Wrist-worn wearables based on force myography: on the significance of user anthropometry

Background Force myography (FMG) is a non-invasive technology used to track functional movements and hand gestures by sensing volumetric changes in the limbs caused by muscle contraction. Force transmission through tissue implies that differences in tissue mechanics and/or architecture might impact FMG signal acquisition and the accuracy of gesture classifier models. The aim of this study is to identify if and how user anthropometry affects the quality of FMG signal acquisition and the performance of machine learning models trained to classify different hand and wrist gestures based on that data. Methods Wrist and forearm anthropometric measures were collected from a total of 21 volunteers aged between 22 and 82 years old. Participants performed a set of tasks while wearing a custom-designed FMG band. Primary outcome measure was the Spearman’s correlation coefficient (R) between the anthropometric measures and FMG signal quality/ML model performance. Results Results demonstrated moderate (0.3 ≤|R| < 0.67) and strong (0.67 ≤ |R|) relationships for ratio of skinfold thickness to forearm circumference, grip strength and ratio of wrist to forearm circumference. These anthropometric features contributed to 23–30% of the variability in FMG signal acquisition and as much as 50% of the variability in classification accuracy for single gestures. Conclusions Increased grip strength, larger forearm girth, and smaller skinfold-to-forearm circumference ratio improve signal quality and gesture classification accuracy.

criteria were self-identified neurological impairments and musculoskeletal barriers to functional movements. The study was approved by the Office of Research Ethics at Simon Fraser University, and all participants provided informed written consent. Testing took place at SFU and Confederation Seniors Centre in British Columbia, Canada.

Anthropometric measures
To investigate and identify the participant-specific measurable features that might cause variability in FMG signals, the following anthropometric measures were considered. Values obtained are reported in Table 1.

Limb length and circumference
Standard positioning for limb length and circumference measurements was a neutral shoulder, 90° flexed elbow, neutral forearm, neutral wrist orientation, and relaxed hand. Forearm length was measured from the olecranon process to the ulnar styloid process and its circumference was measured at the widest part of the forearm. The length of the upper arm was measured from the acromial process to the olecranon process. Wrist circumference was measured within 1 inch proximal to the head of the radius and ulna, i.e., the thinnest part of the wrist. All measurements were done using a standard tape measure and were rounded to the nearest millimeter.

Skinfold thickness
Skinfold thickness was measured from the anterior aspect of the forearm, approximately at its widest part. Methods for taking skinfolds were taken from [42]. In brief, the skin was firmly grasped between the first three digits. The jaws of the calipers were then placed approximately 1 cm from where the skin was grasped, and the skin was released for measurement. An analog Slim Guide Skinfold Caliper (Creative Health Products, Ann Arbor, MI) was used for this purpose, and all measurements were rounded to the nearest millimeter.

Table 1 Participant demographics and anthropometric measures
Values are presented as µ ( σ 2 ), where µ is the mean and σ 2 is the standard deviation

Measure Value
Age (years) 39

Active range of motion
Active range of motion (ROM) of the wrist and the forearm were captured using a goniometer (Jamar Plus + Digital 8″ Goniometer, Patterson Medical, Cedarburg, WI). Wrist ROM for flexion and extension was measured while the shoulder was in neutral position, with the elbow flexed at 90° and the forearm held in neutral orientation. Participants were instructed to actively flex or extend their wrist to its fullest capacity, and measurements for each motion were taken as the smallest relative angle between the second meta-carpal and the central axis of the radius. Measurements were rounded to the nearest degree. Forearm ROM for pronation and supination was measured while the shoulder was in neutral position, with the elbow flexed at 90° and the wrist in neutral orientation. Participants were instructed to hold a pencil in a closed fist while actively pronating or supinating the forearm to its fullest capacity. Measurements for each motion were taken as the smallest relative angle between the starting and end positions of the length of the pencil. Measurements were rounded to the nearest degree.

Grip strength
Grip strength tests were performed with the shoulder and wrist in neutral positions and the elbow flexed at 90°. A non-deformable digital hand dynamometer (HD-BTA, Vernier Software & Technology, Beaverton, OR) was used to measure grip strength. Participants were instructed to the hold the hand dynamometer in a closed fist and to squeeze it with maximum effort (maximum voluntary contraction) for approximately 3 s. This test was repeated 3 times, and the average value was rounded to the nearest 0.01 kg [43].
All measurements were performed on the right side before the start of the protocol.

