Source-linked AI summary
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
TL;DR
Contaminants distort sEMG, while neural SQA methods can lack interpretability and application-specific adaptability. Prism-SQA separates and verifies contaminant components, then computes customizable quality indices. It achieves competitive or superior estimation accuracy relative to a strong black-box baseline across evaluations on Ninapro and clinical SQI data.
Problem
Contaminants distort sEMG, and neural SQA methods can produce opaque scores and fixed quality indices that are difficult to adapt to new applications.
Source: §I · pp. 1–2Method
Prism-SQA separates sEMG into clean and contaminant-specific components, verifies contaminant fingerprints, and computes quality indices from the verified components.
Source: §I · p. 2Results
Experiments on three Ninapro subsets and a clinical SQI dataset demonstrate competitive or superior estimation accuracy relative to a strong black-box baseline.
Source: §VI · p. 13Takeaways & Limitations
The framework provides signal-level interpretability and allows quality criteria to be customized without retraining.
Source: §I · p. 2The authors acknowledge that the study has limitations but do not specify them in the cited passage.
Source: D. Limitations and future research · p. 12
Abstract
from arXiv · showhide
sEMG is vulnerable to various contaminants that distort signal morphology and spectral content. Accurate signal quality assessment (SQA) is essential for identifying such degradation and ensuring reliable clinical analyses and decisions. Recent neural network-based SQA methods achieve accurate quality estimation by learning complex contamination patterns, yet their black-box nature prevents clinicians from understanding or validating the reported scores and limits adaptability to application-specific quality definitions without retraining. To address these limitations, we propose Prism-SQA, an interpretable and adaptable neural framework that reformulates SQA as a physiology-aware source-separation and verification process. Prism-SQA decomposes each input signal into a clean sEMG component and five contaminant-specific components using a U-Net with bidirectional long short-term memory. Each separated contaminant component is examined by a Contaminant Fingerprint Verifier, which enforces physiological plausibility by comparing its temporal and spectral structure with canonical contaminant signatures. This design allows clinicians to inspect how each contaminant affects signal quality, grounding the assessment in transparent, signal-level evidence rather than opaque latent representations. Quality indices computed from the verified components further enable customization of quality criteria across clinical contexts without retraining. We evaluate Prism-SQA on continuous quality-score estimation using synthesized noisy sEMG from public Ninapro datasets and on binary quality classification using a clinical dysphagia dataset. Results show that Prism-SQA achieves competitive or better performance than contemporary black-box neural methods while providing explicit interpretability and adaptability, advancing toward practical and clinically aligned sEMG SQA.
I. INTRODUCTION
Contaminants can distort sEMG and undermine downstream interpretation, while existing neural SQA methods offer limited transparency and adaptation to application-specific quality definitions. Prism-SQA instead separates and verifies contaminant-specific signals, then computes customizable quality indices, and is evaluated on synthesized and clinical data.
I. INTRODUCTION
- Contaminants distort sEMG morphology and spectral content, motivating quality assessment to support reliable interpretation and downstream analysis.Quality scores can inform whether recordings are acceptable for analysis or need denoising or reacquisition.Source: §I · p. 1
- Handcrafted frequency rules can fail on nonstationary, compound, or low-power contamination, while neural methods remain difficult to interpret and adapt without retraining.The stated limitations include opaque quality scores and fixed indices that cannot readily meet new application-specific requirements.Source: §I · p. 2
- Prism-SQA decomposes input into an sEMG component and five contaminant-specific components, then verifies contaminant fingerprints for physiological plausibility.The proposed separation model combines a U-Net with bidirectional LSTM; implausible components are discarded.Source: §I · p. 2
- Quality indices computed from verified components are decoupled from model training, allowing task-specific customization without retraining.The indices can use suitable statistical measures or signal components to represent contamination affecting preprocessing or application performance.Source: §I · p. 2
- Evaluation covers synthesized noisy sEMG from three Ninapro subsets and quality classification on a clinical dysphagia dataset.The experiments include controlled ground-truth quality scores, zero-shot cross-sub-database evaluation, and external clinical validation.Source: §I · pp. 2–3
A. Datasets and preprocessing
The study uses Ninapro recordings as clean sEMG sources for synthesized noisy signals and a clinical dysphagia SQI dataset for quality classification. It describes the datasets, preprocessing, and five modeled contaminant types.
