Boosting training for myoelectric pattern recognition using Mixed-LDA

Pattern recognition based myoelectric prostheses (MP) need a training procedure for calibrating the classifier. Due to the non-stationarity inhered in surface electromyography (sEMG) signals, the system should be retrained day by day in long-term use of MP. To boost the training procedure in later periods, we propose a method, namely Mixed-LDA, which computes the parameters of LDA through combining the model estimated on the incoming training samples of the current day with the prior models available from earlier days. An experiment ranged for 10 days on 5 subjects was carried out to simulate the long-term use of MP. Results show that the Mixed-LDA is significantly better than the baseline method (LDA) when few samples are used as training set in the new (current) day. For instance, in the task including 13 hand and wrist motions, the average classification rate of the Mixed-LDA is 88.74% when the number of training samples is 104 (LDA: 79.32%). This implies that the approach has the potential to improve the usability of MP based on pattern recognition by reducing the training time.