Modern prosthetic hands have more motors than their users can command. Each joint needs its own control signal, and after amputation the remaining muscles often supply only two or three usable signals, recorded through skin electrodes (surface electromyography, or sEMG, which measures the electrical activity of contracting muscle). This mismatch is called the myoelectric control paradox. Clinicians currently learn how much control a particular residual limb can deliver only after a long fitting and training process, and sometimes only when the user abandons the device. We propose to turn that unknown into a rapid diagnostic. A one-minute recording would produce a "control capacity profile": an estimate of how accurately a decoder will eventually separate each candidate movement for this person, and which set of movements is worth assigning to the prosthesis.
The opportunity comes from the rapid-diagnostics literature, which solves a similar problem in a different signal. Labelled clinical data are scarce there too, and the most effective fix has been self-supervised pretraining on large unlabelled corpora from the same domain, followed by a small labelled fine-tune. We carry that recipe over. An sEMG encoder pretrained without labels would turn a few seconds of each attempted movement into an embedding whose geometry predicts what a fully calibrated decoder will later achieve. The claim is new in kind. Prior myoelectric work improves the decoder itself, while this work predicts, before full calibration, the ceiling any decoder will reach for a given limb. It also says which movements to drop so the user stays above a reliability threshold.
Core Claim
From one repetition per attempted movement, a domain-pretrained sEMG encoder can predict each user's full-protocol decoding accuracy, both overall and for each pair of movements. Its predictions will be more accurate than the obvious shortcut of training a decoder on that same scrap of data and reading off its cross-validated accuracy. We then use these predictions to choose movement sets that keep the user above a fixed accuracy threshold.
Formally, let $a_u(S)$ be the accuracy that a reference decoder, trained on the full protocol, reaches for user $u$ on movement set $S$. Let $\hat a_u(S)$ be our estimate from the short recording. The diagnostic reports
$$\hat S_u = \arg\max_{S} |S| \quad \text{subject to} \quad \hat a_u(S) \ge \tau,$$
where $\tau$ is a clinician-chosen reliability threshold, for example 90%. The test succeeds if $a_u(\hat S_u) \ge \tau$ holds for most users and $|\hat S_u|$ comes close to the largest set an oracle with full data would choose.
Why This Sits Where the Two Literatures Meet
Several papers motivate the gap. Castellini and colleagues frame the paradox as hardware outpacing the interface's ability to extract signals (Proceedings of the first workshop on Peripheral Machine Interfaces). Jiang and colleagues show that signal availability collapses at more proximal amputation levels (EMG-based simultaneous and proportional estimation of wrist/hand kinematics in uni-lateral trans-radial amputees). The surgical answers make the stakes of an early capacity estimate concrete. Targeted muscle reinnervation reroutes severed nerves into remaining muscle so it can act as a biological amplifier (Broadband Prosthetic Interfaces). Combined reinnervation and bone anchoring shows long-term benefit, but it carries surgical risk and slow nerve regrowth (Long-term functional and clinical outcome of combined targeted muscle reinnervation and osseointegration). A cheap, non-invasive estimate of how much control the existing limb provides is exactly what a clinician would want before choosing between a better decoder and an operation. We do not evaluate that surgical decision here; it is the downstream use.
The method comes from the diagnostics side. Harvill and colleagues pretrained a predictive-coding model on more than 20,000 unlabelled coughs and fine-tuned it on a small labelled set. Predictive coding here means a network trained to forecast future frames of a signal. They found that pretraining on coughs beat pretraining on generic speech, and that a longer prediction horizon captured more useful structure (Classification of COVID-19 from Cough Using Autoregressive Predictive Coding Pretraining and Spectral Data Augmentation). Mehlman and colleagues supply both the sEMG pretraining corpus and the encoder design. They trained Transformers on 346 hours of raw typing sEMG and found that raw-waveform encoders beat spectrogram inputs by 8–11% relative character error rate (Scaling and Distilling Transformer Models for sEMG). Neither paper predicts anything about an individual's capacity. Ours does not try to improve typing or cough detection; it reuses their pretraining logic for a different output.
