Secondary Logo

Journal Logo

ORIGINAL RESEARCH ARTICLES

Force Myography across Socket Material

Curcio, Brittney C. MSPO; Cirillo, Nicholas V. MSPO; Wininger, Michael PhD

Author Information
Journal of Prosthetics and Orthotics: January 2020 - Volume 32 - Issue 1 - p 52-58
doi: 10.1097/JPO.0000000000000295

Abstract

Force myography (FMG) is an increasingly prevalent technology for detection of neuromuscular activation. FMG yields an inherently low-frequency control signal,1 which reduces computational complexity, and its resiliency in high-moisture environments2–4 makes it an attractive approach for control of prosthetic devices.5,6 However, although FMG has repeatedly demonstrated robustness when applied directly to the skin's surface,1,7–10 the question has not yet been asked: “Is it necessary to apply FMG directly to the skin?”

It may be desirable to place sensors in a way where they are not (or not necessarily) applied directly to the skin, for example, sensors embedded in a sleeve or socket.1,2,11,12 This circumstance creates a dilemma when applying materials to improve fit or comfort. Whereas conventional sensing paradigms (i.e., electromyography) require direct skin contact, they cannot be integrated into the socket in a way that allows ad hoc addition of inserts. Because FMG detects contact forces and not electrical conduction, there is seemingly no strict requirement for direct skin contact. Indeed, if the signal maintains integrity despite indirect contact, it may prove a distinctive advantage in application to prosthetic detection.

Here, we studied the signal from FMG sensors that were obscured from the muscle belly by material insert versus the signal from sensors that were placed directly on the skin's surface. Our objective was to measure the impact of the material insert on the FMG signal to better understand the constraints and opportunities afforded by FMG in prosthetic application.

METHODS

PARTICIPANTS

A convenience sample of 30 healthy volunteers were recruited from the local community. Participants were included if they had at least one full functional arm with musculature below the elbow and were able to provide informed consent; recruited persons were excluded if they were younger than 18 years, unable to provide informed consent, or were legally blind. Otherwise, the inclusion criteria were broad, with an interest in a maximally generalizable study. All participants provided written informed consent before participation in study activities.

MATERIALS

Muscle activity was detected via two force-sensitive resistors (FSRs), 1.72″ on an edge (Adafruit, Model #1075; Adafruit Factory, New York, NY, USA). These sensors were connected to a microcontroller circuit board (Arduino Uno R3; Arduino AG, Zug, Switzerland), as analog input. A custom Arduino script was written to sample at 100 Hz; contact pressure data were manually transferred from the Arduino serial monitor to a text file for retention and analysis.

In order to reproduce the conditions as might be found in clinical application, we selected two interfaces commonly used by persons with amputation: a one-ply cotton stockinet (“Sock,” Alps; St. Petersburg, FL), and a 6-mm gel liner (“Gel,” Alps). A square swatch, approximately 3″ on edge, was cut from both materials. These swatches and the sensors would be secured via a generic strap.

PROTOCOL

After completion of the consent procedure, participants were seated in a way that they could comfortably maintain a relaxed shoulder position with their elbow flexed to 90° and hand at neutral position for an extended time. Participants were palpated to locate the flexor digitorum profundus muscle as follows: the participant was asked to flex fingers 2 to 5 in a supinated position; one of the study team members (listed authors) placed their hands approximately 1 inch distal to the medial side of the forearm in order to identify the largest muscle belly. This method was chosen with the knowledge that the flexor digitorum muscle spanned from the medial epicondyle of the humerus to the phalanges. One sensor was placed at the muscle belly, covered with the interfacial material, and then the second sensor was applied directly over the interface. In addition to the two material conditions (Sock and Gel), a control condition (“Null”) was devised, where the two FMG sensors were overlaid with no material in between. In all three conditions, the strap was secured around the forearm and adjusted until the sensors yielded a preload value of 0.24 to 1.22 V (equivalent to approximately 5% to 25% sensor range).

