A New Hierarchical Approach for Simultaneous Control of MultiJoint Powered Prostheses Aaron J. Young, Student Member, IEEE, Lauren H. Smith, Elliott J. Rouse, Student Member, IEEE, and Levi J. Hargrove, Member, IEEE
Abstract— Advanced upper-limb prostheses capable of actuating multiple degrees of freedom (DOF) are now commercially available. Pattern recognition based algorithms that use surface electromyography (EMG) signals measured from residual muscles show great promise as multi-DOF controllers. Unfortunately, current pattern recognition systems are limited to sequential control of each DOF. This study introduces a hierarchy of linear discriminant analysis (LDA) classifiers arranged to provide simultaneous DOF control. This approach and two other simultaneous control strategies were evaluated using healthy subjects controlling up to four DOFs, where any two DOFs could be controlled simultaneously. The new hierarchical approach was the most promising with classification errors at or below 15% on average for discrete and combined motions. The classification performance was significantly better (p < 0.05) than using a single LDA classifier trained to recognize both discrete and combined motions or classifying each DOF using a set of parallel classifiers. The high accuracy of the hierarchical approach suggests that pattern recognition techniques can be extended to permit simultaneous control, potentially allowing amputees to produce more fluid, life-like movements, ultimately increasing their quality of life.
M
I. INTRODUCTION
YOELECTRIC prostheses, which use surface electromyography (EMG) signals to control joint movement, are often used to effectively treat upperlimb amputation. To restore the form and functionality of the arm, such prostheses must be lightweight, robust, anthropomorphic, and capable of replicating the function of the lost limb. Even simple activities of daily living, such as opening a door, require simultaneous movement of multiple degrees of freedom (DOF). Recently developed multifunction prosthetic hands [1-5] and advanced arm system prototypes described in the literature [6-8] offer the mechanical means to restore such function; however, control
Manuscript received January 31, 2012. This work was supported in part by the NIH under Grant R01-HD-05-8000 and the Rice Foundation. A.J. Young and E.J. Rouse are with the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL 60611 USA and with the Department of Biomedical Engineering at Northwestern University (phone: 312-238-2415; fax: 312-238-208; e-mail:
[email protected] and
[email protected]). L.H. Smith is with the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL 60611 USA and with the Feinberg School of Medicine at Northwestern University (e-mail:
[email protected]). L.J. Hargrove, is with Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL 60611 USA and with the Department of Physical Medicine and Rehabilitation at Northwestern University, Chicago, IL, 60611, USA (e-mail:
[email protected]).
system improvements are required to take full advantage of these devices. Myoelectric control often uses the amplitude of a pair of agonist/antagonist EMG signals to directly control a corresponding DOF—a strategy known as direct control. Clinically, simultaneous control has only been previously implemented using direct control in patients who have undergone targeted muscle reinnervation (TMR) surgery [9]. Several approaches to providing simultaneous multi-DOF commands have been previously investigated to provide simultaneous control to patients without TMR. Artificial neural networks were previously used to predict joint kinematics [10] and kinetics [11] of the wrist. Activation patterns of underlying muscle synergies have been used to predict the movement of multiple wrist DOFs [12, 13] or hand postures [14]. Projection of the EMG signal energy onto an orthonormalized set of principle movement vectors has been investigated to predict combined movements of up to 3 wrist and hand DOFs [15]. These studies have focused primarily on either combined wrist movements or combined finger movements but few investigated combined wrist/hand motions that are frequently used during activities of daily living. Pattern recognition discriminates complex EMG signal patterns into a discrete number of classes. Many different feature set and classifier combinations have been shown to form excellent pattern recognition systems [16]. Currently, only one class may be selected for a given classifier, forcing users to use a combination of sequential DOF movements to command the prosthesis to perform a coordinated task. Such sequential control adds a cognitive burden associated with motion planning. Furthermore, it prevents the user from making fluid, lifelike movements. Notwithstanding these limitations, pattern recognition has been used to successfully control advanced robotic prostheses [17]. Previous attempts to use pattern recognition for simultaneous multi-DOF control have focused on two classifier architectures. Davidge created a single LDA classifier in which both discrete (1 DOF) movements and combined (2 DOF) movements were labeled as unique classes [18]. This classifier successfully classified discrete and combined wrist flexion/extension and hand open/closed movements for three of four combinations. In contrast, Baker et al. used a parallel classification scheme where three separate LDA classifiers were used to predict the motion of three digits simultaneously in a non-human primate [19]. A similar parallel architecture was also recently proposed using
