A Conditional GAN-based Framework for Sparse sEMG Data Augmentation with Muscle Synergy Prior Constraints.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• High-quality surface electromyography (sEMG) data is scarce due to ethical and privacy concerns, hindering deep learning model development.
• Existing generative models struggle with inter-channel correlations and physiological plausibility in multi-channel sEMG signals.
• Error accumulation and lack of biological fidelity limit clinical applications of current sEMG generation techniques.

Purpose of the Study:

• To develop a novel framework for generating physiologically plausible multi-channel sEMG signals.
• To address limitations in existing generative methods, including error accumulation and inadequate modeling of inter-channel relationships.
• To enhance the biological fidelity and clinical utility of synthetic sEMG data.

Main Methods:

• Proposed a Muscle Synergy-Constrained Conditional Generative Adversarial Network (MS-cGAN) framework.
• Introduced a Graph Convolutional Network (GCN)-based generator to model inter-channel relationships in sparse sEMG signals.
• Integrated Muscle Synergy (MS) prior constraints as dynamic loss functions to ensure physiological plausibility.

Main Results:

• MS-cGAN successfully generates multi-channel sEMG signals with improved authenticity and bio-mechanical fidelity.
• The GCN-based generator effectively captures complex inter-channel correlations, mitigating error accumulation.
• Experiments demonstrated superior performance of MS-cGAN over traditional GANs and diffusion models on sEMG datasets.
• Generated data significantly enhanced downstream task performance and kinematic prediction precision.

Conclusions:

• The MS-cGAN framework offers a robust solution for generating high-fidelity synthetic sEMG data.
• This approach effectively supplements scarce sEMG datasets, enabling more reliable deep learning model training.
• The method ensures physiological consistency, making generated sEMG signals suitable for clinical applications and improving predictive accuracy.