Surface Electromyography (sEMG)

• Surface electromyography (sEMG) is a non-invasive technique that records electrical activity from skeletal muscles using skin-surface electrodes.
• Robust acquisition hardware and advanced signal processing, including denoising and feature extraction, enable high-fidelity muscle state and intent inference.
• Integration with machine learning enhances applications in prosthetic control, rehabilitation, and human–machine interfaces through reliable gesture recognition.

Surface electromyography (sEMG) is a non-invasive technique for recording the electrical activity produced by skeletal muscles during contraction from electrodes placed on the skin surface. sEMG has become a core methodology in biomedical engineering, neuroscience, rehabilitation, bionics, and human–machine interface research due to its ability to quantify spatiotemporal patterns of neuromuscular activation with high temporal resolution and its suitability for wearable applications. The sEMG signal is intrinsically stochastic, low in amplitude (10 μV–10 mV), and vulnerable to physiological and technical noise, necessitating robust acquisition hardware and advanced signal processing for high-fidelity muscle state and intent inference (Zheng et al., 2021, Manjarres-Triana et al., 2023).

1. Biophysical Basis and Signal Characteristics

sEMG arises from the summation of motor unit action potentials (MUAPs) generated by individual motor neurons and their associated fibers. The net sEMG waveform $x(t)$ can be modeled as the convolutional sum:

$x(t) = \sum_{i=1}^{M}[a_i(t) * h_i(t)],$

where $a_i(t)$ is the spike train of motor unit $i$, $h_i(t)$ its fiber's impulse response, and $M$ is the number of recruited motor units. Under quasi-stationary assumptions (windowing $<$500 ms), the spectrum typically spans 20–500 Hz, with amplitude attenuated by tissue, fat, and the skin–electrode interface (Manjarres-Triana et al., 2023).

Physiological and anatomical attributes—tissue composition, motor unit recruitment order, firing rates, muscle fiber orientation—directly influence the amplitude, spectral content, and spatial distribution of sEMG. Precise electrode placement, typically midway between the motor point and tendon along the muscle belly, ensures maximal signal while minimizing crosstalk (Gowda et al., 2024). The standard measurement setup involves wet Ag/AgCl electrodes in bipolar or high-density arrays, yielding impedance below 200 Ω for gels, but much higher (kΩ–MΩ) for dry or textile-based electrodes (Zheng et al., 2021, Naim et al., 2020).

2. Acquisition Systems and Hardware Architectures

Contemporary sEMG front-ends fall into several classes:

• Benchtop/clinical systems: Up to 128 channels, wet electrode grids, multi-stage analog pre-filtering, sampling rates typically 1–2 kHz up to 5 kHz (Gowda et al., 2023, Gowda et al., 2024, Gowda et al., 2024).
• Wearable and low-cost modules: Dry or active electrodes, 4–16 channels, power-optimized microcontroller-based acquisition (MCUs or FPGAs), sampling at 250–2,000 Hz (Pires et al., 2018, Sabry, 2023, Naim et al., 2020, Frey et al., 2023).
• Novel form factors: Textile, tattoo, and flexible PCB sensors targeting high wearability at the expense of interface SNR (Zheng et al., 2021, Naim et al., 2020).

Signal conditioning includes:

• Pre-amplification (differential, $\geq$80 dB CMRR),
• Analog band-pass filtering (10–500 Hz or narrower),
• Notch filtering (50/60 Hz line interference),
• Digitization (typically ≥12 bits for sEMG dynamic range).

Front-end technologies range from low-noise CMOS op-amp amplifiers ($\sim$1 μV${}_{\text{rms}}$ input noise) (Sabry, 2023), to active dry-contact or active-shielded electrodes for motion artifact minimization (Naim et al., 2020). Wireless telemetry and onboard MCU signal processing have enabled real-time, portable acquisition with battery lifetimes ranging from hours (continuous US + sEMG acquisition) to multiple days (duty-cycled sEMG-triggered modalities) (Frey et al., 2023, Pires et al., 2018).

3. Preprocessing, Denoising, and Feature Extraction

The core signal processing pipeline involves the following major steps (Manjarres-Triana et al., 2023, Liu et al., 2024, Wang et al., 2024):

1. Preprocessing: DC removal, high- or band-pass filtering ($\sim$20–450 Hz), notch filtering for mains noise, artifact rejection.
2. Denoising: Traditional methods (high-pass, template subtraction) are limited against complex contaminants (e.g., ECG, motion artifacts). Modern neural denoisers—TrustEMG-Net, MSEMG—integrate U-Net and Transformer or Mamba SSM modules, surpassing classical methods by 20–30% in SNR improvement, RMSE, and preservation of clinical features (ARV, mean frequency) while remaining computationally efficient (Liu et al., 2024, Wang et al., 2024).
3. Feature Extraction: Time-domain (TD) features (MAV, RMS, ZC, SSC, WL), frequency-domain features (mean/median frequency from Welch’s or wavelet spectra), SPD covariance features (for Riemannian geometry-based decoding), and time-frequency decompositions (DWT). Hudgins’ TD feature set yields strong baseline discrimination for both biometric and gesture recognition tasks with minimal computational overhead (Pradhan et al., 2021, Suri et al., 2020, Gowda et al., 2023).

