Pattern recognition–based electromyography (EMG) control scheme is an intelligent signal-processing technique that has been applied to the prosthetic control problem for decades, with great promise.1–4 This scheme relies on differentiating between signal patterns of different limb motions.1–4 By using effective feature extraction and multidimensional classifiers, myoelectric pattern recognition is capable of discriminating more prosthetic functions than are current myoelectric control schemes. Many researchers have reported great improvement in the dexterity of control in upper-limb prostheses using EMG pattern recognition, and this method is nearing clinical viability.5
Currently, several factors such as variation in electrode placement6,7 and impedance as well as the effect of socket loading challenge pattern recognition control in clinical settings. The “position effect” is also a challenging problem that has recently been the focus of several researchers.5,8,9 It refers to the degradation of myoelectric pattern recognition performance when the classifier is used with the limb in positions different from the ones used for training or while performing dynamic limb activities.9 This degradation is largely due to the impact of arm position variation on the muscular activation patterns when performing different activities.10–12
A review of the literature shows that resolving the positions effect, which is the focus of this work, has been tried by many researchers. Scheme et al.8 trained a classifier to detect eight motion classes in eight different limb positions in able-bodied subjects. They observed a severe drop in pattern recognition performance, empirically linked to changes in posture and limb position, and demonstrated that training the classifier in multiple positions reduces this degradation to some extent. Their work was replicated by Chen et al.13 using data from transradial amputees, whose results supported the notion that training the classifier in multiple positions reduces the position effect. Scheme et al.,5 in another work, showed that changing the limb position during both static and dynamic activities of daily living (ADLs) also negatively affects myoelectric pattern recognition. They proposed training the classifier with dynamic activities to efficiently reduce these effects. Cipriani et al.14 showed that variations in the weight of the prosthesis and residual limb movements significantly influence the robustness of a classifier and suggested that inertial sensors, capable of monitoring the posture and movement of the limb, could be beneficially combined with myoelectric signals. Boschmann and Platzner15 used a 96-channel high-density electrode array and showed that training the classifier in multiple positions (three positions in their work) in combination with an increased number of EMG channels helps reduce the effect of limb position variation on classification accuracy. Jiang et al.16 showed that changing arm position adversely influences the performance of the myoelectric control algorithm, but their results suggested that this influence may be less pronounced in amputee subjects than in able-bodied subjects. Finally, Radmand et al.17 focused on the impact of limb position variation on features of the EMG, separability and repeatability, on which the robustness of pattern recognition methods rely. In their work, separability index was defined as the distinctness of classes in the feature space, and the repeatability index was defined as the degree of coincidence between training and testing data sets. They used these indices to show that, when the limb is moved to a position different from the one in which the classifier is trained, both the separability and repeatability of the data decrease. The results of their work showed that, by increasing either separability or repeatability, the position effect can be reduced.
In another body of work, accelerometry (ACC) data were used as an additional input to the classifier. Scheme et al.8 showed that integrating ACC data with EMG signals to provide positional context to the classifier helps to reduce the position effect problem. Scheme et al.8 also proposed a dual-stage classification method to partially resolve the position effect, using ACC data to select position-specific classifiers and EMG to classify desired prosthetic function. Similar work was repeated by Fougner et al.,9 but for five limb positions. Geng et al.18 collected EMG and ACC signals simultaneously to decode six classes of motion in five limb positions. They showed that a dual-stage classification, similar to that proposed by Scheme et al.8 (with ACC data for limb position identification and EMG for limb motion classification), may be a promising way to reduce the limb position effect.