Participants were asked to perform two tasks: wrist flexion and fist formation. Each task was performed approximately 5 to 10 times, returning to neutral hand posture between repetitions. Both tasks were demonstrated by a member of the study team, with a rhythm of approximately 1 second per posture; however, participants were not specifically instructed as to how long to hold a posture or how intensely to perform the task; that is, the timing and forcefulness of their muscle contractions were not strictly controlled.

In order to reduce the risk of systematic bias, condition (Null vs. Sock vs. Gel liner), task (wrist flexion vs. fist forming), and sensor order (Sensor A vs. Sensor B contacting skin) were randomly assigned on a per-participant basis. In order to limit fatigue and irritation of donning/doffing, condition and sensor order were randomized as a block and task randomized within this block. Thus, both tasks (a set of wrist flexions or a set of fist formations) were performed in random order within each block; three blocks (Null, Sock, and Gel) were assigned in random order within each participant. Sensor order was randomized in parallel to the condition block (Figure 1).

F1
Figure 1:
Illustration of experimental protocol. Right-handed subject with sensor placed medially (left). Sensor on top of interfacial material (Sock or Gel) on top of sensor on surface of skin (upper right). Protocol randomization scheme (lower right). Within each condition (Null vs. Sock vs. Gel), sensor order was randomized (50–50 chance of Sensor A on skin vs. Sensor B on skin), and task was randomized (50–50 chance of fist clench as first task vs. wrist flexion); order of condition was randomly assigned. Not shown: strap for tightening sensor setup.

HYPOTHESIS TESTING

Before data collection, we established tandem primary hypotheses: 1) intersensor correlation would be between ρ > 0.98 in all conditions (Null, Sock, and Gel), and 2) there would be a significant difference in intersensor correlation between the three conditions. Correlation was defined conventionally; that is, Pearson's product-moment correlation coefficient −1 ≤ ρ ≤ 1, where the ρ = 0 denotes a Near and Far sensor whose signals are completely randomly related (i.e., unrelated), ρ = 1 indicates sensor signals that are exact copies of one another, and ρ = −1 indicates sensor signals that are exact reverses of one another. Hypothesis 1 (ρ > 0.98 in all three conditions) was tested via Student's one-sample t-test; Hypothesis 2 (significant differences in correlation between conditions) was tested via multi-way analysis of variance (ANOVA) with task, participant, and test order as noninteracting co-factors.

In addition, we tested several secondary hypotheses. Hypothesis 2 was replicated with two alternative features: the ratio of area under the curve of sensor voltage over time (since amperage was constant over time, we will refer to this as “power”) and the ratio of sensor range (max value minus min value), both taken as (Near sensor ÷ Far sensor). We also tested whether signal duration (amount of time spent above 2% signal range) was significantly different between Near and Far sensor.

EXPLORATORY HYPOTHESES

Separately, we note that FSRs have been reported to gradually change their signal value over time, possibly due to accommodation to a change in thermal equilibrium (shift from room temperature to body temperature upon application to human subject, a so-called “sensor drift”).13,14 This potential for systematic bias was the principal inspiration for protocol randomization (see above). As a pilot assay, we asked two study participants to don the gel setup for an hour, recording wrist flexions and fist clenches every 10 minutes. Raw (unfiltered) data were analyzed similarly to the data collected in the main study.

Second, we recognized a potential conflict in analytical approach: in clinical science, it is conventional to discard statistical outliers, but in field application, there may not be a straightforward criterion for discriminating incoming signals. In reflection of this tradeoff, we decided that our primary analyses would include all specimens. However, as exploratory analysis, we replicated select analyses with a feature set conditioned for statistical outliers among critical parameters.

Lastly, we devised an exploratory hypothesis related to misclassification of grasp-versus-release, with grasp being defined as either flexion or fist-clench, and release being defined as inactivity, according to <10% signal strength within a data collection epoch. The average correlation among grasps was noted for each subject, and one-sided 95% confidence interval was calculated. Then, each release cycle was assessed for its correlation to each grasp cycle within-subjects. Any release-grasp pairing that yielded a correlation coefficient within the 95% confidence interval for grasp-grasp pairings was considered a “false-positive,” that is, a period of data that may be misconstrued as activation in the absence of true intention to grasp. In order to support correlation of potentially disparate signals, each signal was length-matched to the shorter of the two waveforms.

