Combined classification and regression for simultaneous and proportional EMG control of wrist forces

In this study, a novel method for estimating wrist forces from surface electromyogram (EMG) measured from the upper limb is proposed, which can be applied for unilateral transradial amputees. Three degrees of freedom (DoFs) of wrist including flexion-extension, abduction-adduction, and pronation-supination were used. We first classify feature vectors extracted from the EMG signals into three classes namely positive output, negative output and dead zone output, using a multiple kernel learning (MKL) algorithm. Then for each DoF and each class, a neural network was trained to associate EMG features to their corresponding force outputs. We will show that this classification prior to regression plays an important role in increasing the performance of force estimation. The accuracy of estimation ranges from 90% to 94% (R² index) in 8 able-bodied subjects, which is proved to be significantly higher (p<;0.05) than that of the previous works.