Fig. 1. Block diagrams describing examples of each control strategy for a 3-DOF controller. Each box is an LDA classifier with motion classes from one or more DOFs. The single LDA classification strategy (left) discriminates all discrete and combined classes as separate motions. The parallel classification strategy (middle) discriminates each DOF individually using three LDA classifiers. The hierarchical classification strategy (right) uses the output of DOFs higher in the hierarchy to choose classifiers for DOFs lower in the hierarchy. As an example, first the wrist DOF motion is classified. The output (wrist flexion (WF), wrist extension (WE) or no motion (NM)) is used to choose the second classifier. If the wrist is not moving, a classifier for just rotation is used and its output determines which of three hand classifiers to use. If the wrist is moving (WF or WE), a classifier with all remaining discrete motion classes (wrist pronation (WP), wrist supination (WS), hand open, hand closed, or NM) is used to choose if a second motion class is active.
II. BACKGROUND Three types of pattern recognition strategies capable of classifying combined motions were tested, which are referred to as single LDA classification, parallel classification and hierarchical classification (Fig. 1). The single classification strategy consists of one LDA classifier in which each of the discrete and each of the combined motions are separate classes. The parallel classification strategy uses one LDA classifier for each DOF, where the decision of each classifier is calculated independently. Each classifier consists of three motion classes: the two opposing motion classes of a DOF and no motion. Each motion class is trained using data from its discrete motion and all combined motions in which it participated. The control scheme then outputs combined actions when two of the parallel classifiers have active motion classes as output. The architecture of this strategy is notably different than previously reported [19], which did not use combined motions to train the LDA classifiers. The hierarchical strategy is a new control system for classifying combined motions. This strategy consists of a hierarchy of LDA classifiers similar in structure to a decision tree. In the hierarchical scheme, the highest classifier in the hierarchy is identical to one of the LDA classifiers used in the parallel scheme, and uses both discrete and combined motion data to determine a motion class for a single DOF. The output of this classifier (one of two active
motion classes or no motion) then determines the next classifier used in the hierarchy. If the output is no motion, then the next classifier is used to determine the motion class of a second DOF. This second classifier is conditioned on the decision of the first classifier by removing any combined movements involving active motion in the first DOF from the training data set. If, instead, the output of the first DOF is an active motion class, then the second classifier used will determine whether the current intent is a combined motion. The second classifier used is a single LDA classifier which discriminates between the discrete motion selected by the first classifier and all combined motions in which the first motion classified participates. This same pattern is repeated down the hierarchy with one layer for each DOF, where all classifiers lower in the hierarchy are trained using data (a)
(b)
EMG Feature 2
support vector machines to provide simultaneous control of an elbow and a wrist/hand [20]. In this study, we propose a new method for providing simultaneous control, which uses a hierarchy of pattern classifiers to recognize combined elbow, wrist, and hand motions. We compare the performance of this new method to those of pattern recognition approaches previously described [18-20].
(c)
EMG Feature 1 NM
WE
WF
HO
WE+HO
WF+HO
HC
WE+HC
WF+HC
Fig. 2. Illustration of boundaries used by each of the three classification strategies: single (a), parallel (b), and hierarchical (c). EMG feature 1 and 2 correspond to amplitude on the wrist extension channel and wrist flexion channel respectively. Motion classes are displayed as colored shapes for the 2 DOF configuration (see methods for abbreviation definitions). Solid and dotted lines represent boundaries for different DOFs.
