- Open Access
Non-weight-bearing neural control of a powered transfemoral prosthesis
© Hargrove et al.; licensee BioMed Central Ltd. 2013
Received: 25 June 2012
Accepted: 6 June 2013
Published: 19 June 2013
Lower limb prostheses have traditionally been mechanically passive devices without electronic control systems. Microprocessor-controlled passive and powered devices have recently received much interest from the clinical and research communities. The control systems for these devices typically use finite-state controllers to interpret data measured from mechanical sensors embedded within the prosthesis. In this paper we investigated a control system that relied on information extracted from myoelectric signals to control a lower limb prosthesis while amputee patients were seated. Sagittal plane motions of the knee and ankle can be accurately (>90%) recognized and controlled in both a virtual environment and on an actuated transfemoral prosthesis using only myoelectric signals measured from nine residual thigh muscles. Patients also demonstrated accurate (~90%) control of both the femoral and tibial rotation degrees of freedom within the virtual environment. A channel subset investigation was completed and the results showed that only five residual thigh muscles are required to achieve accurate control. This research is the first step in our long-term goal of implementing myoelectric control of lower limb prostheses during both weight-bearing and non-weight-bearing activities for individuals with transfemoral amputation.
Lower limb amputation is a major cause of disability for millions of people worldwide. It is estimated that there are over 600,000 major lower limb amputees – transfemoral, transtibial, or hip disarticulated patients – living in the United States . Individuals with a transfemoral amputation comprise approximately half of the major lower limb amputee population and the majority of amputations are caused by dysvascular disease. Transfemoral amputations are often treated most effectively through use of a prosthetic leg.
There are three broad categories of prosthetic knees: 1) mechanically passive legs, 2) microprocessor-controlled mechanically passive legs, and 3) microprocessor-controlled mechanically active devices. Mechanically passive legs are readily available from a number of manufacturers and vary in costs and design complexity. In such devices, the movement of the prosthetic joint(s) relies on the properties of its mechanical components, such as hydraulic valves, linkage systems, pneumatic valves, or sliding joints. Prosthesis users must make compensatory movements with their trunk, pelvis, and residual limb to control the prosthesis; however, no sensors are required. Microprocessor-controlled passive prostheses employ sensors and a microcomputer for closed-loop control [2–4]. A finite-state controller receives kinematic and force information from the sensors attached to the prosthesis, detects the gait phase, and adjusts the mechanical impedance of the knee joint through a hydraulic damper  or by modifying a magnetic field . The desired joint impedance is predetermined in each gait phase based on normal gait studies. Studies show that compared to the conventional passive prosthesis, the computerized prosthesis with varied impedance allows reduced energy consumption, improved smoothness of gait, and decreased hip work production during locomotion [7–9]. Neither mechanically passive, nor microprocessor-controlled passive prosthetic legs can generate positive power. This significantly impairs the ability of these prostheses to restore many natural locomotive functions, including ascending stairs and slopes and walking backwards, all of which require significant net positive power at the knee joint, ankle joint, or both [10–15]. Furthermore, mechanically passive prostheses cannot be automatically repositioned during non-weight-bearing situations; they need to be manually manipulated by the user.
Microprocessor-controlled, mechanically active prosthetic legs have recently become commercially available, and several additional prototypes are in various stages of development [16–19]. These devices use a similar high level control strategy to microprocessor-controlled mechanically passive prostheses; high-level state-based controllers interpret signals recorded from mechanical sensors embedded in the prosthesis or from an orthotic placed on the sound limb. The current state of the device then provides control information to lower-level position, force, torque, or impedance controllers to ensure that the actuator behaves properly. The capability of generating positive mechanical power greatly increases the number of locomotion modes which may be restored to the patient; however, improvements to the control system are required before this potential can be realized.
Surface electromyographic (EMG) signals may be decoded to provide an estimate of neural activation and have been used to extract control information for upper limb prostheses for many decades. EMG signals have also been investigated for use with intent recognition for microprocessor-controlled above and below knee prostheses [20–27]. Huang has shown that surface EMG signals can be decoded from transfemoral amputees to predict up to seven locomotion modes with classification accuracies of approximately 80%-95% . Furthermore, it was also found that information extracted from EMG signals were complementary to information extracted from mechanical sensors. Simple sensor fusion improved the accuracy and responsiveness of mode recognition .
