1. Introduction
Myoelectric control is the state-of-the-art control system for a sophisticated upper-limb prosthetic system. This control scheme reflects the user’s intention and uses the muscle’s electrical signals (electromyogram, EMG) to achieve the desired function. Currently, most commercially available upper-limb prostheses use conventional two-site proportional control, on/off control, and direct control, which maps different amplitudes of EMG signals to corresponding mechanical output [
1]. Although traditional myoelectric control schemes can provide reasonable controllability, they have limited functionality that cannot fulfil users’ daily living requirements. According to the prosthetic usage surveys [
2,
3], there has been no significant decrease in the prosthetic abandonment rate since 2007. Users complained most about the lack of functionality and discomfort.
Machine-learning-based myoelectric control systems are the primary developing trend for prosthetics, where pattern recognition (PR) techniques are widely employed [
4] and are now available commercially. PR-based control systems directly decode EMG signals acquired from muscles into different motion patterns and can potentially achieve intuitive control. PR-based control systems perform equivalent to or better than conventional control approaches when used by amputees within their home environment [
5]. EMG signals are stochastic and are sensitive to extrinsic factors (e.g., electrode shift [
6], skin impendence change [
7]) and intrinsic variations (muscle fatigue [
8], mutual adaptation [
9]). Obtaining identical EMG signals consistently is impossible, even with a fixed setup and movements. These factors are evident in home trials and are mitigated using periodic system recalibration [
5]. The performance of PR-based control systems heavily depends on feature extraction and classifications. The change in the EMG signal feature space subsequentially affects the performance. Hence, the influencing factors impose repeated calibration sessions in PR-based control systems.
In the upper-limb prosthesis control, a change in the limb position means using the prosthesis in different arm positions. Amputees need to adjust their limb positions to perform specific daily living activities, such as picking items off the ground or reaching overhead. Different limb positions require corresponding muscle coordination, which generates different EMG signal patterns. In addition, the underlying topography of the muscle fibers and skin may shift relative to the electrode in different limb positions [
10]. Fougner et al. [
11] demonstrated that limb position variation substantially impacted the myoelectric signal classification performance, where the average classification error increased from 3.8% to 18%. They also showed that the errors in intra-position classification (training and testing in same limb position) were relatively lower than in the inter-position classification (training and testing in different limb positions), which were 3.8% and 21.1%, respectively. Jochumsen et al. [
12] compared the effect of limb positions on the surface and intramuscular EMG and showed that both types of EMG were similarly affected by limb positions. The classification accuracies decreased by 12–16% in between-position performance. Due to the requirement of frequently changing limb position for amputees to complete activities of daily living, the mitigation of the effect of limb position on real-time performance is necessary. Teh and Hargrove [
13] demonstrated that limb position significantly affected amputees’ and intact-limb subjects’ real-time virtual prosthesis control performance.
Electrode shift usually happens when the user redons the prosthesis, leading to significant variation in the EMG signal amplitude and can be easily confused with a change in muscle activation patterns. The main reason for the variation is related to the innervation zone of each motor unit [
10]. According to Rainoldi et al. [
14], little disturbances will significantly vary EMG amplitude when electrodes are close to an innervation zone. Young et al. [
15] investigated how the size and orientation of electrodes affect the system’s robustness to electrode shift. Based on displacement up to 2 cm, they found that the direction of the shifts parallel to muscle fibers (5–20% error) had a more negligible effect on classification accuracy than the perpendicular shifts (40% error). High-density (HD) EMG electrodes were proposed to mitigate the effect of shifts [
16,
17]. Nevertheless, the level of performance degradation induced by electrode shifts in HD EMG is sensitive to the density of HD EMG electrodes [
18].