conditioned. See the Appendix for the training and real time algorithm, equations and a flowchart. Fig. 1 provides an example of the hierarchical strategy with wrist flexion/extension highest in the hierarchy followed by wrist rotation followed by hand open/close. Each of the three strategies described uses a different set of boundaries to discriminate between motion classes in feature space. Fig. 2 displays the approach of each classification scheme for the 2 DOF configuration. The single LDA classifier (Fig. 2a) separates in one step each discrete and combined motion as individual classes. In the parallel strategy (Fig. 2b), the LDA classifier for each DOF produces two boundaries for its three motion classes. The union of the bounds for all DOFs produces the overall classification boundaries for parallel controller. Combined motions result when features can be mapped to active motions in both LDA classifiers. In the figure, the solid lines separate the wrist motions, while the dotted lines separate the hand motions. The hierarchical strategy (Fig. 2c) progressively separates feature space as one navigates down the hierarchy. Feature space is first partitioned for the first DOF (wrist motions) in the same way as the parallel strategy (solid lines). The classifiers of the second level of the hierarchy (dotted lines) then discriminate the motions of the second DOF (hand motions) based on the boundaries formed by the first classifier. III. METHODS A. Experimental Protocol Six healthy subjects (three males and three females) completed the following experiment that had been approved by the Northwestern University Institutional Review Board. Four pairs of electrodes were placed around the circumference of the upper forearm approximately 2 cm distal to the elbow such that two pairs were placed on the wrist flexor muscle group and two pairs were placed on the wrist extensor muscle group. Two additional pairs of electrodes were also placed on the biceps and triceps to provide elbow discrimination. A ground electrode was placed on the olecranon, away from the muscles of interest. All data were collected using a Delsys (Boston, MA) Bagnoli-16 Amplifier. Signals were amplified to a convenient value through hardware; using a PC with Matlab (Natick, MA) the signals were digitally sampled at 1000 Hz and high pass filtered at 20 Hz using a 3rd order Butterworth filter to reduce motion artifact. Motions collected were hand open/close (HO/HC), wrist extension/flexion (WE/WF), wrist supination/pronation (WS/WP), elbow extension/flexion (EE/EF), no motion (NM) and all 2-DOF combined motions. Combined motions involving greater than 2 DOFs were not trained due to subjects reporting difficulty in visualizing such complex motions during pilot data collection, and the impracticality of collecting training data for every combined motion involving more than 2 DOFs. The data collection sessions were guided using visual prompts from custom designed
software [21]. The subjects were instructed to make medium, constant force contractions to the best of their ability; however, no feedback was provided to the subjects during the data collection procedure. Ample rest periods were provided during the data collection process to prevent fatigue. TABLE I: DOFS AND AMOUNT OF DATA COLLECTED FOR EACH SUBJECT Number Motions Amount of Data of DOFs 2
HO/HC, WE/WF
108 s (12 s per class)*
HO/HC, WE/WF, 3 228 s (12 s per class)* WS/WP HO/HC, WE/WF, 4 396 s (12 s per class)* WS/WP, EE/EF * Classes include discrete motions, combined motions, and no motion.
EMG data were divided into 250 ms windows with 50 ms frame increment [22, 23] and were represented using time domain (TD) features [24]. Four-fold cross validation was used to train and test each control strategy. For each of the three classification strategies (see Background), three different DOF configurations were tested (Table 1). These strategies controlled the discrete motions for each available DOF and all combined motions where only 2 DOFs were activated simultaneously. The 2 DOF configuration consisted of eight motion classes in which four were discrete and four were combined motion classes. The 3 DOF configuration consisted of fourteen motion classes in which six were discrete and eight were combined motion classes. The 4 DOF configuration consisted of eight discrete motions and twenty combined motion classes. The 2 and 3 DOF configurations used only the four channels around the forearm to provide EMG signals. However, with the addition of elbow flexion/extension in the 4 DOF configuration, two additional channels were used to include biceps and triceps activity. For all classification strategies, all available channels were used for every classification decision. For the parallel classification strategy, the output of the 3 and 4 DOF configurations was limited to only produce the two active motion classes that had the highest probability of correctness. This constraint was only necessary during less than 5% of all classifications. For the hierarchical scheme, the order of the DOFs in the hierarchy was chosen to ensure the best possible order for each user. This ordering of DOFs was tuned to each individual subject by evaluating performance on a test set of data for every possible ordering (see Discussion for explanation of ordering of DOFs). B. Classification Strategy Evaluation Offline classification error, defined as the percent of incorrect classifications, was used to evaluate classifier performances. For each control scheme the classification error resulting from testing discrete motions and combined
