- Email: [email protected]
The development and concurrent validity of a real-time algorithm for temporal gait analysis using inertial measurement units
Accepted Manuscript The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units E. Allseits, J. Lučarević, R. Gailey, V. Agrawal, I. Gaunaurd, C. Bennett PII: DOI: Reference:
S0021-9290(17)30112-4 http://dx.doi.org/10.1016/j.jbiomech.2017.02.016 BM 8136
To appear in:
Journal of Biomechanics
Accepted Date:
11 February 2017
Please cite this article as: E. Allseits, J. Lučarević, R. Gailey, V. Agrawal, I. Gaunaurd, C. Bennett, The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units, Journal of Biomechanics (2017), doi: http://dx.doi.org/10.1016/j.jbiomech.2017.02.016
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units E. Allseits b, J. Lučarević a,d, R. Gailey a , V. Agrawal a,b, I. Gaunaurd a,d , C. Bennett c a
University of Miami Miller School of Medicine Department of Physical Therapy, Miami, Florida, USA b
c
University of Miami, Department of Biomedical Engineering, Coral Gables, Florida, USA
University of Miami, Frost School of Music, Music Engineering Technology, Coral Gables, Florida, USA d
Miami Department of Veterans Affairs Healthcare System, Miami, Florida, USA
Corresponding Author: Dr. Christopher Bennett Email: [email protected] Phone: (305) 284-1275 Keywords: Gait Analysis; Inertial Measurement Unit; Temporal gait parameters; Toe-off identification; Real-time Word Count: 3639
1
The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units E. Allseits b, J. Lučarević a,d, R. Gailey a , V. Agrawal a,b, I. Gaunaurd a,d , C. Bennett c a
University of Miami Miller School of Medicine Department of Physical Therapy, Coral Gables, Florida, USA b c
University of Miami, Department of Biomedical Engineering, Coral Gables, Florida, USA
University of Miami, Frost School of Music, Music Engineering Technology, Coral Gables, Florida, USA d
Miami Department of Veterans Affairs Healthcare System, Miami, Florida, USA
_________________________________________________________________________________________ Abstract The use of inertial measurement units (IMUs) for gait analysis has emerged as a tool for clinical applications. Shank gyroscope signals have been utilized to identify heel-strike and toe-off, which serve as the foundation for calculating temporal parameters of gait such as single and double limb support time. Recent publications have shown that toe-off occurs later than predicted by the dual minima method (DMM) suggests, which has been adopted as an IMU-based gait event detection algorithm. In this study, a real-time algorithm, Noise-Zero Crossing (NZC), was developed to accurately compute temporal gait parameters. Our objective was to determine the concurrent validity of temporal gait parameters derived from the NZC algorithm against parameters measured by an instrumented walkway. The accuracy and precision of temporal gait parameters derived using NZC were compared to those derived using the DMM. The results from Bland-Altman Analysis showed that the NZC algorithm had excellent agreement with the instrumented walkway for identifying the temporal gait parameters of Gait Cycle Time (GCT), Single Limb Support (SLS) time, and Double Limb Support (DLS) time. By utilizing the moment of zero shank angular velocity to identify toe-off, the NZC algorithm performed better than the DMM algorithm in measuring SLS and DLS times. Utilizing the NZC algorithm’s gait event detection preserves DLS time which has significant clinical implications for pathologic gait assessment. Keywords: Gait Analysis; Inertial Measurement Unit; Temporal gait parameters; Toe-off identification; Real-time _____________________________________________________________________________________________
1. Background Gait analysis provides information related to diagnosing pathology (Andriacchi et al., 1977), fall risk (Maki, 1997) and mortality (Fritz et al., 2009). Advances in technology, which include optical motion capture systems, instrumented walkways, and forceplates, have become gold standards (Najafi, 2011), but there use is constrained by cost, time, and physical space. Body worn inertial measurement units (IMU) have arisen as a low-cost, portable alternative to these traditional systems and holds the potential for extended, unsupervised data capture. Traditional supervised techniques may influence how subjects walk, provide limited information regarding inter-cycle variability, and capture relatively small numbers of total gait cycles compared to IMU acquisition (Najafi, 2011). Unsupervised data capture requires extended battery life (Najafi, 2011), methods for limiting sensor placement error (Vanegas and Stirling, 2015), and analytic 2
