SSEMG-Net: A Spectrogram-Based Mamba Network for Surface Electromyography Denoising

I. INTRODUCTION

• ECG (about 0–150 Hz) and sEMG (about 10–500 Hz) bands overlap; high-pass filters remove low-frequency ECG but eliminate important sEMG components, while template subtraction assumes quasi-periodic ECG and zero-mean Gaussian sEMG, rarely satisfied in practice.Source: §I · p. 1
• Waveform-based neural denoising with pointwise losses can have difficulty suppressing low-frequency ECG while retaining high-frequency sEMG and explicitly constraining spectral structure.This limitation matters for frequency-dependent applications such as muscle fatigue assessment.Source: §I · p. 1
• SSEMG-Net reconstructs magnitude and phase in the time–frequency domain and restricts magnitude masking to 0–150 Hz, where ECG contamination is prominent.It also uses composite time- and frequency-domain objectives, including multi-resolution STFT loss.Source: §I · p. 1
• SSEMG-Net is reported to outperform prior methods across SNR levels and preserve spectral features; the authors describe it as the first neural sEMG enhancement framework to denoise explicitly in the time–frequency domain.The reported outcome is presented as robust denoising and superior spectral-feature preservation.Source: §I · pp. 1–2

A. ECG interference removal methods

• High-pass filtering also removes low-frequency sEMG, while template subtraction depends on quasi-periodic ECG morphology and zero-mean Gaussian sEMG assumptions that often fail in real recordings.FCN, SDEMG, and MSEMG are described as waveform-based alternatives that still share pointwise objectives and can be limited in preserving fine-grained spectral characteristics.Source: A. ECG interference removal methods · p. 2

B. Spectrogram-based denoising neural networks

• Spectrogram-based denoising methods use time–frequency patterns; MECG-E applies Mamba and composite temporal and spectral losses to ECG denoising, motivating SSEMG-Net for sEMG.MECG-E is described as achieving state-of-the-art ECG denoising and efficient inference relative to diffusion-based approaches.Source: B. Spectrogram-based denoising neural… · p. 2

III. METHODOLOGY

• SSEMG-Net converts a noisy sEMG segment into a complex time–frequency representation and applies power-law magnitude compression while preserving phase.The representation has time frames and frequency bins; the magnitude compression exponent is c.Source: §III · p. 2

1) Encoder:

• The encoder uses convolutional blocks around a dilated DenseNet to enlarge temporal receptive field and preserve temporal–spectral details, then downsamples frequency to reduce computational cost.The resulting representation has extended temporal context and compact frequency resolution.Source: 1) Encoder · p. 2

2) Time–Frequency bidirectional mamba:

• TF-Bi-Mamba combines temporal bidirectional Mamba with frequency bidirectional Mamba to model temporal dynamics and cross-frequency relations.Forward and reversed sequences are fused through residual and transposed-convolution layers, capturing past–future context and long-range dynamics alongside localized spectral structure.Source: 2) Time–Frequency bidirectional mamba · p. 2

3) Decoder:

• The sub-band magnitude-mask decoder generates a mask with the magnitude spectrogram’s dimensions, while a separate phase decoder estimates wrapped phase.The architecture is included in Fig. 1(a).Source: 3) Decoder · pp. 2–3, Figure 1 · p. 3
• The magnitude mask is fixed to one above cutoff frequency fcut and learned below it, restricting denoising to the sub-band where ECG overlaps and dominates sEMG.The mask uses a learnable sigmoid function on its pre-activation.Source: 3) Decoder · pp. 2–3
• The phase decoder estimates wrapped phase from pseudo-real and pseudo-imaginary outputs using a two-argument arctangent, then combines it with enhanced magnitude to form the final complex spectrogram.The enhanced magnitude is restored by the inverse power-law before combination.Source: 3) Decoder · p. 3

• The objective combines waveform fidelity, complex-spectrogram reconstruction, STFT/inverse-STFT consistency, and multi-resolution STFT supervision.Source: B. Loss function · p. 3
• The notation distinguishes clean and enhanced waveforms and complex spectra, with the STFT defined by analysis parameters including FFT size, hop, and window size.Source: B. Loss function · p. 3
• Multi-resolution STFT loss supervises the model across analysis resolutions, reducing dependence on one fixed representation that can yield suboptimal waveform reconstruction.Source: B. Loss function · p. 3