For each influencing factor, various approaches have been developed to mitigate performance degradation. Classifiers can potentially learn the underlying characteristics of affected data for each motion, which gives inspiration for extensive training [
19,
20,
21]. Gigli et al. [
21] explored a dynamic training protocol to use the EMG signals acquired from a continuous limb movement that contains all motion patterns of interest. The results showed the advantages of the dynamic training protocol, such as satisfying controllability and a less tiresome data collection procedure. The dynamic training protocol also demonstrated a similar level of controllability improvement in amputees [
13]. In addition, some features and adaptive machine-learning techniques are less sensitive to the fluctuations of EMG signals [
22,
23,
24,
25,
26]. Stango et al. [
25] extracted spatial features from HD EMG signals to classify nine and seven classes for intact-limb and amputated subjects, respectively. The results indicated that spatial features were less sensitive to electrode shifts (±1 cm) than classic features. Ameri et al. [
26] manifested a novel self-recalibrated system based on transfer learning with convolutional neural networks (TL-CNN) to reduce the effect of electrode shifts. They compared the adaptive ability to electrode shifts between the proposed TL-CNN with other state-of-the-art techniques. The results showed that TL-CNN was more effective in reducing errors.
Current studies can achieve promising results in classifying motion patterns under the effect of a single influencing factor. However, several factors could appear simultaneously in the practical use of prostheses. For instance, electrode shifts could happen due to the weight of the prosthesis and gravity while the limb position changes. Asogbon et al. [
27] investigated the co-existing impact of mobility of the subject and variation of contraction force on the PR-based prostheses and reduced the factor-induced error by 7.50~17.97% through robust feature extraction. Gu et al. [
28] explored robust features and classifiers under electrode shifts, force variation, limb position change and temporary drift. They utilized an adaptive learning method to mitigate performance degradation. In contrast to the laboratory experiment to simulate the potential factors that could occur during the use of the prostheses in daily life, researchers have attempted to monitor the control of at-home prostheses to improve their functionality and performance [
5,
29] and tried to narrow the gap between academic research and clinical application. They have only investigated how users interact with their prostheses, analyzing the quality of recalibration data and the potential reasons for low-quality EMG signals. Meanwhile, without the support of additional devices or data, the occurrence of influencing factors such as electrode shift sand changes in limb position while using a prosthesis remains unknown. These factors could be the primary causes of recalibration. Hence, it is critical to recognize different influencing factors in EMG signals when performance degradation occurs during daily use to help researchers understand how frequently each factor occurs so that they can adopt the corresponding solution further.
While the above studies [
5,
14,
15,
16,
17,
18,
19,
20,
21,
22,
23,
24,
25,
26,
27,
28,
29] have described performance degradation or proposed mitigating solutions, our study focuses on understanding the causes of classifier failure to inform future adaptive approaches. Also, the extent to which influencing factors affect the feature space and class distributions has received less attention in the literature. Hence, this study aims to quantify the effect of limb positions and electrode shifts on classification. We acquired EMG data from four frequently used limb positions in daily life and from four different electrode shifts that could occur during daily use. To further validate our results in practical settings, we also utilized a dataset from amputees involving four different limb positions. We analyzed how feature space was changed under different factors and demonstrated for the first time that these factors generate additional class instances that make the problem challenging to address from purely a machine-learning perspective.
4. Discussion
This study aimed to quantify how limb positions and electrode shifts affect class distributions. In motion classifications, we used LDA as it is widely used as the baseline in the research field of PR-prosthetic control and is commercialized in the prosthetic control system of COAPT [
37] and OttoBock [
38]. The degradation of results of inter-limb-position classifications was around 15–21%, similar to previous research [
10,
12,
39]. There was no further decrease in accuracy in comparing five and sixteen limb position effects, but forearm orientation had a more significant impact than other limb positions [
10]. In addition, previous studies [
15,
40] have stated that electrodes shifted along the direction of the muscle fibers have a lesser decrease in accuracy than transversal shift and accuracy decreased more with increased distance. Our results demonstrated that electrode shifts affected the classification accuracy, but there was no significant difference between different levels of electrode shifts. The average accuracy of transversal shift (S2 and S4) was slightly lower than longitudinal (S1 and S3) when the shift distances were the same. The reason could be that the displacements induced in our study were too small compared to other studies to reflect the difference, as EMG signals vary more when the electrode is near the innervation zone [
14]. Nevertheless, this shift was enough to induce new class instances.