methods for extracting gait parameters from IMU data. A review of the literature shows a fivesensor system attached to shanks, thighs, and sacrum can provide sufficient information for estimation of center of mass (Westerdijk et al., 2012), spatial gait parameters (Aminian et al., 2002; Li et al., 2009; Sijobert et al., 2015), and joint angle dynamics (Seel et al., 2014), as well as temporal gait parameters (Aminian et al., 2002). A Dual Minima Method (DMM) of detecting heel strike (HS) and toe off (TO) gait events using shank angular velocity is a common method of gait event detection that fits with a such a sensor configuration (Aminian et al., 2002; Greene et al., 2010; Lee and Park, 2011; Salarian et al., 2004; Tong and Granat, 1999). Aminian et al. (2002) originally observed the coincidence of the two minima of the shank angular velocity signal with HS and TO events measured by footswitches placed on the heel and first metatarsal. Accuracy between the IMU and footswitch systems was assessed using RMSE of gait event timing and regression of temporal gait parameters such as gait cycle time (GCT), single limb support time (SLS), and double limb support time (DLS). Subsequently, numerous algorithms have been developed to identify these minima (Aminian et al., 2002; Fraccaro et al., 2014; Greene et al., 2010; Lee and Park, 2011) and observe various pathologic populations (Jasiewicz et al., 2006; Salarian et al., 2004). Recent studies have called into question the validity of this DMM (Bötzel et al., 2016; Greene et al., 2010). In a review of the literature, Bötzel et al. found that while algorithms employing the DMM accurately identified HS, they all identified TO earlier than their criterion measures (Bötzel et al., 2016). In their own analysis using a 200Hz IMU system and an optical motion capture system and pressure sensitive insoles, an average TO discrepancy of {84,67,52} ms at gait velocities of {2,4,6} km/h, respectively was found. They ascribe the inconsistency of the previous results to the variable filtering methods and low filtering cutoffs (~3-5 Hz) employed in 3
some of the earlier studies and point to recent studies that corroborate their findings in both the shank (Greene et al., 2010) and foot (Mariani et al., 2013) angular velocities. While these studies agree with using the first minima for HS, they found that TO occurs later in the gait cycle than the DMM predicts. Currently, no comprehensive analysis has been performed on the impact of the inherent TO detection error of the DMM on clinically relevant gait parameters such as GCT, SLS, and DLS. Likewise, no alternate method for accurate detection of TO has been published. Based on the findings of Bötzel et al., this study proposes a novel algorithm for gait phase detection from the pattern of shank angular velocity using a Noise-Zero Crossing method and compares it to Lee and Park’s implementation of the established DMM. The purpose of this study is to: 1) determine the concurrent validity of temporal gait parameters derived from IMU data using a novel gait phase detection algorithm, and 2) determine whether the proposed algorithm demonstrates greater agreement with the criterion measure than the established DMM.
2. Methods 2.1 Description of Shank Angular Velocity Waveform Gyroscope-derived shank angular velocity during gait has a well-defined, quasi-periodic waveform that consists of a region of positive angular velocity corresponding to advancing the shank relative to the knee and a region of negative angular velocity corresponding to advancing the shank relative to the ground (Fig. 1). In the middle of the negative angular velocity region exists a maximum corresponding to MidStance (MSt) (Hanlon et al., 2009) and in the middle of the positive angular velocity region exists a maximum corresponding to MidSwing (MSw).
4
The DMM posits that the two troughs of the shank angular velocity signal can be used to identify HS and TO. Recent findings have quantified the inaccuracy of the DMM TO identification in the context of the laboratory setting, while validating its HS identification by analyzing each step using a post-processing routine and very high IMU data acquisition sampling rates (Bötzel et al., 2016). Similarly, the reviewed methods require complex filtering techniques to reliably identify HS and TO gait events for the computation of temporal gait parameters. We propose an algorithm using a Noise-Zero Crossing (NZC) method and assess its ability to estimate temporal gait parameters from data acquired under the constraints of Bluetooth Low Energy (BLE).
2.2 Noise-Zero Crossing Algorithm NZC is based in the inverted pendulum model of gait (Aminian et al., 2002; Laudanski et al., 2011), which posits that the leg alternates between advancing as a pendulum during swing phase (pivot at the hip) and as an inverted pendulum during stance phase (pivot at the foot). Under this model, HS is identified by the noise spike in the trough following MSw (HS_NZC) and TO is identified by the moment of zero velocity following the trough before MSw (TO_NZC) as in Figure 1.
2.3 Heel-Strike (NZC) The trough after MSw has been associated with a region of increased noise that has been observed across studies involving healthy and pathologic populations (Aminian et al., 2002; Bötzel et al., 2016; Greene et al., 2010; Salarian et al., 2004) in which wavelet transforms and
5
filtering techniques were used to reduce noise to reliably identify the trough. NZC proposes that the universality of this noise can be exploited to identify HS without filtering.
2.4 Toe-Off (NZC) Bötzel et al. note that TO occurs after the second trough in the shank angular velocity signal, but prior to forward velocity. They explain their findings with the physiologic argument that the minimum of the second trough physically represents the point at which the shank begins to decelerate. They argue that TO should occur sometime after that deceleration begins, but before the limb begins to advance about the opposite pivot, represented by the inception of positive angular velocity. We posit that TO can thus be estimated as the zero-crossing after MSt, which is the precise moment of zero shank angular velocity.
