Surface electromyography (sEMG) signals, the muscles’ electrical activities recorded at the skin surface, are used for the control of multifunction upper-limb prostheses (Parker et al., 2006). Currently, with one recently emerged exception, the techniques used in commercial prostheses provide very limited functionalities which can only be operated over no more than two degrees of freedom (DoF) in a sequential manner with unintuitive switching commands, such as a strong contraction over all EMG channels. One of the techniques used to improve the performance of prosthetic control over the past decades is pattern recognition (Scheme and Englehart, 2011), which is the basis of a recently commercial product, COAPT’s complete control system.¹

For one-to-one mapping of EMG signals to specific contraction types, the pattern of surface electromyography (sEMG) elicited during specific movements are stored and analyzed by the pattern recognition algorithm in the training phase; then, the trained classifier is used to classify the EMG signal during the control phase into the intended movement (Li et al., 2010). To achieve low classification error, the signals or features extracted from the signals, in the control phase should be stationary, or similar to those in the training phase. Currently in the literature, various factors that can induce non-stationary change to EMG and EMG features between the training and control phases have been identified. Consequently, these non-stationarities can significantly affect the performance of pattern recognition-based (PR-based) algorithms in activities of daily living (ADL). This lack of robustness is indeed one of the major obstacles for the pattern recognition algorithms to be implemented in commercial products (Jiang et al., 2012). These non-stationary factors include, but are not limited to, arm and trunk positions (Fougner et al., 2011), electrode shifts (Young et al., 2012), subject learning (He et al., 2015a), and contract levels (Kaufmann et al., 2010). Specifically, muscle contraction level was shown to be one of the factors inducing non-stationarity between the training phase and the actual control phase (Kaufmann et al., 2010). The authors showed that when the contraction level between the training and control phases differed, the performance of a large number of classifiers, which had been previously reported in myoelectric control, dropped significantly. In order to address this type of degradation, these classifiers have to be trained with a large range of contraction levels. However, this approach can be impractical in real-world applications, particularly when there are a large number of classes. Alternatively, features of sEMG or specific portions of sEMG signal that do not change (or have limited change) at different contraction levels is preferable as they can lead to an elegant control scheme that is inherently robust against varying contraction levels between the training and control phases. In He et al. (2015b), a frequency-based feature set is proposed as the first attempt to realize a contraction-level independent myoelectric pattern recognition classification. It was shown that with this new feature set, the classification performance was significantly better than a classic pattern recognition algorithm when presented with contraction levels not seen in the training phase. More recently, Al-Timemy et al. (2015) also proposed a feature set based on spectral moment descriptors to improve the robustness of myoelectric control with the presence of contraction-level variation. However, both of these studies (Al-Timemy et al., 2015; He et al., 2015b) were offline studies, where the advantage may or may not be able to translate to online controllability of the prostheses (Jiang et al., 2014c). Lock et al. (2005) showed that the offline performance of PR-based algorithms was not correlated with its online performance. By comparing three simultaneous and proportional myoelectric control algorithms of non-negative matrix factorization (NMF), linear regression (LR), and artificial neural networks (ANN), Jiang et al. (2014c) showed that the offline performance of the algorithms is weakly correlated to the online performance control. In this sense, the advantage of a feature set, or a signal pre-processing technique for online myoelectric control, is not known from the offline study until it is examined in an online myoelectric control scheme.

More importantly, neither of these two studies (Al-Timemy et al., 2015; He et al., 2015b) explicitly exploited the advantage of the fundamental properties of motor units involved in the generation of sEMG signals. Based on the Henneman’s size principle, the recruitment of motor units at different levels is ordered such that smaller units with lower firing rates are recruited at low contraction levels, and larger units with higher firing rates are recruited progressively at higher contraction levels (Henneman et al., 1965). Indeed, the differences of firing rates of active motor units at different contraction levels can lead to different characteristics among different frequency bands of sEMG. In addition, not only does the power spectrum density function (PSDF) of EMG not change uniformly over all the frequency bands in the case of varying muscle contraction levels but also some frequency bands have less dependency on contraction level (Roman-Liu and Konarska, 2009). As such, extracting frequency band-specific information is beneficial in identifying the changes at different frequency bands and in increasing the robustness of the algorithm against changing contraction levels.

The purpose of this study was to investigate the effect of FDT in an online myoelectric control scheme as the applied contraction level varies in the training and control phases. The online control phase was achieved through goal-oriented tasks. The current study explored the normal and high muscle contraction levels, as they are more likely to be used during wrist contractions.