Study protocol
Participants were asked to use their right hand to perform instructed movements while seated on a chair of standard height and depth. This section describes the selected movements, the experimental task, and the procedure for data collection and analysis in detail.

Selected movements
Common hand gestures and wrist/forearm orientations for performing activities of daily living [44] were selected as the movement set performed by each participant. The selected hand gestures ( Fig. 1 1-7) included: 1. Relax: fingers and thumb are not actively engaged in flexion or extension, 2. Open: fingers and thumb are fully extended and fully abducted, 3. Close: a fist with the buttressing of the distal tips of the phalanges against the central palm and buttressing of the thenar eminence and thumb against the dorsal surfaces of digits 2 and 3 [45] or the lateral aspect of the 2nd digit, 4. Point: only the 2nd digit (index finger) is fully extended, and the pad of the thumb is resting on the lateral aspect of the 3rd digit, 5. Key: the pad of the thumb is in contact with the proximal interphalangeal joint of the 2nd digit, For each of the wrist and forearm orientations (flexion, extension, pronation, supination, and neutral), unless otherwise indicated, participants were instructed to keep the fingers and thumb fully extended and adducted ( Fig. 1 8-12). It is also worthwhile noting that the straight hand gesture (Fig. 1 7) and the neutral wrist/forearm orientations ( Fig. 1 12) are similar. Therefore, when considering a combined set of hand gestures and wrist/forearm orientations, 11 distinct gestures were included.
To evaluate grip strength, participants were instructed to grip a hand dynamometer with the following grips: cylindrical, key and tripod (Fig. 2).

Experimental task
Three groups of tasks were defined based on the selected hand gestures and wrist/forearm orientations ( Fig. 1):   1. Static singleton gestures: participants were instructed to perform each of the hand gestures and wrist/forearm orientations individually, for a total of 11 different classes (gesture groups). Participants performed 5 randomized repetitions of each static gesture for approximately 7 s each. 2. Static compound gestures: participants were instructed to simultaneously perform a hand gesture and a wrist/forearm orientation. For example, for 'point_supination' , participants were instructed to simultaneously point the hand (Fig. 1 4) and supinate the forearm (Fig. 1 11). Participants performed 5 randomized repetitions of each possible combination for approximately 7 s each. 3. Dynamic motions: participants were instructed to either move between two ROM extremes (i.e., from wrist flexion to extension), or to grip while applying minimal to maximal effort. Participants performed one repetition of each dynamic motion for 60 s. Table 2 provides an overview of the specific gestures that constitute each group of tasks.
Participants were in a relaxed posture before the performance of each gesture/motion.

Force myography (FMG) band
A custom force myography (FMG) band was designed to quantify limb volumetric changes while performing the selected movements. Force-sensitive resistors (FSRs) were selected for FMG acquisition due to their low-profile dimensions, flexibility, costeffectiveness, wide-spread availability, and their ease of implementation into a portable and wireless device. The band used in this study utilized 16 FSRs (FSR 400, Interlink Electronics, Inc., Los Angeles, CA) in a staggered design. An FSR consists of a polymer thick film (PTF) circuitry printed on a flexible substrate with a variable force-resistance relationship. FSRs were implemented in series with a 4.6-kΩ resistor, and supplied with a voltage of 3.7 V. An ATMega328 microprocessor (Microchip Technology, Chandler, AZ) was used to facilitate data collection and transmission. Each FSR was sampled at approximately 10 Hz, with raw analog values converted to a digital signal ranging from 0 to 1023 (0.00361 V/bit). Digital values were time stamped and transmitted to an on-site computer via serial connection, then saved onto a .txt file for offline processing. The FSRs were backed with Flex foam and secured to the interior of the band, which was lined with cellulose acetate, a flexible and non-elastic material commonly used in overhead projector film transparencies. This material was used to facilitate a firmer contact between the FSR and the skin, while allowing the band to conform to the shape of the wrist. Participants donned the FMG band on the wrist, 1 to 1.5 in. proximal to the radial and ulnar styloid process surface landmarks. The placement of FSRs on the band is shown in Fig. 3 1. The placement of the band on the right wrist was similar to that shown in [46].
Instructions regarding the gestures were presented as images via a visual interface ( Fig. 3 2) designed in LabVIEW 2014 and displayed in real time on a monitor positioned between eye level and desk level. When necessary, the investigator demonstrated the hand gesture desired at the time and/or corrected the participants' hand gesture to ensure uniformity between all participants.