2.5 Noise-Zero Crossing Algorithm Implementation The NZC algorithm is outlined in Fig. 2. NZC waits for the first zero-crossing after the onset of walking. If the slope is negative, the algorithm waits for the onset of noise, defined as the first sample where the slope of the signal becomes positive, checks whether that sample is the minimum swing period from the last transition point (τୗ , and stores that point as a swing-tostance heel-strike event (SWtoST). If the slope is positive, the algorithm checks whether the current sample is at least the minimum stance period from the last transition point (τୗ ) and stores this point as a stance to swing toe-off event (STtoSW). Transition points consist of STtoSW and SWtoST events. Minimum swing and stance periods were defined a priori using knowledge of gait phase knowledge and manually verified for accuracy. For each sample, gait
6
phase is assigned as stance if the most recent transition point was SWtoST or swing if it was STtoSW. GCT, SLS, and DLS time intervals are computed using the gait phase from both limbs.
2.6 The Dual Minima Method (Lee and Park) Lee and Park (2011) developed an algorithm (LP) that exploits the lag introduced by low pass filtering to identify HS and TO according to the DMM. It iteratively identifies windows on the each side of the MSw peak in which the minima putatively corresponding to HS and TO lie (Lee and Park, 2011). Their algorithm was chosen for analysis because, like NZC, it is compatible with real-time gait analysis.
2.7 Subjects The study protocol was reviewed and approved by the Institutional Review Board at University of Miami. Written informed consent was obtained from all subjects prior to enrollment. Eligible subjects were 18-69 years old and healthy, without neuromusculoskeletal limitations or gait deviation as observed by a licensed physical therapist. All subjects wore standardized footwear (NewBalance420).
2.8 Data Acquisition IMU data were collected using a wireless system developed to transmit data to an iPad Air 2 via BLE and process unconstrained gait kinematics in real-time utilizing a stable, synchronized, 5 Multi-Axial Profile Recorder (MAPR) IMU configuration operating at 50 Hz. MAPRs were attached to thighs and shanks as described by Kim et al. (2016a) with an additional sacral sensor, as described by Kim et al. (2016b). A hand-held button that was synchronized with the MAPR system was used to flag interval times. To reduce errors inherent in sensor placement 7
variability (Vanegas and Stirling, 2015) and slippage, placement was standardized using a virtual alignment algorithm and soft-tissue artifacts (Peters et al., 2010) minimized by securing sensors to an elastic open-patella knee brace (Össur America) with sewn-in straps (2.54 cm wide) in anatomic positions as by Kim et al. (2016a). Data from an alignment trial were processed using a Gauss Newton minimizer (Seel et al., 2012) to virtually align the sensors with the knee joint axis (Fig. 3), which is preferable to manual alignment because it standardizes sensor placement across subjects regardless of anatomical differences while allowing sensors to be attached at a location with minimal soft tissue artifact.
2.9 Instrumented Walkway (Criterion Measure) Vertical ground reaction force (GRF) was captured on a custom-built Matscan System (2.57m long, sensel resolution 0.8382 cm2, Type 3150L sensors; Tekscan) (Kim et al, 2016b) positioned in the center of a 12.25 m walkway with a sampling rate of 50 Hz. The MatScan was chosen as the criterion measure because of its high accuracy (Giacomozzi, 2010), and moderate to good reliability (ICC .44-.95) for measuring plantar pressures during gait (Zammit et al., 2010; Hafer et al., 2013), and ability to capture multiple steps per pass. Temporal parameters captured by the MatScan walkway system were manually determined from force-time curves with HS and TO as the points when the force-time curve passed above and below the noise floor, respectively (Adderson et al., 2007; Mizrahi and Daily, 2012; Zeni et al., 2008). GCT was defined as the duration from HS to subsequent HS of the same foot. HS and TO delimited stance time for each foot, with SLS as the duration when only one limb was in stance and DLS when both limbs were in stance.
8
2.10 Data Collection With the MAPR system donned, subjects performed an alignment trial consisting of a sitstand activity followed by 10 steps. Subjects then walked at a self-selected walking speed (SSWS) in alternating directions along a walkway through three demarcated zones; acceleration and deceleration zones of 2.3 m and a central constant velocity zone of 7.65 m (Bohannon, 1997) with the MatScan placed in the center. Intervals were flagged on the MAPR system when the subject’s trunk broke the plane of each zone. To introduce variability in the recorded measures, subjects were instructed to walk at a perceived 50% of their SSWS (Slow). Each subject performed successive trials until they had ten passes with at least one complete stride recorded by the MatScan for each speed condition.
