SparseEMG: Sparse sEMG for Gesture Decoding

• SparseEMG is a set of methods that impose sparsity on electrode layouts and signal representations to retain key neuromuscular information with fewer channels.
• It integrates approaches like sparse sensor selection, low-density sensing, and event-driven computations to enhance gesture recognition accuracy and reduce computational load.
• Demonstrated in prosthetic control and silent-speech interfaces, SparseEMG improves efficiency, reduces false positives, and accelerates processing without significant loss in decoding quality.

SparseEMG denotes a family of surface electromyography (sEMG) methods in which sparsity is imposed at the level of electrode layouts, signal representations, or computation in order to reduce hardware burden, computation, or both without unacceptable degradation in decoding quality. In its narrowest and most explicit usage, SparseEMG is a computational design tool that generates sparse electrode layouts based on user-selected gesture sets, electrode constraints, and machine learning parameters while also predicting classification performance (Kumar et al., 7 Aug 2025). In the broader literature, closely related work includes sparse sensor selection for prosthetic control, low-density gesture recognition, multiscale operator truncation, sparse autoencoding, spike-driven attention, and manifold methods that remain effective with reduced channel counts (Ren et al., 2022, Hristov et al., 30 Jan 2026, Zhou et al., 2024, Gowda et al., 2023).

1. Scope and organizing concepts

A useful synthesis is to distinguish several recurrent sparsity regimes in the sEMG literature. These regimes differ in what is made sparse—channels, bases, latent activations, or computations—but they share the goal of retaining task-relevant neuromuscular information with fewer degrees of freedom.

Sparsity regime	Main mechanism	Representative work
Layout sparsity	Electrode ranking and subset design	(Kumar et al., 7 Aug 2025, Ren et al., 2022)
Low-density sensing	Accurate decoding with few channels	(Crouch et al., 2021, Hristov et al., 30 Jan 2026)
Representation sparsity	Sparse bases, latent codes, or feature selection	(Zhou et al., 2024, Zheng et al., 2021)
Event-driven sparsity	Spike encoding or active-query computation	(Garg et al., 2021, Zhou et al., 2024)

The review literature places these developments in the context of sEMG as a non-invasive human-machine interface. It describes a canonical pipeline of preprocessing, feature extraction or selection, gesture detection, and classification, and argues that multi-channel recordings often contain high information content in a few key channels or time-points, suggesting that sEMG is often sparse in an appropriate basis (Zheng et al., 2021). The same review also notes logarithmic improvement between channel count and gesture recognition accuracy, with diminishing returns beyond a certain number of channels, which makes sparse sensing and sparse feature selection technically attractive rather than merely economical (Zheng et al., 2021).

2. Sparse sensor selection in prosthetic control

One of the clearest formalizations of SparseEMG appears in work on robotic prosthesis controllers, where sensor selection is posed inside a historical Functional Linear Model (FLM). In that setting, the movement response at time $t_i$ is related to recent EMG history through

$$y_i = \sum_{k=1}^K \int_{\mathcal{T}} X_{ik}(\tau)\,\gamma_k(\tau, z_i)\,d\tau,$$

and sensor selection is induced by a penalized objective in which group LASSO penalties drive entire coefficient functions, and therefore entire sensors, to zero (Ren et al., 2022).

The original Sequential, Adaptive Functional Estimation (SAFE) method repeatedly solves this estimation problem in a multi-stage adaptive fashion, but its dense representation matrix makes the computation expensive. The Multiscale SAFE (MSAFE) variant replaces dense single-scale spline bases with a multiscale piecewise polynomial basis, yielding an approximately sparse representation matrix that can then be truncated. In the notation of the paper, the kernel expansion is

$$\gamma_k(t, z) = \sum_{j=1}^{(p+1)2^n}\sum_{l=1}^q b_{jlk}\,w_j(t)\,s_l(z),$$

and the discretized design matrix is

$$A_{ik}[jl] = s_l(z_i)\int_0^1 X_{ik}(\tau) w_j(\tau)\,d\tau.$$

The truncation step keeps low-level coefficients and zeros the rest, producing a sparse matrix with theoretical error decay of order $2^{-mp}$ (Ren et al., 2022).

