No full-text available
To read the full-text of this research,
you can request a copy directly from the author.
Bayesian Filtering of Myoelectric Signals
Abstract
Surface electromyography is used in research, to estimate the activity of muscle, in prosthetic design, to provide a control signal, and in biofeedback, to provide subjects with a visual or auditory indication of muscle contraction. Unfortunately, successful applications are limited by the variability in the signal and the consequent poor quality of estimates. I propose to use a nonlinear recursive filter based on Bayesian estimation. The desired filtered signal is modeled as a combined diffusion and jump process and the measured electromyographic (EMG) signal is modeled as a random process with a density in the exponential family and rate given by the desired signal. The rate is estimated on-line by calculating the full conditional density given all past measurements from a single electrode. The Bayesian estimate gives the filtered signal that best describes the observed EMG signal. This estimate yields results with very low short-time variability but also with the capability of very rapid response to change. The estimate approximates isometric joint torque with lower error and higher signal-to-noise ratio than current linear methods. Use of the nonlinear filter significantly reduces noise compared with current algorithms, and it may therefore permit more effective use of the EMG signal for prosthetic control, biofeedback, and neurophysiology research.
... A novel Bayesian model for EMG has been introduced, portraying motor intent as a stochastic process characterized by abrupt transitions. [41]. This approach reflects a growing trend in harnessing nature-inspired insights to enhance the precision and sophistication of myoelectric control. ...
... The experiment first tested the effect of EMG decoding in the feedforward branch of the controller. Two models were compared: a plain linear filter (3rd order Butterworth low-pass filter, cut-off frequency 1 Hz) and a Bayesian nonlinear filter as formulated in Sanger ( [41], α = 1e -4, ß = 1e -18, 128-level quantization) (Figure 2). The performance between the two filters is visualized in Figure 5; refer to [41] for detailed analyses. ...
... Two models were compared: a plain linear filter (3rd order Butterworth low-pass filter, cut-off frequency 1 Hz) and a Bayesian nonlinear filter as formulated in Sanger ( [41], α = 1e -4, ß = 1e -18, 128-level quantization) (Figure 2). The performance between the two filters is visualized in Figure 5; refer to [41] for detailed analyses. Both feedforward models were applied on rectified EMG signals from the flexor carpi ulnaris of the amputee. ...
Prosthetic hands are frequently rejected due to frustrations in daily uses. By adopting principles of human neuromuscular control, it could potentially achieve human-like compliance in hand functions, thereby improving functionality in prosthetic hand. Previous studies have confirmed the feasibility of real-time emulation of neuromuscular reflex for prosthetic control. This study further to explore the effect of feedforward electromyograph (EMG) decoding and proprioception on the biomimetic controller. The biomimetic controller included a feedforward Bayesian model for decoding alpha motor commands from stump EMG, a muscle model, and a closed-loop component with a model of muscle spindle modified with spiking afferents. Real-time control was enabled by neuromorphic hardware to accelerate evaluation of biologically inspired models. This allows us to investigate which aspects in the controller could benefit from biological properties for improvements on force control performance. 3 able-bodied and 3 amputee subjects were recruited to conduct a “press-without-break” task, subjects were required to press a transducer till the pressure stabilized in an expected range without breaking the virtual object. We tested whether introducing more complex but biomimetic models could enhance the task performance. Data showed that when replacing proportional feedback with the neuromorphic spindle, success rates of amputees increased by 12.2% and failures due to breakage decreased by 26.3%. More prominently, success rates increased by 55.5% and failures decreased by 79.3% when replacing a linear model of EMG with the Bayesian model in the feedforward EMG processing. Results suggest that mimicking biological properties in feedback and feedforward control may improve the manipulation of objects by amputees using prosthetic hands.
... For this reason, 100[ms] is treated as the standard latency (i.e., sampling time) of the myoprocessor in this study. In a previous study, Sanger [12] used a Bayesian filter to estimate torque by solving the Fokker-Planck partial differential equation for rapid response. However, this filter requires the model parameters and works only for previously defined motion. ...
... where T � is the estimated torque, and γ i = r i μ i is a positive constant. From (12), it can be considered that σ d i is decomposed from the measured sEMG signal and that ν i = σ d i ξ i represents muscle activation (i.e., neural drive), which compensates for the distortion of the sEMG signal. In (12), γ i and ξ i can be estimated because the joint torque T and σ d i can be measured by a torque sensor and from the sEMG signal, respectively. ...
... From (12), it can be considered that σ d i is decomposed from the measured sEMG signal and that ν i = σ d i ξ i represents muscle activation (i.e., neural drive), which compensates for the distortion of the sEMG signal. In (12), γ i and ξ i can be estimated because the joint torque T and σ d i can be measured by a torque sensor and from the sEMG signal, respectively. However, γ i and ξ i cannot be uniquely determined by T and σ d i . ...
In the robotics and rehabilitation engineering fields, surface electromyography (sEMG) signals have been widely studied to estimate muscle activation and utilized as control inputs for robotic devices because of their advantageous noninvasiveness. However, the stochastic property of sEMG results in a low signal-to-noise ratio (SNR) and impedes sEMG from being used as a stable and continuous control input for robotic devices. As a traditional method, time-average filters (e.g., low-pass filters) can improve the SNR of sEMG, but time-average filters suffer from latency problems, making real-time robot control difficult. In this study, we propose a stochastic myoprocessor using a rescaling method extended from a whitening method used in previous studies to enhance the SNR of sEMG without the latency problem that affects traditional time average filter-based myoprocessors. The developed stochastic myoprocessor uses 16 channel electrodes to use the ensemble average, 8 of which are used to measure and decompose deep muscle activation. To validate the performance of the developed myoprocessor, the elbow joint is selected, and the flexion torque is estimated. The experimental results indicate that the estimation results of the developed myoprocessor show an RMS error of 6.17[%], which is an improvement with respect to previous methods. Thus, the rescaling method with multichannel electrodes proposed in this study is promising and can be applied in robotic rehabilitation engineering to generate rapid and accurate control input for robotic devices.
... While our approach could be used as a preprocessing strategy to support classification and 85 control, its primary goals are to produce a high-quality estimate of the underlying muscle activation signal(s), 86 appropriate for understanding basic motor control properties of the brain. As such, we extend a large body of 87 previous literature -particularly, linear filtering or Bayesian filtering -which have largely approached this problem 88 by looking at the activity from a single muscle at a time (Clancy et al., 2001;Hofmann et al., 2016;Sanger, 2007). ...
... We found that AutoLFADS 100 functions adaptively to adjust its frequency response characteristics dynamically according to the time course of 101 different phases of behavior. As with other latent variable models Sanger, 2007), we 102 validated our approach by testing and comparing how well the representations inferred by AutoLFADS correlated 103 with behavioral output. We showed that AutoLFADS enables more accurate single-trial joint angular acceleration 104 predictions from EMG than the predictions of standard filtering techniques (i.e., smoothing, Bayesian filtering). ...
... However, processing this rectified signal to extract muscle activation is non-trivial. A wide 57 variety of filtering approaches have been proposed, ranging from a simple linear, low-pass filter (Clancy et al., 2001; 58 D'Alessio and Conforto, 2001;Hogan and Mann, 1980) to nonlinear, adaptive Bayesian filtering approaches that 59 aim to model the time-varying statistics of the EMG Sanger, 2007). Although these are 60 reasonable approaches, determining the optimal parameters for these filters is challenging, and typically results in 61 arbitrary or non-optimal choices. ...
Objective
To study the neural control of movement, it is often necessary to estimate how muscles are activated across a variety of behavioral conditions. However, estimating the latent command signal that underlies muscle activation is challenging due to its complex relation with recorded electromyographic (EMG) signals. Common approaches estimate muscle activation independently for each channel or require manual tuning of model hyperparameters to optimally preserve behaviorally-relevant features.
Approach
Here, we adapted AutoLFADS, a large-scale, unsupervised deep learning approach originally designed to de-noise cortical spiking data, to estimate muscle activation from multi-muscle EMG signals. AutoLFADS uses recurrent neural networks (RNNs) to model the spatial and temporal regularities that underlie multi-muscle activation.
Main Results
We first tested AutoLFADS on muscle activity from the rat hindlimb during locomotion, and found that it dynamically adjusts its frequency response characteristics across different phases of behavior. The model produced single-trial estimates of muscle activation that improved prediction of joint kinematics as compared to low-pass or Bayesian filtering. We also tested the generality of the approach by applying AutoLFADS to monkey forearm muscle activity from an isometric task. AutoLFADS uncovered previously uncharacterized high-frequency oscillations in the EMG that enhanced the correlation with measured force compared to low-pass or Bayesian filtering. The AutoLFADS-inferred estimates of muscle activation were also more closely correlated with simultaneously-recorded motor cortical activity than other tested approaches.
Significance
Ultimately, this method leverages both dynamical systems modeling and artificial neural networks to provide estimates of muscle activation for multiple muscles that can be used for further studies of multi-muscle coordination and its control by upstream brain areas.
... Traditionally, the amplitude of the sEMG signal is estimated using the Root Mean Square (RMS) or Mean Absolute Value (MAV). Bayesian approaches to sEMG amplitude estimation have been demonstrated to outperform RMS and MAV in stability and responsiveness [19]. More recently, different Bayesian estimation approaches have been adopted to enhance simultaneous and proportional control tasks [9], [20]. ...
... 2) Sanger [19]: A recursive Bayesian estimator which updates the posterior probability density of a supposed 'neural drive' with each updated EMG sample. Drive is modelled as a combined jump and diffusion process. ...
... Rectified EMG is modelled as a random process with exponential density. As with previous applications of the filter, EMG was clipped at ±3 standard deviations to avoid modelling extreme values [9], [19]. Based on pilot data α was set to 10 −5 and jump probability β to 10 −50 . ...
Two recursive Bayesian muscle activation estima-tors were compared against standard linear filtering during use of a myoelectric abstract decoder. The decoder was controlled by intrinsic muscles of the hand. In both experiments the linear filter outperformed the Bayesian methods in terms of general score. The Bayesian muscle decoders were faster to respond to changes in muscle activity and show promise for significantly enhancing overall decoder communication rate.
... A non-linear recursive estimator of the EMG signal for online filtering was recently proposed [28] in which the EMG driving signal is modeled as a mixed jump-diffusion stochastic process. The driving signal is recursively estimated by propagating its probability density using the Fokker-Planck equation, followed by an update of the distribution using Bayes' Rule to account for the EMG measurement. ...
... This algorithm has been used in studies of biofeedback [2] and motor control [31,32]. While one might expect the higher signalto-noise ratio of the Bayesian filter [28] to translate into better control, it cannot be assumed that subjects will make appropriate use of the filtering method, nor can we determine a priori whether the magnitude of the improvement will be sufficient to have a significant effect on performance. We need to have a tool that allows us to directly compare different software applications for online EMG filtering. ...
... We can calculate the information rate using the Fitts' Law regressions and use these to assess and compare the controllability of the two different types of filter output. Based on the evidence from 3 the study proposing the Bayesian algorithm [28], we hypothesize that the Bayesian filtering algorithm will permit users to increase speed for a given difficulty and improve success rate by reducing unwanted noise in the EMG signal without sacrificing rapid intentional changes. While the previous study focused on assessing the Bayesian algorithm as a signal processing algorithm for EMG, this study assesses the use of the Bayesian algorithm for online EMG control. ...
Nonlinear Bayesian filtering of surface electromyography (EMG) can provide a stable output signal with little delay and the ability to change rapidly, making it a potential control input for prosthetic or communication devices. We hypothesized that myocontrol follows Fitts’ Law, and that Bayesian filtered EMG would improve movement times and success rates when compared with linearly filtered EMG. We tested the two filters using a Fitts’ Law speed-accuracy paradigm in a one-muscle myocontrol task with EMG captured from the dominant first dorsal interosseous muscle. Cursor position in one dimension was proportional to EMG. Six indices of difficulty were tested, varying the target size and distance. We examined two performance measures: movement time (MT) and success rate. The filter had a significant effect on both MT and success. MT followed Fitts’ Law and the speed-accuracy relationship exhibited a significantly higher channel capacity when using the Bayesian filter. Subjects seemed to be less cautious using the Bayesian filter due to its lower error rate and smoother control. These findings suggest that Bayesian filtering may be a useful component for myoelectrically controlled prosthetics or communication devices.
NEW & NOTEWORTHY Whereas previous work has focused on assessing the Bayesian algorithm as a signal processing algorithm for EMG, this study assesses the use of the Bayesian algorithm for online EMG control. In other words, the subjects see the output of the filter and can adapt their own behavior to use the filter optimally as a tool. This study compares how subjects adapt EMG behavior using the Bayesian algorithm vs. a linear algorithm.
... This leads to a good de-noising effect on both the synthetic signal and the actual signal. Sanger, T.D. [18] proposes a novel recursive algorithm for the on-line Bayesian iteration of the EMG signal surface. The primary advantage of this algorithm is the possibility of smooth output signals without eliminating the possibility of sudden and large value changes. ...
