Stroke is one of the most prevalent causes of adult disability (Gresham et al., 2004; World Health Organization, 2006). Individuals with stroke often experience impairments in the upper extremity, including a reduction in fine motor control, that contribute to difficulties in completing activities of daily living (ADL), such as dressing, feeding, and home management (Gresham et al., 2004). There is increasing evidence that it is necessary to practice hundreds, if not thousands, of grasp and release motions to optimize hand motor recovery after stroke (Nudo et al., 1996; Murata et al., 2008). As it is unlikely that traditional stroke rehabilitation services can accommodate one-on-one therapies for many hours of practice, this has resulted in increased demand for the creation of wearable sensors to assist therapists and patients in monitoring the large number of repetitions needed to promote motor recovery. Such wearable sensing technologies could be used outside of the clinic to monitor “homework,” as task practice outside of structured therapy sessions can improve neurological recovery and functional abilities (Dobkin, 2004; Harris et al., 2009). Wearable sensors could also be used to encourage individuals with stroke to continue to use their affected limb as part of ADL and therefore avoid the “learned non-use” phenomenon (Taub et al., 2006).

Several types of wearable solutions have been proposed for upper-extremity activity monitoring. One type of sensing system utilizes wrist-worn accelerometers. With appropriate signal processing, accelerometer-based devices have been shown to provide metrics that correlate to the amount of upper-extremity movement achieved by individuals with stroke (Uswatte et al., 2000, 2005, 2006) and the amount of hand use in older adults (Rand and Eng, 2010). However, such devices are unable to determine if the user has successfully grasped an object and are thus unable to distinguish functional use of the upper extremity from non-functional movements, such as shaking or swinging of the arm. Given that the ultimate goal of the rehabilitation process is to restore, enhance, and maintain functional ability (Granger, 1998), it is desirable for rehabilitation sensing devices to be able to distinguish between functional and non-functional movement. To detect functional movement, an activity monitor would need to be able to detect both upper-extremity movement and functional use of the hand.

The use of magnetic sensing systems for detecting wrist and hand movements has also been proposed (Friedman et al., 2014). Such devices use a wrist-worn magnetometer to detect relative movement of a magnetic object donned on a different body part, such as a finger. Friedman et al. have successfully demonstrated the feasibility of such devices at detecting the joint movements associated with the wrist and finger (Rowe et al., 2013; Friedman et al., 2014). However, magnetic sensing systems are susceptible to magnetic interference from other ferromagnetic materials, such as household electronics and hence may have difficulty accurately monitoring functional activity in the home environment (Friedman et al., 2014). Several gloves embedded with sensors for tracking hand postures have also been developed and successfully evaluated (Dipetro et al., 2008). However, such devices can be challenging to don on individuals with limited range of motion and spasticity. In addition, the donning of gloves results in a reduction in palmar sensation and may hamper the completion of fine-motor tasks.

Surface electromyography (sEMG) has been successfully used to detect hand activities in individuals with stroke (Lee et al., 2011; Zhang and Zhou, 2012; Li et al., 2014). Lee et al. (2011) used sEMG data from individuals with stroke to identify six different hand postures. Despite these promising results, challenges remain in the use of sEMG for unobtrusive monitoring of hand motion. sEMG requires the use of bulky electrodes with complex signal acquisition and amplification hardware, and a high signal sampling rate. In addition, given that sEMG is measuring the electrical activity of muscles, its signal-to-noise ratio is sensitive to the positioning of electrodes as well as the skin impedance (Merletti et al., 2001), which can be affected by the presence of hair, sweat, and skin creams. Mechanomyography (MMG) is a complimentary sensing technology, which involves transducing the mechanical oscillations of muscle fibers as they are recruited (Islam et al., 2013). MMG senses the mechanical analog of the motor unit electrical activity measured by sEMG (Islam et al., 2013) and has been shown to be capable of detecting the functional state of the hand using pattern recognition and classification techniques (Islam et al., 2013). However, the performance of MMG is similarly sensitive to sensor placement, and sensor pressure (i.e., adherence) on the surface of the skin (Islam et al., 2013), and requires similarly complex signal-processing methodologies, as those used in sEMG classification.

