Myoelectric control has been used as a prosthetic control strategy for many years. Its popular use dates back to the 1960s and 1970s.1–4 The most straightforward approaches use an estimate of the intensity of the myoelectric signal acquired from one or two surface electrodes placed over distinct remaining muscle sites. In many cases, these intensities are mapped (through various control schemes) to control the speed of the prosthesis. This gives rise to proportional control of the device by varying contraction strength. The optimal method of deriving these proportional control signals was investigated in the 1980s,5,6 with a simple low-pass rectification approach eventually being widely adopted. Since then, the ability to control the rate of motion has been suggested to increase the usability of prosthetic devices, but little work has been published to quantify this.7–9 Consequently, references to myoelectric control in the literature often contain the wording “proportional myoelectric control” regardless of their attention to proportionality. More recently, several groups have looked at using force mapping techniques to elicit proportional and simultaneous control, but these techniques are still very new.10–12
Pattern recognition–based approaches to myoelectric control have also been proposed for many years, and although still in the transition to clinical acceptance, they have found great traction in the research community.13 Improving the classification accuracy of these systems has been the primary focus, and thus, preprocessing and pattern recognition techniques have dominated the literature. It was not until more recently that a shift in focus toward clinical viability began to expand the scope of the research.14 In 2005, Lock et al.15 showed that classification accuracy was not wholly indicative of usability. Deficiencies in clinical robustness have been speculatively linked to several factors, including socket-fitting issues, temporal changes in signals, and an increase in variability of the signals during active prosthesis use. It is this induced variability that is of interest in the current work. Scheme et al.16,17 showed that, if left untreated, variations in limb position greatly affected classification accuracy. Hargrove et al.18–20 examined the effect of electrode shift and rotation, finding similar results.
With all the attention that pattern recognition–based myoelectric control has garnered, there has been little reference to its use with proportional control. In practice, the proportional control signal is often derived from a simple weighted average of the mean absolute value (MAV) of the electromyogram (EMG) channels used for pattern recognition. In 2011, Simon et al.21 were the first to compare this method with conventional direct proportional control and binary on/off pattern recognition control. However, they did not address the interaction between proportional control and classification accuracy. In the influential article of Hudgins et al.22 describing a novel myoelectric control scheme, they presented a time-domain (TD) set of features for classification that has since been widely adopted in myoelectric control research. It can be shown that the discriminatory power of these features predominantly comes from the MAV amplitude feature. As previously stated, the MAV is also used to derive the proportional control signal. This suggests that, although variations in contraction strength can be used to control proportionality, it may be to the detriment of the pattern classification.
In this work, the interaction between proportional control and classification accuracy is illustrated. With the goal of providing a control experience with high usability, best practices in proportional control gain selection and training methods have been developed.
Electromyographic data corresponding to seven classes of motion were collected from 10 healthy, normally limbed subjects (7 men, 3 women) ranging in age from 25 to 50 years. Data were collected over two different sessions, conducted on separate days. All experiments were approved by the University of New Brunswick’s Research Ethics Board.
Subjects were fitted with a cuff made of thermoformable gel that was embedded with eight equally spaced pairs of stainless steel dome electrodes. The cuff was placed around the right forearm (the dominant side for all participants), at approximately one-third of the length of the forearm, around the area of largest circumference. A reference electrode (RedDot by 3 M Health Care, St. Paul, MN) was placed over the back of the hand.
The eight channels of EMG were differentially amplified using remote AC electrode-amplifiers (BE328 by Liberating Technologies, Inc, Holliston, MA) and low pass filtered at 450 Hz with a fifth-order Butterworth filter. Data were sampled with a sampling frequency of 1000 Hz using a 16-bit analog-to-digital converter.
Subjects were prompted to elicit contractions corresponding to the seven classes of motion shown in Figure 1.
DATA ACQUISITION: SESSION 1
Subjects were asked to perform 3-second steady-state contractions of each motion shown in Figure 1 and were given a 2-second rest period between each. Users were prompted to perform the contractions at the following different force levels.
- Two repetitions at a medium level, corresponding to a moderate proportional control velocity.
- Two repetitions at a light level, corresponding to a slow proportional control velocity.
- Two repetitions at a hard level, corresponding to a fast proportional control velocity.
- Two repetitions at 100%, corresponding to a class-specific maximum voluntary contraction (MVC).