One particular mode which seems to be ideally suited for EMG control is during non-weight-bearing situations. The user may wish to reposition their prosthesis to help dress themselves or prepare for transfers. Researchers have previously proposed using EMG pattern recognition to volitionally control knee movements for a transfemoral prosthesis and noted high classification accuracies for knee flexion and extension [28, 29]. Other signal processing techniques such as Kalman filtering have also been proposed to track knee movements using surface EMG signals measured from the thigh muscles. These previous studies were completed using healthy control subjects and did not consider controlling ankle movements. Ha et al. recently showed that EMG signals from the quadriceps and hamstrings could be reliably decoded using quadratic discriminant analysis and used with a knee impedance paradigm to control an amputee’s knee position . This work showed that patients could accurately track a position command in a virtual environment but it was not implemented on a powered prosthesis. We recently completed preliminary research which showed that both knee and ankle movements could be accurately decoded using EMG signals measured from the residual thigh muscles of transfemoral amputees . That work was limited in that it was only completed with four subjects in a virtual environment. In this contribution, we extend our previous research by considering additional amputee patients and demonstrating that the technique can be used to control a powered knee prosthesis.
A. Experiment 1: Non-weight-bearing Control in a Virtual Environment
Prior to any data collection, a therapist helped each patient to imagine performing the following movements: knee flexion, knee extension, ankle plantar flexion, ankle dorsiflexion, femoral rotation, and tibial rotation (Figure 1). EMG signals were viewed in real-time to ensure that electrodes were making good contact with the skin and signal gain was set at an appropriate level. Custom software—Control Algorithms for Prosthetic Systems (CAPS)—instructed the subjects to perform the movements listed previously. Data were also collected for a relaxation state which corresponded to a no movement command. The order that the trials were collected was not randomized for the collection of the classifier offline data and eight repetitions of 3 s each were collected for each motion. Data from offline repetitions 1–4 were used to train a pattern recognition system, and data from offline repetitions 5–8 were used to compute classification accuracy.
The pattern recognition system used was based on time-domain features extracted from 250 ms analysis windows and classified by a linear discriminant analysis classifier. Classifier decisions were made every 50 ms which allowed for 200 ms to be retained in an overlapping window scheme. This system has been used extensively and been shown to provide good classification performance for upper limb amputees . Furthermore, this combination of feature-set and classifier has shown promise in recognizing ambulation modes during weight-bearing activities and performs equivalently to non-linear artificial neural network classifiers . The algorithm and equations required for computing feature extraction and classification are provided in Appendix A.
The pattern recognition system was evaluated on its ability to provide high classification accuracy and real-time controllability. Classification accuracy was defined as the percentage of offline test data correctly predicted by the pattern recognition system trained with the offline training data. Real-time controllability was evaluated by the patient’s performance when completing a virtual environment motion test . The motion test required subjects to replicate motions displayed on a computer screen while real-time position feedback was provided by a virtual environment avatar. Performance metrics for the motion test included motion completion time, motion completion percentage, and motion selection time . Motion completion time was the elapsed time from movement onset until the subject successfully moved the virtual limb through the complete range of motion. The speed of the virtual limb was normalized such that the patient could move the prosthesis through the range of motion in a minimum time of 1 s when the classifier correctly classified the movement (i.e. corresponding to 20 correct decisions). The trial ended unsuccessfully if a patient could not move the virtual limb through the complete range of movement within 15 s. Motion completion percentage was the number of successfully completed motions divided by the total number of trials. Motion selection time was the elapsed time from movement onset to the first correct movement classification.
