Cascaded Adaptation Framework for Fast Calibration of Myoelectric Control

In spite of several decades of intensive research and development, the existing algorithms of myoelectric pattern recognition (MPR) are yet to make significant clinical and commercial impact. This study focuses on the one of the limiting factors of current algorithms: degradation of algorithm performance due to the inherent non-stationarity in electromyography (EMG) signals and the consequent need for frequent re-training and re-calibration. In order to reduce the re-calibration time required for donning/doffing between sessions and avoid to need to re-calibrate within a given session while donning, we propose a cascaded adaptation (CA) framework based on linear discriminant analysis (LDA), which automatically incorporates models from previous sessions in the model calibration for the current session. The framework also updates the model parameters according to new data samples and the corresponding recognized labels. Both off-line analysis (with data from eight intact-limbed subjects and three trans-radial amputees) and online testing with 9 intact-limbed subjects were conducted to evaluate the proposed method. Results show that the LDA embedded with CA (LDA-CA) is able to classify 11 types of motion with a small training data set, beginning from the second session of the experiment. The proposed LDA-CA obtains better performance as compared with three other methods-baseline LDA (LDA-BL), LDA with self-enhancing (LDA-SE), and LDA with domain adaptation (LDA-DA). The online test demonstrates that LDA-CA requiring an initial 1 min training session can be reliably used for 8 h without re-training. The proposed myoelectric control framework with low calibration burden has the potential to move the MPR based prostheses from academic research to clinical application.