Feature selection is application-specific: covariance-based SPD features are optimal for high-dimensional, robust hand-gesture decoding (Gowda et al., 2023, Gowda et al., 2024); TD features suffice for low-latency wearable biometrics (Pradhan et al., 2021).

4. Machine Learning and Decoding Methodologies

Modern sEMG pattern recognition leverages:

• Riemannian geometry/Manifold learning: Multichannel covariance matrices are treated as points on the SPD manifold, with geodesic distances and Fréchet mean-based centroids. Minimum Distance to Mean (MDM), Riemannian SVM, and SPD-networks provide robust classification even under significant subject/session domain shift, offering SOTA accuracy ($>$0.85 for hand gestures, $>$0.8 for diverse populations) at low computational cost (Gowda et al., 2023, Gowda et al., 2024).
• Neural sequence models: Convolutional and recurrent networks—per-channel CNN+LSTM hybrids, TDS-ConvNets (ASR-inspired), Transformer-based architectures—support end-to-end decoding of continuous gestures (e.g., typing, speech, sign language) from raw or spectrogrammed sEMG, with Connectionist Temporal Classification (CTC) loss for alignment-free output (Crouch et al., 2021, Sivakumar et al., 2024). Personalized adaptation, data augmentation (channel rotation, temporal jitter), and domain adaptation layers further improve cross-session/user generalization (Sivakumar et al., 2024, Ketykó et al., 2019).
• Statistical and classical ML classifiers: LDA, SVM, k-NN, PCA, and feature-based multi-stage SVMs (e.g., dual-stage group/intragroup hand-grasp decoding), when combined with low-dimensional or TD feature sets, remain competitive for real-time, resource-constrained settings (Suri et al., 2020, Pradhan et al., 2021).

5. Applications: Biomedical, Rehabilitation, HMIs, and Speech

• Clinical/rehabilitation: sEMG quantifies neuromuscular status (e.g., fatigue, dystonia diagnostics, paralysis assessment), enables objective monitoring in static or dynamic postures via population-trend or weak monotonicity detectors (e.g., downward median frequency drift, WM index) robust to low SNR and subject variability (Guo et al., 2021, Jha et al., 2015, Sabry, 2023).
• Prosthetic and exoskeleton control: Pattern recognition pipelines transform sEMG into multi-DOF motor commands, with state-of-the-art denoising and domain adaptation facilitating robustness under electrode shift, day-to-day variability, and different user demographics (Liu et al., 2024, Ketykó et al., 2019, Naim et al., 2020).
• Human-computer interfaces: Wearable sEMG armbands and wristbands enable gesture, touch, or silent speech decoding for interaction with computers, AR/VR, and mobile devices. End-to-end architectures achieve >90% character-level accuracy in typing, scale to large datasets (e.g., emg2qwerty: 346 h, 5.26 M keystrokes, 108 users), with domain adaptation and personalization crucial for cross-user deployment (Sivakumar et al., 2024, Crouch et al., 2021).
• Speech restoration: Multisite orofacial sEMG (22+ electrodes) supports silent speech decoding, with SPD geometric frameworks enabling cross-individual phoneme and word recognition at unprecedented top-5 decoding accuracy (up to 0.82 in continuous passage) (Gowda et al., 2024).

6. Variability, Domain Adaptation, and Demographic Factors

Inter-session, cross-user, and demographic variability (age, gender, BMI, skin hydration/elasticity) modulate sEMG amplitude and spectral features. Covariance-geometric representations and simple linear adaptation layers can mitigate most domain shift, with classification accuracy losses largely confined to lower frequency bands (20–50 Hz) (Gowda et al., 2024, Ketykó et al., 2019). Quantitative analyses reveal that robust manifold representations and appropriate feature selection (avoidance of low-frequency features) reduce demographic bias and support fair decoding across population subgroups.

7. Future Directions and Challenges

• High-density and flexible sensor development: Textile/tattoo electrodes and power-optimized front ends advance state-of-the-art comfort and wearability, but must further improve SNR and spatial resolution to match gel-based arrays (Zheng et al., 2021, Naim et al., 2020).
• Generalization and fairness: Community-shared, demographically diverse datasets (e.g., 91 adults, (Gowda et al., 2024); 108 users, (Sivakumar et al., 2024)) catalyze research in fair and robust modeling; transfer and domain-adaptive strategies remain active areas.
• Multimodal fusion: Integration of sEMG with IMU, ultrasound, or additional biosignals unlocks deeper intent inference, as in ultra-low-power, EMG-triggered wearable US systems (Frey et al., 2023) or IMU-based sEMG synthesis for envelope estimation (Basak et al., 21 Nov 2025).
• Advanced neural denoising: Emerging neural architectures—TrustEMG-Net, MSEMG—offer generalizable, parameter-efficient denoising across artifact classes and SNR regimes, enabling high-fidelity sEMG for downstream analytics and control (Liu et al., 2024, Wang et al., 2024).
• Real-time implementation and embedded inference: Model compression, quantization, and streaming neural architectures are under active development to support real-time, low-power deployment on-edge in prosthetic, rehabilitation, and assistive devices (Sabry, 2023, Liu et al., 2024).

Surface electromyography thus remains a foundational, highly technical modality for neuromuscular sensing, continuing to bridge advances in physiology, hardware design, signal processing, and machine intelligence to enable next-generation human–machine interaction and biomedical applications.