Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Device Technologies and Biomedical Robotics

Device Technologies and Biomedical Robotics - Poster Session E

V27 - sEMG Hand and Wrist Movement Time‑Series Classification: Evaluating QUANT and Cosine Features

Saturday, October 11, 2025

10:00 AM - 11:00 AM PDT

Location: Exhibit Hall F, G & H

Presenting Author(s)

ZD

Zachary Deutsch

Undergraduate
Duke University, Department of Biomedical Engineering, United States

Co-Author(s)

ZD

Zachary Deutsch

Undergraduate
Duke University, Department of Biomedical Engineering, United States
ET

Eran Tascesme

Master's Student
Tel Aviv University, School of Industrial & Intelligent Systems Engineering, United States

Primary Investigator(s)

NR

Neta Rabin, PhD

Associate Professor
Tel Aviv University, School of Industrial & Intelligent Systems Engineering, United States

Introduction: : Surface electromyography (sEMG) is a non-invasive technique for recording muscle electrical activity and is widely used for human-machine interfaces in rehabilitation and prosthetics. Robust sEMG-based control is crucial for advanced prosthetic hand function, yet roughly one-third of upper-limb amputees abandon myoelectric prostheses due to limited control reliability and intuitiveness. Improving the classification of sEMG signals can enhance the responsiveness and user acceptance of prosthetic devices. While deep networks have recently delivered high accuracy, their computational complexity and hardware requirements limit real-time, on-device use. We therefore propose a novel, lightweight approach to classify hand and wrist movements from multichannel sEMG using quantile-based time-series features. In particular, we evaluate QUANT, a minimalist interval feature extraction method that computes sorted signal values (quantiles) from fixed-length time intervals. Our objective is to assess whether these features, paired with machine-learning classifiers, can improve movement-classification accuracy for prosthetic control applications.

Materials and

Methods: : We used the publicly available NinaPro DB7 dataset [1]. The DB7 dataset contains myoelectric measurements recorded from 20 subjects (Fig. 1). Twelve Delsys Trigno sensors were placed around the forearm—eight sensors equally spaced around the radiohumeral joint and four sensors on the flexor digitorum, extensor digitorum, biceps, and triceps. sEMG signals were sampled at 2 kHz. Subjects performed two exercises: 17 finger/wrist movements and 23 grasping and functional movements.

We extracted quantile-based interval features using QUANT [2]. Each sEMG time series was divided into fixed intervals at multiple depth levels, with depth controlling the number of features. A set of representative quantiles from each sorted interval formed the features, summarizing the distribution of signal amplitudes. We varied the depth to explore the feature space. We evaluated two classifiers: (i) Extremely Randomized Trees (ExtraTrees), and (ii) a Ridge Classifier (Fig. 2). For the ExtraTrees classifier, we varied “max features,” which is the maximum number of features ExtraTrees considers when splitting a node. We also tested cosine similarity features with ExtraTrees to measure inter-channel similarity.

We performed subject-specific classification: for each subject and resample, each class contributed four training trials, one validation trial, and one test trial. We used 30 resamples for ExtraTrees and Ridge. ExtraTrees used 200 estimators, and Ridge used “alpha” equal to 1. We measured classification accuracy and standard deviation for each subject and recorded computational time for feature extraction and classifier training and prediction.

Results, Conclusions, and Discussions:: QUANT-based interval features classified 40 hand/wrist movements from sEMG with high accuracy. QUANT features with the ExtraTrees classifier had test accuracies ranging from 0.8468 ± 0.0304 at depth 1 (138,108 features) to 0.9234 ± 0.0249 at depth 4 (845,964 features) (Fig. 3a). QUANT features with the Ridge classifier had test accuracies ranging from 0.8554 ± 0.0388 at depth 1 to 0.8956 ± 0.0397 at depth 4 (Fig. 3b). Cosine similarity features (66) with ExtraTrees yielded test accuracies ranging from 0.8473 ± 0.0463 with max features set to 16 to 0.8150 ± 0.0500 with max features set to 1024 (Fig. 3c). ExtraTrees consistently yielded higher classification accuracy than Ridge on QUANT features. Cosine similarity features alone resulted in modest accuracy.

We observed a trade-off between classification accuracy and computational efficiency. All experiments were conducted on an 8-core CPU laptop. The highest-accuracy (0.9234) configuration, 845,964 QUANT features with ExtraTrees, had an 88.64-minute runtime for all 20 subjects. Of that total runtime, transform training took 85.76 minutes, 2.08 minutes on fitting the classifier, and 0.71 minutes on prediction. However, a similar classification accuracy (0.90) configuration of only 138,108 QUANT features with ExtraTrees had a 17.28-minutes compute time for all 20 subjects. Transform training took 15.94 minutes, 0.92 minutes on fitting the classifier, and 0.35 minutes on prediction. Using 66 cosine similarity features with ExtraTrees yielded a 1.36-minutes runtime for all 20 subjects. Transform training took 0.33 minutes, 0.69 minutes on fitting the classifier, and 0.27 minutes on prediction.

QUANT delivers state-of-the-art accuracy for multi-class sEMG and is computationally lightweight. Cosine similarity features with the ExtraTrees classifier delivers moderate-accuracy and low runtime, a suitable alternative for quick prosthetic calibration or low-resource settings. Practically, a prosthetic controller could use QUANT features with the ExtraTrees classifier to maximize accuracy, or fall back to cosine features alone for computational speed. Improving dependable sEMG pattern recognition along this accuracy–latency continuum can reduce control frustration, potentially lowering prosthesis rejection and enhancing user quality of life.

Acknowledgements and/or References (Optional): : [1] M. Atzori, A. Gijsberts, C. Castellini, B. Caputo, A.-G. Hager, S. Elsig, et al., “Electromyography data for non-invasive naturally-controlled robotic hand prostheses,” Scientific Data, vol. 1, no. 1, Art. no. 140053, Dec. 2014, doi: 10.1038/sdata.2014.53.
[2] J. Lines, M. Baydoğan, and A. Bagnall, “QUANT: A Minimalist Interval Method for Time Series Classification,” arXiv preprint arXiv:2308.00928, 2023.
[3] T. Georgoulas and V. Sapsanis, “sEMG for Basic Hand Movements Data Set,” UCI Machine Learning Repository, 2014. https://archive.ics.uci.edu/ml/datasets/SEMG+for+Basic+Hand+movements.
[4] R. N. Khushaba, S. Kodagoda, D. Liu, and G. Dissanayake, “Classification of hand and wrist movements via surface electromyogram using the random convolutional kernels transform,” PLOS ONE, vol. 18, no. 4, Art. no. e0283386, Apr. 2023, doi: 10.1371/journal.pone.0283386.