1) sEMG database:
- Ninapro DB2, DB3, and DB4 provide clean recordings from intact participants and transradial amputees, while the SQI dataset contains expert-labeled good- and poor-quality dysphagia recordings.The SQI dataset has three subsets; purely silent poor-quality segments were excluded from evaluation.Source: 1) sEMG database · p. 3
A. Datasets and preprocessing
- Ninapro signals were bandpass-filtered from 20 to 500 Hz, downsampled from 2 kHz to 1 kHz, segmented into two-second windows, and stripped of silent segments.The SQI subsets were resampled to 1 kHz for compatibility with models trained on Ninapro data.Source: 1) sEMG database · p. 3
2) sEMG contaminants dataset:
- The synthesized data include low-frequency noise, powerline interference, ECG interference, spike-like artifacts, and high-frequency noise.LFN spans baseline wander, electrode motion, motion artifacts, a step function, and low-pass-filtered noise; PLI includes 50/60 Hz tones, drift, and harmonics.Source: 2) sEMG contaminants dataset · pp. 3–4
- Spike-like artifacts include instantaneous pulses and simulated motor unit action potentials, while HFN includes white and bandpass-filtered noise.Compound contaminants combine two to five noise components with equal signal power per component.Source: 2) sEMG contaminants dataset · p. 4
- Contaminants were scaled and added to clean sEMG at specified SNRs to synthesize noisy signals.The resulting signals support evaluation against known quality targets.Source: 3) Noisy sEMG dataset · p. 4
3) Noisy sEMG dataset:
The study synthesizes noisy sEMG by mixing clean signals with contaminants at controlled SNRs and builds training targets from separated sources. Testing uses mismatched Ninapro conditions and a clinical SQI dataset.
3) Noisy sEMG dataset:
- The study caps synthesized-signal SNR at 30 dB, a level described as often indicating a clean signal.The passage defines clean-signal and contaminant powers and their waveforms in the SNR calculation.Source: 3) Noisy sEMG dataset · p. 4
B. Data preparation
- Training and validation use 38,787 and 10,221 clean segments from Ninapro DB2, respectively.Segments came from Channels 1–8 during Exercise B, with 25 subjects for training and five for validation.Source: B. Data preparation · p. 4
- Dynamic mixing samples zero to five contaminant types and SNRs from −20 to 30 dB for each training segment.The selected contaminants are randomly drawn from PLI, LFN, ECG, SPK, and HFN.Source: B. Data preparation · p. 4
- Ground-truth separation targets include clean and contaminant waveforms, and six quality targets comprise overall SNR plus five signal-to-contaminant ratios.Absent contaminants receive zero-vector targets, and each noisy input is normalized with the same scale applied to its targets.Source: B. Data preparation · p. 4
- Ninapro test sets vary subjects, exercises, channels, and, for some subsets, participant condition or recording device; the clinical SQI dataset tests quality classification.The synthesized test sets use five SNR levels and one to five contaminant types, with clean segments also included.Source: B. Data preparation · pp. 4–5
III. METHODOLOGY
Prism-SQA separates each noisy sEMG into clean and contaminant-specific components, then verifies contaminant plausibility before quality estimation. Its architecture and training objectives are designed to preserve component structure, amplitude relationships, and contaminant presence.
1) Physiology-aware source separation:
- A U-Net with BiLSTM layers decomposes noisy sEMG into six outputs: one clean signal and five contaminant-specific components.The outputs correspond to sEMG, PLI, LFN, ECG, SPK, and HFN.Source: 1) Physiology-aware source separation · p. 5
- The training objective combines L1, scale-dependent SD-SDR, and component-presence classification losses to support physiologically meaningful decomposition.SD-SDR penalizes amplitude-scaling errors, while the presence loss encourages near-zero outputs for absent contaminants.Source: 1) Physiology-aware source separation · p. 6
- The Contaminant Fingerprint Verifier checks separated contaminants against electrophysiological fingerprints before quality estimation.Table III summarizes the fingerprints and verification thresholds used.Source: 2) Contaminant fingerprint verification · p. 6, Table III · p. 6
2) Contaminant fingerprint verification:
CFV verifies separated contaminants by matching their spectral or temporal fingerprints to contaminant-specific criteria. It is activated for preliminary SNR below 0 dB, and quality indices are computed from the validated components.
2) Contaminant fingerprint verification:
- CFV uses frequency-band power ratios for PLI, LFN, ECG, and SPK, and waveform kurtosis for HFN.The bands are 0–30 Hz for LFN, 50/60 Hz and harmonics (±2 Hz) for PLI, 5–150 Hz for ECG, and 200–500 Hz for SPK.Source: 2) Contaminant fingerprint verification · p. 6
- Components pass through a threshold gate or are zeroed; CFV activates only when preliminary SNR is below 0 dB.The thresholds were selected to avoid near-universal passage or rejection, and sensitivity analysis found stable performance across a broad range.Source: 2) Contaminant fingerprint verification · p. 6
- Six quality indices are extracted from CFV-validated components and can be redefined or extended without retraining the neural network.The indices include overall SNR and five contaminant-specific SNRs.Source: 3) Quality index extraction and… · p. 7
B. Implementation details
The implementation uses early stopping and scheduled learning-rate reductions, while evaluation uses metrics suited to continuous quality estimation and binary classification. Comparisons include a capacity-matched neural baseline and a feature-based baseline, with subject-aware statistical procedures.
