# Does Training on Intact Limbs Help Decode Amputee EMG Gestures?

**Yes — but only after at least a few labeled calibration repetitions from the amputee user.** A preprint posted today by Jethro Odeyemi and W. J. Zhang (arXiv:2609.20297) tested a montage-agnostic cross-user surface-EMG encoder on eleven transradial amputees and found that zero-shot cross-population transfer fails outright. Once three labeled repetitions are collected from the new user, however, the encoder reaches **0.779 macro-F1** — compared to **0.589 macro-F1** for a standard per-user classifier fit from scratch. That is a **+0.190 macro-F1 advantage** that held for every subject in the eleven-person cohort.

The study also settles a practical sourcing question for developers building prosthetic control pipelines: training on **forty intact subjects** produced better transfer to a new amputee than training on **ten amputees**, and combining both source pools outperformed either alone. These findings are directly relevant to anyone designing calibration workflows for myoelectric prosthetic hands — a space where data collection from amputees is expensive, slow, and clinically constrained.

*Note: This is a preprint feasibility study with eleven participants. Results have not been peer-reviewed and should not be interpreted as clinical validation.*

---

## The Evidence Gap This Study Targets

Odeyemi and Zhang conducted a systematic review of **1,077 studies** in surface-EMG gesture decoding before running their experiment. Their headline finding from that review: amputees appear in approximately **one in six** of those studies. The field has overwhelmingly developed and benchmarked decoders on intact-limb participants, then assumed — often incorrectly — that performance generalizes.

This matters because amputee EMG is physiologically distinct. The residual limb musculature is reorganized, electrode placement options are constrained, and signal separability between gesture classes is generally reduced compared to intact hands. A model optimized on intact-limb data carries assumptions that can collapse when applied to a fundamentally different signal source.

The authors' pre-registered hypothesis was that amputee EMG is *less separable*, which would explain why cross-population transfer underperforms. That prediction held — but only partially, and only under specific conditions.

---

## What the Encoder Actually Does

The encoder is described as "montage-agnostic," meaning it does not require a fixed electrode placement grid — an important property for clinical deployment where amputee limb geometry varies substantially between patients. It was originally trained on intact-limb data and is being extended here to the amputee population for the first time.

The experimental protocol was deliberately matched to the intact-limb training protocol, controlling for the most obvious confound in cross-population comparisons.

**Key performance results (from source text):**

| Condition | Macro-F1 |
|---|---|
| Per-user classifier (baseline) at 3 reps | 0.589 |
| Cross-user encoder at 3 reps | 0.779 |
| Advantage | +0.190 |

The +0.190 gap held for **every subject** in the cohort — not just on average. That consistency across individuals, rather than aggregate improvement driven by a few easy subjects, is the more clinically meaningful result.

---

## Where the Pre-Registered Hypothesis Breaks Down

The authors are transparent about where their account of the data fails. The prediction that amputee EMG is less separable — and that this explains transfer difficulty — only holds after:

1. A few labeled repetitions from the amputee user become available, **and**
2. The source pool is enriched with additional amputee data

At **a single labeled repetition**, the prediction fails. At every calibration budget below a source pool matched to the intact-limb comparison, it also fails. The authors describe this as "locating the boundary of the proposed account" — a rigorous acknowledgment that the mechanistic explanation is incomplete, even if the practical performance advantage is robust.

This kind of pre-registration and honest failure reporting is uncommon in the EMG decoding literature and strengthens confidence in the positive findings.

---

## Calibration Budget: The Clinical Bottleneck

The three-repetition threshold is operationally significant. In a clinical fitting session for a myoelectric prosthetic hand, three labeled gesture repetitions per class is a plausible ask — far less burdensome than the extended calibration sessions that have historically limited real-world adoption of pattern-recognition control.

The field, including companies like [BrainRobotics](https://bciintel.com/companies/brainrobotics) and the broader myoelectric prosthetics ecosystem, has long struggled with the gap between laboratory decoding accuracy and at-home performance. A persistent contributor to that gap is calibration burden: users who perform well in a structured session deteriorate rapidly as electrode shift, sweat, and fatigue alter the signal distribution. An encoder that requires only three repetitions to outperform a full per-user pipeline meaningfully changes that calculus.

