Objective. Myoelectric prostheses use electromyographic (EMG) signals to control movement of prosthetic joints. Clinically available myoelectric control strategies do not allow simultaneous movement of multiple degrees of freedom (DOFs); however, the use of implantable devices that record intramuscular EMG signals could overcome this constraint. The objective of this study was to evaluate the real-time simultaneous control of three DOFs (wrist rotation, wrist flexion/extension, and hand open/close) using intramuscular EMG. Approach. We evaluated task performance of five able-bodied subjects in a virtual environment using two control strategies with fine-wire EMG: (i) parallel dual-site differential control, which enabled simultaneous control of three DOFs and (ii) pattern recognition control, which required sequential control of DOFs. Main results. Over the course of the experiment, subjects using parallel dual-site control demonstrated increased use of simultaneous control and improved performance in a Fitts' Law test. By the end of the experiment, performance using parallel dual-site control was significantly better (up to a 25% increase in throughput) than when using sequential pattern recognition control for tasks requiring multiple DOFs. The learning trends with parallel dual-site control suggested that further improvements in performance metrics were possible. Subjects occasionally experienced difficulty in performing isolated single-DOF movements with parallel dual-site control but were able to accomplish related Fitts' Law tasks with high levels of path efficiency. Significance. These results suggest that intramuscular EMG, used in a parallel dual-site configuration, can provide simultaneous control of a multi-DOF prosthetic wrist and hand and may outperform current methods that enforce sequential control.
One of the hottest topics in rehabilitation robotics is that of proper control of prosthetic devices. Despite decades of research, the state of the art is dramatically behind the expectations. To shed light on this issue, in June, 2013 the first international workshop on Present and future of non-invasive peripheral nervous system (PNS)-Machine Interfaces (MI; PMI) was convened, hosted by the International Conference on Rehabilitation Robotics. The keyword PMI has been selected to denote human-machine interfaces targeted at the limb-deficient, mainly upper-limb amputees, dealing with signals gathered from the PNS in a non-invasive way, that is, from the surface of the residuum. The workshop was intended to provide an overview of the state of the art and future perspectives of such interfaces; this paper represents is a collection of opinions expressed by each and every researcher/group involved in it.
Fine-wire intramuscular electrodes were used to obtain EMG signals from six extrinsic hand muscles associated with the thumb, index, and middle fingers. Subjects' EMG activity was used to control a virtual three-DOF hand as they conformed the hand to a sequence of hand postures testing two controllers: direct EMG control and pattern recognition control. Subjects tested two conditions using each controller: starting the hand from a pre-defined neutral posture before each new posture and starting the hand from the previous posture in the sequence. Subjects demonstrated their ability to simultaneously, yet individually, move all three DOFs during the direct EMG control trials, however results showed subjects did not often utilize this feature. Performance metrics such as failure rate and completion time showed no significant difference between the two controllers.
Human joint impedance is the dynamic relationship between the differential change in the position of a perturbed joint and the corresponding response torque; it is a fundamental property that governs how humans interact with their environments. It is critical to characterize ankle impedance during the stance phase of walking to elucidate how ankle impedance is regulated during locomotion, as well as provide the foundation for future development of natural, biomimetic powered prostheses and their control systems. In this study, ankle impedance was estimated using a model consisting of stiffness, damping and inertia. Ankle torque was well described by the model, accounting for 98 ±1.2% of the variance. When averaged across subjects, the stiffness component of impedance was found to increase linearly from 1.5 to 6.5 Nm/rad/kg between 20% and 70% of stance phase. The damping component was found to be statistically greater than zero only for the estimate at 70% of stance phase, with a value of 0.03 Nms/rad/kg. The slope of the ankle's torque-angle curve-known as the quasi-stiffness-was not statistically different from the ankle stiffness values, and showed remarkable similarity. Finally, using the estimated impedance, the specifications for a biomimetic powered ankle prosthesis were introduced that would accurately emulate human ankle impedance during locomotion.
Targeted muscle reinnervation (TMR) is a surgical technique that creates myoelectric prosthesis control sites for high-level amputees. The electromyographic (EMG) signal patterns provided by the reinnervated muscles are well-suited for pattern recognition control. Pattern recognition allows for control of a greater number of degrees of freedom (DOF) than the conventional, EMG amplitude-based approach. Previous pattern recognition studies have shown benefit in placing electrodes directly over the reinnervated muscles. Localizing the optimal TMR locations is inconvenient and time consuming. In this contribution, we demonstrate that a clinically practical grid arrangement of electrodes yields real-time control performance that is equivalent to, or better than, the site-specific electrode placement for simultaneous control of multiple DOFs using pattern recognition. Additional findings indicate that grid-like electrode arrangement yields significantly lower classification errors for classifiers with a large number of movement classes ( > 9). These findings suggest that a grid electrode arrangement can be effectively used to control a multi-DOF upper limb prosthesis while reducing the time and effort associated with fitting the prosthesis due to clinical localization of control sites on amputee patients.
Pattern recognition (PR) based strategies for the control of myoelectric upper limb prostheses are generally evaluated through offline classification accuracy, which is an admittedly useful metric, but insufficient to discuss functional performance in real time. Existing functional tests are extensive to set up and most fail to provide a challenging, objective framework to assess the strategy performance in real time.