The offline training data were used to create two pattern recognition systems. The first pattern recognition system was trained to recognize a 2 degree of freedom (DOF) subset of all the collected movements: knee flexion/extension, ankle plantar flexion/dorsi-flexion and a no movement class. After the system was trained, the patient completed a real-time control test using feedback provided by the virtual limb. The motion test configuration consisted of nine real-time trials of each of the four movements (the no motion class was not tested) which were presented in random order. The second pattern recognition system was trained using the offline training data to recognize all movements. This motion test configuration consisted of 3 real-time trials of each of the eight movements which were presented in a random order.
An additional offline analysis was performed to determine the number of channels required to achieve acceptable pattern recognition accuracy for transfemoral amputees only. Exhaustive searches of the optimal channel subsets ranging from lengths 1–9 were made for the 2 and 4 DOF pattern recognition system . Optimal channel subsets of length m were defined as the subset of channels that produced the highest classification accuracy when only m channels were used as input to the pattern recognition system. Additionally, the classification accuracies were evaluated when only data from the rectus femoris and semitendinosus were used as this configuration most resembled the EMG input to the two channel system described by Ha et al. . Due to time constraints during the experiments, motion tests were not performed for different channel subset lengths.
B. Experiment 2: Non-weight-bearing Control using a Physical Prosthesis
Three of the six transfemoral amputee participants returned to complete a second experiment to evaluate their performance when controlling a powered knee prosthesis. A prosthetist created a transparent test socket for each patient with stainless steel dome electrodes (Motion Control Inc.) embedded on the interior socket wall at the same electrode locations as Experiment 1. A male snap was threaded on the exterior socket wall to provide convenient connections to the same Bagnoli-16 EMG amplifier used in the previous experiment. Once again, data were amplified by a factor of 1000, sampled by a 16 bit analog-to-digital converter and streamed across a CAN bus to CAPS software.
When the prosthesis was initially powered on, the tuning parameters (k and b) were set to 0 such that a joint torque command of 0 Nm was sent to the device while training and testing data were collected. The data were collected using the same procedure as used in Experiment 1 except only knee flexion, knee extension, ankle plantar flexion, and ankle dorsiflexion, and no motion data were collected. Next, the myoelectric impedance control parameters were tuned empirically. The values of k k and b k were slowly adjusted until the subject could move the knee through the full range of motion at a comfortable speed with a smooth kinematic profile. Since the prosthesis did not contain an ankle actuation unit, the ankle tuning parameters, k a and ba, remained at 0. These parameters would also need to be adjusted in order to control an ankle actuation unit. Subjects practiced controlling the knee for several minutes prior to completing motion tests with the physical prosthesis.
The motion tests were very similar to those described in Experiment 1 except that the order of motions was not randomized; knee flexion and extension were tested first. Subjects were cued by the experimenter to perform the appropriate motion and move the knee joint through the full range of motion. Ankle motion tests were completed with the prosthetic knee positioned at 90 degrees of knee flexion (i.e. neutral position when sitting) and at 45 degrees of knee flexion. Testing in the two different positions allowed us to determine if the pattern recognition system could still recognize ankle motions when the knee was repositioned. Feedback was provided to the subject by both the virtual environment and the physical prosthesis: the output of the pattern recognition classifier was displayed on a computer monitor and if the pattern recognition system erroneously decoded a knee command, then the prosthesis would move incorrectly. The performance metrics of the motion tests were motion completion percentage and motion completion time.
A. Experiment 1: Non-weight-bearing Control in a Virtual Environment
Virtual prosthesis performance metrics
Knee F/E and ankle DF/PF
Knee F/E, femoral rotation, ankle DF/PF and ankle rotation
(n = 6)
(n = 6)
(n = 6)
(n = 6)
Classification accuracy, %
Motion selection time, s
Motion completion time, s
Motion completion percentage, %
B. Experiment 2: Non-weight-bearing Control using a Physical Prosthesis
Comparison of physical and virtual prosthesis performance metrics
Classification accuracy, %
Completion time, s
Ankle (knee at 90 deg)
Ankle (knee at 45 deg)
Completion rate, %
Ankle (knee at 90 deg)
Ankle (knee at 45 deg)
Accurate classification and controllability of knee and hip movements was expected because the EMG signals were recorded from physiologically appropriate residual limb muscles. Accurate classification and controllability of ankle motions was unexpected; the muscles that control the ankle are located below the knee and were lost as a result of amputation. Nonetheless, transfemoral amputees were generating distinct and repeatable co-activity patterns even after an average time of 26.3 (17.7) years post-amputation. These patterns were properly interpreted by the pattern recognition system. This is analogous to recognizing subtle differences in hand grasp patterns using only the extrinsic forearm muscles of transradial amputees .