For engineers and researchers at the intersection of EMG-based prosthetic control and robotic arm systems, [humanoidintel.ai](https://humanoidintel.ai) tracks parallel developments in dexterous manipulation and neural-to-robotic hand control pipelines that intersect with this decoding work.

---

## Implications for BCI and Neuroprosthetics Development

**Training data strategy:** The finding that forty intact subjects outperform ten amputees as a transfer source is counterintuitive but practically useful. Intact-limb data is far cheaper and faster to collect at scale. This gives developers a principled justification for building large intact-limb datasets as the primary pretraining corpus, supplemented by smaller amputee datasets.

**Zero-shot is not viable:** Any product roadmap that assumes a pre-trained model will work out-of-the-box on a new amputee user should be revised. The data are unambiguous: the encoder requires labeled calibration from the new user before it begins decoding at all. Zero-shot cross-population transfer fails outright.

**Clinical translation timeline:** Surface-EMG gesture decoding for prosthetic control has been "almost ready" for broad clinical deployment for over a decade. The calibration burden problem has been a consistent brake. A three-repetition calibration budget that yields 0.779 macro-F1 — if it replicates in larger, more diverse cohorts — would represent a meaningful step toward routine clinical deployment, though the eleven-subject cohort here is too small to draw definitive conclusions.

**Regulatory path:** Surface-EMG prosthetic controllers that incorporate adaptive, transfer-learning-based decoders will face scrutiny around how calibration session data affects device performance and whether the decoder's behavior is sufficiently predictable across the patient population. FDA's guidance on adaptive algorithms in software as a medical device (SaMD) will be relevant to any commercialization effort.

---

## Key Takeaways

- A cross-user surface-EMG encoder tested on **11 transradial amputees** reaches **0.779 macro-F1** with three labeled calibration repetitions, versus **0.589** for a standard per-user pipeline — a **+0.190** advantage consistent across every subject
- **Zero-shot cross-population transfer fails entirely** — the encoder requires at least some labeled data from the new user before it decodes at all
- Training on **40 intact subjects** transfers better to new amputees than training on **10 amputees**; combining both sources outperforms either alone
- A systematic review of **1,077 studies** found amputees represented in only about **one in six**, quantifying how intact-limb-centric the field's evidence base remains
- The authors' pre-registered mechanistic hypothesis holds only partially — it fails at a single repetition and at low calibration budgets, an honest finding that constrains interpretation
- This is a preprint feasibility study (n=11); peer review and larger replication cohorts are required before clinical or commercial conclusions

---

## Frequently Asked Questions

**What is cross-population transfer in surface-EMG decoding?**
It refers to training a gesture decoder on one group of users (e.g., people with intact limbs) and applying it to a different group (e.g., amputees) without retraining from scratch. This study found that zero-shot transfer — applying the model with no new data at all — fails completely in the amputee population.

**How many labeled repetitions does the encoder need before it outperforms a per-user classifier?**
According to this study, three labeled repetitions per gesture class were sufficient for the cross-user encoder to exceed the per-user baseline by 0.190 macro-F1, with the advantage holding for all eleven subjects in the cohort.

**Why is intact-limb training data more useful than amputee training data for transfer?**
The study found that forty intact subjects produced better transfer to a new amputee than ten amputees did. The authors suggest this is partly a scale effect — larger intact-limb datasets are easier to collect — but the mechanistic explanation for why intact data transfers better is not fully resolved and is identified as a boundary of their theoretical account.

**What does macro-F1 mean in gesture decoding studies?**
Macro-F1 averages the F1 score (a harmonic mean of precision and recall) equally across all gesture classes, regardless of class frequency. It is a more demanding metric than accuracy when gesture classes are imbalanced, which is common in clinical populations.

**Is this research clinically validated?**
No. This is a preprint with eleven participants that has not undergone peer review. The results are from a feasibility study and should not be interpreted as clinical validation. Larger, controlled trials with diverse amputee cohorts would be required before these findings inform clinical or regulatory decisions.