Myoelectric control has been used for decades to control powered upper limb prostheses. Conventional, amplitude-based control has been employed to control a single prosthesis degree of freedom (DOF) such as closing and opening of the hand. Within the last decade, new and advanced arm and hand prostheses have been constructed that are capable of actuating numerous DOFs. Pattern recognition control has been proposed to control a greater number of DOFs than conventional control, but has traditionally been limited to sequentially controlling DOFs one at a time. However, able-bodied individuals use multiple DOFs simultaneously, and it may be beneficial to provide amputees the ability to perform simultaneous movements. In this study, four amputees who had undergone targeted motor reinnervation (TMR) surgery with previous training using myoelectric prostheses were configured to use three control strategies: 1) conventional amplitude-based myoelectric control, 2) sequential (one-DOF) pattern recognition control, 3) simultaneous pattern recognition control. Simultaneous pattern recognition was enabled by having amputees train each simultaneous movement as a separate motion class. For tasks that required control over just one DOF, sequential pattern recognition based control performed the best with the lowest average completion times, completion rates and length error. For tasks that required control over 2 DOFs, the simultaneous pattern recognition controller performed the best with the lowest average completion times, completion rates and length error compared to the other control strategies. In the two strategies in which users could employ simultaneous movements (conventional and simultaneous pattern recognition), amputees chose to use simultaneous movements 78% of the time with simultaneous pattern recognition and 64% of the time with conventional control for tasks that required two DOF motions to reach the target. These results suggest that when amputees are given the ability to control multiple DOFs simultaneously, they choose to perform tasks that utilize multiple DOFs with simultaneous movements. Additionally, they were able to perform these tasks with higher performance (faster speed, lower length error and higher completion rates) without losing substantial performance in 1 DOF tasks.
Common goals in the development of human-machine interface (HMI) technology are to reduce cognitive workload and increase function. However, objective and quantitative outcome measures assessing cognitive workload have not been standardized for HMI research. The present study examines the efficacy of a simple event-related potential (ERP) measure of cortical effort during myoelectric control of a virtual limb for use as an outcome tool. Participants trained and tested on two methods of control, direct control (DC) and pattern recognition control (PRC), while electroencephalographic (EEG) activity was recorded. Eighteen healthy participants with intact limbs were tested using DC and PRC under three conditions: passive viewing, easy, and hard. Novel auditory probes were presented at random intervals during testing, and significant task-difficulty effects were observed in the P200, P300, and a late positive potential (LPP), supporting the efficacy of ERPs as a cognitive workload measure in HMI tasks. LPP amplitude distinguished DC from PRC in the hard condition with higher amplitude in PRC, consistent with lower cognitive workload in PRC relative to DC for complex movements. Participants completed trials faster in the easy condition using DC relative to PRC, but completed trials more slowly using DC relative to PRC in the hard condition. The results provide promising support for ERPs as an outcome measure for cognitive workload in HMI research such as prosthetics, exoskeletons, and other assistive devices, and can be used to evaluate and guide new technologies for more intuitive HMI control.
Powered prostheses are controlled using electromyographic (EMG) signals, which may introduce high levels of uncertainty even for simple tasks. According to Bayesian theories, higher uncertainty should influence how the brain adapts motor commands in response to perceived errors. Such adaptation may critically influence how patients interact with their prosthetic devices; however, we do not yet understand adaptation behavior with EMG control. Models of adaptation can offer insights on movement planning and feedback correction, but we first need to establish their validity for EMG control interfaces. Here we created a simplified comparison of prosthesis and able-bodied control by studying adaptation with three control interfaces: joint angle, joint torque, and EMG. Subjects used each of the control interfaces to perform a target-directed task with random visual perturbations. We investigated how control interface and visual uncertainty affected trial-by-trial adaptation. As predicted by Bayesian models, increased errors and decreased visual uncertainty led to faster adaptation. The control interface had no significant effect beyond influencing error sizes. This result suggests that Bayesian models are useful for describing prosthesis control and could facilitate further investigation to characterize the uncertainty faced by prosthesis users. A better understanding of factors affecting movement uncertainty will guide sensory feedback strategies for powered prostheses and clarify what feedback information best improves control.
Lower limb prostheses that can generate net positive mechanical work may restore more ambulation modes to amputees. However, configuration of these devices imposes an additional burden on clinicians relative to conventional prostheses; devices for transfemoral amputees that require configuration of both a knee and an ankle joint are especially challenging. In this paper, we present an approach to configuring such powered devices. We developed modified intrinsic control strategies--which mimic the behavior of biological joints, depend on instantaneous loads within the prosthesis, or set impedance based on values from previous states, as well as a set of starting configuration parameters. We developed tables that include a list of desired clinical gait kinematics and the parameter modifications necessary to alter them. Our approach was implemented for a powered knee and ankle prosthesis in five ambulation modes (level-ground walking, ramp ascent/descent, and stair ascent/descent). The strategies and set of starting configuration parameters were developed using data from three individuals with unilateral transfemoral amputations who had previous experience using the device; this approach was then tested on three novice unilateral transfemoral amputees. Only 17% of the total number of parameters (i.e., 24 of the 140) had to be independently adjusted for each novice user to achieve all five ambulation modes and the initial accommodation period (i.e., time to configure the device for all modes) was reduced by 56%, to 5 hours or less. This approach and subsequent reduction in configuration time may help translate powered prostheses into a viable clinical option where amputees can more quickly appreciate the benefits such devices can provide.
Human locomotion is a rhythmic task in which patterns of muscle activity are modulated by state-dependent feedback to accommodate perturbations. Two popular theories have been proposed for the underlying embodiment of phase in the human pattern generator: a time-dependent internal representation or a time-invariant feedback representation (i.e., reflex mechanisms). In either case the neuromuscular system must update or represent the phase of locomotor patterns based on the system state, which can include measurements of hundreds of variables. However, a much simpler representation of phase has emerged in recent designs for legged robots, which control joint patterns as functions of a single monotonic mechanical variable, termed a phase variable. We propose that human joint patterns may similarly depend on a physical phase variable, specifically the heel-to-toe movement of the Center of Pressure under the foot. We found that when the ankle is unexpectedly rotated to a position it would have encountered later in the step, the Center of Pressure also shifts forward to the corresponding later position, and the remaining portion of the gait pattern ensues. This phase shift suggests that the progression of the stance ankle is controlled by a biomechanical phase variable, motivating future investigations of phase variables in human locomotor control.