Note that although the subjects were instructed to elicit an intended force level, their interpretation of that level was determined subjectively without feedback. After the collection of the 100% repetitions, successive repetitions were collected as a percentage of MVC, while providing real-time proportional control feedback. Users were asked to track specific amplitude targets corresponding to percentage levels of effort (normalized by their 100% class-specific levels). The collection order was randomized but included two repetitions each of the following:
- 20%, 30%, 40%, 50%, 60%, 70%, and 80% of MVC.
- A dynamic ramp from 20% to 80% of MVC.
This entire process was then repeated (with the exception of the 100% repetitions), resulting in four repetitions of each level and a total of 46 repetitions of each motion class.
DATA ACQUISITION: SESSION 2
As in session 1, subjects were asked to elicit the seven motions from Figure 1 at different force levels to be used as training data. This time, the users were prompted to perform the following repetitions without feedback, as they saw fit:
- Two repetitions at a medium level, corresponding to a moderate proportional control velocity
- Two repetitions of a ramp, dynamically increasing intensity from a light to a hard level
Upon completion of this training data collection, session 2 consisted of a real-time tracking test, shown in Figure 2, similar to those used by Simon et al.21 and Corbett et al.23 Users were instructed to use class-specific proportional position control to track a sinusoid of randomly varying height and frequency. The control output was low pass filtered at 15 Hz to smooth the visualization (and to emulate the inertial smoothing effect incurred when driving mechanical devices). The target height was bound to the dynamic range of the control signal, and the target signal was band-limited to 10 Hz to allow for theoretically perfect tracking. Because the target motion changed during the test, the user was required to elicit the required proportional control output in the context of the target motion contraction. The users’ reticle (shown as a red circle in Figure 2) was hidden when the user-elicited contraction was classified as something other than the target motion class.
The tracking test consisted of tracking each active motion (excluding no motion) for 10 seconds with a 4-second pause between classes. The test was conducted under varying combinations of two conditions:
- Training data (medium or ramp repetitions).
- Automatic gain selection.
Because the training and configuration of pattern recognition–based systems can be cumbersome and complex, it is desirable to minimize the requirements. By automatically determining class-specific gains using the training data, the burden on the clinician and patient could be reduced. Herein, gain selection for the proportional control signal was done by autoscaling the mean of the class-specific proportional control signal (calculated using data from the training repetitions) to seven different levels corresponding to 40%, 50%, 60%, 70%, 80%, 90%, and 100% of the full dynamic range. This was done as follows:
where Nj is the number of samples for class j, xi,j is the data from channel i for class j, and MAVij is its MAV. The class-specific mean proportional control value, PCj, was calculated using
where Ni is the number of channels of EMG used for pattern recognition. Finally, the class-specific gain was determined using:
where TL is the desired target level (as a percentage of full scale) to map the class-specific mean proportional control value. During feed-forward use, the real-time proportional control value, PCcur, was determined using:
where GainPR is the class-specific gain for the current pattern recognition output, Ni is the number of channels, Nf is the frame length used for pattern recognition, and xi is the raw EMG data for channel i.
Three trials of the test were repeated for each combination of training data and gain setting, resulting in 42 total trials per subject.
Many different myoelectric control schemes have been proposed in the literature. The current state of the art is, debatably, based on the work of Englehart and Hudgins.24 They showed that combining a linear discriminant analysis (LDA) classifier with TD features provided effective real-time myoelectric control. Because of its minimal processing requirements, its simplicity, and its favorable performance, this system has been widely adopted13 and was used in this work.
Sampled EMG data were notch filtered at 60 Hz using a third-order Butterworth filter and segmented for feature extraction using 160-millisecond windows, with processing increments of 16 milliseconds. During the real-time tracking portion of session 2, a nine-decision majority vote window was used to smooth intermittent decisions. This was done by overwriting the instantaneous output decision by the decision occurring most frequently during the previous nine outputs.
Results from session 1 were compared using standard classification error. In session 2, performance during the real-time amplitude tracking test was evaluated using classification error and the R2 error:
where a[n]— is the user generated amplitude, a[n] is the target amplitude, and a[n]— is the mean target amplitude.
The R2 statistical analyses were conducted using a multiway analysis of variance (ANOVA) with a post hoc multiple comparison (Tukey-Kramer) test. The α value used was 0.001.
