Real-time implementation of spatial convolutional network to control myoelectric prostheses

Convolutional Neural Networks (CNNs) have been used to classify electromyogram (EMG) signal patterns for diverse assistive technology applications. However, most of these CNN-based models have used pseudo-images as input. Such an approach entails significant computational cost, which hinders the real-life applications. We employed an Arduino-based commercially available board, Nano 33 BLE Sense, to implement all stages, including EMG signal recording, filtering, windowing, feature extraction, and classification, for which we used Tiny Machine Learning (TinyML) to deploy a 1D CNN in hardware. We conducted a real-time gesture recognition experiment with 7 able-bodied participants. Using the proposed embedded system, an accuracy of 94.22% was achieved.