Spending Electrodes on Spikes, Not Waveforms: Task-Targeted Channel Selection for Motor Unit Decomposition

Electrode grids on the skin can now record a muscle at 256 points, and decoding these signals recovers the firing times of individual spinal motor neurons. Each neuron drives a bundle of muscle fibres, and together they form a "motor unit". Going from 64 to 256 electrodes roughly doubles how many motor units can be recovered, and very dense grids reveal small, early-recruited units that sparser grids miss (Larger and Denser). The cost is large. Each channel is stacked with many time-delayed copies of itself, so the decoder works on vectors with thousands of dimensions. Real-time decoders have to calibrate once and then hope nothing changes (I-Spin live). The natural response is to pick a smaller subset of electrodes using experimental design, the statistics of choosing measurements that most reduce uncertainty. The textbook choice, D-optimal design, picks the measurements that shrink the total volume of the uncertainty region over all model parameters.

This note argues that D-optimal design aimed at the wrong parameters is the wrong tool here. A better tool already exists in another field and has never been applied to neural signals. The physical model behind decomposition has an enormous number of waveform parameters, and most of them barely affect the recorded signal. In the language of "sloppy" models, the measurement is sensitive to a few combinations of parameters and nearly blind to the rest. In sloppy models, maximizing total information pushes measurements toward regimes where the model's simplifications break (The Limitations of Model-Based Experimental Design and Parameter Estimation in Sloppy Systems). The quantity we care about is not the waveform. It is whether each unit's spikes can be detected. We propose choosing electrodes so that every motor unit, including the smallest, keeps a guaranteed floor of information about its own spike times, using as few channels as possible. We then ask how much of that floor can be supplied for free by a realistic model of how motor neurons fire, as opposed to the memoryless assumptions decoders use today.

Core Claim

Channel selection for motor unit decomposition should be posed as information matching on spike detectability: find the sparsest set of electrodes for which each unit's information about its own spike times exceeds a target. Three things follow if this is right. First, at a fixed channel budget it should keep more units than heuristic sub-grids and than D-optimal criteria. The advantage should be largest among small, early-recruited units, because an aggregate criterion trades them away for large units. Second, it should be more robust than waveform-level D-optimal design when the muscle departs from the calibration conditions. That design spends channels on waveform detail, which is exactly what changes when force level or joint angle shifts. Third, a discharge model with refractoriness and spike-frequency adaptation can replace part of the electrode information. Spike-frequency adaptation is the progressive slowing of firing driven by slow potassium currents (Muller et al.). Replacing part of the electrode information this way lowers the number of channels needed to meet the same error rate.

The method comes from An information-matching approach to optimal experimental design and active learning. That work minimizes the L1 norm of experiment weights subject to a matrix inequality, so that information is supplied only along prescribed quantities of interest. It showed this on power grids, underwater acoustics and interatomic potentials. It did not treat convolutive mixtures, spike trains or electrode arrays, and its quantities of interest were static predictions rather than per-source detectability. The motivation comes from White et al., who showed that pursuing full-parameter identifiability in sloppy models can blow up the best-fit error from about 7,000 to over 96,000 in an EGFR signalling model. Their study diagnosed the problem and offered no constructive remedy for a signal-separation setting.

Research Program

The forward model is the standard convolutive mixture used by MUedit. Each channel records the sum of every unit's spike train filtered through that unit's action-potential waveform at that electrode, plus noise. A short pilot decomposition of the full grid gives spike-triggered estimates of each unit's waveform. Let $h_{j,r} \in \mathbb{R}^N$ collect the waveform values of all $N$ units at channel $j$ and lag $r$, and let $\sigma_j^2$ be that channel's noise variance. Treat the spike amplitudes of all units at a given instant as the parameters of interest, and give channel $j$ a selection weight $w_j \in [0,1]$. The Fisher information about those amplitudes is then linear in the weights:

$$J(w) = \sum_{j=1}^{M} \frac{w_j}{\sigma_j^2} \sum_{r=0}^{R-1} h_{j,r} h_{j,r}^{\top},$$

where $M$ is the number of electrodes and $R$ the number of lags. The marginal precision for unit $i$ after accounting for interference from all other units is $1/[J(w)^{-1}]_{ii}$. We select channels by solving

$$\min_{w} \|w\|_1 \quad \text{subject to} \quad [J(w)^{-1}]_{ii} \le \frac{1}{d_i^2} \ \text{for all units } i, \quad 0 \le w \le 1.$$

Here $d_i$ is the detection margin each unit needs, in units of noise standard deviation. Each per-unit constraint can be written as a linear matrix inequality, so the problem is a semidefinite program. This is the same convex structure information matching exploits. We then round the weights to a channel set and polish the set with a few swaps. Rounding is necessary because exact discrete D-optimal-type selection is NP-hard and does not admit constant-factor approximation (D-Optimal Data Fusion).