motions were also reported. Statistical comparisons were conducted using a general linear model with classification error as the response variable, DOF configuration, classifier strategy, error type (discrete or combined) as fixed factors, and subject as a random factor. Post-hoc comparisons with a Bonferroni correction factor were conducted to analyze differences between classifier strategy, DOF configurations and error type. A sequential LDA classifier was also trained on the discrete motion classes, and its performance was compared to the discrete motion classification error of the combined motion classification schemes. IV. RESULTS
Discrete Motion Classification Error (%)
A. Effect of Classification Strategy Classifier strategy and the number of DOFs controlled significantly affected overall classification error (p < 0.01 for both) (Fig. 3). In the post-hoc test, the single and the hierarchical strategies performed better than the parallel strategy (p < 0.01). The hierarchical classifier performed 40
parallel
single
hierarchical
sequential LDA
35
B. Classification Error for Parallel Classifier LDAs Classification errors for each of the classifiers in the parallel strategy are displayed in Table II. The hand open/close DOF had the highest error for all DOF configurations. TABLE II: CLASSIFICATION ERROR (%) FOR EACH DOF IN PARALLEL STRATEGY Number of DOFS 2 3 4
Wrist DOF
Hand DOF
Rotation DOF
2.2 ± 0.5
14.3 ± 2.7
-
10.5 ± 1.1 25.9 ± 2.8 12.6 ± 2.7 20.7 ± 3.6 29.3 ± 2.8 21.1 ±1.9 ± indicates standard error of the mean
Elbow DOF 9.5 ± 2.2
30 25
V. DISCUSSION
20
Simultaneous myoelectric control has been previously implemented clinically using direct control of multiple independent EMG control sites [9]. However, pattern recognition control offers many benefits, such as the control of a greater number of DOFs. To date, pattern recognition control has been most often presented at the cost of simultaneous control of multiple DOFs. This study introduced a new strategy and evaluated its performance when compare to two other classification strategies for providing simultaneous control of powered prostheses. The results from this study lay the foundation for the development of a clinically viable simultaneous pattern recognition control strategy using surface EMG. The hierarchical control strategy not only achieved an overall classification error of 15% or less for each DOF configuration tested for discrete and combined motions, but also showed the lowest discrete motion classification error of the three methods tested (Fig. 3a). Minimizing discrete motion error is important, as ideally, the addition of a combined motion classifier to the current control scheme should not sacrifice the ability to achieve sequential, discrete motion control. Furthermore, clinical experience with multiDOF direct control techniques [9] also suggests that misclassifications of discrete actions into combined actions is very detrimental to performance and is frustrating for the user. None of the combined motion controllers presented here had discrete classification error as low as sequential control. The hierarchical strategy provided had the least discrete motion classification error on average compared to
15 10 5 0 2 DOF
3 DOF
4 DOF
(a)
Combined Motion Classification Error (%)
better than the single classifier (p < 0.05). The 2 DOF configurations performed better than the 3 DOF or 4 DOF configurations (p < 0.01), and the 4 DOF configuration performed significantly better (p < 0.01) than the 3 DOF configuration. For discrete motions, the standard sequential classifier outperformed all the combined motion classifiers (p < 0.01 for all strategies) (Fig. 3a). The classification error for discrete movements was found to be significantly better (p < 0.05) than classification error for combined movements (Fig. 3a vs. Fig. 3b).
40
parallel
single
hierarchical
35 30 25 20 15 10 5 0 2 DOF
3 DOF
4 DOF
(b) Fig. 3. Discrete motion classification error (a) and combined motion classification error (b) of the three combined motion classifiers for each DOF option. The discrete motion classification error was also compared to a typical LDA sequential pattern recognition controller. Results are an average of six subjects. Error bars show +/- 1 SEM.
