Boosting training for myoelectric pattern recognition using Mixed-LDA.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Signal Processing

Background:

• Pattern recognition-based myoelectric prostheses (MP) require classifier calibration.
• Surface electromyography (sEMG) signal non-stationarity necessitates daily retraining for long-term MP use.
• Current training methods can be time-consuming and inefficient.

Purpose of the Study:

• To propose and evaluate Mixed-LDA, a novel method to optimize the retraining procedure for MP.
• To reduce the training time and improve the classification accuracy of MP systems.
• To address the challenges posed by sEMG signal non-stationarity in long-term prosthesis use.

Main Methods:

• Developed Mixed-LDA by combining current day's training samples with prior day's models.
• Conducted a 10-day experiment with 5 subjects to simulate long-term MP usage.
• Compared Mixed-LDA performance against the baseline Linear Discriminant Analysis (LDA) method.

Main Results:

• Mixed-LDA demonstrated significantly better performance than LDA when using limited training samples.
• In a 13-motion task, Mixed-LDA achieved an average classification rate of 88.74% with 104 training samples (LDA: 79.32%).
• The proposed method shows substantial improvement in classification accuracy with reduced training data.

Conclusions:

• Mixed-LDA effectively mitigates the impact of sEMG non-stationarity in myoelectric prostheses.
• The approach has strong potential to enhance MP usability by decreasing the required training duration.
• This method offers a practical solution for more efficient and accurate myoelectric prosthesis control.