People with amputations can now control robotic hands through small pieces of muscle that surgeons wire onto severed nerves. The muscle works as a living amplifier. It turns faint nerve impulses of a few microvolts into millivolt-level muscle signals that implanted electrodes can read reliably for years. When the signal degrades, though, nobody can tell why. The electrode may have moved, the graft may have died from poor blood supply, scar tissue may have replaced muscle, or nerve fibers may have pulled back. Telling these apart today requires surgery, or in animals, sacrificing the subject for histology, meaning microscopic examination of stained tissue sections. We propose to train a model that reads the graft's own electrical activity and predicts what histology would show. The predicted quantities are the number of nerve cells feeding the muscle, the fraction of scar tissue, and the mix of fast and slow muscle fibers. The readout would come from a few-minute recording through the electrodes already implanted, so it serves as a rapid bedside diagnostic.
The central bet is that the size of the muscle signal is the wrong thing to watch. In rats, a muscle rewired with a nerve carrying many more axons received roughly 15 times more functional reinnervation and over 10 times more motor neuron sources than a low-capacity transfer, yet its evoked muscle response amplitude stayed comparable to controls (Peripheral neural interfaces: Skeletal muscles are hyper-reinnervated according to the axonal capacity of the surgically rewired nerves). Amplitude therefore misses an order-of-magnitude change in structure. The fine grain of the signal should not. That fine grain includes the number and size distribution of motor units (one nerve cell plus the muscle fibers it drives), which can be pulled apart from the recording, and the spectral shifts that accompany the fast-to-slow fiber conversion that same study saw under the microscope. We will test whether these features recover histology in rats, where paired tissue is plentiful. We will then test whether the same mapping holds in a larger animal before it is trusted in humans.
Core Claim
A model built on motor-unit-level features of implanted muscle recordings, trained against matched histology in rats, can predict four graft quantities in held-out animals. These are motor neuron source count, regenerated axon density, collagen fraction, and slow-fiber fraction. It should predict them substantially better than signal amplitude, signal-to-noise ratio, or time since surgery. It should also separate electrode faults from tissue faults. With only per-species normalization, and no retraining, the mapping should rank graft health correctly in a large animal. If amplitude-level features match the motor-unit features on held-out rats, the claim is false.
Where This Sits in Prior Work
Two rat studies supply the experimental designs and the histology targets. Both treat histology as a terminal endpoint. Tereshenko et al. varied donor nerve capacity and measured the outcome with five readouts. Masson-Goldner trichrome staining quantified muscle and collagen. Fast Blue tracer counted motor neurons. Neurofilament and choline acetyltransferase staining measured axon densities. Myosin heavy chain staining tracked fiber type. Evoked muscle responses recorded function. Tang et al. varied repair delay (immediate, 2 weeks, 4 weeks) and found that delay reduced response amplitude and force and caused atrophy, measured by intramuscular EMG and spinal cord synaptic staining after sacrifice. We reuse both manipulations as sources of controlled structural variation. Where those studies stop at the endpoint, we add a predictive mapping from signal to structure.
Human work motivates the problem. Long-term upper-extremity prosthetic control using regenerative peripheral nerve interfaces and implanted EMG electrodes reports signal-to-noise ratios above 15 for up to 1,054 days and decoding above 94% for 604 days without recalibration in two participants. Graft health in that work is inferred only from signal quality. Regenerative peripheral nerve interfaces (RPNIs) and implanted electrodes improve online control of prostheses for hand and wrist * shows implanted electrodes holding 78.3% and 75.1% accuracy during body movement, against 12.3% and 22.0% for dry surface electrodes. Implanted recording is the channel worth monitoring, and it is already in place. The review Machine learning for Regenerative Peripheral Nerve Interface-based prosthetic control: current applications and clinical translation traces the platform from macaques to humans. It describes machine learning used only for decoding intended movement, never for diagnosing the tissue.
The method for extracting motor units comes from Decoding motor neuron activity from epimysial thin-film electrode recordings following targeted muscle reinnervation. That work separates individual motor neuron firing from implanted recordings over reinnervated muscle. We use its decomposition as a feature extractor rather than a control signal.
