Feature extraction for LSTM-based finger-joint angle estimation from sEMG

Gerda de Jager; Marelie H Davel; Ian Thomson

cs.LGNov 2, 2026

Feature extraction for LSTM-based finger-joint angle estimation from sEMG

Authors: Gerda de Jager, Marelie H Davel, Ian Thomson

Organizations: Faculty of Engineering, North-West University, Potchefstroom, South Africa · Centre for Artificial Intelligence Research, South Africa · National Institute for Theoretical and Computational Sciences, South Africa

Abstract

Continuous estimation of finger-joint angles from surface electromyography (sEMG) is a promising approach for controlling dexterous myoelectric prostheses. Deep time series models have been applied successfully in this context, typically relying on feature extraction rather than raw signals, given limited subject-specific data. Various time-domain and frequency-domain features can be used, extracted for a given window and step size. This study systematically investigates these three design choices for Long Short-Term Memory (LSTM) regression on the Ninapro DB2 dataset. Window sizes from 300 to 1200 ms and step sizes from 10 to 100 ms were evaluated in combination with six time- and frequency-domain features. Models were evaluated using regression performance metrics, inference time, and memory footprint to assess feasibility on embedded microcontroller hardware suitable for a prosthetic device. A 300 ms window with a 100 ms step, paired with either the Mean Absolute Value (MAV) or Root Mean Squared (RMS) feature, yielded the best performance (test PCC of 0.779 +/- 0.080 / 0.779 +/- 0.085). Window size, step size, and feature choice were found to interact non-trivially: jointly tuning them can improve both accuracy and computational feasibility for real-time, embedded myoelectric control.

Figures & tables

	Protocol A		Protocol B
Parameter	Search Space	Distribution	Search Space	Distribution
Dropout	$[0.0,0.1,0.2,0.3]$	–	$[0.0,0.1,0.2,0.3]$	–
Learning rate	$\sigma=1,\ \mu=\log(10^{-3})$	Log-normal	$\sigma=1,\ \mu=\log(10^{-3})$	Log-normal
Width	$[64,256]$	–	$[64,256]$	–
Weight decay	–	–	$\sigma=1,\ \mu=\log(10^{-5})$	Log-normal
Batch size	–	–	$[4,64,1024]$	–

Table 2: HPs and corresponding search spaces for HP tuning Protocols A and B.

Parameter	Value	Parameter	Value
Stacked models	5	Task	Regression
Number of inputs	12 channels	Stateful-ness	Stateless
Number of outputs	10 finger angles	Loss function	MSE
Maximum epochs	200	Extracted feature	RMS
Batch size	64	Optimiser	Adam
Weight decay	$3\times 10^{-6}$	Layernorm	True

Table 3: Static HPs for the LSTM model under Protocol A. Extracted feature, weight decay, and batch size vary under Protocol B (see Table 2 ).

Panel of Figure 2 — Figure 2: Performance of window and step size combinations, using the RMS feature, evaluated on the validation set. For all graphs, darker blue indicates better performance and darker red indicates worse performance.

Panel of Figure 3 — Figure 3: Subject-specific performance per experiment, with the average and standard deviation of validation loss calculated across all experiments. The x-axis ranks the subjects in order of performance for each experiment, with the worst-performing subject plotted on the left (position 6) and the best-performing subject to the right (position 1).

Table 7

Figure 7: Extraction and inference time for each of the six features, at each of the three window-step combinations. Blue indicates feature extraction time, green indicates inference time, and red indicates the total time (the sum of the two).

Test (35)
Work	S - HPP	Loss	$R^{2}$	PCC	RMSE
Long et al. baseline	35 -	-	0.632 $\pm 0.030$	0.749 $\pm 0.017$	9.666 $\pm 0.305$
Long et al. LSTA-conv	35 -	-	0.727 $\pm 0.041$	0.825 $\pm 0.015$	7.926 $\pm 0.325$
Test (12)
Phase 1 Baseline	12 - B	218.775 $\pm$ 91.774	0.389 $\pm$ 0.220	0.693 $\pm$ 0.112	13.868 $\pm$ 2.844
Phase 5 MAV	12 - B	168.550 $\pm$ 70.535	0.544 $\pm$ 0.183	0.779 $\pm$ 0.080	12.181 $\pm$ 2.568

Table 6: Summary of model test and validation performance. The number of subjects (S) and the HP tuning protocol (HHP) is noted along with the four performance metrics, loss,

R^{2}

, PCC, and RMSE. The best validation and test (for 12 subjects) performance is written in boldface, together with all values that lie within one standard error of the best value.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Training
Work	S - HPP	Loss	$R^{2}$	PCC	RMSE
Phase 1 Replication	35 -	11.526 $\pm$ 11.531	0.962 $\pm$ 0.043	0.982 $\pm$ 0.021	2.935 $\pm$ 1.530
Phase 1 Baseline	35 - B	9.395 $\pm$ 7.535	0.973 $\pm$ 0.021	0.989 $\pm$ 0.008	2.735 $\pm$ 1.150
Phase 2 Baseline	6 - A	9.434 $\pm$ 6.471	0.975 $\pm$ 0.018	0.990 $\pm$ 0.008	2.764 $\pm$ 1.180
Phase 4 MAV	6 - A	13.380 $\pm$ 3.319	0.960 $\pm$ 0.015	0.983 $\pm$ 0.007	3.548 $\pm$ 0.428
Phase 4 RMS	6 - A	12.286 $\pm$ 6.529	0.967 $\pm$ 0.011	0.985 $\pm$ 0.005	3.339 $\pm$ 0.847

Appendix

Table 7: Model training performance in the same format as Table 6 .