SESSION 1: INTERLEVEL AND INTRALEVEL ANALYSIS
A separate classifier was trained using data from each different contraction level (11 classifiers in total) and tested using data from all levels. Testing was repeated in a leave-two-out fashion, each time training with two repetitions and testing with the other two. The results were averaged and are shown in Table 1. Lower levels of error are shown as lighter colors, whereas higher errors are shown as darker colors. It can be seen that the intralevel errors (those on the diagonal) are lower, with a mean error of 9.3%, whereas the mean interlevel error is 31.2%. Of note is the mean interlevel error when training with a ramp contraction, which is 18.9%. This is lower than any other single level.
From Figure 3, it can be seen that increasing training contraction strength distances the features from the no motion data. Considering that the LDA classifier boundaries are determined using these training clusters, it is understandable how a classifier trained with, for example, the 80% clusters could erroneously misclassify data from the 20% clusters as no motion. Note that the ramp data largely encompass the areas ranging from the 20% to the 80% clusters and in between.
Figure 4 shows the results of a multivariate analysis of variance when training with each single training level and testing with all levels. The results shown were calculated using a post hoc multiple comparison test and showed a significant training effect (p < 0.001). The horizontal interval bars were computed using MATLAB’s (The MathWorks Inc, Natick, MA) “multcompare” function such that two estimates are significantly different if they have nonoverlapping intervals. Note again that the ramp contraction significantly outperformed all other levels. It should be noted that, although combining multiple discrete training levels also performed well, collection of multiple training repetitions at various levels is a clinically undesirable requirement.
Figure 5 shows the mean normalized contraction level with respect to the 100% class-specific MVC. Note that, on average, the light, medium, and hard levels roughly correspond to 20%, 40%, and 70% MVC, respectively. The wrist supination class light and medium levels tended to be higher than those of the other classes.
SESSION 2: TRACKING TEST ANALYSIS
Figure 6 shows the classification error for all combinations of training method and gain level. Note, again, that the gain levels depicted along the horizontal axis correspond to the gain necessary to map a class-specific mean to that percentage of the dynamic range.
Figure 7 shows the corresponding R2 tracking score. A multiway ANOVA (with post hoc multiple comparison test) indicated a significant interaction (p < 0.001) between training method and gain level. As a result, an f test was performed comparing the variability in the R2 tracking score of the two training methods caused by changes in gain level. The results showed that training with ramp data produced a control scheme that was significantly more tolerant to variations in gain level.
The ANOVA showed that training with the dynamic ramp data significantly improved (p < 0.001) both classification error (14.84% ± 0.60% and 11.16% ± 0.54% for medium and ramp training, respectively) and R2 tracking score (0.1458 ± 0.0368 and 0.3279 ± 0.0276 for medium and ramp training, respectively) during the tracking task. It also showed that the selection of gain level had a significant effect (p < 0.001) on both classification error (Figure 8) and R2 tracking score (Figure 9). In both cases, selecting a gain level that mapped the class-specific mean proportional control value to 60% of the full dynamic range provided the best performance. Again, note that in Figures 8–11, nonoverlapping intervals indicate a significant difference between comparisons.
Figures 10 and 11 show the classification error and R2 tracking score, respectively, broken up by class. Note that the wrist extension class was significantly worse than the other classes in both cases. Wrist flexion was significantly better than the other classes with respect to classification accuracy, but not so for R2 tracking.
This work was divided into two sections. The first section, session 1, looked at classification error when training with one contraction level but testing with others. This is analogous to common clinical pattern recognition studies where users are instructed to elicit a repeatable medium-level contraction during training then asked to use proportional control during usability testing. As seen in Figure 4, these medium-level contractions indeed offer the best performance out of the static level choices. It is evident, however, that performance quickly degrades as static training levels deviate from the 40% to 50% range. The ramp training scheme significantly outperformed all level-based schemes. This is likely because of the presence of exemplars from many contraction levels in the resultant training data. In addition, the dynamic nature of the ramp contraction bodes well for suitability during equally dynamic proportional control tasks.
