GNN for sEMG gesture recognition: 99% accuracy, 48ms

Pragatheeswaran Vipulanandan, Kamal Premaratne and Manohar Murthi submitted an arXiv paper on 8 July 2026 introducing a graph neural network model for real-time gesture recognition based on surface electromyography signals. The authors evaluated the method on sEMG recorded with a myoband (8 electrodes around the forearm) from eight healthy subjects and report an average classification accuracy of 99% and an average time of 48ms for graph construction plus prediction on an M1 Pro CPU.

How does the model represent sEMG signals?

It converts the myoband sEMG channel readings into graph networks that encode muscle activation patterns across the forearm, then applies a graph neural network for classification. The paper describes a novel sEMG representation using graphs which contain information about muscle activation patterns, built from the myoband’s 8 electrodes placed around the forearm, prior to GNN inference.

The representation step produces a structured input that the authors feed to a GNN tailored for gesture classification. The arXiv entry lists the system pipeline as: sEMG capture from the myoband electrodes, graph construction capturing spatial activation relationships, and GNN-based real-time prediction.

How well does the model perform?

The proposed method achieved an average classification accuracy of 99% on data from eight healthy subjects and required an average of 48ms for graph construction and prediction when run on an M1 Pro CPU, the paper states. The authors add that this performance surpasses the performance of state-of-the-art techniques.

Those two concrete numbers anchor the evaluation: 99% average accuracy and 48ms average latency for the combined graph construction and prediction step. The dataset used for evaluation was explicitly sEMG signals acquired with a myoband device that has 8 electrodes placed around the forearm, and the participant pool comprised eight healthy subjects.

Why does it matter?

The paper frames the work around a practical need: "For seemless control of advanced hand prostheses and augmented reality, accurate and immediate hand gestures recognition is essential." Low latency and high accuracy are both needed for responsive prosthesis control and interactive AR experiences. Achieving 99% accuracy with a 48ms end-to-end graph-plus-inference time on a laptop-class M1 Pro suggests the approach could be deployed on consumer or embedded hardware for real-time control.

Beyond immediate device responsiveness, representing sEMG as graphs emphasizes spatial activation patterns rather than treating channels in isolation, which can improve robustness to electrode placement and inter-subject variability if validated at scale. The authors also connect the work to prior conference material via a related DOI listed on the arXiv page (https://doi.org/10.1109/EMBC53108.2024.10781990).

What to watch

Look for public release of code, models or datasets linked from the arXiv entry: the paper’s arXiv page includes toggles and links for code and data platforms such as Hugging Face, DagsHub and Replicate. Replication on larger, more diverse subject pools and independent comparisons against established sEMG benchmarks will be the next concrete tests of the method and of the claim that it "surpass[es] the performance of state-of-the-art techniques."