EdgeEMG: On-Device Neural Network Training for Real-Time EMG Pattern Recognition

Jiahua Tang, Philip Liang, Zhuwei Qin

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Gesture Recognition, sEMG, Machine Learning, Neural Network

TL;DR: This paper presents EdgeEMG, the first fully on-device deep learning system for sEMG gesture recognition, capable of real-time training and inference on low-power microcontrollers without external computation

Abstract:

Surface electromyography (sEMG) is a non-invasive
technique that records bioelectrical signals generated by muscle
activity via electrodes placed on the skin. Its ability to capture
a user’s motor intent in real time has enabled a wide range of
applications, including prosthetic control, rehabilitation robotics,
and human-computer interaction. Recent advances in machine
learning (ML), particularly deep learning (DL), have enabled
automated processing of complex biosignals. While DL-based
approaches for sEMG gesture recognition have shown strong
performance on embedded systems, they typically rely on pre-
training models on high-performance computing platforms (e.g.,
PCs or supercomputers) before deployment to low-power devices.
This off-device pretraining limits portability and adaptability, as
it requires prior collection and processing of sEMG data on non-
portable hardware. In this paper, we present EdgeEMG, the first
fully on-device training approach for sEMG gesture recognition
using a deep neural network. Our approach is implemented using
the AIfES library, developed by the Fraunhofer Institute, which
provides efficient operations for feedforward neural networks
and supports deployment on resource-constrained microcon-
trollers. We validate our system on the Sony Spresense MCU
and benchmark it against a conventional Linear Discriminant
Analysis (LDA) classifier. Experimental results demonstrate that
our approach achieves an average real-time accuracy of 70% across
different hand gestures. These findings highlight the feasibility of real-time,
user-adaptive EMG decoding entirely on embedded hardware,
without reliance on external compute resources.

Track: 3. Signal processing, machine learning, deep learning, and decision-support algorithms for digital and computational health

NominateReviewer: Zhuwei Qin, [e-mail]

Submission Number: 142