2.11 Data Analysis IMU data captured outside the constant velocity zone was discarded. Since literature agrees that the DMM is valid for HS, NZC accuracy in HS detection was assessed through stride-bystride concurrence comparison to LP. Agreement of temporal gait parameters between each algorithm and the criterion measure was examined using Bland-Altman analysis (Bland and Altman, 1999; Cole et al., 2014; Hanlon and Anderson, 2009). Reported reproducibility coefficient (RPC) was calculated as 1.96 times standard deviation. Concurrence of HS and accuracy of SLS and DLS times provides sufficient information to determine the accuracy of TO estimation, as shown in Equations 1-2. ்ை ுௌ ௌ௧
(1)
ௌ௧ 2∆ௌ ∆ௌௌ (2) Kendall’s tau was computed to determine heteroscedasticity with a cutoff of τ >0.1 (Brehm et al., 2012). Although IMU systems are indirect, noisy measures of gait phenomena, they capture 9
more steps per pass than traditional gait analysis techniques. To account for the discrepancy in observation number between the MAPR and MatScan systems, the average value per pass of each temporal gait parameter observed by each system was used for analysis. The strength of IMU systems is their ability to collect clean data from large numbers of steps, while traditional techniques are unable to capture more than a handful of consecutive strides per “clean pass” and can require multiple passes before a “clean pass” is captured. IMU-based systems are less precise than traditional techniques because they are noisy and indirect measures of gait. This noise can be addressed using windowing, averaging, and outlier-rejection techniques to obtain more accurate “representative” estimates of temporal gait parameters without altering the mean estimate of the parameters. Thus, the average accuracy of NZC and LP, as opposed to the strideby-stride precision, is a better method for comparing each IMU system to the instrumented walkway. Coefficient of variation (CoV) analysis was performed on temporal gait parameters derived from SSWS trials for each patient and Wilcoxon sign-ranked test was used to determine significant differences between each IMU algorithm and the instrumented walkway. CoV was defined as the standard deviation over the mean scaled by 100 (Hausdorff et al., 1998; Riva et al., 2014). All data analysis was completed in MATLAB 2015a (MathWorks Inc, Natick, MA).
3. Results 3.1 Subjects and Data Acquisition Eleven subjects participated in the study (5 females) with a mean age of 32.4±6.9yrs (20.345.9yrs), height of 173.4±11.4cm, and weight of 75.9±23.2kg. The IMU system captured roughly 6 times as many measurements of GCT and SLS and 3 times as many measurements of DLS as the MatScan (Table 1). 10
3.2 Temporal Gait Parameters CoV was calculated for each temporal gait parameter across subjects under the SSWS condition (Table 2). The Wilcoxon sign-ranked test found no significant difference between NZC and the criterion measure for GCT and DLS. Both NZC and LP were found to have a significant difference from the criterion measure for the CoV of SLS. NZC showed good concurrence with LP for HS event timing with a mean difference of -1.5 samples (~30ms) and 72% of gait events within ±1 sample (±20ms) of simultaneity across gait speeds (Table 3). Bland Altman analysis demonstrated the validity of gait intervals of SLS, DLS, and GCT for each algorithm in comparison to the instrumented walkway (Fig. 4; Table 4). Performance of NZC and LP was almost identical for GCT and had the highest agreement between the algorithms and the criterion measure (mean difference ~10ms). The NZC algorithm had excellent agreement with the criterion measure for SLS and DLS, as noted by the mean difference near zero and narrow limits of agreement (LoA). The LP algorithm had poor agreement with the criterion measure for SLS (mean difference 170ms; LoA [50ms,280ms]) and DLS (mean difference 170ms; LoA [-280ms,-66ms]). Kendall’s tau showed that all Bland-Altman analyses were homoscedastic except for DLS via LP, where τ =.57, p<0.001.
4. Discussion The results show that while both NZC and LP yield accurate and precise estimates of GCT, only NZC accurately estimates SLS and DLS times. NZC identifies HS concurrently with LP, validating our earlier hypothesis that mean difference in SLS and DLS times from the instrumented walkway reflects the accuracy of TO event detection. Using Bland-Altman results and Eq. 2, LP identifies TO 170 ms early, while NZC identifies TO only 27 ms (~1 sample) late, 11
demonstrating that the zero-crossing can be accurately used to practically identify TO. Likewise, the underestimation of DLS by LP (Table 2) supports the findings of Bötzel et al. that the DMM identifies TO early.
4.1 Temporal Gait Parameters Regression, as well as correlation coefficients, coefficients of determination, and comparison of means, have been shown to be inappropriate statistical methods for measuring concurrent validity between two systems of measurement. Nonetheless, in a systematic review by Zaki et al. (2012), 10% of reviewed articles only utilized these inappropriate statistical tests for assessing agreement between medical devices. Bland-Altman analysis, however, was developed to overcome the limitations of regression analysis (Bland and Altman, 1999). While many studies have used regression to show concurrent validity of GCT, stance, and swing time (Aminian et al, 2002; Aminian et al, 2004;Trojaniello et al., 2014), a distinct absence of regression or BlandAltman plots for sub-stance gait phases has been noted. Bland-Altman analysis demonstrated that while NZC showed good concurrence with the instrumented walkway for determining DLS and SLS times, LP shows poor concurrence. NZC was accurate within roughly one sample for DLS (-23 ms) and SLS (19 ms) and half a sample for GCT (9.7 ms) (Fig. 4). The LoA for NZC are within ± 4 samples (55 ms, 73 ms, 81 ms) for DLS, SLS, and GCT. However, GCT calculated from HS using both LP and NZC were determined with equal accuracy and precision (Fig. 4). Since GCT depends only on the identification of a single gait event type (HS), GCT CoV functions as a measure of regularity of that identification process (Marcar et al., 2014). Likewise, since SLS and DLS are delimited by both HS and TO, they represent the regularity of HS and TO. While GCT CoV from the NZC algorithm was not found to be significantly different than 12
that of the instrumented walkway, the GCT CoV of the LP algorithm was significantly larger. This suggests that the HS identification process of LP may be less reproducible and stable than that of NZC and may explain why 28% of HS events identified by the two IMU algorithms were more than 1 sample different. This difference was greater in the Slow condition (40%) than SSWS (12.5%). The increased CoV and LoA of SLS and DLS (Fig. 4) provide further evidence that LP, not NZC, had issues reproducibly identifying gait events. The increased variance in the Slow condition compared to the SSWS condition for LP corroborates that this effect was larger at slow speeds. Manual inspection of NZC and LP identification on the angular velocity trace showed that LP sometimes had difficulty identifying the center of the trough because of noise peaks, a problem which was exacerbated at slower speeds.