Generally, most of the studies involving the position effect can be categorized into one of the following three categories:
- Single-stage classification, training in multiple positions using EMG data only
- Single-stage classification, training in multiple positions using EMG + ACC data
- Dual-stage classification, using ACC data to detect position and EMG data to detect motion
In most of the abovementioned works, it was assumed that the classifier was trained with data from all possible limb positions. Since, because of the excessive training time, it is not always clinically practical to train the classifier with data from all possible positions, training in a limited number of limb positions is desired. Although Fougner et al.9 studied training in a subset of the possible positions using EMG data, the suitability of this approach has not been investigated before for integration of ACC data and EMG data. The aim of this work was first to compare the performance of the previously proposed methods, when the training occurs only in a subset of tested positions. The focus is on the efficacy of combining ACC and EMG data for reducing the position effect problem. It is shown in this work that, unless the classifier is trained in most of the possible positions, the integration of ACC data with EMG data through a single-stage classification approach can significantly degrade performance compared with EMG alone. Second, a practical dynamic training method that moves the residual limb through the regions of interest is introduced. This approach is shown to minimize training time while improving performance when combining ACC and EMG data.
POPULATION AND DATA ACQUISITION
Electromyography data corresponding to eight classes of motion were collected from 10 right-handed, healthy subjects with normal limbs (9 men, 1 woman) within the age range of 19 to 32 years. All experiments were approved by the University of New Brunswick’s Research Ethics Board.
A Trigno Wireless System (Delsys Inc, Natick, MA, USA) was used to record surface myoelectric signals. Six wireless electrodes were equally spaced around the dominant forearm, proximal to the elbow, at the position with the largest muscle bulk. The six channels of EMG were band-pass filtered (20–450 Hz Butterworth) and sampled at 1 kHz by a custom data collection system. Accelerometry data were acquired by two CHRobotics UM6 inertial measurement units (IMUs; CHRobotics LLC, Payson, UT, USA) and sampled at 1 kHz. The first IMU was affixed on the forearm, over the brachioradialis muscle, and the second was placed over the biceps brachii, as shown in Figure 1.
The subjects were prompted to elicit contractions corresponding to the eight classes of motion including wrist flexion/extension, wrist supination/pronation, power grip, pinch grip, hand open, and no movement. Each contraction was sustained for 3 seconds, and a 3-second rest was given between subsequent contractions. The subjects were instructed to perform contractions at a moderate and repeatable force level and given rest periods between trials to avoid fatigue.
This set of contractions was repeated during three sessions, each involving a different form of positional variation.
SESSION 1: STATIC POSITIONS
The motions were sustained while holding the arm in 1 of the 16 static limb positions shown in Figure 2. It is assumed that these positions cover the work space in which most ADLs are performed. Positions with odd numbers were located on a plane parallel to the sagittal plane, passing through the subject’s humerus, and positions with even numbers were located on the sagittal plane. P3 to P6 were carried out with the elbow bent, P9 to P16 were carried out with the elbow straight, and the rest were carried out somewhere in between.
To ensure that all subjects moved their arms to the same set of 16 positions, the subjects were asked to stand in front of a white board, on which a grid of eight cells corresponding to positions P1 to P8 was drawn, and move the limb to reach the center of each cell. When data from all eight of these positions were collected, the board was moved away from the subject to elicit the other eight positions (P9–P16), and data were again collected from these positions. Before each session, the height of the board and the spatial distribution of the cells were normalized to the height and reach of the subject.
SESSION 2: ACTIVITIES OF DAILY LIVING
Perhaps a more meaningful assessment of the usability of a control system is its accuracy while performing functional tasks: ADLs. The ADLs shown in Figure 3 were completed while holding each of the eight classes of motion:
- A1: P13 to P5 (tabletop to drink)
- A2: P1 to P13 (neutral to tabletop)
- A3: P1 to P15 (neutral to cupboard)
Contraction classes were held for 4 seconds during performing ADLs, with 3-second interclass delays.
SESSION 3: DYNAMIC TRAINING MOTION
Rather than have the users move the arm into each of the static positions, one dynamic motion was empirically chosen that would encompass the regions of interest. The subjects were asked to smoothly move the arm through a circular trajectory in the plane of the humerus, shown in Figure 4, while holding each of the classes of motion. Contraction classes were sustained for 15 seconds while performing the dynamic motion, with a 3-second interclass delay.