the other combined motion classification strategies, further suggesting it as a preferred method for pattern recognitionbased simultaneous control. The results presented in this study suggest important characteristics of classifier architecture for discriminating the investigated motions. The training accuracies in the single classification scheme were observed to be high (94% for the 2 DOF configuration, 92% for the 3 DOF configuration, and 93% for the 4 DOF configuration). These results suggest that it is possible to linearly separate the discrete and combined motions investigated in this study. This is consistent with previous studies that have shown linear classifiers to perform as well as more complex nonlinear control systems in discriminating motion classes [25]. Each of the three classification schemes presented used linear boundaries to discriminate between motion classes, yet produced significantly different classification performances (p < 0.05). The classification scheme with the lowest classification error was the hierarchical strategy, which outperformed both the single and parallel strategies (Figs. 3 and 4). This variation in classification performance is likely due to the differences in the schemes' methods for creating these bounds. The single LDA classifier makes no assumptions on how classes are grouped in feature space, and therefore treats single motion and combined motion classes equally when calculating classification boundaries. The entire set of training data is used to calculate 2N2DOF boundaries to partition 2N2DOF 1 classes (where NDOF equals the number of DOFs available in the control scheme). This method is most similar to pattern recognition schemes currently used for sequential control of DOFs and performed well in this simultaneous control application. In contrast, the parallel scheme assumes that motion classes that share at least one active DOF motion (e.g. WF, WF/HO, and WF/HC for a 2 DOF configuration) will be well separated in feature space. The parallel scheme also assumes that these groupings will be linearly separable from the grouping for the antagonistic movement (e.g. WE, WE/HO, and WE/HC) and grouping for no motion in that DOF (e.g. NM, HO, and HC). The overall poorer performance of the parallel method suggests that this assumption is not always valid. In particular the classification accuracies for each DOF classifier in the parallel scheme (Table II) demonstrate that while some DOFs may be well discriminated by this assumption (wellseparable DOFs), others, such as the hand DOF, are not (less-separable DOFs). The parallel scheme’s assumption also has important implications on the number of classification boundaries calculated, as only 2NDOF linear boundaries to separate 2N2DOF 1 classes. Therefore, as one increases the number of degrees of freedom, the total number of motion classes quickly surpasses the number of boundaries the parallel scheme calculates to separate these classes. This disparity between the number of classes and number of boundaries is not true for the single LDA or the
hierarchical strategies (which both produce 2N2DOF boundaries), and may explain the substantial increases in classification error produced by the parallel method with higher numbers of DOFs. Despite the poorer performance of the parallel strategy, the success of hierarchical strategy suggests that using groups of similar motions for classification may be advantageous, if groupings are applied appropriately. The hierarchical strategy uses the output of classifiers that group motions within more-separable DOFs (higher on the hierarchy) to choose more optimal classifiers for discriminating between the classes of less-separable DOFs (lower on the hierarchy). As a result, the hierarchy of LDA classifiers produces 2N2DOF classification boundaries that are specifically intended to better separate motions that share a similar DOF activation. Given that decisions lower on the hierarchy are conditioned by decisions made higher in the hierarchy, it is also important that the error of the earlier decisions is minimized. Because DOFs evaluated earlier in the hierarchy include more motion classes in the groupings, DOFs should be arranged where more-separable DOFs are evaluated before the less-separable DOFs. The hierarchical and other combined motion control strategies at each of the DOF configurations investigated may provide simultaneous control of up to 2 DOFs for a variety of amputation levels and prostheses types. Both the 2-DOF and the 3-DOF controllers are applicable to transradial amputees that use a myoelectric prosthesis to control both hand movements and wrist movements (WF/WE and/or WP/WS). The successful classification of combined wrist/hand movements in this study is particularly important, given the lack of significant previous work on controlling wrist and hand simultaneously. The 3 DOF controller studied allows for control of emerging advanced arms that allow 2 DOF wrist motion. In addition, the 4 DOF configuration, which adds elbow flexion/extension control in combination with any wrist or hand movement, suggests that the hierarchical strategy may be used by transhumeral and shoulder disarticulation amputees who have undergone TMR surgery [21]. Control of elbow movement is an important functional capability for this patient population. The parallel scheme’s ability to classify elbow motions with simultaneous control of other DOFs is therefore a promising result. It is interesting to note the decrease in classification error in the 4 DOF configuration compared to the 3 DOF configuration. This occurred because the discrete and combined motions involving the elbow joint were easily discriminated by each of the combined motions classification strategy, decreasing the overall classification error for the 4 DOF configuration. It is also useful to note that, the hierarchical strategy may be expanded to control advanced hands entering the market, as different hand grasp patterns can be incorporated as subclasses of the hand close motion class. One significant limitation to the simultaneous control
schemes presented here is the requirement that training data be collected for the combined classes. As the number of DOFs increase, the number of combined classes increases quadratically. Therefore, it would be ideal if only the discrete motion classes were collected and the combined classes were calculated as a combination of the discrete classes. The architecture of a parallel classification scheme theoretically supports the development of such a system; however, the parallel scheme investigated in this study did not limit training to discrete motions. Other approaches, such as the use of muscle synergies may also prove beneficial for simultaneous pattern recognition control to reduce the amount of training data needed. The promising results of this study demonstrate the need for additional investigation into the benefits and practicality of using simultaneous pattern recognition control, in particular the hierarchical strategy. However, this study is limited by the use of healthy control subjects and lack of online performance measures. Future studies will therefore use a simultaneous pattern recognition controller online, first in control subjects and then with both transradial and TMR subjects at the Rehabilitation Institute of Chicago.
