Multiday Evaluation of Techniques for EMG-Based Classification of Hand Motions

Effectiveness of Immersive Virtual Reality-Based Hand Rehabilitation Games for Improving Hand Motor Functions in Subacute Stroke Patients

Stroke rehabilitation faces challenges in attaining enduring improvements in hand motor function and is frequently constrained by interventional limitations. This research aims to present an innovative approach to the integration of cognitive engagement within visual feedback incorporated into fully immersive virtual reality (VR) based games to achieve enduring improvements. These innovative aspects of interaction provide more functional advantages beyond motivation to efficiently execute repeatedly hand motor tasks. The effectiveness of virtual reality games incorporated with innovative aspects has been investigated for improvements in hand motor functions. A randomized controlled trial was conducted, a total of (n=56) subacute stroke patients were assessed for eligibility and (n=52) patients fulfilled the inclusion criteria. (n=26) patients were assigned to the experimental group and (n=26) patients were assigned to the control group. VR intervention involves four VR based games, developed based on hand movements including flexion/extension, close/open, supination/pronation and pinch. All patients got therapy of 24 sessions, lasting 4 days/week for a total of 6 weeks. Five clinical outcome measures were Fugl- Meyer Assessment-Upper Extremity, Action Research Arm Test, Box and Block Test, Modified Barthel Index, and Stroke-Specific Quality of Life were assessed to evaluate patients' performance. Results revealed that after therapy there was significant improvement between the groups (p<0.05) and within groups (p<0.05) in all assessment weeks in all clinical outcome measures however, improvement was observed significantly greater in the experimental group due to fully immersive VR-based games. Results indicated that cognitive engagement within visual feedback incorporated in VR-based hand games effectively improved hand motor functions.

On Automated Object Grasping for Intelligent Prosthetic Hands Using Machine Learning

Prosthetic technology has witnessed remarkable advancements, yet challenges persist in achieving autonomous grasping control while ensuring the user's experience is not compromised. Current electronic prosthetics often require extensive training for users to gain fine motor control over the prosthetic fingers, hindering their usability and acceptance. To address this challenge and improve the autonomy of prosthetics, this paper proposes an automated method that leverages computer vision-based techniques and machine learning algorithms. In this study, three reinforcement learning algorithms, namely Soft Actor-Critic (SAC), Deep Q-Network (DQN), and Proximal Policy Optimization (PPO), are employed to train agents for automated grasping tasks. The results indicate that the SAC algorithm achieves the highest success rate of 99% among the three algorithms at just under 200,000 timesteps. This research also shows that an object's physical characteristics can affect the agent's ability to learn an optimal policy. Moreover, the findings highlight the potential of the SAC algorithm in developing intelligent prosthetic hands with automatic object-gripping capabilities.

Robust myoelectric pattern recognition methods for reducing users' calibration burden: challenges and future

Myoelectric pattern recognition (MPR) has evolved into a sophisticated technology widely employed in controlling myoelectric interface (MI) devices like prosthetic and orthotic robots. Current MIs not only enable multi-degree-of-freedom control of prosthetic limbs but also demonstrate substantial potential in consumer electronics. However, the non-stationary random characteristics of myoelectric signals poses challenges, leading to performance degradation in practical scenarios such as electrode shifting and switching new users. Conventional MIs often necessitate meticulous calibration, imposing a significant burden on users. To address user frustration during the calibration process, researchers have focused on identifying MPR methods that alleviate this burden. This article categorizes common scenarios that incur calibration burdens as based on data distribution shift and based on dynamic data categories. Then further investigated and summarized the popular robust MPR algorithms used to reduce the user's calibration burden. We categorize these algorithms as based on data manipulate, feature manipulation and, model structure. And describes the scenarios to which each method is applicable and the conditions required for calibration. Finally, this review is concluded with the advantages of robust MPR and the remaining challenges and future opportunities.

One-shot random forest model calibration for hand gesture decoding

Objective.Most existing machine learning models for myoelectric control require a large amount of data to learn user-specific characteristics of the electromyographic (EMG) signals, which is burdensome. Our objective is to develop an approach to enable the calibration of a pre-trained model with minimal data from a new myoelectric user.Approach.We trained a random forest (RF) model with EMG data from 20 people collected during the performance of multiple hand grips. To adapt the decision rules for a new user, first, the branches of the pre-trained decision trees were pruned using the validation data from the new user. Then new decision trees trained merely with data from the new user were appended to the pruned pre-trained model.Results.Real-time myoelectric experiments with 18 participants over two days demonstrated the improved accuracy of the proposed approach when compared to benchmark user-specific RF and the linear discriminant analysis models. Furthermore, the RF model that was calibrated on day one for a new participant yielded significantly higher accuracy on day two, when compared to the benchmark approaches, which reflects the robustness of the proposed approach.Significance.The proposed model calibration procedure is completely source-free, that is, once the base model is pre-trained, no access to the source data from the original 20 people is required. Our work promotes the use of efficient, explainable, and simple models for myoelectric control.