Two results constrain the design. Qamar and colleagues' dual-pathway spectrogram network falls behind off-the-shelf ImageNet backbones on amputees, with 85.36% against VGG19's 87.96% on NinaPro DB3 (EMG gesture signal analysis towards diagnosis of upper limb). Custom architectures trained from scratch therefore generalize poorly to damaged muscle, which argues for pretraining over architectural cleverness. Second, Jiang, Lorrain and Farina calibrate a working controller from only 48 seconds of data using simple features and linear discriminant analysis, a linear classifier (A state-based, proportional myoelectric control method). Short calibrations clearly carry a lot of information, so our method must beat the "just train on the short data" baseline rather than ignore it.
Research Program
We build three things. The first is a raw-waveform sEMG encoder. We pretrain it without labels on emg2qwerty using next-segment prediction with a horizon of roughly 100 ms, following the horizon result from the cough work. We then continue pretraining on the unlabelled portions of NinaPro DB1, DB2 and DB8 recordings from users held out of the evaluation fold.
The second is the capacity estimator. It embeds the short calibration recording, computes per-movement-pair separability in embedding space, and maps these quantities to a predicted pairwise confusion matrix. Separability is measured by between-class distance relative to within-class spread, estimated from the few available windows. The mapping is a small calibrated regressor trained across users, and it converts the predicted pairwise confusions into $\hat a_u(S)$ for any movement set $S$.
The third is the selection procedure above. A greedy search over movement sets is enough at NinaPro's scale.
One methodological concern is circularity. The ground truth $a_u(S)$ depends on which reference decoder we choose. We therefore report everything against two reference decoders: linear discriminant analysis on Hudgins time-domain features, and a convolutional network. We check whether the predicted ranking of users and movements holds for both.
We also create capacity variation under control, because amputee cohorts are small. NinaPro DB3 has 10–11 subjects. Following the channel-masking logic that Parés-Pujolràs and colleagues used to simulate cervical spinal injury (Dealing with Target Uncertainty in a Reaching Control Interface), we mask electrode subsets in intact-subject data from DB1 and DB2. Each intact user then becomes a family of synthetic "limbs" with known, graded loss of recording sites. This gives hundreds of user-condition units for correlation analysis, while DB3 remains the honest test on real scarred tissue.
Baselines
The naive reference point is a population prior. It predicts each new user's accuracy as the training-fold mean for their group (intact or amputee) and assigns the standard movement set to everyone. Any useful diagnostic must beat it on per-user error. Otherwise the test adds nothing beyond knowing the amputation status.
The strongest competitor is short-calibration cross-validation. We train the reference decoder itself on the same one-repetition-per-movement budget and use its leave-one-window-out accuracy as $\hat a_u(S)$. This is what a clinician would do with the Jiang-style 48-second pipeline (A state-based, proportional myoelectric control method), and it shares our data budget exactly. A stronger variant extrapolates this estimate with a power-law learning curve across half and full repetitions. To beat these baselines, our estimator must achieve significantly lower mean absolute error in predicted accuracy, and higher Spearman rank correlation with true accuracy, under leave-one-subject-out evaluation on both DB3 and the masked-channel cohorts. Significance is judged by paired bootstrap over users.
For context on the target decoder's ceiling, the best reported amputee gesture accuracy in the file is VGG19 at 87.96% on 23 NinaPro DB3 gestures (EMG gesture signal analysis towards diagnosis of upper limb). For the diagnostic task itself, predicting per-user decodability from a short recording, we found no reported result on Lacuna, so all comparisons are run by us under one protocol. Every baseline gets the same calibration windows, the same folds, and the same hyperparameter search budget, measured in configurations tried on inner validation folds.
Three ablations isolate the key ingredient. In the first, the separability estimator runs on Hudgins features instead of learned embeddings. In the second, it runs on a randomly initialized encoder. The third swaps domain-specific sEMG pretraining for an ImageNet backbone on spectrograms. That last comparison tests whether the DB3 advantage of vision backbones carries over to capacity prediction, and it addresses the raw-versus-spectrogram tension between Qamar and colleagues and Mehlman and colleagues.
