Certify the Update, Not the Weights: Bounded Recalibration for Myoelectric Decoders

A continuous myoelectric decoder is a model that reads electrical activity from forearm muscles and outputs smooth joint angles for a prosthetic hand. Such decoders degrade from one day to the next because electrodes shift, skin impedance changes, and muscles tire. The engineering fix is to recalibrate often. Medical device regulation, however, certifies a fixed piece of software, and changing the weights can trigger recertification. This clash is one concrete, measurable instance of regulatory lag: the device that stays accurate is not the device that was approved. We propose to split the decoder into two parts. The first is a nonlinear decoder that is certified once and never retrained. The second is a very small, physically motivated input-alignment layer. This layer is the only thing allowed to change, and each change must pass a pre-registered statistical acceptance test on a short calibration block before it goes live. The question is empirical. Does such a restricted, auditable update recover most of the accuracy that full retraining recovers, while keeping a stated per-joint error bound valid across sessions?

In two sentences: sEMG drift across sessions is mostly a change in how muscle activity reaches the electrodes, not a change in what the hand is doing. So a certified decoder plus a tiny, pre-approved "re-alignment" knob with a statistical admission test should stay accurate and stay within its certified error claim, without the regulator having to re-approve a new model.

Core Claim

Current continuous decoders are optimized only for accuracy. KinEMbed aligns sEMG and joint-angle embeddings with a contrastive loss and then regresses joint angles with a temporal convolutional network. It reaches mean $R^2$ = 0.732 on NinaPro DB8 under a cross-session split, 0.769 for able-bodied subjects, and 0.568 for the two subjects with limb differences. Continual Imitation Learning for Prosthetic Limbs keeps a gait regressor above $R^2$ = 0.95 under distribution shift, but it does so by continually updating the whole network. Neither work asks what, exactly, a regulator would be approving.

Policy work motivates the gap without supplying a method. Busch et al. describe how the EU AI Act is harmonized with the Medical Device Regulation around conformity assessment. Luna et al. document how static legal definitions decay as models change. Khan et al. find that AI bills stall in specialist committees (odds ratio 0.34 for reaching a floor vote), which is consistent with, though it does not prove, a shortage of testable technical definitions. The evidence dilemma in the First International AI Safety Report frames the cost: regulators either act without data or wait for harm.

Our claim is that for sEMG decoding, a certifiable object exists that is narrower than "the model" and broader than "a frozen model." That object is an update class together with an admission test. We will show whether this object keeps accuracy close to full retraining. We make no claim about how regulators would actually respond to it.

Research Program

The decoder is written as $\hat y_t = f(A_s x_t + b_s)$. Here $x_t$ is the multichannel sEMG feature vector at time $t$, $f$ is the certified nonlinear decoder, and $A_s, b_s$ are a channel-mixing matrix and offset for session $s$. The decoder $f$ is either a KinEMbed-style encoder with a temporal convolutional network or a plain temporal convolutional network, frozen after training on sessions 1 and 2. The alignment layer starts at $A_s = I$ and $b_s = 0$.

The update class is deliberately physical, and we study it at three sizes. The smallest is diagonal gains, which model impedance changes. The middle is a banded mixing matrix, which models electrode rotation or displacement onto neighboring muscles. The largest is full channel mixing. Before each session the user performs a short calibration block of guided movements, and $A_s, b_s$ are fit on it with a proximity penalty toward the identity.

The certified claim is a per-joint error bound $q_j$ at miscoverage level $\alpha$. It is set at approval time by split conformal prediction on held-out training-session data. Split conformal prediction is a distribution-free method that turns held-out residuals into an error band with guaranteed coverage when the data are exchangeable. A session's update is admitted only if two conditions hold on a disjoint slice of that session's calibration block. First, the conformal $(1-\alpha)$ residual quantile must not exceed $q_j$ for any joint:

$$\hat Q^{(s)}_{1-\alpha}\big(|y_{t,j}-f(A_s x_t+b_s)_j|\big)\le q_j \quad \text{for all joints } j.$$

Second, $\|A_s - I\|$ must stay inside a pre-declared radius. If either condition fails, the device falls back to identity alignment and flags the session.

The reason for this design is that $f$'s behavior on any given input stays byte-identical. The audit surface therefore shrinks to a handful of interpretable gains and mixing weights plus a logged pass/fail test. That is the sense in which the proposal answers the "static rules versus adaptive systems" problem.