The class-specific user-elicited contraction levels are shown in Figure 5. When prompted to elicit hard contractions without feedback, the average corresponding normalized levels reached between 60% and 70%. This implies that anything higher than 60% is simply unsustainable for most users. Similarly, user-elicited light contraction, described to users as “as light as possible while still performing the class contraction,” was limited to approximately 20%. These results yield a suggested usable range of 20% to 60% of class-specific MVC. Users often stated difficulty in achieving the higher level targets when using the wrist supination class. This is thought to be caused by a reduced dynamic range in the forearm signals corresponding to wrist supination. The theory is supported by the shift in user-elicited contraction levels compared with the other classes (Figure 5). This could be alleviated by including electrodes over the biceps (the prime movers for supination), but it is not a clinically desirable arrangement.
The second part of the study, session 2, used lessons learned during session 1 to perform a real-time amplitude tracking test with feedback. The tracking cursor was hidden when the classifier concluded that the user was performing an erroneous class contraction. This behavior was chosen because a prosthetic user would be unable to track a given degree of freedom while unintentionally activating another. Consequently, users were asked to focus on reestablishing the correct classification before concentrating on tracking. This predictably created a strong relationship between classification error and tracking score.
As learned in session 1, using a ramp contraction during training greatly improves the classifier’s performance when testing with various contraction levels. This theory held true during the proportional tracking test, yielding a significant improvement over the standard medium-level contraction. Furthermore, it was shown to provide superior resilience to suboptimal gain selection.
Gain selection was also shown to have a significant effect on classification accuracy and tracking ability. This is an important finding of this work, as clinical practice currently consists of manually setting gains based on empirical observation. These results indicate that, on top of simplifying the clinical configuration process, autogain selection could ensure optimal performance of a pattern recognition–based myoelectric control system.
There is also a tradeoff between gain level and user effort level. Users indicated a preference for higher gains because of ease of contraction but complained of an associated “jumpiness” or lack of controllability. Conversely, users expressed strong aversion to lower gain levels because of an inability to produce higher control signals. When using lower gains, many users resorted to eliciting pulsatile/burst contractions to approach higher targets, and some gave up altogether. Consequently, the level of user effort must be taken into consideration when choosing gain levels. Fortunately, the algorithmically determined optimal gain levels correspond with those empirically preferred by the users.
It is instructive to note that, in this work, as in previous work examining the effect of limb position on classification error,16,17 the robustness of myoelectric control seems to be related to the proper population of feature space. Training data collected using “repeatable medium-level” contractions with the limb in a rested position (as is often the norm in myoelectric controls research) can produce good offline classification accuracies. The resultant feature space is, however, sparse and the true class boundaries are underdefined. Under more challenging usage scenarios, these boundaries are tested by dynamically varying contractions and class transitions, resulting in undefined classifier behavior. It is the opinion of the authors that proper population of feature space and definition of class boundaries will play an integral role in the deployment of clinically viable pattern recognition–based myoelectric control systems.
Finally, it should be noted that this study was conducted using normally limbed subjects. Because this study provided an introductory look at the effects of force variation on classification, the data collection protocol was onerous, at times requiring intense contraction of the muscles. Completion of the protocol, especially by amputee patients, was demanding, something we endeavor to shelter our patients from. In previous work by Scheme et al.,25 it was shown that the relative trends when comparing myoelectric control schemes were consistent between normal and amputee populations. Although the end goal of any prostheses research remains to develop and test algorithms for the end-user, we believe that using normally limbed subjects in the early investigation and validation of control algorithms is justified.
This work has shown that the use of proportional control has a significant effect on classifier performance. Consequently, the authors are currently performing an ongoing study looking at algorithmic approaches to alleviating proportional control-induced classification errors. Alternate methods of deriving the proportional control signal are also being investigated.
This work is part of a larger investigation aimed at understanding and mitigating the factors that limit the clinical robustness of myoelectric control.
1. Bottomley AH. Working model of a myoelectric control
system. In: Proceedings International Symposium on the Application of Automatic Control in Prosthetics Design, Belgrade, Yugoslavia, 1962.
2. Herbets P. Myoelectric signals in control of prostheses
. Acta Orthop Scandinavica
1969; 124 (suppl 3):83.
3. Dorcas DS, Scott RN. A three-state myoelectric control
. Med Biol Eng
1966; 4: 367–372.
4. Finley RR, Wirta RW. Myocoder studies of multiple myocoder response. Arch Phys Med Rehabil
1967; 48: 598.
5. Hogan N, Mann R. Myoelectric signal processing: optimal estimation applied to electromyography, part 1: derivation of the optimal myoprocessor. IEEE Trans Biomed Eng
1980; 27: 382–395.