A. Training Algorithm for Hierarchical Strategy ← ←
Evaluate Equation 1 Class 1 = Motion 1 of DOF #1
Evaluate Equation 2
1: _
,
1 2 3
| | |
1: 3
_
REFERENCES [1]
_
if 2 ∙ NDOF _ _ _ _ |
1 1:
,
_
1: ,
_ |
[2]
2
|
[3]
3
_
_
1
[4]
end end
[5]
B. Real-Time Algorithm for Hierarchical Strategy _ 1:
for
1 1
if
_
; ∈ , _
[6] ,
[7]
_
1 _
_
1 2 ,
_
_
,
[8]
_
end
[9]
_ end
Evaluate Equation 2
The authors would like to acknowledge Dr. Todd Kuiken for his expert advice regarding this study. We would also like to thank Blair Lock for his contributions in developing the project.
1: 3 ←
← # ←
Evaluate Equation 2
ACKNOWLEDGMENT
_
Class 1 = Motion 2 of DOF #1 Class 1 = No Motion in DOF #1
for
C. Equations and Flow Charts for Hierarchical Strategy is the EMG feature vector Equation 1: 1 arg max #1| Equation 2: 2 arg max #2 | , #1 1 Flow Chart:
Fig. 4. Flowchart of hierarchical strategy for two degrees of freedom. Each additional layer of the hierarchy for more than 2 DOFs is a repeat of the second layer, conditioned on all layers higher in the hierarchy.
APPENDIX NDOF ← #
_
1
,
_
,
C. Medynski and B. Rattray, "Bebionic Prosthetic Design," in 2011 MyoElectric Controls/Powered Prosthetics Symposium, Fredericton, New Brunswick, Canada, 2011. S. Schulz, "First Experiences With The Vincent Hand," in 2011 MyoElectric Controls/Powered Prosthetics Symposium, Fredericton, New Brunswick, Canada, 2011. O. Van Der Niet Otr, et al., "The i-LIMB hand and the DMC plus hand compared: a case report," Prosthetics and Orthotics International, vol. 34, p. 216, 2010. B. Waryck, "Comparison Of Two Myoelectric MultiArticulating Prosthetic Hands," in 2011 MyoElectric Controls/Powered Prosthetics Symposium, Fredericton, New Brunswick, Canada, 2011. J. M. Miguelez, "Clinical Experiences With The Michelangelo Hand, A Four-Year Review," in 2011 MyoElectric Controls/Powered Prosthetics Symposium, Fredericton, New Brunswick, Canada, 2011. L. Resnik, "Research update: VA study to optimize DEKA arm," J Rehabil Res Dev, vol. 47, 2010. A. Harris, et al., "Revolutionizing Prosthetics software technology," in 2011 IEEE International Conference on Systems, Man, and Cybernetics, Anchorage, AK, 2011, pp. 2877-2884. P. Kyberd, et al., "Two-degree-of-freedom powered prosthetic wrist," Journal of rehabilitation research and development, vol. 48, p. 609, 2011. T. A. Kuiken, et al., "The use of targeted muscle reinnervation for improved myoelectric prosthesis control in a bilateral
[10]
[11]
[12]
[13]
[14] [15]
[16]
[17] [18]
[19]
[20]
[21]
[22]
[23]
[24] [25]
shoulder disarticulation amputee," Prosthetics and Orthotics International, vol. 28, pp. 245-53, Dec 2004. S. Muceli, et al., "Multichannel Surface EMG Based Estimation of Bilateral Hand Kinematics during Movements at Multiple Degrees of Freedom," in 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Buenos Aires, Argentina, 2010, pp. 6066-6069. J. L. G. Nielsen, et al., "Simultaneous and Proportional Force Estimation for Multifunction Myoelectric Prostheses Using Mirrored Bilateral Training," Ieee