The rapid-diagnostics literature mostly supplies a cautionary template. Compact image classifiers such as COVID-CAPS reached 98.3% accuracy after domain pre-training, but sensitivity fell from 90% to 80%. Headline accuracy can therefore hide the missed failures that matter most in a monitoring tool, and we report sensitivity on the failure class explicitly. Self-supervised pre-training on unlabeled audio rescued a cough classifier trained on only 75 positives (Classification of COVID-19 from Cough Using Autoregressive Predictive Coding Pretraining and Spectral Data Augmentation). That is a method we borrow, because unlabeled EMG from implanted animals is abundant while paired histology is scarce.
Research Program
The rat arm generates paired data. Grafts are made with high-capacity facial or low-capacity ulnar donor nerves, following Tereshenko. Repairs are done immediately or after 2 or 4 weeks of delay, following Tang. A further arm uses deliberately oversized free muscle grafts to induce central necrosis. A separate group with healthy grafts receives controlled electrode displacement, so that electrode faults can be distinguished from tissue faults. All animals carry implanted intramuscular electrodes.
Each animal undergoes a standardized five-minute protocol at several timepoints between 2 and 12 weeks. The protocol records electrode impedance, a graded nerve-stimulation series, spontaneous resting activity, and voluntary activity. Cohorts are sacrificed at staggered timepoints, so that every histology label has a recording from the same day. The first two resting recordings matter because spontaneous fibrillation at rest signals denervated fibers. The graded stimulation series reveals stepwise motor unit recruitment.
Features fall into three groups. The first is decomposed motor unit statistics: count, size distribution, and firing regularity. The second is spectral and conduction features expected to shift with fiber type. The third is impedance and noise features expected to reflect the electrode rather than the tissue. Because cohorts will be tens of animals, the predictor is deliberately small. It is a multi-output regression for the four histology quantities plus a four-way fault classifier (healthy, denervated or atrophic, necrotic or fibrotic, electrode fault). An encoder pre-trained self-supervised on all unlabeled recordings can be swapped in, following the predictive-coding recipe of Wani et al.
Cross-species transfer is handled by expressing features relative to each animal's own early post-operative baseline and to known muscle size. The model then predicts structural change rather than absolute values that depend on scale. Rodent regeneration finishes in weeks, while large-animal and human regeneration takes months. Recordings are therefore indexed by regeneration stage, estimated from recruitment onset, not by calendar days.
The large-animal arm uses pigs or macaques with the same protocol. Needle biopsies are taken at a small number of timepoints rather than terminal sections. Following the cross-species scaling logic of Lyophilized apoptotic vesicle-encapsulated adhesive hydrogel sponge as a rapid hemostat for traumatic hemorrhage in coagulopathy, which moved from rat to rabbit because rodent vessels could not reproduce human loads, we treat the large animal as the decisive test, not the rat. Sample sizes there will be small, realistically N=3 to 6. We therefore commit in advance to rank-based tests.
Finally, the human check is retrospective and purely predictive. If similar longitudinal recordings from implanted participants can be accessed, we will test whether the model's structural estimates foreshadow later drops in decoding accuracy before the signal-to-noise ratio falls. Histology is never available there.
Baselines
The naive reference is time since surgery combined with surgical group. This model predicts each histology quantity from the surgical arm and the elapsed weeks alone. It is stronger than it sounds, because both prior rat studies show large group-level effects of donor nerve and delay. Any signal-based model must beat it on held-out animals within the same group. Otherwise it is only re-deriving the experimental design.
The strongest existing practice is amplitude-level monitoring, the evoked response amplitude plus signal-to-noise ratio used to judge graft health in Vu et al.. The cited number is a signal-to-noise ratio above 15 sustained to 1,054 days. That is a stability figure, not a diagnostic accuracy. I found no reported result on Lacuna for predicting graft histology or fault type from implanted EMG. We will fit this baseline with the same regressors, the same data, and the same cross-validation over animals. Beating it means an absolute gain in held-out rank correlation for motor neuron count of at least 0.2 with non-overlapping animal-bootstrap intervals, plus a fault-classification AUROC improvement on the necrosis-versus-electrode distinction. Tereshenko's finding of comparable amplitude despite 10-fold more motor neuron sources predicts this baseline will fail specifically on innervation count.
The third baseline is motor unit counting alone. It uses decomposition-derived counts with no spectral, resting-activity, or impedance features, and it isolates whether the extra feature groups earn their place.