Functionally Adaptive Myosite Selection Using High-Density sEMG for Upper Limb Myoelectric Prostheses

OBJECTIVE

Our study defines a novel electrode placement method called Functionally Adaptive Myosite Selection (FAMS), as a tool for rapid and effective electrode placement during prosthesis fitting. We demonstrate a method for determining electrode placement that is adaptable towards individual patient anatomy and desired functional outcomes, agnostic to the type of classification model used, and provides insight into expected classifier performance without training multiple models.

METHODS

FAMS relies on a separability metric to rapidly predict classifier performance during prosthesis fitting.

RESULTS

The results show a predictable relationship between the FAMS metric and classifier accuracy (3.45%SE), allowing estimation of control performance with any given set of electrodes. Electrode configurations selected using the FAMS metric show improved control performance ( ) for target electrode counts compared to established methods when using an ANN classifier, and equivalent performance ( R2 ≥ .96) to previous top-performing methods on an LDA classifier, with faster convergence ( ). We used the FAMS method to determine electrode placement for two amputee subjects by using the heuristic to search through possible sets, and checking for saturation in performance vs electrode count. The resulting configurations that averaged 95.8% of the highest possible classification performance using a mean 25 number of electrodes (19.5% of the available sites).

SIGNIFICANCE

FAMS can be used to rapidly approximate the tradeoffs between increased electrode count and classifier performance, a useful tool during prosthesis fitting.

Concurrent Prediction of Dexterous Finger Flexion and Extension Force via Deep Forest

Neuromuscular injuries can impair hand function and profoundly impacting the quality of life. This has motivated the development of advanced assistive robotic hands. However, the current neural decoder systems are limited in their ability to provide dexterous control of these robotic hands. In this study, we propose a novel method for predicting the extension and flexion force of three individual fingers concurrently using high-density electromyogram (HD-EMG) signals. Our method employs two deep forest models, the flexor decoder and the extensor decoder, to extract relevant representations from the EMG amplitude features. The outputs of the two decoders are integrated through linear regression to predict the forces of the three fingers. The proposed method was evaluated on data from three subjects and the results showed that it consistently outperforms the conventional EMG amplitude-based approach in terms of prediction error and robustness across both target and non-target fingers. This work presents a promising neural decoding approach for intuitive and dexterous control of the fingertip forces of assistive robotic hands.

Simultaneous assessment and training of an upper-limb amputee using incremental machine-learning-based myocontrol: a single-case experimental design

BACKGROUND

Machine-learning-based myocontrol of prosthetic devices suffers from a high rate of abandonment due to dissatisfaction with the training procedure and with the reliability of day-to-day control. Incremental myocontrol is a promising approach as it allows on-demand updating of the system, thus enforcing continuous interaction with the user. Nevertheless, a long-term study assessing the efficacy of incremental myocontrol is still missing, partially due to the lack of an adequate tool to do so. In this work we close this gap and report about a person with upper-limb absence who learned to control a dexterous hand prosthesis using incremental myocontrol through a novel functional assessment protocol called SATMC (Simultaneous Assessment and Training of Myoelectric Control).

METHODS

The participant was fitted with a custom-made prosthetic setup with a controller based on Ridge Regression with Random Fourier Features (RR-RFF), a non-linear, incremental machine learning method, used to build and progressively update the myocontrol system. During a 13-month user study, the participant performed increasingly complex daily-living tasks, requiring fine bimanual coordination and manipulation with a multi-fingered hand prosthesis, in a realistic laboratory setup. The SATMC was used both to compose the tasks and continually assess the participant's progress. Patient satisfaction was measured using Visual Analog Scales.

RESULTS

Over the course of the study, the participant progressively improved his performance both objectively, e.g., the time required to complete each task became shorter, and subjectively, meaning that his satisfaction improved. The SATMC actively supported the improvement of the participant by progressively increasing the difficulty of the tasks in a structured way. In combination with the incremental RR-RFF allowing for small adjustments when required, the participant was capable of reliably using four actions of the prosthetic hand to perform all required tasks at the end of the study.

CONCLUSIONS

Incremental myocontrol enabled an upper-limb amputee to reliably control a dexterous hand prosthesis while providing a subjectively satisfactory experience. The SATMC can be an effective tool to this aim.

Myoelectric Pattern Recognition Using Gramian Angular Field and Convolutional Neural Networks for Muscle-Computer Interface

In the field of the muscle-computer interface, the most challenging task is extracting patterns from complex surface electromyography (sEMG) signals to improve the performance of myoelectric pattern recognition. To address this problem, a two-stage architecture, consisting of Gramian angular field (GAF)-based 2D representation and convolutional neural network (CNN)-based classification (GAF-CNN), is proposed. To explore discriminant channel features from sEMG signals, sEMG-GAF transformation is proposed for time sequence signal representation and feature modeling, in which the instantaneous values of multichannel sEMG signals are encoded in image form. A deep CNN model is introduced to extract high-level semantic features lying in image-form-based time sequence signals concerning instantaneous values for image classification. An insight analysis explains the rationale behind the advantages of the proposed method. Extensive experiments are conducted on benchmark publicly available sEMG datasets, i.e., NinaPro and CagpMyo, whose experimental results validate that the proposed GAF-CNN method is comparable to the state-of-the-art methods, as reported by previous work incorporating CNN models.