Why not just…?
Why not just train the decoder on the short calibration and read its accuracy? This is the real threat. With very few windows per movement, cross-validated accuracy is noisy and biased downward, and a linear model cannot exploit structure shared across users. Pretraining is meant to supply exactly that shared structure. Still, the 48-second results in A state-based, proportional myoelectric control method show how much short data can already reveal. If our estimator does not beat this baseline on DB3 and the masked cohorts, we will report the diagnostic as a standardized short-calibration protocol, without claiming a learned estimator.
Why not just use simple signal-quality measures such as channel amplitude or the count of active electrodes? These measures capture how many sites are live but not whether the movements they encode are distinct. Two strong channels driven by the same remnant muscle give one control signal, not two. The masked-channel experiment tests this directly. If the active-channel count alone tracks true accuracy as well as our estimator under masking, the learned embedding is unnecessary for site loss. Its value would then have to appear on DB3, where scarring degrades separability without removing channels.
Why not just run the full calibration? Full calibration answers the question only after the expensive steps of socket fitting, training sessions and device configuration have already happened. It also requires the user to complete a long protocol that the most impaired users struggle with. Wearable-sleeve work shows that even fast decoders still depend on lengthy lab sessions and caregiver help (Decoding hand and wrist movement intention from chronic stroke survivors). The diagnostic is worth having only if it is cheaper than the thing it predicts. That is why the budget is fixed at one repetition per movement.
Proposed Experiments
On NinaPro DB1, DB2 and DB3, the calibration input is the first repetition of each of the 23 common gestures. The ground truth is the reference decoders' accuracy on held-out later repetitions, both overall and for each gesture pair. We report mean absolute error and Spearman correlation per user. We also report pairwise-confusion prediction error, which provides hundreds of gesture-pair data points even in the small DB3 cohort. For selection, we report the fraction of users for whom $a_u(\hat S_u) \ge \tau$ at $\tau$ = 85% and 90%, and the ratio of $|\hat S_u|$ to the size of the oracle set.
On NinaPro DB8, we test whether a short slice of session one predicts cross-session degradation, measured as the R² of a continuous finger-angle decoder on session three. KinEMbed's reported 0.732 overall and 0.568 for the two limb-difference users (KinEMbed) serve as the reference decoder's expected range. The masked-channel cohorts provide the dose-response check. Predicted capacity should fall monotonically as sites are removed, and its slope should match the slope of true accuracy.
Two outcomes would settle the question. The first is whether the pretrained estimator beats short-calibration cross-validation on DB3 per-pair confusion prediction under leave-one-subject-out evaluation. The second is whether the advantage persists when the reference decoder is swapped. If both hold, a one-minute recording gives a clinically usable capacity estimate. If the estimator wins only with one reference decoder, it has learned that decoder's quirks rather than the limb's capacity. We would report that outcome as negative.
References
- Proceedings of the first workshop on Peripheral Machine Interfaces: going beyond traditional surface electromyography
- EMG-based simultaneous and proportional estimation of wrist/hand kinematics in uni-lateral trans-radial amputees
- Broadband Prosthetic Interfaces: Combining Nerve Transfers and Implantable Multichannel EMG Technology to Decode Spinal Motor Neuron Activity
- Long-term functional and clinical outcome of combined targeted muscle reinnervation and osseointegration for functional bionic reconstruction in transhumeral amputees: a case series
- Classification of COVID-19 from Cough Using Autoregressive Predictive Coding Pretraining and Spectral Data Augmentation
- Scaling and Distilling Transformer Models for sEMG
- EMG gesture signal analysis towards diagnosis of upper limb using dual-pathway convolutional neural network
- A state-based, proportional myoelectric control method: online validation and comparison with the clinical state-of-the-art
- Dealing with Target Uncertainty in a Reaching Control Interface
- Decoding hand and wrist movement intention from chronic stroke survivors with hemiparesis using a user-friendly, wearable EMG-based neural interface
- KinEMbed: Decoding Kinematics from Electromyography via Cross-Modal Contrastive Learning