Four sets of contractions were collected for each of the static positions and the dynamic motion. Two of these sets were used for training and two were used for testing.
The mean duration of the experiment was approximately 120 minutes per subject (approximately 30 minutes for setup, 60 minutes for static tasks, 15 minutes for ADLs, and 15 minutes for dynamic motion). Consequently, some patients noted minor shoulder (deltoid) fatigue, although no forearm fatigue was noted.
A control scheme, previously described by Englehart and Hudgins,3 was used in this study. They showed that a simple time-domain (TD) feature extraction combined with a linear discriminant analysis (LDA) classifier could be used as an effective real-time control scheme for myoelectric control. Because of its relative ease of implementation and high performance, this system has been widely accepted19 and was therefore adopted in the present study.
Electromyography data were digitally notch filtered at 60 Hz using a third-order Butterworth filter to remove any power line interference. Data were segmented for feature extraction using 200-millisecond windows, with processing increments of 100 milliseconds. Time-domain features were extracted from the EMG data, and the mean value of the ACC data was calculated for each processing window.
APPROACHES TO REDUCE THE POSITION EFFECT
When a classifier is trained with the limb in one position, the EMG of each motion class shapes a cluster in feature space. Ideally, the EMG patterns of a testing data set collected in the same position should coincide with those of the training clusters. In this case, given that the classes are distinct, the classifier would be capable of correctly identifying those patterns. When the limb position changes, the EMG may be affected and the location of the resulting features might be different from those of the training clusters, as illustrated in Figure 5. Therefore, adding data from multiple positions to the testing data set will have the effect of decreasing the repeatability, as defined by Radmand et al.17 To avoid this problem, data from several limb positions can be incorporated into the training data set.
In this section, classifiers were trained in one position to begin with, and more positions were successively added for training until all 16 positions were represented in the training data set. A brute force method was applied, meaning that, for each distinct number of training positions, the results were acquired for every possible subset of the 16 positions, and the mean was computed. To compare this approach with other methods proposed in previous works, testing the classifier using data from all positions was considered.
To investigate the effect of providing positional context to the classifier, through adding ACC data to the control scheme, four different strategies shown in Table 1 were used for training the classifier using data from multiple positions.
To achieve robust classification, EMG feature patterns of different motion classes need to be not only repeatable but also separable.17 As explained in the previous section, patterns of data collected in different positions may not coincide in the feature space. Therefore, by adding data from more positions to a data set, the variance of the clusters increases. As shown by Radmand et al.,17 this effect reduces class separability and may affect the performance of the classifier. To solve this problem, a dual-stage classification scheme can be applied.
In a dual-stage classification approach, as previously described by Scheme et al.,8 multiple positions are still involved in training, but their data are used to train multiple position-specific classifiers. Therefore, training clusters within each classifier are smaller in size and more separable in feature space, as illustrated in Figure 6.
This multiclassifier approach requires that, first, the originating limb position of a given test sample be determined using ACC features. Once this is known, an EMG motion classifier, which is trained using data from only the detected position, is used to classify the test sample.
It is hypothesized, in this work, that, to effectively integrate ACC data with EMG signals to provide the classifier with positional information about the limb, the training data set should contain examples from most of the possible limb positions. Training the classifier in all possible positions imposes a rather extensive training session for the user and, therefore, is not clinically practical. Moreover, in reality, the limb is used during dynamic ADLs in addition to static tasks. It was hypothesized by Scheme et al.5 that dynamic training would be a more effective paradigm for training a system, especially when the system would be used for performing ADLs. However, they applied this dynamic training approach using only EMG signals. In this work, we applied dynamic training for the integration of ACC data and EMG signals.
