The aim of this work was to investigate the feasibility of using force myography signals collected from the forearm as an input method to control a robotic hand orthosis worn on the ipsilateral hand. We collected signals from ten participants performing four gestures in nine different arm configurations, i.e., seven static positions and two dynamic movements, corresponding to typical hand use in daily life. We determined corresponding offline classification accuracies and assessed the impact of individual arm configurations in the classifier training data on the overall classifier performance.
Feasibility of controlling a RHO
The overall achieved average classification accuracy of 92.9% (individual classification
RO/TC: 95.5%,
RC/TO: 90.3%) exceeds the target accuracy of 90% proposed by Scheme and Englehart for reasonable, non-frustrating use in upper-limb prosthetics [
31]. The total response time of the system, i.e., the time from movement intention to classification, consists of four aspects: the time from movement intention (measurable by electroencephalography) to electrical muscle activation (measurable by EMG), typically around 15–25 ms [
32], the time between the onset of electrical muscle activity and the onset of muscle contraction (measurable by FMG), i.e., the electromechanical delay of the muscles, typically around 50 ms [
33], the delay by the sampling of the FMG signal, i.e., 10 ms at 100 Hz, as well as the inference time. Thus, even considering the largest measured inference time of 0.26 ms, the total speed of response adds up to only approximately 85 ms, which is still more than fast enough to produce predictions in a real-time application, assuming that the bandwidth of human hand movement is typically below 4.5 Hz [
34]. We can assume that this holds true even when switching to a microcontroller on-board the RHO with less processing power than the processor used for the analysis in this study, since the inference time makes only a very small fraction of the overall response time. Accordingly, classification accuracy and response time indicate general feasibility, but the fact that one of the classification accuracies only marginally falls above that limit calls for further investigation.
FMG has previously been examined for the control of an RHO by Yap et al. [
15] who reported 95% online classification accuracy distinguishing four hand gestures in a fixed elbow configuration in three neurologically-intact participants in a setting where FMG was measured from the forearm contralateral to the RHO. Such a contralateral setting could foster bilateral movement training, which in turn can promote functional recovery in the contralateral paretic hand [
35]. However, for practical control of an assistive RHO in activities of daily living, it is important that FMG data is acquired on the forearm ipsilateral to a worn RHO so that it does not limit bimanual activities. No previous studies have reported quantitative performance metrics such as classification accuracies using FMG in the ipsilateral setting. Xiao et al. [
17] employed the ipsilateral setting in their study to help the participants familiarize themselves with the control of a RHO by using FMG signals collected from their ipsilateral forearm. Esposito et al. [
18] implemented both, contra- and ipsilateral control and selected between those settings depending on the quality of the measurable FMG signal of the user. However, neither of these studies reported any quantitative results of the ipsilateral setting.
Despite only conducting a presumably simple binary classification compared to previous works distinguishing more gestures, the classification accuracies achieved in our work were not notably superior to other studies that employed FMG for gesture recognition. Using a similar sensor setup as the one used for our work, Xiao et al. [
19] detected grasping actions in pick and place tasks with a comparable accuracy of 92% in neurologically-intact participants; Jiang et al. [
29] were able to distinguish between 48 hand gestures in a cross-trial evaluation with an accuracy of 83.5%. The main reason that we could not achieve notably higher accuracies than these studies is likely the inherent physical restrictions posed by the RHO on the hand. Therefore, the gestures to classify were not as distinguishable as, e.g., a fully open hand and a closed fist, but rather resembled an isometric muscle contraction in the transition phases between the current states of the RHO (trying to close while RHO open/trying to open while RHO closed). In addition, Xiao et al. [
19] also reported significantly higher prediction accuracies when using FMG signals collected from the wrist instead of the forearm. However, for many RHO, the wrist is covered by the device and does therefore not allow placement of FMG sensors in that area. All these observations support that wearing the RHO makes it more difficult to get distinguishable signals and thus correctly classify a desired action compared to gesture classification without wearing a RHO.
The significantly worse performance of the classification “RHO closed” compared to “RHO open” matched the expectations from observations during data collection and unstructured feedback by the participants. In neurologically-intact participants, finger flexors are usually stronger than finger extensors [
36]. This leads to the expectation that during flexion (i.e., gesture
TC) larger and hence more distinguishable FMG signals could be measurable than during extension (i.e., gesture
TO). Further, when the RHO was open, the participants only had to counteract the stiffness of the springs in the fingers to perform the gesture
TC. However, when the RHO was closed, it provided an additional force intended for grasp assistance, which the participants had to counteract in order to perform the gesture
TO. Thus, it might be that the maximum force applied by the participants, as instructed by the experimenters, was too small to achieve a more distinct volumetric change in the muscles.
Looking at the different error modes in Fig.
4, we can distinguish between two types of failure for each classification: when a desired change of RHO-state is wrongly detected or when an actual desired change is not detected. The most frequently occurring error was found when the participants tried to change the state from RHO closed to open (true class
TO), yet the RHO stayed closed (predicted class
RC). While this type of failure might be annoying, it can usually be solved by just trying to conduct the desired gesture again. On the other hand, a misclassification when intending to keep the RHO closed yields an unintended opening (true class
RC, predicted class
TO). When the user is holding an object, this leads to dropping it. This failure mode is therefore considered to be the most critical. Although this failure mode occurred less than
\(7\%\) of the time, it could be considered to adapt the decision threshold of the classifier for future iterations, in order to reduce the risk even further at the expense of a potential increase of undetected opening attempts [
37].
