Housekeeping Loci for Motor Units: Shift-Stable Identity Profiles for Tracking Neurons Across Sessions

High-density surface electromyography records muscle electrical activity with grids of 64 to 256 skin electrodes. Decomposition algorithms can separate that signal into the firing times of individual motor neurons, which gives a noninvasive view of the spinal cord's output. The field has little longitudinal use of this ability for a practical reason: it cannot reliably recognize the same neuron from one week to the next. The standard method compares each unit's spatial waveform "fingerprint" across sessions and accepts a match when the 2D cross-correlation is at least 0.8. With that rule, only 38–40% of units survive one week and 16–21% survive three sessions (Martinez-Valdes et al.). The waveform shape changes whenever the grid shifts by a few millimetres, so the fingerprint breaks at exactly the moment it is needed.

Microbial epidemiology solved a structurally similar problem decades ago. Multilocus sequence typing (MLST) identifies bacterial lineages by a short list of "housekeeping" genes rather than by fast-mutating surface antigens. Mutations in those genes are purged by selection, so lineage identity stays stable while the surface features drift (Tracking Microbial Lineages Using Conserved Housekeeping Genes). The loci are also chosen to vary independently, which makes their combined discriminating power multiply. We propose to build the motor-unit equivalent. Each unit gets a short profile of physiological and biophysical features chosen by two explicit criteria. The first is stability under electrode shift and session change. The second is mutual non-redundancy, measured with variance inflation factors (VIF), which say how much of one feature is predictable from the others. We then test, on benchmarks with known ground truth and with subject-level held-out evaluation, whether such profiles track more units at a controlled false-match rate than waveform correlation does.

Core Claim

Motor-unit tracking fails for the same reason serotyping fails: it keys on the most variable observable. Our claim is that a small set of unit features can be identified empirically that behave like housekeeping loci. These features vary a lot between units but little within a unit across grid shifts and days. Candidates include conduction velocity, innervation-zone position relative to the grid after registration, estimated source depth, recruitment threshold, and firing rate at a fixed force level. A profile built from such features, pruned so that no feature is largely predictable from the others, should retain substantially more units than 2D cross-correlation at the same false-match rate.

The new element is not the analogy. It is the selection rule. Features are admitted as "loci" only if they pass a stability test and a redundancy test, and both tests are estimated on some subjects and applied to different subjects. We will also measure, rather than assume, whether discretizing features into MLST-style "alleles" helps. The honest prior is that it may not.

What Exists and What Does Not

The research file lists this intersection as open. Its closest description is a "Gap 2" sketch that proposes quantizing motor-unit features into a barcode but reports no implementation. Three bodies of work shape the design.

The first is Martinez-Valdes et al.. They track vastus medialis and lateralis units with semi-blind decomposition and a cross-correlation threshold, and they show that tracked units give more reliable measurements (ICCs 0.63–0.99 versus 0.39–0.95 for the full detected population). This work motivates ours by establishing both the value of tracking and the retention problem. It also supplies our main baseline. Their method uses one continuous, shift-sensitive feature, and we replace it with a selected multi-feature profile.

The second is BMISS, which fits a biophysical forward model (BioMime) to recover each unit's depth, angle, size, innervation zone, fiber length, and conduction velocity. On noiseless synthetic data it reports a parameter condition error of 0.0106. BMISS supplies both a method and a candidate set of loci, because anatomical parameters of a fixed muscle fiber bundle should not change when the grid moves. BMISS was never tested on real recordings or across sessions, and we treat its parameter stability as a hypothesis to test rather than a given.

The third is Cho's reanalysis of the Mice Protein Expression data, which constrains how we evaluate. It showed two things. First, splitting replicates from the same animal across training and test sets inflates performance. Second, collinear predictors produce suppression effects, where a truly informative variable (SOD1) looks unimportant because correlated partners absorb its credit. Both carry over directly. Stability and redundancy statistics estimated and tested on the same participants would be a form of single-subject leakage. Collinear features would inflate apparent profile distinctiveness without adding real discrimination.