Force myography (FMG) is an alternative technique that involves monitoring the force, or pressure, at the surface of the limb, as a means to characterize the state of the underlying musculotendinous complex (Wininger et al., 2008). FMG provides a simple-to-use, inexpensive, and unobtrusive method for sensing muscle activity. In contrast to electromyography, the FMG signal can be extracted using off-the-shelf force-sensing elements and does not require complex signal-processing circuitry (Dementyev and Paradiso, 2014). In addition, FMG does not require the sensors to be placed at specific anatomical points on the body to ensure adequate signal acquisition (Castellini et al., 2014), as is required in sEMG (Merletti et al., 2001) and MMG (Islam et al., 2013). Given these considerations, a growing amount of research is exploring the use of FMG for activity-tracking, gesture-recognition, and the detection of functional tasks. FMG systems have been used to predict grip strength (Wininger et al., 2008) and single-finger forces (Castellini and Ravindra, 2014). More recently, FMG has been combined with machine learning and pattern recognition techniques to create systems that can detect upper-extremity postures (Xiao and Menon, 2014), hand gestures (Dementyev and Paradiso, 2014), and repetitions of grasp and move tasks in healthy participants (Sadarangani and Menon, 2015). Dementyev et al. demonstrated classification accuracy in excess of 80% when classifying six different grasp types using an FMG band with healthy participants (Dementyev and Paradiso, 2014). Li et al. (2012) demonstrated classification accuracy of 99% at classifying 17 different grasp types using a high-density FMG array with healthy participants. The results of these studies give credence to the concept of classifying the FMG signal to detect grasping.

Despite these promising results, little work has been done to establish the feasibility of using FMG for detecting grasping in populations with upper-extremity impairments, who might ultimately benefit from this technology. Yungher and Craelius (2012) have shown that FMG is an effective method for providing biofeedback to improve fine motor function, as part of structured upper-extremity rehabilitation for individuals with upper-extremity impairment, including individuals with stroke and individuals with traumatic brain injury. In their study, the authors requested participants to pinch at predetermined intervals and used the difference between FMG signatures for the requested pinch gesture and a template pinch gesture to provide biofeedback to participants. Experimental results indicate that gesture recognition-based biofeedback improved the outcomes of repetitive task training. While the results of this study give credence to the concept of using FMG sensing for providing biofeedback in rehabilitation applications, the study did not evaluate the utility of FMG for monitoring and distinguishing between grasping and no grasping, in the presence of other upper-extremity movement which the FMG signal has also been shown to be sensitive to Xiao and Menon (2014). Instead, the proposed system involved monitoring the FMG signature of a single grasp type, which was elicited at a known time, in the absence of any other upper-extremity movement. To be suitable for monitoring functional use of the hand in daily living, for rehabilitation applications, FMG sensing systems will need to be able to detect grasping in individuals with stroke, without knowing when the grasp may be elicited, despite the presence of other confounding upper-extremity movement.

Several characteristics associated with individuals with stroke may impact the feasibility of using FMG for detecting grasping in this population. Individuals with stroke have reduced muscular strength (Bohannona, 1987) and greater muscle spasticity (Watkins et al., 2002) which may affect the magnitude and quality of the FMG signal. In addition, movements completed with the stroke-affected limb have reduced range of motion, are less smooth, and involve variations in speed and acceleration when compared to the less affected limb (Kamper et al., 2002). Given that the FMG signal is sensitive to postures of the hand, wrist, forearm, and elbow (Xiao and Menon, 2014), the movement features of the stroke-affected limb, and the associated compensatory mechanisms employed, could introduce challenges when detecting grasping in the presence of movement in the three-dimensional workspace. Furthermore, for FMG to be feasible in busy clinical settings, the initial setup of the device (i.e., number of grasp repetitions required to train the grasp detection classifier) would need to be minimal in time and in repetitions.

