Role of the Window Length for Myoelectric Pattern Recognition in Detecting User Intent of Motion

In this study, the problem of electromyographic (EMG) based motion intention detection (MID) of upper limb was addressed, investigating the role of the window length for feature extraction. Two pattern recognition experiments were performed taking into account eight healthy subjects. The first involved the direct comparison of classification performances using feature computed over 150, 100 and 50 ms window length for eight class of shoulder movements. In the second one, a feature fusing scheme, based on canonical correlation analysis (CCA), was used to investigate whether pattern recognition architectures (PRAs), i.e. support vector machine, were able to boost their performances when 50 ms features were used as testing set. The rationale behind such investigations grounds on the lack of consensus regarding the most suitable window length for myoelectic pattern recognition. No drop of accuracy was observed in the first experiment for the three different windows length, maintaining values around 90%. Moreover, as observed in the second experiment, the CCA feature fusing scheme enhanced the performances of the PRAs when working over 50 ms features, reaching comparable results with feature at 150 ms. The proposed approach can be suitable for MID in real-time scenario, where the computational represents a central issue.