High-Density Intramuscular MEAs

• High-density intramuscular microelectrode arrays are implantable devices that record spatially resolved muscle action potentials with high precision using dense, flexible microfabricated contacts.
• They leverage advanced signal processing and blind source separation techniques to demix multi-channel EMG signals, achieving extraction of 51–196 motor units per task.
• These arrays facilitate robust neural decoding for fine motor control, offering improved performance for prosthetic applications and fundamental motor research.

High-density intramuscular microelectrode arrays (HD-iEMG MEAs) are implantable devices designed for the spatially resolved recording of action potentials and motor unit (MU) activity within skeletal muscle. These arrays enable precise, real-time access to neural drive signals from spinal motoneuron pools, thus providing a direct interface to the inputs governing fine motor control tasks such as finger flexion and extension. Leveraging advanced microfabrication and sophisticated signal processing, HD-iEMG MEAs have become a cornerstone technology in invasive neural interfacing, neural prosthetics, and high-fidelity decomposition of muscle activation patterns for both clinical and research applications (Baracat et al., 4 Sep 2025, Grison et al., 2024).

1. Array Design, Fabrication, and Surgical Deployment

HD-iEMG MEAs typically comprise multiple linear arrays, each with 40 platinum or platinum-coated microcontacts. The contacts measure 140 μm × 40 μm and are distributed along a 2 cm flexible polyimide shank, yielding an inter-electrode spacing (IED) of 0.5 mm (Grison et al., 2024). These arrays are fabricated via photolithography and metal lift-off, followed by laser cutting to define the shank geometry. Acute percutaneous deployment is achieved by back-loading the array into a 25-gauge hypodermic needle and advancing it to the target muscle using real-time ultrasound and MRI-based anatomical landmarks.

Implantation targets often include the flexor digitorum superficialis and various extensor compartments of the forearm to capture motor output to individual fingers. Electrodes are positioned within the muscle belly at a depth of several millimeters beneath the fascia, with percutaneous lead wires stabilized by adhesive dressings. Placement fidelity is confirmed by maximum voluntary contraction (MVC) testing (Baracat et al., 4 Sep 2025, Grison et al., 2024).

2. Signal-Acquisition Electronics and Preprocessing Pipeline

The arrays interface with multichannel amplifiers (e.g., OT-Bioelettronica Quattrocento) providing 16-bit resolution and 10,240 Hz sampling with hardware and digital band-pass filtering (20–2,000 Hz for HD-iEMG; 100–4,400 Hz in some studies). Recording is performed in a bipolar configuration with a common wrist reference (Grison et al., 2024). The resulting signals are characterized by high spatial density, ensuring detection of distinct MU action potential (MUAP) signatures on multiple channels and facilitating high-yield MU decomposition.

Platinum contacts on polyimide have impedances typically within 50 kΩ–1 MΩ; practical systems achieve input-referred noise below 5 μV RMS. Although not always explicitly measured, frequency-dependent impedance often follows $Z(f) \approx Z_0 f^{-\alpha}$, with $\alpha \approx 0.5$–$1.0$ and $Z_0$ in the $10^5$–$10^6\,\Omega$ range at 1 Hz (Baracat et al., 4 Sep 2025, Grison et al., 2024).

3. Motor Unit Extraction via Blind Source Separation

Motor unit decomposition from HD-iEMG relies on multichannel blind-source-separation (BSS) algorithms. A convolutional mixing model is adopted:

$$\widetilde{\mathbf{x}}(k) = \widetilde{\mathbf{H}} \widetilde{\mathbf{s}}(k) + \widetilde{\mathbf{n}}(k),\quad k=1,\dots,D_R$$

where $\widetilde{\mathbf{x}}(k)$ (observations) are high-dimensional HD-iEMG samples, $\widetilde{\mathbf{s}}(k)$ are underlying MU spike trains, and $\widetilde{\mathbf{H}}$ encodes the spatial–temporal MUAP templates. Swarm-Contrastive Decomposition (SCD) or silhouette-guided BSS methods optimize a demixing matrix $\mathbf{W}$ such that $$\mathbf{y}(k) = \mathbf{W} \widetilde{\mathbf{x}}(k)$$ yields sparse, temporally precise spike estimates (Grison et al., 2024, Baracat et al., 4 Sep 2025). Post-BSS, K-means clustering, silhouette scoring ($>0.85$ in (Baracat et al., 4 Sep 2025), $>0.9$ in (Grison et al., 2024)), and filtering on firing-rate (4–50 Hz) select physiologically plausible motor units.