There was no significant difference (p > 0.1) between the performances of the able-bodied control participants and the amputee participants in terms of overall classification accuracy or motion completion rate for the 2 and 4 DOF pattern recognition systems. We believe this is primarily because only above knee muscles were sampled from the control subjects; however, we did expect that the control participants would perform slightly better than the amputee participants because visual and proprioceptive feedback make it easier to perform repeatable movements. There was some evidence (but not statistically significant) to suggest that control subjects completed motions faster than amputee subjects, but the differences in time were very small (<0.5 s). We believe that the therapist’s instructions prior to data collection helped to compensate for the lack of feedback available to the amputee participants. This is consistent with upper-limb pattern recognition work in which we have noted that patient training helps to improve control, especially for novice users [36, 37]. Both amputee patients and control subjects took longer to select motions for the 4 DOF control system. We believe that the patients had to think longer about which motion they were being asked to control in the 4 DOF test.
Previous studies have demonstrated impressive offline classification accuracies [28, 29] or position tracking of knee movements  with healthy control subjects. The capability for transfemoral amputees to control knee movements in a virtual environment has also been investigated but the patients were not wearing the prosthesis . In the current study, only 2 DOFs were evaluated with the physical prosthesis. Saggital plane movements of the knee and ankle were selected because there are currently 2 DOF mechanically powered legs under development. Although these results are preliminary, they are promising. All subjects were able to reliably to control the knee in real time. Furthermore, the pattern recognition system properly interpreted ankle commands when the prosthesis was repositioned to a 45 degree angle, suspending freely in space. This suggests that myoelectric signal changes resulting from dynamic loading on the socket do not degrade lower limb pattern recognition performance. Further testing with additional amputees is required to see if this result can be generalized across subjects. It also should be noted that only changes in the knee angle were tested and not changes in the position of the residual limb.
Proportional control estimates of knee velocity were not incorporated into the control system, and the parameters of the myoelectric impedance controller were adjusted empirically by the experimenter. Proportional control signals may be added by taking a simple average of EMG amplitudes  or by using a weighted average of EMG amplitudes determined by principle component analysis . Smoother kinematic profiles may be obtained by optimizing the selection of the impedance parameters—the objective of ongoing research.
The channel subset analysis showed that only 4 channels of EMG data are required to achieve low classification errors when only 2 DOFs are controlled. Using fewer EMG channels simplifies the socket modification required to collect the signals and reduces the chances of a wire failure. The optimal channel subset was slightly different across the six transfemoral amputees for the 2 and 4 DOF systems. Selecting the most common five muscles, including the semitendiosus, biceps femoris, sartorius, gracilis and tensor fasciae latae, resulted in a system with only slight increases in classification error above the optimal five-channel set. The performance obtained when using only the semitendinosus and rectus femoris channels was significantly different than using the optimal 4–6 channel subsets. This is somewhat to be expected since Ha et al. only investigated 1 DOF, a much more simple system than the current study’s 2 and 4 DOF systems.
The benefits of intuitive non-weight-bearing control are that it allows the patient to reposition the limb for increased comfort or in preparation for a transition. These pattern recognition systems are quite insensitive to muscle crosstalk ; however they are sensitive to electrode shifts with respect to the residual limb , changes in residual limb position , and changes in electrode impedance. These issues may be mitigated by collecting more training data or altering the electrode configuration . The proposed algorithm was only developed for use during non-weight-bearing situations. While this type of control is important, our ultimate goal is to investigate the benefits of using EMG signals to supplement mechanical sensor data in weight-bearing situations. A different signal processing approach, such as making use of neuromuscular mechanical sensor fusion , is necessary for weight-bearing situations, especially when the patient is ambulating, as there may be large movement artifacts on the signal or the electrodes may even lose contact with the socket.