This paper presents the design and experimental implementation of a novel feedback control strategy that regulates effective shape on a powered transfemoral prosthesis. The human effective shape is the effective geometry to which the biological leg conforms - through movement of ground reaction forces and leg joints - during the stance period of gait. Able-bodied humans regulate effective shapes to be invariant across conditions such as heel height, walking speed, and body weight, so this measure has proven to be a very useful tool for the alignment and design of passive prostheses. However, leg joints must be actively controlled to assume different effective shapes that are unique to tasks such as standing, walking, and stair climbing. Using our previous simulation studies as a starting point, we model and control the effective shape as a virtual kinematic constraint on the powered Vanderbilt prosthetic leg with a custom instrumented foot. An able-bodied subject used a by-pass adapter to walk on the controlled leg over ground and over a treadmill. These preliminary experiments demonstrate, for the first time, that effective shape (or virtual constraints in general) can be used to control a powered prosthetic leg.
Recently developed powered lower limb prostheses allow users to more closely mimic the kinematics and kinetics of non-amputee gait. However, configuring such a device, in particular a combined powered knee and ankle, for individuals with a transfemoral amputation is challenging. Previous attempts have relied on empirical tuning of all control parameters. This paper describes modified stance phase control strategies - which mimic the behavior of biological joints or depend on the instantaneous loads within the prosthesis - developed to reduce the number of control parameters that require individual tuning. Three individuals with unilateral transfemoral amputations walked with a powered knee and ankle prosthesis across five ambulation modes (level ground walking, ramp ascent/descent, and stair ascent/descent). Starting with a nominal set of impedance parameters, the modified control strategies were applied and the devices were individually tuned such that all subjects achieved comfortable and safe ambulation. The control strategies drastically reduced the number of independent parameters that needed to be tuned for each subject (i.e., to 21 parameters instead of a possible 140 or approximately 4 parameters per mode) while relative amplitudes and timing of kinematic and kinetic data remained similar to those previously reported and to those of non-amputee subjects. Reducing the time necessary to configure a powered device across multiple ambulation modes may allow users to more quickly realize the benefits such powered devices can provide.
Recently developed lower-limb prostheses are capable of actuating the knee and ankle joints, allowing amputees to perform advanced locomotion modes such as stepover- step stair ascent and walking on sloped surfaces. However, transitions between these locomotion modes and walking are neither automatic nor seamless. This study describes methods for construction and training of a high-level intent recognition system for a lower-limb prosthesis that provides natural transitions between walking, stair ascent, stair descent, ramp ascent, and ramp descent. Using mechanical sensors onboard a powered prosthesis, we collected steady-state and transition data from six transfemoral amputees while the five locomotion modes were performed. An intent recognition system built using only mechanical sensor data was 84.5% accurate using only steady-state training data. Including training data collected while amputees performed seamless transitions between locomotion modes improved the overall accuracy rate to 93.9%. Training using a single analysis window at heel contact and toe off provided higher recognition accuracy than training with multiple analysis windows. This study demonstrates the capability of an intent recognition system to provide automatic, natural, and seamless transitions between five locomotion modes for transfemoral amputees using powered lower limb prostheses.
The simultaneous control of multiple degrees of freedom (DOFs) is important for the intuitive, life-like control of artificial limbs. The objective of this study was to determine whether the use of intramuscular electromyogram (EMG) improved pattern classification of simultaneous wrist/hand movements compared to surface EMG. Two pattern classification methods were used in this analysis, and were trained to predict 1-DOF and 2-DOF movements involving wrist rotation, wrist flexion/extension, and hand open/close. The classification methods used were (1) a single pattern classifier discriminating between 1-DOF and 2-DOF motion classes, and (2) a parallel set of three classifiers to predict the activity of each of the 3 DOFs. We demonstrate that in this combined wrist/hand classification task, the use of intramuscular EMG significantly decreases classification error compared to surface EMG for the parallel configuration (p<0.01), but not for the single classifier. We also show that the use of intramuscular EMG mitigates the increase in errors produced when the parallel classifier method is trained without 2-DOF motion class data.
Few studies have directly compared real-time control performance of pattern recognition to direct control for the next generation of myoelectric controlled upper limb prostheses. Many different implementations of pattern recognition control have been proposed, with minor differentiations in the feature sets and classifiers. An objective and generalizable evaluation tool quantifying the control performance, other than classification accuracy, is needed. This paper used the implementation of such a tool through the design of a target acquisition test, similar to a Fitts law test, relating movement time of the target acquisition to the difficulty of the target, for a given control strategy. Performance metrics such as throughput (bits/sec), completion rate (%) and path efficiency (%) allow for a complete evaluation of the described strategies. We compared direct control and pattern recognition control with the proposed test and found that 1) the test was valid for control system evaluation by following Fitts law with high coefficients of determination for both types of control and 2) that pattern recognition significantly outperformed direct control in throughput with similar completion rates and path efficiencies. In this framework, the present pilot study supports pattern recognition as a promising strategy and forms a basis for the development of a general and objective tool for the performance evaluation of upper limb control strategies.
Pattern recognition myoelectric control shows great promise as an alternative to conventional amplitude based control to control multiple degree of freedom prosthetic limbs. Many studies have reported pattern recognition classification error performances of less than 10% during offline tests; however, it remains unclear how this translates to real-time control performance. In this contribution, we compare the real-time control performances between pattern recognition and direct myoelectric control (a popular form of conventional amplitude control) for participants who had received targeted muscle reinnervation. The real-time performance was evaluated during three tasks; 1) a box and blocks task, 2) a clothespin relocation task, and 3) a block stacking task. Our results found that pattern recognition significantly outperformed direct control for all three performance tasks. Furthermore, it was found that pattern recognition was configured much quicker. The classification error of the pattern recognition systems used by the patients was found to be 16% ±(1.6%) suggesting that systems with this error rate may still provide excellent control. Finally, patients qualitatively preferred using pattern recognition control and reported the resulting control to be smoother and more consistent.