Transactions on Biomedical Engineering, vol. 58, pp. 681-688, Mar 2011. C. Choi and J. Kim, "Development of a Myoelectric Joystick: a Preliminary Study," in 2010 3rd IEEE RAS and EMBS International Conference on Biomedical Robotics and Biomechatronics, Tokyo, Japan, 2010, pp. 173-178. N. Jiang, et al., "Extracting Simultaneous and Proportional Neural Control Information for Multiple-DOF Prostheses From the Surface Electromyographic Signal," Ieee Transactions on Biomedical Engineering, vol. 56, pp. 1070-1080, Apr 2009. A. Ajiboye and R. Weir, "Muscle synergies as a predictive framework for the EMG patterns of new hand postures," Journal of neural engineering, vol. 6, p. 036004, 2009. D. Yatsenko, et al., "Simultaneous, proportional, multi-axis prosthesis control using multichannel surface EMG," in 29th Annual International Conference fot eh IEEE EMBS, Lyon, France, 2007, pp. 6133-6136. E. Scheme and K. Englehart, "EMG Pattern Recognition for the Control of Powered Upper Limb Prostheses: State-of-the-Art and Challenges for Clinical Use," Journal of Rehabilitation Research and Development, vol. in Submission, 2010. T. A. Kuiken, et al., "Targeted muscle reinnervation for realtime myoelectric control of multifunction artificial arms," JAMA, vol. 301, pp. 619-28, Feb 11 2009. K. Davidge, "Multifunction Myoelectric Control Using a Linear Electrode Array," Masters of Science in Engineering, Electrical and Computer Engineering, University of New Brunswick, Fredericton, New Brunswick, Canada, 1999. J. J. Baker, et al., "Continuous detection and decoding of dexterous finger flexions with implantable myoelectric sensors," Ieee Transactions on Neural Systems and Rehabilitation Engineering, vol. 18, pp. 424-432, 2010. A. Boschmann, et al., "Development Of A Pattern RecognitionBased Myoelectric Transhumeral Prosthesis With Multifunctional Simultaneous Control Using A Model-Driven Approach For Mechatronic Systems," in 2011 MyoElectric Controls/Powered Prosthetics Symposium, Fredericton, New Brunswick, Canada, 2011. T. A. Kuiken, et al., "Targeted Muscle Reinnervation for Realtime Myoelectric Control of Multifunction Artificial Arms," Journal of the American Medical Association, vol. 301, pp. 619628, Feb 2009. K. Englehart and B. Hudgins, "A robust, real-time control scheme for multifunction myoelectric control," IEEE Transactions on Biomedical Engineering, vol. 50, pp. 848-54, Jul 2003. L. H. Smith, et al., "Determining the Optimal Window Length for Pattern Recognition-Based Myoelectric Control: Balancing the Competing Effects of Classification Error and Controller Delay," Ieee Transactions on Neural Systems and Rehabilitation Engineering, vol. 19, pp. 186-192, Apr 2011. B. Hudgins, et al., "A new strategy for multifunction myoelectric control," IEEE Transactions on Biomedical Engineering, vol. 40, pp. 82-94, Jan 1993. L. J. Hargrove, et al., "A comparison of surface and intramuscular myoelectric signal classification," Ieee Transactions on Biomedical Engineering, vol. 54, pp. 847-853, May 2007.
