Technological advances and clinical research have shown that body-worn sensors measuring angular velocity (gyroscopes) and/or the acceleration of the trunk can accurately quantify balance during stance and gait tasks [
The sensors used for these purposes must be accurate over different ranges of angular velocity, low velocity ranges (<0.5°/s) for stance tests on a firm surface [
A further method to reduce the costs of balance measuring devices is to use a sensor system measuring only roll and pitch motion thereby ignoring yaw motion (Figure
(a) SwayStar (FOG) mounted on a converted motorcycle belt with a Valedo (MEMS) sensor attached to its side. (b) The SwayStar system mounted on a subject. The SwayStar motion measurement axes (pitch and roll) are, as shown, sensor-based.
In this study, we investigated whether a 2D or 3D MEMS motion sensor could be used as a cheaper lightweight alternative to measuring balance control in the form of angular sway velocity at the lower trunk with accurate FOGs. As small sensors can be placed easily at other locations on the body, an affirmative result would pave the way for the use of such sensors in different body locations and provide the basis for a comprehensive body-mounted motion analysis system. Our primary hypothesis was that a MEMS system would provide a comparable level of accuracy (kappa > 0.8) in classifying normal balance test results as a FOG system. We did not compare the MEMS motion to optical motion capture system because, unlike the 2 systems we compared in this study, motion capture systems are not portable and not quick to start, requiring the attachment of several optical “markers.”
A fiber-optic gyroscope system SwayStar, manufactured by Balance International Innovations GmbH (Switzerland), was used as this is supplied with an extensive healthy control reference database of several clinical stance and gait balance tasks for subjects in the age range of 6 to 80 years [
For the microelectromechanical system (MEMS), one sensor system from the Valedo® products, developed and manufactured by Hocoma AG (Switzerland), was used. The standard application of these sensors is to measure pelvic and spinal movements in order to assess movement parameters and to provide training as part of a physiotherapy plan [
To obtain 3D angular velocity, the received quaternion samples were differentiated with respect to time [
The lower trunk angles measured with the MEMS system were calculated using two methods. The first technique involved applying the 3D Tilt/Twist extraction based on the orientation of the sensor [
From the two sets of sampled sensor data, the following measures were extracted for analysis: peak-to-peak range (difference between maximum and minimum value during the task) and 90% range (difference between 95% and 5% percentile values when the peak-to-peak range of sampled values was divided into 40 bins and a histogram of the task recording samples built after assigning samples to these bins), for both angular velocities and angles in the pitch (sagittal plane) and roll (lateral plane) direction. Therefore, the data extraction yielded the following 8 measures: Peak-to-peak range, roll, angle 90% range, roll, angle Peak-to-peak range, pitch, angle 90% range, pitch, angle Peak-to-peak range, roll, angular velocity 90% range, roll, angular velocity Peak-to-peak range, pitch, angular velocity 90% range, pitch, angular velocity
During the clinical stance and gait tasks, a Valedo (MEMS) sensor was held on the side of the SwayStar (FOG) sensor as shown in Figure
Data of 10 young healthy subjects (8 male, 2 female, age: 19–34 years) were recorded with the FOG and MEMS sensor systems simultaneously. We planned to compare between Valedo and SwayStar sensor measurements for 10 subjects and then if several trends for differences were observed to expand the data set to 20 subjects. As described below, the results showed either statistically significant differences or no differences, with a few trends. Therefore, an expansion of the data set was not considered necessary. The 9 tasks evaluated with both sensor systems are listed below in the order these were performed, that is, in the same order as for the reference database [ Standing on two legs with eyes open, on a normal (firm) surface Standing on two legs with eyes closed, on a foam surface Standing on one leg with eyes closed, on a normal surface Walking 8 tandem steps with eyes open Getting up from a stool and walking 3 meters Walking 3 meters while pitching the head up and down Walking 3 meters with eyes closed Walking up and down a set of stairs (2 steps up and 2 down) Walking 8 meters with eyes open
If subjects were not able to complete a task (due to loss of balance which mostly occurred for the “standing on one leg, eyes closed,” task for which the mean duration for healthy young subjects is 12 sec [
For the data comparison, the differences between the FOG and MEMS measurements were expressed as absolute values as well as percentage values. The 8 extracted measures from both sensor systems were compared with reference data from 88 age- and gender-matched healthy subjects recorded by Hegeman et al. [
Results of 3 of the 9 tasks performed are presented here in detail. These 3 tasks cover the range from low body dynamics (represented by “standing on two legs with eyes open”) to high dynamics (“get up and go, and then walk 3 meters”). All graphs and tables present both the 2D MEMS and 3D MEMS data. In the sections “2D data processing” and “2D vs. 3D data processing,” the comparison between the 2D and 3D angle calculations is described in further detail for all tasks.