The trajectories of goal-oriented tasks, from one subject for the combinations of process and training methods, are presented in Figure 2. As shown in the figure, the performance of the runs using FDT was better than the performance of the runs with the bandpass method. The median CR of runs with the FDT method was 95% across all non-perfect runs and was 77.5% with the bandpass method. More importantly, the variability in CR for FDT was 3.45%, which was strikingly smaller than the bandpass method with the variability of 12.87% (p < 0.001). In fact, the CR of 50% of the FDT runs lies between 90 and 95%, whereas for bandpass it lies between 66.25 and 85%. Subsequent statistical analysis supported this observation, indicating FDT provided a clear advantage in real-time control over bandpass such that the participants were able to complete the tasks more successfully. Figure 3 clearly presents that the CR of the non-perfect runs are higher with lower variability when the FDT method was used. For the successful trials, in case of found significance, the other performance indices are summarized below. For type II targets, when train-high data were used, the T2R of the FDT method was 3.26 s, which was lower than the bandpass method with a T2R of 4.16 s. The mean of TP for type II targets with the FDT method was 1.64 bit/s, while with the bandpass method it was only 1.41 bit/s. The FDT method outperformed the bandpass method in throughput (TP = 2.10 vs. 1.73 bit/s) for type I targets in train-normal runs. In train-normal runs, the Γ with FDT and bandpass were 58.23 and 51.46%, respectively.

There were significant two-way interactions between the target type and training method for T2R, TP, and NM with corresponding p-values of 0.002, 0.016, and 0.014, respectively, and also between the process and training methods with p-value of 0.003 for Γ. Therefore, a series of focused ANOVA were performed, which are summarized in Table 1. For T2R, a focused ANOVA was performed with fixed target type. Only for type II targets with train-high runs, the T2R of FDT was found to be statistically better than the T2R of bandpass (p = 0.003). For type I and type III targets, no significance between the two process methods (p = 0.681 and p = 0.085, respectively) was found. For TP, the focused ANOVA with only type I targets found significant interaction in train-low runs between the two main factors of process and training methods (p = 0.005). In this case, no significance between the two process methods (p = 0.695) was found using train-high data. For type II targets, FDT was found to be statistically better than bandpass (p = 0.006). The TP of runs with FDT and bandpass was not found to be statistically different for type III targets (p = 0.469). For Γ, a focused ANOVA was performed, and the test found that FDT and bandpass were statistically different in all train-normal runs. In this case, a Tukey comparison revealed that FDT resulted in significantly higher path efficiency than bandpass with a p-value of 0.001. For NM, in neither of the target types (type I, type II, and type III), was significance found (p = 0.996, p = 0.841, and p = 0.195, respectively). Figure 4 presents the result of the statistical analysis for the first four performance indices.

In summary, for the key performance indices of TP, T2R, and Γ, FDT outperformed for at least one of the target types. For train-normal runs, FDT resulted in higher TP for type I and type II targets, and it performed with higher Γ for all target types. In other words, the advantage of FDT was more pronounced when the contraction level of the training data differed from the contraction level of the real-time control tasks; for train-normal runs, FDT had a tendency to have better performance than bandpass. This result was due to the high muscle contraction that the subjects applied during the runs as they were asked to complete the trials as fast as they can. Consequently, the muscle contraction level used during the real-time trials was closer to train-high trials than train-normal trials. The outperformance of FDT over bandpass, while the contraction level of the training data and real-time control were different, clearly highlights the robustness of the FDT method against the contraction level; and this result is in agreement with the secondary hypothesis. The comparison of the process methods revealed that FDT has advantage over bandpass. Also, this advantage was supported by the significantly lower CR variability that FDT had with respect to bandpass. This low variability in CR for FDT supported the main hypothesis of this experiment, which was that FDT provides better online performance compared to bandpass. It is important to note that these comparisons were performed only with successful trials (failure trials were only used for performance index of CR). These analyses implicitly penalized FDT, which had a significantly higher CR. Because if the maximal allowed task completion time was set at a longer time, more trials with bandpass with longer completion time, worse path efficiency, and more overshoot would have been included.