Powered lower limb prostheses, capable of multiple locomotion modes, are being developed for transfemoral amputees. Current devices do not seamlessly transition between modes such as level walking, stairs and slopes. The purpose of this study was to develop an intent recognition system and test its performance across five different modes. A Dynamic Bayesian Network (DBN) was used for classification of neural and mechanical signals while four amputees completed a circuit containing level-walking, ramp ascent, ramp descent, stair ascent and stair descent. Our results indicate that transitional and steady-state stair steps had a high recognition rate (>99%), while ramp steps were significantly more difficult to classify (p<0.01) (13.7% error on transition steps and 1.3% on steady-state steps). With all five modes trained into the same system, the transitional error rate was 11.3%. Transitional error could be reduced by 31% by training the ramp ascent mode as level walking, and 92% by training both ramp ascent and descent as level walking. This is a viable solution when the level-walking mode can accommodate ramp modes which is currently the case with the ramp ascent. The high recognition rates for recognizing stairs shown in this study demonstrates the potential for an intent recognition system using neural information to allow amputees to naturally transition between locomotion modes on powered prostheses.
The clinical application of robotic technology to powered prosthetic knees and ankles is limited by the lack of a robust control strategy. We found that the use of electromyographic (EMG) signals from natively innervated and surgically reinnervated residual thigh muscles in a patient who had undergone knee amputation improved control of a robotic leg prosthesis. EMG signals were decoded with a pattern-recognition algorithm and combined with data from sensors on the prosthesis to interpret the patients intended movements. This provided robust and intuitive control of ambulation--with seamless transitions between walking on level ground, stairs, and ramps--and of the ability to reposition the leg while the patient was seated.
Mechanical and neurological couplings exist between musculotendon units of the human hand and digits. Studies have begun to understand how these muscles interact when accomplishing everyday tasks, but there are still unanswered questions regarding the control limitations of individual muscles. Using intramuscular electromyographic (EMG) electrodes, this study examined subjects ability to individually initiate and sustain three levels of normalized muscular activity in the index and middle finger muscle compartments of extensor digitorum communis (EDC), flexor digitorum profundus (FDP), and flexor digitorum superficialis (FDS), as well as the extrinsic thumb muscles abductor pollicis longus (APL), extensor pollicis brevis (EPB), extensor pollicis longus (EPL), and flexor pollicis longus (FPL). The index and middle finger compartments each sustained activations with significantly different levels of coactivity from the other finger muscle compartments. The middle finger compartment of EDC was the exception. Only two extrinsic thumb muscles, EPL and FPL, were capable of sustaining individual activations from the other thumb muscles, at all tested activity levels. Activation of APL was achieved at 20 and 30% MVC activity levels with significantly different levels of coactivity. Activation of EPB elicited coactivity levels from EPL and APL that were not significantly different. These results suggest that most finger muscle compartments receive unique motor commands, but of the four thumb muscles, only EPL and FPL were capable of individually activating. This work is encouraging for the neural control of prosthetic limbs because these muscles and compartments may potentially serve as additional user inputs to command prostheses.
New computerized and powered lower limb prostheses are being developed that enable amputees to perform multiple locomotion modes. However, current lower limb prosthesis controllers are not capable of transitioning these devices automatically and seamlessly between locomotion modes such as level-ground walking, stairs and slopes. The focus of this study was to evaluate different intent recognition interfaces, which if configured properly, may be capable of providing more natural transitions between locomotion modes. Intent recognition can be accomplished using a multitude of different signals from mechanical sensors on the prosthesis. Since these signals are non-stationary over any given stride, and gait is cyclical, time history information may improve locomotion mode recognition. The authors propose a dynamic Bayesian network classification strategy to incorporate prior sensor information over the gait cycle with current sensor information. Six transfemoral amputees performed locomotion circuits comprising level-ground walking and ascending/descending stairs and ramps using a powered knee and ankle prosthesis. Using time history reduced steady-state misclassifications by over half (p < 0.01), when compared to strategies that did not use time history, without reducing intent recognition performance during transitions. These results suggest that including time history information across the gait cycle can enhance locomotion mode intent recognition performance.
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.
The mechanical properties of human joints (i.e., impedance) are constantly modulated to precisely govern human interaction with the environment. The estimation of these properties requires the displacement of the joint from its intended motion and a subsequent analysis to determine the relationship between the imposed perturbation and the resultant joint torque. There has been much investigation into the estimation of upper-extremity joint impedance during dynamic activities, yet the estimation of ankle impedance during walking has remained a challenge. This estimation is important for understanding how the mechanical properties of the human ankle are modulated during locomotion, and how those properties can be replicated in artificial prostheses designed to restore natural movement control. Here, we introduce a mechatronic platform designed to address the challenge of estimating the stiffness component of ankle impedance during walking, where stiffness denotes the static component of impedance. The system consists of a single degree of freedom mechatronic platform that is capable of perturbing the ankle during the stance phase of walking and measuring the response torque. Additionally, we estimate the platforms intrinsic inertial impedance using parallel linear filters and present a set of methods for estimating the impedance of the ankle from walking data. The methods were validated by comparing the experimentally determined estimates for the stiffness of a prosthetic foot to those measured from an independent testing machine. The parallel filters accurately estimated the mechatronic platforms inertial impedance, accounting for 96% of the variance, when averaged across channels and trials. Furthermore, our measurement system was found to yield reliable estimates of stiffness, which had an average error of only 5.4% (standard deviation: 0.7%) when measured at three time points within the stance phase of locomotion, and compared to the independently determined stiffness values of the prosthetic foot. The mechatronic system and methods proposed in this study are capable of accurately estimating ankle stiffness during the foot-flat region of stance phase. Future work will focus on the implementation of this validated system in estimating human ankle impedance during the stance phase of walking.