Data processing
FMG signals were collected while participants performed instructed hand and wrist gestures ( Table 2), then normalized based on the global minimum and maximum of acquired data. We then generated several ML models to establish trends of FMG behavior that were persistent despite selected models and chosen parameters. More specifically, four multi-output ML models, namely artificial neural network (ANN), extreme machine learning (ELM), linear discriminant analysis (LDA), and support vector machine (SVM), were generated and trained using built-in machine learning functions in MATLAB R2016a.

Gesture classification using machine learning
All ML models were trained using data collected while performing static singleton gestures (Table 2). These models previously demonstrated good accuracy in hand gesture classification based on FMG data [17,47,48]. ML models were then used to identify the singleton gestures performed during compound gestures or dynamic motions. For instance, during the 'Wrist Flexion/Extension' task, the model output is expected to be either flexion, neutral, or extension, based on the orientation of the hand with respect to the forearm. An exploratory analysis considering 11 static gestures ( Table 2) as the  ML model output showed that training the ELM model using raw FMG data results in higher classification accuracy, while the classification accuracy of SVM, LDA, and ANN models is higher if normalized FMG data are used as the model input.

Outcome measures
The primary outcome measure for this study was the correlation between anthropometric measures ("Grip strength" section) and FMG signal quality/ML model performance. Spearman's correlation coefficient (R) was used to evaluate such a correlation: a smaller R value, where |R| < 0.33, was considered a weak correlation, a medium range R value, 0.33 ≤ |R| < 0.67, represented a moderate correlation, and 0.67 ≤ |R| demonstrated a strong correlation. Anthropometric measures considered were grip strength, skinfold thickness, and wrist and forearm circumferences. Moreover, several measurements were combined to create the following measures: 1. Ratio of skinfold thickness to forearm circumference (skinfold:forearm), which is an indicator of the ratio of bone/muscle to skin/adipose tissue and has a value in the range of [0,1], 2. Ratio of wrist circumference to forearm circumference (wrist:forearm) which is an indicator of muscle bulk around the forearm and has a value in the range of [0,1].
The quality of acquired FMG signals was quantified using two measures: 1. separability of gesture class clusters: to quantify the linear separateness of the data clusters used in classification, the separability index, J B/W , was defined as where are within-class and between-class scatter matrices, respectively, with I. ω i : class label II. m i : mean of class ω i III. K i : number of samples in class ω i IV. m : overall mean V. K : overall number of samples VI. P i : the a priori probability of class ω i = K i K VII. S i : scatter (covariance) matrix for class ω i Intuitively, the optimal separability, i.e., larger values of J B/W , is achieved by maximizing the between-class variance and minimizing the within-class variance. 2. Sensitivity: this measure was defined as the magnitude of FMG response to incremental changes in orientation. Sensitivity was represented as the root mean square (RMS) of change in the FMG signal magnitude from baseline values across all sensors.
The performance of the developed machine learning models in detecting static and dynamic gestures was quantified using the following measures: 1. In the case of static gestures, the model performance was represented by accuracy, i.e., ratio of correctly classified gestures over total number of performed gestsures. 2. In the case of dynamic motions, the model performance was represented by the following measures: (1) the number of correct classes, (2) whether these classes overlapped with the training region. For instance, during the 'Wrist Flexion/Extension' task, the number of correct classes is three (flexion, neutral, and extension), and as long as the wrist is flexed, the output class is expected to be 'flexion' .

Participants' anthropometric measures and range of motion
Anthropometric measures were determined for each participant as detailed in "Methods" section. Table 1 summarizes the average and standard deviation for those measurements. During a number of dynamic motions, e.g., wrist flexion/extension, participants were asked to move through their full range of motion. The ROM values for different joints, measured at the beginning of the data collection session, are listed in Table 3.

Correlation between anthropometric measures and separability
At neutral wrist, the mean (standard deviation) separability of FMG clusters across the seven hand gestures (Fig. 1 1-7) was 124.79 (96.35). With a straight hand gesture, the mean (standard deviation) separability of FMG clusters across the five wrist/forearm positions (Fig. 1 8-12) was 76.21 (41.28). Mean correlations between user anthropometry and data cluster separability are tabulated in Table 4.

Correlation between anthropometric measures and sensitivity
Mean correlations between user anthropometry and sensitivity of acquired FMG signals are tabulated in Table 5. Grip strength and ratio of skinfold thickness to forearm circumference demonstrated the greatest relationship with FMG responsiveness.