B. Implementation details
- Prism-SQA and QASE-Net training and validation losses show stable convergence under the reported implementation and training procedure.Training used Adam, early stopping, and learning-rate reductions after 20, 200, and 400 epochs.Source: B. Implementation details · p. 7
1) Evaluation metrics:
- Continuous quality estimation is assessed with LCC, SRCC, and MAE, while binary classification is assessed with AUC.Higher LCC and SRCC and lower MAE indicate better estimation; AUC measures threshold-independent discrimination.Source: 1) Evaluation metrics · p. 7
- QASE-Net is capacity-matched as the quality-estimation baseline, and variance-based SQA is added for the classification experiment.Subject-level paired t-tests with adjusted p-values are used for first-experiment comparisons; second-experiment AUC confidence intervals use a subject-cluster bootstrap.Source: 2) Baseline SQA approach · p. 7, 1) Evaluation metrics · p. 7
2) Baseline SQA approach:
Across Ninapro DB2–DB4, Prism-SQA generally matched or outperformed QASE-Net in quality-score correlations and MAE across the evaluated scenarios.
A. Quality score estimation
- LCCs exceeded 0.97 and SRCCs exceeded 0.90 across six indices on DB2–DB4; Prism-SQA significantly led in most overall comparisons.The figure compares overall and contaminant-included scenarios by dataset and correlation metric.Source: Figure 3 · p. 8, A. Quality score estimation · p. 9
- Under contaminant-included conditions, average LCC/SRCC exceeded 0.96; Prism-SQA led most comparisons, while QASE-Net significantly won only SHR on DB4 (∆LCC = −0.009; ∆SRCC = −0.011).Other comparisons either favored Prism-SQA or showed no significant difference.Source: Figure 3 · p. 8, A. Quality score estimation · p. 9
- Overall, Prism-SQA's MAE was significantly lower except for SHR on DB4; it led 14/18 contaminant-included and 9/18 contaminant-excluded comparisons.QASE-Net had significantly lower MAE in one contaminant-excluded case, SSR on DB2.Source: Table IV · p. 8, A. Quality score estimation · pp. 9–10
- The three scenarios cover all samples, samples containing the relevant contaminant, and samples without it; absent-contaminant targets were constant at 30 dB.The contaminant-excluded scenario assesses whether the model avoids spurious estimates when the relevant contaminant is absent.Source: A. Quality score estimation · p. 9
B. Quality classification
For clinical quality classification, Prism-SQA uses estimated SNR as a decision variable and explores an adapted index that excludes LFN. Its zero-shot AUCs were numerically higher than QASE-Net's on two datasets, with overlapping confidence intervals.
B. Quality classification
- Adapted SNR omits LFN from the noise-power term because SQI-DB1/DB3 annotations can label LFN-containing segments as good quality.The adaptation is exploratory and no adapted result is reported for SQI-DB2.Source: B. Quality classification · p. 10
- AUCs for Prism-SQA were 0.963 (0.924–0.989), 1.000 (1.000–1.000), and 0.835 (0.729–1.000) on SQI-DB1/2/3; QASE-Net matched DB2.The confidence intervals overlap, so the comparison does not establish a statistically supported difference between the two neural methods.Source: B. Quality classification · p. 10
- The conventional feature-based SQA method achieved AUCs of 0.952 (0.910–0.980), 0.792 (0.611–0.938)3, and 0.595 (0.000–0.757) on SQI-DB1, SQI-DB2, and SQI-DB3, respectively.It performed competitively on SQI-DB1, while performance dropped substantially on SQI-DB2 and SQI-DB3.Source: B. Quality classification · p. 10
- Adapted Prism-SQA yielded AUCs of 0.977 (0.949–0.993) on SQI-DB1 and 0.907 (0.829–1.000) on SQI-DB3, numerically above its unadapted results.The LFN exclusion was motivated by those datasets' annotation characteristics.Source: B. Quality classification · p. 10
C. Ablation study
Ablations compare objective-function components and the CFV module on Ninapro DB2–DB4; a separate analysis tests sensitivity to CFV thresholds on DB2.