Empirically, MSAFE saves 85%$\sim$90% computing time compared with SAFE while producing better sensor selection and comparable accuracy on real hand movement data (Ren et al., 2022). The focused summary reports an example in which SAFE required 505 min and MSAFE 47 min for Constant #1 finger movement, and states that sparse matrices accelerate matrix operations by at least an order of magnitude (Ren et al., 2022). In simulation with correlated noise, MSAFE selected fewer spurious sensors, reduced the mean number of false positive sensors by up to 90% versus SAFE, and produced mean cross-validation errors that were 70–80% those of SAFE across varied noise scenarios (Ren et al., 2022). A common misconception is that sparsification necessarily trades accuracy for speed; in this FLM setting, truncation also acted as a filtering mechanism and improved robustness against noise (Ren et al., 2022).

3. SparseEMG as a computational layout design tool

The paper “SparseEMG: Computational Design of Sparse EMG Layouts for Sensing Gestures” defines SparseEMG as a design environment for sparse electrode layout synthesis rather than a single classifier (Kumar et al., 7 Aug 2025). Electrode selection is cast as a feature subset selection problem, and the system systematically evaluates 28 combinations formed by 4 electrode selection schemes and 7 classifiers across six datasets: CSL-HDEMG, Hyser, DELTA, Nizamis et al., GrabMyo, and putEMG (Kumar et al., 7 Aug 2025).

The four selection schemes are Mutual Information, Permutation Importance, Root Mean Square Importance, and SHapley Additive exPlanations; the seven classifiers are Random Forest, Support Vector Classifier, Logistic Regression, K-Nearest Neighbors, Gaussian Naive Bayes, XGBoost, and Multi-Layer Perceptron (Kumar et al., 7 Aug 2025). For a candidate layout with $E$ electrodes, the tool optimizes a custom trade-off metric,

$$\text{Sparsity Score}(E)=w_1\cdot(100-\text{Accuracy}(E))+w_2\cdot E,\qquad w_1+w_2=1,$$

and selects the lowest-scoring configuration (Kumar et al., 7 Aug 2025).

Across the multi-dataset evaluation, Permutation Importance paired with Random Forest was identified as the best practical combination, reducing the number of electrodes by 53.5% (Kumar et al., 7 Aug 2025). The detailed evaluation reports, for CSL-HDEMG, a reduction from 192 to 76 electrodes with peak accuracy 94.04%, and under a practical limit of at most 20 electrodes, 88.33% accuracy with 19 electrodes (Kumar et al., 7 Aug 2025). The same source states that SparseEMG layouts outperformed conventional band layouts by +9% with 16 electrodes, +3% with 8 electrodes, and +1% with 4 electrodes on average, with the largest benefit in the few-electrode regime (Kumar et al., 7 Aug 2025).

SparseEMG supports 50+ unique gestures and is validated in three real-world applications using different hardware setups (Kumar et al., 7 Aug 2025). It also allows user constraints such as maximum number of electrodes, allowed placement regions, and spacing constraints, and it can generate placement stencils tailored to anatomy (Kumar et al., 7 Aug 2025). The reported transfer results indicate that layouts generated by the design tool are transferable across users with only minimal variation in gesture recognition performance, with within-dataset drops typically in the 2–4% range and some cases up to 6–8% (Kumar et al., 7 Aug 2025). This suggests that sparse layout design can generalize even when end-to-end cross-subject decoding remains difficult.

4. Low-density sensing and the feasibility of few-channel EMG

A central question in SparseEMG is how far electrode count can be reduced before task performance degrades materially. Several studies indicate that low-density operation is feasible when the model architecture is matched to the temporal structure of the signal. In natural typing recognition from forearm sEMG, a temporal-first architecture operating on raw 2000 Hz signals achieved mean character-level accuracy of 90.93% on held-out data, while synthetic spatial downsampling showed that high accuracy was retained with as few as 12 channels per arm and about 85% accuracy with only 4 channels per arm (Crouch et al., 2021). The same study found that recognition remained above 85% down to 200 Hz sampling, which implies that sparse hardware and lower-bandwidth acquisition can remain viable if sequence modeling is sufficiently strong (Crouch et al., 2021).