Accelerometry features and therefore patterns collected while doing dynamic activities are different from those collected while doing static tasks. This is because ACC data have the gravity component only while performing static tasks, but it will have the movement component as well during dynamic activities. To have comparable ACC feature patterns for both static and dynamic tasks, only the gravity component of ACC data was kept. For this aim, ACC data were low-pass filtered before being integrated with EMG signals.
To investigate the effect of adding ACC data as an additional input to myoelectric pattern recognition, the four strategies shown in Table 1 were used and the results were compared.
STATIC TRAINING RESULTS
Two approaches to resolve the positions effect were compared in this work: multiple-position training and dual-stage classification. These were performed when the classifier was trained using data from only a subset (1–16) of possible static positions but was tested in all positions. The results are shown in Figure 7A, in which the performance was evaluated using classification error. As a baseline for comparison, the mean classification error, when the classifier is trained using data from only 1 static position but is tested in all 16 positions, is also shown in the plot. Figures 7B and C show how different the results of each method are when only one of the accelerometers is used compared with when both of them are used.
A multivariate analysis of variance (ANOVA) was completed using the classification error for both methods. A general linear model was used with subject as a random factor and method as a fixed factor. The results suggest that multiple-position training using only EMG data significantly (p < 0.05) outperforms dual-stage classification for every given number of training positions.
To study the value of increasing the number of accelerometers, an ANOVA test was completed using the classification error for the four single-stage training strategies, shown in Table 1. A general linear model was used with subject as a random factor and training strategy as a fixed factor. The ANOVA showed that, for fewer than five training positions, adding ACC data from either one or two accelerometers significantly (p < 0.05) increases the classification error. In addition, for five training positions, using ACC data from both accelerometers has the same effect. The addition of a humeral accelerometer never significantly improved the results over the forearm ACC alone. However, integrating ACC data from the forearm accelerometer for more than 10 training positions, and from both accelerometers for more than 9, significantly reduced (p < 0.05) the classification error, compared with using EMG data only.
Figure 8 shows the effect of adding ACC data to EMG features on the classifier performance in the optimal/extreme case, when the classifier is trained in all positions.
DYNAMIC TRAINING RESULTS
Figure 9 shows a comparison of single-stage static position and dynamic movement training methods, with and without using ACC data as a complementary source of information. Static training was completed in two ways: one using data from all possible limb positions (as the best possible case for static training) and one using data from only three positions to equalize the training time with that of the dynamic training method. For the latter case, data from positions P1, the neutral position; P3, at which the elbow was bent; and P13, at which the arm was straight, were used for training. Regardless of the method of training, the classifiers were tested using static activities, ADLs, and dynamic movements, as previously described. The baseline comparison was the mean classification error for classifiers trained with data from only one static position.
The results in Figure 9 and an ANOVA test suggest that, in all test cases, both static training using data from all positions and dynamic training significantly (p < 0.05) reduce the position effect. Static training using data from only three positions, however, does not always make a clear improvement and, when combining ACC data with EMG signals, may even degrade the results.
In addition, a comparison of dynamic training with static training using data from all positions shows that, although static training significantly outperforms dynamic training when tested with static tasks, dynamic training is significantly better when tested with dynamic tasks. None of these methods hold a clear advantage over the other one when tested with ADLs. Dynamic training, however, takes much shorter time for training. This method of data collection reduced the training time to 18 seconds for each class of motion, compared with the static data collection in all 16 positions, which required 96 seconds (holding the arm for 3 seconds in each position, with a 3-second interposition delay) to perform each class.
Finally, Figure 9 shows that, for both dynamic and static training in all positions, adding ACC data from only the forearm accelerometer or from both accelerometers is significantly better than using only the upper-arm accelerometer or using no accelerometer. However, using both accelerometers does not hold a significant advantage over using only the forearm accelerometer. It should be noted that the forearm accelerometer could be easily implemented into an existing transradial socket design and should therefore be preferred.