This work demonstrates the feasibility of extracting neural information from above knee muscles that is suitable for controlling both knee and ankle movements of powered lower limb prostheses during non-weight bearing activates. To our knowledge, this is the first time EMG signals from the residual muscles of transfemoral amputees have been used to directly control both knee and ankle movements. Current work is focusing on modifying the powered knee prosthesis control system so that neural information can be used during weight-bearing activities.
Time-domain feature extraction
The time-domain statistics proposed by Hudgins et al. consisted of the MAV, the number of zero crossings, the number of slope sign changes, and the waveform length. These four statistics were computed for each EMG input channel and are concatenated to form a feature vector.
Mean absolute value
Where x(n) is the nth sample in analysis window i, and I is the total number of analysis windows over the entire sampled signal.
Number of zero crossings
The value of ϵ is dependent on the system noise and needs to be selected appropriately by examining the noise levels of the data.
Slope sign changes
The value of ϵ is once again dependent on the noise and should be selected appropriately.
Where Δx n = x n -xn-1, is the difference between consecutive signal samples. The resultant value provides a measure of waveform amplitude, frequency, and duration all within a single parameter.
Linear discriminant analysis (LDA)
for each of the M classes and choosing the maximum value.
After the weights and the offset have been calculated from an appropriate set of training data, feed-forward classification using a LDA is computationally simple.
This work was supported by the US Army’s Telemedicine and Advanced Technology Research Center (TATRC) contract WW81XWH-09-2-0020.
- Ziegler-Graham K, MacKenzie EJ, Ephraim PL, Travison TG, Brookmeyer R: Estimating the prevalence of limb loss in the United States: 2005 to 2050. Arch Phys Med Rehabil 2008, 89: 422-429. 10.1016/j.apmr.2007.11.005View ArticlePubMedGoogle Scholar
- Grimes DL, Flowers WC, Donath M: Feasibility of an active control scheme for above knee prostheses. J Biomech Eng 1977,99(4):215-221. 10.1115/1.3426293View ArticleGoogle Scholar
- Peeraer L, Tilley K, Van der Perre G: A computer controlled knee prosthesis: a preliminary report. J Eng Technol 1988, 13: 134-135.Google Scholar
- Aeyels B, Peeraer L, Vander Sloten J, Van der Perre G: Development of an above-knee prosthesis equipped with a microcomputer-controlled knee joint: first test results. J Biomed Eng 1992, 14: 199-202. 10.1016/0141-5425(92)90052-MView ArticlePubMedGoogle Scholar
- I. Otto Bock Orthopedic Industry: Manual for the 3c100 Otto Bock C-LEG. Germany: Duderstadt; 1998.Google Scholar
- OSSUR: RHEO KNEE. http://www.ossur.com/pages/12702
- Johansson JL, Sherrill DM, Riley PO, Bonato P, Herr H: A clinical comparison of variable-damping and mechanically passive prosthetic knee devices. Am J Phys Med Rehabil 2005, 84: 563-575. 10.1097/01.phm.0000174665.74933.0bView ArticlePubMedGoogle Scholar
- Taylor MB, Clark E, Offord EA, Baxter C: A comparison of energy expenditure by a high level trans-femoral amputee using the Intellegent Prosthesis and conventionally damped prosthetic limbs. Prosthet Orthot Int 1996, 20: 116-121.PubMedGoogle Scholar
- Schmalz T, Blumentritt S, Jarasch R: Energy expenditure and biomechanical characteristics of lower limb amputee gait: the influence of prosthetic alignment and different prosthetic components. Gait Posture 2002, 16: 255-263. 10.1016/S0966-6362(02)00008-5View ArticlePubMedGoogle Scholar
