Centimeter Differences in Wrist Electrode Placement Significantly Impact Myoelectric Performance

Abstract

The long-term goal of this research is to establish electromyography (EMG) as an intuitive and dexterous control interface for human-computer interaction. EMG is an established technique for classifying hand gestures and motions, used often in prosthetics and orthotics. Recently, there has been a shift towards recording EMG at the wrist, instead of at the forearm, to yield a more socially acceptable form factor for consumer applications. EMG within the size of a watch or bracelet means fewer electrodes and more variable placement with respect to the underlying muscle anatomy. Here, we explore how differences in location along the wrist impact EMG quality and myoelectric control. We recorded EMG and compared myoelectric performance across three different regions of electrodes (distal, central, and proximal) using electrode arrays at both the wrist and the forearm. We found that a small 4.3 cm shift proximally on the wrist yields significant improvements in EMG information content and myoelectric performance. When trained on a k-Nearest Neighbors model, classification accuracy increased from 79.3% at the distal wrist region to 83.7% at the proximal wrist position. EMG from the proximal wrist region also had significantly more information content, as indicated by greater variance outside of the first principal component and by more frequently selected channels via a minimum-redundancy-maximum-relevance selection approach. These findings indicate that the spatial position of electrodes at the wrist has a noticeable impact on myoelectric control in a way not seen in traditional EMG recordings from the forearm. This can inform the design of future wrist-worn EMG devices, which in turn may lead to more robust control for partial hand prostheses, hand orthoses, and augmented/virtual reality.

Clinical Relevance—

A subtle change in the position of electromyographic electrodes on the wrist can yield significant improvements in the control of technology, like prostheses, exoskeletons, and virtual/augmented reality.

I. Introduction

As technology becomes more present, how we interact with it has evolved. The traditional mouse and keyboard work well to control classic computers, but a different interface is needed when computers are embedded within our thermostats, refrigerators, and sound systems [1]. Touch screens have become ubiquitous for devices such as smartphones and tablets [2]. Still, with augmented and virtual reality becoming increasingly popular, there is a need for a human-machine interface (HMI) that can intuitively interact with an intangible environment [3]. Voice control is effective but works best in environments with precise, command-based interaction (such as choosing a specific song to be played on the speakers) [4]. For continuous control of spatial environments, gesture control provides an interaction analogous to a computer mouse or touchscreen [5], [6].

There have been several approaches to refine gesture control for everyday use. The Apple Watch (Apple Inc., Cupertino, CA) uses accelerometers, gyroscopes, and photoplethysmography (optical heart rate sensor) to distinguish simple gestures such as grasp and pinch [7]. The new Mudra Band (Wearable Devices Ltd., Yokneam Illit, Israel) is an Apple Watch wristband that uses accelerometers and “Surface Nerve Conductance” (SNC) sensors to provide fine control of cursors [6]. The now discontinued Myo Armband (Thalmic Labs, New York City, NY) was an attempt to bring the idea of intuitive gesture control into mainstream use through surface electromyographic signals (sEMG) around the forearm [8].

EMG has historically been used for controlling prostheses [9]. Traditionally, EMG is recorded at the muscle belly, where the signals are the strongest [10]. For the muscles that control the hand, the muscle belly is located in the proximal portion of the forearm, and as such, this has been the predominant recording location for upper-limb myoelectric prostheses [11]. In contrast to this conventional approach, there has been a recent surge of interest in the idea of EMG at the wrist [12], [13], [14], since it is already socially accepted as a location for wearable smart technology [15]. Indeed, Meta’s EMG Wristband represents one of the first large-scale applications of EMG recorded from the wrist (Meta, Menlo Park, CA) [16].

Several studies have demonstrated the effectiveness of wrist-based EMG [12], [13], [14], [17], and others have analyzed the importance of electrode positioning in general [18], [19], but there has not yet been an analysis of how the different positioning of electrodes at the wrist contributes to the overall performance. Precise electrode placement across the forearm to target specific muscles has significant effects on myoelectric controllability [20], and high-density (HD) arrays make it simpler to sample from the ideal locations without precise placement [21]. However, for wrist EMG to be effective in the long term for consumer use, the electrodes should ideally fall within the average width of a watch (around 20 mm) while still maintaining robust performance.

In this paper, we investigate how subtle changes in the placement of a ring of electrodes contribute to overall myoelectric performance. Our findings show that a ring of electrodes around the wrist can achieve significantly better performance by moving proximally by as little as 43 mm (1.7 in). These findings can inform the design of future wrist EMG devices to maximize performance within the small, socially acceptable form factor of a wrist-worn device.