Pattern recognition of myoelectric signals for prosthesis control has been extensively studied in research settings and is close to clinical implementation. These systems are capable of intuitively controlling the next generation of dexterous prosthetic hands. However, pattern recognition systems perform poorly in the presence of electrode shift, defined as movement of surface electrodes with respect to the underlying muscles. This paper focused on investigating the optimal interelectrode distance, channel configuration, and electromyography feature sets for myoelectric pattern recognition in the presence of electrode shift. Increasing interelectrode distance from 2 to 4 cm improved pattern recognition system performance in terms of classification error and controllability (p < 0.01). Additionally, for a constant number of channels, an electrode configuration that included electrodes oriented both longitudinally and perpendicularly with respect to muscle fibers improved robustness in the presence of electrode shift (p < 0.05). We investigated the effect of the number of recording channels with and without electrode shift and found that four to six channels were sufficient for pattern recognition control. Finally, we investigated different feature sets for pattern recognition control using a linear discriminant analysis classifier and found that an autoregressive set significantly (p < 0.01) reduced sensitivity to electrode shift compared to a traditional time-domain feature set.
Despite high classification accuracies (~95%) of myoelectric control systems based on pattern recognition, how well offline measures translate to real-time closed-loop control is unclear. Recently, a real-time virtual test analyzed how well subjects completed arm motions using a multiple-degree of freedom (DOF) classifier. Although this test provided real-time performance metrics, the required task was oversimplified: motion speeds were normalized and unintended movements were ignored. We included these considerations in a new, more challenging virtual test called the Target Achievement Control Test (TAC Test). Five subjects with transradial amputation attempted to move a virtual arm into a target posture using myoelectric pattern recognition, performing the test with various classifier (1- vs 3-DOF) and task complexities (one vs three required motions per posture). We found no significant difference in classification accuracy between the 1- and 3-DOF classifiers (97.2% +/- 2.0% and 94.1% +/- 3.1%, respectively; p = 0.14). Subjects completed 31% fewer trials in significantly more time using the 3-DOF classifier and took 3.6 +/- 0.8 times longer to reach a three-motion posture compared with a one-motion posture. These results highlight the need for closed-loop performance measures and demonstrate that the TAC Test is a useful and more challenging tool to test real-time pattern-recognition performance.
In order to provide natural, biomimetic control to recently developed powered ankle prostheses, we must characterize the impedance of the ankle during ambulation tasks. To this end, a platform robot was developed that can apply an angular perturbation to the ankle during ambulation and simultaneously acquire ground reaction force data. In this study, we detail the design of the platform robot and characterize the impedance of the ankle during quiet standing. Subjects were perturbed by a 3° dorsiflexive ramp perturbation with a length of 150 ms. The impedance was defined parametrically, using a second order model to map joint angle to the torque response. The torque was determined using the inverted pendulum assumption, and impedance was identified by the least squares best estimate, yielding an average damping coefficient of 0.03 ± 0.01 Nms/° and an average stiffness coefficient of 3.1 ± 1.2 Nm/°. The estimates obtained by the proposed platform robot compare favorably to those published in the literature. Future work will investigate the impedance of the ankle during ambulation for powered prosthesis controller development.
Prosthetic hands are becoming more advanced and gaining degrees-of-freedom similar to their human counterparts. However, the command interface enabling control of these prostheses needs to be improved for more intuitive functional use. One barrier to using electromyographic (EMG) signals as the command interface is measuring independent muscle control sites in the residual limb. Surface electrodes are commonly used to detect muscle activity in the forearm; however, the measured signals are often comprised of EMG signals from multiple muscles that are close together. This study investigated the suitability of the index and middle finger compartments of the extrinsic muscles as control sites for prostheses using a direct myocontrol interface. Fine-wire intramuscular electrodes were inserted into seven subjects and their ability to achieve isolated activations of each compartment was tested. The results showed five of the six compartments yield signals suitable for independent volitional control. The middle finger compartment of extensor digitorum communis was found to be incapable of isolated contractions and is therefore not recommended as a control site for direct myocontrol prostheses. A cross-correlation threshold was used to verify that simultaneously measured EMG signals were free from crosstalk and were therefore attributed to muscle co-activations.
Myoelectric pattern recognition control can potentially provide upper limb amputees with intuitive control of multiple prosthetic functions. However, the lack of robustness of myoelectric pattern recognition algorithms is a barrier for clinical implementation. One issue that can contribute to poor system performance is electrode shift, which is a change in the location of the electrodes with respect to the underlying muscles that occurs during donning and doffing and daily use. We investigated the effects of interelectrode distance and feature choice on system performance in the presence of electrode shift. Increasing the interelectrode distance from 2 cm to 4 cm significantly (p<0.01) improved classification accuracy in the presence of electrode shifts of up to 2 cm. In a controllability test, increasing the interelectrode distance from 2 cm to 4 cm improved the users ability to control a virtual prosthesis in the presence of electrode shift. Use of an autoregressive feature set significantly (p<0.01) reduced sensitivity to electrode shift when compared to use of a traditional time-domain feature set.
Machine learning methods for interfacing humans with machines is an emerging area. Here we propose a novel algorithm for interfacing humans with powered lower limb prostheses for restoring control of naturalistic gait following amputation. Unlike most previous neural machine interfaces, our approach fuses control information from the user with sensor information from the prosthesis to approximate the closed loop behavior of the unimpaired sensorimotor system. We present a Bayesian framework to control an artificial knee by probabilistically mixing of process state estimates from different Kalman filters, each addressing separate regimes of locomotion such as level ground walking, walking up a ramp, and walking down a ramp. We show its utility as a mode classifier that is tolerant to temporary sensor faults which are frequently experienced in practical applications.