This study explores the feasibility of using FMG, with machine learning techniques, for grasp detection in the presence of upper-extremity movements, for individuals with stroke who have arm and hand impairments. As noted, the use of FMG classification for grasp detection continues to be an active area of research in healthy volunteers. However, given that the FMG signal also captures information related to upper-extremity movements and postures other than grasping, the use of FMG classification for grasp detection in individuals with stroke, who may have altered muscular characteristics and movement patterns, has yet to be established. Feasibility was assessed by (1) establishing and comparing the accuracies of FMG-based grasp detection for individuals with stroke and healthy individuals, with FMG data corresponding to grasping, releasing, moving while grasping, and moving without grasping, using linear and non-linear classifiers and (2) determining and comparing the amount of training data necessary to accurately classify a grasp with participants with stroke and healthy participants. We considered the device to be feasible if the grasp detection accuracy was at least 90%, and if the training data required for the classifier were at most 50% of the size of the testing data. We hypothesized that (1) the use of FMG for grasp detection in individuals with stroke would be feasible, yielding an accuracy of at least 90% with training data that are at most 50% of the size of testing data, in a controlled environment, (2) classification accuracy would be lower for participants with stroke when compared to healthy participants, and (3) more data would be required for accurate grasp detection for participants with stroke when compared to healthy participants.

All eight participants with stroke and eight healthy participants successfully completed the data collection protocol. The Kolmogorov–Smirnov test confirmed that the acquired data were normally distributed, and hence suitable for classification using LDA.

Figure 4 depicts the predicted class label in relation to the true class label for the RBF-SVM and LDA classifiers for an exemplary case of Task 1 for the first participant with stroke. Table 1 shows the accuracies associated with grasp detection for each task, for participants with stroke and healthy participants using the RBF-SVM and LDA.

Grasp detection accuracy was lower for participants with stroke, when compared to healthy volunteers. Average grasp detection accuracy was 92.2% (σ = 3.5%) and 91.6% (σ = 3.8%) across all tasks for participants with stroke using the RBF-SVM and LDA, respectively. Average grasp detection accuracy was 96.0% (σ = 1.6%) and 94.9% (σ = 2.3%) across all tasks for healthy participants using the RBF-SVM and LDA, respectively. The two-way ANOVA revealed that the higher classification accuracies obtained for healthy individuals when compared to individuals with stroke was statistically significant (F_2,31 = 11.65, p < 0.005). There was no significant difference in classification performance between the RBF-SVM and LDA classifiers, and there was no interaction effect between participant type and classifier type.

Classification accuracies obtained were lower for moving upwards (Task 2), or forwards (Task 3), when compared to moving laterally (Task 1) for both participants with stroke and healthy participants. To explore the possibility of using a single model to detect grasping across all movement directions, a single classifier model was constructed using data from all three tasks. In specific 50% of the repetitions from all three tasks were used for training, and 50% of the repetitions from all three task were used for testing. A two-fold cross-validation scheme was employed. The classification accuracies obtained were lower than those obtained when using a single classifier model for each task. Classification accuracies for participants with stroke were 91.6% (σ = 3.4%) and 90.9% (σ = 3.4%) using the RBF-SVM and LDA classifiers, respectively. The classification accuracies obtained for healthy participants were 94.9% (σ = 2.5%) and 94.0% (σ = 2.8%), using the RBF-SVM and LDA classifiers, respectively.