DATA CONDITIONING AND ANALYSIS

In order to ensure adequate sampling, datasets containing fewer than five specimens were removed from the analysis stream. All data were conditioned and analyzed using custom scripts in R (v3.4.3, http://r-project.org), including a routine designed to partition individual task cycles (“specimens”) based on identification of a local signal minimum.15–17 Protocol randomization was also accomplished using a custom script in R.

RESULTS

DESCRIPTIVE STATISTICS

Participants' demographic parameters were as follows: 9 male/21 female, 39 ± 22 years old (range, 18–87 years), 66 ± 4 inches tall (range, 61–77 inches), 155 ± 41 lb (range, 99–285 lb), 26 right-handed, 3 left-handed, and 1 ambidextrous. A total of 1,797 specimens were recorded, enduring for 1.1 ± 0.5 seconds, with 0.1 ± 0.1 of seconds rest in between cycles of repetition. Total protocol completion time was 12.1 ± 6.7 minutes, averaged across participants. Breakout statistics by task-condition are shown in Table 1.

T1
Table 1:
Number of specimens total per-task per-condition, with values averaged across participants

That fewer than 30 datasets populate a given cell is the result of two circumstances: 1) several datasets were corrupted in the manual transfer from the Arduino serial monitor to the text editor, and thus were not recoverable, or 2) datasets were removed from the analysis stream if containing fewer than five specimens. In filtering for statistical outliers, 101 specimens were removed due to duration, 72 specimens were removed due to outlier values in power ratio, and 29 specimens were removed due to outliers in correlation; all criteria were based on exceeding two standard deviations beyond the study-wide mean.

SENSOR COMPARISON

Across all task conditions (n = 1,797 specimens), the Near and Far sensors averaged a range of 1.3 ± 0.5 and 1.1 ± 0.5 volts, respectively, with a ratiometric value (Near divided by Far) of 1.3 ± 0.3, P < 0.001, paired t-test. Signal power for Near and Far sensors was 2.2 × 102 ± 1.3 × 102 and 1.7 × 102 ± 0.98 × 102 volt-seconds, respectively, with a ratiometric value (Near divided by Far) of 1.3 ± 0.4, P < 0.001, paired t-test (Figure 2). Breakout statistics by task-condition are shown in Table 2.

F2
Figure 2:
Sample sensor readings from Near sensor (lower trace, black) and Far sensor (upper trace, gray) in Gel condition, Fist formation task. Correlation ρ = 0.986 (Specimen 1), ρ = 0.991 (Specimen 3), and ρ = 0.994 (Specimen 5). Traces offset for clarity.
T2
Table 2:
Range (volts units) and power (volt-seconds) for near and far sensor

HYPOTHESIS TESTING

Our primary hypotheses test confirmed the null in both cases: correlation between Near and Far sensor was significantly greater than ρ > 0.98: 0.991 ± 0.019, P < 0.001; and the correlation varied by condition: 0.990 ± 0.03 (Null) versus 0.994 ± 0.01 (Sock) versus 0.989 ± 0.01 (Gel), P < 0.001. Secondary hypothesis testing revealed significant difference in power between conditions: 25% ± 36% increase Area in Near versus Far sensor (Null) versus 32% ± 38% (Sock) and 36% ± 42% (Gel), P < 0.001. There was also a significant difference in signal range between conditions: 25% ± 28% increase in Near versus Far (Null) versus 24% ± 33% (Sock) and 34% ± 30% (Gel). We found no significant difference in signal duration between Near versus Far sensor.

EXPLORATORY

Sensor drift was observed in this study. In a multi-way ANOVA adjusting for intersubject differences and task, intersensor correlation decreased approximately 0.06% per 10 minutes (equivalently: 0.36% per hour). This effect fluctuated across the seven measurements in the hour, and therefore was not statistically significant. However, in a separate analysis looking only at time of donning versus 1 hour, the effect was significant (P < 0.001). Although we did not find intersensor differences, significant differences in range, neither in time-series nor before-after analysis, signal power decreased both for the Near sensor and Far sensor in both time-series and before-after (P < 0.05).