1) Ablation study on the objective function:
- Versus L1 alone, the full loss improved overall ∆LCC = 0.007, ∆SRCC = 0.022, and ∆MAE = −0.757 dB.Adding LSD-SDR to L1 yielded LCC = 0.986, SRCC = 0.961, and MAE = 1.161 dB.Source: 1) Ablation study on the objective… · p. 10
- Adding Lcls to L1 + LSD-SDR improved overall LCC by ∆LCC = 0.002, SRCC by ∆SRCC = 0.005, and MAE by ∆MAE = −0.102 dB.The MAE reduction was primarily in contaminant-excluded cases (∆MAE = −0.447 dB), whereas included-case MAE increased (∆MAE = 0.110 dB).Source: 1) Ablation study on the objective… · p. 10
2) Effect of the CFV module:
- CFV improved overall ∆LCC = 0.0021, ∆SRCC = 0.0043, and ∆MAE = −0.064 dB; included-condition differences were numerically small.The largest MAE reduction was in the contaminant-excluded condition (∆MAE = −0.155 dB).Source: 2) Effect of the CFV module · pp. 10–11
D. Sensitivity analysis of CFV threshold values
- Performance remained stable across thresholds of approximately 0.1–0.6 for spectral-power-ratio thresholds and 0.4–1.0 for the kurtosis threshold.Thresholds were varied independently on DB2; adopted values fell within these stable ranges, though they were not always optimal.Source: Figure 6 · p. 11, D. Sensitivity analysis of CFV… · p. 11
A. Quality estimation performance of Prism-SQA
Prism-SQA combines physiology-aware decomposition and verification with signal-level quality indices. It reports strong estimation alongside interpretable outputs and quality-index customization, while acknowledging limitations.
A. Quality estimation performance of Prism-SQA
- Across synthetic and clinical evaluations, Prism-SQA provides stable, meaningful quality estimates and strong performance while preserving interpretability and adaptability.The authors describe its clinical SQI results as preliminary evidence of potential generalization beyond controlled synthetic scenarios.Source: A. Quality estimation performance of… · p. 11
B. Interpretability of Prism-SQA
- Separated components provide a waveform-level rationale: PLI, ECG, and HFN match contaminant patterns, while absent LFN and SPK remain near zero.Fig. 7 contrasts these inspectable components with black-box scores that lack interpretable representations.Source: Figure 7 · p. 11, B. Interpretability of Prism-SQA · p. 12
- After CFV rejects a spurious PLI component, the quality-score error falls from 16.07 dB to 0 dB.Without CFV, the SPR is 13.93 dB against a 30 dB ground-truth score; CFV rejects the component because it lacks narrowband peaks at 50/60 Hz and harmonics.Source: B. Interpretability of Prism-SQA · p. 12, Figure 8 · p. 12
C. Adaptability of Prism-SQA
- Users can redefine quality indices without retraining; exploratory SQI results show higher AUCs on SQI-DB1 and SQI-DB3 and closer alignment with expert annotations.The framework can compute alternative index definitions from separated components, unlike fixed-index black-box models that require retraining.Source: C. Adaptability of Prism-SQA · p. 12
D. Limitations and future research
- The authors acknowledge limitations to the framework.Source: D. Limitations and future research · p. 12
1) Generalization to unseen contaminant types and characteristics:
Prism-SQA’s generalization is limited for out-of-distribution contaminants, its verifier remains a proof of concept, and its computational cost constrains real-time use. The conclusion reports competitive or superior accuracy relative to a strong black-box baseline.
1) Generalization to unseen contaminant types and characteristics:
- The supervised separator may mishandle untrained contaminant characteristics: atypical 209.5 Hz and 419 Hz PLI remained partly in reconstructed sEMG and leaked into HFN.The authors propose broader real-world contaminant training and hybrid separation with signal-processing priors as future directions.Source: 1) Generalization to unseen… · pp. 12–13, Figure 9 · p. 13
2) Simplicity of CFV module:
- CFV is a proof-of-concept verifier limited by its simple spectral or temporal fingerprints; richer contaminant properties and filtering could refine verification.Source: 2) Simplicity of CFV module · p. 13
3) High computational complexity for real-time or wearable usage:
- Prism-SQA is feasible for offline analysis on high-performance hardware but runs approximately five to six times slower than QASE-Net.The authors identify efficiency improvements as necessary for real-time or embedded applications.Source: 3) High computational complexity for… · p. 13
VI. CONCLUSION
- Across three Ninapro subsets and a clinical SQI dataset, Prism-SQA achieves competitive or superior estimation accuracy relative to a strong black-box baseline.The conclusion describes its signal-level interpretability and quality-index adaptability alongside this performance.Source: §VI · p. 13