Figure
Angular velocity data (a) and angles (b) in the lateral/roll plane (upper), sagittal/pitch plane (middle), and the axial/yaw plane (lower) of the FOG and MEMS sensors for standing on two legs with eyes open on a normal surface task. The red lines depict the 2D FOG data, the green lines the 2D MEMS, and the blue lines the 3D MEMS angle calculations. Note that the velocity traces overlay. The pitch angle 2D traces for FOG and MEMS 2D also overlay. For the yaw angular velocity and yaw angle, only the MEMS data are available.
In Figure
Table
FOG to MEMS comparison results for the task “standing on two legs with eyes open” of all recordings.
Value | PtP Ro A (°) | 90 Ro A (°) | PtP Pi A (°) | 90 Pi A (°) | PtP Ro V (°/s) | 90 Ro V (°/s) | PtP Pi V (°/s) | 90 Pi V (°/s) |
---|---|---|---|---|---|---|---|---|
Mean normal reference | 0.493 | 0.368 | 1.250 | 1.004 | 1.742 | 0.604 | 3.311 | 1.336 |
FOG mean | 0.450 | 0.322 | 1.352 | 1.108 | 1.553 | 0.609 | 2.884 | 1.404 |
FOG SD | 0.364 | 0.258 | 0.481 | 0.412 | 0.759 | 0.261 | 0.980 | 0.544 |
MEMS 2D mean | 0.427 | 0.298 | 1.337 | 1.102 | 1.525 | 0.595 | 2.763 | 1.368 |
MEMS 2D SD | 0.338 | 0.261 | 0.484 | 0.408 | 0.783 | 0.266 | 0.967 | 0.536 |
Error between 2D FOG and 2D MEMS relative to mean normal reference | 4.66% | 6.46% | 1.20% | 0.63% | 1.61% | 2.34% | 3.67% | 2.68% |
|
0.242 | 0.044 |
0.528 | 0.739 | 0.635 | 0.073 | 0.079 | <0.001 |
MEMS 3D mean | 0.420 | 0.307 | 1.299 | 1.068 | ||||
MEMS 3D SD | 0.229 | 0.171 | 0.338 | 0.325 | ||||
Error between 2D FOG and 3D MEMS relative to mean normal reference | 6.16% | 4.03% | 4.26% | 3.97% | ||||
|
0.654 | 0.769 | 0.273 | 0.340 |
PtP: peak-to-peak range, 90 : 90% range (95%–5% percentiles); Ro: roll; Pi: pitch; A: angle in degrees; V: angular velocity in degrees/seconds.
Note that the differences in 3D are only presented in Table
For the other stance tasks, the following was observed: “Standing on two legs with eyes closed, on a foam surface” showed significant differences between FOG and 2D MEMS for both roll and pitch angular velocities (roll:
Figure
Angular velocity data (a) and angles (b) in the roll plane (upper) and pitch plane (middle) and yaw plane (lower) of the FOG and MEMS sensors for the “get up and go 3 meters” task. The red lines illustrate the 2D FOG data, the green lines the 2D MEMS, and the blue lines the 3D MEMS angle calculations. Note that the velocity traces overlay. The 2D FOG and 2D MEMS roll traces overlay. The pitch angle 2D traces for FOG and MEMS also overlay with the 3D MEMS traces. For the yaw angular velocity and yaw angle, only the MEMS data are available.
FOG to MEMS comparison results for the task “get up and go 3 meters.”
Value | PtP Ro A (°) | 90 Ro A (°) | PtP Pi A (°) | 90 Pi | PtP Ro V (°/s) | 90 Ro V (°/s) | PtP Pi V (°/s) | 90 Pi V (°/s) |
---|---|---|---|---|---|---|---|---|
Mean normal reference | 6.451 | 5.201 | 45.95 | 41.90 | 53.78 | 29.61 | 191.7 | 126.5 |
FOG mean | 5.646 | 4.347 | 34.59 | 31.44 | 50.61 | 28.16 | 139.6 | 93.18 |
FOG SD | 1.931 | 1.252 | 5.858 | 5.615 | 23.13 | 9.960 | 33.03 | 27.74 |
MEMS 2D mean | 5.927 | 4.645 | 34.57 | 31.55 | 48.08 | 27.78 | 137.3 | 92.60 |
MEMS 2D SD | 1.966 | 1.428 | 5.876 | 5.599 | 21.78 | 9.768 | 32.48 | 26.97 |
Error between 2D FOG and 2D MEMS relative to mean normal reference | 4.35% | 5.73% | 0.04% | 0.26% | 4.69% | 1.25% | 1.18% | 0.46% |
|
0.077 | 0.024 |
0.614 | 0.167 | 0.278 | 0.260 | <0.001 |
0.343 |
MEMS 3D mean | 6.746 | 5.323 | 34.54 | 31.50 | ||||
MEMS 3D SD | 1.962 | 1.532 | 5.997 | 5.722 | ||||
Error between 2D FOG and 3D MEMS relative to mean normal reference | 17.0% | 18.7% | 0.12% | 0.15% | ||||
|
0.002 |
0.002 |
0.512 | 0.572 |
PtP: peak-to-peak range, 90 : 90% range (95%–5% percentiles); Ro: roll; Pi: pitch; A: angle in degrees; V: angular velocity in degrees/seconds. Note that the differences in 3D are only presented for the angle values; the pitch and roll angular velocities are equal for 2D and 3D.