II. Methods

A. Participants

Twelve participants were recruited for this study. The participants’ ages ranged from 18 to 29 years old, with a mean age of 22. All participants were neurologically healthy, and 75% were male. Across all participants, the average wrist and forearm circumferences were 17.4 cm (6.84 in) and 27.6 cm (10.89 in), respectively. Participants gave consent prior to the experiment under the University of Utah Institutional Review Board number 00098851.

B. Wrist and Forearm EMG Arrays

EMG data were recorded using two custom-made surface EMG electrode sleeves [22]. The two arrays were positioned on the arm around the wrist and the upper forearm, as shown in (Fig. 1a). The arrays are comprised of nickel-plated brass electrodes soldered to a Samtec connector (Samtec, New Albany, IN). The electrodes have an average inter-electrode distance of 25 mm (0.98 inches) and a diameter of 15 mm (0.6 inches). Electrodes were affixed to neoprene and were secured to the wrist or forearm using Velcro straps. The ground and reference electrodes were aligned with the ulna near the styloid process to minimize electrical activity [23].

Fig 1.

Experiment Setup. (a) Diagram showing the setup of the wrist and forearm sEMG arrays, with the different regions of interest labeled. (b) Layout and dimensions of the wrist array. (c) Layout of the forearm array (dimensions match the wrist array). G and R denote the ground and reference electrodes, respectively.

The wrist array was made up of 18 electrodes divided into three rings (distal, central, and proximal) that span the wrist’s circumference. The distal ring was closest to the hand (distal to the elbow), and the proximal ring was closest to the elbow. The central ring was positioned between the distal and the proximal rings and was staggered laterally by 12.5 mm (0.49 in) (Fig. 1b). The electrodes closest to the ground and reference electrodes on the central ring were excluded from this study, so each ring on the wrist examined a total of five electrodes. The array is placed on the wrist such that it sits against the wrist flexion crease.

The forearm array was comprised of 34 electrodes divided into three rings (Fig. 1c). Like the wrist array, the forearm array was divided into three rings (distal, central, and proximal), with the distal ring closer to the wrist (distal to the elbow) and the proximal ring closer to the elbow. The central ring was between the distal and outer rings, and the three rings were the same distance apart from one another as the wrist array. The array was positioned around the muscle belly of the forearm, where the circumference was largest. The ground and reference electrodes were located on the distal and proximal rings and placed along the ulna. Because the forearm array had more electrodes than the wrist array, the best five electrodes from each of the three rings were selected using a channel selection algorithm described later. This resulted in an equal comparison of 15 electrodes at the wrist against 15 electrodes at the forearm.

C. Experimental Design

Each participant performed 17 different hand or wrist gestures 10 times while EMG was recorded simultaneously from the forearm and wrist. Participants were instructed to perform gestures in synchrony with those displayed by a virtual hand (Modular Prosthetic Limb, MSMS; Johns Hopkins Applied Physics Lab, Baltimore, MD) on a computer monitor. The virtual hand repeated the same gesture 10 times in a row before advancing to the next type of gesture. Each gesture consisted of a 0.7-second transition into the gesture, a 3-second hold of the gesture, and then a 0.7-second transition back to a neutral resting position. A 2-second rest period occurred between each gesture. The total data collection process took approximately 18 minutes.

The gestures recorded were comprised of single and multi-finger gestures, as well as wrist gestures. Finger gestures are denoted by D1 (thumb) through D5 (little finger). The gestures used were: D1 flexion (D1F), D2 flexion (D2F), D3 flexion (D3F), D4 flexion (D4F), D5 flexion (D5F), D1 extension (D1E), D2 extension (D2E), D1-D5 extension (open hand; D1–5E), D1-D2 flexion (pinch, D1–2F), D1-D3 flexion (tripod punch; D1–3F), D1-D5 flexion (grasp, D1–5F), wrist flexion (WF), wrist extension (WE), wrist radial deviation (WRD), wrist ulnar deviation (WUD), wrist pronation (WP), and wrist supination (WS).

D. EMG Signal Acquisition and Processing

EMG was sampled at 1 kHz and filtered using the Summit Neural Interface Processor (Ripple Neuro Med LLC, Salt Lake City, UT). EMG was filtered through a bandpass filter with cutoff frequencies of 15 and 375 Hz and a notch filter at 60, 120, and 180 Hz. The mean absolute value (MAV) was then calculated at 30 Hz. The MAV was smoothed using an overlapping 300-ms window. The resulting EMG feature set consisted of the 300-ms smoothed MAV on 16 wrist EMG channels and the 32 forearm EMG channels, calculated at 30 Hz [24].

