Modem Myoelectric Control - Is it Time to Change the Algorithmic Focus?

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyogram (EMG) signal analysis is crucial for hand, wrist, and finger movement recognition.
• Deep learning (DL) algorithms are increasingly used but have high computational costs, limiting clinical translation.
• Research is exploring reduced-complexity models and traditional/hybrid methods for EMG pattern recognition.

Purpose of the Study:

• To compare the performance of a state-of-the-art DL algorithm (ROCKET) against traditional EMG feature extraction methods.
• To evaluate lightweight methods (WLPHASOR, RMSPHASOR) and a novel hybrid approach (MSWL) for EMG pattern recognition.
• To challenge the narrative of DL dominance by assessing the performance gap in EMG analysis.

Main Methods:

• Comparative study using EMG data from 22 participants performing 11 hand/wrist movements.
• Utilized two EMG armbands (10 and 8 channels) and the open-source LibEMG toolbox.
• Compared Random Convolutional Kernel Transform (ROCKET) with Waveform Length Phasors (WLPHASOR), Root-Mean-Squared Phasor (RMSPHASOR), and Multi-Signal Waveform Length (MSWL).

Main Results:

• No significant accuracy differences between ROCKET, WLPHASOR, and RMSPHASOR (87% average accuracy).
• MSWL significantly improved performance to 90% average accuracy.
• The combination of ROCKET+MSWL achieved the highest average accuracy at 91%.

Conclusions:

• Lightweight traditional and hybrid EMG feature extraction methods offer competitive performance compared to advanced DL.
• The proposed MSWL method enhances EMG pattern recognition accuracy, especially when combined with DL.
• Findings suggest a re-evaluation of algorithmic focus in EMG analysis, balancing performance with computational efficiency.