Predicting wrist kinematics from motor unit discharge timings for the control of active prostheses

Five normally-limbed men and one woman, aged 24–38 years, participated in the study. Moreover, a 57 years-old man with a transradial amputation that occurred 37 years before the experiment was also recruited. He has been a daily user of a myoelectric prosthesis since the amputation.

Depending on the anatomy of the subjects, two or three high-density electrode grids (ELSCH064NM3, OT Bioelettronica) were mounted around the dominant forearm (normally limbed subjects) or the residual limb (transradial amputee). The centerline of the grid was at the distal end of the proximal third of the forearm for each subject (Fig. 1A). Each grid consisted of a matrix of 8 × 8 concentric electrodes with a 10 mm diameter and a 10 mm inter-electrode distance. The electrode grids were connected to pre-amplifiers (AD1x64SD5, OT Bioelettronica) and a laboratory EMG amplifier (EMGUSB2, OT Bioelettronica). The EMG signals were recorded in monopolar mode with the sampling frequency of 2048 Hz, 2nd order band-pass filtered between 3 and 900 Hz, and A/D converted to 12 bits. The ground and reference electrodes were placed around the wrists. The ground electrode in the limb-deficient participant was mounted on the lateral elbow epicondyle of the ipsilateral side. A motion capture system (Xsens Technologies B.V., MTx) was used to track wrist kinematics during the performed tasks, and to provide visual feedback to the subjects (Fig. 1). Three pods were attached to the subjects on the dorsal side of the palm, on the wrist and the upper arm just above the elbow. For the subject with limb deficiency, the motion capture equipment was mounted on the contralateral arm. During the experiments, subjects were seated comfortably with their arms relaxed in the neutral position at the side of the body pointing downwards with no additional constrains to any of the joints.

The subjects performed movements of one DoF at a time guided by a visual cue (Fig. 1B). Horizontal movements of an arrow shown on a computer screen corresponded to flexion/extension, vertical movements to adduction/abduction, and rotation to pronation/supination. Visual feedback on the current wrist position was provided by a second arrow. Subjects were instructed to match the two arrows.

For each DoF, the cue prescribed a triangular trajectory at constant speed for both directions of the DoF and the full range of movement. One run consisted of three of these trials for each DoF (random order across DoFs and trials). The subjects performed three runs at three speeds, corresponding to a duration of the ramps of 5 s (slow speed), 2.5 s (medium speed), and 1 s (high speed). The subject with limb deficiency was instructed to perform the movements in a mirrored fashion with both limbs concurrently. Otherwise, the procedures were the same as for the normally-limbed subjects.

In addition to the analogue filtering, the EMG signal was digitally band-pass filtered using a zero-phase filter of the 5th order with cut-off frequencies 20 Hz and 500 Hz, as commonly used in the myocontrol literature [17]. Signals were visually inspected and noisy channels, which occurred rarely (< 5 channels per recording), were excluded. Then, as recommended [18], the signal was windowed at 100 ms intervals, with 10 ms of overlap resulting in the new feature vector being obtained each 90 ms. The following time-domain features were calculated for each window across all considered channels: root mean square, slope sign changes, zero crossings, and waveform length [3]. The necessary threshold parameters were selected manually for each subject based on visual inspection. The selected values were on average below 10% of the full scale of the amplified signal, and were similar for all subjects. Principal Component Analysis (PCA) was then performed on the extracted feature space containing all trials considered for the controller training, so that the resulting principal components of the features retained 98% of the original variance, as it was previously suggested [19]. This reduced-dimensionality time-domain signal description will be referred to as the TD feature set.

All high-density EMG signals recorded during the contractions could be decomposed using the blind source separation algorithm (an example is reported in Fig. 5). Table 1 shows the number of active motor units during each DoF, including motor units that were active in multiple DoFs. The number of decomposed motor units decreased with the ramp duration.

The full ANOVA detected statistically significant effects between Features (p = 0.0025) and a significant three-way interaction (p < 0.001). Fixing the ramp duration and performing two-way ANOVAs, we found that while the significant effect of features remained present in all comparisons (p < 0.01), there were significant two-way interactions between the factors “Subject” and “Feature”, across all levels. Therefore, the feature performance for each subject and ramp duration was analyzed separately, as described in the Methods. The post-hoc tests identified statistically significant differences between the TD and the neural feature sets in all subjects and ramp durations, with an average R² of 0.77 for the neural features and 0.70 for TD. Similarly, the average R² value of subject D1 improved from 0.52 to 0.64 with the proposed method. Differences between Ramp-Durations were not statistically significant. An example of the regression results is shown in Fig. 6.

We repeated the statistical analysis including the two additional neural feature sets AM1 and AM2. The ANOVA detected statistically significant effects of Features (p = 0.0025), significant two-way interaction between the factors Subject-Ramp Duration (p < 0.001) and Subject-Feature (p = 0.02), as well as a significant three-way interaction (p < 0.001). The post-hoc analysis showed that the proposed methods significantly outperformed both AM1 and AM2 in most cases, and never underperformed them significantly (Fig. 7). As for the proposed method, AM1 also provided an improvement over TD consistently for all conditions.

