One Electrode Grid, Two Species: A Motor-Unit-Level Biomarker of Plasmalemmal Electrical Failure

When a muscle is forcibly lengthened while contracting, dystrophic muscle loses most of its force within tens of contractions — and the surprising finding of the last decade is that this force loss is electrical, not mechanical. In mice lacking the dystrophin glycoprotein complex (DGC), the membrane scaffolding that anchors muscle fiber surface membranes to the surrounding matrix, lengthening contractions drive sustained depolarization that locks voltage-gated sodium channels into inactivation; the muscle's contractile machinery is intact but receives no electrical command. Cory Baumann and colleagues showed that in three dystrophic mouse lines, "≥98% of the variance in eccentric torque loss could be explained by the variance in M-wave RMS" — the root-mean-square amplitude of the nerve-evoked compound muscle action potential — whereas in wild-type mice the same measure explained almost nothing ($R^2 = 0.211$; $p = 0.155$) (Contraction-Induced Loss of Plasmalemmal Electrophysiological Function Is Dependent on the Dystrophin Glycoprotein Complex). The problem is that this result rests on two tools that do not travel: supramaximal nerve stimulation under anesthesia, and a lumped compound potential that cannot say whether the whole membrane fails at once or a subpopulation of motor units is silenced first.

I propose to replace that lumped readout with high-density surface electromyography (HD-sEMG) — grids of 64–128 skin electrodes spaced millimeters apart — decomposed by blind source separation into individual motor unit spike trains, and to run that single measurement stack in mice and humans within the three-part cross-species validation scheme of Cross-species neuroscience: closing the explanatory gap: validate the non-invasive signal against invasive ground truth within one species, apply the identical tool across species, and translate both into a shared analytical space. That framework was built for cortex and fMRI and has never been operationalized in the peripheral neuromuscular system; dystrophic eccentric failure is the ideal first use case, because the mechanism is established in mice, the human phenotype diverges notoriously (the mdx mouse lives a near-normal lifespan while human Duchenne patients lose ambulation), and no stimulation-free electrical biomarker of lengthening-induced failure exists in any species.

Core Claim

The claim has two parts. First, in dystrophic muscle, plasmalemmal failure under lengthening strain is not uniform across motor units: decomposed motor unit spike trains and their fiber conduction velocities (MFCV — the speed at which a motor unit's action potential travels along the muscle fiber, an immediate index of membrane excitability) will show a collapsing, heterogeneous conduction velocity distribution — concentrated in high-threshold, fast units — that the compound M-wave cannot resolve. This is motivated, not merely hypothesized: in healthy humans, conduction velocity changes during mild fatiguing activity are already known to be non-uniform across low-threshold motor units, driven by local extracellular potassium accumulation around the most active fibers (Adjustments Differ Among Low-Threshold Motor Units During Intermittent, Isometric Contractions). Second, a stimulation-free, decomposition-derived failure signature validated in mice will predict eccentric force loss in awake human dystrophic muscle, something nerve-evoked M-waves cannot do because supramaximal nerve stimulation is neither tolerable nor ethical in fragile dystrophic patients.

A key methodological constraint shapes the design: conduction velocity, unlike firing rate, is a peripheral readout that is not contaminated by changes in central drive. Experimental muscle pain reduces motor unit firing rates through central inhibition while leaving conduction velocity completely unaltered (Effect of Experimental Muscle Pain on Motor Unit Firing Rate and Conduction Velocity), and higher discharge rates accelerate, rather than confuse, the membrane-level decline we want to measure (Low-threshold motor unit membrane properties vary with contraction intensity during sustained activation with surface EMG visual feedback). This makes the decomposed conduction velocity distribution the right observable when comparing anesthetized, stimulated mice with awake, voluntary humans.

Research Program

The program follows the three pillars of the cross-species framework directly. Tier one is within-species validation in mice. We build miniaturized HD-sEMG arrays scaled to the mouse tibialis anterior and record simultaneously with intramuscular fine-wire electrodes and nerve-stimulated M-waves during the exact 50-contraction in vivo eccentric protocol of Baumann et al., in wild-type, mdx, Sgcb$^{-/-}$, and mL172H animals (Contraction-Induced Loss of Plasmalemmal Electrophysiological Function Is Dependent on the Dystrophin Glycoprotein Complex). The fine-wire spikes serve as ground truth for validating decomposition accuracy, replicating the spike-triggered averaging logic used in humans by Farina, Arendt-Nielsen, and Graven-Nielsen to extract single motor unit surface footprints and track conduction velocity (Effect of temperature on spike-triggered average torque and electrophysiological properties of low-threshold motor units). That temperature study is also our calibration reference: it shows conduction velocity moves coherently with twitch mechanics (conduction velocity up ~29% from 33°C to 45°C, peak twitch torque up ~94%, correlation with twitch amplitude $r^2 = 0.48$), which is the physiological coupling our failure index should run in reverse.