Even more aggressive reduction appears in low-density prosthetic control. A Convolutional Sparse Autoencoder (CSAE) framework trained on only two sEMG channels achieved a multi-subject F1-score of 94.3% $\pm$ 0.3% on a 6-class gesture set (Hristov et al., 30 Jan 2026). Its few-shot transfer protocol improved performance on unseen subjects from a baseline of 35.1% $\pm$ 3.1% to 92.3% $$y_i = \sum_{k=1}^K \int_{\mathcal{T}} X_{ik}(\tau)\,\gamma_k(\tau, z_i)\,d\tau,$$0 0.9% with minimal calibration data, and an incremental learning strategy extended the system to a 10-class set with a 90.0% $$y_i = \sum_{k=1}^K \int_{\mathcal{T}} X_{ik}(\tau)\,\gamma_k(\tau, z_i)\,d\tau,$$1 0.2% F1-score without full retraining (Hristov et al., 30 Jan 2026). The reported deployment profile—0.74 MB static model size, 0.93 MB runtime footprint, 0.3 GFLOPS, and 6.4 ms inference latency on a standard CPU—shows that sparse sensing and sparse latent structure can be coupled with stringent compute budgets (Hristov et al., 30 Jan 2026).

Hardware-oriented evidence points in the same direction. An in-house 8-channel surface-EMG system for upper- and lower-limb movements found that small, well-placed channel subsets were often sufficient, and that for many tasks the number of optimally placed electrodes matched or slightly exceeded the number of movement classes (K et al., 17 Sep 2025). Reported examples include up to 100% accuracy for knee flexion versus extension with 2 electrodes, about 90% for distinguishing elbow linear versus rotational movements with 2 electrodes, and around 80% for three-class shoulder and ankle movement groupings with 3 electrodes (K et al., 17 Sep 2025). These results do not imply that all tasks are equally compressible; rather, they show that channel count interacts strongly with gesture taxonomy and placement.

5. Sparse representations, sparse computation, and event-driven models

SparseEMG is not limited to channel pruning. It also includes methods in which the representation itself is compressed or made sparse. The review literature explicitly frames sparse representation, sparse coding, and sparsity-enforced algorithms as natural responses to high-dimensional multi-channel recordings, and discusses LASSO-style objectives, sparse channel selection, pruning, compressed sensing, and on-sensor machine learning as part of the sEMG design space (Zheng et al., 2021).

One recent example is SAFE-Net for sEMG-based joint angle estimation. SAFE-Net introduces a Spiking Sparse Attention Encoder (SSAE) that compresses sEMG signals into neural spiking forms and performs attention only for active queries, replacing many floating-point multiplications with sparse addition, followed by a Spiking Attentional Feature Decomposition module that separates kinematic features from biological features (Zhou et al., 2024). On two datasets, SSAE achieved significant power consumption savings of 39.1% and 37.5% over Informer and Spikformer, respectively, in terms of inference costs, while SAFE-Net surpassed both baselines in recognition accuracy (Zhou et al., 2024). The same report states that RMSE was consistently lower, with reductions of 12–59% depending on dataset, joint, and setting (Zhou et al., 2024).

Event-driven sparsity is developed differently in neuromorphic reservoir computing. In “Signals to Spikes for Neuromorphic Regulated Reservoir Computing and EMG Hand Gesture Recognition,” continuous sEMG is converted into UP and DOWN spike trains through temporal contrast encoding, and the resulting spike streams are processed by a biologically inspired spiking reservoir regulated by the CRITICAL algorithm to operate at the edge of chaos (Garg et al., 2021). The regulated reservoir achieved 89.72% on the Roshambo EMG dataset and 70.6% on the EMG subset of the Sensor Fusion dataset, exceeding both the spike-encoding baseline and prior spiking or convolutional neural models reported in that comparison (Garg et al., 2021). Here sparsity is computational rather than anatomical: activity is generated only for significant temporal changes.