... In order to get an accurate and stable measurement result, we need hundreds of measurements to filter out results, which is inconvenient when the experimental data is incomplete. Inspired by similar methods [12][13][14][15][16][17][18][19], we refer to the theory of Bayesian estimation and use an effective estimation method to improve accuracy in the presence of noise. Since only the frequency of the valley point of the S-parameter is required for demodulation, all of the parameters in the algorithm are related to the frequency of the valley point of S-parameter. ...
This paper proposes a novel iteration Bayesian reweighed (IBR) algorithm to obtain accurate estimates of a measurement parameter that uses only a few noisy measurement data. The method is applied to optimize the frequency fluctuation in an optical carrier-based microwave interferometry (OCMI) system. The algorithm iteratively estimates the frequency of the S-parameter valley point by collecting training samples to rebalance the weights between prior samples, which reduces the impact of noise in the system. Simulation shows that the estimated result of this algorithm is closer to the true value than that of the maximum likelihood estimation (MLE) using the same amount of measured data. Under the influence of system noise, this algorithm optimizes the frequency fluctuation of the S-parameter and reduces the impact of individual measured data. In this study, we applied the algorithm in the strain sensing experiment and compared it with the MLE. When axial strain changes 240 με, the IBR algorithm yields a deviation of 36 με, which is a significant reduction from 138 με (using the MLE method). Moreover, the average error rate decreases from 25% to 3% (with the MLE method), suggesting that the linear fitting degree of the estimated results and accuracy of the system are improved. Moreover, the algorithm has a wide range of applicability, for it can handle different application models in the OCMI system and the systems with frequency fluctuation problems.
... Some of the methods to determine the start of the user command demodulate the recordings to obtain the envelope. These techniques incorporate low pass filtering which was found by Sanger [63] to be a limiting factor because it introduces an inherent delay. He proposed an alternative with the use of a recursive filter using Bayesian estimation approach to identify the background activity from the command associated signal. ...
... To overcome this limitation, Bayesian statistical approach to identify the start of the command was proposed with the advantage over GMM that it does not have to assume Gaussianity. The assumption of the signal property may be a problem because it has been shown that the probability distribution of sEMG at low levels of contraction is non-Gaussian [65] and may be better represented by the exponential model proposed in [63]. ...
... Since the research of Inman et al. (Inman et al., 1952) in 1952, a plethora of studiesutilizing a variety of modeling methods-have related surface electromyogram (EMG) activity to force/torque generated about a joint (Buchanan et al., 2004, Staudenmann et al., 2010. Various strategies have emerged to improve the fidelity of the EMG-force relationship, including: techniques to reduce the variability of the processed EMG (Clancy and Hogan, 1994, 1995, Hashemi et al., 2015, Hogan and Mann, 1980a, b, Parker et al., 2006, Potvin and Brown, 2004, Sanger, 2007, modeling agonist and antagonist muscles about a joint , Clancy and Hogan, 1997, Messier et al., 1971, Solo mo now et al., 1986, applying system identification methods that adapt to each subject , Hasan and Enoka, 1985, Hashemi et al., 2012, Thelen et al., 1994, incorporating dynamic changes in force (Gottlieb and Agarwal, 1971, Hashemi, Morin, 2015, Hashemi, Morin, 2012, and considering variations in joint angle (Doheny et al., 2008, Hashemi et al., 2013, Hof and Van den Berg, 1981, Liu et al., 2013b. These models have been utilized in numerous application areas, such as ergonomics assessment (Hagg et al., 2004, Kumar andMital, 1996), clinical bio mechanics (Disselhorst-Klug et al., 2009, Doorenbosch andHarlaar, 2003) and motor control research (Ostry and Feldman, 2003). ...
... First, we used simple modeling (linear regression) and EMG processing, so as to focus our effort on the novel evaluation of electrode site selection for these hand-wrist tasks. More advanced models are known to reduce error (An, Cooney, 1983, Clancy and Hogan, 1994, 1995, 1997, Dai, Bardizbanian, 2017, Gottlieb and Agarwal, 1971, Hasan and Enoka, 1985, Hashemi, Morin, 2015, Hashemi, Morin, 2012, Hogan and Mann, 1980a, b, Messier, Duffy, 1971, Potvin and Brown, 2004, Sanger, 2007, Solo mo now, Guzzi, 1986, Thelen, Schultz, 1994, and can be added in future work. Second, we studied fixed-posture, dynamic contraction with a bandwidth of 0.75 Hz. ...
Few studies have related the surface electromyogram (EMG) of forearm muscles to two degree of freedom (DoF) hand-wrist forces; ones that have, used large high-density electrode arrays that are impractical for most applied biomechanics research. Hence, we researched EMG-force in two DoFs—hand open-close paired with one wrist DoF—using as few as four conventional electrodes, comparing equidistant placement about the forearm to optimized site selection. Nine subjects produced 1-DoF and 2-DoF uniformly distributed random forces (bandlimited to 0.75 Hz) up to 30% maximum voluntary contraction (MVC). EMG standard deviation (EMGσ) was related to force offline using linear dynamic regression models. For 1-DoF forces, average RMS errors using two optimally-sited electrodes ranged from 8.3 to 9.0 %MVC, depending on the DoF. For 2-DoFs, overall performance was best when training from both 1- and 2-DoF trials, giving average RMS errors using four optimally-sited electrodes of 9.2 %MVC for each DoF pair (hand open-close paired with one wrist DoF). For each model, additional optimally-sited electrodes showed little statistical improvement. Electrodes placed equidistant performed noticeably poorer than an equal number of electrodes that were optimally sited. The results suggest that reliable 2-DoF hand-wrist EMG-force with a small number of electrodes may be feasible.
... In both applications, yaw and pitch's FOD and EMG ABS data of the Myo worn by a performer were mapped to parameters for controlling the intensity of vibrational motors of other Myos worn by the second performer. At TaikaBox's DigiDance workshop 22 , dancers controlled sound and video projections. MM was used to stream yaw, pitch and EMG ABS data to Isadora 23 and Live 24 . ...
... Future development of Myo Mapper will include GUI improvements, including native support for multiple Myos, and a more intuitive organisation of items in the Features Selection window. Future MM releases will implement additional feature extractors such as EMG RMS [10], Bayesian filter [22] and EMG Maximum Voluntary Contraction (MVC) [1]. Improvements will include MIDI and MPE (MIDI Polyphonic Expression) output, allowing the streaming of gestural data to non-OSC audio applications that accept only MIDI messages for external control. ...
Myo Mapper is a free and open source cross-platform application to map data from the gestural device Myo armband into Open Sound Control (OSC) messages. It provides an easy to use tool for musicians to explore the Myo's potential for creating new gesture-based musical interfaces. Together with details of the software, this paper reports on projects realised with the Myo Mapper as well as a qualitative evaluation. We propose guidelines for using Myo data in interactive artworks based on insight gained from the works described and the evaluation. We show that Myo Mapper empowers artists and non-skilled developers to easily take advantage of raw data from the Myo data and work with high-level signal features for the realisation of interactive artistic and musical works. Myo Mapper: 1) Solves an IMU drift problem to allow multimodal interaction; 2) Facilitates an clear workflow for novice users; 3) Includes feature extraction of useful EMG features; and 4) Connects to popular machine learning software for bespoke gesture recognition.
... For isometric contractions, which have numerous applications in rehabilitation, Zhang et al. (2011) developed a torque tracking control for a functional electrical stimulation (FES) device, based on a time-variable black-box ankle model, using the dynamometer-measured torque as the external sensor measurement. Also for the isometric case, Sanger (2007) proposed a nonlinear recursive filter based on Bayesian estimation, that is generalization of KF for nonlinear and non-uniformly distributed variables, to find the signal that best describes neuromuscular excitation from EMGs, but not muscle dynamic states. ...
... The noise contaminating the dynamic process and measurements should be Gaussian. This is not the case for rectified EMG signals, that can be considered, at the best, "Half-Gaussian" (Sanger 2007). Noise, input and measurements are all derived from EMG and presumably present some degree of correlation. ...
State-space control of myoelectric devices and real-time visualization of muscle forces in virtual rehabilitation require measuring or estimating muscle dynamic states: neuromuscular activation, tendon force and muscle length. This paper investigates whether regular (KF) and extended Kalman filters (eKF), derived directly from Hill-type muscle mechanics equations, can be used as real-time muscle state estimators for isometric contractions using raw electromyography signals (EMG) as the only available measurement. The estimators’ amplitude error, computational cost, filtering lags and smoothness are compared with usual EMG-driven analysis, performed offline, by integrating the nonlinear Hill-type muscle model differential equations (offline simulations—OS). EMG activity of the three triceps surae components (soleus, gastrocnemius medialis and gastrocnemius lateralis), in three torque levels, was collected for ten subjects. The actualization interval (AI) between two updates of the KF and eKF was also varied. The results show that computational costs are significantly reduced (70x for KF and 17\(\times \) for eKF). The filtering lags presented sharp linear relationships with the AI (0–300 ms), depending on the state and activation level. Under maximum excitation, amplitude errors varied in the range 10–24% for activation, 5–8% for tendon force and 1.4–1.8% for muscle length, reducing linearly with the excitation level. Smoothness, measured by the ratio between the average standard variations of KF/eKF and OS estimations, was greatly reduced for activation but converged exponentially to 1 for the other states by increasing AI. Compared to regular KF, extended KF does not seem to improve estimation accuracy significantly. Depending on the particular application requirements, the most appropriate KF actualization interval can be selected.
... The Terminal contains an active differential surface electrode to record the EMG activity of the target muscle, and a vibrator motor, so that the feedback occurred directly at the site of the target muscle, making the stimulus salient and relevant (Figure 1(A)). The Control Unit computes the amplitude of the EMG signal of the target muscle through Bayesian estimation (Borish et al., 2018;Sanger, 2007) and actuates a silent vibration motor with a rotation speed and amplitude proportional to the magnitude of the EMG. The fast processor and the use of a nonlinear filter allowed the device to implement online proportional vibratory biofeedback (Bloom et al., 2010;Casellato et al., 2019 for details). ...
Aims
The objective of this case series was to examine the feasibility of vibrotactile EMG-based biofeedback (BF) as a home-based intervention tool to enhance sensory information during everyday motor activities and to explore its effectiveness to induce changes in active ankle range of motion during gait in children with spastic cerebral palsy (CP).
Methods
Ten children ages 6 to 13 years with spastic CP were recruited. Participants wore two EMG-based vibro-tactile BF devices for at least 4 hours per day for 1-month on the ankle and knee joints muscles. The device computed the amplitude of the EMG signal of the target muscle and actuated a silent vibration motor proportional to the magnitude of the EMG.
Results
Our results demonstrated the feasibility of the augmented sensory information of muscle activity to induce changes of the active ankle range of motion during gait for 6 children with an increase ranging from 8.9 to 51.6% compared to a one-month period without treatment.
Conclusions
Preliminary findings of this case series demonstrate the feasibility of vibrotactile EMG-based BF and suggest potential effectiveness to increase active ankle range of motion, therefore serving as a promising therapeutic tool to improve gait in children with spastic CP.
... The neuromorphic chip integrated key models of NRC, which included a motoneuron pool, a skeletal muscle, and an associated muscle spindle. Electromyography (EMG) signal from residual muscle of the individual with amputation was filtered into alpha motor command by adopting a non-linear Bayesian algorithm (Sanger, 2007). Alpha motor command is the entrance to the BC. ...
The human hand has compliant properties arising from muscle biomechanics and neural reflexes, which are absent in conventional prosthetic hands. We recently proved the feasibility to restore neuromuscular reflex control (NRC) to prosthetic hands using real-time computing neuromorphic chips. Here we show that restored NRC augments the ability of individuals with forearm amputation to complete grasping tasks, including standard Box and Blocks Test (BBT), Golf Balls Test (GBT), and Potato Chips Test (PCT). The latter two were more challenging, but novel to prosthesis tests. Performance of a biorealistic controller (BC) with restored NRC was compared to that of a proportional linear feedback (PLF) controller. Eleven individuals with forearm amputation were divided into two groups: one with experience of myocontrol of a prosthetic hand and another without any. Controller performances were evaluated by success rate, failure (drop/break) rate in each grasping task. In controller property tests, biorealistic control achieved a better compliant property with a 23.2% wider range of stiffness adjustment than that of PLF control. In functional grasping tests, participants could control prosthetic hands more rapidly and steadily with neuromuscular reflex. For participants with myocontrol experience, biorealistic control yielded 20.4, 39.4, and 195.2% improvements in BBT, GBT, and PCT, respectively, compared to PLF control. Interestingly, greater improvements were achieved by participants without any myocontrol experience for BBT, GBT, and PCT at 27.4, 48.9, and 344.3%, respectively. The functional gain of biorealistic control over conventional control was more dramatic in more difficult grasp tasks of GBT and PCT, demonstrating the advantage of NRC. Results support the hypothesis that restoring neuromuscular reflex in hand prosthesis can improve neural motor compatibility to human sensorimotor system, hence enabling individuals with amputation to perform delicate grasps that are not tested with conventional prosthetic hands.
... Alpha motor command is at the entrance to the biomimetic controller. sEMG of wrist flexor was filtered into the alpha motor command and scaled to 0 -1 by adopting a nonlinear Bayesian algorithm (drift term: α = 1e -4; jumping term: ß = 1e -18; 128-level quantization) that has proven its advantage in myoelectric control applications [31]. The motor command enters the biomimetic reflex loop activating the motoneuron pool, which is then converted into a muscle force of contraction through muscle model; meanwhile, the calculation of muscle force is constantly adjusted by a biomimetic spindle. ...
Restoring neuromuscular reflex properties in the control of a prosthetic hand may potentially approach human-level grasp functions in the prosthetic hand. Previous studies have confirmed the feasibility of real-time emulation of a monosynaptic spinal reflex loop for prosthetic control [1]. This study continues to explore how well the biomimetic controller could enable the amputee to perform force-control tasks that required both strength and error-tolerance. The biomimetic controller was programmed on a neuromorphic chip for real-time emulation of reflex. The model-calculated force of finger flexor was used to drive a torque motor, which pulled a tendon that flexed prosthetic fingers. Force control ability was evaluated in a “press-without-break” task, which required participants to press a force transducer toward a target level, but never exceeding a breakage threshold. The same task was tested either with the index finger or the full hand; the performance of the biomimetic controller was compared to a proportional linear feedback (PLF) controller, and the contralateral normal hand. Data from finger pressing task in 5 amputees showed that the biomimetic controller and the PLF controller achieved 95.8% and 66.9% the performance of contralateral finger in success rate; 50.0% and 25.1% in stability of force control; 59.9% and 42.8% in information throughput; and 51.5% and 38.4% in completion time. The biomimetic controller outperformed the PLF controller in all performance indices. Similar trends were observed with full-hand grasp task. The biomimetic controller exhibited capacity and behavior closer to contralateral normal hand. Results suggest that incorporating neuromuscular reflex properties in the biomimetic controller may provide human-like capacity of force regulation, which may enhance motor performance of amputees operating a tendon-driven prosthetic hand.
... Force and EMG data were sampled at 1 KHz using an analog-to-digital interface (Power 1401, CED Technologies Inc., UK) and custom data acquisition software. The force data were low-pass filtered (2 nd order Butterworth, 1 Hz cutoff), while a non-linear Bayesian filter (α = 1e-4, β = 1e-18, 128-bin histogram) was applied to the rectified EMG signals [39]. Cursor position was proportional to either the actual force recorded by the transducer (force-control), or the force estimated in real-time from the recorded and rectified EMGs or synergies [27]. ...
The design of myocontrolled devices faces particular challenges in children with dyskinetic cerebral palsy because the electromyographic signal for control contains both voluntary and involuntary components. We hypothesized that voluntary and involuntary components of movements would be uncorrelated and thus detectable as different synergistic patterns of muscle activity, and that removal of the involuntary components would improve online EMG-based control. Therefore, we performed a synergy-based decomposition of EMG-guided movements, and evaluated which components were most controllable using a Fitts’ Law task. Similarly, we also tested which muscles were most controllable. We then tested whether removing the uncontrollable components or muscles improved overall function in terms of movement time, success rate, and throughput. We found that removal of less controllable components or muscles did not improve EMG control performance, and in many cases worsened performance. These results suggest that abnormal movement in dyskinetic CP is consistent with a pervasive distortion of voluntary movement rather than a superposition of separable voluntary and involuntary components of movement.
... Because raw EMG signals are intrinsically noisy, we do not include them in our feature vector. From EMG we calculate gesture power [5] from muscle exertion, by tracking the amplitude envelope from each EMG channel with a Bayesian filter [19] to probabilistically predict the amplitude envelope. In addition to orientation, we take the first-order difference between the current orientation frame and the previous frame (e.g., x d , y d , z d , w d ) to represent the motion of the forearm. ...
This chapter explores three systems for mapping embodied gesture, acquired with electromyography and motion sensing, to sound synthesis. A pilot study using granular synthesis is presented, followed by studies employing corpus-based concatenative synthesis, where small sound units are organized by derived timbral features. We use interactive machine learning in a mapping-by-demonstration paradigm to create regression models that map high-dimensional gestural data to timbral data without dimensionality reduction in three distinct workflows. First, by directly associating individual sound units and static poses (anchor points) in static regression. Second, in whole regression a sound tracing method leverages our intuitive associations between time-varying sound and embodied movement. Third, we extend interactive machine learning through the use of artificial agents and reinforcement learning in an assisted interactive machine learning workflow. We discuss the benefits of organizing the sound corpus using self-organizing maps to address corpus sparseness, and the potential of regression-based mapping at different points in a musical workflow: gesture design, sound design, and mapping design. These systems support expressive performance by creating gesture-timbre spaces that maximize sonic diversity while maintaining coherence, enabling reliable reproduction of target sounds as well as improvisatory exploration of a sonic corpus. They have been made available to the research community, and have been used by the authors in concert performance.
... To obtain the predicted gesture for the current 200ms bin of data, first the top 10 features are calculated (as established in Part 1processing before game play). EMG features are processed through a Bayesian recursive filter [25]. This reduces noise for on-line prediction while remaining sensitive to rapid changes as expected to control the game. ...
Movement-based video games can provide engaging practice for repetitive therapeutic gestures towards improving manual ability in youth with cerebral palsy (CP). However, home-based gesture calibration and classification is needed to personalize therapy and ensure an optimal challenge point. Nineteen youth with CP controlled a video game during a 4-week home-based intervention using therapeutic hand gestures detected via electromyography and inertial sensors. The in-game calibration and classification procedure selects the most discriminating, person-specific features using random forest classification. Then, a support vector machine is trained with this feature subset for in-game interaction. The procedure uses features intended to be sensitive to signs of CP and leverages directional statistics to characterize muscle activity around the forearm. Home-based calibration showed good agreement with video verified ground truths (0.86±0.11, 95%CI=0.93-0.97). Across participants, classifier performance (F1-score) for the primary therapeutic gesture was 0.90±0.05 (95%CI=0.87-0.92) and, for the secondary gesture, 0.82±0.09 (95%CI=0.77-0.86). Features sensitive to signs of CP were significant contributors to classification and correlated to wrist extension improvement and increased practice time. This study contributes insights for classifying gestures in people with CP and demonstrates a new gesture controller to facilitate home-based therapy gaming.
... Simply applying a low-pass filter to the signal to reduce undesired noise may result in the loss of sharp onsets describing rapid movement and may also introduce latency when processing the signal in real time. Adopting a nonlinear recursive filter based on Bayesian estimation [36] significantly reduces the noise while allowing very rapid changes in the signal, greatly improving the quality of the signal for real-time gestural interaction. ...
This chapter presents an overview of Interactive Machine Learning (IML) techniques applied to the analysis and design of musical gestures. We go through the main challenges and needs related to capturing, analysing, and applying IML techniques to human bodily gestures with the purpose of performing with sound synthesis systems. We discuss how different algorithms may be used to accomplish different tasks, including interacting with complex synthesis techniques and exploring interaction possibilities by means of Reinforcement Learning (RL) in an interaction paradigm we developed called Assisted Interactive Machine Learning (AIML). We conclude the chapter with a description of how some of these techniques were employed by the authors for the development of four musical pieces, thus outlining the implications that IML have for musical practice.
... Raw EMG signals are intrinsically noisy, and we do not include them in our feature vector. Instead, we use a Bayesian [8] filter to probabilistically predict the amplitude envelope for each electrode in the armband. The sum of all eight amplitude envelopes is also included in the input feature vector, along with a new feature we have developed called "vector sum." ...
This paper presents a method for mapping embodied gesture , acquired with electromyography and motion sensing, to a corpus of small sound units, organised by derived timbral features using concatenative synthesis. Gestures and sounds can be associated directly using individual units and static poses, or by using a sound tracing method that leverages our intuitive associations between sound and embodied movement. We propose a method for augmenting corporal density to enable expressive variation on the original gesture-timbre space.
... S INCE the research of relationship between surface electromyogram (EMG) and force proposed by Inman et al. in 1952, a diversity of studies have contributed to this research field to improve the precision of the EMG-to-force model [1] [2], which includes: developing models to describe the non-linearity of the relationship [3], establishing models for the joint using agonist and antagonist muscle activations [4] [5] [6] [7], decreasing the variability of the processed EMG signals [8] [9] [10] [11] [12] [13], removing various noises and artifacts [14], considering the influences generated by different joint angles [15], developing individual system identification methods for a specific subject [3] [16] [17] [18], etc. These researches have been used in wide application areas, such as: ergonomics [19], human-robot collaboration systems [20] and motor control [21]. ...
Surface electromyography (sEMG) signal is one of the widely applied biological signals in the research field of the force intention prediction. However, due to the severe cross-talk issue of sEMG signals during fine hand contractions, few studies have related sEMG to multiple degree-of-freedom (DoF) force prediction of individual fingers simultanously. Accordingly, this study proposed methods mainly based on neural networks: Convolutional neural Network (CNN) and Recurrent Neural Network (RNN) to achieve better prediction results. Several improvements on traditional methods are also proposed in this article such as: Common Spatial Pattern (CSP), Softmax function and several new channel selection standards to solve the cross-talk issues for the estimation of EMG-force during multiple finger contractions. High-density sEMG signals of forearm extensor muscles were obtained, and experimental data from seven able-bodied subjects were analyzed. Subjects produced 1-DoF and Multi-DoF forces up to 30% maximum voluntary contraction (MVC). Then, the root-mean-square values of sEMG were related to joint force. To realize a better practical use, the EMG-to-force models were trained with minimal numbers of trials (using 1-DoF trials only), then assessed on multi-DoF trials. Our results showed that the proposed modifications on traditional method also made an improvement on the prediction results. Our findings suggest that Multi-DoF control for individual fingers with minimal training procedure (using 1-DoF trials only) may be feasible for practical use. Furthermore, methods based on neural networks greatly outperform traditional methods and the combination of CNN and LSTM showed the best performance.
... The terminal contains an active differential surface electrode to record the EMG activity of the target muscle, and a vibration motor, so that the feedback occurs directly at the site of the target muscle, making the stimulus salient and relevant. The Control unit computes the amplitude of the EMG signal of the target muscle through Bayesian estimation [27] and actuates a silent vibration motor with a rotation speed and amplitude proportional to the magnitude of the EMG. The fast processor and the use of a nonlinear filter allow the device to implement online proportional biofeedback. ...
Background:
This study is aimed at better understanding the role of a wearable and silent ElectroMyoGraphy-based biofeedback on motor learning in children and adolescents with primary and secondary dystonia.
Methods:
A crossover study with a wash-out period of at least 1 week was designed; the device provides the patient with a vibration proportional to the activation of an impaired target muscle. The protocol consisted of two 5-day blocks during which subjects were trained and tested on a figure-8 writing task: their performances (at different levels of difficulty) were evaluated in terms of both kinematics and muscular activations on day 1 and day 5, while the other 3 days were purely used as training sessions. The training was performed with and without using the biofeedback device: the week of use was randomized. Data were collected on 14 subjects with primary and secondary (acquired) dystonia (age: 6-19 years).
Results:
Results comparing kinematic-based and EMG-based outcome measures pre- and post-training showed learning due to practice for both subjects with primary and secondary dystonia. On top of said learning, an improvement in terms of inter-joint coordination and muscular pattern functionality was recorded only for secondary dystonia subjects, when trained with the aid of the EMG-based biofeedback device.
Conclusions:
Our results support the hypothesis that children and adolescents with primary dystonia in which there is intact sensory processing do not benefit from feedback augmentation, whereas children with secondary dystonia, in which sensory deficits are often present, exhibit a higher learning capacity when augmented movement-related sensory information is provided. This study represents a fundamental investigation to address the scarcity of noninvasive therapeutic interventions for young subjects with dystonia.
... We initially developed the first interaction based on the muscle activity and the acceleration captured by the Myo and sent to a MaxMSP patch. Data is filtered through a a Bayesian filter [55] and used to train a gesture recognition algorithm developed by [22] on the movement qualities of various dance sequences. When the dancer is performing a sequence, the algorithm would recognize its qualities and trigger a corresponding video. ...
I describe the research and creation journey of a choreographic dance piece called SKIN that I made with another choreographer, 3 dancers, 1 musician and 1 developer. The performance integrates interactive technologies mapping inner movement to sound and video on stage. We followed a research though practice method that includes iterative cycles of choreographic practice and interaction design. This generated a set of research questions that I address through experience explicitation interviews of both audience and creative team members. The interviews allow me to investigate the lived experience of making and attending the performance and the emergent relationships between dance, media and interaction as well as the tensions and negotiations that emerged from integrating technology in art. I discuss my approach as anti-solutionist and argue for more openness in HCI to allow artists to contribute to knowledge by embracing the messiness of their practice.
... We calculate the first order differences (x d , y d , z d ) of these angles, which are correlated with direction and speed of displacement, and augment our regression feature vector with historical data. We detect gesture power [4] by tracking muscle exertion, following the amplitude envelope of four (of the Myo's 8) EMG channels with a Bayesian filter [21]. ...
We present a system that allows users to try different ways to train neural networks and temporal modelling to associate gestures with time-varying sound. We created a software framework for this and evaluated it in a workshop-based study. We build upon research in sound tracing and mapping-by-demonstration to ask participants to design gestures for performing time-varying sounds using a multimodal, inertial measurement (IMU) and muscle sensing (EMG) device. We presented the user with two classical techniques from the literature, Static Position regression and Hidden Markov based temporal modelling, and propose a new technique for capturing gesture anchor points on the fly as training data for neural network based regression , called Windowed Regression. Our results show trade-offs between accurate, predictable reproduction of source sounds and exploration of the gesture-sound space. Several users were attracted to our windowed regression technique. This paper will be of interest to musicians engaged in going from sound design to gesture design and offers a workflow for interactive machine learning.
... EMG signals were then normalized to the maximum activation during the movement, resulting in signals ranging from 0 to 1. A nonlinear recursive filter based on Bayesian estimation was applied (Filt) [9]. In order to detect the frequency features related to the motor outcome on the EMG signals, spectral analysis was applied to kinematic (Ytablet and Xtablet) and EMG data (normalized Bayesian filtering outputs) [4]. ...
... Electromyography (EMG) has been widely employed to estimate hand movements in wrist-worn and forearm-worn devices. Using EMG signals from arm muscles as input, various conventional and machine learning techniques have been used for classifying hand gestures [11]- [13] including algorithms using neural networks [14], Bayesian filtering [15], and hidden Markov models [16]. Muscle synergies have also been used by clustering groups of muscle EMG signals related to classify specific hand gestures [17]. ...
This paper presents a new approach to wearable hand gesture recognition and finger angle estimation based on modified barometric pressure sensing. Barometric pressure sensors were encased and injected with VytaFlex rubber such that the rubber directly contacted the sensing element allowing pressure change detection when the encasing rubber was pressed. A wearable prototype consisting of an array of 10 modified barometric pressure sensors around the wrist was developed and validated with experimental testing for three different hand gesture sets and finger flexion/extension trials for each of the five fingers. Overall hand gesture recognition classification accuracy was 94%. Further analysis revealed that the most important sensor location was the underside of the wrist and that when reducing the sensor number to only 5 optimally placed sensors, classification accuracy was still 90%. For continuous finger angle estimation, aggregate R2 values between actual and predicted angles were thumb: 0.81±0.10, index finger: 0.85±0.06, middle finger: 0.77±0.08, ring finger: 0.77±0.12, and pinkie finger: 0.75±0.10, and the overall average was 0.79±0.05. These results demonstrate that a modified barometric pressure wristband can be used to classify hand gestures and to estimate individual finger joint angles. This approach could serve to improve clinical treatment for upper extremity deficiencies, such as for stroke rehabilitation, by providing objective patient motor control metrics to inform and aid physicians and therapists throughout the rehabilitation process.
... Mean Absolute Value (MAV), Root Mean Square (RMS), linear envelope (ENV)) suffer from high variability and strongly depend on the selected time window [20]. Recently, non-linear biological-inspired descriptors of EMG amplitude have been shown to outperform the classical linear estimators [21]- [24]. For instance, the so-called EMGto-Muscle Activation (ACT) is a model-driven feature that has been successfully used in EMG-based joint kinematics reconstruction [25]. ...
A hand impairment can have a profound impact on the quality of life. This has motivated the development of dexterous prosthetic and orthotic devices. However, their control with neuromuscular interfacing remains challenging. Moreover, existing myocontrol interfaces typically require an extensive calibration. We propose a minimally supervised, online myocontrol system for proportional and simultaneous finger force estimation based on ridge regression using only individual finger tasks for training. We compare the performance of this system when using two feature sets extracted from high-density EMG recordings: EMG linear envelope (ENV) and non-linear EMG to Muscle Activation mapping (ACT). Eight intact-limb participants were tested using online target reaching tasks. On average, the subjects hit 85 × 9% and 91 × 11% of single finger targets with ENV and ACT features respectively. The hit rate for combined finger targets decreased to 29 × 16% (ENV) and 53 × 23% (ACT). The non-linear transformation (ACT) therefore improved the performance, leading to higher completion rate and more stable control, especially for the non-trained movement classes (better generalization). These results demonstrate the feasibility of proportional multiple finger control by regression on non-linear EMG features with a minimal training set of single finger tasks.
... The activations of the SOL, GAS, and TA muscles were estimated from the sEMG signals using the activation dynamics model in Krishnaswamy et al. (2011) with diffusion drift rate ( = 50), Poisson jump rate (b = 10 À27 ) (Sanger, 2007), and activation (s act = 9 ms) and deactivation (s deact = 45 ms) time constants taken from Winters and Woo (1990). ...
Surface electromyography driven models are desirable for estimating subject-specific muscle forces. However, these models include parameters that come from an array of sources, thus creating uncertainty in the model-estimated force. In this study, we used Monte-Carlo simulations to evaluate the sensitivity of Hill-based model muscle forces to changes in 11 parameters in the muscle-tendon unit morphological properties and in the model force-length and force-velocity relationships. We decomposed the force variability and ranked the sensitivity of the model to the underlying parameters using the Variogram Analysis of Response Surfaces. For the analyzed running experiments and the adopted Hill model structure, our results show that the parameters are separable into four groups, where the parameters in each group have a synergistic contribution to the model global sensitivity. The first group consists of the maximum isometric force and the pennation angle. The second group contains the optimal fiber length, the tendon slack length, the tendon reference strain and the tendon shape factor. The third group contains the width and shape at the extremities of the active contractile element, along with the maximum contraction velocity and the curvature constant in the force-velocity curve. The fourth group consisted only of the force enhancement during eccentric contraction. The first two groups revealed the largest influence on the output force sensitivity. As many input parameters are difficult to measure and impact estimated forces, we propose that model estimates be presented with confidence intervals as well as inter-parameter relationships, to encourage users to explicitly consider the model uncertainty.
... At the same time, the problem of the correct estimation of sEMG amplitude is still important, as highlighted in recent studies aiming at extracting neural control strategies from the sEMG signal (Schuman et al., 2017). Among all the solutions presented in literature (Clancy, 1999;Fullmer et al., 1984;D'Alessio, 1984;D'Alessio et al., 2001;Sanger, 2007), the adaptive iterative procedure proposed in (D'Alessio et al., 2001) seems to be the most appropriate for both static and dynamic contractions. The key advantage of this algorithm relies on the fact that it is able to find the point by point optimal window length for both the MA and the RMS approach, based on the properties of both the original signal and the estimated envelope at each iteration of the algorithm. ...
Surface ElectroMyography (sEMG) is widely used as a non-invasive tool for the assessment of motor control strategies. However, the standardization of the methods used for the estimation of sEMG amplitude is a problem yet to be solved; in most cases, sEMG amplitude is estimated through the extraction of the envelope of the signal via different low-pass filtering procedures with fixed cut-off frequencies chosen by the experimenter. In this work, we have shown how it is not possible to find the optimal choice of the cut-off frequency without any a priori knowledge on the signal; considering this, we have proposed an updated version of an iterative adaptive algorithm already present in literature, aiming to completely automatize the sEMG amplitude estimation. We have compared our algorithm to most of the typical solutions (fixed window filters and the previous version of the adaptive algorithm) for the extraction of the sEMG envelope, showing how the proposed adaptive procedure significantly improves the quality of the estimation, with a lower fraction of variance unexplained by the extracted envelope for different simulated modulating waveforms (p < 0.005). The definition of an entropy-based convergence criterion has allowed for a complete automatization of the process. We infer that this algorithm can ensure repeatability of the estimation of the sEMG amplitude, due to its independence from the experimental choices, so allowing for a quantitative interpretation in a clinical environment.
... Traditional methods such as time-domain and frequencydomain analyses have been widely utilized in EMG pattern recognition [13], and they have a good capability to track muscular changes. Other methods like Bayesian estimation [15] and linear filtering also achieve good estimates of muscle forces. We first consider simple and robust linear models with one estimator per channel. ...
Spinal cord stimulation has enabled humans with motor complete spinal cord injury (SCI) to independently stand and recover some lost autonomic function. Quantifying the quality of bipedal standing under spinal stimulation is important for spinal rehabilitation therapies and for new strategies that seek to combine spinal stimulation and rehabilitative robots (such as exoskeletons) in real time feedback. To study the potential for automated electromyography (EMG) analysis in SCI, we evaluated the standing quality of paralyzed patients undergoing electrical spinal cord stimulation using both video and multi-channel surface EMG recordings during spinal stimulation therapy sessions. The quality of standing under different stimulation settings was quantified manually by experienced clinicians. By correlating features of the recorded EMG activity with the expert evaluations, we show that multi-channel EMG recording can provide accurate, fast, and robust estimation for the quality of bipedal standing in spinally stimulated SCI patients. Moreover, our analysis shows that the total number of EMG channels needed to effectively predict standing quality can be reduced while maintaining high estimation accuracy, which provides more flexibility for rehabilitation robotic systems to incorporate EMG recordings.
... Traditional methods such as time-domain and frequencydomain analyses have been widely utilized in EMG pattern recognition [13], and they have a good capability to track muscular changes. Other methods like Bayesian estimation [15] and linear filtering also achieve good estimates of muscle forces. We first consider simple and robust linear models with one estimator per channel. ...
Spinal cord stimulation has enabled humans with motor complete spinal cord injury (SCI) to independently stand and recover some lost autonomic function. Quantifying the quality of bipedal standing under spinal stimulation is important for spinal rehabilitation therapies and for new strategies that seek to combine spinal stimulation and rehabilitative robots (such as exoskeletons) in real time feedback. To study the potential for automated electromyography (EMG) analysis in SCI, we evaluated the standing quality of paralyzed patients undergoing electrical spinal cord stimulation using both video and multi-channel surface EMG recordings during spinal stimulation therapy sessions. The quality of standing under different stimulation settings was quantified manually by experienced clinicians. By correlating features of the recorded EMG activity with the expert evaluations, we show that multi-channel EMG recording can provide accurate, fast, and robust estimation for the quality of bipedal standing in spinally stimulated SCI patients. Moreover, our analysis shows that the total number of EMG channels needed to effectively predict standing quality can be reduced while maintaining high estimation accuracy, which provides more flexibility for rehabilitation robotic systems to incorporate EMG recordings.
... We recorded the 8 EMG sensors signals and the orientation quaternion from the Myo device placed on the left forearm. To filter the raw EMG data we applied a Bayesian filtering [8] to estimate the activity of muscles. We capture the accelerometer and gyroscope from the Riot sensor placed in the right hand. ...
We train and evaluate two machine learning models for predicting fingering in violin performances using motion and EMG sensors integrated in the Myo device. Our aim is twofold: first, provide a fingering recognition model in the context of a gamification virtual violin application where we measure both right hand (i.e. bow) and left hand (i.e. fingering) gestures, and second, implement a tracking system for a computer assisted pedagogical tool for self-regulated learners in high-level music education. Our approach is based on the principle of mapping-by-demonstration in which the model is trained by the performer. We evaluated a model based on Decision Trees and compared it with a Hidden Markovian Model.
... To that end, a highly novel recent development has to do with the use of Bayesian probabilities 38,39 to smooth the EMG signal. As mentioned previously, the myoelectric control signal-raw EMG signal-is noisy and can be considered for most purposes to be a Gaussian random signal. ...
Extrapolation of Emerging Technologies and Their Long-Term Implications for Myoelectric versus Body-Powered Prostheses: An Engineering Perspective
Richard F. ff. Weir, PhD
ABSTRACT
The field of prosthetic rehabilitation is at the cusp of a revolution in upper-limb prosthetic techniques and treatment options. After 50 years of largely incremental developments in the design of both body-powered and myoelectric upper extremity prostheses, new technologies are coming of age that will provide sensory feedback to the user. This, in turn, will promote embodiment of the prosthesis, allowing users to believe the device is a true extension of themselves. This will facilitate the incorporation of the prosthesis into their body image and allow users to finally begin to think of the prosthesis as a true limb replacement rather than as a tool. This review surveys innovations in upper-limb prosthetic rehabilitation from an engineering perspective.
(J Prosthet Orthot. 2017;29:P63–P74)
KEY INDEXING TERMS: cable operated, EMG, external power, electric powered, extended physiological proprioception, osseointegration, neural interface, rapid prototyping, implantable sensors, sensors, cineplasty, kineplasty, sensory feedback, body-powered, externally-powered, limb transplantation, tissue printing, bio-printing, limb regrowth
When using EMG biofeedback to control the grasping force of a myoelectric prosthesis, subjects need to activate their muscles and maintain the myoelectric signal within an appropriate interval. However, their performance decreases for higher forces, because the myoelectric signal is more variable for stronger contractions. Therefore, the present study proposes to implement EMG biofeedback using nonlinear mapping, in which EMG intervals of increasing size are mapped to equal-sized intervals of the prosthesis velocity. To validate this approach, 20 able-bodied subjects performed force-matching tasks using Michelangelo prosthesis with and without EMG biofeedback with linear and nonlinear mapping. Additionally, four transradial amputees performed a functional task in the same feedback and mapping conditions. The success rate in producing desired force was significantly higher with feedback (65.4±15.9%) compared to no feedback (46.2±14.9%) as well as when using nonlinear (62.4±16.8%) versus linear mapping (49.2±17.2%). Overall, in able-bodied subjects, the highest success rate was obtained when EMG biofeedback was combined with nonlinear mapping (72%), and the opposite for linear mapping with no feedback (39.6%). The same trend was registered also in four amputee subjects. Therefore, EMG biofeedback improved prosthesis force control, especially when combined with nonlinear mapping, which showed to be an effective approach to counteract increasing variability of myoelectric signal for stronger contractions.
The contribution of different brain regions to movement abnormalities in children with dystonia is unknown. Three awake subjects undergoing depth electrode implantation for assessments of potential deep brain recording targets performed a rhythmic figure-8 drawing task. Two subjects had dystonia, one was undergoing testing for treatment of Tourette Syndrome and had neither dystonia nor abnormal movements during testing. Movement-related signals were evaluated by determining the magnitude of task-related frequency components. Brain signals were recorded in globus pallidus internus (GPi), the ventral oralis anterior/posterior (VoaVop) and the ventral intermediate (Vim) nuclei of the thalamus. In comparison to the subject without dystonia, both children with dystonia showed increased task-related activity in GPi and Vim. This finding is consistent with a role of both basal ganglia and cerebellar outputs in the pathogenesis of dystonia. Our results further suggest that frequency analysis of brain recordings during cyclic movements may be a useful tool for analysis of the presence of movement-related signals in various brain regions.
Objective:
To study the neural control of movement, it is often necessary to estimate how muscles are activated across a variety of behavioral conditions. One approach is to try extracting the underlying neural command signal to muscles by applying latent variable modeling methods to electromyographic (EMG) recordings. However, estimating the latent command signal that underlies muscle activation is challenging due to its complex relation with recorded EMG signals. Common approaches estimate each muscle activation independently or require manual tuning of model hyperparameters to preserve behaviorally-relevant features.
Approach:
Here, we adapted AutoLFADS, a large-scale, unsupervised deep learning approach originally designed to de-noise cortical spiking data, to estimate muscle activation from multi-muscle EMG signals. AutoLFADS uses recurrent neural networks (RNNs) to model the spatial and temporal regularities that underlie multi-muscle activation.
Main results:
We first tested AutoLFADS on muscle activity from the rat hindlimb during locomotion and found that it dynamically adjusts its frequency response characteristics across different phases of behavior. The model produced single-trial estimates of muscle activation that improved prediction of joint kinematics as compared to low-pass or Bayesian filtering. We also applied AutoLFADS to monkey forearm muscle activity recorded during an isometric wrist force task. AutoLFADS uncovered previously uncharacterized high-frequency oscillations in the EMG that enhanced the correlation with measured force. The AutoLFADS-inferred estimates of muscle activation were also more closely correlated with simultaneously-recorded motor cortical activity than were other tested approaches.
Significance:
This method leverages dynamical systems modeling and artificial neural networks to provide estimates of muscle activation for multiple muscles. Ultimately, the approach can be used for further studies of multi-muscle coordination and its control by upstream brain areas.
Model-based biomimetic control with neuro-muscular reflex requires accurate representation of muscle fascicle length, which affects both force generation capability of muscle and dynamics of muscle spindle. However, physiological data are insufficient to guide the selection of range of fascicle length for task control. Here a reverse engineering approach was used to investigate the effects of different fascicle length range on controller's force control ability, so as to justify the selection of operating range of muscle length for a grasp force task. We compared 3 different ranges of fascicle length for their effects on force generation, i.e. R1: 0.5 - 1.0 Lo, R2: 0.5 - 1.3 Lo and R3: 0.5 - 1.6 Lo. The rationale to test these range selections was based on both physiological realism and engineering considerations. The steady state force output and transient force responses were evaluated with a range of step inputs as controller input. Results show that the prosthetic finger can produce a linear steady state force response with all 3 ranges of fascicle length. Peak force was the largest with R3. Fascicle length range had no significant effect on the rise time in force generation tasks. Results suggest that a wider range of fascicle length may be more favorable for force capacity, since the contact point of force control may well fall near the optimal length (Lo) region.
Objective:
The implementation of somatosensory feedback in upper limb myoelectric prostheses is an important step towards the restoration of lost sensory-motor functions. EMG feedback is a recently proposed method for closing the control loop wherein the myoelectric signal that drives the prosthesis is also used to generate the feedback provided to the user. Therefore, the characteristics of the myoelectric signal (variability and sensitivity) are likely to significantly affect the subject's ability to utilize feedback for online control. In the present study, we investigated how the cutoff frequency of the low-pass filter (0.5, 1 and 1.5 Hz) and normalization value (20, 40 and 60% of maximum voluntary contraction) affect the quality of closed-loop control with EMG feedback. Lower cutoff and normalization decrease the variability of EMG but also increase the time lag between the contraction and the feedback (cutoff) as well as the sensitivity of myoelectric signal (normalization).
Approach:
Ten participants were asked to generate three grasp force levels with a myoelectric prosthetic hand, while receiving 5-level vibrotactile EMG feedback, over nine experimental runs (all parameter combinations). The outcome measure was the success rate in achieving the appropriate level of myoelectric signal (primary outcome) and grasping force (secondary outcome).
Main results:
Overall, the experiments demonstrated that EMG feedback provided robust control across conditions. Nevertheless, the performance was significantly better for the lowest cutoff (0.5 Hz) and higher normalization (40 and 60%). The highest success rate for the EMG was 72% in the condition (40% MVC, 0.5 Hz), which was 24% higher than that in the condition (20% MVC, 1.5 Hz) with the lowest performance. The success rate for the force followed a similar trend.
Significance:
This is the first study that systematically explored the parameter space for calibration of EMG feedback, which is a critical step for the future clinical application.
Continuous movement intent decoders are critical for precise control of hand and wrist prostheses. Noise in biological signals (e.g., myoelectric or neural signals) can lead to undesirable jitter in the output of these types of decoders. A low-pass filter (LPF) at the output of the decoder effectively reduces jitter, but also substantially slows intended movements. This paper introduces an alternative, the latching filter (LF), a recursive, nonlinear filter that provides smoothing of small-amplitude jitter but allows quick changes to its output in response to large input changes. The performance of a Kalman filter (KF) decoder smoothed with an LF is compared with that of both an KF decoder without an additional smoother and a KF decoder smoothed with a LPF. These three algorithms were tested in real-time on target holding and target reaching tasks using surface electromyographic signals recorded from 5 non-amputee subjects, and intramuscular electromyographic and peripheral neural signals recorded from an amputee subject. When compared with the LPF, the LF provided a statistically significant improvement in amputee and non-amputee subjects’ ability to hold the hand steady at requested positions and achieve movement goals faster. The KF decoder with LF provided a statistically significant improvement in all subjects’ ability to hold the prosthetic hand steady, with only slightly lower speeds, when compared to the unsmoothed KF.
Current control of prosthetic hands is ineffective when grasping deformable, irregular, or heavy objects. In humans, grasping is achieved under spinal reflexive control of the musculotendon skeletal structure, which produces a hand stiffness commensurate with the task. We hypothesize that mimicking reflex on a prosthetic hand may improve grasping performance and safety when interacting with human. Here, we present a design of compliant controller for prosthetic hand with a neuromorphic model of human reflex. The model includes 6 motoneuron pools containing 768 spiking neurons, 1 muscle spindle with 128 spiking afferents, and 1 modified Hill-type muscle. Models are implemented using neuromorphic hardware with 1 kHz real-time computing. Experimental tests showed that the prosthetic hand could sustain a 40 N load compared to 95 N for an adult. Stiffness range was adjustable from 60 to 640 N/m, about 46.6% of that of human hand. The grasping velocity could be ramped up to 14.4 cm/s, or 24% of the human peak velocity. The complaint control could switch between free movement and contact force when pressing a deformable beam. The amputee can achieve a 47% information throughput of healthy humans. Overall, the reflex-enabled prosthetic hand demonstrated the attributes of human compliant grasping with the neuromorphic model of spinal neuromuscular reflex.
Cerebral palsy (CP) is the most common childhood disability. CP can impact a person’s motor control, perception, intellectual function, ability to perform daily activities and participation in society. Persons with CP frequently have impaired hand function affecting motor activities. Usually improving motor activities require frequent and intense use of the affected hand. Importantly, biofeedback, where a person receives information about their body state can help improve function by informing the individual to how their body is moving. This type of information can readily be delivered through interactive computer play (ICP) technologies, that is where one interacts with virtual objects using therapeutic movements, to motivate highly repetitive practice at home.In this thesis an evidence-based biofeedback strategy is identified by systematic review and embedded into a novel ICP technology via co-creation with individuals with CP and clinicians. Feasibility of the resulting technology is evaluated by 19 young people with CP during a 1-month home-based intervention.The systematic review showed generally positive but very low quality of evidence due to the number of non-controlled studies and the heterogeneity of outcome measures. While interventions consistently showed improvements in measures of motor activity pre-post-intervention, they frequently implemented characteristics of biofeedback incongruent with motor learning principles expected to facilitate sustained results. Interventions using biofeedback would be improved by using a strategy that facilitates self-regulation, varies in timing and presentation, and closely connects the game and movement goals.The co-creation process described in this thesis resulted in design recommendations and practical tools for integrating biofeedback into therapy games. These are compiled in an infographic and detailed in tables to support interdisciplinary knowledge sharing with different audiences. Participant use of the biofeedback implementation proved efficient (i.e. participants reduced compensatory arm movements by 10.2±4.0%), effective (i.e. participants made higher quality gestures over time) and engaging (i.e. participants reviewed biofeedback 65.4±22.4% of the time). Participants found the game usable and enjoyable. Home-based calibration showed good agreement with video verified true labels (0.86 ± 0.11, 95%CI = 0.93 - 0.97). Across participants, classifier accuracy in the target therapeutic gesture was 0.91±0.07, 95%CI = 0.87 - 0.94 (0.80±0.14, 95%CI = 0.73 - 0.87 and 0.75±0.23, 95%CI = 0.64 - 0.86 in secondary gestures). Features expected to be sensitive to neuromotor signs of CP were significant contributors to classification and correlated with practice wrist extension improvement.Evaluation of the a priori feasibility success criteria showed that the 1-month home-based intervention protocol was feasible with minor modifications. Recruitment response (31%) and assessment completion (84%) rates were good and 74% of participants reached self-identified practice goals and 83% of technical issues were resolved immediately. Moderate effects were observed in Body Function measures (active wrist extension: SMD = 1.82, 95%CI = 0.85 – 2.78; Grip Strength: SMD = 0.63, 95%CI = 0.65 – 1.91; Box and Blocks: Hedge’s g = 0.58, 95%CI = -0.11 – 1.27) and small-moderate effects in Activities and Participation measures (AHA: Hedge’s g = 0.29, 95%CI = -0.39 – 0.97, COPM: r = 0.60, 95%CI = 0.13 – 0.82, SEAS: r = 0.24, 95%CI = -0.25 – 0.61). A definitive RCT is warranted following improvements in technical robustness and increased clinician involvement in the protocol.This thesis provides strategies and practical tools to enhance the efficacy of ICP technologies as a home-based support to manual therapy activities.
Objective
Surface electromyography (sEMG) is a potentially useful signal that can provide therapeutic biofeedback. However, sEMG signal processing is difficult because of the low signal-to-noise ratio and non-stationarity of the raw signal. Conventional online filters often suffer from a compromise between smoothness and responsiveness. Here we propose a new particle filtering method for sEMG processing and compare it to established filtering methods.
Methods
A wrist apparatus measuring isometric wrist extension/flexion force was developed. Six filters (moving average windowing (MAW), adaptive-MAW, 3-layer, Kalman, Bayes and particle filters) were tested on forearm sEMG collected with a Myo armband. Fourteen subjects performed two visuomotor tracking tasks (square and sine wave tracking). Tracking error, measured as the root mean square error (RMSE2), was used as a metric to compare the influence of different filters on overall performance.