Intramuscular EMG feature extraction and evaluation at different arm positions and hand postures based on a statistical criterion method

Prostheses control using electromyography signals have shown promising aspects in various fields including rehabilitation sciences and assistive technology controlled devices. Pattern recognition and machine learning methods have been observed to play a significant role in evaluating features and classifying different limb motions for enhanced prosthetic executions. This paper proposes feature extraction and evaluation method using intramuscular electromyography (iEMG) signals at different arm positions and hand postures based on the RES Index value statistical criterion method. Sixteen-time domain features were selected for the study at two main circumstances; fixed arm position (FAP) and fixed hand posture (FHP). Eight healthy male participants (30.62 ± 3.87 years) were asked to execute five motion classes including hand grip, hand open, rest, hand extension, and hand flexion at four different arm positions that comprise of 0°, 45°, 90°, and 135°. The classification process is accomplished via the application of the k-nearest neighbor (KNN) classifier. Then RES index was calculated to investigate the optimal features based on the proposed statistical criterion method. From the RES Index, we concluded that Variance (VAR) is the best feature while WAMP, Zero Crossing (ZC), and Slope Sign Change (SSC) are the worst ones in FAP conditions. On the contrary, we concluded that Average Amplitude Change (AAC) is the best feature while WAMP and Simple Square Integral (SSI) resulted in least RES Index values for FHP conditions. The proposed study has possible iEMG based applications such as assistive control devices, robotics. Also, working with the frequency domain features encapsulates the future scope of the study.

Automated Bone Marrow Cell Classification for Haematological Disease Diagnosis Using Siamese Neural Network

The critical structure and nature of different bone marrow cells which form a base in the diagnosis of haematological ailments requires a high-grade classification which is a very prolonged approach and accounts for human error if performed manually, even by field experts. Therefore, the aim of this research is to automate the process to study and accurately classify the structure of bone marrow cells which will help in the diagnosis of haematological ailments at a much faster and better rate. Various machine learning algorithms and models, such as CNN + SVM, CNN + XGB Boost and Siamese network, were trained and tested across a dataset of 170,000 expert-annotated cell images from 945 patients' bone marrow smears with haematological disorders. The metrics used for evaluation of this research are accuracy of model, precision and recall of all the different classes of cells. Based on these performance metrics the CNN + SVM, CNN + XGB, resulted in 32%, 28% accuracy, respectively, and therefore these models were discarded. Siamese neural resulted in 91% accuracy and 84% validation accuracy. Moreover, the weighted average recall values of the Siamese neural network were 92% for training and 91% for validation. Hence, the final results are based on Siamese neural network model as it was outperforming all the other algorithms used in this research.

Surface EMG Statistical and Performance Analysis of Targeted-Muscle-Reinnervated (TMR) Transhumeral Prosthesis Users in Home and Laboratory Settings

A pattern-recognition (PR)-based myoelectric control system is the trend of future prostheses development. Compared with conventional prosthetic control systems, PR-based control systems provide high dexterity, with many studies achieving >95% accuracy in the last two decades. However, most research studies have been conducted in the laboratory. There is limited research investigating how EMG signals are acquired when users operate PR-based systems in their home and community environments. This study compares the statistical properties of surface electromyography (sEMG) signals used to calibrate prostheses and quantifies the quality of calibration sEMG data through separability indices, repeatability indices, and correlation coefficients in home and laboratory settings. The results demonstrate no significant differences in classification performance between home and laboratory environments in within-calibration classification error (home: 6.33 ± 2.13%, laboratory: 7.57 ± 3.44%). However, between-calibration classification errors (home: 40.61 ± 9.19%, laboratory: 44.98 ± 12.15%) were statistically different. Furthermore, the difference in all statistical properties of sEMG signals is significant (p < 0.05). Separability indices reveal that motion classes are more diverse in the home setting. In summary, differences in sEMG signals generated between home and laboratory only affect between-calibration performance.

Random Channel Masks for Regularization of Least Squares-Based Finger EMG-Force Modeling to Improve Cross-Day Performance

Estimating the finger forces from surface electromyography (sEMG) is essential for diverse applications (e.g., human-machine interfacing). The performance of pre-trained sEMG-force models degenerates significantly when applied on a second day, due to the large cross-day variation of sEMG characteristics. Previous studies mainly employed transfer learning algorithms to tackle this problem. However, transfer learning algorithms normally require data collected on the second day for model calibration, increasing the inconvenience in practical use. In this work, we investigated the effect of model regularization on this issue. Specifically, 256-channel high-density sEMG (HDsEMG) signals with varying finger forces were collected on different days (3-25 days apart). We applied randomly generated channel perturbations ("masks") to feature maps of randomly selected channels in training dataset. The channel masks of the training set were generated randomly and independently in each narrow time window (~20 ms). We assumed that by learning from randomly masked feature maps (randomness is the central aspect), the model would not be biased by a small number of features but would be based on learning from a global perspective, therefore avoiding overfitting to the within-day EMG patterns. Moore-Penrose inverse model regularization was also employed as a baseline method, with results showing that cross-day EMG-force models require a higher tolerance parameter compared with within-day applications. In combination with the Moore-Penrose inverse model regularization, further applying random channel masks to the training set significantly improved model performance in cross-day validation.

