Deep Neural Networks for Learning Intent from sEMG Signals to Support Hardware Devices for Post-Stroke Neurorehabilitation
Zakariyya Brewster, Divy Wadhwani, Emily Yan, Aidan Wang, Karma Namgyal, Shuting Xie, Markiyan Konyk, Tala Abdelmaguid
cs.LG
Quick verdict
Post-stroke sEMG five-finger intent decoding distilled to a 4-channel ONNX model; honest systems study, but single-split evidence and modest accuracy.
Summary
The paper formulates five-finger multilabel motor-intent decoding from impaired-arm high-density sEMG as a hardware-aware machine-learning problem, using PhysioMio, a bilateral longitudinal dataset of 48 stroke patients (64 electrodes, 16 gestures). A shared pipeline — label alignment, 20–450 Hz Butterworth filtering, Symlet-4 wavelet denoising, 200 ms overlapping windows, twelve time/frequency features — feeds LSTM, CNN, and GNN decoders under patient-level 70/10/20 splits. Direct baselines are complementary: LSTM leads subset accuracy (0.545), GNN leads macro F1 (0.706) and macro AUPRC (0.776). An Optuna architecture search finds CNN-Large best among CNNs (0.593 subset accuracy, 0.714 macro F1), with the 158K-parameter CNN-Micro nearly matching it. For a four-sensor design (ECRB, ECRL, FDS, FDP), cross-channel knowledge distillation from a 64-channel teacher outperforms direct training and first-layer-sliced transfer, reaching 0.5219±0.0114 subset accuracy and 0.6095±0.0058 macro F1 across five seeds; the 123K-parameter student is exported to ONNX. Healthy-to-impaired pretraining helps CNN-Base but hurts LSTM on ranking metrics. Limitations: most tables are single-split, single-run estimates without significance tests; no classical ML baseline on identical features; the 64-channel reference is single-seed; no on-device latency; and 'intent' labels derive from performed gesture annotations rather than attempted movement in weak movers.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- On impaired-arm PhysioMio HD-sEMG, LSTM and GNN decoders are complementary: LSTM attains the highest subset accuracy (0.545) while GNN attains the highest macro F1 (0.706) and macro AUPRC (0.776) (Table 2).
- Optuna architecture search identifies CNN-Large as the strongest CNN (0.593 subset accuracy, 0.714 macro F1) and CNN-Micro as a compact near-equal alternative at ~1/5 the INT8 footprint (Table 3).
- Cross-channel knowledge distillation from a 64-channel teacher improves the four-channel CNN-Micro over direct training and sliced-transfer initialization across five seeds, reaching 0.5219±0.0114 subset accuracy and 0.6095±0.0058 macro F1 (Table 5).
- Healthy-to-impaired transfer helps the tuned CNN-Base (0.465→0.585 subset accuracy) but not the LSTM, whose ranking metrics degrade (Table 4).
- The resulting 123K-parameter four-channel model, exported to ONNX, establishes a software interface contract for Raspberry Pi 5 integration — explicitly not a real-time performance claim.
Strengths
Honest scoping of claims
The paper explicitly labels Tables 2–4 as single deterministic-split estimates that 'do not support statistical-significance claims' (§4.3), states the ONNX export establishes 'interface compatibility, not real-time hardware performance' (§6), and lists specific limitations including the untested electrode maps and missing cross-validation.
Unusually complete method specification
§3 documents the frozen preprocessing config (4th-order zero-phase Butterworth 20–450 Hz, Symlet-4 four-level soft thresholding), twelve named features, exact channel indices ([14,15,8,0] zero-based), distillation hyperparameters (T=2, α=0.5), and patient counts (34/5/9; 2656/432/720 samples), plus a cited public code repo (ref [6]).
Multi-seed evaluation for the deployment-critical experiment
Clean training-mode ablation with negative results reported
Hardware-aware model selection discipline
Weaknesses
Thin statistical support
No classical ML baseline
Single-seed 64-channel reference
'Intent' framing exceeds the label source
No timing evidence despite latency-driven design
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper frames five-finger multilabel intent decoding from impaired-arm high-density sEMG as a hardware-aware machine-learning problem, using PhysioMio (48 stroke patients, 64 electrodes, 16 gestures, bilateral longitudinal; ref [10]). Three questions drive the design: (1) which inductive bias — LSTM, CNN, or GNN — best suits multilabel finger prediction on a shared feature representation; (2) how much do CNN architecture search, ResNet teachers, and healthy-to-impaired transfer add; (3) how much performance survives reduction to four hardware-feasible channels (ECRB, ECRL, FDS, FDP), and can cross-channel distillation recover the loss. Headline claims: model-family complementarity (LSTM best subset accuracy 0.545; GNN best macro F1 0.706 / AUPRC 0.776, Table 2), CNN-Large as strongest single-split CNN (0.593/0.714, Table 3), and a distilled four-channel CNN-Micro at 0.5219±0.0114 subset accuracy and 0.6095±0.0058 macro F1 over five seeds (Table 5), exported as a 123K-parameter ONNX graph. One framing caveat: labels derive from annotations of executed gestures (§3.2), so "intent" here means movement-labeled activation, not decoding of attempted-but-unachieved movement.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is the finger-intent decoder for stroke patients?
On held-out stroke patients, the best 64-channel model (CNN-Large, single split) reaches 0.593 subset accuracy — meaning all five finger labels are simultaneously correct — and 0.714 macro F1. The deployable four-channel distilled model reaches 0.522 subset accuracy and 0.6095 macro F1 averaged over five random seeds.
What dataset does the paper use?
PhysioMio: a bilateral, longitudinal high-density sEMG dataset from 48 stroke patients with 64 electrodes and 16 hand gestures, recorded repeatedly during inpatient recovery, published in Scientific Data (2026). The paper uses the impaired arm with patient-level train/validation/test splits.
Is the code or the trained model available?
The paper cites a public GitHub repository (post-stroke-rehab/psr-pipeline) as the project software implementation, and the selected 123K-parameter four-channel model was exported to an ONNX file, which is a standard exchange format for running models on embedded hardware.
Does the model run in real time on a Raspberry Pi 5?
No. The paper explicitly reports no on-device timing; the ONNX export establishes interface compatibility only, and measuring Raspberry Pi 5 latency — including the zero-phase filtering and wavelet denoising overhead — is left to future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure graphics (Figs. 1–4) cannot be visually assessed from text extraction; all quantitative claims in this review were verified against the extracted tables and body text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.