In this study, two sEMG signal pre-processing methods, namely bandpass and FDT, were compared to investigate their robustness against varying contraction levels and consequently the accuracy of PR-based myoelectric control algorithms in the presence of varying contraction levels. A goal-directed motor task involving wrist contractions were used, similar to several previous studies (Ameri et al., 2014a,b; Jiang et al., 2014a,b,c) of multi-DoF real-time myoelectric control. Five performance indices of a goal-oriented online control tasks, time to reach (T2R), throughput (TP), path efficiency (Γ), near miss (NM), and completion rate (CR), were used in this study. It was previously reported that the performance of pattern recognition algorithms is seriously affected by varying muscle contraction levels (Kaufmann et al., 2010; He et al., 2015b). This was confirmed by our results: significantly lower accuracy was observed when the contraction level between the training and control phases was different. Therefore, a more robust signal processing technique that does not lead to online performance degradation with the presence of variation in muscle contraction is highly desirable. In particular, it would be practically appealing to perform the training phase for as few contraction levels as possible.

One of the important outcomes in this study is demonstrating that in an online control scheme, FDT provides a more robust prosthetic control against changing contraction levels than bandpass. Indeed, FDT increases the robustness against contraction level without significantly increasing the computational cost, as only time-domain features are used in the method. The experiment protocol designed for this study intentionally allowed the subjects to have freedom in terms of muscle contraction level in the control phase, as they would in ADLs. This design allowed the subjects to exert high contraction levels during the control phase. As such, the advantage of FDT over bandpass was more pronounced in the train-normal runs where the control phase and the training phase data had the most discrepancy in contraction levels. Another reason for this design was to minimize the correlation between the performances of the subjects in the training and control phases. In fact, the subjects were asked to maintain a constant muscle contraction level in the training phases (normal contraction in the normal runs and high contraction in the high runs), and they had no such restriction on the contraction level during the control phases.

Another important outcome of this study is that the performance of PR-based myoelectric control was higher in general when FDT was used in data processing than bandpass. This conclusion comes from the fact that FDT resulted in lower variability in CR and also higher CR even though the statistical analyses were biased toward the bandpass method by excluding failed trials (as discussed in Section 3).

In this study, similar to Jiang et al. (2014b), the process methods were measured in terms of relative advantage with the performance indices and the control parameters of the tasks (relative area of circular targets, the minimum acceptable time of being inside the circle for success, and the maximum time allowed). To monitor the relative differences, all the subjects’ runs had similar control parameters. Hence, in this paper, although the absolute values of the performance indices were provided, only the statistical difference of the performance indices between the process methods were emphasized.

Considering that all the subjects attended in the study were naïve, the experiment protocol had two sessions for each subject. The first session helped the subjects to become familiar with the protocol of the experiment as well as with using myoelectric control to achieve the goal-oriented tasks.

The main limitation of this study is that the change of feature space is not investigated in the performances. In He et al. (2015b), changes of feature space were studied whenever normalization of data was possible. However, such systematic analysis in feature space change for online performances cannot be conducted as standardization of online data is not possible as in offline analysis.

The current study is only limited to intact-limb subjects. Based on Jiang et al. (2014b), the online performance of PR-based myoelectric control in upper-limb amputees is similar to the performance of intact-limb subjects. Hence, in presence of the amputee subject limitation, it is reasonable to expect a similar advantage of FDT over bandpass in trans-radial amputee subjects. However, this needs to be validated with further experiments.

In this paper, we analyzed two signal pre-processing techniques for PR-based myoelectric control, aiming at improving system robustness against variation of muscle contraction levels. We investigated the performance of the FDT and bandpass methods in presence of muscle contraction variations between the training and control phases, and we found that FDT significantly outperformed bandpass for PR-based prosthetic myoelectric control, with higher online control performance in general, and a significantly smaller inter-subject variability. More importantly, the online control is significantly more robust against changing contraction levels with the FDT method. FDT can increase the freedom of the subject to vary the muscle contraction level used for control, which provides comfort for the user. To emphasize the advantage of FDT over bandpass, future work would include force/torque measurement, as well as standardized trials in the control phase to systematically study the feature space change with the presence of contraction-level change, as well as validation with trans-radial amputees.

The experiment protocol was approved by the University of Waterloo Research Ethics Committee (#21387). Each subject read and signed the informed consent prior to the experiments.

BT contributed to the following: ethical application process, design of the experiment, implementation of the filtering methods, data acquisition through experiment, statistical analysis of the data, drafting the manuscript, and approval of the final version of the manuscript before submission. NJ contributed to the following: ethical application process, concept of the study, design of the experiment, development of the methodology, statistical analysis of the data, revision of the manuscript, and approval of the final version of the manuscript before submission. Both authors, BT and NJ, are responsible for the accuracy of the analysis and conclusions presented in this study.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.