The contrast with D-optimal design is in the constraint. Maximizing $\log\det J(w)$ rewards total volume. It can be satisfied by making large superficial units extremely precise while leaving a small deep unit undetectable. The per-unit floor forbids that trade.

The second component sets $d_i$ from the discharge model rather than fixing it. A spike detector at time $t$ combines channel evidence with prior log-odds that a spike occurs, given the unit's history. Motor neurons fire fairly regularly, are refractory, and adapt. A history-dependent intensity model therefore concentrates prior mass sharply, and a smaller margin achieves the same false-positive and miss rates. The history model is a point-process generalized linear model with a post-spike filter capturing refractoriness and adaptation (Pillow et al.). We compare three detectors: no prior, which is today's thresholding or k-means step; a renewal prior that conditions only on time since the last spike; and the non-renewal history prior. Each choice yields its own required $d_i$ and hence its own minimal channel set. The non-renewal prior is motivated by evidence that spike-time dependence reaches across several past intervals (Bayesian nonparametric (non-)renewal processes). We also use it because renewal approximations produce artefacts during transients (Inference of a mesoscopic population model from population spike trains). We choose the simpler model with a parametric history filter, rather than the nonparametric process, so that it can run in a real-time loop. Goodness of fit of each discharge model is checked with time-rescaling tests (Time-rescaling methods).

The third component is an analysis, and its outcome could undercut the motivation. The research file's description of waveform parameters as sloppy is a hypothesis, not a measured fact. We will compute the full Fisher information over waveform taps from pilot data and report its eigenvalue spectrum. We expect the multi-decade decay seen in other mechanistic models (Universally Sloppy Parameter Sensitivities in Systems Biology Models). We will also check whether the sloppy directions are the ones that change between calibration and test conditions. If the spectrum is not sloppy, the White-type failure argument for this setting fails. The per-unit floor would then still be justified by the representativeness argument alone, but we would report that the sloppiness story does not transfer.

Baselines

The naive reference points are the full 256-channel decomposition and fixed spatial sub-grids at the same channel budget. The full grid is the ceiling we try to approach. The sub-grids are uniform every-other-electrode decimation and a contiguous block. For physical grids rather than subsets, Larger and Denser reports an average of 56 motor units per participant with 256 channels at 4 mm spacing, against roughly 30 with 64 channels, on tibialis anterior. On simulated surface EMG, MUedit reports a rate of agreement above 95% with ground truth. All selected subsets are decoded with the identical MUedit pipeline, with the same lag count, iteration limits and automated editing rules, so that only the channel set differs. Beating the naive baseline means recovering significantly more ground-truth units than decimation at budgets of 32, 64 and 96 channels, without lowering rate of agreement.

The strongest existing selection method is an adaptation of classification-based selection from Optimal Adaptive Electrode Selection to Maximize Simultaneously Recorded Neuron Yield. That method picks the subset that maximizes waveform discriminability measured by linear discriminant analysis on a pilot sort, and it gained 41–85% in neuron yield over checkerboard layouts on a Neuropixels probe. It is the strongest competitor because it is data-driven, uses the same pilot information we use, and has proven large gains. We will extend its objective to lagged multi-channel windows and give it the same pilot data and the same greedy-swap search budget. A second heuristic competitor is the power-correlation ratio and frequency-domain principal component selection of Automated Channel Selection in High-Density sEMG. It achieved about a 30% reduction in force-estimation error with 3 of 7 channels per muscle, a different task from decomposition. We found no reported result on Lacuna for any channel-subset selection method evaluated on motor unit decomposition yield, so every subset-selection number will come from our own runs.

Two ablations isolate the key ingredients. The first is D-optimal selection on the same spike-amplitude information, maximizing $\log\det J(w)$ at the same budget. It shares our task targeting but drops the per-unit floor, so any gap measures the value of the floor. The second is classical D-optimal selection on the full waveform-tap Fisher information. It drops task targeting entirely, so it is the direct test of the sloppy-model failure mode. For the discharge-prior component, the reference is the I-Spin live detector with no history model. That detector reports 93.6% agreement with offline decomposition on surface tibialis anterior and firing-rate error below 2 pulses per second. The renewal prior serves as the ablation for the non-renewal prior.

Why not just…?

Why not just use all 256 channels and buy faster hardware? This is a genuine option for offline work, and we treat the full grid as the ceiling, not a straw man. The case for selection is real time and wearable use. I-Spin live already has to freeze separation vectors after calibration, and channel count multiplied by lag count sets the size of the matrices it must handle. If a 64-channel targeted subset recovers nearly all of the full-grid units, re-calibration can happen online and grids can be cheaper. If it does not, the experiments will show that, and the honest conclusion will be that density cannot be replaced.