Myoelectric Pattern Recognition Performance Enhancement Using Nonlinear Features

The multichannel electrode array used for electromyogram (EMG) pattern recognition provides good performance, but it has a high cost, is computationally expensive, and is inconvenient to wear. Therefore, researchers try to use as few channels as possible while maintaining improved pattern recognition performance. However, minimizing the number of channels affects the performance due to the least separable margin among the movements possessing weak signal strengths. To meet these challenges, two time-domain features based on nonlinear scaling, the log of the mean absolute value (LMAV) and the nonlinear scaled value (NSV), are proposed. In this study, we validate the proposed features on two datasets, the existing four feature extraction methods, variable window size, and various signal-to-noise ratios (SNR). In addition, we also propose a feature extraction method where the LMAV and NSV are grouped with the existing 11 time-domain features. The proposed feature extraction method enhances accuracy, sensitivity, specificity, precision, and F1 score by 1.00%, 5.01%, 0.55%, 4.71%, and 5.06% for dataset 1, and 1.18%, 5.90%, 0.66%, 5.63%, and 6.04% for dataset 2, respectively. Therefore, the experimental results strongly suggest the proposed feature extraction method, for taking a step forward with regard to improved myoelectric pattern recognition performance.

A Hand-Modeled Feature Extraction-Based Learning Network to Detect Grasps Using sEMG Signal

Recently, deep models have been very popular because they achieve excellent performance with many classification problems. Deep networks have high computational complexities and require specific hardware. To overcome this problem (without decreasing classification ability), a hand-modeled feature selection method is proposed in this paper. A new shape-based local feature extractor is presented which uses the geometric shape of the frustum. By using a frustum pattern, textural features are generated. Moreover, statistical features have been extracted in this model. Textures and statistics features are fused, and a hybrid feature extraction phase is obtained; these features are low-level. To generate high level features, tunable Q factor wavelet transform (TQWT) is used. The presented hybrid feature generator creates 154 feature vectors; hence, it is named Frustum154. In the multilevel feature creation phase, this model can select the appropriate feature vectors automatically and create the final feature vector by merging the appropriate feature vectors. Iterative neighborhood component analysis (INCA) chooses the best feature vector, and shallow classifiers are then used. Frustum154 has been tested on three basic hand-movement sEMG datasets. Hand-movement sEMG datasets are commonly used in biomedical engineering, but there are some problems in this area. The presented models generally required one dataset to achieve high classification ability. In this work, three sEMG datasets have been used to test the performance of Frustum154. The presented model is self-organized and selects the most informative subbands and features automatically. It achieved 98.89%, 94.94%, and 95.30% classification accuracies using shallow classifiers, indicating that Frustum154 can improve classification accuracy.

Towards zero retraining for multiday motion recognition via a fully unsupervised adaptive approach and fabric myoelectric armband

Surface electromyogram pattern recognition (EMG-PR) requires time-consuming training and retraining procedures for long-term use, hindering the usability of myoelectric control. In this paper, we design a fabric myoelectric armband to reduce the electrode shifts. Furthermore, we propose a fully unsupervised adaptive approach called hybrid serial classifier (HSC) to eliminate the burden of retraining over multiply days. We investigated the performance of our approach with a dataset of ten types of forearm motion from ten male subjects over eight weeks (total ten days, including: from day 1 to day 7, day 14, day 28, day 56). The average inter-day classification accuracies of HSC without any new retraining data are 86.61% when trained exclusively with the first day's EMG data, and 94.77% when trained with other nine days' data. We compare our proposed HSC algorithm with linear discriminant analysis (LDA) without recalibration (BLDA) and supervised adaption LDA (ALDA) with just one trial of new retraining data. The inter-day classification accuracy of HSC is significantly higher than that of BLDA and ALDA. These results indicate that our novel armband sEMG device is feasible for long-term use in conjunction with the proposed HSC algorithm.

Classification of high knee flexion postures using EMG signals

BACKGROUND

High knee flexion postures are often adopted in occupational settings and may lead to increased risk of knee osteoarthritis. Pattern recognition algorithms using wireless electromyographic (EMG) signals may be capable of detecting and quantifying occupational exposures throughout a working day.

OBJECTIVE

To develop a k-Nearest Neighbor (kNN) algorithm for the classification of eight high knee flexion activities frequently observed in childcare.

METHODS

EMG signals from eight lower limb muscles were recorded for 30 participants, signals were decomposed into time- and frequency-domain features, and used to develop a kNN classification algorithm. Features were reduced to a combination of ten time-domain features from 8 muscles using neighborhood component analysis, in order to most effectively identify the postures of interest.

RESULTS

The final classifier was capable of accurately identifying 80.1%of high knee flexion postures based on novel data from participants included in the training dataset, yet only achieved 18.4%accuracy when predicting postures based on novel subject data.