- Winter DA: The biomechanics and motor control of human gait: normal, elderly and pathological. 2nd edition. Kitchner, ON, Canada: Waterloo Biomechanics; 1991.Google Scholar
- Nadeau S, McFadyen BJ, Malouin F: Frontal and sagittal plane analyses of the stair climbing task in healthy adults aged over 40 years: what are the challenges compared to level walking? Clin Biomech 2003, 18: 950-959. 10.1016/S0268-0033(03)00179-7View ArticleGoogle Scholar
- Reiner R, Rabuffetti M, Frigo C: Joint powers in stair climbing at different slopes. Proc IEEE Int Conf Eng Med Biol 1999, 1: 530.Google Scholar
- Prilutsky BI, Petrova LN, Raitsin LM: Comparison of mechanical energy expenditure of joint moments and muscle forces during human locomotion. J Biomech 1996, 29: 405-416. 10.1016/0021-9290(95)00083-6View ArticlePubMedGoogle Scholar
- DeVita P, Torry M, Glover KL, Speroni DL: A functional knee brace alters joint torque and power patterns during walking and running. J Biomech 1996, 29: 583-588. 10.1016/0021-9290(95)00115-8View ArticlePubMedGoogle Scholar
- Nagano A, Ishige Y, Fukashiro S: Comparison of new approaches to estimate mechanical output of individual joints in vertical jumps. J Biomech 1998, 31: 951-955. 10.1016/S0021-9290(98)00094-3View ArticlePubMedGoogle Scholar
- Varol HA, Sup F, Goldfarb M: Multiclass Real-Time Intent Recognition of a Powered Lower Limb Prosthesis. IEEE Trans Biomed Eng 2010, 57: 542-551.PubMed CentralView ArticlePubMedGoogle Scholar
- Holgate MA, Hitt JK, Bellman RD, Sugar TG, Hollander KW: The SPARKy (Spring Ankle with Regenerative kinetics) project: Choosing a DC motor based actuation method. In Biomedical Robotics and Biomechatronics, 2008. BioRob 2008. Scottsdale, AZ, USA: 2nd IEEE RAS & EMBS International Conference; 2008:163-168.Google Scholar
- Martinez-Villalpando EC, Weber J, Elliott G, Herr H: Design of an agonist–antagonist active knee prosthesis, in Biomedical Robotics and Biomechatronics. In BioRob 2008. Scottsdale, AZ, USA: 2nd IEEE RAS & EMBS International Conference; 2008:529-534.Google Scholar
- OSSUR The POWER KNEE. The POWER KNEE http://www.ossur.com/?PageID=14255
- Peeraer L, Aeyels B, Van der Perre G: Development of EMG-based mode and intent recognition algorithms for a computer-controlled above-knee prosthesis. J Biomed Eng 1990, 12: 178-182. 10.1016/0141-5425(90)90037-NView ArticlePubMedGoogle Scholar
- Au S, Berniker M, Herr H: Powered ankle-foot prosthesis to assist level-ground and stair-descent gaits. Neural Netw 2008, 21: 654-666. 10.1016/j.neunet.2008.03.006View ArticlePubMedGoogle Scholar
- Ha K, Varo HA, Goldfarb M: Volitional Control of a Prosthetic Knee Using Surface Electromyography. IEEE Trans Biomed Eng 2011, 58: 144-151.View ArticlePubMedGoogle Scholar
- Jin D, Yang J, Zhang R, Wang R, Zhang J: Terrain Identification for Prosthetic Knees Based on Electromyographic Signal Features. Tsinghua Sci Tech 2006, 11: 74-79. 10.1016/S1007-0214(06)70157-2View ArticleGoogle Scholar
- Huang H, Kuiken T, Lipschutz R: A Strategy for Identifying Locomotion Modes using Surface Electromyography. IEEE Trans Biomed Eng 2009, 56: 65-73.PubMed CentralView ArticlePubMedGoogle Scholar
- Huang H, Zhang F, Hargrove LJ, Zhi D, Rogers DR, Englehart KB: "Continuous Locomotion-Mode Identification for Prosthetic Legs Based on Neuromuscular Mechanical Fusion," Biomedical Engineering . IEEE Transactions 2011, 58: 2867-2875.Google Scholar