Tier two is cross-species tool application: the same array geometry (scaled in electrode spacing to muscle length, kept identical in channel count and montage logic), the same convolution-kernel-compensation decomposition pipeline, and the same eccentric protocol logic — but voluntary submaximal lengthening contractions in human participants: healthy controls first, then ambulatory Becker muscular dystrophy and limb-girdle muscular dystrophy type 2E patients, whose DGC defects mirror the mouse lines. Becker patients are the right first population precisely because the mouse data show a dose effect: mL172H mice with ~40% dystrophin lose less torque and less M-wave RMS than mdx mice, and $\beta$-sarcoglycan deficiency is as damaging as complete dystrophin deficiency (differences between mdx and Sgcb$^{-/-}$ at $p \ge 0.337$). We therefore expect graded, genotype-ordered failure trajectories in humans, which is a falsifiable ordering the common-space analysis can test.

Tier three is the common analytical space. Because raw conduction velocities, forces, and array geometries do not compare across a 100-fold body-mass difference — the volume conductor of a mouse leg differs from a human limb by orders of magnitude in field decay and capacitive filtering — we translate everything into dimensionless failure coordinates:

$$f(n) = \frac{\bar{v}(n) - \bar{v}(0)}{\bar{v}(0)}, \qquad \tau(n) = \frac{T(n)}{T(0)}$$

where $\bar{v}(n)$ is the median single-unit conduction velocity across all decomposed motor units after the $n$-th eccentric contraction, and $T(n)$ is the simultaneously measured torque. Every animal and human session becomes a trajectory in the $(f, \tau)$ plane, compared with representational-similarity-style analysis across genotype and species. This is the direct peripheral analog of the third pillar in Cross-species neuroscience: closing the explanatory gap, which we re-use as method rather than re-invent.

A fourth, payoff experiment closes a standing mechanistic gap. Kristen Baltgalvis and colleagues rescued eccentric force loss in mdx muscle by overexpressing $\gamma$-cytoplasmic actin ~200-fold, producing "72% to 263% higher" force across a 10-contraction ex vivo series, but they never measured the membrane electrically (Transgenic overexpression of γ-cytoplasmic actin protects against eccentric contraction-induced force loss in mdx mice). Given Baumann's finding that eccentric force loss is ≥98% electrical in this genotype, the rescue must operate on the membrane or be redundant with it. Running our tier-one pipeline on mdx-TG animals distinguishes the two: preserved conduction velocity trajectories under load mean the actin scaffold stabilizes the membrane electrically; collapsing trajectories with preserved torque would mean Baumann's regression does not generalize to this intervention.

Proposed Experiments

The mouse experiment (E1) records four genotypes × ~10 animals over the 50-contraction eccentric series, with three concurrent streams: HD-sEMG grid, intramuscular fine-wire pairs for decomposition ground truth, and the standard nerve-evoked M-wave. Primary metrics are pulse-to-pulse decomposition agreement against the fine-wire triggers, the full conduction velocity distribution per contraction (not just its mean), the fraction of decomposable motor units that become electrically silent mid-series, and the regression of torque loss on these features. The human experiment (E2) records the same grid on the tibialis anterior during submaximal voluntary lengthening contractions on an isokinetic dynamometer at intensities chosen to avoid injury in patients, tracking voluntary motor unit trajectories. The common-space analysis (E3) tests whether $(f, \tau)$ trajectories cluster by DGC status across species and whether a failure index calibrated in mice predicts human force loss with sign and ordering preserved. Secondary analyses handle the two known confounds head-on: anesthesia suppresses spinal drive in the mouse tier, which is why the primary observable is conduction velocity rather than firing rate; and dystrophic fibrosis and fat infiltration alter the volume conductor, which is why tier-one validation exists at all. As a pipeline sanity check on inter-session robustness, we benchmark the decomposition stability of our grid montage on the CapgMyo database, where published inter-session gesture accuracies fall to 35.1% on DB-b and DB-c even with deep domain adaptation (Surface EMG-Based Inter-Session Gesture Recognition Enhanced by Deep Domain Adaptation) — a sobering reference point for how much session-to-session drift we must engineer around.

Baselines

The naive reference point is torque loss itself, regressed on raw grid-level HD-sEMG amplitude (root-mean-square over all channels). No reported result found on Lacuna for grid-RMS amplitude predicting eccentric torque loss in dystrophic muscle, so this baseline will be established de novo on our data; it tells us whether any decomposition machinery is needed at all beyond "the surface signal got smaller." The strongest existing method is the Baumann-style M-wave RMS regression: nerve-evoked compound potential root-mean-square, which in dystrophic mouse lines explained ≥98% of eccentric torque-loss variance ($p < 0.001$), with M-wave RMS deficits exceeding 50% in all dystrophic lines ($p \le 0.007$) and only an 8% non-significant drop in wild-type ($p = 0.080$) (Contraction-Induced Loss of Plasmalemmal Electrophysiological Function Is Dependent on the Dystrophin Glycoprotein Complex). It is the strongest competitor because it is already at the ceiling for torque prediction in stimulated mice, and honesty requires saying so: in E1 our decomposition features need only match its predictive $R^2$ there. The comparison is run fairly by recording both signals on the same animals in the same session, fitting both through the identical regression model family with the same number of free parameters and the same nested cross-validation, and giving both the same tuning budget. Beating this baseline is defined on three counts that M-waves cannot reach: prediction under voluntary activation without nerve stimulation, resolution of non-uniform failure across the motor unit population (a bimodal or threshold-ordered conduction velocity collapse invisible in the compound potential), and transferability of the calibrated failure index to human E2 data. A third baseline isolates the proposal's key ingredient: the same grid, same session, same regression, but with features taken from global amplitude and spectral statistics rather than decomposed motor unit trains. If decomposition adds nothing over this ablation, the blind source separation machinery is not earning its place, and the finding reduces to "surface signal amplitude tracks force loss," which the field already suspected.