CONCLUSIONS

EMG based classification of high flexion postures may be possible within occupational settings when the model is first trained on sample data from a given individual. The developed algorithm may provide quantitative measures leading to a greater understanding of occupation specific postural requirements.

Inter-Subject Domain Adaptation for CNN-Based Wrist Kinematics Estimation Using sEMG

Recently, convolutional neural network (CNN) has been widely investigated to decode human intentions using surface Electromyography (sEMG) signals. However, a pre-trained CNN model usually suffers from severe degradation when testing on a new individual, and this is mainly due to domain shift where characteristics of training and testing sEMG data differ substantially. To enhance inter-subject performances of CNN in the wrist kinematics estimation, we propose a novel regression scheme for supervised domain adaptation (SDA), based on which domain shift effects can be effectively reduced. Specifically, a two-stream CNN with shared weights is established to exploit source and target sEMG data simultaneously, such that domain-invariant features can be extracted. To tune CNN weights, both regression losses and a domain discrepancy loss are employed, where the former enable supervised learning and the latter minimizes distribution divergences between two domains. In this study, eight healthy subjects were recruited to perform wrist flexion-extension movements. Experiment results illustrated that the proposed regression SDA outperformed fine-tuning, a state-of-the-art transfer learning method, in both single-single and multiple-single scenarios of kinematics estimation. Unlike fine-tuning which suffers from catastrophic forgetting, regression SDA can maintain much better performances in original domains, which boosts the model reusability among multiple subjects.

EMG based classification for pick and place task

The hand amputee is deprived of number of activities of daily living. To help the hand amputee, it is important to learn the pattern of muscles activity. There are several elements of tasks, which involve forearm along with the wrist and hand. The one very important task is pick and place activity performed by the hand. A pick and place action is a compilation of different finger motions for the grasping of objects at different force levels. This action may be better understood by learning the electromyography signals of forearm muscles. Electromyography is the technique to acquire electrical muscle activity that is used for the pattern recognition technique of assistive devices. Regarding this, the different classification characterizations of EMG signals involved in the pick and place action, subjected to variable grip span and weights were considered in this study. A low-level force measuring gripper, capable to bear the changes in weights and object spans was designed and developed to simulate the task. The grip span varied from 6 cm to 9 cm and the maximum weight used in this study was 750 gms. The pattern recognition classification methodology was performed for the differentiation of phases of the pick and place activity, grip force, and the angular deviation of metacarpal phalangeal (MCP) joint. The classifiers used in this study were decision tree (DT), support vector machines (SVM) and k-nearest neighbor (k-NN) based on the feature sets of the EMG signals. After analyses, it was found that k-NN performed best to classify different phases of the activity and relative deviation of MCP joint with an average classification accuracy of 82% and 91% respectively. However; the SVM performed best in classification of force with a particular feature set. The findings of the study would be helpful in designing the assistive devices for hand amputee.

Evaluating User and Machine Learning in Short- and Long-Term Pattern Recognition-Based Myoelectric Control

Proper training is essential to achieve reliable pattern recognition (PR) based myoelectric control. The amount of training is commonly determined by experience. The purpose of this study is to provide an offline validation method that makes the offline performance transferable to online control and find the proper amount of training that achieves good online performance. In the offline experiment, eight able-bodied subjects and three amputees participated in a ten-day training. Repeatability index (RI) and classification error (CE) were used to evaluate user learning and machine learning, respectively. The performance of cross-validation (CV) and time serial related validation (TSV) was compared. Learning curves were established with different training trials by TSV. In the online experiment, sixteen able-bodied subjects were randomly divided into two groups with one- or five-trial training, respectively, followed by participating in the test with and without classifier-output feedback. The correlation between offline and online tests was analyzed. Results indicated that five-trial training was proper to train the user and the classifier. The long-term retention of skills could not shorten the learning process. The correlation between CEs of TSV and the online test was strong ( r=0.87 ) with five-trial training, while the correlation between CEs of CV and the online test was weak ( r=0.30 ). Outcomes demonstrate that offline performance evaluated by TSV is transferable to online performance and the learning process can guide the user to achieve good online myoelectric control with minimum training.

Activities of daily living with bionic arm improved by combination training and latching filter in prosthesis control comparison

BACKGROUND

Advanced prostheses can restore function and improve quality of life for individuals with amputations. Unfortunately, most commercial control strategies do not fully utilize the rich control information from residual nerves and musculature. Continuous decoders can provide more intuitive prosthesis control using multi-channel neural or electromyographic recordings. Three components influence continuous decoder performance: the data used to train the algorithm, the algorithm, and smoothing filters on the algorithm's output. Individual groups often focus on a single decoder, so very few studies compare different decoders using otherwise similar experimental conditions.

METHODS

We completed a two-phase, head-to-head comparison of 12 continuous decoders using activities of daily living. In phase one, we compared two training types and a smoothing filter with three algorithms (modified Kalman filter, multi-layer perceptron, and convolutional neural network) in a clothespin relocation task. We compared training types that included only individual digit and wrist movements vs. combination movements (e.g., simultaneous grasp and wrist flexion). We also compared raw vs. nonlinearly smoothed algorithm outputs. In phase two, we compared the three algorithms in fragile egg, zipping, pouring, and folding tasks using the combination training and smoothing found beneficial in phase one. In both phases, we collected objective, performance-based (e.g., success rate), and subjective, user-focused (e.g., preference) measures.