4.3 Advantages and Future Work Bötzel et al. (2016) argue that the prominent studies supporting the DMM suffered from flawed methodological and signal processing techniques, which resulted in the misidentification of the TO landmark in the shank gyroscope signal. This study supports those findings and further 1) demonstrates the adverse impact these findings have on the calculation of clinically relevant temporal gait parameters from IMUs, 2) proposes the NZC algorithm as a suitable replacement to calculate these parameters in real-time, and 3) validates the accuracy and precision of NZC when compared to an instrumented walkway. Since NZC operates on the unfiltered signal, we may have avoided the pitfalls of previous studies suggested by Bötzel et al. Also, unlike previous studies, our methodology standardized sensor placement across subjects and minimized gaitphase dependent soft-tissue artifact using a virtual alignment protocol.
13
While NZC is as accurate as the instrumented walkway, it does not suffer from the limitations of traditional gait analysis techniques. It can be operated without a clinician present, it minimizes subject fatigue by not requiring multiple passes to get an acceptable or “clean” pass, it is portable and user friendly, it does not require the extensive post-processing time of the instrumented mat (~8 hours per subject), and it captures ~4 times the number of observations per pass (Table 1), which may account for the relatively low CoV, for example in SLS times. Studies critiquing the lack of standardization in gait variability analysis have demonstrated that hundreds of consecutive steps are required to stably measure gait cycle variability using IMU measurements (Riva et al., 2014) or instrumented walkways (Hollman et al., 2010). While Hollman et al note that this kind of data collection is onerous and nearly impossible in the clinical setting using traditional techniques, this study demonstrates the feasibility of collecting and processing vast amounts of accurate temporal gait parameters in real-time. This has a broad impact on future work because accurate estimation of temporal gait parameters is a prerequisite for developing models for the calculation of more complex spatial parameters. Future studies, however, should compare variability measures across measurement systems using the sample sizes recommended by Hollman et al and Riva et al. Since the growing consensus is that the DMM is inaccurate, this study suggests that current methods of spatial gait parameter calculation developed using the DMM should be reevaluated using temporal estimates derived via NZC.
5. Conclusion The NZC method, although an indirect measure of temporal gait parameters, has been shown to be concurrently valid with the direct measure instrumented walkway. NZC produces stable and repeatable measures of gait event landmarks that yield accurate and precise estimates of gait
14
phase intervals that can be acquired in the subject’s everyday life. The ability of NZC to accurately detect GCT variability and DLS interval times allows for clinicians to utilize an IMU system for gait assessment. Additionally, since temporal gait parameters are the foundation for computing higher-level gait features, the accuracy of NZC over the DMM implies a need for the reevaluation of current techniques for the calculation of spatial gait parameters such as stride length, step length, and gait deviation.
Conflict of Interest All authors declare that there is no proprietary, financial, professional or other personal interest of any nature or kind in any product, service or company that could be construed as influencing the position presented in this manuscript.
Acknowledgments The Authors would like to thank the Joint Incentive Fund for supporting this work under the DoD-VA Mobile Device Outcomes-based Rehabilitation Program.