Removing statistical outliers yielded highly similar results as to the full-set analysis. Each cell in Table 2 was to within 10% tolerance between the two analyses and generally to within a 5% tolerance. In the case of correlation coefficients, the values were typically within 2 thousandths of each other.

CLASSIFICATION

A total of 42,741 grasp-versus-grasp pairings yielded high waveform correlations: the average lower-bound of the 95% confidence interval was 0.63 ± 0.07 (range, 0.47–0.75). A total of 21,075 release-versus-grasp pairings were tested for waveform similarity in order to estimate classification error of the FMG system in resting state. A total of 924 (4.6%) of release epochs breached the 95% confidence limit for like-kind correlation among grasps. We show correlation heat maps for two representative datasets in Figure 3.

F3
Figure 3:
Representative correlation matrices hierarchically clustered according to signal similarity. Blue = highly positively correlated; white = not at all correlated; red = highly negatively correlated.

DISCUSSION

MAIN FINDINGS

In descriptive view, we report that healthy persons fitted with commercially available FMG sensors to their flexor digitorum profundus, with the sensors preloaded to low-moderate range, yield a signal value of approximately 0.8 V to 1.6 V. The values within this range are dependent upon the task and interfacial material introduced. We observed moderate (approximately 30%) degradation in signal strength at the Far sensor, measured both by signal amplitude and area under the curve. The Far sensor may require some amplification if there is a desire to reach the same value as the Near sensor. Yet, the correlation between the two sensors is high (ρ > 0.98), suggesting that the signal quality is excellent across the interface.

Exploratory analysis on sensor drift yielded inconsistent results, varying by analytic approach (time-series vs. before-and-after, and across three features). Given the small sample size (n = 2) and the lack of consistent findings, conclusive inference is not supported. Lastly, we report that the main scientific findings of this article do not show sensitivity to the statistical outliers in our dataset. This suggests that both clinical science (application to human body) and device development are mutually well supported in this study.

CLINICAL APPLICATION

Force myography is growing in its perception as an attractive technology for prosthetic control in device development laboratories.1,5,10,18,19 While we are not aware of an incorporation of FMG as primary control technology into commercially available prostheses, we observe that many companies use force-sensitive resistors in their professional demonstration kits, and some devices are switch-controlled by FMG technology.20 Moreover, preliminary study into multifaceted approaches into prosthetic detection suggest that FMG leads to improved signal classification as an adjuvant strategy with EMG.2,8,21 The field is very strongly interested in FMG; we believe that wider adoption of FMG in patient prosthesis control is a certainty.

Regarding our exploratory work in misclassification of resting state as grasp: while our classification error rate was low (<5%), this study was not designed to test classification accuracy as a mainline hypothesis. As such, our analysis was somewhat narrow in scope, and therefore substantial additional work remains to 1) test this more stringently in the laboratory setting, and 2) expand the analysis to include patients in real-world settings. Our approach here provides excellent preliminary insight into the promise of FMG as a potentially robust sensing paradigm, but more declarative statements are simply beyond the scope of this study's design.

LIMITATIONS

Foremost, our study included only limb-intact individuals and did not include any limb-deficient participants. Our recruitment was from a convenience sample. Laudably, our participant pool was relatively large and diverse in demography and body composition, and therefore the findings are more generalizable. Testing among healthy persons has an additional advantage of potentially greater feasibility of replication. On the other hand, we acknowledge that the most relevant population for testing prosthetic technology is among limb-deficient individuals. Nevertheless, because our objective was to test an effect of material interface on the sensor signal, we determined that maximizing the participant pool was the more critical consideration. We note that force-sensing approaches have shown successful application to limb-deficient persons,18,19,22 and that there is no reason to believe that the context for our study, that is, FMG response to interfacial materials, would not transfer to the patient population similarly.