The number of decomposed spike trains depended on the ramp duration (Table 1), so that more spike trains were identified at higher speeds than at lower ones. This indicates that at least in some phases of the movement at higher speeds the contraction level was higher, and thus more motor units were recruited within the pick-up area of the surface electrodes. We also observed that the muscle activations during the tasks were selective, i.e. there were just a few motor units detected in more than one DoF (Table 1). These units have probably been activated as a part of neural control strategies and had an influence on the wrist joint stiffness. The limitation of the applied regressor is that it interpreted these as concurrent activation of DoFs (Fig. 6).

Neural information extracted from MUAP trains showed better regression performance than traditional TD features, for both able-bodied subjects and a subject with limb deficiency. The tests using the alternative methods showed that both the inclusion of the residual EMG and the proposed dimensionality reduction contributed to the achievement of superior performance.

One plausible reason for the relatively low performance of purely spike-based features is the imperfect decomposition, since including the residual information outperformed traditional features consistently. Additionally, the matching of the MUs across trials was done using a fixed threshold value, which could potentially benefit from case-specific statistical optimization. Another possibility is the non-linear relation between motor unit spike trains and wrist kinematics [33]. Nevertheless, the observation that the proposed method as well as AM1 outperformed TD indicates that spike trains do carry additional information that could not be extracted with traditional features, even for such high channel numbers.

The overall best performing feature set was the proposed model-based neural set (Fig. 7), although a significant advantage resulting from the model-based dimensionality reduction was only observed at the shortest ramp duration (the fastest speed). This is likely due to the decomposition being less accurate for faster contractions, in which case a model-based approach could recover more information. At the same time, TD might benefit from more careful DoF-wise channel selection in addition to PCA. At this stage, the computational load required for extracting neural features is much greater compared to the TD features. The implementation and results presented in this study aimed at a rigorous testing of the concept, to prove the feasibility of the neural approach. Future work should explore online controllers and test their clinical validity with the focus on the implications of the observed increase in offline performance.

The linear regression on spike trains (AM2) was not sufficient to achieve R² performance superior to TD. Regression on spike trains including the residual EMG (AM1), however, outperformed traditional features, although the proposed model-based approach further improved performance. In particular, the model-based approach made the most out of the larger population of decomposed MUs available during the shortest ramp duration. This indicates that the proposed physiologically inspired dimensionality reduction method partly counteracted decomposition inaccuracies.

There were no statistically significant differences in regression performance between ramp durations. This was an unexpected finding since the number of spike trains, and therefore decomposition complexity increases with movement speed. The significant three-way interaction showed that the effect of ramp duration on performance was subject-specific, indicating that multiple factors influenced regression in addition to decomposition complexity. One confounding factor is the biased sample of the motor units detected by decomposition. Because higher threshold units tend to have action potentials with greater energy than lower threshold units, the sample of decoded units mainly comprises of high threshold units, for which the decomposition and waveform comparison task are more accurate [34]. It is also possible that the action potential shape of some lower threshold units was considerably changed due to muscle motion relative to the electrode, resulting in these units not being detected by the decomposition. Moreover, recruitment threshold may vary with contraction speed [35] which may also negatively influence regression performance. Other confounding factors include the variability in subject anatomy and the properties of the tissue layers between the muscle fibres and the electrodes. These factors influence decomposition accuracy in a subject-specific manner, which might have translated to differences in regression performance.

The main limitation of the study is that we used an offline automatic EMG decomposition method, which is not invariant to the movements of muscles relative to the skin surface since it has been developed for low to medium force isometric contractions and has been shown to be only partly effective for dynamic contractions [36]. We also do recognize that the improvements in the offline control do not necessarily result in the increase of clinical scores [37]. Based on the statistically significant improvement in the offline scores, it is indeed difficult to conclude how beneficial the observed increase in clinical performance will be. However, in this study, we aimed to investigate whether the information gained from EMG decomposition can in principle benefit myoelectric control. Moreover, the presented evaluation is not dependent on the data acquisition method, and can be used with any method for extracting spike trains of motor unit populations, including future online EMG decomposition algorithms of surface or intramuscular EMG [38], as well as spike sorting from other signals such as peripheral nerve recordings [39‐41]. It should also be noted that an online implementation of the method used here is feasible [42], and it is also possible to implement an MU tracking algorithm [20] that can provide continuous information on the activity of the relevant MUs while at the same time reducing the computational time needed for signal decomposition.

Another limitation is that we only included single DoF contractions. Although we acknowledge the importance of simultaneous control of multiple DoFs, the present study on single DoFs is a necessary first step for future developments of multi-DoF control based on motor unit activity. Similarly, we have only conducted tests on the tasks that the estimators were familiar with, while additional investigation on handling spurious activity originating from untrained DoFs will be done in future work. Finally, the inclusion of a subject with limb deficiency showed the feasibility of motor unit recordings and regression in the target population of prosthesis users, but these data are not sufficient to demonstrate general clinical applicability or to make more general claims about the observed performance. Overall, having shown that we are able to precisely regress the recovered neural information we have established a framework for the development of more efficient and ultimately real-world viable control systems.