The results in Figure 7 demonstrate that, although dual-stage classification minimizes the position effect through the use of position-specific classifiers, it never outperforms the multiple-position training method for classification. Considering the fact that dual-stage classification also needs more hardware as well as processing time and memory because of its dual-stage nature, this method does not present a preferred method of control. Radmand et al.17 showed that, assuming perfect position classification, dual-stage classification improves the classifier performance by increasing separability between the motion classes in the feature space. However, at the same time, it decreases repeatability between training and test data.17 Therefore, the results of the current work suggest that the benefit gained in increasing separability by doing position-specific classification is outweighed by the reduction in repeatability and also by any error incurred in positions classification using ACC data.
The results in Figure 7 also suggest that integrating ACC data with EMG signals in a single-stage classification approach actually degrades the classifier performance when the training data set contains data from only a small subset of possible limb positions (fewer than six in this case). As the number of positions included in the training data set increases, and the class boundaries become defined, the added information can significantly improve the results. To explain the degradation observed with few positions, two important factors should be studied: the separability between the classes as well as the coincidence between the training and test data sets, before and after addition of ACC data. Figure 10 illustrates an example of the changes in the feature space with and without addition of ACC data, when only a few positions are included in the training data set but all the positions are used for test.
As it can be seen in the figure, the difference between training and test data clusters is minor when using EMG data. This is because the effect of limb position on the EMG is substantially less than the effect of the deliberate motion class contraction. Consequently, the trained classification boundaries still apply to the test data, to a good extent. However, when using combined ACC and EMG data, the training and test data sets are very different. This is because, unlike in the EMG case, in which the position information is secondary to the class information (EMG was recorded from the forearm only, not the shoulder complex or the biceps), the primary information gained from the ACC is the actual position itself. This large effect of position on the ACC features (relative to the small effect of movement class) introduces an unpredictable bias in the classifier boundaries, unless all positions are represented in the training set. Therefore, although the classes are more separable using integration of ACC and EMG data compared with using EMG only, the trained classification boundaries are highly different from the optimum boundaries. This confuses the classifier and degrades the classification performance. By increasing the number of training positions, however, training and test data will coincide with each other. At that point, higher separability between the classes, resulting from the additional information provided by ACC data, improves the overall performance of the classifier. Therefore, unless the classifier is provided with examples of ACC data from adequate numbers of the positions, its inclusion obscures the class-specific differences in EMG and confuses the classifier.
Although training a classifier using data from multiple positions considerably lengthens the training session, the results in Figure 9 demonstrate that dynamic training might be an effective compromise to this problem (especially if the control is used during dynamic ADLs). These results show that, through the dynamic training method, ACC data from most of the possible limb positions can be collected in a reasonable time, which then makes it a useful complementary control input in addition to EMG.
Finally, the results of this work indicate that, although use of ACC data can provide useful information about the limb position and motion, it does not seem to always be a wise solution to the position effect problem. It is important to understand the function of accelerometers properly and to modify the training protocol accordingly to achieve the desired results.
There is a significant body of research describing the use of myoelectric pattern recognition to control upper-limb prostheses. However, the clinical robustness of this approach is still being improved. There are several factors that must be accommodated to achieve truly robust performance of myoelectric control in real task-oriented use. Limb position change for performing dynamic activities is an example factor that affects the robustness of myoelectric pattern recognition. Many researchers have tried to resolve the position effect problem by combining ACC data and EMG. In this work, we studied two commonly proposed methods of integrating ACC data with EMG and investigated the efficacy of those methods when the classifier is trained only in a subset of possible limb positions. It was shown that, unless the classifier is trained in most of the positions, this approach can significantly degrade performance. Training the classifier in several positions, however, is not clinically practical because of the excessive training time. A practical dynamic training method was proposed that moves the residual limb through the regions of interest. This method was shown to minimize training time while improving performance when combining ACC and EMG data.