Few studies have focused on proportional control with multi-channel electromyographic (EMG) pattern recognition systems. In a simple proportional control algorithm, movement speed is often calculated by averaging the mean absolute values of all EMG channels. The aim of our study was to compare the performance of two types of pattern recognition control (simple proportional and binary on/off) to direct proportional control. Six EMG channels were collected from non-targeted forearm muscles of four healthy subjects. Subjects were prompted to perform eight medium force isometric repetitions of the following contractions: wrist flexion/extension, wrist pronation/supination, hand open/close, and no movement (rest). Control performances were measured during a one-dimensional position-tracking task using a custom-made graphical user interface. The results show that a simple proportional control algorithm for the pattern recognition system outperformed binary on/off control and was comparable to the performance achieved with direct proportional control.
Movement misclassifications often occur during real-time pattern recognition control. Majority vote and a decision-based velocity ramp are two different post-processing methods that have been suggested to improve real-time control. With majority vote, spurious misclassifications are removed at the expense of an additional controller delay. With a decision-based velocity ramp, the effect of misclassifications is minimized by attenuating movement speed following a change in decision from the classifier. The goal of the study was to determine which, if any, post-processing method improved real-time control above a baseline condition that did not involve post-processing. Five non-amputee subjects controlled a virtual prosthesis in real time using pattern recognition. While performing a challenging target achievement test in a virtual environment, subjects had significantly higher completion rates (p < 0.04) and more direct paths to the target (p < 0.02) while using the velocity ramp than while using majority vote or the control condition. There were no significant differences in completion rate or path efficiency between the majority vote conditions and the control condition (p > 0.6). The benefits of removing misclassifications through majority vote may be offset by the added controller delay. These results highlight the need for real-time performance measures, as methods that have been shown to reduce errors during offline analysis may not improve real-time control.
In this study, we developed an algorithm based on neuromuscular-mechanical fusion to continuously recognize a variety of locomotion modes performed by patients with transfemoral (TF) amputations. Electromyographic (EMG) signals recorded from gluteal and residual thigh muscles and ground reaction forces/moments measured from the prosthetic pylon were used as inputs to a phase-dependent pattern classifier for continuous locomotion-mode identification. The algorithm was evaluated using data collected from five patients with TF amputations. The results showed that neuromuscular-mechanical fusion outperformed methods that used only EMG signals or mechanical information. For continuous performance of one walking mode (i.e., static state), the interface based on neuromuscular-mechanical fusion and a support vector machine (SVM) algorithm produced 99% or higher accuracy in the stance phase and 95% accuracy in the swing phase for locomotion-mode recognition. During mode transitions, the fusion-based SVM method correctly recognized all transitions with a sufficient predication time. These promising results demonstrate the potential of the continuous locomotion-mode classifier based on neuromuscular-mechanical fusion for neural control of prosthetic legs.
Myoelectric pattern recognition systems for prosthesis control are often studied in controlled laboratory settings, but obstacles remain to be addressed before they are clinically viable. One important obstacle is the difficulty of maintaining system usability with socket misalignment. Misalignment inevitably occurs during prosthesis donning and doffing, producing a shift in electrode contact locations. We investigated how the size of the electrode detection surface and the placement of electrode poles (electrode orientation) affected system robustness with electrode shift. Electrodes oriented parallel to muscle fibers outperformed electrodes oriented perpendicular to muscle fibers in both shift and no-shift conditions (p < 0.01). Another finding was the significant difference (p < 0.01) in performance for the direction of electrode shift. Shifts perpendicular to the muscle fibers reduced classification accuracy and real-time controllability much more than shifts parallel to the muscle fibers. Increasing the size of the electrode detection surface was found to help reduce classification accuracy sensitivity to electrode shifts in a direction perpendicular to the muscle fibers but did not improve the real-time controllability of the pattern recognition system. One clinically important result was that a combination of longitudinal and transverse electrodes yielded high controllability with and without electrode shift using only four physical electrode pole locations.
Real-time pattern recognition control is frequently affected by misclassifications. This study investigated the use of a decision-based velocity ramp that attenuated movement speed after a change in classifier decision. The goal was to improve prosthesis positioning by minimizing the effect of unintended movements. Non-amputee and amputee subjects controlled a prosthesis in real-time using pattern recognition. While performing a target achievement test in a virtual environment, subjects had a significantly higher completion rate (p < 0.05) and a more direct path (p < 0.05) to the target with the velocity ramp than without it. Using a physical prosthesis, subjects stacked a greater average number of 1 cubes (p < 0.05) in three minutes with the velocity ramp than without it (76% more blocks for non-amputees; 89% more blocks for amputees). Real-time control using the velocity ramp also showed significant performance improvements above using majority vote. Eighty-three percent of subjects preferred to control the prosthesis using the velocity ramp. These results suggest that using a decision-based velocity ramp with pattern recognition may improve user performance. Since the velocity ramp is a post-processing step, it has the potential to be used with a variety of classifiers for many applications.
Pattern recognition-based control of myoelectric prostheses has shown great promise in research environments, but has not been optimized for use in a clinical setting. To explore the relationship between classification error, controller delay, and real-time controllability, 13 able-bodied subjects were trained to operate a virtual upper-limb prosthesis using pattern recognition of electromyogram (EMG) signals. Classification error and controller delay were varied by training different classifiers with a variety of analysis window lengths ranging from 50 to 550 ms and either two or four EMG input channels. Offline analysis showed that classification error decreased with longer window lengths (p < 0.01 ). Real-time controllability was evaluated with the target achievement control (TAC) test, which prompted users to maneuver the virtual prosthesis into various target postures. The results indicated that user performance improved with lower classification error (p < 0.01 ) and was reduced with longer controller delay (p < 0.01 ), as determined by the window length. Therefore, both of these effects should be considered when choosing a window length; it may be beneficial to increase the window length if this results in a reduced classification error, despite the corresponding increase in controller delay. For the system employed in this study, the optimal window length was found to be between 150 and 250 ms, which is within acceptable controller delays for conventional multistate amplitude controllers.