E. Signal-to-Noise Ratio Calculation

Each of the participants performed 10 repetitions of 17 different gestures, for a total of 170 movements per participant. The signal-to-noise ratio (SNR) was calculated using the formula

$$\mathit{\text{SNR}}=20\times \mathit{\text{log}}\frac{{\mathit{\text{RMS}}}_{\mathit{\text{activated}}}}{{\mathit{\text{RMS}}}_{\mathit{\text{unactivated}}}}$$

where the RMS_activated was calculated when participants were actively performing gestures, and RMS_unactivated was calculated over a period of rest that occurs before the gestures are recorded to measure baseline activation [12], [22]. The SNR was calculated for each gesture across each individual electrode, then averaged across locations (distal, central, and proximal rings), and each repeated gesture was averaged together for one value per gesture. The resulting data set was 17 SNR calculations (one per gesture) aggregated across the 12 participants for each of the six recording locations.

F. Information Content

Information content, or the variance and uniqueness of data between EMG input channels, across different EMG regions was calculated using principal component analysis (PCA) to calculate eigenvectors (or principal components), which define directions for projecting the multidimensional dataset to preserve maximum variance. The first principal component captures the most variance when all data points are projected onto it. Each subsequent principal component captures progressively less variance within the dataset [25]. If PCA were calculated on identical EMG channels, it would capture all variance within the first principal component because the combined dataset would be perfectly linear. Therefore, the percentage of variance captured outside the first principal component over a set of EMG channels serves as a straightforward measure of the EMG data’s variability across those channels. Using PCA to evaluate the variance of EMG datasets has been previously established [25], [26].

G. Electrode Importance Mapping

Across the wrist and forearm arrays, the five best were selected using the Minimum Redundancy Maximum Relevance (mRMR) algorithm [27]. The mRMR algorithm selects the best electrode according to how well it corresponds with the training labels (maximum relevance) and continues to select subsequent channels, prioritizing ones that maximize relevance and minimize redundant data represented in previously selected channels. After selection, the five best electrodes were summed across all the participants for both arrays, and the overall results are reflected in a heatmap to show which areas contribute best overall across participants.

H. Classification Accuracy

EMG features were classified using a k-Nearest Neighbor (kNN) model, with k-value of 10 and 5-fold cross-validation. The kNN is a basic model that measures the Euclidean distance between a test point and its nearest k neighbors. The test point is classified into the same class as the majority of its neighbors [28]. The kNN has been used extensively to analyze EMG performance because of its speed and simplicity [29], [30]. The kNN was used in this study because it is deterministic, widely used, and effective with small amounts of data.

I. Statistical Analysis

Data were screened for normality using the Anderson-Darling test. The SNR and information content were each analyzed separately using a Kruskal-Wallis test. The classification accuracy was analyzed with a single-factor repeated measure ANOVA. Post-hoc pairwise comparisons were performed between the different electrode regions at the wrist and forearm (but not across both), using the Tukey-Kramer correction for multiple comparisons [31].

III. Results

A. Forearm SNR Was Generally Higher than Wrist SNR

We calculated the SNR for each of the 17 gestures, aggregated across all 12 participants for six different electrode regions: distal, central, proximal locations on the wrist and forearm (Fig. 2). The forearm SNR was generally higher than the wrist SNR, although this was not statistically significant. Among the three wrist regions, the SNR of the central electrodes (7.4 dB) was significantly higher than that of the distal electrodes (5.6 dB, p < 0.05, Kruskal-Wallis). The SNR of the proximal electrodes (5.9 dB) was not significantly different than either the distal or central electrodes. Among the three forearm regions, the distal region had significantly higher SNR than the proximal region (11.1 and 8.7 dB, respectively, p < 0.05, Kruskal-Wallis). The forearm central region (9.6 dB) was not significantly different than the other forearm regions.

Fig 2.

Signal-to-Noise Ratio (SNR) of different gestures.

At the wrist, the central electrodes displayed a significantly higher SNR than the distal electrodes. The forearm distal region was significantly better than the forearm proximal region. Data shows the average SNR across all participants for each gesture (N=17) in each of the six regions. Boxplots show median, IQR, and most extreme non-outlier values. Outliers are shown as circles. Asterisk (*) denotes statistical significance (p < 0.05, Kruskal-Wallis with Tukey-Kramer correction for multiple comparisons).