Experiment	Loss	R2	PCC	RMSE	Inference	Epoch
					time
Validation
Tuned Long et al.	116.280	0.543	0.744	10.229	0.003	94
+ z-norm	91.499	0.642	0.806	9.104	0.002	25
+ $\mu$ -law removed	80.466	0.689	0.840	8.514	0.002	126
Training

Appendix

Table 8: Performance when HP tuning the three pre-processing approaches: pre-processing method of Long et al. vs replacing min-max normalisation with z-norm vs also removing

\mu

-law scaling; result for subject 20, using protocol B.

Surface Electromyography

See all 35 papers

Decoding High-Dimensional Finger Motion from EMG Using Riemannian Features and RNNs
Apr 24, 2026Martin Colot, Cédric Simar, Guy Cheron +23D Hand Pose Estimation Neural Decoding
Machine Learning Group, ULB, Bruxelles, Belgium · § These authors contributed equally to this work and share ﬁrst authorship. · Neuromove, ULB Neuroscience Institute, ULB, Brussels, Belgium +1
Development of a Low-Cost Prosthetic Hand Using Electromyography and Machine Learning
Nov 23, 2024Mosab Diab, Ashraf Mohammed, Yinlai JiangAssistive Robotics Robotic Hand Design
Department of Electrical and Electronic Engineering, Faculty of Engineering, University of Khartoum, Khartoum, Sudan · Center for Neuroscience and Biomedical Engineering, The University of Electro-Communications, Chofu, Tokyo, Japan
Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven
May 20, 2026L. F. Salazar Álvarez, D. Escobar-Saltarén, M. B. Salazar Sánchez +1EMG-Based Gesture Recognition Clustering
In2Lab, Engineer Faculty, Universidad de Antioquia. Medellín, Calle 70 No. 52-21, A.A. 1226 · School of Engineering and Sciencies, Tecnológico de Monterrey. Guadalajara, México

Apr 24, 2026J/K move · Enter open · S save

Decoding High-Dimensional Finger Motion from EMG Using Riemannian Features and RNNs

Martin Colot, Cédric Simar, Guy Cheron, Ana Maria Cebolla Alvarez+1

Machine Learning Group, ULB, Bruxelles, Belgium · § These authors contributed equally to this work and share ﬁrst authorship. · Neuromove, ULB Neuroscience Institute, ULB, Brussels, Belgium+1

3D Hand Pose Estimation Neural Decoding Wearable Sensing

Open paper AskPDF arXiv

Continuous estimation of high-dimensional finger kinematics from forearm surface electromyography (EMG) could enable natural control for hand prostheses, AR/XR interfaces, and teleoperation. However, the complexity of human hand gestures and the entanglement of forearm muscles make accurate recognition intrinsically challenging. Existing approaches typically reduce task complexity by relying on classification-based machine learning, limiting the controllable degrees of freedom and compromising on natural interaction. We present an end-to-end framework for continuous EMG-to-kinematics regression using only consumer-grade hardware. The framework combines an 8-channel EMG armband, a single webcam, and an automatic synchronization procedure, enabling the collection of the EMG Finger-Kinematics dataset (EMG-FK), a 10-h dataset of synchronized EMG and 15 finger joint angles from 20 participants performing rich, unconstrained right-hand motions. We also introduce the Temporal Riemannian Regressor (TRR), a lightweight GRU-based model that uses sequences of multi-band Riemannian covariance features to decode finger motion. Across EMG-FK and the public emg2pose benchmark, TRR outperforms state-of-the-art methods in both intra- and cross-subject evaluation. On EMG-FK, it reaches an average absolute error of

9.79 °\pm 1.48

in intra-subject and

16.71 °\pm 3.97

in cross-subject. Finally, we demonstrate real-time deployment on a Raspberry Pi 5 and intuitive control of a robotic hand; TRR runs at nearly 10 predictions/s and is roughly an order of magnitude faster than state-of-the-art approaches. Together, these contributions lower the barrier to reproducible, real-time EMG-based decoding of high-dimensional finger motion, and pave the way toward more natural and intuitive control of embedded EMG-based systems.

Abstract

Surface Electromyography

Explore similar work

Decoding High-Dimensional Finger Motion from EMG Using Riemannian Features and RNNs

Development of a Low-Cost Prosthetic Hand Using Electromyography and Machine Learning

Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven