Influence of Joint Angle on EMG-Torque Model During Constant-Posture, Torque-Varying Contractions

Improved Surface Electromyogram-Based Hand-Wrist Force Estimation Using Deep Neural Networks and Cross-Joint Transfer Learning

Deep neural networks (DNNs) and transfer learning (TL) have been used to improve surface electromyogram (sEMG)-based force estimation. However, prior studies focused mostly on applying TL within one joint, which limits dataset size and diversity. Herein, we investigated cross-joint TL between two upper-limb joints with four DNN architectures using sliding windows. We used two feedforward and two recurrent DNN models with feature engineering and feature learning, respectively. We found that the dependencies between sEMG and force are short-term (<400 ms) and that sliding windows are sufficient to capture them, suggesting that more complicated recurrent structures may not be necessary. Also, using DNN architectures reduced the required sliding window length. A model pre-trained on elbow data was fine-tuned on hand-wrist data, improving force estimation accuracy and reducing the required training data amount. A convolutional neural network with a 391 ms sliding window fine-tuned using 20 s of training data had an error of 6.03 ± 0.49% maximum voluntary torque, which is statistically lower than both our multilayer perceptron model with TL and a linear regression model using 40 s of training data. The success of TL between two distinct joints could help enrich the data available for future deep learning-related studies.

Evaluation of generic EMG-Torque models across two Upper-Limb joints

Advanced single-use dynamic EMG-torque models require burdensome subject-specific calibration contractions and have historically been assumed to produce lower error than generic models (i.e., models that are identical across subjects and muscles). To investigate this assumption, we studied generic one degree of freedom (DoF) models derived from the ensemble median of subject-specific models, evaluated across subject, DoF and joint. We used elbow (N = 64) and hand-wrist (N = 9) datasets. Subject-specific elbow models performed statistically better [5.79 ± 1.89 %MVT (maximum voluntary torque) error] than generic elbow models (6.21 ± 1.85 %MVT error). However, there were no statistical differences between subject-specific vs. generic models within each hand-wrist DoF. Next, we evaluated generic models across joints. The best hand-wrist generic model had errors of 6.29 ± 1.85 %MVT when applied to the elbow. The elbow generic model had errors of 7.04 ± 2.29 %MVT when applied to the hand-wrist. The generic elbow model was statistically better in both joints, compared to the generic hand-wrist model. Finally, we tested Butterworth filter models (a simpler generic model), finding no statistical differences between optimum Butterworth and subject-specific models. Overall, generic models simplified EMG-torque training without substantive performance degradation and provided the possibility of transfer learning between joints.

Myoelectric Control Performance of Two Degree of Freedom Hand-Wrist Prosthesis by Able-Bodied and Limb-Absent Subjects

Recent research has advanced two degree-of-freedom (DoF), simultaneous, independent and proportional control of hand-wrist prostheses using surface electromyogram signals from remnant muscles as the control input. We evaluated two such regression-based controllers, along with conventional, sequential two-site control with co-contraction mode switching (SeqCon), in box-block, refined-clothespin and door-knob tasks, on 10 able-bodied and 4 limb-absent subjects. Subjects operated a commercial hand and wrist using a socket bypass harness. One 2-DoF controller (DirCon) related the intuitive hand actions of open-close and pronation-supination to the associated prosthesis hand-wrist actions, respectively. The other (MapCon) mapped myoelectrically more distinct, but less intuitive, actions of wrist flexion-extension and ulnar-radial deviation. Each 2-DoF controller was calibrated from separate 90 s calibration contractions. SeqCon performed better statistically than MapCon in the predominantly 1-DoF box-block task (>20 blocks/minute vs. 8-18 blocks/minute, on average). In this task, SeqCon likely benefited from an ability to easily focus on 1-DoF and not inadvertently trigger co-contraction for mode switching. The remaining two tasks require 2-DoFs, and both 2-DoF controllers each performed better (factor of 2-4) than SeqCon. We also compared the use of 12 vs. 6 optimally-selected EMG electrodes as inputs, finding no statistical difference. Overall, we provide further evidence of the benefits of regression-based EMG prosthesis control of 2-DoFs in the hand-wrist.

Leveraging Joint Mechanics Simplifies the Neural Control of Movement

Behaviors we perform each day, such as manipulating an object or walking, require precise control of the interaction forces between our bodies and the environment. These forces are generated by muscle contractions, specified by the nervous system, and by joint mechanics, determined by the intrinsic properties of the musculoskeletal system. Depending on behavioral goals, joint mechanics might simplify or complicate control of movement by the nervous system. Whether humans can exploit joint mechanics to simplify neural control remains unclear. Here we evaluated if leveraging joint mechanics simplifies neural control by comparing performance in three tasks that required subjects to generate specified torques about the ankle during imposed sinusoidal movements; only one task required torques that could be generated by leveraging the intrinsic mechanics of the joint. The complexity of the neural control was assessed by subjects' perceived difficulty and the resultant task performance. We developed a novel approach that used continuous estimates of ankle impedance, a quantitative description of the joint mechanics, and measures of muscle activity to determine the mechanical and neural contributions to the net ankle torque generated in each task. We found that the torque resulting from changes in neural control was reduced when ankle impedance was consistent with the task being performed. Subjects perceived this task to be easier than those that were not consistent with the impedance of the ankle and were able to perform it with the highest level of consistency across repeated trials. These results demonstrate that leveraging the mechanical properties of a joint can simplify task completion and improve performance.

Simplified Optimal Estimation of Time-Varying Electromyogram Standard Deviation (EMGσ): Evaluation on Two Datasets

To facilitate the broader use of EMG signal whitening, we studied four whitening procedures of various complexities, as well as the roles of sampling rate and noise correction. We separately analyzed force-varying and constant-force contractions from 64 subjects who completed constant-posture tasks about the elbow over a range of forces from 0% to 50% maximum voluntary contraction (MVC). From the constant-force tasks, we found that noise correction via the root difference of squares (RDS) method consistently reduced EMG recording noise, often by a factor of 5-10. All other primary results were from the force-varying contractions. Sampling at 4096 Hz provided small and statistically significant improvements over sampling at 2048 Hz (~3%), which, in turn, provided small improvements over sampling at 1024 Hz (~4%). In comparing equivalent processing variants at a sampling rate of 4096 Hz, whitening filters calibrated to the EMG spectrum of each subject generally performed best (4.74% MVC EMG-force error), followed by one universal whitening filter for all subjects (4.83% MVC error), followed by a high-pass filter whitening method (4.89% MVC error) and then a first difference whitening filter (4.91% MVC error)-but none of these statistically differed. Each did significantly improve from EMG-force error without whitening (5.55% MVC). The first difference is an excellent whitening option over this range of contraction forces since no calibration or algorithm decisions are required.

Data Management for Transfer Learning Approaches to Elbow EMG-Torque Modeling

The performance of single-use subject-specific electromyogram (EMG)-torque models degrades significantly when used on a new subject, or even the same subject on a second day. Improving the generalization performance of models is essential but challenging. In this work, we investigate how data management strategies contribute to the performance of elbow joint EMG-torque models in cross-subject evaluation. Data management can be divided into two parts, namely data acquisition and data utilization. For data acquisition, analysis of data from 65 subjects shows that training set data diversity (number of subjects) is more important than data size (total data duration). For data utilization, we propose a correlation-based data weighting (COR-W) method for model calibration which is unsupervised in the modeling stage. We first evaluated the domain shift level between data in each training trial (source domain) and data acquired from a new subject (target domain) via the mismatch of feature correlation, using only EMG signals in the target domain without the synchronized torque values (hence unsupervised during model training). Data weights were assigned to each training trial according to different domain shift levels. The weighted least squares method using the obtained data weights was then employed to develop a calibrated EMG-torque model for the new subject. The COR-W method can achieve a low root mean square error (9.29% maximum voluntary contraction) in cross-subject evaluation, with significant performance improvement compared to models without calibration. Both the data acquisition and utilization strategies contribute to the performance of EMG-torque models in cross-subject evaluation.

EMG-Force and EMG-Target Models During Force-Varying Bilateral Hand-Wrist Contraction in Able-Bodied and Limb-Absent Subjects

System identification models relating forearm electromyogram (EMG) signals to phantom wrist radial-ulnar deviation force, pronation-supination moment and/or hand open-close force (EMG-force) are hampered by lack of supervised force/moment output signals in limb-absent subjects. In 12 able-bodied and 7 unilateral transradial limb-absent subjects, we studied three alternative supervised output sources in one degree of freedom (DoF) and 2-DoF target tracking tasks: (1) bilateral tracking with force feedback from the contralateral side (non-dominant for able-bodied/ sound for limb-absent subjects) with the contralateral force as the output, (2) bilateral tracking with force feedback from the contralateral side with the target as the output, and (3) dominant/limb-absent side unilateral target tracking without feedback and the target used as the output. "Best-case" EMG-force errors averaged ~ 10% of maximum voluntary contraction (MVC) when able-bodied subjects' dominant limb produced unilateral force/moment with feedback. When either bilateral tracking source was used as the model output, statistically larger errors of 12-16 %MVC resulted. The no-feedback alternative produced errors of 25-30 %MVC, which was nearly half the tested force range of ± 30 %MVC. Therefore, the no-feedback model output was not acceptable. We found little performance variation between DoFs. Many subjects struggled to perform 2-DoF target tracking.

Efficiently Training Two-DoF Hand-Wrist EMG-Force Models

Single-use EMG-force models (i.e., a new model is trained each time the electrodes are donned) are used in various areas, including ergonomics assessment, clinical biomechanics, and motor control research. For one degree of freedom (1-DoF) tasks, input-output (black box) models are common. Recently, black box models have expanded to 2-DoF tasks. To facilitate efficient training, we examined parameters of black box model training methods in 2-DoF force-varying, constant-posture tasks consisting of hand open-close combined with one wrist DoF. We found that approximately 40-60 s of training data is best, with progressively higher EMG-force errors occurring for progressively shorter training durations. Surprisingly, 2-DoF models in which the dynamics were universal across all subjects (only channel gain was trained to each subject) generally performed 15-21% better than models in which the complete dynamics were trained to each subject. In summary, lower error EMG-force models can be formed through diligent attention to optimization of these factors.

Normalising surface EMG of ten upper-extremity muscles in handcycling: Manual resistance vs. sport-specific MVICs

Muscular activity in terms of surface electromyography (sEMG) is usually normalised to maximal voluntary isometric contractions (MVICs). This study aims to compare two different MVIC-modes in handcycling and examine the effect of moving average window-size. Twelve able-bodied male competitive triathletes performed ten MVICs against manual resistance and four sport-specific trials against fixed cranks. sEMG of ten muscles [M. trapezius (TD); M. pectoralis major (PM); M. deltoideus, Pars clavicularis (DA); M. deltoideus, Pars spinalis (DP); M. biceps brachii (BB); M. triceps brachii (TB); forearm flexors (FC); forearm extensors (EC); M. latissimus dorsi (LD) and M. rectus abdominis (RA)] was recorded and filtered using moving average window-sizes of 150, 200, 250 and 300 ms. Sport-specific MVICs were higher compared to manual resistance for TB, DA, DP and LD, whereas FC, TD, BB and RA demonstrated lower values. PM and EC demonstrated no significant difference between MVIC-modes. Moving average window-size had no effect on MVIC outcomes. MVIC-mode should be taken into account when normalised sEMG data are illustrated in handcycling. Sport-specific MVICs seem to be suitable for some muscles (TB, DA, DP and LD), but should be augmented by MVICs against manual/mechanical resistance for FC, TD, BB and RA.

Implementation and experimental validation of surface electromyogram and force model of Tibialis Anterior muscle for examining muscular factors

This study reports a surface electromyogram and force of contraction model. The objective was to investigate the effect of changes in the size, type and number of motor units in the Tibialis Anterior muscle to surface electromyogram and force of dorsiflexion. A computational model to simulate surface electromyogram and associated force of contraction by the Tibialis Anterior muscle was developed. This model was simulated for isometric dorsiflexion, and comparative experiments were conducted for validation. Repeated simulations were performed to investigate the different parameters and evaluate inter-experimental variability. An equivalence statistical test and the Bland–Altman method were used to observe the significance between the simulated and experimental data. Simulated and experimentally recorded data had high similarity for the three measures: maximal power of power spectral density ( p < 0.0001), root mean square of surface electromyogram ( p < 0.0001) and force recorded at the footplate ( p < 0.03). Inter-subject variability in the experimental results was in-line with the variability in the repeated simulation results. This experimentally validated computational model for the surface electromyogram and force of the Tibialis Anterior muscle is significant as it allows the examination of three important muscular factors associated with ageing and disease: size, fibre type and number of motor units.

Optimal Estimation of EMG Standard Deviation (EMG σ ) in Additive Measurement Noise: Model-Based Derivations and Their Implications

Typical electromyogram (EMG) processors estimate EMG signal standard deviation (EMG σ ) via moving average root mean square (RMS) or mean absolute value (MAV) filters, whose outputs are used in force estimation, prosthesis/orthosis control, etc. In the inevitable presence of additive measurement noise, some processors subtract the noise standard deviation from EMG RMS (or MAV). Others compute a root difference of squares (RDS)-subtract the noise variance from the square of EMG RMS (or MAV), all followed by taking the square root. Herein, we model EMG as an amplitude-modulated random process in additive measurement noise. Assuming a Gaussian (or, separately, Laplacian) distribution, we derive analytically that the maximum likelihood estimate of EMG σ requires RDS processing. Whenever that subtraction would provide a negative-valued result, we show that EMG σ should be set to zero. Our theoretical models further show that during rest, approximately 50% of EMG σ estimates are non-zero. This result is problematic when EMG σ is used for real-time control, explaining the common use of additional thresholding. We tested our model results experimentally using biceps and triceps EMG from 64 subjects. Experimental results closely followed the Gaussian model. We conclude that EMG processors should use RDS processing and not noise standard deviation subtraction.

Two degrees of freedom, dynamic, hand-wrist EMG-force using a minimum number of electrodes

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.

Single-trial estimation of quasi-static EMG-to-joint-mechanical-impedance relationship over a range of joint torques

Joint mechanical impedance is commonly measured by applying dynamic perturbations about a joint at a fixed operating point/background torque, and quantifying torque change vs. angle change. Impedance characterization in functional tasks, therefore, requires multiple experimental trials over a range of operating points-a cumbersome, invasive, time-consuming and impractical task. As an alternative, studies have related EMG to impedance, after which EMG can estimate impedance without applying joint perturbations. But, the cumbersome calibration trials are still required. We describe a method of single contraction perturbations in which the background torque slowly ramps over the operating range, with EMG simultaneously acquired. Using one such "quasi-static" contraction for model training and another for testing, we show this method to be a reasonable surrogate for traditional second-order, linear impedance modeling. A simple, short-duration calibration results. We compared our single-trial ramp method to multiple constant background torque trials at 10, 20, 30, and 40% maximum effort (extension and flexion), finding only limited differences in traditional vs. EMG-based ramp impedance estimates (12-22%, most prominent at the two lower contraction levels). Such constant force and slowly-variable force contractions are relevant to many practical applications, including ergonomics assessment, prosthetic control and clinical biomechanics.

A Review on the Control of the Mechanical Properties of Ankle Foot Orthosis for Gait Assistance

In the past decade, advanced technologies in robotics have been explored to enhance the rehabilitation of post-stroke patients. Previous works have shown that gait assistance for post-stroke patients can be provided through the use of robotics technology in ancillary equipment, such as Ankle Foot Orthosis (AFO). An AFO is usually used to assist patients with spasticity or foot drop problems. There are several types of AFOs, depending on the flexibility of the joint, such as rigid, flexible rigid, and articulated AFOs. A rigid AFO has a fixed joint, and a flexible rigid AFO has a more flexible joint, while the articulated AFO has a freely rotating ankle joint, where the mechanical properties of the AFO are more controllable compared to the other two types of AFOs. This paper reviews the control of the mechanical properties of existing AFOs for gait assistance in post-stroke patients. Several aspects that affect the control of the mechanical properties of an AFO, such as the controller input, number of gait phases, controller output reference, and controller performance evaluation are discussed and compared. Thus, this paper will be of interest to AFO researchers or developers who would like to design their own AFOs with the most suitable mechanical properties based on their application. The controller input and the number of gait phases are discussed first. Then, the discussion moves forward to the methods of estimating the controller output reference, which is the main focus of this study. Based on the estimation method, the gait control strategies can be classified into subject-oriented estimations and phase-oriented estimations. Finally, suggestions for future studies are addressed, one of which is the application of the adaptive controller output reference to maximize the benefits of the AFO to users.

Real-time, simultaneous myoelectric control using a convolutional neural network

The evolution of deep learning techniques has been transformative as they have allowed complex mappings to be trained between control inputs and outputs without the need for feature engineering. In this work, a myoelectric control system based on convolutional neural networks (CNN) is proposed as a possible alternative to traditional approaches that rely on specifically designed features. This CNN-based system is validated using a real-time Fitts' law style target acquisition test requiring single and combined wrist motions. The performance of the proposed system is then compared to that of a standard support vector machine (SVM) based myoelectric system using a set of time-domain features. Despite the prevalence and demonstrated performance of these well-known features, no significant difference (p>0.05) was found between the two methods for any of the computed control metrics. This demonstrates the potential for automated learning approaches to extract complex and rich information from stochastic biological signals. This first evaluation of the usability of a CNN in a real-time myoelectric control environment provides a basis for further exploration.

Two degrees of freedom quasi-static EMG-force at the wrist using a minimum number of electrodes

Surface electromyogram-controlled powered hand/wrist prostheses return partial upper-limb function to limb-absent persons. Typically, one degree of freedom (DoF) is controlled at a time, with mode switching between DoFs. Recent research has explored using large-channel EMG systems to provide simultaneous, independent and proportional (SIP) control of two joints-but such systems are not practical in current commercial prostheses. Thus, we investigated site selection of a minimum number of conventional EMG electrodes in an EMG-force task, targeting four sites for a two DoF controller. In a laboratory experiment with 10 able-bodied subjects and three limb-absent subjects, 16 electrodes were placed about the proximal forearm. Subjects produced 1-DoF and 2-DoF slowly force-varying contractions up to 30% maximum voluntary contraction (MVC). EMG standard deviation was related to forces via regularized regression. Backward stepwise selection was used to retain those progressively fewer electrodes that exhibited minimum error. For 1-DoF models using two retained electrodes (which mimics the current state of the art), subjects had average RMS errors of (depending on the DoF): 7.1-9.5% MVC for able-bodied and 13.7-17.1% MVC for limb-absent subjects. For 2-DoF models, subjects using four electrodes had errors on 1-DoF trials of 6.7-8.5% MVC for able-bodied and 11.9-14.0% MVC for limb-absent; and errors on 2-DoF trials of 9.9-11.2% MVC for able-bodied and 15.8-16.7% MVC for limb-absent subjects. For each model, retaining more electrodes did not statistically improve performance. The able-bodied results suggest that backward selection is a viable method for minimum error selection of as few as four electrode sites for these EMG-force tasks. Performance evaluation in a prosthesis control task is a necessary and logical next step for this site selection method.

Dynamic Elbow Flexion Force Estimation Through a Muscle Twitch Model and sEMG in a Fatigue Condition

We propose a joint force estimation method to compute elbow flexion force using surface electromyogram (sEMG) considering time-varying effects in a fatigue condition. Muscle fatigue is a major cause inducing sEMG changes with respect to time over long periods and repetitive contractions. The proposed method composed the muscle-twitch model representing the force generated by a single spike and the spikes extracted from sEMG. In this study, isometric contractions at six different joint angles (10 subjects) and dynamic contractions with constant velocity (six subjects) were performed under non-fatigue and fatigue conditions. Performance of the proposed method was evaluated and compared with that of previous methods using mean absolute value (MAV). The proposed method achieved average 6.7 ± 2.8 %RMSE for isometric contraction and 15.6 ± 24.7%RMSE for isokinetic contraction under fatigue condition with more accurate results than the previous methods.

Comparison of Constant-Posture Force-Varying EMG-Force Dynamic Models About the Elbow

Numerous techniques have been used to minimize error in relating the surface electromyogram (EMG) to elbow joint torque. We compare the use of three techniques to further reduce error. First, most EMG-torque models only use estimates of EMG standard deviation as inputs. We studied the additional features of average waveform length, slope sign change rate and zero crossing rate. Second, multiple channels of EMG from the biceps, and separately from the triceps, have been combined to produce two low-variance model inputs. We contrasted this channel combination with using each EMG separately. Third, we previously modeled nonlinearity in the EMG-torque relationship via a polynomial. We contrasted our model versus that of the classic exponential power law of Vredenbregt and Rau (1973). Results from 65 subjects performing constant-posture, force-varying contraction gave a "baseline" comparison error (i.e., error with none of the new techniques) of 5.5 ± 2.3% maximum flexion voluntary contraction (%MVC_F). Combining the techniques of multiple features with individual channels reduced error to 4.8 ± 2.2 %MVC_F, while combining individual channels with the power-law model reduced error to 4.7 ± 2.0 %MVC_F. The new techniques further reduced error from that of the baseline by ≈ 15 %.