RESULTS

Phase one showed that combination training improved prosthesis control accuracy and speed, and that the nonlinear smoothing improved accuracy but generally reduced speed. Phase one importantly showed simultaneous movements were used in the task, and that the modified Kalman filter and multi-layer perceptron predicted more simultaneous movements than the convolutional neural network. In phase two, user-focused metrics favored the convolutional neural network and modified Kalman filter, whereas performance-based metrics were generally similar among all algorithms.

CONCLUSIONS

These results confirm that state-of-the-art algorithms, whether linear or nonlinear in nature, functionally benefit from training on more complex data and from output smoothing. These studies will be used to select a decoder for a long-term take-home trial with implanted neuromyoelectric devices. Overall, clinical considerations may favor the mKF as it is similar in performance, faster to train, and computationally less expensive than neural networks.

Decoding of Ankle Joint Movements in Stroke Patients Using Surface Electromyography

Stroke is a cerebrovascular disease (CVD), which results in hemiplegia, paralysis, or death. Conventionally, a stroke patient requires prolonged sessions with physical therapists for the recovery of motor function. Various home-based rehabilitative devices are also available for upper limbs and require minimal or no assistance from a physiotherapist. However, there is no clinically proven device available for functional recovery of a lower limb. In this study, we explored the potential use of surface electromyography (sEMG) as a controlling mechanism for the development of a home-based lower limb rehabilitative device for stroke patients. In this experiment, three channels of sEMG were used to record data from 11 stroke patients while performing ankle joint movements. The movements were then decoded from the sEMG data and their correlation with the level of motor impairment was investigated. The impairment level was quantified using the Fugl-Meyer Assessment (FMA) scale. During the analysis, Hudgins time-domain features were extracted and classified using linear discriminant analysis (LDA) and artificial neural network (ANN). On average, 63.86% ± 4.3% and 67.1% ± 7.9% of the movements were accurately classified in an offline analysis by LDA and ANN, respectively. We found that in both classifiers, some motions outperformed others (p < 0.001 for LDA and p = 0.014 for ANN). The Spearman correlation (ρ) was calculated between the FMA scores and classification accuracies. The results indicate that there is a moderately positive correlation (ρ = 0.75 for LDA and ρ = 0.55 for ANN) between the two of them. The findings of this study suggest that a home-based EMG system can be developed to provide customized therapy for the improvement of functional lower limb motion in stroke patients.

Evaluation of windowing techniques for intramuscular EMG-based diagnostic, rehabilitative and assistive devices

Objective.Intramuscular electromyography (iEMG) signals, invasively recorded, directly from the muscles are used to diagnose various neuromuscular disorders/diseases and to control rehabilitative and assistive robotic devices. iEMG signals are potentially being used in neurology, kinesiology, rehabilitation and ergonomics, to detect/diagnose various diseases/disorders. Electromyography-based classification and analysis systems are being designed and tested for the classification of various neuromuscular disorders and to control rehabilitative and assistive robotic devices. Many studies have explored parameters such as the pre-processing, feature extraction and selection of classifiers that can affect the performance and efficacy of iEMG-based classification systems. The pre-processing stage includes the removal of any unwanted noise from the original signal and windowing of the signal.Approach.This study investigated and presented the optimum windowing configurations for robust control and better performance results of an iEMG-based analysis system based on the stationarity rate (SR) and classification accuracy. Both disjoint and overlap, windowing techniques with varying window and overlap sizes have been investigated using a machine learning-based classification algorithm called linear discriminant analysis.Main results.The optimum window size ranges are from 200-300 ms for the disjoint and 225-300 ms for the overlap windowing technique, respectively. The inferred results show that for the overlap windowing technique the optimum range of overlap size is from 10%-30% of the length of window size. The mean classification accuracy (MCA) and mean stationarity rate (MSR) were found to be lower in the disjoint windowing technique compared to overlap windowing at all investigated overlap sizes. Statistical analysis (two-way analysis of variance test) showed that the MSR and MCA of the overlap windowing technique was significantly different at overlap sizes of 10%-30% (p-values < 0.05).Significance.The presented results can be used to achieve the best possible classification results and SR for any iEMG-based real-time diagnosis, detection and control system, which can enhance the performance of the system significantly.

Harnessing tactile waves to measure skin-to-skin interactions

Skin-to-skin touch is an essential form of tactile interaction, yet there is no known method to quantify how we touch our own skin or someone else's skin. Skin-to-skin touch is particularly challenging to measure objectively, since interposing an instrumented sheet, no matter how thin and flexible, between the interacting skins is not an option. To fill this gap, we explored a technique that takes advantage of the propagation of vibrations from the locus of touch to pick up a signal that contains information about skin-to-skin tactile interactions. These "tactile waves" were measured by an accelerometer sensor placed on the touching finger. Applied pressure and speed had a direct influence on measured signal power when the target of touch was the self or another person. The measurements were insensitive to changes in the location of the sensor relative to the target. Our study suggests that this method has potential for probing behaviour during skin-to-skin tactile interactions and could be a valuable technique to study social touch, self-touch, and motor control. The method is non-invasive, easy to commission, inexpensive, and robust.