Some variation in performance between participants was observed. Out of the ten participants, only six achieved an acceptable classification accuracy above
\(90\%\) for both classifications “RHO open” and “RHO closed”. Further, although the classification “RHO open” performing overall significantly better than “RHO closed”, there were two participants for which the contrary was observed. Besides the participant’s individual ability to generate consistent muscle activations, further sources potentially introducing variability could be differences in band tightness based on the participant’s oral feedback [
7], inconsistency in the sensor locations as previously investigated in EMG [
38], e.g., after prolonged wearing time, or the amount of force applied against the finger mechanism of the RHO to perform the intended gesture.
An approach to try to compensate for the observed inter-participant variability is to use individually optimized data processing pipelines, i.e., combinations of preprocessing steps and classification algorithms, instead of a general pipeline for all participants. However, when investigating this option in a preliminary analysis, we found that such individual pipelines only performed marginally better in validation and, when translating to unseen data during testing, they performed even slightly worse than the general pipeline which was used in this work. Using the same pipeline across all participants leads to simpler processing and is consistent with other studies involving neurologically-intact participants [
15,
19,
29]. However, it should be investigated whether an individual pipeline could provide a meaningful improvement in accuracy in case of lower muscle activations, e.g., for users with sensorimotor hand impairments.
Arm configuration analysis
Analysing the contribution of individual arm configurations to the overall classification performance, we found that the dynamic movement 9, i.e., the one horizontally on table level, contributed the highest. Including dynamic movements provides data with a higher variability in the training set, which in turn can make a model more suitable for testing in scenarios which also include such variabilities, such as eating or other activities of daily living. The three highest contributing static positions 6, 7, and 2 required the participants to extend the elbow straight to the left humeral plane (Positions 6 and 2) or to mouth level on the sagittal plane close to the body with a flexed elbow. Including these three static positions on different vertical levels yielded in a significant improvement in classification accuracy compared to only using one dynamic movement on table level, matching the suggestion by Radmand et al. [
27] to include positions in both, straight and bent elbow configurations. Introducing the second dynamic movement, 8, resulted in a further significant increase for classification “RHO closed”. This observation matches our expectation, since adding this movement not only introduced dynamic, and thus highly variable data, but also, for the first time, introduces data from the right humeral plane. Accordingly, the data from static positions 5, 1 and 4, which all lie in the same plane as movement 8 might be partially redundant and therefore don’t contribute to further significant improvements. However, surprisingly, introducing the dynamic movement 8 led to a decrease in accuracy for classification “RHO open”. Yet, as the decrease was not significant and the achieved accuracy is still relatively high (
\(>95\%\)) we assume that in this case the additionally introduced variability in data was not required since a plateau was already achieved earlier.
For this work, no data collection time constraints (i.e., large number of repetitions and arm configurations) were considered for the sake of achieving a large training set for the investigations. However, in real-life applications, such a long training time is critical as the users may lose motivation and experience fatigue during prolonged training. These findings lead to the conclusion that the static positions 4, 1, and 5 (in that order) could be removed in a future training data collection without notably decreasing the classification accuracy, reducing the overall training time and therefore the burden on the participants.
Limitations and future work
While this study provided valuable novel insights and a first indication towards the feasibility of using FMG to control a RHO, important aspects need to be taken into account when transferring these findings to practical applications. In a first step, the feasibility should also be investigated for online control. Previous works using EMG have found significantly worse performance in online classification compared to offline [
39]. Further, all study participants were neurologically-intact, male, and represented a relatively low variability in age and forearm circumference. Unfortunately, there is only very limited data available on FMG signals collected from people with neurological hand impairments. For grasping detection of one grasp type, Sadarangani et al. [
40] found that, compared to neurologically-intact individuals, individuals after stroke achieved inferior, but still acceptable (
\(>90\%\)) classification accuracies. For people with hand impairments due to spinal cord injury, no such data is available. However, in this population, the residual muscle activity (and therefore the volumetric change) in the forearm depends on the type and level of lesion which could impact the distinctness of measurable FMG signals. For both these reasons, we expect that the classification accuracy in online control for people with neurological hand impairment would be lower than what was reported in this study. In order to improve online classification accuracy to be suitable for this population, some improvements on the hardware and the data collection are needed. On the hardware side, fusing data from multiple sensor types could be considered. Previous works have shown superior performance when using EMG and FMG data simultaneously compared to only EMG or only FMG [
13,
41]. Further, the inclusion of inertial measurement units could identify dynamic motions in order to avoid misclassifications when transitioning between different hand positions during activities of daily living [
42]. From an application point of view, different gestures could be used to trigger the desired actions. While the gestures used in this study were selected to resemble the targeted action (e.g., “trying to close” in order to close the hand), other less intuitive gestures such as, e.g., a simple co-contraction of the forearm, might be easier to perform for the participants and produce better distinguishable FMG signals [
43]. Further, with data from a larger number of participants, more sophisticated signal analysis such as transfer learning could further allow for improved classification accuracies while keeping the required amount of training in an acceptable range [
44].
Conclusions
This work, for the first time, assessed the feasibility of classifying opening and closing based on FMG data from the forearm while wearing a robotic hand orthosis on the ipsilateral hand in neurologically-intact participants. In an offline analysis, we found that using FMG could be a viable intention detection strategy for such assistive devices, yet for a more conclusive statement, further investigations involving people with hand impairments are required. Additionally, this work identified trade-offs between gesture recognition accuracy and the burden on the user during collection of training data and determined methods to optimize the training procedure and time without reducing gesture classification performance. Based on these findings, methods were identified which could potentially overcome challenges arising when transferring such technologies to their intended context of use, i.e., assisting people with sensorimotor hand impairments during activities of daily living.