Why not just adapt classification-based selection, which already works on probes? This is the most serious threat, because a discriminability objective may capture most of what our information criterion captures. Two differences are testable. That method optimizes an aggregate discriminability score, which, like log-determinant, can trade small units away. Our floor is per unit, so the comparison will be stratified by unit amplitude and recruitment threshold. The prediction is that the gap appears among the early-recruited, small units that Larger and Denser found make up 29% of the sample on 2 mm grids. If the adapted method matches us in every stratum, the per-unit formulation adds nothing beyond convexity and speed, and we will say so.

Why not just maximize the log-determinant of the spike-amplitude information, since that is already task-targeted? That is our first ablation. If it matches information matching in yield and small-unit coverage, the specific contribution reduces to "target spikes, not waveforms", which is still a finding, but a smaller one. Such a result would also weaken the claim that per-unit guarantees matter.

Why not just add a refractory-period rule to delete impossible spikes after decoding? Post-hoc deletion removes false positives but cannot recover missed spikes, and it cannot lower the channel count needed. The renewal-prior arm approximates this simpler fix inside the detector. If the full adaptation filter does not beat it on held-out force ramps, the non-renewal part of the proposal is unnecessary.

Proposed Experiments

The first test runs on simulation with ground truth. We use the 200-unit digital muscle behind the MUedit benchmarks and replace its discharge generator with two versions: renewal discharges, and adapting discharges with serial interval correlations. At budgets of 32, 64 and 96 channels out of 256, we report units recovered, rate of agreement, and recovery of the smallest-amplitude quartile, for every baseline and ablation. The decisive check is this: at 64 channels, information matching must recover a clearly larger fraction of the units found by the full 256-channel decomposition than decimation does, and must beat the log-determinant ablation in the small-unit quartile. If neither holds, the per-unit floor is not doing useful work.

The second test probes robustness under mismatch, using NeuroMotion (NeuroMotion), which simulates waveform morphing during movement. Channels are selected and separation vectors calibrated on an isometric segment, then decoding is evaluated on changed force levels and dynamic segments. The sloppy-model hypothesis predicts that waveform-tap D-optimal subsets degrade most, because their information sits in the waveform details that morph. If the waveform-tap spectrum turns out not to be sloppy, or if this ordering does not appear, we retract the White-type explanation for this domain.

The third test uses in vivo tibialis anterior recordings on 256-electrode grids under the protocols of Larger and Denser and MUedit. We use released data where available and otherwise collect new recordings with simultaneous intramuscular wires. Subset decompositions are scored against the full-grid decomposition and the intramuscular reference, following agreement and editing conventions from the CEDE single motor unit matrix.

The final test isolates the discharge prior. For each detector (no prior, renewal, non-renewal), we find the smallest channel budget that reaches a fixed agreement level, on both stationary contractions and force ramps. The non-renewal claim stands only if it lowers that budget on ramps beyond what the renewal prior achieves.

References

  1. Larger and Denser: An Optimal Design for Surface Grids of EMG Electrodes to Identify Greater and More Representative Samples of Motor Units
  2. I-Spin live, an open-source software based on blind-source separation for real-time decoding of motor unit activity in humans
  3. The Limitations of Model-Based Experimental Design and Parameter Estimation in Sloppy Systems
  4. Spike-Frequency Adapting Neural Ensembles: Beyond Mean Adaptation and Renewal Theories
  5. An information-matching approach to optimal experimental design and active learning
  6. Tutorial on MUedit: An open-source software for identifying and analysing the discharge timing of motor units from electromyographic signals
  7. D-Optimal Data Fusion: Exact and Approximation Algorithms
  8. Maximum Likelihood Estimation of a Stochastic Integrate-and-Fire Neural Encoding Model
  9. Bayesian nonparametric (non-)renewal processes for analyzing neural spike train variability
  10. Inference of a Mesoscopic Population Model from Population Spike Trains
  11. Time-rescaling methods for the estimation and assessment of non-Poisson neural encoding models
  12. Universally Sloppy Parameter Sensitivities in Systems Biology Models
  13. Optimal Adaptive Electrode Selection to Maximize Simultaneously Recorded Neuron Yield
  14. Automated Channel Selection in High-Density sEMG for Improved Force Estimation
  15. NeuroMotion: Open-source Simulator with Neuromechanical and Deep Network Models to Generate Surface EMG signals during Voluntary Movement
  16. Consensus for experimental design in electromyography (CEDE) project: Single motor unit matrix