B. Proximal Wrist Electrodes Contain More Information than Distal Wrist Electrodes

We calculated the information content for each region by summing up the variance outside the first principal component. We found that, at the wrist but the forearm, the distance between the regions had a significant impact on information content. The proximal electrodes contained significantly more variance outside the first principal component (21.5%, IQR 4.9) than the distal electrodes (8.5% IQR 8.9, p < 0.05, Kruskal-Wallis). The central electrodes were not significantly different from either (14.4% IQR 19.8). In contrast, at the forearm, the best five electrodes from each region (distal, central, proximal) were not statistically different among regions (18.6% IQR 9.5, 11.3% IQR 18.3, 12.4% IQR 20.0 respectively; Fig. 3a).

Fig 3.

Information content. (a) Variance outside the first principal component for different regions. (b) Importance map of electrodes for the wrist array, showing the number of times that each electrode was selected as one of the five best electrodes using the mRMR algorithm. (c) Importance map of forearm electrodes. Boxplots show median, IQR, and most extreme non-outlier values. Asterisk (*) denotes statistical significance (p < 0.05, Kruskal-Wallis with Tukey-Kramer correction for multiple comparisons).

C. Electrode Importance Mapping Indicates Location Matters More at the Wrist

The best five electrodes were selected for each participant using the mRMR algorithm. These selections were summed across all participants to identify particular regions of importance in the wrist (Fig. 3b) and forearm (Fig. 3c). Of the 60 electrodes selected (5 best electrodes for each of the 12 participants) at the wrist, 15 were distal, 20 were central, and 25 were proximal. Of the 60 electrodes chosen at the forearm, 11 were distal, 26 were central, and 23 were proximal.

D. Proximal Wrist Electrodes Performed Best at Classifying Gestures

Each electrode group was used to train a kNN classification model. At the wrist, the electrode region showed a significant impact on classification performance, but no difference was seen across electrode regions at the forearm (Fig. 4). The kNN model trained with proximal wrist electrodes performed significantly better (83.7%) at classifying gestures than the model trained with distal wrist electrodes (79.3%, p < 0.05, ANOVA), despite only being 4.3 cm more proximal. The model trained on the central wrist electrodes (82.2%) showed no significant difference from either the distal wrist or proximal wrist electrodes. The kNN models trained on the forearm EMG showed no significant difference between the distal, central, and proximal regions (79.7%, 80.3%, and 79.7%, respectively), but all the forearm regions performed significantly worse than the proximal wrist region (p < 0.05, ANOVA). No noticeable differences were seen in the accuracy of the individual gestures; rather, differences in accuracy were reflected across all gestures more broadly.

Fig 4.

Classification accuracy for each electrode region. The proximal wrist region significantly outperforms the distal wrist region, whereas the proximal forearm region shows no significant improvement over the distal forearm region. Violin plots show data distribution with median and mean values labeled. Asterisk (*) denotes statistical significance (p < 0.05, ANOVA with Tukey-Kramer correction for multiple comparisons).

IV. Discussion

In this study, we compared six different electrode regions, three around the wrist and three around the forearm, to evaluate how the positioning of these electrodes affects overall myoelectric performance. We found that, although the forearm contains higher SNR than the wrist, this does not affect overall performance. The proximal wrist electrodes significantly outperformed the distal wrist electrodes, while the forearm showed no such improvement over the same distance between the distal and proximal forearm electrodes. This indicates that the precise positioning of a ring of electrodes is more critical at the wrist than the forearm and that a small number of electrodes (five, in this case) can provide sufficient control in this region.

The higher SNR in the forearm follows conventional wisdom that the muscle belly would contain higher amplitude signals than the wrist. Interestingly, these results are different from reports found in other studies. Botros et al. reported that the wrist, in fact, had better SNR across select gestures but showed no significant difference overall [12]. Our prior work comparing wrist and forearm EMG in stroke patients showed no significant difference between either location for the non-paretic side of the body [22]. Variations in the types of gestures performed could lead to the differences in SNR reported across studies. Variations in electrode density could also lead to variable SNR across studies. For example, our previous work used a different forearm array that contained 32 electrodes distributed across the entire forearm. In contrast, our current study used a new forearm array that matched the electrode density of the wrist array. This new array was constructed to eliminate certain confounding factors and lead to a more direct comparison. Nevertheless, whether or not wrist EMG has worse or better SNR than forearm EMG is somewhat irrelevant, as studies agree that the wrist is a viable recording location for sEMG recording [12], [13], [14], [17], and there is no strong correlation between SNR and myoelectric performance [22]. Since the SNR is based purely on the amplitude of the signal, it makes sense that it isn’t a strong indicator of predictive power.

