Comparison of Surface Vs. Intramuscular EMG for Multifunctional Prosthesis Control

The use of surface electromyograms (EMG) to control a multiple degree-of-freedom prosthesis has been investigated for several decades. A variety of approaches have been employed with groups using different numbers of input channels (1-2), feature extraction methods (2-5) and pattern recognition algorithms (2,6-7). While much work has been done, all of these efforts have used surface EMG as the control signal.

It has been hypothesized that, due to the potential to provide a larger number of independent control sites and the ability to record selectively from muscles of the forearm (in particular the deep muscles), intramuscular EMG should be advantageous for multifunctional prosthesis control (8).

Admittedly, the technology has not existed for chronic intramuscular recordings to be clinically feasible for prosthetic use. The Implantable Myoelectric Sensor (IMES) that is being developed at the Northwestern University Prosthetic Research Laboratory is an implantable sensor that will allow for intramuscular EMGs to be chronically recorded directly from muscles in the forearm (Fig.1) (9).

Figure 1 – A) A rendering of the proposed Implantable Myoelectric Sensor (IMES). B) An illustration of several IMES implanted into individual muscles of the forearm. The IMES will both receive power and transmit the recorded EMG signals without any wires passing through the skin. Used with permission of NUPRL.

We hypothesize and hope to demonstrate that by utilizing intramuscular EMG recordings it will be possible to substantially increase the ability of a multifunctional prosthesis controller to recognize the intended movement of the user, or ‘classification accuracy.’

Both surface and intramuscular EMG data were collected from the forearm of each subject. Ten pairs of fine wire bipolar EMGs were recorded from 10 muscles in the forearm of four subjects. These pairs of wires were separated by approximately 13 mm, to mimic those signals that would be recorded by the IMES sensor. The muscles that were targeted were the extensor carpi radialis, extensor carpi ulnaris, extensor digitorum communis, extensor pollicis longus, flexor carpi radialis, flexor carpi ulnaris, flexor digitorum superficialis, flexor pollicis longus, pronator teres, and supinator. In addition, six surface electrodes were placed in an equally spaced array around the circumference of the forearm (Fig. 2).

Figure 2 - A photograph of the location of the surface electrodes and fine wire insertion sites. The writing on the forearm indicates preliminary markings to assist in locating the fine-wire sites. Used with the permission of NUPRL.

EMG was collected as the subjects produced a series of contractions for each of the movements being investigated: hand close, hand open, pronation, supination, wrist extension and wrist flexion. For each trial the subject would produce four five-second contractions of the same movement, each time starting from and returning to rest. Four of these trials were collected for each movement with two trials used as training data for the pattern recognition system and the other two used to test the efficacy of the system.

The signal amplitude was calculated for each channel. Additionally, because it has been demonstrated that it is possible to significantly increase classification accuracies by presenting EMG signal features in addition to signal amplitude (3-6+8) to the pattern recognition system. In this study we chose to use auto-regressive (AR) parameters as our additional signal feature.

Four sets of input data were created for analysis:

1. Six surface channels: amplitude only

2. Six surface channels: amplitude + AR

3. Ten intramuscular channels: amplitude only

4. Ten intramuscular channels: amplitude + AR

5. Six intramuscular channels: amplitude only

6. Six intramuscular channels: amplitude + AR

Note: The subsets of six of the ten intramuscular channels used in input sets #5 and #6 were selected using a process called multinomial logistic regression.

The classification accuracy was determined by comparing the intended movement of the user to the output of a controller that uses a pattern recognition algorithm called a linear discriminant analyzer (LDA).

The classification accuracies that resulted from the six input sets described above are shown in Figure 3. When utilizing only EMG amplitude the surface data produced the worst classification accuracy of any data set (78.9%). Adding the AR coefficients to the input data set increased the classification accuracy substantially (89.7%).

Figure 3 – Classification accuracies of the surface and intramuscular EMG. The surface recordings are shown in black, the full set of 10 intramuscular recordings are shown in light gray and a subset of six of the ten intramuscular channels are shown in dark gray. The left column of each pair shows classification accuracies when only EMG amplitude is used and the right column shows classification accuracies from using autoregressive coefficients in addition to EMG amplitude. Used with the permission of NUPRL.

