A PCA-assisted EMG-driven model to predict upper extremities' joint torque in dynamic movements

To relate electromyographic signals (EMG) to net joint torque, different approaches have been taken into account. In this regard, some researchers chose to use Principal Component Analysis (PCA). A Study in 2001 reported a linear relationship between the PCA-processed EMG data and the joint torque while investigating isometric movements. In this project we questioned the possibility to use this method for free dynamic tasks. Four healthy subjects participated in the current study, performing three sets of Dumbbell Kick Back exercise for five different dumbbell weights. The net joint torque was calculated using the kinematic data in an inverse dynamics model. Meanwhile the EMG data were processed with a PCA method, and then were input to the model to estimate the joint torque. In order to predict this torque, we used two models; a single-input model that was fed with the PCA-processed EMG of the all corresponding muscles; and a double-input model that utilized the PCA-processed EMG data of the agonist and antagonist muscles separately. The results demonstrated that both the single-input and double-input models are capable of predicting the torque for both isometric and free dynamic tasks. Employing a paired t-test we found that the double-input model was significantly more successful in estimating the torque comparing to the single-input model (p <; 0.005). The other factor (the movement type) proved to also have a significant effect on the estimation outcome (p <; 0.0005). In general, this study suggests that a linear relationship exists between PCA-processed EMG data and the joint torque in both isometric and free dynamic movements; however, in order to have a better estimate of the net joint torque, distinguishing the agonist-antagonist muscle groups' generated torques may be beneficial.