Using recurrent artificial neural network model to estimate voluntary elbow torque in dynamic situations

Muscle modelling is an important component of body segmental motion analysis. Although many studies had focused on static conditions, the relationship between electromyographic (EMG) signals and joint torque under voluntary dynamic situations has not been well investigated. The aim of this study was to investigate the performance of a recurrent artificial neural network (RANN) under voluntary dynamic situations for torque estimation of the elbow complex. EMG signals together with kinematic data, which included angle and angular velocity, were used as the inputs to estimate the expected torque during movement. Moreover, the roles of angle and angular velocity in the accuracy of prediction were investigated, and two models were compared. One model used EMG and joint kinematic inputs and the other model used only EMG inputs without kinematic data. Six healthy subjects were recruited, and two average angular velocities (60 degrees s(-1) and 90 degrees s(-1)) with three different loads (0 kg, 1 kg, 2 kg) in the hand position were selected to train and test the RANN between 90 degrees elbow flexion and full elbow extension (0 degrees). After training, the root mean squared error (RMSE) between expected torque and predicted torque of the model, with EMG and joint kinematic inputs in the training data set and the test data set, were 0.17 +/- 0.03 Nm and 0.35 +/- 0.06 Nm, respectively. The RMSE values between expected torque and predicted torque of the model, with only EMG inputs in the training data set and the test set, were 0.57 +/- 0.07 Nm and 0.73 +/- 0.11 Nm, respectively. The results showed that EMG signals together with kinematic data gave significantly better performance in the joint torque prediction; joint angle and angular velocity provided important information in the estimation of joint torque in voluntary dynamic movement.