When compared to the surface with AR data set the intramuscular data had slightly lower classification accuracies when using only signal amplitude with 6 (87.9%) or 10 (86.8%) channels. It was also demonstrated that the use of the AR parameters again reduced the error substantially when applied to the intramuscular input sets 6 (92.8%) or 10 (93.7) channels.

The increased accuracy that is seen by implanting the electrodes is encouraging. However, these results are based on a comparison of non-targeted surface channels to targeted intramuscular channels and to achieve a truly fair comparison of surface vs. intramuscular EMG we feel we need to target both recordings. In the near future we will be able to compare targeted surface recordings with the targeted intramuscular data.

It was also interesting to note the considerable improvement that is achieved by adding the auto-regressive parameters to each input set. The classifier error was reduced by more than half in two instances with the error being decreased by 51.7% for the surface inputs and 52.8% for the 10 channel intramuscular inputs.

Finally, it was also interesting to observe that the smaller subsets of six intramuscular channels performed as well as the full set of ten channels. This indicates that it may only be necessary to record from a subset of forearm muscles to maximize classification accuracy for this six-class problem. The muscles that were contained in these six-channel subsets were not consistent for each subject however when a fixed subset of six channels was used for all subjects it performed comparably (91.0 %) to the sets that were customized for each subject (92.8%).

Intramuscular and surface EMG data has been collected and preliminary analysis performed on four normal subjects. The analysis shows that adding additional signal features to the input set increases the ability of the classifier to accurately predict the subjects intended movement when the surface or intramuscular EMG data is used. At this point only non-site-specific surface EMG recordings have been obtained and this prevents any firm conclusions from being drawn regarding the abilities of classifiers that use intramuscular versus surface EMG inputs. Future experiments will target the surface EMG sites and allow for a comparison of the classification accuracies of the two approaches.

The authors would like to acknowledge A. Bolu Ajiboye and Dr. Todd Kuiken for their assistance with this work. This work was partially funded by the National Institute on Disability and Rehabilitation Research (NIDRR) of the Department of Education under grant number H133E030030. The opinions contained in this paper are those of the grantee and do not necessarily reflect those of the Department of Education. This work was also supported in part the by the National Institutes of Health (NIBIB/NICHD) under Grant 1 R01 EB01672 and by the Department of Veterans Affairs, Rehabilitation Research and Development Service and is administered through the Jesse Brown Veterans Affairs Medical Center, Chicago, IL.

1. Lawrence, PD, Kadefors, R. Classification of Myoelectric Patterns for the Control of a Prosthesis. The Control of Upper-Extremity Prostheses and Orthoses Springfield, IL: Charles C Thomas, 1973.

2. Hudgins B, Parker P, Scott, RN. A new strategy for multifunction myoelectric control. IEEE Trans Biomed Eng, 1993; 40:1:82-94.

3. Saridis GN, Gootee TP. EMG Pattern-Analysis and Classification for a Prosthetic Arm. IEEE Transactions on Biomedical Engineering, 1982; 29:6:403-412

4. Englehart K, Hudgins B, Parker PA. A wavelet-based continuous classification scheme for multifunction myoelectric control. IEEE Trans Biomed Eng 2001; 48:3:302-311

5. Graupe D, Cline WK, Functional Separation of Emg Signals Via Arma Identification Methods for Prosthesis Control Purposes. IEEE Transactions on Systems Man and Cybernetics 1975; 5:2:252-259.

6. Ajiboye AB, Weir RW. A Heruistic Fuzzy Logic Approach to EMG Pattern Recognition for Multifunctional Prosthesis Control. IEEE Transactions on Neural Systems and Rehabilitation Engineering (In Press) Englehart K, Hudgins B, Parker PA., & Stevenson, M. Classification of the myoelectric signal using time frequency based representations. Med Eng Phys 1999; 21:6-7:431-438

7. Reilly RE., Implantable Devices for Myoelectric Control. Conf. on the Control of Upper-Extremity Prostheses and Orthoses, Goteborg, Sweden: Charles C.

8. Thomas, Springfield, IL, 1973: 22-33. Comparison of Surface Vs Intramuscular EMG For Multifunctional Prosthesis Control Todd Farrell and Richard Weir

9. Weir RF, Troyk PR, DeMichele G, and Kuiken T. Implantable myoelectric sensors (IMES) for upper-extremity prosthesis control – preliminary work. Proceedings of the 25th Silver Anniversary International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), Cancun, Mexico, September 17th – 21st, 2003