The single-case literature is the reason the held-out-subject design matters rather than being a formality. The 22-month intracortical interface study of participant T15 (A Personalized and Adaptable User Interface) and the 7-participant wristband study both warn that designs tuned to individuals may not transfer. A tracking rule tuned on one person's grid and anatomy is exactly such a design.

Research Program

The program has three parts.

The first part assembles a feature pool for each decomposed unit. It includes the conventional spatial waveform, and the waveform's centroid and innervation-zone location after rigid registration of the grid between sessions. It adds conduction velocity estimated along fiber-aligned columns, recruitment and derecruitment thresholds as a percentage of maximal force, and mean discharge rate on the plateau of a standardized ramp. Where the forward-model fit converges, it also includes BMISS-style depth and fiber-length estimates.

The second part selects loci on training subjects only. For each feature $f$ we compute a stability ratio

$$S_f = \frac{\sigma^2_{\text{between units}}(f)}{\sigma^2_{\text{within unit, across shifts or sessions}}(f)},$$

which is the between-unit variance divided by the variance of the same unit across repeated observations. A high $S_f$ means the feature discriminates units and survives perturbation, which is the analogue of a housekeeping gene. We rank features by $S_f$. We then add them greedily, admitting a feature only if its VIF against the features already chosen stays below a cutoff that we treat as a tuned hyperparameter rather than the conventional 5 or 10. This mirrors the MLST design logic that loci should be independent, so that each one adds discrimination.

The third part matches units using the selected profile. We compare a continuous version (a shrinkage Mahalanobis distance followed by one-to-one assignment) with a discretized version, in which each feature is binned into "alleles" using bin widths set by its within-unit spread. The discretized version accepts a match when at least $k$ of $n$ loci agree, which is the analogue of grouping sequence types into clonal complexes. Discrimination within a session is reported with the Simpson-type diversity index used in typing,

$$D = 1 - \frac{\sum_j n_j (n_j - 1)}{N(N-1)},$$

where $N$ is the number of units and $n_j$ is the number of units sharing profile $j$. $D$ is the probability that two distinct units receive distinct profiles. For scale, the 7-locus S. parasuis MLST scheme reaches $D = 0.9821$ over 255 isolates (Developing genome typing strategies for S. parasuis). That is the kind of number a useful profile needs, although the context is different.

Ground Truth Without New Hardware

Tracking studies usually cannot say which matches are correct. We will build three levels of ground truth.

The first level is simulation. We generate BioMime-style synthetic recordings of the kind used to benchmark BMISS, with known unit identities, controlled grid translations of 0–10 mm, noise levels down to low signal-to-noise ratios, and small per-session changes in conduction velocity and recruitment.

The second level is a shifted sub-grid test on real data. From a 256-electrode recording, such as the tibialis anterior protocol of Caillet et al., we extract two overlapping 64-electrode sub-grids offset by one or two electrode rows. We decompose each independently and define true matches by discharge-time agreement between the two decompositions. This gives a real physiological recording with a known, exact spatial shift and known identities.

The third level is a target–decoy control, borrowed from proteomics practice. We match each subject's units against a decoy library of units from other subjects. Any decoy match is by definition false, so the procedure gives an empirical false-match rate on real longitudinal data where true identities are unknown.

On longitudinal data, meaning vastus recordings of the Martinez-Valdes type if access can be obtained, or a new small collection otherwise, the decoy false-match rate and the ICC of tracked units' properties become the primary outcomes.

Baselines

The naive reference point is no tracking. Longitudinal comparisons are then made on population averages of all decomposed units in each session, which is the practice tracking is meant to improve. Its known cost is the lower ICC range (0.39–0.95) reported by Martinez-Valdes et al.. In that paper, tracking reached a conduction-velocity effect size of 2.1 in the endurance training cohort. We will report the untracked effect size alongside it on our own data.

The main published competitor is the Martinez-Valdes method itself: semi-blind decomposition plus 2D waveform cross-correlation with an acceptance threshold of 0.8. It reports 38.3% (vastus medialis) and 40.1% (vastus lateralis) retention over one week, and 21.0% and 16.3% across three sessions. These are retention fractions without verified correctness, so they cannot be read as accuracy. We will rerun this method on our own data under identical conditions and sweep its threshold, so that both methods are compared at matched decoy false-match rates.