Results
For sine wave tracking tasks (representing continuous trajectory control), the particle filter (RMSE2: 53.30 ± 15.69 pixels) had the lowest tracking error. For the square wave tracking task (representing discrete endpoint control), the Bayes filter (RMSE2: 37.82 ± 23.53 pixels) had the lowest tracking error. With respect to computational requirements, the Kalman filter was the most efficient.
Conclusion
Our results indicate that the filter requirements for sEMG controllers are task specific, but the new particle filtering method presented here represents a good compromise for the different types of motor control tested here.
Significance
The particle filter has the potential to improve sEMG based therapeutic biofeedback.
Myocontrol holds great promise because it has the potential to provide flexible and accurate prosthetic control that approaches the quality of normal movement. Speed and accuracy are important factors to consider when applying myoelectric signals to external devices. Fitts's law can be used to assess the speed-accuracy trade-off. We hypothesized that speed is affected not only by accuracy as prescribed by Fitts's law, but also by target distances independent of target size. A total of 12 healthy adult subjects were studied. Subjects controlled the vertical movement of a computer cursor by contracting their dominant first dorsal interosseus muscle to reach targets as quickly as possible. We manipulated twenty-five different target distances in order to obtain five indices of difficulty, as defined by Fitts's law, combined with five target widths. We tested the relationship between movement time and the index of difficulty as predicted by Fitts's law among different combinations of target distance and widths. Results showed a significant linear regression for all conditions, with the exception of a significantly longer movement time than predicted for targets close to the start point. Movements to these targets showed significantly higher relative variance during stabilization, higher overshoot, and lower success. Therefore, we found that with comparable index of difficulty, small distance movements had a higher variability, slower movement, and higher rates of error compared to larger distance movements. Our results are consistent with our hypothesis that low muscle activation required for short distances results in higher variability and low controllability in reaching the target as required by the task demand. Neurophysiological mechanisms underlying the violation of the Fitts's law relationship are discussed. These results have significance for myocontrol applications, and we suggest that such applications require control signals with sufficient recruitment to reduce variability at lower levels of muscle activation.
Neurobiological systems operate at power levels that are unattainable by modern electronic systems while exhibiting broader information processing capabilities for a number of important tasks. A variety of engineered systems designed for energy efficiency or hardware simplicity use time-based signal representations, which share similar mathematical principles with those that arise naturally in biology. In general, time-based signal representations refer to embedding information into the timing, density, or duration of a predetermined, and often bipolar, waveform. In mammalian nervous systems, it is generally accepted that neurons embed information into the timing and firing density of the sudden changes in their membrane potential, or spikes. Similarly, many low-power electronic systems use signal representations that embed information in the timing, repetition frequency, or duration of simple pulse waveforms. Despite their apparent similarities, such signal representations are often studied in different contexts.
A direct, ready-to-use surface electromyogram (sEMG) pattern classification algorithm that does not require prerequisite training, regardless of the user, is proposed herein. In addition to data collection, conventional supervised learning approaches for sEMG require labeling and segmenting the data and additional time for the learning algorithm. Consequently, these approaches cannot cope well with sEMG patterns during motion transitions of various movement speeds. The proposed unsupervised and self-adaptive method employs an iterative self-adaptive procedure realized by the probabilistic methods of diffusion, updating, and registration to cluster the activation patterns simultaneously in real time, and classify the current sEMG as new clustered patterns. Experiments demonstrated that even for the same motion, the proposed method could autonomously detect changes in muscular activation patterns varying with the speed of motion. Furthermore, some patterns of both steady- and transient-state motions could be distinguished. In addition, it was verified that the classified sEMG pattern could be correlated consistently with the actual motion, thereby realizing a high level of motion classification.
Humans consistently coordinate their joints to perform a variety of tasks. Computational motor control theory explains these stereotypical behaviors using optimal control. Several cost functions have been used to explain specific movements, which suggests that the brain optimizes for a combination of costs and just varies their relative weights to perform different tasks. In the case of tunable human-machine interfaces, we hypothesize that the human-machine interface should be optimized according to the costs that the user cares about when making the movement. Here, we study how the relative weights of individual cost functions in a composite movement cost affect the optimal control signal produced by the user and the mapping between the user's control signals and the machine's output, using prosthesis control as a specific example. This framework was tested by building a hierarchical optimization model that independently optimized for the user control signal and the virtual dynamics of the device. Our results indicate the feasibility of the approach and show the potential for using such a model in prosthesis tuning. This method could be used to allow clinicians and users to tune their prosthesis based on costs they actually care about; and allow the platforms to be customized for the unique needs of every patient.
Spinal cord stimulation (SCS) has recently enabled humans with motor complete spinal cord injury (SCI) to independently stand and recover some lost autonomic function. However, the nature of the recovered motor activity and the interplay between SCS and motor training are not well understood. Understanding the effect of stand training and spinal stimulation on motor activity during bipedal standing is important for designing spinal rehabilitation therapies that seek to combine spinal stimulation and rehabilitative robots. In this study, we examined electromyography (EMG) data gathered from two SCI patients and six healthy subjects as they attempted standing. We analyzed the muscle activation patterns and EMG waveform shape to quantify both the changes in SCI patient motor activity with training, and the differences between healthy motor activity and SCI patient motor activity under stimulation. We also looked for correlations between the similarity in SCI patients' motor activity to healthy subjects and their overall standing ability. We found that good standing in SCI patients does not emulate healthy standing muscle activity. Furthermore, patient stand training heavily influenced motor activation patterns, but not in ways that improved standing ability. These results indicate that current training techniques do not optimally influence motor activity, and robotic rehabilitation strategies for SCI patients should target essential features of motor activity to optimize functional performance, rather than emulate healthy activity.
L'amputation du membre supérieur, dont la prévalence est comparable à celle des maladies orphelines, induit chez les patients une perte considérable d'autonomie dans la majorité des tâches simples de la vie quotidienne. Pour pallier ces difficultés, les prothèses myoélectriques actuelles proposent une multitude de mouvements possibles. Cependant, leur contrôle non intuitif et lourd cognitivement requiert un apprentissage long et difficile, qui pousse une proportion importante de patients amputés à l'abandon de la prothèse. Dans cette thèse, nous avons cherché à identifier l'origine des difficultés et les manques du contrôle myoélectrique en comparaison au contrôle sensorimoteur naturel, dans le but à terme de proposer de meilleures solutions de restitution et de suppléance. Pour cela, nous avons manipulé diverses conditions expérimentales dans un contexte d'interface homme-machine simplifié où des sujets non amputés contrôlent un curseur sur un écran à partir de contractions isométriques, i.e. des contractions qui n'engendrent pas de mouvement. Cette condition isométrique nous a permis de nous approcher de la condition de la personne amputée contrôlant sa prothèse à partir de l'activité électrique (EMG) de ses muscles résiduels, en absence de mouvement articulaire. Durant une tâche d'atteinte de cible, nous avons entre autre démontré le bénéfice d'une adaptation conjointe du décodeur qui traduit les activités EMG en mouvement du curseur, venant s'ajouter à la propre adaptation du plan de mouvement des sujets en réponse à des perturbations orientées. De plus, il a été mis en évidence que ce bénéfice est d'autant plus important que la dynamique d'adaptation artificielle du décodeur s'inspire de celle de l'Homme. Dans des tâches d'acquisition et de poursuite de cible, impliquant davantage les mécanismes de régulation en ligne du mouvement, nous avons mis en évidence l'importance d'une congruence immédiate entre les informations sensorimotrices et la position du curseur à l'écran pour permettre des corrections rapides et efficaces. Dans une condition où le niveau de bruit du système est relativement faible, comme avec l'utilisation du signal de forces plus stable que l'habituel signal EMG, cette congruence explique, en partie, la supériorité d'un contrôle d'ordre 0 (i.e. position) sur un contrôle d'ordre 1 (i.e.} vitesse). Cependant, dès lors que le niveau de bruit est trop important, ce qui est le cas avec le signal EMG, le filtrage induit par l'intégration nécessaire au contrôle vitesse fait que celui-ci devient plus performant que le contrôle position. L'ensemble de ces résultats suggèrent qu'un décodeur adaptatif et intuitif, respectant et suppléant au mieux les boucles du contrôle sensorimoteur naturel, est le plus à même de faciliter le contrôle des futures prothèses.
In this paper we propose a novel synergy-based myocontrol scheme for finger force estimation and classification which is able to simultaneously control 4 fingers with a training phase based only on individual-finger data. The proposed method has been tested using the online-available NinaPro database and validated in a preliminary experiment conducted with the use of a hand-exoskeleton. Results show how the presented approach outperforms considerably the linear regression method which is considered standard approach in myoelectric control. The low error rate obtained (smaller than 10% of the targeted force) and the effectiveness in decreasing the number of false activation open the possibilities for future uses in fields such as haptics and neuro-rehabilitation.
During a sustained muscle contraction, the amplitude of electromyographic (EMG) signals increases and the spectrum of the EMG signal shifts toward lower frequencies. These effects are due to muscular fatigue and can cause problems in the control of myoelectric prostheses and in the estimation of contraction level from the EMG signal. It has been well known that the fatigue effects can be explained by the conduction velocity changes during the fatigue process and by the idea that the conduction velocity is linearly proportional to the median frequency of EMG signals. Hence the fatigue process can be monitored by measuring the median frequency. A fatigue compensation preprocessor has been developed. It uses the widely accepted power spectrum density model of EMG signals that contains the conduction velocity as a measure of fatigue. It was verified that the preprocessor scales down the amplitude of the fatigued EMG signal and decompresses the spectrum. Hence, the preprocessor eliminates the increase in amplitude and the shift in frequency and enables consistent EMG signals to be used to control prostheses.
A systematic, experimental study of the influence of smoothing window length on the signal-to-noise ratio (SNR) of electromyogram (EMG) amplitude estimates is described. Surface EMG waveforms were sampled during nonfatiguing, constant-force, constant-angle contractions of the biceps or triceps muscles, over the range of 10%-75% maximum voluntary contraction. EMG amplitude estimates were computed with eight different EMG processor schemes using smoothing length durations spanning 2.45-500 ms. An SNR was computed from each amplitude estimate (deviations about the mean value of the estimate were considered as noise). Over these window lengths, average +/- standard deviation SNR's ranged from 1.4 +/- 0.28 to 16.2 +/- 5.4 for unwhitened single-channel EMG processing and from 3.2 +/- 0.7 to 37.3 +/- 14.2 for whitened, multiple-channel EMG processing (results pooled across contraction level). It was found that SNR increased with window length in a square root fashion. The shape of this relationship was consistent with classic theoretical predictions, however none of the processors achieved the absolute performance level predicted by the theory. These results are useful in selecting the length of the smoothing window in traditional surface EMG studies. In addition, this study should contribute to the development of EMG processors which dynamically tune the smoothing window length when the EMG amplitude is time varying.
Previous research showed that whitening the surface electromyogram (EMG) can improve EMG amplitude estimation (where EMG amplitude is defined as the time-varying standard deviation of the EMG). However, conventional whitening via a linear filter seems to fail at low EMG amplitude levels, perhaps due to additive background noise in the measured EMG. This paper describes an adaptive whitening technique that overcomes this problem by cascading a nonadaptive whitening filter, an adaptive Wiener filter, and an adaptive gain correction. These stages can be calibrated from two, five second duration, constant-angle, constant-force contractions, one at a reference level [e.g., 50% maximum voluntary contraction (MVC)] and one at 0% MVC. In experimental studies, subjects used real-time EMG amplitude estimates to track a uniform-density, band-limited random target. With a 0.25-Hz bandwidth target, either adaptive whitening or multiple-channel processing reduced the tracking error roughly half-way to the error achieved using the dynamometer signal as the feedback. At the 1.00-Hz bandwidth, all of the EMG processors had errors equivalent to that of the dynamometer signal, reflecting that errors in this task were dominated by subjects' inability to track targets at this bandwidth. Increases in the additive noise level, smoothing window length, and tracking bandwidth diminish the advantages of whitening.
This work represents an ongoing investigation of dexterous and natural control of powered upper limbs using the myoelectric signal. When approached as a pattern recognition problem, the success of a myoelectric control scheme depends largely on the classification accuracy. A novel approach is described that demonstrates greater accuracy than in previous work. Fundamental to the success of this method is the use of a wavelet-based feature set, reduced in dimension by principal components analysis. Further, it is shown that four channels of myoelectric data greatly improve the classification accuracy, as compared to one or two channels. It is demonstrated that exceptionally accurate performance is possible using the steady-state myoelectric signal. Exploiting these successes, a robust online classifier is constructed, which produces class decisions on a continuous stream of data. Although in its preliminary stages of development, this scheme promises a more natural and efficient means of myoelectric control than one based on discrete, transient bursts of activity.