Why not just…?

Why not just keep using M-wave RMS? In anesthetized, nerve-stimulated mice it is close to unbeatable at predicting torque loss, and I will not claim otherwise. It falls short on exactly the two things this proposal needs: it requires supramaximal nerve stimulation, which cannot be delivered to awake dystrophic patients, and it is a spatially and electrically lumped sum that cannot say whether failure is uniform or concentrated in a motor unit subpopulation — the mechanistic question the fatigue literature already answers non-uniformly for submaximal isometric work (Adjustments Differ Among Low-Threshold Motor Units During Intermittent, Isometric Contractions). The E1 design includes M-wave RMS as the strongest baseline on identical data, so the head-to-head is built in rather than argued around.

Why not just run HD-sEMG decomposition on Becker or Duchenne patients directly and skip the mice? This is the most serious alternative, and it would produce some result quickly. But without within-species ground truth, that result is uninterpretable: dystrophic muscle replaces fibers with fat and connective tissue, and subcutaneous adipose is precisely the spatial low-pass filter that attenuates the high spatial frequencies blind source separation depends on; deep compartment activity is known to be misattributed or lost at the surface even in healthy muscle (The effect of wrist posture on extrinsic finger muscle activity during single joint movements; Recruitment of motor units in two fascicles of the semispinalis cervicis muscle). A changed decomposition signal in patients could reflect membrane failure or simply a rewritten volume conductor, and nothing in a human-only study could separate them. The mouse tier, with fine-wire ground truth in the same animals, is what makes the human signal legible — that is the entire first pillar of the framework we are importing (Cross-species neuroscience: closing the explanatory gap).

Why not just train a deep network to classify dystrophic from healthy sEMG? The gesture-recognition literature shows this is technically feasible — an attention-based CNN-RNN reaches 87.0% on NinaPro DB1 and 94.5% on CSL-HDEMG (A novel attention-based hybrid CNN-RNN architecture for sEMG-based gesture recognition) — but the same literature shows inter-session collapse (35.1% on CapgMyo DB-b/DB-c even with domain adaptation), and the transfer-learning survey notes that deep models on wearables suffer catastrophic forgetting and heavy compute (Transfer Learning on Electromyography (EMG) Tasks: Approaches and Beyond). More fundamentally, a black-box classifier answers "diseased or not," which we already know clinically; it does not answer "is this lengthening contraction electrically failing the membrane," which is the question with a mechanistic answer and a therapeutic use. I will still run a learned classifier as a secondary upper-bound comparator on the same data, and if it predicts force loss better than the mechanistic failure index, that is a finding worth reporting, not hiding.

Why not just do ex vivo EDL muscles with added electrophysiology? The Baltgalvis protocol rescued force in a bath but removed the vascular supply, neural drive, and natural volume conductor, and ex vivo conditions plausibly exaggerate membrane stress in exactly the genotype under study (Transgenic overexpression of γ-cytoplasmic actin protects against eccentric contraction-induced force loss in mdx mice). More decisively, an isolated bath cannot validate a non-invasive sensor, which is the deliverable. Our E4 reuses their transgenic animals, not their preparation, to answer the question they left open.

Decisive Checks

Three outcomes would decide the proposal's fate. First, if decomposed motor unit conduction velocities in dystrophic mice fall uniformly rather than in a threshold-ordered or bimodal pattern, the first core claim is wrong, and the paper becomes a careful negative result that still validates the cross-species measurement stack. Second, if the mouse-calibrated failure index fails to sign or order correctly in Becker and LGMD2E patients in E3, the common-space mapping — not the biology — is the suspect, and the $(f, \tau)$ coordinate choice must be revisited before any DGC interpretation is drawn. Third, if in E4 the $\gamma$-actin rescue preserves torque while conduction velocities collapse on the same trajectories as unrepaired mdx muscle, then Baumann's ≥98% electrical regression is genotype- or condition-specific, and the field's single most-cited mechanism for eccentric weakness needs re-examination. Any one of these is a publishable, honest result; all three together would give neuromuscular medicine what it currently lacks — a stimulation-free, membrane-specific, cross-species-validated early warning that a lengthening contraction is failing the muscle electrically.