Adaptive EMG Pattern Classification via Probabilistic Knowledge Transfer With Scale Mixture-Based Bayesian Sequential...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyogram (EMG) signals are crucial for controlling devices like myoelectric prostheses.
• Temporal variations in EMG signals due to factors like muscle fatigue and electrode shift degrade classification accuracy over time.
• Existing EMG interfaces struggle with continuous adaptation to these signal changes.

Purpose of the Study:

• To develop an adaptive method for robust EMG signal classification.
• To address the challenge of performance degradation in EMG-based interfaces.
• To improve the reliability and accuracy of human-device interaction using EMG signals.

Main Methods:

• Integration of a scale mixture classification model (SMCM) with Bayesian sequential self-training (BSST).
• Sequential updating of model parameters using Bayesian updates and pseudo-labels based on prediction confidence.
• Utilizing SMCM for variance uncertainty modeling to represent EMG signal distributions and enhance confidence estimation.

Main Results:

• The proposed SMCM-BSST method demonstrated superior classification accuracy compared to conventional methods.
• The method effectively mitigated accuracy degradation over short-term and long-term (30 days) datasets.
• Improved reliability in prediction confidence estimation was observed.

Conclusions:

• The combination of SMCM and BSST offers effective adaptation to EMG signal variations.
• This approach provides a practical solution for developing reliable and continuously performing EMG-based interfaces.
• The study highlights the potential for advanced machine learning techniques to overcome limitations in bio-signal processing.