Figure 5 depicts the average accuracies taken across both folds and all tasks, for training sets of differing sizes for participants with stroke and healthy participants. The accuracies associated with each task are depicted in Figure 6. The accuracy was dependent on the size of the training set provided. The correlation coefficients between the number of training samples and the average accuracy for participants with stroke are 0.8606 (p = 0.001) using the RBF-SVM, and 0.8488 (p = 0.002) using the LDA. The correlation coefficients between the number of training samples and the accuracy obtained for healthy participants are 0.7604 (p = 0.011) using the RBF-SVM, and 0.6254 (p = 0.05) using the LDA. The use of two and four repetitions was necessary for achieving greater than 90% accuracy for participants with stroke using the RBF-SVM and LDA, respectively (Figure 5). The use of one and two repetitions was sufficient for achieving greater than 90% accuracy for healthy participants using the RBF-SVM and LDA, respectively (Figure 5).

The overall objective of this study was to explore the feasibility of using FMG for grasp detection in individuals with stroke. Grasp detection accuracy was evaluated using an experimental protocol that involved grasping, releasing, moving while grasping, and moving without grasping, in the three planes of movement within which we perform our daily activities. Additional variability was introduced by the variety of movement trajectories that were employed by participants to achieve the tasks, as participants were free to employ varying wrist, elbow, and shoulder motions to complete the target movement required for each task. In addition, participants were also able to vary the grasp type used, as long as they picked up the cup using the instrumented handle. Variability was also introduced into the paradigm by the functional impairments that the participants with stroke demonstrated, which are reflected in the clinical scores. For example, a Chedoke Hand Score of 5 would represent an individual who has difficulty extending their wrist and spreading the fingers apart. Grasp classification accuracy was established and compared for participants with stroke and healthy participants using both RBF-SVM and LDA classifiers. The susceptibility of classification accuracy to the amount of training data provided was also established.

Grasp detection accuracy was lower for participants with stroke when compared to healthy participants. A potential explanation for the lower grasp detection accuracy is the empirically observed variability in grasp types, movement trajectories, and additional compensatory mechanisms demonstrated by participants with stroke. This variability may have resulted in a more challenging classification problem. In addition, participants with stroke often had difficulty in grasping and completely releasing the cup at the start and end positions, which potentially made classification of those points of time more challenging. It is noteworthy that there is an age difference in the groups of participants enrolled in this study. Aging can affect the properties of muscle fibers (Deschenes, 2004) and subcutaneous fat, which may impact the quality of muscle based signals (Vaillancourt et al., 2003). Hence, the difference in accuracies observed between the participants with stroke and healthy individuals may not be related only to the existence of upper-extremity impairment due to stroke. Instead, these results show the accuracy obtainable, within the confines of the experimental protocol, with FMG-based grasp detection for the target population of individuals with stroke, in comparison to the population of healthy, young individuals, in which FMG has been, and continues to be, widely studied (Dementyev and Paradiso, 2014; Xiao and Menon, 2014; Sadarangani and Menon, 2015).

The use of a linear classification scheme did result in lower classification performance for both participants with stroke and healthy participants. However, the differences in classification performance were not statistically significant. Average grasp detection accuracy with the LDA remained above 90% for both participants with stroke and healthy participants, suggesting that linear classification may be adequate for FMG-based grasp detection.

It is noteworthy that classification accuracies were lower for moving upwards (Task 2), or forwards (Task 3), when compared to moving laterally (Task 1) for both participants with stroke and healthy participants. Furthermore, the classification accuracies obtained when using a single classifier model for all three tasks were lower than those obtained when using a single classifier for each task. These variations in accuracy suggest that FMG grasp classification accuracy may be sensitive to movement direction and upper-extremity postures. Despite the noted reductions, average grasp classification accuracy remained above 90% for this combined model, suggesting that the use of a single training round for predicting grasping in all three directions may be feasible.

Overall, FMG-based grasp detection demonstrated high accuracy of 92.2% (σ = 3.5%) with participants with stroke. These promising results met our first feasibility criteria and suggest that the use of FMG for grasp detection in individuals with stroke, in the presence of upper-extremity movement, may be feasible.