Moreover, several aspects of our interface depart from what would typically be found on a patient. We intentionally tested our participants without sockets, due to the impracticality of producing 30 custom sockets. We note that FMG has been tested previously in the socket environment.2 Furthermore, complete samples of sock and gel liners were not utilized within the protocol. Given the high frequency of donning and doffing, we sought to streamline the experimental protocol, minimizing participant discomfort and maximizing efficiency. Thus, we cut a square swatch from the conical liner and tube-shaped stockinet and discarded the remainder. While we expect that there is no meaningful difference in the material properties of the sock, we recognize that the elastomeric properties of the gel in particular are such that dilatational stress may change the material's durometer: stretched gel is likely stiffer and less pliant. Whether our findings would change if the gel were made more rigid through stretching cannot be known without formal study. We speculate that a more rigid material would transfer forces more directly, and thus the signal at the Far sensor would even be more reflective of that at the Near sensor. Thus we believe that our results may underestimate the performance of FMG in this context.

DESIGN CONSIDERATIONS

Testing of only two materials and only two tasks was a concession of study design meant to reduce burden on the participants and enhance study feasibility over a large number of participants. This reasoning also applies to the decision to randomize sensor order with the Condition block, rather than within the block: once the sensors and liner had been applied, we preferred not to bother the participants with an on-the-fly swapping of the sensors.

The flexor digitorum muscle was selected as the measurement site due to its proximal location and likeliness to be intact within a patient population of persons with transradial amputation. We recorded only a single muscle for the sake of study scope: our study design is already multifactorial; adding an additional muscle would only increase complexity and without clear value. We believe the flexor digitorium is adequately representative of the other muscles available for measurement in this study. Nevertheless, we believe that our findings will extend well into other superficial muscles including application to lower limb.7,23

FMG SENSOR CALIBRATION

The force-sensitive resistor yields an inherently nonlinear response to force. We intentionally tested the raw data without mapping against a calibration curve for the sake of maximum generalizability. The experimental protocol include sensor preload to a reasonable window (5% to 25% sensor range), and most specimens spanned a further 20% to 25% from min to max value. Thus, our data were in a reasonably linear epoch near the middle aspect of the sensor response curve. Indeed, though the Near sensor was approximately 30% more sensitive (higher amplitude), increasing the likelihood of sensor values in disparate regions of the sensor response profile, we nevertheless found very high correlation to the Far sensor. We believe this is further indication of robustness of FMG in the presence of socket lining materials.

FILTERING

It is conventional to process digitized sensor data streaming from a transducer. We decided against filtering in our primary analyses so that any findings of high concordance between Near and Far sensor could not be construed as an artifact due to information loss in the filtering process. In informal review, we found that application of a conventional filter did not appear to alter the outcomes of this study: data were plotted bidirectionally with a first-order, low-pass Butterworth's filter with 10-Hz cutoff. For brevity, we omit the results in this article, but show exemplars (Figure 4).

F4
Figure 4:
Sample sensor readings shown in Figure 2, following application of a low-pass Butterworth filter. Correlation remains high, despite the modest change in signal character.

FUTURE WORK

We encourage others to attempt reproduction our findings, either through complete replication (de novo data collection) or by extending our analysis on our published dataset. We provide a diagrammed study setup as supplemental material to this article (Supplementary Digital Content, https://links.lww.com/JPO/A32). We believe that testing more varieties of inserts (thickness, durometer, etc.) would provide greater insight into the response of FMG to these materials. We believe that testing on a patient population in-socket would add valuable clinical connection and create additional opportunity to establish early best practices in FMG application to clinical prosthetics.

ACKNOWLEDGMENTS

The authors thank the contributions of Melanie Davidson, Laura Faubion, Holly Hendrix, Abby Hoffman-Finitsis, and Sarah Leonard.