This paper describes a novel pattern recognition based myoelectric control system that uses parallel binary classification and class specific thresholds. The system was designed with an intuitive configuration interface, similar to existing conventional myoelectric control systems. The system was assessed quantitatively with a classification error metric and functionally with a clothespin test implemented in a virtual environment. For each case, the proposed system was compared to a state-of-the-art pattern recognition system based on linear discriminant analysis and a conventional myoelectric control scheme with mode switching. These assessments showed that the proposed control system had a higher classification error ( p < 0.001) but yielded a more controllable myoelectric control system ( p < 0.001) as measured through a clothespin usability test implemented in a virtual environment. Furthermore, the system was computationally simple and applicable for real-time embedded implementation. This work provides the basis for a clinically viable pattern recognition based myoelectric control system which is robust, easily configured, and highly usable.
Pattern recognition myoelectric control in combination with targeted muscle reinnervation (TMR) may provide better real-time control of upper limb prostheses. Current pattern recognition algorithms can classify movements with an off-line accuracy of approximately 95%. When amputees use these systems to control prostheses, motion misclassifications may hinder their performance. This study investigated the use of a decision based velocity profile that limited movement speed when there was a change in classifier decision. The goal of this velocity ramp was to improve prosthesis positioning by minimizing the effect of unintended movements. Two patients who had undergone TMR surgery controlled either a virtual or physical prosthesis. They completed a Target Achievement Control Test where they commanded a virtual prosthesis into a target posture. Participants showed improved performance metrics of 34% increase in completion rate and 13% faster overall time with the velocity ramp compared to without the velocity ramp. One participant controlled a physical prosthesis and in three minutes was able to create a tower of 1" cubes seven blocks tall with the velocity ramp compared to a tower of only two blocks tall in the control condition. These results suggest that using a pattern recognition system with a decision based velocity profile may improve user performance.
Targeted muscle reinnervation has been introduced as an effective neural machine interface. In the case of a shoulder disarticulation patient, an effective site for a nerve transfer involves the pectoralis muscles, as these perform little useful function with a missing limb. Consequently, the myoelectric signals measured from the reinnervated muscles may be corrupted by a large amount of ECG interference. This paper investigates the effect of ECG upon the accuracy of a pattern-classification-based scheme for myoelectric control of powered upper limb prostheses. The results suggest that ECG interference, at levels typically encountered in a clinical measurement, has little effect upon classification accuracy, but can affect the estimate of myoelectric activity used to convey the velocity of motion (commonly referred to as proportional control). High-pass filtering at approximately 100 Hz appears to effectively mitigate the effect of ECG interference.
Information extracted from multiple channels of the surface myoelectric signal (MES) recording sites can be used as inputs to control systems for powered upper limb prostheses. For small, closely spaced muscles, such as the muscles in the forearm, the detected MES often contains contributions from more than one muscle, the contribution from each specific muscle being modified by the dispersive propagation through the volume conductor between the muscle and the detection points. In this paper, the measured raw MES signals are rotated by class-specific principal component matrices to spatially decorrelate the measured data prior to feature extraction. This "tunes" the data to allow a pattern recognition classifier to better discriminate the test motions. This processing technique was used to significantly ( p < 0.01) reduce pattern recognition classification error for both intact limbed and transradial amputee subjects.
Current state-of-the-art upper limb myoelectric prostheses are limited by only being able to control a single degree of freedom at a time. However, recent studies have separately shown that the joint angles corresponding to shoulder orientation and upper arm EMG can predict the joint angles corresponding to elbow flexion/extension and forearm pronation/ supination, which would allow for simultaneous control over both degrees of freedom. In this preliminary study, we show that the combination of both upper arm EMG and shoulder joint angles may predict the distal arm joint angles better than each set of inputs alone. Also, with the advent of surgical techniques like targeted muscle reinnervation, which allows a person with an amputation intuitive muscular control over his or her prosthetic, our results suggest that including a set of EMG electrodes around the forearm increases performance when compared to upper arm EMG and shoulder orientation. We used a Time-Delayed Adaptive Neural Network to predict distal arm joint angles. Our results show that our networks root mean square error (RMSE) decreases and coefficient of determination (R(2)) increases when combining both shoulder orientation and EMG as inputs.
It is possible to replace amputated limbs with mechatronic prostheses, but their operation requires the users intentions to be detected and converted into control signals sent to the actuators. Fortunately, the motoneurons (MNs) that controlled the amputated muscles remain intact and capable of generating electrical signals, but these signals are difficult to record. Even the latest microelectrode array technologies and targeted motor reinnervation (TMR) can provide only sparse sampling of the hundreds of motor units that comprise the motor pool for each muscle. Simple rectification and integration of such records is likely to produce noisy and delayed estimates of the actual intentions of the user. We have developed a novel algorithm for optimal estimation of motor pool excitation based on the recruitment and firing rates of a small number (2-10) of discriminated motor units. We first derived the motor estimation algorithm from normal patterns of modulated MN activity based on a previously published model of individual MN recruitment and asynchronous frequency modulation. The algorithm was then validated on a target motor reinnervation subject using intramuscular fine-wire recordings to obtain single motor units.