References Adderson, J., Parker, K., Macleod, D., Kirby, R., McPhail, C., 2007. Effect of a shock-absorbing pylon on transmission of heel strike forces during the gait of people with unilateral trans-tibial amputations: a pilot study. Prosthet. Orthot. Int. 31, 384–393. doi:10.1080/03093640701254018 Aminian, K., Najafi, B., Büla, C., Leyvraz, P.F., Robert, P., 2002. Spatio-temporal parameters of gait measured by an ambulatory system using miniature gyroscopes. J. Biomech. 35, 689-699. Aminian, K., Trevisan, C., Najafi, B., Dejnabadi, H., Frigo, C., Pavan, E., Telonio, A., Cerati, F., Marinoni, E., Robert, P., 2004. Evaluation of an ambulatory system for gait analysis in hip osteoarthritis and after total hip replacement. Gait Posture 20, 102-107. Andriacchi, T., Ogle, J., Galante, J., 1977. Walking speed as a basis for normal and abnormal gait measurements. J. Biomech. 10, 261-268. 15
Barker, S., Craik, R., Freedman, W., Herrmann, N., Hillstrom, H., 2006. Accuracy, reliability, and validity of a spatiotemporal gait analysis system. Med. Eng. Phys. 28, 460-467. Bland, J.M., Altman, D.G., 1999. Measuring agreement in method comparison studies. Stat. Methods in Med. Res. 8, 135-160. Bohannon, R.W., 1997. Comfortable and maximum walking speed of adults aged 20—79 years: reference values and determinants. Age Ageing 26, 15-19. Bötzel, K., Marti, F.M., Rodríguez, M.Á.C., Plate, A., Vicente, A.O., 2016. Gait recording with inertial sensors - How to Determine Initial and Terminal Contact. J. Biomech. 49, 332–337. Brehm, M.A., Scholtes, V.A., Dallmeijer, A.J., Twisk, J.W., Harlaar, J., 2012. The importance of addressing heteroscedasticity in the reliability analysis of ratio-scaled variables: An example based on walking energy-cost measurements. Dev. Med. Child Neurol. 54, 267–273. Cole, M.H., Van Den Hoorn, W., Kavanagh, J.K., Morrison, S., Hodges, P.W., Smeathers, J.E., Kerr, G.K., 2014. Concurrent validity of accelerations measured using a tri-axial inertial measurement unit while walking on firm, compliant and uneven surfaces. PLoS One 9, 1–12. doi:10.1371/journal.pone.0098395 Fraccaro, P., Walsh, L., Doyle, J., O’Sullivan, D., 2014. Real-world Gyroscope-based Gait Event Detection and Gait Feature Extraction. eTELEMED 2014, Sixth Int. Conf. eHealth, Telemedicine, Soc. Med. 247–252. Fritz, S., Lusardi, M., 2009. White paper: “Walking speed: The sixth vital sign”. J. Geriatr. Phys. Ther. 32, 2-5. Giacomozzi, C., 2010. Appropriateness of plantar pressure measurement devices: A comparative technical assessment. Gait Posture 32, 141-144. Greene, B.R., McGrath, D., O’Neill, R., O’Donovan, K.J., Burns, A., Caulfield, B., 2010. An adaptive gyroscope-based algorithm for temporal gait analysis. Med. Biol. Eng. Comput. 48, 1251-1260. Hafer, J.F., Lenhoff, M.W., Song, J., Jordan, J.M., Hannan, M.T., Hillstrom, H.J., 2013. Reliability of plantar pressure platforms. Gait Posture 38, 544-548. Hanlon, M., Anderson, R., 2009. Real-time gait event detection using wearable sensors. Gait Posture 30, 523-527. Hausdorff, J.M., Cudkowicz, M.E., Firtion, R., Wei, J.Y., Goldberger, A.L., 1998. Gait variability and basal ganglia disorders: Stride-to-stride variations of gait cycle timing in Parkinson’s disease and Huntington's disease. Mov. Disord. 13, 428–437. Hollman, J.H., Childs, K.B., McNeil, M.L., Mueller, A.C., Quilter, C.M., Youdas, J.W., 2010. Number of strides required for reliable measurements of pace, rhythm and variability parameters of gait during normal and dual task walking in older individuals. Gait Posture 32, 23–28. doi:10.1016/j.gaitpost.2010.02.017 Hordacre, B.G., Barr, C., Patritti, B.L., Crotty, M., 2015. Assessing Gait Variability in Transtibial Amputee Fallers Based on Spatial-Temporal Gait Parameters Normalized for Walking Speed. Arch. Phys. Med. Rehabil. 96, 1162–1165. Jasiewicz, J.M., Allum, J.H.J., Middleton, J.W., Barriskill, A., Condie, P., Purcell, B., Li, R.C.T., 2006. Gait event detection using linear accelerometers or angular velocity transducers in able-bodied and spinal-cord injured individuals. Gait Posture 24, 502–509. doi:10.1016/j.gaitpost.2005.12.017 Kim, K.J., Agrawal, V., Gaunaurd, I., Gailey, R., Bennett, C., 2016a. Missing Sample Recovery for Wireless Inertial Sensor-Based Human Movement Acquisition. IEEE Trans. Neural Syst. Rehabil. Eng. 4320, 1–1. Kim, K.J., Lucarevic, J., Bennett, C., Gaunaurd, I., Gailey, R., Agrawal, V., 2016b. Testing the assumption of normality in body sway area calculations during unipedal stance tests with an inertial sensor. 2016 IEEE EMBC. Lee, J.K., Park, E.J., 2011. Quasi real-time gait event detection using shank-attached gyroscopes. Med. Biol. Eng. Comput. 49, 707-712.