Correlation between anthropometric measures and model accuracy in classification of static gestures
To train the selected ML models, data collected from each participant were randomly split such that 60% of the collected data corresponding to each gesture were assigned to the training set, and the remaining 40% was used for testing the trained models. Mean training accuracies were 88.06, 90.38, 99.08, and 99.92 for SVM, LDA, ELM, and ANN, respectively. These models achieved mean testing classification accuracies of 87.95, 89.96, 98.89, and 99.82, respectively. It should be noted that an accuracy of 100% demonstrates perfect classification performance. Correlations between anthropometric measures and testing accuracy of considered ML models are tabulated in Table 6. Results  show that the accuracy of ELM and NN models weakly correlated with anthropometric measures. However, the skinfold thickness and ratio of skinfold thickness to forearm circumference highly affected the accuracy of SVM and LDA models, while other anthropometric measures had moderate-to-low correlations with the accuracy of these models.

Correlation between anthropometric measures and model performance in classification of dynamic tasks
The ML models trained for classification of static gestures were applied to the dynamic tasks presented in Table 2. Figure 4 shows the static gestures used to train models (blue box), non-static tasks onto which the trained models were applied (orange boxes), and the expected classification outputs based on the considered static gestures (gray boxes). Figure 5 shows an example of the performance of trained models in identifying the dynamic wrist extension/flexion motion. In the shown case, the number of classes is three: flexion, neutral, extension (marked by number 1 on Fig. 5). The identified classes overlap with corresponding actual wrist orientations in regions where circular markers have the same color as the background shade (an example is marked by a purple circle numbered 2 on Fig. 5). Moreover, the variability within each cluster is characterized by the relative number of model outputs that did not correspond to the actual wrist orientation (an example is marked by a purple circle numbered 3 on Fig. 5).
Considering the data collected from all participants performing the dynamic tasks, Fig. 6 1 shows the accuracy of trained models in identifying the correct number of gestures for each task. Generally speaking, the correct number of gestures was correctly identified in at least 57% of cases across different tasks and trained models. Moreover, the number of gestures corresponding to dynamic wrist flexion/extension motion and forearm pronation/supination motion was correctly identified in more that 80% of cases using any of the trained models. Figure 6 2 shows the proportion of correct overlap between the identified gestures and the corresponding actual orientation (training region) in wrist flexion/extension and forearm pronation/supination motions. Defining 'Extremes' as the fully flexed/extended state for the wrist and the fully pronated/supinated for the forearm, and 'Neutral' as the  neutral wrist and neutral forearm position, the percentage of correct overlap is higher in the 'Neutral' , reaching 96% using the trained SVM model. In the case of 'Extremes' , SVM shows a high overlap percentage of 93%. Figure 6 3 shows that continuous identification of close, key, and tripod grips during dynamic cylindrical grip, key grip, and tripod grip tasks is quite poor. This is indicated by low correct overlap (between 11% and 28%) between the identified gesture and its actual counterpart. Such a poor performance might be attributed to the participants' handling of an object (a hand dynamometer) in the dynamic tasks, which was not present when performing the static gestures All model performance measures reported in this section had low correlations (|R| < 0.3) with anthropometric measures.

Additional results
Although this study was not focused on the effects of age and gender on the outcome measures, we investigated further. Specifically, participants were divided into two age  groups, namely seniors (above 60 years old) and non-seniors (between 19 and 60 years old). Table 7 provides age-and gender-specific details of participant demographics and anthropometric measures. Student's t test and ANOVA were implemented to reveal whether age and gender affect the quality of acquired FMG signals and the performance of the developed machine learning models. This analysis showed age-associated differences in the accuracy of classification, and in the separability of hand gestures and wrist gestures (p-value < 0.005). However, it should be noted that the number of participants was not balanced in age and gender groups (Table 7). Therefore, these results might have been skewed. This aspect is addressed further in the Discussion.