1. Merletti R, Parker P. Electromyography: Physiology, Engineering, and Noninvasive Applications
. Hoboken, NJ: IEEE Press; Wiley-Interscience; 2004.
2. Hudgins B, Parker P, Scott RN. A new strategy for multifunction myoelectric control. IEEE Trans Biomed Eng
1993; 40( 1): 82–94.
3. Englehart K, Hudgins B. A robust, real-time control scheme for multifunction myoelectric control. IEEE Trans Biomed Eng
2003; 50( 7): 848–854.
4. Hargrove L, Losier Y, Lock B, et al. A real-time pattern recognition based myoelectric control usability study implemented in a virtual environment. IEEE Eng Med Biol Mag
2007; 1: 4842.
5. Scheme E, Biron K, Englehart K. Improving myoelectric pattern recognition positional robustness using advanced training protocols. Paper presented at: the Engineering in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the IEEE
6. Biron K, Englehart K. EMG pattern recognition adaptation. Paper presented at: the XVIII Congress of the International Society for Electrophysiology and Kinesiology
; 2010; Aalborg, Denmark.
7. Young AJ, Hargrove LJ, Kuiken TA. Improving myoelectric pattern recognition robustness to electrode shift by changing interelectrode distance and electrode configuration. IEEE Trans Biomed Eng
2012; 59( 3): 645–652.
8. Scheme E, Fougner A, Stavdahl Ø, et al. Examining the adverse effects of limb position on pattern recognition based myoelectric control. Paper presented at: the Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE
9. Fougner A, Scheme E, Chan ADC, et al. Resolving the limb position effect
in myoelectric pattern recognition. IEEE Trans Neural Syst Rehabil Eng
2011; 19( 6): 644–651.
10. Howard JD, Hoit JD, Enoka RM, Hasan Z. Relative activation of two human elbow flexors under isometric conditions: a cautionary note concerning flexor equivalence. Exp Brain Res
1986; 62( 1): 199–202.
11. Jamison JC, Caldwell GE. Muscle synergies and isometric torque production: influence of supination and pronation level on elbow flexion. J Neurophysiol
1993; 70( 3): 947–960.
12. Mountjoy K, Morin E, Hashtrudi-Zaad K. Use of the fast orthogonal search method to estimate optimal joint angle for upper limb hill-muscle models. IEEE Trans Biomed Eng
2010; 57( 4): 790–798.
13. Chen L, Geng Y, Li G. Effect of upper-limb positions on motion pattern recognition using electromyography. Paper presented at: Image and Signal Processing (CISP), 2011 4th International Congress; 2011
14. Cipriani C, Sassu R, Controzzi M, et al. Preliminary study on the influence of inertia and weight of the prosthesis on the EMG pattern recognition robustness. Paper presented at: the 2011 Myoelectric Controls/Powered Prosthetics Symposium
; 2011; University of New Brunswick, Fredericton.
15. Boschmann A, Platzner M. Reducing the limb position effect
in pattern recognition based myoelectric control using a high density electrode array. Paper presented at: the Biosignals and Biorobotics Conference (BRC), 2013 ISSNIP
16. Jiang N, Muceli S, Graimann B, Farina D. Effect of arm position on the prediction of kinematics from EMG in amputees. Med Biol Eng Comput
2013; 51( 1–2): 1–2.
17. Radmand A, Scheme E, Englehart K. A characterization of the effect of limb position on EMG features to guide the development of effective prosthetic control schemes. Paper presented at the Engineering in Medicine and Biology Society, EMBC, 2014 Annual International Conference of the IEEE; 2014.
18. Geng Y, Zhou P, Li G. Toward attenuating the impact of arm positions on electromyography pattern-recognition based motion classification in transradial amputees. J Neuroeng Rehabil
2012; 9: 74.
19. Scheme E, Englehart K. Electromyogram pattern recognition for control of powered upper-limb prostheses: state of the art and challenges for clinical use. J Rehabil Res Dev
2011; 48( 6): 643–660.