Electrical signals recorded by means of surface electromyography (SEMG) contain some useful information for a better understanding of strategies underlying human movement. In particular, a great contribution to biomechanic studies may be provided by a correct estimation of the amplitude of SEMG signals that is related to the force exerted by muscles. This information could, in fact, represent an indirect assessment of muscular force obtained without using invasive measurement techniques. This article presents a new fully automatic estimation technique adaptively working on SEMG signal characteristics. The discussion of the theoretical background of the estimator together with a feasibility study demonstrates the usefulness of its application. An example of the application to signals recorded during dynamic protocols is also shown
The sections above have described an EMG amplitude estimator and an initial application of this estimator to the EMG-torque problem. The amplitude estimator consists of six stages. In the first stage, motion artifact and power-line interference are attenuated. Motion artifact is typically removed with a highpass filter. Elimination of power-line noise is more difficult. Commercial systems tend to use notch filters, accepting the concomitant loss of "true" signal power in exchange for simplicity and robustness. Adaptive methods may be preferable, however, to preserve more "true" signal power. In stage two, the signal is whitened. One fixed whitening technique and two adaptive whitening methods were described. For low-amplitude levels, the adaptive whitening technique that includes adaptive noise cancellation may be necessary. In stage three, multiple EMG channels (all overlying the same muscle) are combined. For most applications, simple gain normalization is all that is required. Stage four rectifies the signal and then applies the power law required to demodulate the signal. In stage six, the inverse of the power law is applied to relinearize the signal. Direct comparison of MAV (first power) to RMS (second power) processing demonstrates little difference between the two. Therefore, unless there is reason to believe that the EMG density departs strongly from that found in the existing studies, RMS and MAV processing are essentially identical. In stage five, the demodulated samples are averaged across all channels and then smoothed (time averaged) to reduce the variance of the amplitude estimate, but at the expense of increasing the bias. For best performance, the window length that best trades off variance and bias error is selected. The advanced EMG processing was next applied to dynamic EMG-torque estimation about the elbow joint. Results showed that improved EMG amplitude estimates led to improved EMG-torque estimates. An initial comparison of different system-identification techniques and model orders was reported. It is expected that these advanced processing and identification algorithms will also improve performance in other EMG applications, including myoelectrically controlled prostheses, biofeedback, and ergonomic assessment.
Intuitive myoelectric prosthesis control is difficult to achieve due to the absence of proprioceptive feedback, which forces the user to monitor grip pressure by visual information. Existing myoelectric hand prostheses form a single degree of freedom pincer motion that inhibits the stable prehension of a range of objects. Multi-axis hands may address this lack of functionality, but as with multifunction devices in general, serve to increase the cognitive burden on the user. Intelligent hierarchical control of multiple degree-of-freedom hand prostheses has been used to reduce the need for visual feedback by automating the grasping process. This paper presents a hybrid controller that has been developed to enable different prehensile functions to be initiated directly from the user's myoelectric signal. A digital signal processor (DSP) regulates the grip pressure of a new six-degree-of-freedom hand prosthesis thereby ensuring secure prehension without continuous visual feedback.
A widely used signal processing paradigm is the state-space model. The state-space model is defined by two equations: an observation equation that describes how the hidden state or latent process is observed and a state equation that defines the evolution of the process through time. Inspired by neurophysiology experiments in which neural spiking activity is induced by an implicit (latent) stimulus, we develop an algorithm to estimate a state-space model observed through point process measurements. We represent the latent process modulating the neural spiking activity as a gaussian autoregressive model driven by an external stimulus. Given the latent process, neural spiking activity is characterized as a general point process defined by its conditional intensity function. We develop an approximate expectation-maximization (EM) algorithm to estimate the unobservable state-space process, its parameters, and the parameters of the point process. The EM algorithm combines a point process recursive nonlinear filter algorithm, the fixed interval smoothing algorithm, and the state-space covariance algorithm to compute the complete data log likelihood efficiently. We use a Kolmogorov-Smirnov test based on the time-rescaling theorem to evaluate agreement between the model and point process data. We illustrate the model with two simulated data examples: an ensemble of Poisson neurons driven by a common stimulus and a single neuron whose conditional intensity function is approximated as a local Bernoulli process.
This paper represents an ongoing investigation of dexterous and natural control of upper extremity prostheses using the myoelectric signal (MES). The scheme described within uses pattern recognition to process four channels of MES, with the task of discriminating multiple classes of limb movement. The method does not require segmentation of the MES data, allowing a continuous stream of class decisions to be delivered to a prosthetic device. It is shown in this paper that, by exploiting the processing power inherent in current computing systems, substantial gains in classifier accuracy and response time are possible. Other important characteristics for prosthetic control systems are met as well. Due to the fact that the classifier learns the muscle activation patterns for each desired class for each individual, a natural control actuation results. The continuous decision stream allows complex sequences of manipulation involving multiple joints to be performed without interruption. Finally, minimal storage capacity is required, which is an important factor in embedded control systems.
The dependence of the form of the EMG-force relation on key motoneuron and muscle properties was explored using a simulation approach. Surface EMG signals and isometric forces were simulated using existing motoneuron pool, muscle force, and surface EMG models, based primarily on reported properties of the first dorsal interosseous (FDI) muscle in humans. Our simulation results indicate that the relation between electrical and mechanical properties of the individual motor unit level plays the dominant role in determining the overall EMG amplitude-force relation of the muscle, while the underlying motor unit firing rate strategy appears to be a less important factor. However, different motor unit firing rate strategies result in substantially different relations between counts of the numbers of motoneuron discharges and the isometric force. Our simulation results also show that EMG amplitude (estimated as the average rectified value) increases as a result of synchronous discharges of different motor units within the pool, but the magnitude of this increase is determined primarily by the action potential duration of the synchronized motor units. Furthermore, when the EMG effects are normalized to their maximum levels, motor unit synchrony does not exert significant effects on the form of the EMG-force relation, provided that the synchrony level is held similar at different excitation levels.
A system for electromyographic (EMG) triggering of robot-assisted therapy (dubbed the EMG game) for stroke patients is presented. The onset of a patient's attempt to move is detected by monitoring EMG in selected muscles, whereupon the robot assists her or him to perform point-to-point movements in a horizontal plane. Besides delivering customized robot-assisted therapy, the system can record signals that may be useful to better understand the process of recovery from stroke. Preliminary experiments aimed at testing the proposed system and gaining insight into the potential of EMG-triggered, robot-assisted therapy are reported.
This paper introduces and evaluates the use of Gaussian mixture models (GMMs) for multiple limb motion classification using continuous myoelectric signals. The focus of this work is to optimize the configuration of this classification scheme. To that end, a complete experimental evaluation of this system is conducted on a 12 subject database. The experiments examine the GMMs algorithmic issues including the model order selection and variance limiting, the segmentation of the data, and various feature sets including time-domain features and autoregressive features. The benefits of postprocessing the results using a majority vote rule are demonstrated. The performance of the GMM is compared to three commonly used classifiers: a linear discriminant analysis, a linear perceptron network, and a multilayer perceptron neural network. The GMM-based limb motion classification system demonstrates exceptional classification accuracy and results in a robust method of motion classification with low computational load.
Epoch-based electromyogram (EMG) amplitude estimates have not incorporated signal whitening, even though whitening has demonstrated significant improvements for stream-based estimates. This work presents new epoch-based algorithms, for both single- and multiple-channel EMG, which include a whitening stage. The best multiple-channel whitening processor provided a 21.4%-22.5% improvement over single-channel unwhitened estimation in an EMG-to-torque application.
This article reviews the experimental foundations of EMG biofeedback with the upper extremity. Considered are investigations on recruitment and training of single and multiple motor units in both normal and nonnormal subjects, on transfer of training effects from trained to untrained muscles, and on the relationship of reduced muscle output to relaxation. Examined are procedures, results, and conclusions of these basic studies. Problems noted in the research are discussed and suggestions are made for further work in the area.
The use of EMG as a proportional control signal for prostheses is reviewed. The lack of success of existing proportional EMG-controlled prostheses is shown to be attributable in the greater part to the presence of low frequency "noise" in the processed EMG signal. If EMG is to yield useful proportional information, this noise must be eliminated, and to this end the methods of processing EMG are reviewed and the sources of noise examined. A number of methods for its removal are discussed.
A comparison of signal-to-noise ratios and rise times was performed on several myoelectric filters used for muscle-force estimation and prosthesis control. Linear, averaging, and adaptive filters were compared using single as well as multiple electrode pairs (spatial filtering). The filters were matched for having the same rise time (0-95%) and the signal-to-noise ratios were measured off-line using the same myoelectric signal recording. The linear filter was a low-pass filter with a time constant of 80 ms. The averaging filter had an averaging time of 250 ms. The adaptive filter was the same as is used in the Utah Artificial Arm. The adaptive filter varied its time constant according to the rate of change of the signal mean. If the rate was high, the time constant was set low. If the rate was low, the time constant was set high. Spatial filtering is where the myoelectric signals from four cutaneous sites over the same muscle were summed, that is, spatially filtered, and the resultant signal was smoothed by the linear, averaging, or adaptive filter. Significant improvement in the signal-to-noise ratio has been shown over conventional linear or averaging filters when using spatial and adaptive filtering, both when used separately and when used together.
For pt.I see ibid., vol.35, no.4, p.230-7, 1988. The successful application of functional neuromuscular stimulation to the muscles of paraplegics depends to a large extent on the adequate provision of a means by which the subject can exercise control over the resulting movement. The use of above-lesion electromyographic signals as a solution to the control problem is considered. A practical demonstration of signature discrimination of electromyographic signals, based on the parameters of an autoregressive time series, is presented. The success of this method is dependent on the choice of electrode location. Various experiments to determine such an electrode location are described.
The successful application of functional neuromuscular stimulation to the muscles of paraplegics depends to a large extent on the adequate provision of a means by which the subject can exercise control over the resulting movement. The use of above-lesion electromyographic signals as a solution to the control problem is considered. A number of criteria for such a control system are defined. The general concepts underlying time-series analysis are described and the suitability of this method as a means of processing electromyographic signals is investigated. The electromyogram, which exhibits weak stationarity over short time intervals, is represented by a fourth-order autoregressive model. A sequential least-squares algorithm is used to determine the model parameters, which are then used to achieve signature discrimination.
A critical review is presented of studies utilizing EMG biofeedback for relaxation of upper extremity musculature. Examined are experimental investigations with normal subjects and those with psychological problems, and clinical applications of the methodology for treatment of involuntary movements and anxiety. Articles are reviewed in terms of procedures, controls and results. It is determined that few valid conclusions can be drawn regarding the efficacy of upper extremity EMG biofeedback for relaxation and that further research is required prior to utilizing the technique clinically. Suggestions are offered for areas of investigation.
An analytic study was initiated to investigate whether the normalized surface myoelectric signal vs. normalized force relationship varies in different human muscles and whether it is dependent on training level and rate of force production. The data were obtained from experiments that involved the biceps, deltoid, and first dorsal interosseous of three pianists, four long-distance swimmers, three power lifters, and six normal subjects. The elite performers (among the world's best) were chosen because they exhibited varying degrees of fine motor control, endurance training, and power training in different muscles. Approximately 200 isometric linearly force-varying contractions peaking at 80% of the maximal voluntary contraction level were processed. The results indicated that the myoelectric signal-force relationship was primarily determined by the muscle under investigation and was generally independent of the subject group and the force rate. Whereas this relationship was quasilinear for the first dorsal interosseous, it was nonlinear for the biceps and deltoid. Several possible physiological causes of the observed behavior of the myoelectric signal-force relationship are discussed.
Temporal whitening of individual surface electromyograph (EMG) waveforms and spatial combination of multiple recording sites have separately been demonstrated to improve the performance of EMG amplitude estimation. This investigation combined these two techniques by first whitening, then combining the data from multiple EMG recording sites to form an EMG amplitude estimate. A phenomenological mathematical model of multiple sites of the surface EMG waveform, with analytic solution for an optimal amplitude estimate, is presented. Experimental surface EMG waveforms were then sampled from multiple sites during nonfatiguing, constant-force, isometric contractions of the biceps or triceps muscles, over the range of 10-75% maximum voluntary contraction. A signal-to-noise ratio (SNR) was computed from each amplitude estimate (deviations about the mean value of the estimate were considered as noise). Results showed that SNR performance: 1) increased with the number of EMG sites, 2) was a function of the sampling frequency, 3) was predominantly invariant to various methods of determining spatial uncorrelation filters, 4) was not sensitive to the intersite correlations of the electrode configuration investigated, and 5) was best at lower levels of contraction. A moving average root mean square estimator (245-ms window) provided an average +/- standard deviation (A +/- SD) SNR of 10.7 +/- 3.3 for single site unwhitened recordings. Temporal whitening and four combined sites improved the A +/- SD SNR to 24.6 +/- 10.4. On one subject, eight whitened combined sites were achieved, providing an A +/- SD SNR or 35.0 +/- 13.4.