Despite investigating the effect of combined factors [
27,
28,
30] on the overall classification, classifying different influencing factors has not been attempted to the best of our knowledge. LDA can also be selected to classify factors. Its simplicity and fast processing time allow for the potential use of complex algorithms that depend on the results of factor identification. This enables adaptive classification to run within the limited time frame required for real-time control. Our results indicate that influence factors (limb positions vs. electrode shifts) are classifiable with high accuracy. The classification performance of the influencing factors based on every single motion (Type 1–2) with two selected classifiers and feature sets supports our hypothesis. Nevertheless, the motion did not affect the degree of classification. Upper extremity muscles exhibit similar synergistic activation patterns during the same hand movements [
41].
Additionally, changes in limb position can cause alterations in muscle synergy recruitment due to factors such as gravity and variations in proprioception [
42]. The variation in EMG signals induced by electrode shifts depends on the direction and magnitude of the shift. Consequently, when electrode shifts or changes in limb position occur during EMG recordings, new class instances are generated based on the different types and levels of muscle patterns.
In the feature space, the RI was low when it was measured from the same motion within the same factor. After inducing different levels of limb positions and electrode shifts, the consistency of the same motion became lower, indicating the potential existence of factor clusters. Changes in mSI support our classification results of poor performance (due to poor separability) when many factors are combined.
To demonstrate the existence of new class instances, we also classified motion trials (repetitions), and the results revealed accuracies below chance. This is further shown in
Figure 5 and
Figure 6, where factors are well separated graphically, but repetitions are not. From the perspective of feature distribution, each level of influencing factor was well separated for classification Type 1–2, where each level of the same influence factor was closely distributed. When all motions were combined (
Figure 6a,b), the feature distributions of each level of influencing factor became closer. Because of the divergence of feature distribution of different motions, influencing factors were distributed in several areas in the feature space. They caused overlaps, which caused classification performance degradation subsequentially, making classification tasks challenging.
Furthermore, in the analysis of amputees’ data, we found that the limb position did have the same extent of degradation on the performance of BFC as the able-bodied subjects. The influence factor classification performance showed no significant difference between amputees and able-bodied subjects, with promising accuracy (>98%). It demonstrated that the limb position caused by a change in feature distribution would also form corresponding class instances in amputees. The number of amputees was limited; therefore, more validation should be conducted on amputee-related datasets.
On one side, the results of this research provide a fundamental basis for designing better adaptive mechanisms that should account for the significant statistical changes in the feature space. These mechanisms should be able to separate classes when they translate into space. Dual-stage classification schemes were investigated to mitigate the effect of limb position by classifying limb position before the hand motion classification. Previous studies [
43,
44] stated that classifying accelerometer data can identify limb positions with promising results. However, introducing additional data could increase the potential computation cost. On the other hand, this research raises the question of the importance of recalibration when using actual prostheses. In the studies conducted by Hargrove et al. [
45] and Simon et al. [
5], participants recalibrated their PR-based myoelectric prostheses 32.6 ± 8.2 times and a median value 18 times with an interquartile range of 11.75 to 36 over 8 weeks home trial, respectively. Notably, the reason for recalibration was not recorded in those prior studies. It is reasonable to assume that some recalibrations were habitual, whereas others were necessitated by poor control. EMG signal quality helps in noise estimation and onset detection. If factors that cause performance degradation can be identified, it can give researchers more profound insight into the reasons for prosthetic recalibration.
The high accuracy of influencing factor classification was achieved in the current study, showing its impact. However, the study is at an early stage, and its clinical translation for determining the reason(s) behind changes in EMG signals necessitates further investigation. Factor identification would require collecting data on different factors for training the classifier, which is time-consuming. Additionally, some within/between day factors [
10] and co-existing factors may contribute to changes in the signal characteristics due to the time gap between the influencing factor data acquisition and the analysis. These issues will be addressed in future research. Moreover, as this research is limited to two factors and conducted offline, we recommend future studies where class distributions are affected in real-time with visual feedback and where the limb is in a constant state of motion. This will allow quantification of class distribution both at the initial condition when the factor occurs and while the user attempts to correct the mistake.