When the nervous system contracts a muscle, it sends roughly the same control signal to many motor neurons at once. This shared signal, the "common neural drive," is what researchers estimate from electromyography (EMG, electrical recordings of muscle activity). The standard estimate is the cumulative spike train: identify each motor unit's firing times, then add them up. Today every such pipeline assumes that all electrode channels arrive at one processor. That assumption breaks for the recording systems people now want to build, such as wireless sleeves with many electrode patches spread over several muscles, each patch with its own battery and radio. Massive-antenna base stations hit the same wall years ago, and wireless engineers solved it with decentralized baseband processing: each cluster of antennas does local math and sends only small summaries to a central node.
Our proposal starts from one observation. A base station must recover every user's symbols. A neural-drive decoder only needs one quantity, the sum over motor units. Because that target is a single linear function of the unknown sources, the best linear estimator of it splits exactly across electrode clusters. Each cluster multiplies its own samples by a fixed weight vector and sends one number per time step, and the central node adds those numbers. The weights are computed centrally from small per-cluster summary matrices that are refreshed only when the muscle-to-electrode geometry changes. We propose to build this estimator, find where it stops matching a centralized decoder, and test whether it keeps the drive's higher-frequency content (the 15–30 Hz beta band). Cheap amplitude-envelope summaries, the obvious low-bandwidth alternative, should lose exactly that content.
Core Claim
Write the multichannel EMG as $y(t) = H s(t) + n(t)$. Here $y(t)$ stacks all $M$ electrode channels, together with a few time-lagged copies of each channel so that the temporal shape of each motor unit's action-potential waveform fits a linear model. The vector $s(t)$ holds the $K$ motor units' spike indicators, $H$ holds their spatial and temporal waveform templates, and $n(t)$ is noise. The neural drive is $d(t) = \mathbf{1}^T s(t)$. For uncorrelated unit-variance sources and white noise of variance $\sigma^2$, the linear minimum mean-square-error estimate of $d(t)$ is
$$ \hat d(t) = \mathbf{1}^T (H^T H + \sigma^2 I)^{-1} H^T y(t) = \sum_{c=1}^{C} w_c^T y_c(t), \qquad H^T H = \sum_{c=1}^{C} H_c^T H_c , $$
where $y_c$ and $H_c$ are the rows belonging to electrode cluster $c$, and $w_c = H_c (H^T H + \sigma^2 I)^{-1}\mathbf{1}$.
This is the "partially decentralized" structure from wireless equalization. Decentralized Baseband Processing for Massive MU-MIMO Systems and Decentralized Equalization With Feedforward Architectures for Massive MU-MIMO show that summing small local Gram matrices, the per-cluster products $H_c^T H_c$, reproduces the centralized linear detector exactly. Our contribution in that setting is to collapse the per-sample message from $K$ values to one scalar per cluster, which is possible because the target is a single sum rather than all $K$ unit signals.
The claim becomes testable where the idealized model fails. Motor units too small to decompose act as spatially correlated interference that spreads across clusters through the tissue. The waveform templates also drift with electrode shift and fatigue. The wireless analyses predict that such inter-cluster correlation is exactly what degrades decentralized fusion (Decentralized MIMO Systems With Imperfect CSI Using LMMSE Receivers), and we expect the same here.
Research Program
The first component is the estimator and its update protocol. During a calibration window, a convolutive blind source separation run gives the motor-unit waveform templates. Blind source separation here means recovering the individual units' firing from mixed channels without knowing the mixing in advance. Each cluster then computes its local Gram matrix and sends it once to the central node. The central node solves the $K\times K$ system and returns each cluster's $w_c$.
In steady state, each cluster streams one scalar per sample, or a few scalars if we also want residual-energy statistics that flag template drift. Design Trade-offs for Decentralized Baseband Processing in Massive MU-MIMO Systems found that sending Gram matrices beats sending local estimates once the channel stays stable for more than about 16 symbols. The EMG analogue of that stable period is how long the templates remain valid, and we will measure it directly. We will also test the same paper's reduced-precision result (fp8 cut data transfer 4x with negligible error loss) on the Gram and scalar messages.
The second component handles the model's known misfits. It adds a noise-covariance correction built from each cluster's residual. It also compares two ways of defining clusters: grouping electrodes anatomically, following each muscle's innervation zones, versus splitting the grid into equal blocks. Wireless work reports that equal-size antenna clusters give the worst signal quality and that adding clusters always hurts (Decentralized Equalization With Feedforward Architectures for Massive MU-MIMO, Decentralized MIMO Systems With Imperfect CSI Using LMMSE Receivers). We will test whether this holds when motor units are spatially localized. For a single common drive, clusters that align with anatomy might approach the "uncorrelated clusters" regime, in which that paper proves suboptimal fusion reaches optimal quality.
The third component is an optional central nonlinear stage. Spikes are non-negative and sparse, so a linear estimate leaves accuracy on the table. A small network that sees only the $C$ incoming scalar streams and their recent history can correct some of this without increasing what the clusters transmit.
Baselines
The naive reference is the per-cluster amplitude envelope. Each cluster sends a rectified, low-pass-filtered root-mean-square value per channel group, and a linear regression fitted on calibration data fuses them into a drive estimate. This is what a practitioner would build first. It uses the same or lower bandwidth than our method, and it will be fitted on the same calibration windows.