An end-to-end convolutional network on raw recordings, given an equal hyperparameter search budget, tests whether hand-designed physiology helps at this sample size.
Two ablations isolate the key ingredients. Removing decomposition features tests the central claim. Removing per-animal baseline normalization tests whether cross-species transfer depends on it.
Why not just…?
Why not just watch amplitude and signal-to-noise ratio, as clinicians already do? That approach is cheap and already deployed, and for gross graft death it may be enough. Its blind spot is documented, however: a roughly 15-fold reinnervation difference produced comparable amplitudes in Tereshenko's rats. Amplitude also cannot separate a displaced electrode from a dying graft, since both reduce it. The amplitude baseline will settle how much this matters. If it ties our model on the fault classifier, we will narrow the claim to innervation count only.
Why not just use classical motor unit number estimation? This is the real threat. For the innervation quantities it may capture most of the benefit, and the motor-unit-counting baseline tests exactly that. Our expectation is that counts alone cannot distinguish fibrosis or necrosis from denervation, and cannot flag electrode faults. If the count baseline matches the full model on all four outputs, the proposal reduces to validating that standard technique against histology across species. That would be a smaller but still useful result, and we would report it as such.
Why not just image the graft with ultrasound or biopsy it? Biopsy is the invasive procedure we aim to avoid, and repeated sampling damages small grafts. Bedside ultrasound has standardized protocols for other uses (Developing a Consensus-Based POCUS Protocol for Critically Ill Patients During Pandemics: A Modified Delphi Study), but I found no reported result on Lacuna for ultrasound assessment of nerve-muscle grafts. In the large-animal arm, where feasible, we will record ultrasound echo intensity at biopsy timepoints as an additional comparator rather than assume it is inferior.
Why not just train a big end-to-end network? With tens of labeled animals, an unconstrained network will learn which animal or group it is looking at. The end-to-end baseline, evaluated with animal-level held-out splits, will show this directly.
Proposed Experiments
The primary rat experiment uses leave-animals-out cross-validation. The four histology targets are motor neuron count from Fast Blue tracing, axon density from neurofilament and choline acetyltransferase staining, collagen fraction from trichrome staining, and slow-fiber fraction from myosin heavy chain staining. Each is scored by Spearman correlation and R² against held-out measurements. The four-way fault classifier is scored by macro AUROC, with sensitivity at fixed specificity reported for the necrotic class, given the sensitivity loss seen in COVID-CAPS.
The decisive cross-species check trains on rats only and applies the model, frozen, to large-animal recordings. It asks whether predicted rankings of graft health across biopsy timepoints agree with biopsy rankings better than the amplitude baseline does. A second variant adds few-shot recalibration of one animal's baseline.
A falsification test asks whether the model's advantage over amplitude is concentrated in the high-capacity donor arm, as the hyper-reinnervation result predicts. If the advantage is spread evenly across arms, our explanation of why it works is wrong even if the model works.
The prospective test in the oversized-graft and delayed-repair arms asks whether the electrical readout flags tissue failure at least one recording session before amplitude falls. That lead time is the property that would make this a useful rapid diagnostic rather than a post-hoc correlate.
References
- Peripheral neural interfaces: Skeletal muscles are hyper-reinnervated according to the axonal capacity of the surgically rewired nerves
- Improving nerve and muscle function: an exploration of targeted nerve function replacement following differential delay periods in a rat model
- Long-term upper-extremity prosthetic control using regenerative peripheral nerve interfaces and implanted EMG electrodes
- Regenerative peripheral nerve interfaces (RPNIs) and implanted electrodes improve online control of prostheses for hand and wrist *
- Machine learning for Regenerative Peripheral Nerve Interface-based prosthetic control: current applications and clinical translation
- Decoding motor neuron activity from epimysial thin-film electrode recordings following targeted muscle reinnervation
- COVID-CAPS: A capsule network-based framework for identification of COVID-19 cases from X-ray images
- Classification of COVID-19 from Cough Using Autoregressive Predictive Coding Pretraining and Spectral Data Augmentation
- Lyophilized apoptotic vesicle-encapsulated adhesive hydrogel sponge as a rapid hemostat for traumatic hemorrhage in coagulopathy
- Developing a Consensus-Based POCUS Protocol for Critically Ill Patients During Pandemics: A Modified Delphi Study