Effect of Window Conditioning Parameters on the Classification Performance and Stability of EMG-Based Feature Extraction Methods

For upper limb multiple degrees of freedom prosthesis to be clinically viable, its control performance should be accurate and consistently stable over time. Factors such as the feature extraction methods and window conditioning parameters play an important role in this context. To provide information on optimal feature/windowing parameters, this study investigated the accuracy and stability of notable time-domain (TD) and frequency-domain (FD) features across different windowing conditions. Specifically, the interaction effect of a range of window length (50ms~300ms) and window increment of 25ms, 50ms, and 100ms, on the classification performance, stability, and computation time of TD and FD features were examined based on electromyogram (EMG) recordings of four able-bodied subject performing seven classes of limb motions. Experimental results show that TDAR (consisting of 4th Order autoregressive coefficient and root mean square) achieved the lowest classification error (CE) among the TD features at an optimal window size of 300ms and increment of 100ms, while MNP (mean power) recorded the best accuracy among the FD features. Despite the significant reduction in CE of TDAR and MNP over the other features, their computation time were observed to be relatively high thereby indicating a trade-off between accuracy and computation time amongst the different feature extraction methods. Thus, the findings from this study may provide potential insight on the proper choice of features and window conditioning parameters in the context of research and practical applications in myoelectric control systems.