REFERENCES

1. Wininger M, Kim NH, Craelius W. Pressure signature of forearm as predictor of grip force. J Rehabil Res Dev 2008;45(6):883–892.
2. Radmand A, Scheme E, Englehart K. High-density force myography: a possible alternative for upper-limb prosthetic control. J Rehabil Res Dev 2016;53(4):443–456.
3. Kadkhodayan A, Jiang X, Menon C. Continuous prediction of finger movements using force myography. J Med Biol Eng 2016;36(4):594–604.
4. Ahmadizadeh C, Merhi LK, Pousett B, et al. Toward intuitive prosthetic control: solving common issues using force myography, surface electromyography, and pattern recognition in a pilot case study. IEEE Robotics & Automation Magazine 2017;24(4):102–111.
5. Castellini C, Artemiadis P, Wininger M, et al. Proceedings of the first workshop on peripheral machine interfaces: going beyond traditional surface electromyography. Front Neurorobot 2014;8:22.
6. Rasouli M, Ghosh R, Lee WW, et al. Stable force-myographic control of a prosthetic hand using incremental learning. Conf Proc IEEE Eng Med Biol Soc 2015;2015:4828–4831.
7. Yungher DA, Wininger MT, Barr JB, et al. Surface muscle pressure as a measure of active and passive behavior of muscles during gait. Med Eng Phys 2011;33(4):464–471.
8. Connan M, Ruiz Ramírez E, Vodermayer B, Castellini C. Assessment of a wearable force- and electromyography device and comparison of the related signals for Myocontrol. Front Neurorobot 2016;10:17.
9. Sakr M, Menon C. Regressing force-myographic signals collected by an armband to estimate torque exerted by the wrist: a preliminary investigation. IEEE 2016;1–4.
10. Sadarangani GP, Jiang X, Simpson LA, et al. Force myography for monitoring grasping in individuals with stroke with mild to moderate upper-extremity impairments: a preliminary investigation in a controlled environment. Front Bioeng Biotechnol 2017;5:42.
11. Li N, Yang D, Jiang L, et al. Combined use of FSR sensor array and SVM classifier for finger motion recognition based on pressure distribution map. J Bionic Eng 2012;9(1):39–47.
12. Xiao ZG, Menon C. Towards the development of a wearable feedback system for monitoring the activities of the upper-extremities. J Neuroeng Rehabil 2014;11(1):2.
13. Dabling JG, Filatov A, Wheeler JW. Static and cyclic performance evaluation of sensors for human interface pressure measurement. Conf Proc IEEE Eng Med Biol Soc 2012;2012:162–165.
14. Lebosse C, Renaud P, Bayle B, de Mathelin M. Modeling and evaluation of low-cost force sensors. IEEE Transactions on Robotics 2011;27(4):815–822.
15. Wininger M, Kim N-H, Craelius W. Spatial resolution of spontaneous accelerations in reaching tasks. J Biomech 2009;42(1):29–34.
16. Wininger M. A generalized strategy for measuring performance in the velocity-position phase plane. J Mot Behav 2012;44(4):269–273.
17. Beed AT, Peduzzi P, Guarino P, Wininger M. A partitioning algorithm for extracting movement epochs from robot-derived kinematic data. Frontiers in Robotics and AI 2017;4.
18. Cho E, Chen R, Merhi L-K, et al. Force myography to control robotic upper extremity prostheses: a feasibility study. Front Bioeng Biotechnol 2016;4:18.
19. Ferigo D, Merhi LK, Pousett B, et al. A case study of a force-myography controlled bionic hand mitigating limb position effect. J Bionic Eng 2017;14(4):692–705.
20. Krajbich JI, Pinzur MS, Potter BK, et al, eds. Pediatrics. 4th Ed. Rosemont, IL: American Academy of Orthopaedic Surgeons; 2016.
21. Sanford J, Patterson R, Popa DO. Concurrent surface electromyography and force myography classification during times of prosthetic socket shift and user fatigue. J Rehabil Assist Technol Eng 2017;4:205566831770873.
22. Phillips SL, Craelius W. Residual kinetic imaging: a versatile interface for prosthetic control. Robotica 2005;23(3):277–282.
23. Jiang X, Chu K, Khoshnam M, Menon C. A wearable gait phase detection system based on force myography techniques. Sensors (Basel) 2018;18(4):E1279.
Keywords:

force myography; FMG; prosthetics; socket; liner; detection; control; upper limb

Supplemental Digital Content

Copyright © 2019 American Academy of Orthotists and Prosthetists