Yield varies with task and anatomical region: typical extractions are 51–196 MUs per task (spread across flexion/extension, subjects), with average firing rates of 11.0 ± 3.9 Hz (Baracat et al., 4 Sep 2025, Grison et al., 2024). Spatial localization is accomplished by mapping rectified, normalized STA-derived MUAP amplitudes across channels, revealing muscle subregion specificity for each MU.

4. Computational Decoding and Integration with Neural Networks

Decoded MU spike trains and raw HD-iEMG signals can serve as input to neural decoders for classification or proportional control. Spiking neural networks (SNNs) deployed for this purpose use leaky-integrate-and-fire (LIF) and leaky integrator (LI) neuron models with discrete time updates:

$$I_{\mathrm{syn},j}[t+1] = \alpha I_{\mathrm{syn},j}[t] + \sum_i W_{ji} x_i[t+1]$$

$U_j[t+1] = \beta U_j[t] + I_{\mathrm{syn},j}[t+1]$

with $\alpha = \exp(-\Delta t/\tau_{\mathrm{syn}})$, $\beta = \exp(-\Delta t/\tau_m)$, and spiking if $U_j$ exceeds threshold.

For finger-force decoding, two SNN input modes are contrasted: (1) direct MU spike trains (one neuron per MU) and (2) spike-encoded HD-iEMG (LIF encoder per channel, with per-electrode threshold). Architectures are feedforward (inputs to outputs, one per finger) with LI or LIF readouts. Surrogate-gradient descent (MSE loss, Adam optimizer) trains both weights and output time constants; inference is performed at 10 ms resolution with 10 s training windows, 50% overlap (Baracat et al., 4 Sep 2025).

MU discharge timings may alternatively be classified for gesture identification, leveraging innervation pulse train features and C-support vector machine (C-SVM) classifiers (Grison et al., 2024).

5. Decoding Performance, Robustness, and Comparative Metrics

Decoding tasks include isometric contraction at 15% MVC for single- and multi-finger force output and gesture classification among up to 20 pre-defined movement classes. Performance metrics encompass RMSE, $R^2$, memory and latency, and classification accuracy:

Decoding Task	Input mode	RMSE (\% MVC) S1	RMSE (\% MVC) S2	Accuracy (classes)	Latency	Memory
SNN (LI) MU spikes	MU spike trains	1.73 ± 0.23	2.31 ± 0.34	—	~10 ms	2.5 kB + BSS
SNN (LIF) MU spikes	MU spike trains	1.78 ± 0.27	2.12 ± 0.41	—	~10 ms	2.6 kB + BSS
SNN (LI) spike-EMG	Spike-EMG	2.55 ± 1.13	3.41 ± 1.23	—	~10 ms	4.8 kB
MU timing classifier	MU spike timings	—	—	100% (12), 96.1% (16)	—	—

Loss of up to 50% of MU spike input at inference increases SNN RMSE from $\sim$1.7 to 3.2% MVC, but LIF readouts display the highest robustness. Classification accuracies for 16 hand tasks reach 100% (S1) and 96.1% (S2) using MU timings, outperforming r.m.s. HD-iEMG and HD-sEMG by 0.3–28.6% depending on the task (Grison et al., 2024, Baracat et al., 4 Sep 2025).

6. Implications and Comparison to Conventional EMG Interfaces

HD-iEMG systems enable direct mapping of spinal MU recruitment with temporal and spatial specificity unachievable by global or surface EMG. Classification, regression, and proportional decoding tasks benefit from the array’s dense sampling: MU yield, spatial precision, and demixing quality determine downstream decoder fidelity. Lower-density or surface EMG alternatives exhibit substantial drops in decoding accuracy, with bipolar iEMG (2 cm IED) suffering 40% accuracy loss on 16-class tasks (Grison et al., 2024).

These advances permit physiologically grounded, interpretable neural interfaces suitable for assistive prosthetics and research in fundamental motor control. Chronic implantation potential and biocompatibility remain under study, with acute (hours-to-days) recordings demonstrating robust performance and task reproducibility. A plausible implication is the extension of this approach to clinical neuroprosthesis and long-term human–computer interfacing, contingent on further development of stable chronic interface technology (Grison et al., 2024).

7. Future Directions and Methodological Considerations

Key limitations include the acute implantation window, absence of chronic stability data, and lack of absolute impedance or crosstalk quantification. Standardization of spike-detection quantification and task-specific validation (e.g., robust force decoding in dynamic contractions, out-of-laboratory contexts) are open areas. Improvement of blind-source-separation methods, further reduction in memory/latency, and investigation of robust decoding with minimal preprocessing are active research frontiers. Additionally, fine mapping of recruitment overlap, synergy encoding, and afferent/efferent modulation may illuminate broader motor control principles beyond current hand gesture classification models (Baracat et al., 4 Sep 2025, Grison et al., 2024).