Pattern recognition can provide intuitive control of myoelectric prostheses. Currently, screen-guided training (SGT), in which individuals perform specific muscle contractions in sync with prompts displayed on a screen, is the common method of collecting the electromyography (EMG) data necessary to train a pattern recognition classifier. Prosthesis-guided training (PGT) is a new data collection method that requires no additional hardware and allows the individuals to keep their focus on the prosthesis itself. The movement of the prosthesis provides the cues of when to perform the muscle contractions. This study compared the training data obtained from SGT and PGT and evaluated user performance after training pattern recognition classifiers with each method. Although the inclusion of transient EMG signal in PGT data led to decreased accuracy of the classifier, subjects completed a performance task faster than when compared to using a classifier built from SGT data. This may indicate that training data collected using PGT that includes both steady state and transient EMG signals generates a classifier that more accurately reflects muscle activity during real-time use of a pattern recognition-controlled myoelectric prosthesis.
Advanced upper limb prostheses capable of actuating multiple degrees of freedom (DOFs) are now commercially available. Pattern recognition algorithms that use surface electromyography (EMG) signals show great promise as multi-DOF controllers. Unfortunately, current pattern recognition systems are limited to activate only one DOF at a time. This study introduces a novel classifier based on Bayesian theory to provide classification of simultaneous movements. This approach and two other classification strategies for simultaneous movements were evaluated using nonamputee and amputee subjects classifying up to three DOFs, where any two DOFs could be classified simultaneously. Similar results were found for nonamputee and amputee subjects. The new approach, based on a set of conditional parallel classifiers was the most promising with errors significantly less (p < 0.05) than a single linear discriminant analysis (LDA) classifier or a parallel approach. For three-DOF classification, the conditional parallel approach had error rates of 6.6% on discrete and 10.9% on combined motions, while the single LDA had error rates of 9.4% on discrete and 14.1% on combined motions. The low error rates demonstrated suggest than pattern recognition techniques on surface EMG can be extended to identify simultaneous movements, which could provide more life-like motions for amputees compared to exclusively classifying sequential movements.
The ankle contributes the majority of mechanical power during walking and is a frequently studied joint in biomechanics. Specifically, researchers have extensively investigated the torque-angle relationship for the ankle during dynamic tasks, such as walking and running. The slope of this relationship has been termed the "quasi-stiffness." However, over time, researchers have begun to interchange the concepts of quasi-stiffness and stiffness. This is an especially important distinction as researchers currently begin to investigate the appropriate control systems for recently developed powered prosthetic legs. The quasi-stiffness and stiffness are distinct concepts in the context of powered joints, and are equivalent in the context of passive joints. The purpose of this paper is to demonstrate the difference between the stiffness and quasi-stiffness using a simple impedance-controlled inverted pendulum model and a more sophisticated biped walking model, each with the ability to modify the trajectory of an impedance controllers equilibrium angle position. In both cases, stiffness values are specified by the controller and the quasi-stiffness are shown during a single step. Both models have widely varying quasi-stiffness but each have a single stiffness value. Therefore, from this simple modeling approach, the differences and similarities between these two concepts are elucidated.
It is possible to replace amputated limbs with mechatronic prostheses, but their operation requires the users intentions to be detected and converted into control signals to the actuators. Fortunately, the motoneurons (MNs) that controlled the amputated muscles remain intact and capable of generating electrical signals, but these signals are difficult to record. Even the latest microelectrode array technologies and targeted motor reinnervation can provide only sparse sampling of the hundreds of motor units that comprise the motor pool for each muscle. Simple rectification and integration of such records is likely to produce noisy and delayed estimates of the actual intentions of the user. We have developed a novel algorithm for optimal estimation of motor pool excitation based on the recruitment and firing rates of a small number (2-10) of discriminated motor units. We first derived the motor estimation algorithm from normal patterns of modulated MN activity based on a previously published model of individual MN recruitment and asynchronous frequency modulation. The algorithm was then validated on a target motor reinnervation subject using intramuscular fine-wire recordings to obtain single motor units.
Pattern recognition based control of powered upper limb myoelectric prostheses offers a means of extracting more information from the available muscles than conventional methods. By identifying repeatable patterns of muscle activity across multiple muscle sites rather than relying on independent EMG signals it is possible to provide more natural, reliable control of myoelectric prostheses. The purposes of this study were to (1) determine if participants can perform distinctive muscle activation patterns associated with multiple wrist and hand movements reliably and (2) to show that high density EMG can be applied individually to determine the electrode location of a clinically acceptable number of electrodes (maximally eight) to classify multiple wrist and hand movements reliably in transradial amputees. Eight normally limbed subjects (five female, three male) and four transradial amputee subjects (two traumatic and congenital) subjects participated in this study, which examined the classification accuracies of a pattern recognition control system. It was found that tasks could be classified with high accuracy (85-98%) with normally limbed subjects (10-13 tasks) and with amputees (4-6) tasks. In healthy subjects, reducing the number of electrodes to eight did not affect accuracy significantly when those electrodes were optimally placed, but did reduce accuracy significantly when those electrodes were distributed evenly. In the amputee subjects, reducing the number of electrodes up to 4 did not affect classification accuracy or the number of tasks with high accuracy, independent of whether those remaining electrodes were evenly distributed or optimally placed. The findings in healthy subjects suggest that high density EMG testing is a useful tool to identify optimal electrode sites for pattern recognition control, but its use in amputees still has to be proven. Instead of just identifying the electrode sites where EMG activity is strong, clinicians will be able to choose the electrode sites that provide the most important information for classification.
Related JoVE Video
Journal of Visualized Experiments
What is Visualize?
JoVE Visualize is a tool created to match the last 5 years of PubMed publications to methods in JoVE's video library.
How does it work?
We use abstracts found on PubMed and match them to JoVE videos to create a list of 10 to 30 related methods videos.
Video X seems to be unrelated to Abstract Y...
In developing our video relationships, we compare around 5 million PubMed articles to our library of over 4,500 methods videos. In some cases the language used in the PubMed abstracts makes matching that content to a JoVE video difficult. In other cases, there happens not to be any content in our video library that is relevant to the topic of a given abstract. In these cases, our algorithms are trying their best to display videos with relevant content, which can sometimes result in matched videos with only a slight relation.