Previous investigators have experimentally demonstrated and/or analytically predicted that temporal whitening of the surface electromyograph (EMG) waveform prior to demodulation improves the EMG amplitude estimate [1]-[6]. However, no systematic study of the influence of various whitening filters upon amplitude estimate performance has been reported. This paper describes a phenomenological mathematical model of a single site of the surface EMG waveform and reports on experimental studies which examined the performance of several temporal whitening filters. Surface EMG waveforms were sampled during nonfatiguing, constant-force, isometric contractions of the biceps or triceps muscles, over the range of 10-75% maximum voluntary contraction. A signal-to-noise ratio (SNR) was computed from each amplitude estimate (deviations about the mean value of the estimate were considered as noise). A moving average root mean square estimator (245ms window) provided an average +/- standard deviation (A +/- SD) SNR of 10.7 +/- 3.3 for the individual recordings. Temporal whitening with one fourth-order whitening filter designed per site improved the A +/- SD SNR to 17.6 +/- 6.0.
This paper describes an experimental study which relates simultaneous elbow flexor-extensor electromyogram (EMG) amplitude to joint torque. Investigation was limited to the case of isometric, quasi-isotonic (slowly force-varying), nonfatiguing contractions. For each of the flexor and extensor muscle groups, the model relationship between muscle group torque contribution and EMG amplitude was constrained to be a sum of basis functions which had a linear dependence on a set of fit parameters. With these constraints, the problem of identifying the EMG-to-torque relationship was reduced to a linear least squares problem. Surface EMG's from elbow flexors and extensors, and joint torque were simultaneously recorded for nonfatiguing, quasi-isotonic, isometric contractions spanning 0-50% maximum voluntary contraction. Single-/multiple-channel unwhitened/whitened/adaptively whitened EMG amplitude processors were used to identify an EMG-to-torque relation, and then estimate joint torque based on this relation. Each unwhitened multiple-channel EMG-to-torque estimator had a standard error (SE) approximately 70% of its respective single-channel estimator. The adaptively whitened multiple-channel joint torque estimator had an SE approximately 90% of the unwhitened multiple-channel estimator, providing an estimation error approximately 3% of the combined flexion/extension torque range. The experimental studies demonstrated that higher fidelity EMG amplitude processing led to improved joint torque estimation.
When the surface electromyogram (EMG) generated from constant-force, constant-angle, nonfatiguing contractions is modeled as a random process, its density is typically assumed to be Gaussian. This assumption leads to root-mean-square (RMS) processing as the maximum likelihood estimator of the EMG amplitude (where EMG amplitude is defined as the standard deviation of the random process). Contrary to this theoretical formulation, experimental work has found the signal-to-noise-ratio [(SNR), defined as the mean of the amplitude estimate divided by its standard deviation] using mean-absolute-value (MAV) processing to be superior to RMS. This paper reviews RMS processing with the Gaussian model and then derives the expected (inferior) SNR performance of MAV processing with the Gaussian model. Next, a new model for the surface EMG signal, using a Laplacian density, is presented. It is shown that the MAV processor is the maximum likelihood estimator of the EMG amplitude for the Laplacian model. SNR performance based on a Laplacian model is predicted to be inferior to that of the Gaussian model by approximately 32%. Thus, minor variations in the probability distribution of the EMG may result in large decrements in SNR performance. Lastly, experimental data from constant-force, constant-angle, nonfatiguing contractions were examined. The experimentally observed densities fell in between the theoretic Gaussian and Laplacian densities. On average, the Gaussian density best fit the experimental data, although results varied with subject. For amplitude estimation, MAV processing had a slightly higher SNR than RMS processing.
Experimental electromyogram (EMG) data from the human biceps brachii were simulated using the model described in [10] of this work. A multichannel linear electrode array, spanning the length of the biceps, was used to detect monopolar and bipolar signals, from which double differential signals were computed, during either voluntary or electrically elicited isometric contractions. For relatively low-level voluntary contractions (10%-30% of maximum force) individual firings of three to four-different motor units were identified and their waveforms were closely approximated by the model. Motor unit parameters such as depth, size, fiber orientation and length, location of innervation and tendonous zones, propagation velocity, and source width were estimated using the model. Two applications of the model are described. The first analyzes the effects of electrode rotation with respect to the muscle fiber direction and shows the possibility of conduction velocity (CV) over- and under-estimation. The second focuses on the myoelectric manifestations of fatigue during a sustained electrically elicited contraction and the interrelationship between muscle fiber CV, spectral and amplitude variables, and the length of the depolarization zone. It is concluded that a) surface EMG detection using an electrode array, when combined with a model of signal propagation, provides a useful method for understanding the physiological and anatomical determinants of EMG waveform characteristics and b) the model provides a way for the interpretation of fatigue plots.
The last few decades have produced significant improvements in the design of upper limb prostheses through the increasing use of technology. However the limited function exhibited by these devices remains rooted in their single degree of freedom format. Commercial myoelectric hand prostheses warrant high grip forces to ensure stable prehension due to a planar pincer movement. Hence precise and conscious effort is required on the part of the user to ensure optimum grip. Consumers have shown dissatisfaction with the status quo due to the excessive weight and poor function of existing artificial hands. Increasing the number of grasping patterns and improving the visual feedback from an object in the hand are cited as key objectives. This paper outlines the development of the six-axis Southampton-Remedi hand prosthesis that addresses these design issues by maintaining stable prehension with minimal grip force. Constraints such as modularity, anthropomorphism, and low weight and power consumption are factors that have been adhered to throughout the design process.
Neural receptive fields are plastic: with experience, neurons in many brain regions change their spiking responses to relevant stimuli. Analysis of receptive field plasticity from experimental measurements is crucial for understanding how neural systems adapt their representations of relevant biological information. Current analysis methods using histogram estimates of spike rate functions in nonoverlapping temporal windows do not track the evolution of receptive field plasticity on a fine time scale. Adaptive signal processing is an established engineering paradigm for estimating time-varying system parameters from experimental measurements. We present an adaptive filter algorithm for tracking neural receptive field plasticity based on point process models of spike train activity. We derive an instantaneous steepest descent algorithm by using as the criterion function the instantaneous log likelihood of a point process spike train model. We apply the point process adaptive filter algorithm in a study of spatial (place) receptive field properties of simulated and actual spike train data from rat CA1 hippocampal neurons. A stability analysis of the algorithm is sketched in the. The adaptive algorithm can update the place field parameter estimates on a millisecond time scale. It reliably tracked the migration, changes in scale, and changes in maximum firing rate characteristic of hippocampal place fields in a rat running on a linear track. Point process adaptive filtering offers an analytic method for studying the dynamics of neural receptive fields.
We consider the problem of propagating the conditional probability density associated with the movement parameters (position, heading, velocity, etc.) of an animal, given the responses of an ensemble of place cells. While we are not the first to look at this question, ours seems to be the first treatment that incorporates a general Markov process model for the motion parameters and a general observation model postulating place cells centered in a lower dimensional 'measurement space' formed from combinations of the Markovian variables. An important part of our analysis involves the determination of a suitable set of sufficient statistics for propagating the conditional density in this context. Making use of these results we are led to approximations which greatly simplify the estimation problem and various aspects of its neuroscientific interpretation.
This paper reviews data acquisition and signal processing issues relative to producing an amplitude estimate of surface EMG. The paper covers two principle areas. First, methods for reducing noise, artefact and interference in recorded EMG are described. Wherever possible noise should be reduced at the source via appropriate skin preparation, and the use of well designed active electrodes and signal recording instrumentation. Despite these efforts, some noise will always accompany the desired signal, thus signal processing techniques for noise reduction (e.g. band-pass filtering, adaptive noise cancellation filters and filters based on the wavelet transform) are discussed. Second, methods for estimating the amplitude of the EMG are reviewed. Most advanced, high-fidelity methods consist of six sequential stages: noise rejection/filtering, whitening, multiple-channel combination, amplitude demodulation, smoothing and relinearization. Theoretical and experimental research related to each of the above topics is reviewed and the current recommended practices are described.
Many facets of neuromuscular activation patterns and control can be assessed via electromyography and are important for understanding the control of locomotion. After spinal cord injury, muscle activation patterns can affect locomotor recovery. We present a novel application of reversible jump Markov chain Monte Carlo simulation to estimate activation patterns from electromyographic data. We assume the data to be a zero-mean, heteroscedastic process. The variance is explicitly modeled using a step function. The number and location of points of discontinuity, or change-points, in the step function, the inter-change-point variances, and the overall mean are jointly modeled along with the mean and variance from baseline data. The number of change-points is considered a nuisance parameter and is integrated out of the posterior distribution. Whereas current methods of detecting activation patterns are deterministic or provide only point estimates, ours provides distributional estimates of muscle activation. These estimates, in turn, are used to estimate physiologically relevant quantities such as muscle coactivity, total integrated energy, and average burst duration and to draw valid statistical inferences about these quantities.
Accurate and computationally efficient means of classifying surface myoelectric signals has been the subject of considerable research effort in recent years. The aim of this paper is to classify myoelectric signals using new fuzzy clustering neural network (NN) architectures to control multifunction prostheses. This paper presents a comparative study of the classification accuracy of myoelectric signals using multilayered perceptron NN using back-propagation, conic section function NN, and new fuzzy clustering NNs (FCNNs). The myoelectric signals considered are used in classifying six upper-limb movements: elbow flexion, elbow extension, wrist pronation and wrist supination, grasp, and resting. The results suggest that FCNN can generalize better than other NN algorithms and help the user learn better and faster. This method has the potential of being very efficient in real-time applications.
Surface EMG is an important tool in biomechanics, kinesiology and neurophysiology. In neurophysiology the concept of high-density EMG (HD-EMG), using two dimensional electrode grids, was developed for the measurement of spatiotemporal activation patterns of the underlying muscle and its motor units (MU). The aim of this paper was to determine, with the aid of a HD-EMG grid, the relative importance of a number of electrode sensor configurations for optimizing muscle force estimation. Sensor configurations are distinguished in two categories. The first category concerns dimensions: the size of a single electrode and the inter electrode distance (IED). The second category concerns the sensor's spatial distribution: the total area from which signals are obtained (collection surface) and the number of electrodes per cm(2) (collection density). Eleven subjects performed isometric arm extensions at three elbow angles and three contraction levels. Surface-EMG from the triceps brachii muscle and the external force at the wrist were measured. Compared to a single conventional bipolar electrode pair, the force estimation quality improved by about 30% when using HD-EMG. Among the sensor configurations, the collection surface alone appeared to be responsible for the major part of the EMG based force estimation quality by improving it with 25%.
This paper represents an ongoing investigation of dexterous and natural control of upper extremity prostheses using the myoelectric signal. The scheme described within uses a hidden Markov model (HMM) to process four channels of myoelectric signal, with the task of discriminating six classes of limb movement. The HMM-based approach is shown to be capable of higher classification accuracy than previous methods based upon multilayer perceptrons. The method does not require segmentation of the myoelectric signal data, allowing a continuous stream of class decisions to be delivered to a prosthetic device. Due to the fact that the classifier learns the muscle activation patterns for each desired class for each individual, a natural control actuation results. The continuous decision stream allows complex sequences of manipulation involving multiple joints to be performed without interruption. The computational complexity of the HMM in its operational mode is low, making it suitable for a real-time implementation. The low computational overhead associated with training the HMM also enables the possibility of adaptive classifier training while in use.
This paper represents an ongoing investigation for surface myoelectric signal segmentation and classification. The classical moving average technique augmented with principal components analysis and time-frency analysis were used for segmentation. Multiresolution wavelet analysis was adopted as an effective feature extraction technique while artificial neural networks were used for classification. Results of classifying four elbow and wrist movement signals recorded from biceps and triceps gave 5.1% classification error when two channels were used.
Numerous studies have investigated the relationship between surface electromyogram (EMG) and torque exerted about a joint. Most studies have used conventional EMG amplitude (EMGamp) processing, such as rectification followed by low-pass filtering, to pre-process the EMG before relating it to torque. Recently, advanced EMGamp processors that incorporate signal whitening and multiple-channel combination have been shown to significantly improve EMGamp processing. In this study, we compared the performance of EMGamp-torque estimators with and without these advanced EMGamp processors. Fifteen subjects produced constant-posture, non-fatiguing, force-varying contractions about the elbow while torque and biceps/triceps EMG were recorded. EMGamp was related to torque using a linear FIR model. Both whitening and multiple-channel combination reduced EMG-torque errors and their combination provided an additive benefit. Using a 15th-order linear FIR model, EMG-torque errors with a four-channel, whitened processor averaged 7.3% of maximum voluntary contraction (MVC) (or 78% of variance accounted for). By comparison, the equivalent single-channel, unwhitened (conventional) processor produced an average error of 9.9% of MVC (variance accounted for of 55%). In addition, the study describes the occurrence of spurious peaks in estimated torque when the torque model is created from data with a sampling rate well above the bandwidth of the torque. This problem occurs when the torque data are sampled at the same rate as the EMG data. The problem is corrected by decimating the EMGamp prior to relating it to joint torque, in our case to an effective sampling rate of 40.96 Hz.