Exploration of Data Dimensionality Reduction Methods for Improving Classification Performance of Voluntary Movements
Electromyography (EMG)-based pattern recognition is one of the motor function restoration approaches for the amputees and the hemiparetic patients who suffered from stroke, spinal cord injury or brain injury. To improve the classification performance of multiple voluntary upper limb movements, high-density EMG was often used which may include some redundant information and increase the computational loads. For this reason, a common spatial filter (CSP)- based data dimensionality reduction method was proposed in this study, and the motion classification performance using multi-class CSP was compared with that when using universal principal component analysis (PCA) and the individual PCA. 22 classes of 56-channel EMG data that recorded from the upper limb of eight brain injured patients were used. The results showed that CSP decreased the motion classification error by 2.9% in comparison to that when using all EMG data, and the CSP was significantly better than the two PCA-based data dimensionality reduction methods in terms of classification error.
KeywordsCSP PCA pattern recognition high-density EMG brain injury
Unable to display preview. Download preview PDF.