6. Evans H, Pan Z, Parker P, Scott R. Signal processing for proportional myoelectric control
. IEEE Trans Biomed Eng
1984; 31: 207–211.
7. Peerdeman B, Boere D, Witteveen H, et al.. Myoelectric forearm prostheses
: state of the art from a user-centered perspective. J Rehabil Res Dev
2011; 48: 719–738.
8. Gordon K, Ferris D. Proportional myoelectric control
of a virtual object to investigate. Exp Brain Res
2004; 159: 478–486.
9. Sears H, Shaperman J. Proportional myoelectric hand control: an evaluation. Am J Phys Med Rehabil
1991; 70: 20–28.
10. Jiang N, Vest-Nielsen J, Muceli S, Farina D. EMG
-based simultaneous and proportional estimation of wrist/hand dynamics in unilateral trans-radial amputees. J Neuroeng Rehabil
2012; 9: 42.
11. Muceli S, Farina D. Simultaneous and proportional estimation of hand kinematics from EMG
during mirrored movements at multiple degrees-of-freedom. Trans Neural Syst Rehabil Eng
2012; 20: 371–378.
12. Nielsen J, Holmgaard S, Jiang N, et al.. Simultaneous and proportional force estimation for multifunction myoelectric prostheses
using mirrored bilateral training. Trans Biomed Eng
2011; 58: 681–688.
13. Oskoei M. Myoelectric control
systems—a survey. Biomed Signal Process Control
2007; 2: 275–294.
14. Scheme E, Englehart K. EMG pattern recognition
for the control of powered upper-limb prostheses
: state of the art and challenges for clinical use. J Rehabil Res Dev
2011; 48: 643–660.
15. 15. Lock B, Englehart K, Hudgins B. Real-time myoelectric control
in a virtual environment to relate usability vs. accuracy. Presented at the Myoelectric Controls Symposium, August 17–19, 2005, Fredericton, New Brunswick, Canada.
16. 16. Scheme E, Fougner A, Stavdahl Ø, et al.. Examining the adverse effects of limb position on pattern recognition
based myoelectric control
. Paper presented at: Conference Proceeding of the 32nd annual conference of the IEEE Engineering in Medicine and Biology Society, August 31–September 4, 2010, Buenos Aires, Argentina, vol. 32, pp 6337–6340.
17. Fougner A, Scheme E, Chan A, et al.. Resolving the limb position effect in myoelectric pattern recognition
. IEEE Trans Neural Syst Rehabil Eng
2011; 19: 644–651.
18. 18. Hargrove L, Englehart K, Hudgins B. The effect of electrode displacements on pattern recognition
based myoelectric control
. Proceedings of the 28th annual conference of the IEEE Engineering in Medicine and Biology Society, August 30–September 3, 2006, New York, pp 2203–2206.
19. Hargrove L, Englehart K, Hudgins B. A training strategy to reduce classification degradation due to electrode displacements in pattern recognition
based myoelectric control
. Biomed Signal Process Control
2008; 3: 175–180.
20. Young A, Hargrove L, Kuiken T. The effects of electrode size and orientation on the sensitivity of myoelectric pattern recognition
systems to electrode shift. IEEE Trans Biomed Eng
2011; 58–59: 2537–2543.
21. 21. Simon A, Stern K, Hargrove L. A comparison of proportional control
methods for pattern recognition
control. Proceedings of the 33rd annual conference of the IEEE Engineering in Medicine and Biology Society, August 30–September 3, 2011, Boston, MA, pp 3354–3357.
22. Hudgins B, Parker P, Scott RN. A new strategy for multifunction myoelectric control
. IEEE Trans Biomed Eng
1993; 40: 82–94.
23. Corbett E, Perreault E, Kuiken T. Comparison of electromyography and force as interfaces for prosthetic control. J Rehabil Res Dev
2011; 48: 629–642.
24. Englehart K, Hudgins B. A robust, real-time control scheme for multifunction myoelectric control
. IEEE Trans Biomed Eng
2003; 50: 848–854.
25. Scheme E, Englehart K, Hudgins B. Selective classification for improved robustness of myoelectric control
under non-ideal conditions. IEEE Trans Biomed Eng
2011; 58: 1698–1705.
Keywords:© 2013 American Academy of Orthotists & Prosthetists
prostheses; myoelectric control; EMG; proportional control; pattern recognition