16
Li, Q., Young, M., Naing, V., Donelan, J.M., 2009. Walking speed and slope estimation using shankmounted inertial measurement units. 2009 IEEE Int. Conf. Rehabil. Robot. ICORR 2009 839–844. Maki, B.E., 1997. Gait changes in older adults: predictors of falls or indicators of fear? J. Am. Geriatr. Soc. 45, 313-320. Marcar, V.L., Bridenbaugh, S.A., Kool, J., Niedermann, K., Kressig, R.W., 2014. A simple procedure to synchronize concurrent measurements of gait and brain electrical activity and preliminary results from a pilot measurement involving motor-cognitive dual-tasking in healthy older and young volunteers. J. Neurosci. Methods 228, 46–49. doi:10.1016/j.jneumeth.2014.03.003 Mariani, B., Rouhani, H., Crevoisier, X., Aminian, K., 2013. Quantitative estimation of foot-flat and stance phase of gait using foot-worn inertial sensors. Gait Posture 37, 229–234. doi:10.1016/j.gaitpost.2012.07.012 Mizrahi, J., Daily, D., 2012. Modeling the Foot-Strike Event in Running Fatigue via Mechanical Impedances. Inj. Skelet. Biomech. 153–170. Najafi, B., 2011. Laboratory in a Box: Wearable Sensors and Its Advantages for Gait Analysis, IEEE EMBS. pp. 6507–6510. Peters, A., Galna, B., Sangeux, M., Morris, M., Baker, R., 2010. Quantification of soft tissue artifact in lower limb human motion analysis: a systematic review. Gait Posture 31, 1-8. Riva, F., Bisi, M.C., Stagni, R., 2014. Gait variability and stability measures: Minimum number of strides and within-session reliability. Comput. Biol. Med. 50, 9–13. Salarian, A., Russmann, H., Vingerhoets, F.J.G., Dehollain, C., Blanc, Y., Burkhard, P.R., Aminian, K., 2004. Gait assessment in Parkinson's disease: toward an ambulatory system for long-term monitoring. Biomedical Engineering, IEEE Trans. on Biomed. Eng. 51, 1434-1443. Seel, T., Schauer, T., 2012. Joint Axis and Position Estimation from Inertial Measurement Data by Exploiting Kinematic Constraints, IEEE. Intl. Conf. Contr. pp. 0–4. Seel, T., Raisch, J., Schauer, T., 2014. IMU-based joint angle measurement for gait analysis. Sensors (Basel). 14, 6891–909. doi:10.3390/s140406891 Sijobert, B., Benoussaad, M., Denys, J., Pissard-Gibollet, R., Geny, C., Azevedo Coste, C., 2015. Implementation and Validation of a Stride Length Estimation Algorithm, Using a Single Basic Inertial Sensor on Healthy Subjects and Patients Suffering from Parkinson’s Disease. Health (Irvine. Calif). 7, 704–714. doi:10.4236/health.2015.76084 Tong, K., Granat, M.H., 1999. A practical gait analysis system using gyroscopes. Med. Eng. Phys. 21, 87–94. doi:10.1016/S1350-4533(99)00030-2 Trojaniello, D., Cereatti, a, Mori, L., Pelosin, E., Della Croce, U., 2014. Stride-by-stride gait spatiotemporal parameters estimation from shank-mounted magneto-inertial sensors Application to healthy and hemiparetic gait. XIII Int. Symp. 3D Anal. Hum. Mov. 11–14. Vanegas, M., Stirling, L., 2015. Characterization of inertial measurement unit placement on the human body upon repeated donnings, IEEE Intl. Conf. BSN, Cambridge, MA, pp. 1-6. Westerdijk, M.J., Schepers, H.M., Veltink, P.H., Asseldonk, E.H.F. Van, Buurke, J.H., 2012. Use of Inertial Sensors for Ambulatory Assessment of Center-of-Mass Displacements During Walking. IEEE Trans. Biomed. Eng. 59, 2080–2084. Zaki, R., Bulgiba, A., Ismail, R., Ismail, N.A., 2012. Statistical methods used to test for agreement of medical instruments measuring continuous variables in method comparison studies: A systematic review. PLoS One 7, 1–7. Zammit, G. V, Menz, H.B., Munteanu, S.E., 2010. Reliability of the TekScan MatScan(R) system for the measurement of plantar forces and pressures during barefoot level walking in healthy adults. J. Foot Ankle Res. 3, 11. Zeni, J.A., Richards, J.G., Higginson, J.S., 2008. Two simple methods for determining gait events during treadmill and overground walking using kinematic data. Gait Posture 27, 710–714.
17
Figures and Tables Figure 1. Comparison of Gait Phase events detected using two different gait event detection methods a) Represented by red lines DMM = Dual Minima Method, b) Represented by blue lines NZC = Noise-Zero Crossing method; HS = Heel-Strike, TO = Toe-Off Figure 2. Flowchart of the Noise-Zero Crossing (NZC) Algorithm for calculating gait phase Figure 3. MAPR Alignment Algorithm virtually aligns the orientation of the thigh and shank IMUs. Figure 4. Bland-Altman Plots assessing the performance of the LP and NZC algorithms when compared to the MatScan for each temporal gait parameter. Each point represents the results from a single trial. Red lines represent performance of the LP algorithm, and blue lines represent the NZC algorithm, + = Slow, □= SSWS. Mean Difference is represented by the solid line, and the limits of agreement are represented by the dotted line. GCT=gait cycle time, SLS=single limb support time, DLS= double limb support time.