Rather than focusing on the signal quality measured by the SNR as an indicator of performance, we investigated the richness of the information content in each electrode region. Previous work has used PCA-based methods for investigating information content, as the reported variance across principal components can indicate data uniqueness and richness across channels [26]. Our approach was simple in scope, but the PCA-calculated information content tells a similar story to the overall classification performance. That is, the proximal wrist contains more information and provides better classification accuracy.

The electrode importance heatmaps found the best five electrodes from each array (regardless of distal, central, or proximal region) and summed them across participants. These results further support the preference for the proximal wrist position. At the wrist, the proximal electrodes were most often selected as the top five electrodes per participant. In contrast, at the forearm, no region stands out as prominently as it does for the wrist. Still, certain regions are more obvious contributors to overall performance, indicating ideal muscle groups that extend beyond a single row. The clustering of commonly chosen electrodes resembles heat maps shown previously [32].

The improved performance of the kNN classifier at the wrist is consistent with our previous work [22]. Indeed, prior work shows that although the wrist is a viable recording location, it excels over the forearm in controlling fine-finger movements [12], [14]. Our results here highlight the sensitivity of the spatial positioning of electrodes at the wrist compared to the forearm, which appears to have less dependence on longitudinal change along the length of the limb. This is consistent with the heat maps shown in Chamberland et al., and is likely due to the shrinking musculature as the forearm muscles approach the wrist [32]. Admittedly, these analyses are done over relatively small longitudinal distances (a few inches at most), and a longer displacement across the forearm may result in more noticeable changes. Still, it is interesting that our results show that a small difference (as little as 4.3 cm) at the wrist can significantly impact performance. This fact indicates that further analysis of the wrist is needed, and that assumptions about EMG performance at the wrist cannot necessarily be based on past findings from the forearm.

In this work, we used a simple kNN model, but other models, such as convolutional neural networks and recurrent neural networks, exist for EMG control [33], [34]. More complex neural network models trained on large amounts of data may be able to provide robust control that generalizes across variable positions [16]. Nevertheless, optimal placement may enable more efficient models and simplify calibration, both of which are critical for consumer applications of wearable EMG.

Surprisingly, the five electrodes in the distal wrist region are more effective than the five best electrodes selected from each of the forearm regions, even when there are so many more electrodes on the forearm to choose from. We suspect that one of the reasons for this is that the five electrodes at the wrist can practically encircle the limb, while the five electrodes at the forearm (recording higher SNRs) sample fewer muscle groups. The participants’ wrist circumferences ranged from 15.24 cm - 19.69 cm (6 in - 7.75 in) for this study, which covers the average wrist size for adult men and women [35]. However, these results may not be consistent for individuals whose wrist circumference falls outside the normal healthy distribution.

Another possible reason why the wrist achieves greater classification accuracy than the forearm may be attributed to reduced crosstalk between channels. Muscles are smaller at the wrist and the EMG SNR is lower at the wrist; this might lead to less spatiotemporal summation of the EMG signal, making individual electrode recordings more distinct. This is further supported by our prior work in stroke patients, where classification accuracy was uniquely better at the wrist for the paretic arms that had substantial spasticity and co-contractions [22]. It is also possible that electrical artifacts generated by the tendons moving in the wrist could add a new dimension to the EMG signal that correlates with motor intent. However, our prior work with stroke patients would imply electrical artifacts from tendon motion is not the predominant factor at play; classification accuracy was greater from wrist EMG in stroke patients even when they could not produce visually apparent movement due to paresis.

V. Conclusion

As wrist EMG becomes more ubiquitous, the need to gather as much information from the smallest form factor becomes an important issue. Our results show how moving the recording location at the wrist as little as 4.3 cm (Fig. 5) yields significantly higher information content and myoelectric performance. This work highlights an important aspect of how the anatomy of the wrist affects EMG in a unique way not seen in the forearm. These findings will help to improve wrist-worn EMG sensing, which may in turn allow for more intuitive interaction within AR/VR environments and more robust control for exoskeletons and remote-controlled robotics.

Fig 5.

Distal versus proximal wrist positions. (a) A watch worn in the distal wrist position. (b) A watch worn in the proximal wrist position. The proximal wrist position is ideal for EMG sensing.