On the Selection of Neural Network Architecture for Supervised Motor Unit Identification from High-Density Surface EMG

In the last decade, accurate identification of motor unit (MU) firings received a lot of research interest. Different decomposition methods have been developed, each with its advantages and disadvantages. In this study, we evaluated the capability of three different types of neural networks (NNs), namely dense NN, long short-term memory (LSTM) NN and convolutional NN, to identify MU firings from high-density surface electromyograms (HDsEMG). Each type of NN was evaluated on simulated HDsEMG signals with a known MU firing pattern and high variety of MU characteristics. Compared to dense NN, LSTM and convolutional NN yielded significantly higher precision and significantly lower miss rate of MU identification. LSTM NN demonstrated higher sensitivity to noise than convolutional NN.Clinical Relevance-MU identification from HDsEMG signals offers valuable insight into neurophysiology of motor system but requires relatively high level of expert knowledge. This study assesses the capability of self-learning artificial neural networks to cope with this problem.

A Multiday Evaluation of Real-Time Intramuscular EMG Usability with ANN

Recent developments in implantable technology, such as high-density recordings, wireless transmission of signals to a prosthetic hand, may pave the way for intramuscular electromyography (iEMG)-based myoelectric control in the future. This study aimed to investigate the real-time control performance of iEMG over time. A novel protocol was developed to quantify the robustness of the real-time performance parameters. Intramuscular wires were used to record EMG signals, which were kept inside the muscles for five consecutive days. Tests were performed on multiple days using Fitts' law. Throughput, completion rate, path efficiency and overshoot were evaluated as performance metrics using three train/test strategies. Each train/test scheme was categorized on the basis of data quantity and the time difference between training and testing data. An artificial neural network (ANN) classifier was trained and tested on (i) data from the same day (WDT), (ii) data collected from the previous day and tested on present-day (BDT) and (iii) trained on all previous days including the present day and tested on present-day (CDT). It was found that the completion rate (91.6 ± 3.6%) of CDT was significantly better (p < 0.01) than BDT (74.02 ± 5.8%) and WDT (88.16 ± 3.6%). For BDT, on average, the first session of each day was significantly better (p < 0.01) than the second and third sessions for completion rate (77.9 ± 14.0%) and path efficiency (88.9 ± 16.9%). Subjects demonstrated the ability to achieve targets successfully with wire electrodes. Results also suggest that time variations in the iEMG signal can be catered by concatenating the data over several days. This scheme can be helpful in attaining stable and robust performance.

Classification of Electromyographic Hand Gesture Signals Using Modified Fuzzy C-Means Clustering and Two-Step Machine Learning Approach

Understanding and classifying electromyogram (EMG) signals is of significance for dexterous prosthetic hand control, sign languages, grasp recognition, human-machine interaction, etc.. The existing research of EMG-based hand gesture classification faces the challenges of unsatisfied classification accuracy, insufficient generalization ability, lack of training data and weak robustness. To address these problems, this paper combines unsupervised and supervised learning methods to classify an EMG dataset consisting of 10 classes of hand gestures. To lessen the difficulty of classification, clustering methods including subtractive clustering and fuzzy c-means (FCM) clustering algorithms are employed first to obtain the initial partition of the inputs. In particular, modified FCM algorithm is proposed to accustom the conventional FCM to the multi-class classification problem. Based on the grouping information obtained from clustering, a type of two-step supervised learning approach is proposed. Specifically, a top-classifier and three sub-classifiers integrated with windowing method and majority voting are employed to accomplish the two-step classification. The results demonstrate that the proposed method achieves 100% test accuracy and the strongest robustness compared to the conventional machine learning approaches, which shows the potential for industrial and healthcare applications, such as movement intention detection, grasp recognition and dexterous prostheses control.

Performance Evaluation of Convolutional Neural Network for Hand Gesture Recognition Using EMG

Electromyography (EMG) is a measure of electrical activity generated by the contraction of muscles. Non-invasive surface EMG (sEMG)-based pattern recognition methods have shown the potential for upper limb prosthesis control. However, it is still insufficient for natural control. Recent advancements in deep learning have shown tremendous progress in biosignal processing. Multiple architectures have been proposed yielding high accuracies (>95%) for offline analysis, yet the delay caused due to optimization of the system remains a challenge for its real-time application. From this arises a need for optimized deep learning architecture based on fine-tuned hyper-parameters. Although the chance of achieving convergence is random, however, it is important to observe that the performance gain made is significant enough to justify extra computation. In this study, the convolutional neural network (CNN) was implemented to decode hand gestures from the sEMG data recorded from 18 subjects to investigate the effect of hyper-parameters on each hand gesture. Results showed that the learning rate set to either 0.0001 or 0.001 with 80-100 epochs significantly outperformed (p < 0.05) other considerations. In addition, it was observed that regardless of network configuration some motions (close hand, flex hand, extend the hand and fine grip) performed better (83.7% ± 13.5%, 71.2% ± 20.2%, 82.6% ± 13.9% and 74.6% ± 15%, respectively) throughout the course of study. So, a robust and stable myoelectric control can be designed on the basis of the best performing hand motions. With improved recognition and uniform gain in performance, the deep learning-based approach has the potential to be a more robust alternative to traditional machine learning algorithms.

