We presented a novel state-based myoelectric control approach and the online control results from 9 intact-limbed subjects and one subject with upper limb deficiency using the proposed system. In the goal-directed online tests, the users would activate proportionally (sequentially) the two most important DoFs for trans-radial amputees: hand open/close and pronation/supination. The aim was to replicate the functionalities of commercial prostheses while providing a more natural switch between the two functions, without the need for an unintuitive and error-prone switching command such as co-contraction. Two online tests, a target reaching test, and a grasping test, were performed by the subjects. The target reaching test was chosen as it corresponds to one of the most representative tasks during the daily use of prostheses, i.e., reaching a precise target. In the current study, the subjects were free to operate the system as they preferred during the process of reaching, as opposite to tests based on sequences of predefined motions where the subject has very precise instructions at any point in time . The completion rates for 1-DoF and 2-DoF target reaching tests were on average, respectively, 96.2% and 91.5%, demonstrating that precise positioning control can be achieved within a short time (average: 3.0 s and 4.8 s). Further, the proposed approach was shown to significantly reduce the difficulty of 2-DoF tasks, when compared with the industrial SOA approach and the pure pattern recognition method. The grasping test was chosen to evaluate the potential of the algorithm in force control. Grasping is the most relevant function for proportional force control, as the manipulation of different objects requires maintaining different force levels. The completion rate in the grasping test was above 97% on average, showing that the subjects were able to control force accurately using the proposed system. The time required to maintain the instructed force level was on average <4 s. Misclassification and, especially, grasping failure occurred rarely (Table 2). This indicates that the system has good reliability in grasping. For both tasks, the performance of the subject with limb deficiency was similar to the average performance of the able-bodied subjects.
The subject with limb deficiency, along with subjects #3 and #5, expressed tiredness towards the end of the experiment. This most likely accounted for their greater misclassification and/or grasp failure occurrences. It is important to note that comparing with other studies on the topic, the training data used in this study were very short (<1 min recordings for 4 active classes and resting class). This suggests that it is possible to recalibrate the proposed algorithm daily (or even more often when necessary). Alternatively, it is expected that adaptive extensions to the proposed algorithm could lead to an even better performance. For example, the parameter VRC could be adaptively varied to account for slow changes in the signal characteristics, which may occur because of factors such as fatigue.
The overall performance in all tests was better for the 3 experienced subjects compared to the 6 naïve subjects, although the differences were not statistically significant, likely due to the small subject sample. Nevertheless, it suggests that the results presented could be improved by subject training. The system indeed involves classification that requires the contractions to be as repeatable as possible to limit the variability between training and testing. Experienced myoelectric control subjects are “trained” to produce repeatable contractions to obtain the best results. Along with experience, the online feedback, as provided in this study, most likely facilitated the adaptation of the subject to the myoelectric system, and thus it is of primary importance in myoelectric control research . For clinical applications, the improved performance with experience confirms that rehabilitation and training of the patients are key factors toward the successful use of myoelectric control for prosthesis control . In this study, the results showed that, with experienced subjects, the proposed algorithm allowed very high performance in both reaching tasks (completion rate 1-DoF: >98%, 2-DoF: >95% on average) and grasping task (>98% on average), did not have any grasping failure, and resulted in nearly no misclassification during grasping. In addition, despite no experience in myoelectric control, the performance of naïve subjects was reasonably good. These results suggest that the proposed system offers intuitive control. The subject with limb deficiency showed comparable performance as the naïve able-bodied subjects. The lower completion rate in the grasping test is to be related with the specific anatomy of the subject, which results in very limited surface EMG activity during attempts to operate hand function or wrist flexion/extension. This accounts for the larger target range used in the grasping test with the subject with limb deficiency.
In the current state-of-the-art methods in pattern recognition based myoelectric control algorithms, most of the classification errors occur at the transition between classes, during which the misclassification of resting state into active motions has the most detrimental effect on the usability of the algorithms [12, 20]. Various approaches have been proposed to address this problem. These include simple methods, such as majority vote , and more complex approaches, such as velocity ramp  and confident-based rejection [21, 22]. The general approach of these methods is to utilize the history or prior information of the system, at the output of the classifier. The current approach exploited the prior information before the classification stage. The advantage of this approach is that the structure of the classifier is reduced because the detection stage effectively removes the necessity of a ‘resting class’, which was shown to be associated with the majority of classification errors . Further, classification is only necessary when a state transition is detected, making the entire system more efficient. The direct comparison of the proposed method against the pure pattern recognition method (with MV post-processing) showed a significantly smaller nICT using the proposed method, supporting the advantage of the proposed method over a pure pattern recognition method.
The goal of myoelectric systems is to extract natural neural control information from EMG and to provide intuitive and reliable control of multiple functions to the prosthetic users. To this end, in this study we presented a real time state-based proportional control system. This system was shown to be reliable and to allow the subjects precise control of the feedback position and grasping force. With respect to commercial systems, the proposed algorithms provided a more natural and intuitive control, where the user can switch among available functions as it would be done naturally. This is in contrast with the conventional co-contraction based switching method, available in commercial systems, where the user needs to constantly monitor the active function, and perform additional strong contractions to switch between the available functions. Indeed, the subject with limb deficiency commented on the easiness of switching between wrist and hand functions of the proposed approach, as compared to the co-contraction switching mode in his commercial prosthesis. It is likely that the proposed approach would be more advantageous over the co-contraction-based switching when more functions need to be articulated since switching or circling through more than two functions would be mentally much more demanding.
In conclusion, we proposed a state-based myoelectric control system that was shown to be reliable and effective on the investigated situations, and to provide intuitive control to the subjects with minimal training. This system could provide a suitable alternative for the control of commercial prostheses.