Around this we build a software stress-test harness, a modest biosignal analogue of the aviation sandbox in LINA. LINA builds physical test ranges so regulators can accept evidence about adaptive UAV controllers. We are not building hardware. The harness injects synthetic channel gain changes, adjacent-channel mixing, and additive noise into recorded sessions, so that we can see where the admission test correctly accepts and correctly refuses updates.

Baselines

The naive reference point is the frozen decoder with no adaptation, evaluated on session 3 of NinaPro DB8. This quantifies how much cross-session drift costs and how often the certified bound $q_j$ is violated when nothing is done.

The strongest existing decoder is KinEMbed. Its reported NinaPro DB8 cross-session result is mean $R^2$ = 0.732 overall and 0.568 on the limb-difference subjects, where CEBRA reported 0.570, all from KinEMbed. We will reimplement it on the identical split (train on sessions 1–2, test on session 3). The reimplementation must land near 0.732 before any comparison counts. Every method gets the same training data, the same calibration-block length, and the same hyperparameter search budget.

The strongest adaptation competitor is full fine-tuning of the whole decoder on the same calibration block. As a second variant we use a continual-learning update in the style of Dey et al., who report $R^2$ above 0.95 on ENABL3S and the Embry gait data. We have no reported result on Lacuna for any sEMG input-alignment or conformal-admission method on NinaPro DB8. Our method therefore "wins" on either of two conditions:

  • It recovers at least roughly 80% of the $R^2$ gain that full fine-tuning achieves over the frozen decoder.
  • It matches that gain while keeping empirical per-joint coverage at or above $1-\alpha$, where fine-tuning without an admission test does not.

The key ablation removes the admission test while keeping the alignment layer. A second ablation sweeps the update-class size from diagonal to full mixing. Together these isolate whether the benefit comes from restricting the update, from the statistical gate, or from both.

Why not just…?

Why not just recalibrate the whole decoder every session? This is the real threat to the accuracy half of the claim. Full retraining will likely be the most accurate option, and we expect it to win on raw $R^2$. The proposal survives only if the restricted update closes most of the gap. If full fine-tuning beats banded alignment by a wide margin on session 3, then drift is not mostly input-side, and we will report that as a negative finding. The regulatory argument for full retraining is weak regardless: every retraining produces a new $f$, which is exactly the static-certification problem Busch et al. and Lin et al. describe.

Why not just clip outputs to anatomical joint limits? Clipping prevents impossible poses. It says nothing about errors inside the valid range, such as a hand closing 20 degrees too far on a paper cup. That graded regime is precisely why continuous decoding replaced classification in Luu et al.. We include clipping in every condition, so it cannot explain any difference between them.

Why not just put conformal error bands on the frozen model? Conformal coverage assumes exchangeable data, and cross-session drift breaks that assumption. The frozen model's bands will either under-cover on session 3 or have to be widened until they are useless. We test this directly by reporting the coverage of frozen-model bands on session 3. If they happen to hold at usable width, then the alignment layer is unnecessary, and we will say so.

Why not just use an interpretable linear decoder like PLS? A linear decoder is easy to audit. However, KinEMbed reports that PCA and PLS are outperformed by nonlinear representations on DB8, especially for thumb joints. Our design keeps the nonlinear decoder frozen and confines interpretability to the part that changes. We report PLS with the same alignment layer as a reference.

Proposed Experiments

The primary experiment is on NinaPro DB8: 11 subjects, 16 sEMG channels, 5 degrees of actuation, cross-session split. For each condition we report per-joint $R^2$, per-joint empirical coverage of $q_j$, admission rate, and the size of the admitted update $\|A_s - I\|$. We also report one number we call "certified operating fraction." It is the fraction of test time during which the device runs inside its original certified claim without any change to $f$.

We vary calibration-block length to find the smallest block at which alignment is admitted reliably. We report the two limb-difference subjects separately and descriptively; with N=2, no statistical claim is possible.

The stress-test harness supplies the decisive check on the admission test. It should accept synthetic gain and neighbor-mixing perturbations that alignment can undo, and refuse perturbations that it cannot, such as noise or channel dropout beyond the declared radius. We measure its false-accept rate against the nominal $\alpha$.

A secondary experiment on HYGRIP applies the same scheme to continuous grip-force regression from forearm sEMG, a harder target where intermediate force levels are difficult to decode. We found no reported baseline result on Lacuna for HYGRIP force regression, so its frozen-decoder result becomes the reference. On the ENABL3S gait data we compare against Dey et al.'s continual updating to see whether input-side alignment is enough outside the hand.

The idea is falsified if either of two things happens on DB8. One is that banded alignment recovers less than half of fine-tuning's gain. The other is that the admission test's coverage fails at the nominal level on real session-3 data.