Table
For the other gait tasks listed below, the differences between the 2D FOG and 3D MEMS roll angles were not significant: Walking 3 meters while pitching the head up and down (not significant (ns) with Walking 3 meters with eyes closed (ns, Walking up and down a set of stairs (ns, Walking 8 meters with eyes open (ns,
For the pitch angles, no significant differences were observed.
Figure
Angular velocity data (a) and angles (b) in the roll plane (upper) and pitch plane (middle) and yaw plane (lower) of the FOG and MEMS sensors for walking 8 tandem steps with eyes open task. The red lines represent the 2D FOG, the green lines the 2D MEMS, and the blue lines the 3D MEMS angle calculations. Note that roll angle traces overlay as do pitch angle FOG and MEMS 2D traces. All velocity traces overlay. The yaw angle is only available for MEMS 3D.
Table
FOG to MEMS comparison results for the task “walking 8 tandem steps with eyes open.”
Value | PtP Ro A (°) | 90 Ro A (°) | PtP Pi A (°) | 90 Pi A (°) | PtP Ro V (°/s) | 90 Ro V (°/s) | PtP Pi V (°/s) | 90 Pi V (°/s) |
---|---|---|---|---|---|---|---|---|
Mean normal reference | 6.324 | 4.714 | 6.920 | 5.160 | 33.86 | 18.49 | 37.92 | 21.03 |
FOG mean | 5.200 | 3.718 | 5.706 | 4.126 | 35.93 | 19.21 | 31.90 | 17.14 |
FOG SD | 1.904 | 1.494 | 1.328 | 1.181 | 12.47 | 5.055 | 5.813 | 3.407 |
MEMS 2D mean | 5.134 | 3.657 | 5.717 | 4.127 | 34.18 | 18.76 | 30.66 | 16.59 |
MEMS 2D SD | 1.919 | 1.524 | 1.336 | 1.189 | 11.46 | 4.963 | 5.076 | 3.156 |
Error between 2D FOG and 2D MEMS relative to mean normal reference | 1.04% | 1.29% | 0.16% | 0.01% | 5.17% | 2.48% | 3.28% | 2.62% |
|
0.202 | 0.155 | 0.675 | 0.984 | 0.051 | <0.001 |
0.216 | 0.003 |
MEMS 3D mean | 5.126 | 3.641 | 5.957 | 4.229 | ||||
MEMS 3D SD | 1.968 | 1.534 | 1.132 | 1.113 | ||||
Error between 2D FOG and 3D MEMS relative to mean normal reference | 1.17% | 1.64% | 3.63% | 1.98% | ||||
|
0.496 | 0.447 | 0.132 | 0.412 |
PtP: peak-to-peak range, 90 : 90% range (95%–5% percentiles); Ro: roll; Pi: pitch; A: angle in degrees; V: angular velocity in degrees/seconds.
For all tasks, except trials were subjects lost their balance control, the eight extracted measures for both the 2D FOG and 2D MEMS were checked for lying within or outside the normal reference range defined by the 95% limits of the reference database. If the data are within the reference 95% range, clinically, the recording would be considered normal [
Because of the loss of balance, six recordings were not taken into account in the analyses (
Table
Contingency table for agreement on values lying within or outside the range of 95% limit of the reference data.
MEMS 2D | FOG 2D | |
Inside range | Outside range | |
|
||
Inside range | 637 | 1 |
Outside range | 1 | 33 |
The single measurement that was inside the range as measured by the MEMS but outside with the FOG was a peak-to-peak value of the angular velocity (no differences were detected for the corresponding 90% range values because single peaks or outliers are filtered out when calculating the 90% range value.). Note that as we compared with 95% reference range values, some values outside the normal range are to be expected.