Discussion
Force myography is a non-invasive method to measure volumetric changes in muscle cross-sectional area, which are transmitted through tissue. The resultant surface pressure changes can be used to track and identify movements and gestures. Due to the nature of force transduction through underlying tissue, user anthropometry may influence the quality of FMG signal acquisition and FMG-based modeling. Understanding how anthropometry can confound FMG performance is important for the creation of robust user-centered designs for wearable technologies. This exploratory study investigated the relationship between user anthropometric measures and FMG acquisition/ modeling, and further quantified its magnitude. With respect to FMG sensitivity, the greatest significant contributors to variability in FMG responsiveness were ratio of skinfold thickness to forearm circumference (R = − 0.55), and grip strength (R = 0.55). Grip strength is an anthropometric measure directly related to muscle fiber cross-sectional area. As FMG measures the volumetric changes that occur with activity, it is understandable that lower grip strengths would result in lower magnitudes of changes, which is supported by our results. When performing static gestures, increased grip strength was associated with improved hand gesture separability (R = 0.53). The relationship between grip strength and wrist motions was not as strong (R = 0.34). This difference is attributed to the muscle-tendon organization of the forearm and recruitment during wrist/hand movements. In the distal forearm, the anterior flexor compartment is composed of tendons responsible for finger flexion. Relative to motions of the wrist, these tendons are more engaged in the cylindrical grip during grip strength measurements, as well as in 5 of the 7 hand gestures, hence the stronger correlation. Despite the differences in the separability of wrist gestures versus hand gestures, the strong relationship between grip strength and gesture classification accuracy has implications for using FMG with certain populations.
Another variable related to muscle cross-sectional area and grip strength was the ratio of wrist to forearm circumference. Previous studies show a correlation between grip strength and forearm circumference [37,38]. In the present study, the greater the forearm circumference relative to the wrist circumference, the better the improvements in sensitivity to change (R = − 0.39), which further highlights the significance of strength, as well as the underlying musculoskeletal bulk in FMG signal acquisition and gesture classification.
The correlation between the ratio of skin fold thickness to forearm circumference and FMG sensitivity had a moderate and negative correlation (R = − 0.55). As non-rigid structures, the skin and subcutaneous fat act as dampeners of the forces produced by muscle fibers as they are transmitted to the FSR sensors. Thus, an increase in adipose tissue likely increases the dampening effect on FMG signals and decrease the discriminability of gestures. This ratio also takes into consideration the combined effects of the skin/fat dampening and muscle cross-sectional area. A high ratio of skinfold thickness to forearm circumference also strongly correlated with lower classification accuracy, i.e.,  Using models trained with static gestures to identify expected similar gestures performed during non-static activity presented moderate-to-high success (Fig. 6 1). However, this performance was not consistent throughout the full range of motion or effort and was not necessarily related to training conditions. Clusters of correctly identified activity were more consistent at the extremes of movements, but were least likely to overlap with the training region for that intended gesture. At neutral wrist/forearm, there was a high degree of variability and misclassification, however it is surmised that this is due to the inherently low separability of hand gestures. When the models were trained by performing hand gestures without objects, there was a low probability of correctly identifying that gesture in the presence of an object (< 27%). The performance of static models during non-static activity had little to no relationship with anthropometry measures and is therefore more likely related to the nature of the machine learning paradigm itself.
The identification of anthropometric measures that can influence FMG signal acquisition quality and modeling accuracy has the potential to enhance the development of FMG technology. It is recommended to explore these factors as weighting factors for classification and regression algorithms. In addition, this knowledge would be beneficial for sensor selection and calibration in future designs. However, the authors acknowledge that the current study has limitations. First, the number of participants is relatively low, and more participants would increase statistical power. In addition, the protocol was performed using the right arm, regardless of hand dominance. Grip strength has a documented relationship with hand dominance [39], which was not addressed in the current work. It is recommended that future iterations of this work be performed on both dominant and non-dominant hands and evaluate intra-subject variances.
The participants recruited for this study covered a wide age range ( Table 7). The emphasis of this study was not on age and gender effects, although previous research reported a relationship between age, gender and weight, and anthropometry measures such as skinfold thickness [40]. Given the potential for FMG technology to be used as an assistive tool for senior populations [41], we thus encourage future studies to consider the effects of age and gender on the relationship between anthropometric measures and FMG signal acquisition quality and modeling accuracy.

Conclusions
We presented data and analysis identifying and quantifying how individual differences in anthropometry measures can affect the quality of FMG signal acquisition and FMGbased modeling. Sixteen force-sensing resistors were arranged as a grid in a portable and wearable band to track hand gestures and wrist/forearm orientations. Participants perform a pre-selected set of gestures commonly used to perform activities of daily living. The protocol as presented identified three key anthropometric measures, namely grip strength, ratio of wrist to forearm circumference, and ratio of skinfold thickness to forearm circumference, as having moderate (0.3 < |R| < 0.6) influence on FMG signal acquisition and modeling. Increased grip strength, larger forearm girth, and smaller