Prism-SQA: An Interpretable and Adaptable Neural Framework for Surface Electromyography Quality Assessment
Kuan-Chen Wang, Kai-Chun Liu, Ping-Cheng Yeh, Sheng-Yu Peng, Yu Tsao
eess.SP · cs.LG
Quick verdict
Proposes decompose-then-verify sEMG quality scoring; rigorous stats, honest limits, but evidence mainly synthetic with a single baseline; clinical gain modest.
Summary
Prism-SQA reformulates surface EMG signal quality assessment (SQA) as physiology-aware source separation plus verification. A U-Net with a BiLSTM bottleneck decomposes each 2-s waveform into a clean sEMG component and five contaminant-specific components (PLI, LFN, ECG, SPK, HFN); a rule-based Contaminant Fingerprint Verifier (CFV) discards components failing spectral/temporal fingerprints (e.g., narrowband peaks for PLI, kurtosis for HFN), and quality indices (overall SNR plus five contaminant-specific ratios) are computed analytically from verified components, decoupling index definition from training. Trained on synthesized noisy Ninapro DB2 data, it is evaluated zero-shot on DB2/DB3 (amputees)/DB4 (different electrodes) and on the clinical dysphagia SQI dataset. Against a capacity-matched QASE-Net (19.05M vs 16.53M parameters), Prism-SQA achieves overall LCC > 0.97 and SRCC > 0.90 with significantly lower MAE in most subject-level paired comparisons (BH-FDR controlled); the notable exception is the high-frequency-noise index (SHR) on DB4, where QASE-Net is significantly better. Clinically it attains AUCs of 0.963/1.000/0.835 on SQI-DB1/2/3 versus QASE-Net's 0.938/1.000/0.815, with overlapping confidence intervals; an exploratory adapted SNR excluding LFN raises AUC to 0.977/0.907. Main limitations: ground truth exists only for synthetic contamination, a single NN baseline, CFV gating applies only below 0 dB SNR, out-of-distribution contaminants (e.g., 209.5 Hz PLI) are mishandled, and inference is 5-6x slower than QASE-Net. Code is publicly released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
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
- Reformulating sEMG SQA as physiology-aware source separation plus fingerprint verification yields accuracy competitive with or better than black-box neural regression while exposing signal-level evidence.
- Decoupling quality-index computation from network training enables application-specific index redefinition without retraining, demonstrated by an adapted SNR improving clinical AUCs.
- The CFV module significantly improves accuracy, chiefly by suppressing spurious outputs for absent contaminants (contaminant-excluded ΔMAE = −0.155 dB).
- Prism-SQA generalizes zero-shot across subjects, exercises, channels, an amputee population, and a different electrode system, while remaining slower (5-6x) than the baseline.
Strengths
Unusually rigorous statistical methodology
Subject-level paired t-tests with effect sizes and 95% CIs, Benjamini-Hochberg FDR control at 0.05 across three pre-specified families (126/14/7 comparisons), and subject-cluster bootstrap (10,000 resamples) for AUC CIs (Sec. III-C); even the significant loss to QASE-Net (SHR on DB4, ΔMAE = 0.377 dB) is reported plainly.
Design choices are each empirically validated
Loss ablation (Table V: removing SD-SDR raises overall MAE 1.059 → 1.816 dB) and CFV ablation (Table VI: ΔMAE −0.155 dB contaminant-excluded) isolate each component's contribution; a threshold sensitivity analysis (Fig. 6) shows stable performance across 0.1–0.6 (power ratios) and 0.4–1.0 (kurtosis).
Genuine cross-domain zero-shot evaluation
Adaptability claim is demonstrated concretely
Honest out-of-distribution failure analysis
Weaknesses
Ground truth exists only for synthesized contamination
Single NN baseline, modified by the authors themselves
Post hoc quality-index adaptation on the evaluation data
CFV activation gate is an un-ablated heuristic
Compute cost undercuts the clinical/wearable motivation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Prism-SQA targets surface EMG signal quality assessment (SQA), arguing that black-box neural regressors (e.g., QASE-Net) cannot be clinically audited or retuned. Its remedy is a three-stage framework: (1) a U-Net with a six-layer BiLSTM bottleneck separates each 2-s waveform into clean sEMG plus five contaminant components (PLI, LFN, ECG, SPK, HFN); (2) a rule-based Contaminant Fingerprint Verifier (CFV) discards separated components failing physiological fingerprints (band-power ratios, kurtosis); (3) an overall SNR and five contaminant-specific ratios are computed analytically from verified components. Central claims: competitive-or-better accuracy versus black-box peers, signal-level interpretability, retraining-free index adaptability, and zero-shot generalization across subjects, populations, and devices. All four receive at least partial empirical support.
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
What is Prism-SQA and what problem does it solve?
Prism-SQA is a neural framework for assessing the quality of surface electromyography (sEMG) recordings. Instead of a black-box network directly outputting a quality score, it separates the signal into a clean sEMG component and five contaminant components (powerline interference, low-frequency noise, ECG, spikes, high-frequency noise), verifies each against physiological fingerprints, and computes quality indices analytically — so users can see which contaminant degraded the signal and can redefine quality criteria without retraining.
How accurate is Prism-SQA compared to previous methods?
On synthesized noisy sEMG from Ninapro DB2/DB3/DB4 it reaches overall correlations above 0.97 (LCC) and 0.90 (SRCC) and lower error than the capacity-matched QASE-Net baseline in most subject-level tests; the main exception is the high-frequency-noise index on DB4, where QASE-Net is significantly better. On the clinical dysphagia SQI dataset it reaches AUCs of 0.963/1.000/0.835 (DB1/DB2/DB3), numerically above QASE-Net's 0.938/1.000/0.815 but with overlapping confidence intervals.
What data was Prism-SQA trained on, and is the code public?
The separation model was trained on Ninapro DB2 sEMG (channels 1–8, Exercise B, 25 subjects) mixed on-the-fly with contaminants from MIT-BIH databases (baseline wander, electrode motion, ECG) and simulated powerline, spike, and high-frequency noise. Yes — the code is released on GitHub (eric-wang135/Prism-SQA), and all datasets are publicly available.
What are the main limitations of Prism-SQA?
Ground-truth quality scores exist only for synthetic contamination; the clinical evaluation is binary classification only; it fails on out-of-distribution contaminants (e.g., powerline interference at 209.5 Hz); the fingerprint verifier activates only below 0 dB SNR; and inference is about 5-6x slower than QASE-Net (37.66 vs 6.33 ms per 2-s segment on GPU).
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.
- Figures 1–9 are available only as captions and in-text descriptions; their visual content (waveforms, ROC curves, example decompositions) could not be inspected from the text extraction, so qualitative interpretability claims (Figs. 7–9) rest on the authors' descriptions.
- Table values were extracted inline from the PDF and could contain OCR digit errors; spot-checks against the in-text statistics (e.g., the reported ΔMAE and CI values) were consistent.
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.