The 3D angles and angular velocity measures measured with the MEMS and 2D FOG measures were compared with the reference database and the FOG similar to the 2D MEMS comparisons presented in the previous paragraph. For stance tasks that have low ranges, the differences between the 2D and 3D calculations were in the same range as the noise level of the MEMS sensors because the tasks involved limited axial rotation. Thus, divergences in comparison with the reference database were not expected. In contrast, in some recordings of the “get up and go 3 meters” and walking tasks, axial rotation caused a significant “cross-talk” between roll and pitch angles that resulted in a slightly higher number (6) of false-negatives when comparing the angles with the normal reference values. Nonetheless, the Kappa value is 0.868, which is also considered as an almost perfect agreement [
In Figure
(a) Regression of peak-to-peak roll angle FOG vs MEMS 2D (blue circles) and MEMS 3D (red crosses). (b) Peak-to-peak roll angular velocity FOG vs MEMS.
Contingency table for agreement on values lying within or outside the range of 95% limit of reference data.
MEMS 3D | FOG 2D | |
Inside range | Outside range | |
|
||
Inside range | 636 | 6 |
Outside range | 2 | 28 |
In this study, we have tested whether low-cost MEMS motion sensors can provide comparable accuracy as highly accurate fiber-optic gyroscopes to assess balance tasks, which require low noise, minimum drift, and a high resolution across the range of angular sway and sway velocity induced by the balance tasks. We could also assess whether cross-talk errors on pitch and roll angular measures due to not recording yaw angular velocity are significant. If comparable in accuracy and with insignificant cross-talk errors, then MEMS motion sensors can be used to compare extracted balance measures with reference values obtained with highly accurate fiber-optic gyroscopes recording pitch and roll angular velocities. Our main findings were, firstly, that except for the get up and go test, there were no significant differences between 2D FOG and 3D MEMS roll and pitch angle measures. Secondly, angular velocities were slightly underestimated with the MEMS system. Thus, the analyses of the 2D MEMS data showed almost perfect agreement with the FOG data with an interrater classification accuracy of
Statistically significant differences were found between the 3D MEMS roll angles in comparison with the 2D FOG values for the most dynamic gait task “get up and go 3 m,” that is, with the greatest range of pitch angular velocity (over 100 deg/s, Table
Angular velocities measured with the MEMS sensors were obtained by differentiating the processed quaternion output with respect to time. Even if the small differences noted (less than 3% of normal reference values) are not clinically relevant, angular velocities tended to be underestimated by the MEMS. A cause of this difference could be related to mechanical misalignment of the two sensor systems, which is estimated to be around 1 degree [
One of the drawbacks of our study is the limited range of subject ages (19–34 years) we considered. We have compared the accuracy of the two systems in relation to the healthy control reference database of Hegeman et al. [
There are three analytical and clinical areas, which should be considered for future studies. As indicated above, the MEMS system tends to underestimate values of pitch and roll velocities. Thus, the cause of this difference should be examined and established if this underestimate is due to signal processing or sensor alignment. If these causes are ruled out, then attention should be placed on examining patient groups with ataxia that are known to have higher velocity trunk sway during gait trials with tandem steps or eyes closed [
In this study, the MEMS sensors were attached directly onto the FOG system, which was mounted on a converted motorcycle kidney belt. Therefore, both sensor systems measured the same angular movements of the pelvis and lower back. This ensured that movement of the skin during the tasks had no effect in comparing the measurements in the pitch and roll planes between the sensors systems. MEMS sensors can be mounted with double-sided adhesive tape directly on the skin or with an elasticated belt around the waist. These later methods of mounting can cause distinctive soft tissue artefacts compared to the relatively rigid converted motorcycle belt used for the FOG. Additionally, the significant difference in weight between the two sensor systems can influence the effect of soft tissue movements on the outcome measures. For walking tasks, typical roll and pitch soft tissue errors are of the order of 1–2 degrees [
In conclusion, except for tests that involve large yaw movements, there were no significant differences between 2D FOG and 3D MEMS roll and pitch angle measures, although angular velocities were slightly underestimated with the MEMS system. Therefore, 2D MEMS data showed almost perfect agreement with the 2D FOG data. In summary, although for some tasks and some measures statically significant differences were found, further analysis showed that all these differences were within a few percent of the reference values and therefore these differences were assumed not to be clinically relevant. Future studies could consider placing two MEMS sensors side-by-side on a belt, thereby reducing skin artifacts and providing increased accuracy.
The data used to support the findings of this study are available from the corresponding author upon request.
This study was approved by the local ethical committee responsible for the University of Basel Hospital, Ethical Committee of North-Central Switzerland (approval EKNZ 2015-071).
D. Roetenberg and C. Höller are employees of Hocoma AG producing the Valedo inertial sensors used in this study. J. H. J. Allum is a consultant for the company Balance International Innovation GmbH producing the SwayStar equipment used in this study.
This research was partially funded by the Department of Life Sciences, Fachhochschule Nordwestschweiz, Munchenstein, Switzerland, and the Department of ORL, University of Basel Hospital, Switzerland.