remain a significant unmet need across wearable electrophysiological monitoring.

GenENet is a self-supervised generative algorithm that is pre-trained on a 32-channel stretchable EMG array and then integrated with a compact 6-channel wearable device to reconstruct muscle activity in regions not covered by the smaller sensor. The 6-channel device achieved comparable performance to the full 32-channel system for both sign language gesture recognition and gait force prediction, while reducing power consumption by approximately 71%. Consequently, GenENet has the potential to make high-quality electrophysiological monitoring practical in everyday wearable form factors and is applicable to other sensing domains such as electrocardiography, electroencephalography, and pressure sensing.

Stage of Development:
Proof of Concept