The strongest accuracy competitor is the centralized deep convolutional network trained to output spike counts directly from all channels. On gastrocnemius HD-EMG, A deep CNN framework for neural drive estimation from HD-EMG across contraction intensities and joint angles reports correlation "consistently higher than 0.96" and nRMSE "< 7%" against the blind-source-separation reference. On vastus medialis and lateralis, Toward a generalizable deep CNN for neural drive estimation across muscles and participants reports leave-one-participant-out correlation of 0.985±0.005. We will retrain this architecture, with four convolutional and two fully connected layers on 40-sample windows, on our exact training splits and give it the same hyperparameter-search budget as our central fusion stage. Its advantage is full access to raw channels at a raw-channel bandwidth cost. We count our method as beating it if it matches its correlation within its reported spread while transmitting at least an order of magnitude less per sample, and if it preserves beta-band coherence at least as well.
The third baseline is a naive fully decentralized design: a local network in each cluster estimates spike counts, and the central node adds them. It isolates double counting of motor units that appear in several clusters. The fourth is iterative consensus, with ADMM or conjugate gradient across clusters, which Decentralized Baseband Processing for Massive MU-MIMO Systems shows reaches centralized performance in 2 to 3 iterations. It serves as the reference for what exchanging messages repeatedly per sample buys.
The key ablation replaces the Gram-derived weights $w_c$ with weights computed from local data only. This isolates the contribution of the once-per-calibration global exchange. For decentralized EMG decoding there is no reported result found on Lacuna, so every decentralized comparison is new.
Why not just…?
Why not just send amplitude envelopes, since bandwidth is the concern? Envelopes are a real threat for slow force tracking, and we expect them to be competitive below a few hertz. The question is the drive's content above that range. Factors Influencing the Estimates of Correlation between Motor Unit Activities in Humans shows that pooling spike trains is what exposes beta-band common input: a five-unit cumulative spike train raised beta coherence magnitude by about 163% over single pairs. Envelopes of the interference pattern carry no spike timing. Changes in high-frequency neural inputs to muscles during movement cancellation uses exactly this beta and gamma content to detect cognitive states. Our experiments report error separately by frequency band. If envelopes match us in the beta band too, the decentralized estimator is not worth its calibration cost, and we will say so.
Why not just run the centralized network, since EMG data rates are small? For one wired grid this objection is correct, and we do not claim otherwise. The case for our method is multi-patch wireless systems, where each patch's radio power scales with the samples it transmits, and dense intramuscular arrays of the kind High performance sorting of motor unit action potentials with EMUsort targets. We will report a bandwidth–accuracy curve rather than a single point, so a reader can see at what channel count and radio budget the centralized network stops being viable.
Why not just add local spike counts? A single motor unit's action potential spreads across neighboring clusters, so local detectors count it more than once, and the error depends on where each unit sits. The naive decentralized baseline measures this directly. The analysis of spatially correlated clusters in Decentralized MIMO Systems With Imperfect CSI Using LMMSE Receivers predicts that naive fusion loses the most when clusters overlap.
Proposed Experiments
We will begin in simulation, where the true drive is known. We will use a leaky integrate-and-fire motoneuron pool of the kind used by Estimation of the firing behaviour of a complete motoneuron pool by combining electromyography signal decomposition and realistic motoneuron modelling, in which each simulated neuron sums its input until it crosses a threshold and fires. We will pair it with the simulated recordings from EMUsort's MUsim, driving the pool with a known common input that contains low-frequency and beta components plus independent noise. Simulation lets us vary three things the real data cannot: the number of undecomposed units, how far each unit's waveform spreads, and electrode shift.
We will then move to HD-EMG recordings of the types described in the file: isometric gastrocnemius contractions at several intensities and ankle angles, and vastus medialis and lateralis sessions about 20 months apart. The reference drive is the cumulative spike train from blind source separation. We will split grids into two to eight clusters, both uniformly and anatomically.
The metrics are correlation and nRMSE against the reference drive, error within bands (0–5 Hz, beta, gamma), and the coherence between the estimated drive and the reference cumulative spike train. We will also measure latency and the bits sent per cluster per second. Force prediction through a musculoskeletal model tests downstream use: the raw-sample reference in the virtual-pool paper is $r^2 \approx 0.85$ with 20% nRMSE, and the complete virtual pool reaches $r^2 \approx 0.97$ with 7.5% nRMSE.
The decisive checks are four. First, in simulation with white noise and no drift, the one-scalar decentralized estimator must match the centralized linear estimator to numerical precision; this confirms the exact decomposition. Second, as the number of undecomposed units grows, we will measure how fast the decentralized estimator falls behind centralized estimators that can model correlated interference. We will also test whether anatomical clustering slows that loss, as the uncorrelated-cluster result predicts. Third, on real data, we will test whether our method beats envelope fusion in the beta band at equal bandwidth. Fourth, we will measure how long waveform templates remain valid before recalibration is needed; this sets the true per-sample communication cost.
If the first check fails, the formulation is wrong. If the third fails, the practical case collapses to envelope fusion, and we would report that outcome.
References
- Decentralized Baseband Processing for Massive MU-MIMO Systems
- Decentralized Equalization with Feedforward Architectures for Massive MU-MIMO
- Decentralized MIMO Systems With Imperfect CSI Using LMMSE Receivers
- Design Trade-offs for Decentralized Baseband Processing in Massive MU-MIMO Systems
- A deep CNN framework for neural drive estimation from HD-EMG across contraction intensities and joint angles
- Toward a generalizable deep CNN for neural drive estimation across muscles and participants
- Factors Influencing the Estimates of Correlation between Motor Unit Activities in Humans
- Changes in high-frequency neural inputs to muscles during movement cancellation
- High performance sorting of motor unit action potentials with EMUsort
- Estimation of the firing behaviour of a complete motoneuron pool by combining electromyography signal decomposition and realistic motoneuron modelling