- Hargrove L, Simon A, Lipschutz R, Finucane S, Kuiken T: Real-Time Myoelectric Control of Knee and Ankle Motions for Transfemoral Amputees. JAMA 2011, 305: 1442-1444.View ArticleGoogle Scholar
- Dyck WR, Onyshko S, Hobson DA, Winter DA, Quanbury AO Bull Prosthet Res. In A voluntarily controlled electrohydraulic above-knee prosthesis. Spring; 1975:169-186.Google Scholar
- Myers D, Moskowitz GD: Myoelectric Pattern Recognition for Use in the Volitional Control of Above-Knee Prostheses. IEEE Trans Syst Man Cybern 1981, 11: 296-302.View ArticleGoogle Scholar
- Hillstrom H, Moskowitz GD: Robust Intent Recognition for Prosthesis Control. Paris, France: 14th Annual International Conference of the IEEE Engineering in Medicine and Biology Society; 1992:1448-1449.Google Scholar
- Delis A, Carvalho J, Nascimento , Borges G: Myoelectric Knee Angle Estimation Algorithms for Control of Active Transfemoral Leg Prostheses. InTech; 2011.Google Scholar
- Losier Y, Wilson A: Moving Towards an Open Standard: The UNB Prosthetic Device Communication Protoco. Leipzig, Germany: International Society for Prosthetics & Orthotics (ISPO) World Congress; 2010.Google Scholar
- Kuiken TA, Li G, Lock BA, Lipschutz RD, Miller LA, Stubblefield KA, Englehart KB: Targeted muscle reinnervation for real-time myoelectric control of multifunction artificial arms. JAMA 2009, 301: 619-628. 10.1001/jama.2009.116PubMed CentralView ArticlePubMedGoogle Scholar
- Hargrove L, Englehart K, Hudgins B: A Comparison of Surface and Intramuscular Myoelectric Signal Classification. IEEE Trans Biomed Eng 2007, 54: 847-853.View ArticlePubMedGoogle Scholar
- Sup F, Bohara A, Goldfarb M: Design and control of a powered transfemoral prosthesis. Int J Robotics Res 2008, 27: 263-273. 10.1177/0278364907084588View ArticleGoogle Scholar
- Li G, Schultz AE, Kuiken TA: Quantifying pattern recognition-based myoelectric control of multifunctional transradial prostheses. IEEE Trans Neural Syst Rehabil Eng 2010, 18: 185-192.PubMed CentralView ArticlePubMedGoogle Scholar
- Bunderson N, Kuiken T: Quantification of feature space changes with experience during electromyogram pattern recognition control. IEEE Trans Neural Syst Rehabil Eng 2012, 20: 239-246.View ArticlePubMedGoogle Scholar
- Simon A, Lock BA, Stubblefield K: Patient Training for Functional Use of Pattern Recognition-Controlled Prostheses. J Prosthet Orthot 2012, 24: 56-64. 10.1097/JPO.0b013e3182515437PubMed CentralView ArticlePubMedGoogle Scholar
- Hargrove L, Zhou P, Englehart K, Kuiken T: The Effect of ECG Interference on Pattern Recognition Based Myoelectric Control For Targeted Muscle Reinnervated Patients. IEEE Trans Biomed Eng 2009, 56: 2197-2201.View ArticlePubMedGoogle Scholar
- 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. 10.1016/j.bspc.2007.11.005View ArticleGoogle Scholar
- Scheme E, Fougner A, Stavdahl O, Chan A, Englehart K: Examining the adverse effects of limb position on pattern recognition based myoelectric control. Buenos Aires, Argentina: 32nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society; 2010.View ArticleGoogle Scholar
- Young A, Hargrove L, Kuiken T: Improving Myoelectric Pattern Recognition Robustness to Electrode Shift by Changing Interelectrode Distance and Electrode Configuration. IEEE Trans Biomed Eng 2012, 59: 645-652.PubMed CentralView ArticlePubMedGoogle Scholar
- Hudgins B, Parker P, Scot RN: A new strategy for multifunction myoelectric control. IEEE Trans Biomed Eng 1993, 40: 82-94. 10.1109/10.204774View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.