Selection of Features and Classifiers for EMG-EEG-Based Upper Limb Assistive Devices-A Review

Bio-signals are distinctive factors in the design of human-machine interface, essentially useful for prosthesis, orthosis, and exoskeletons. Despite the progress in the analysis of pattern recognition based devices; the acceptance of these devices is still questionable. One reason is the lack of information to identify the possible combinations of features and classifiers. Besides; there is also a need for optimal selection of various sensors for sensations such as touch, force, texture, along with EMGs/EEGs. This article reviews the two bio-signal techniques, named as electromyography and electroencephalography. The details of the features and the classifiers used in the data processing for upper limb assist devices are summarised here. Various features and their sets are surveyed and different classifiers for feature sets are discussed on the basis of the classification rate. The review was carried out on the basis of the last 10-12 years of published research in this area. This article also outlines the influence of modality of EMGs and EEGs with other sensors on classifications. Also, other bio-signals used in upper limb devices and future aspects are considered.

Comparing EMG-Based Human-Machine Interfaces for Estimating Continuous, Coordinated Movements

Electromyography (EMG)-based interfaces are trending toward continuous, simultaneous control with multiple degrees of freedom. Emerging methods range from data-driven approaches to biomechanical model-based methods. However, there has been no direct comparison between these two types of continuous EMG-based interfaces. The aim of this study was to compare a musculoskeletal model (MM) with two data-driven approaches, linear regression (LR) and artificial neural network (ANN), for predicting continuous wrist and hand motions for EMG-based interfaces. Six able-bodied subjects and one transradial amputee subject performed (missing) metacarpophalangeal (MCP) and wrist flexion/extension, simultaneously or independently, while four EMG signals were recorded from forearm muscles. To add variation to the EMG signals, the subjects repeated the MCP and wrist motions at various upper extremity postures. For each subject, the EMG signals collected from the neutral posture were used to build the EMG interfaces; the EMG signals collected from all postures were used to evaluate the interfaces. The performance of the interface was quantified by Pearson's correlation coefficient (r) and the normalized root mean square error (NRMSE) between measured and estimated joint angles. The results demonstrated that the MM predicted movements more accurately, with higher r values and lower NRMSE, than either LR or ANN. Similar results were observed in the transradial amputee. Additionally, the variation in r across postures, an indicator of reliability against posture changes, was significantly lower (better) for the MM than for either LR or ANN. Our findings suggest that incorporating musculoskeletal knowledge into EMG-based human-machine interfaces could improve the estimation of continuous, coordinated motion.

Online Learning and Classification of EMG-Based Gestures on a Parallel Ultra-Low Power Platform Using Hyperdimensional Computing

This paper presents a wearable electromyographic gesture recognition system based on the hyperdimensional computing paradigm, running on a programmable parallel ultra-low-power (PULP) platform. The processing chain includes efficient on-chip training, which leads to a fully embedded implementation with no need to perform any offline training on a personal computer. The proposed solution has been tested on 10 subjects in a typical gesture recognition scenario achieving 85% average accuracy on 11 gestures recognition, which is aligned with the state-of-the-art, with the unique capability of performing online learning. Furthermore, by virtue of the hardware friendly algorithm and of the efficient PULP system-on-chip (Mr. Wolf) used for prototyping and evaluation, the energy budget required to run the learning part with 11 gestures is 10.04 mJ, and 83.2  μJ per classification. The system works with a average power consumption of 10.4 mW in classification, ensuring around 29 h of autonomy with a 100 mAh battery. Finally, the scalability of the system is explored by increasing the number of channels (up to 256 electrodes), demonstrating the suitability of our approach as universal, energy-efficient biopotential wearable recognition framework.

Feature Selection Based on Binary Tree Growth Algorithm for the Classification of Myoelectric Signals

Electromyography (EMG) has been widely used in rehabilitation and myoelectric prosthetic applications. However, a recent increment in the number of EMG features has led to a high dimensional feature vector. This in turn will degrade the classification performance and increase the complexity of the recognition system. In this paper, we have proposed two new feature selection methods based on a tree growth algorithm (TGA) for EMG signals classification. In the first approach, two transfer functions are implemented to convert the continuous TGA into a binary version. For the second approach, the swap, crossover, and mutation operators are introduced in a modified binary tree growth algorithm for enhancing the exploitation and exploration behaviors. In this study, short time Fourier transform (STFT) is employed to transform the EMG signals into time-frequency representation. The features are then extracted from the STFT coefficient and form a feature vector. Afterward, the proposed feature selection methods are applied to evaluate the best feature subset from a large available feature set. The experimental results show the superiority of MBTGA not only in terms of feature reduction, but also the classification performance.