A. Dataset preprocessing and preparation

• NINAPro DB2 provides 12-channel recordings from 40 subjects across three sessions; sEMG is bandpass-filtered, downsampled to 1 kHz, normalized, and segmented into 10-second samples.Source: A. Dataset preprocessing and preparation · pp. 3–4
• ECG interference is drawn from MIT-BIH NSRD Channel 1 recordings, filtered to 10–200 Hz, and mixed into sEMG at multiple SNR levels.Source: A. Dataset preprocessing and preparation · p. 4
• Training and validation use NINAPro DB2 sEMG from 30 subjects, while testing uses different subjects, movements, channels, and SNRs.Test sEMG comes from Channels 9–12 of Exercise 2 and 10 subjects; ECG comes from three remaining NSRD subjects.Source: A. Dataset preprocessing and preparation · p. 4

B. Implementation details

• The magnitude mask uses a 150 Hz cutoff and deterministic pass-through above it; MR-STFT loss uses three STFT parameter sets.Training ran for 30 epochs with AdamW; the encoder and decoder settings and loss weights are specified in the implementation details.Source: B. Implementation details · p. 4

C. Evaluation metrics

• The evaluation reports SNR improvement and waveform RMSE for reconstruction, plus ARV and mean-frequency RMSE for feature extraction.Higher SNR improvement and lower errors indicate more effective artifact suppression and better preservation of sEMG characteristics.Source: C. Evaluation metrics · p. 4

A. Overall comparison with prior methods

• SSEMG-Net achieves the strongest performance across all reported metrics against representative single-channel baselines and recent neural methods.Source: Table I · p. 4, A. Overall comparison with prior methods · p. 4
• Feature-level errors decrease consistently, especially RMSEMF; mean frequency is used as a frequency-domain biomarker in fatigue-related analysis.Source: A. Overall comparison with prior methods · p. 4
• SSEMG-Net outperforms competing approaches across a wide range of input SNRs, spanning severe and mild contamination.Spectrogram comparisons also show more faithful time-frequency structures, less residual low-frequency interference, and no visible broadband spectral warping.Source: A. Overall comparison with prior methods · p. 4

B. Ablation studies

• In the highlighted spectrogram region, SSEMG-Net reconstructs the clean reference more closely than MSEMG.The blue rectangle marks the comparison region.Source: Figure 3 · p. 5
• Each Table II row changes one system element while keeping the data construction, training schedule, and evaluation pipeline fixed.This setup is intended to attribute observed differences to the targeted factor rather than confounded training effects.Source: B. Ablation studies · p. 5
• Removing explicit phase estimation causes the largest degradation, supporting it as a necessary component for feature-faithful denoising.Phase errors can distort the reconstructed waveform and redistribute spectral energy, contributing to mean-frequency drift and larger feature error.Source: Table II · p. 5, B. Ablation studies · p. 5
• Removing MR-STFT loss reduces performance across metrics, indicating that pointwise or single-resolution supervision alone does not sufficiently constrain spectral structure.Multi-resolution STFT supervision encourages agreement across analysis scales and reduces resolution-specific fitting that can bias frequency-domain features.Source: Table II · p. 5, B. Ablation studies · p. 5
• The sub-band mask restricts learnable suppression to the low-frequency region where ECG interference dominates, protecting mid- and high-frequency sEMG content.The unrestricted variant can slightly improve pointwise reconstruction metrics but may attenuate or reshape sEMG-dominant frequencies, compromising mean-frequency fidelity.Source: B. Ablation studies · p. 5, Table II · p. 5

VI. CONCLUSION

• SSEMG-Net combines spectrogram-based denoising with physiologically aware sub-band masking, a bidirectional Mamba backbone, and dual-head decoding.The model targets low-frequency ECG artifacts while preserving mid- and high-frequency sEMG content.Source: §VI · p. 5
• SSEMG-Net consistently improves signal quality and feature preservation over existing waveform-based denoising approaches.The reported findings support spectrogram-domain modeling combined with physiological priors as a framework for sEMG enhancement.Source: §VI · p. 5
• Future work will explore extending SSEMG-Net to other sEMG contaminants and adapting it for streaming inference to enable real-time applications.Source: §VI · p. 5