The strongest competitor we must build ourselves is the same cross-correlation, but with an explicit search over spatial shifts of the waveform map before correlation. This directly addresses the failure mode we attribute to the baseline. No reported result for shift-compensated cross-correlation tracking was found on Lacuna.

A second strong competitor is continuous Mahalanobis matching on the full feature pool, with covariance shrinkage and no stability or VIF selection.

All methods will use the same decompositions, the same registration step, and the same subjects. Hyperparameters (the thresholds, VIF cutoff, bin widths, and $k$) are tuned by nested leave-one-subject-out cross-validation with equal search budgets.

Two ablations isolate the key ingredients. One removes the stability criterion and keeps VIF pruning. The other removes VIF pruning and keeps stability ranking. A third comparison sets the discretized profile against its continuous counterpart.

Beating the baseline means higher recall of true matches at a decoy false-match rate of 5% or below on held-out subjects. On the shifted sub-grid test, the advantage must hold for shifts of one row or more. On longitudinal data, retention must be higher at matched false-match rate, with tracked-unit ICCs no worse.

Why not just…?

Why not just search over spatial shifts before cross-correlating waveforms? This is the most serious threat. If the only problem is rigid grid translation, aligning the waveform map may recover most lost matches, and our feature selection would add nothing. It is not the only problem, though. Martinez-Valdes et al. attribute losses also to changes in skin impedance and recruitment variability, and a waveform shape changes under rotation and changes in tissue conduction, not just under translation. The shifted sub-grid test settles this directly. If shift-compensated correlation matches our profile at every offset and noise level, the waveform is sufficient and we will report that the housekeeping-loci idea is unnecessary for pure translation. The profile's case then rests entirely on longitudinal data, where non-rigid changes dominate.

Why not learn an embedding end to end, for example with domain adaptation? Learned alignment is the default modern answer to session shift. The evidence available here is not encouraging for raw high-density signals: multi-stream adaptive batch normalization reached only 35.1% inter-session accuracy on CapgMyo DB-b (Surface EMG-Based Inter-Session Gesture Recognition). That result concerns gesture labels, not unit identity, so it constrains rather than refutes the approach. The bigger issue is data: small-cohort longitudinal studies provide a few hundred tracked units, far too few to train a metric without the single-subject overfitting that Cho documented. We will include a small learned metric (a shallow twin network over the same feature pool) trained under leave-one-subject-out as a check. If it wins on held-out subjects, it should replace our selection rule.

Why not skip VIF and let a regularized distance handle collinearity? A shrinkage Mahalanobis distance already down-weights correlated directions, so an explicit VIF step may be redundant for matching accuracy. That is plausible, and the ablation that removes VIF pruning tests it. What VIF adds is a readable, reportable locus set, as MLST schemes have, which a shrinkage distance does not produce. If accuracy is unchanged without VIF, we will keep it only as a reporting tool and say so.

Why discretize at all, when binning throws information away? Discrete alleles made bacterial typing portable across laboratories, but that portability came from nucleotide sequences being discrete by nature. Motor-unit features are continuous, so binning costs information. The discrete versus continuous comparison is included precisely because we expect the continuous version may win on accuracy. Discretization would earn its place only if it gives comparable recall with better cross-laboratory reproducibility of profiles, which we can test by decomposing the same recordings with two different decomposition pipelines.

Decisive Checks

The idea stands or falls on three results, all on held-out subjects.

First, on the shifted sub-grid benchmark, the selected profile must beat shift-compensated cross-correlation at one-row and two-row offsets, at matched false-match rate.

Second, on longitudinal data, it must retain clearly more than the roughly 40% one-week retention of waveform tracking at a decoy false-match rate of 5% or below. Tracked-unit ICCs must be at least as high.

Third, the features selected on training subjects must be largely the same features that are most stable on held-out subjects. If each person needs their own loci, the method has simply reproduced the single-case limitation it was designed to escape, and we will report that as the finding.