Sparse or lower-dimensional intermediate targets also appear in speech-related sEMG. An encoding study comparing Speech Articulatory Coding (SPARC) features with phoneme one-hot features showed that SPARC linearly predicts sEMG envelopes more accurately across aloud, mimed, and subvocal speech, contributes substantially more unique variance than phoneme features, and yields anatomically interpretable weight maps stable across modes (Le et al., 20 Apr 2026). Because SPARC is lower-dimensional and more directly related to muscle movements, the authors argue that it supports sparser and more interpretable sEMG modeling for silent-speech interfaces (Le et al., 20 Apr 2026).

6. Geometry, transfer, and demographic robustness

SparseEMG methods face a persistent tension between transferability of layouts and non-transferability of raw signal geometry. Manifold-based work on sEMG covariance structure provides a precise account of this issue. In “Topology of surface electromyogram signals: hand gesture decoding on Riemannian manifolds,” each trial is embedded as a covariance matrix

$$y_i = \sum_{k=1}^K \int_{\mathcal{T}} X_{ik}(\tau)\,\gamma_k(\tau, z_i)\,d\tau,$$2

which lies on the manifold of symmetric positive definite matrices (Gowda et al., 2023). On 12-channel datasets, Minimum Distance to Mean and SVM classifiers operating in Cholesky space achieved accuracies such as 0.92 and 0.93 on Ninapro, and 0.82 and 0.86 on UCD-MyoVerse-Hand-0, while remaining computationally light and interpretable (Gowda et al., 2023). The same study quantified distribution shift across individuals, showed that direct classification across subjects without alignment fails, and used parallel transport to map gesture clusters into a common reference frame (Gowda et al., 2023).

The demographic dataset of 91 adults extends this point by explicitly benchmarking dense and sparse layouts under inter-individual variability. Using 12 electrodes, 8 forearm electrodes, or 4 wrist electrodes, the reported decoding accuracies in the 110–230 Hz band were 0.825, 0.67, and 0.55, respectively, against random chance 0.1 for 10-way classification (Gowda et al., 2024). The same work found that power spectrum is concentrated at low frequencies but higher bands provide better decodability, and that demographic confounds were small above 110 Hz: the regression of k-medoids accuracy on age, skin hydration, elasticity, and BMI was significant only for the 20–50 Hz band, with $$y_i = \sum_{k=1}^K \int_{\mathcal{T}} X_{ik}(\tau)\,\gamma_k(\tau, z_i)\,d\tau,$$3 and $$y_i = \sum_{k=1}^K \int_{\mathcal{T}} X_{ik}(\tau)\,\gamma_k(\tau, z_i)\,d\tau,$$4, whereas higher bands showed smaller and non-significant effects (Gowda et al., 2024). A key conclusion was that high-frequency components above 110 Hz provide better decoding and are robust to demographic and anatomical variation (Gowda et al., 2024).

These findings complicate a common assumption that sparse layouts automatically solve generalization. The layout optimization results in SparseEMG indicate cross-user transferability of recommended electrode sets with minimal variation in recognition performance (Kumar et al., 7 Aug 2025), but the manifold and demographic studies show that the underlying covariance bases are highly individual-specific, and that cross-subject generalization of spatial sEMG patterns is not feasible in a naive zero-shot sense (Gowda et al., 2024, Gowda et al., 2023). Complementary work on domain adaptation for sparse sEMG gesture recognition addresses this gap directly: a two-stage recurrent model with a linear adaptation layer improved inter-subject accuracy on NinaPro DB-1 from 35.1% without adaptation to 65.2% using only 50% of a new user’s data for adaptation (Ketykó et al., 2019).

Taken together, the literature suggests that SparseEMG is best understood not as a single method but as a design principle for balancing information content, sensor burden, and computational load in sEMG systems. Sparse channel selection, sparse latent structure, and sparse computation have all been shown to preserve or improve performance under particular modeling assumptions, but the success of sparsification depends on gesture set, anatomy, spectral regime, and adaptation strategy (Zheng et al., 2021, Kumar et al., 7 Aug 2025).