Grasp detection accuracy was significantly dependent on training set size in all cases. However, a larger training set size was required to achieve 90% accuracy with participants with stroke when compared to healthy participants. It is possible that a larger training set size was necessary for the classifier to generalize the variability in movement trajectories and grip patterns observed in participants with stroke. Despite the reduction in accuracies, the use of a training set that was 50% the size of the testing set resulted in 91.7% (σ = 3.9%) with participants with stroke, and 95.6% (σ = 1.6%) with healthy participants using the RBF-SVM (Figure 5); and this met our second feasibility criteria. These promising results are indicative of the capabilities of FMG-based sensing systems for grasp detection in individuals with stroke, with minimal setup, in controlled environments.

It is noteworthy that the grasp detection accuracy obtained for this preliminary experimental protocol, corresponds to the detection of grasping of a single object of constant weight and shape. While the use of a variety of movement trajectories, wrist, elbow and shoulder postures, and grasp types introduced variability within the data, this study did not investigate the effect of the weight of the object on FMG grasp detection accuracy. Furthermore, this study also did not seek to establish the effect of removing and redonning the force-sensing band at different positions, and with different levels of tightness, as would be expected in daily use. These variations in object weight, band donning position and band tightness could introduce variability in the FMG signal acquired and impact grasp detection accuracy. It is also possible that other sensing modalities, such as sEMG or MMG may prove more repeatable, and less susceptible to these additional sources of variability when compared to FMG. Based on the promising results obtained in this preliminary investigation, future work should investigate the impact of the aforementioned variables on FMG-based grasp detection accuracy.

The scope of this study is limited to establishing the preliminary feasibility of classifying the FMG signal for grasp detection with participants with stroke. The study evaluated grasp detection accuracy with a single object, of constant weight, in a controlled environment. Future work should explore robustness of FMG signal-processing and feature-extraction techniques for differing objects, various grasp types and movement trajectories in an uncontrolled environment. In addition, FMG-based grasp detection with participants with stroke with moderate to severe impairments should also be evaluated. Furthermore, robust FMG sensing bands and systems will also need to be developed to be practically deployable for grasp detection in uncontrolled environments. This includes an analysis and selection of various force-sensing elements, band backing materials, and battery-based power management systems to make the band wireless. Furthermore, the impact of removal and redonning of the FMG sensing band should be investigated and minimized. Finally, the efficacy of FMG-based grasp detection systems for encouraging grasping and functional activity, as part of daily living, in individuals with stroke who are actively undergoing rehabilitation should also be assessed.

This study investigated the feasibility of classifying FMG for grasp detection, in individuals with stroke, with a data set that included variations in upper-extremity movement and postures, as would be expected in daily living. FMG classification accuracy was lower and required more training data to achieve the target 90% accuracy for participants with stroke when compared to healthy participants. Despite the noted considerations, FMG-based sensing achieved a high accuracy of approximately 92% for detecting grasping in the presence of other upper-extremity movements with participants with stroke. In addition, a training set size that was 50% of the testing set size was sufficient to achieve approximately 91.5% grasp detection accuracy with participants with stroke. These promising results indicate that FMG sensing may be feasible for monitoring grasping in the presence of upper-extremity movements, in individuals with stroke and mild to moderate upper-extremity impairment.

The study received ethics approval from the SFU and UBC Research Ethics Boards. All participants provided informed consent for their participation in the study.

GS designed the experimental device and experimental protocol, implemented data collection software, performed experiments, analyzed experimental results, and participated in manuscript preparation. XJ performed experiments, analyzed experimental results, and participated in manuscript preparation. LS facilitated participant enrollment, administered functional tests, and participated in manuscript preparation. JE contributed to discussions, analysis, and participated in manuscript revisions. CM supervised the project, contributed to discussions, analysis, and participated in manuscript revisions. All the authors read and approved the submitted manuscript.

The principal investigator, CM, and members of his research team have a vested interest in commercializing the technology tested in this study, if it is proven to be successful and may benefit financially from its potential commercialization. The data are readily available upon request.