Robotic systems have been used in stroke rehabilitation to restore upper limb functions. In this study, an innovative myoelectrically controlled robotic system with one degree of freedom (DOF) was developed. The axis of the robotic system was aligned with the elbow joint and a torque from a servo motor was applied directly to the elbow based on the electromyographic (EMG) signal of the subject’s affected muscle at the elbow joint. This could help subjects after stroke to perform active elbow training in the horizontal plane. Two control strategies were investigated for the robotic system:​ the recurrent artificial neural network (RANN) model and proportional control. The RANN model was investigated on six subjects without impairment and three subjects after stroke. After training, the average cross-correlation coefficients between the expected and the predicted torque of subjects without impairment were 0.97±0.01 in the training data and 0.92±0.03 in the test data, respectively. It appeared that the output of the RANN was highly correlated to the expected torque. However, the performance of the RANN model on the three subjects after stroke did not show results as good as that on the subjects without impairment. The average cross-correlation coefficients of the subjects after stroke were 0.73±0.10 in the training data and 0.41±0.07 in the test data, respectively. Proportional control with a resistive load was used as an alternative control strategy. With the application of proportional control, the system could provide assistive extension torque which was proportional to the amplitude of the subject’s processed and normalized triceps EMG. The EMG-torque gain was set at 0%, 50%, 100% and 150% for the assistive torque. The system could also provide a resistive load, the level of which ranged from 0%-20% of the maximum isometric voluntary extension (MIVE) torque and the maximum isometric voluntary flexion (MIVF) torque of the affected elbow when the elbow angle was at 90 deg. Effects of the resistive loads and EMG-torque gains on the performance of the elbow extension were investigated on the affected arms of nine subjects after stroke in a tracking experiment. Results showed that the design could enable eight subjects with weak triceps to extend their affected elbows to a more extended position with the assistance of the myoelectrically controlled robotic system except for one subject who could already extend her elbow to the full extension position (0 deg) without the assistance of the system. There was a significant decrease in triceps EMG along with the increase in the EMG-torque gain during the elbow movement from 90 deg to 60 deg, which implied that it took less effort for subjects after stroke to perform the same movement with a larger gain. Since the myoelectrically controlled robotic system could facilitate elbow movement, its long-term training effect on the functional improvement of the affected arm in three subjects after stroke was investigated in a 20-session training program for four weeks. In each session, there were 18 trials with different combinations of the EMG-torque gain and the resistive load. In each trial, the subject was asked to follow a target trajectory which ranged from 0 deg to 90 deg, and complete five-cycle repetitive elbow flexion and extension with the myoelectrically controlled robotic system. Outcome measurements on the muscle strength at the elbow joint showed that there were increases in the MIVE and MIVF torques of the affected arms of all the subjects after the four-week rehabilitation training. The subjects could also reach a more extended position without the assistance of the robotic system after the four-week rehabilitation training. Moreover, there were a decrease in the modified Ashworth scale and an increase in the Fugl-Meyer score for all three subjects after the four-week rehabilitation training.

In addition, another sinusoidal arm tracking experiment was designed to quantitatively evaluate the elbow control function on nine subjects after stroke in dynamic situations. The movement performance was analyzed in terms of three parameters:​ root mean square error (RMSE) between the actual elbow angle and the target angle, root mean square jerk (RMSJ) and response delay (RD) at six velocities (10, 20, 30, 40, 50 and 60 deg/s). Results showed the RMSE and RMSJ increased in both the affected and the unaffected arms with the increase in the tracking velocity. The RMSE and RMSJ of the unaffected arms were significantly lower than those of the affected arms at all the velocities studied. The RD of the affected arms was larger than that of the unaffected arms at the velocities of 20, 30, 40 and 60 deg/s. There were significant correlations between the RMSJ and the modified Ashworth scale at the velocities of 10, 20, 30, 40 and 60 deg/s. The sinusoidal arm tracking experiment was also conducted on the three subjects after stroke before and after the four-week training. Results showed that there were decreases in the RMSE and RD of the affected arms after the four-week training, which indicated the improvement of the elbow control function in the affected arms for the three subjects.