Table 1. Comparison of Observation Counts of Temporal Gait Parameters Between the MatScan and MAPR Systems Table 2. Average Coefficient of Variation for each measurement system during Self Selected Walking Speed. Table 3. Stride-by-stride analysis of HS concurrence between the NZC and LP algorithms in number of samples (sample = 20ms) for Slow and SSWS conditions, as well as cumulative analysis. Mean difference is average difference between observed HS identification times. RMSE is root mean square error between the observed HS identification times. % ±1 sample is the percent of strides where NZC and LP identified HS within one sample of each other. Table 4. Temporal Gait Analysis Performance Comparison of NZC and LP Algorithms: Agreement with the Criterion Measure. RPC= reproducibility coefficient, LoA = limits of agreement, CV=coefficient of variation.
18
Figguree 1
19
Figure 2
20
Figure 3
21
Figure 4
22
Table 1 Slow Temporal Gait Parameters
SSWS
Cumulative
MatScan
MAPR
MatScan
MAPR
MatScan
MAPR
GCT
213
1169
125
876
338
2045
SLS
232
1169
136
876
368
2045
DLS
338
1169
243
876
581
2045
GCT = Gait Cycle Time, SLS = Single Limb Support Time, DLS = Double Limb Support Time,
Table 2 Coefficient of Variation (%) Temporal Gait Parameters
p values
MatScan
LP
NZC
MatScan vs LP
MatScan vs NZC
GCT
3.50 ± 0.74
4.60 ± 1.40
3.73 ± 1.12
0.0418*
0.6458
SLS
9.20 ± 1.47
6.84 ± 1.95
6.81 ± 1.95
0.0071**
0.0071**
DLS
15.0 ± 3.67
58.33 ± 27.01
16.2 ± 6.66
<0.0001**
0.8955
GCT = Gait Cycle Time, SLS = Single Limb Support Time, DLS = Double Limb Support Time, LP=Lee&Park, NZC = Noise Zero Crossing * = p<0.05; ** = p<.01
Table 3 Slow
SSWS
Cumulative
Mean Difference (samples)
-2.71
-.06
-1.54
RMSE (samples)
5.15
2.29
4.14
59.7%
87.5%
72.0%
±1 Sample (percent)
RMSE=root mean square error, 1 sample = 20ms
23
Table 4 Mean
LoA
RPC
CV
Kendall’s τ
LP
-11
[-93, 71]
82 (6.2%)
3.1%
0.064 (p = 0.17)
NZC
-9.7
[-91, 71]
81 (6.1%)
3.1%
-0.12 (p = 0.0068)
LP
170
[50, 280]
120 (24%)
12%
-0.61 (p < 0.001)
NZC
19
[-54, 91]
73 (18%)
9%
-0.21 (p < 0.001)
LP
-170
[-280, -66]
110 (58%)
30%
0.57 (p < 0.001)
NZC
-23
[-78, .32]
55 (22%)
11%
-0.017 (p = 0.71)
GCT
SLS
DLS GCT = Gait Cycle Time, SLS = Single Limb Support Time, DLS = Double Limb Support Time, LP=Lee & Park, NZC = Noise Zero Crossing, LoA= Bland-Altman Analysis limits of agreement, RPC= Reproducibility Coefficient, CV= coefficient of variation
24
S0021-9290(17)30112-4 http://dx.doi.org/10.1016/j.jbiomech.2017.02.016 BM 8136
To appear in:
Journal of Biomechanics
Accepted Date:
11 February 2017
Please cite this article as: E. Allseits, J. Lučarević, R. Gailey, V. Agrawal, I. Gaunaurd, C. Bennett, The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units, Journal of Biomechanics (2017), doi: http://dx.doi.org/10.1016/j.jbiomech.2017.02.016
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units E. Allseits b, J. Lučarević a,d, R. Gailey a , V. Agrawal a,b, I. Gaunaurd a,d , C. Bennett c a
University of Miami Miller School of Medicine Department of Physical Therapy, Miami, Florida, USA b
c
University of Miami, Department of Biomedical Engineering, Coral Gables, Florida, USA
University of Miami, Frost School of Music, Music Engineering Technology, Coral Gables, Florida, USA d
Miami Department of Veterans Affairs Healthcare System, Miami, Florida, USA
Corresponding Author: Dr. Christopher Bennett Email: [email protected] Phone: (305) 284-1275 Keywords: Gait Analysis; Inertial Measurement Unit; Temporal gait parameters; Toe-off identification; Real-time Word Count: 3639
1
The Development and Concurrent Validity of a Real-time Algorithm for Temporal Gait Analysis using Inertial Measurement Units E. Allseits b, J. Lučarević a,d, R. Gailey a , V. Agrawal a,b, I. Gaunaurd a,d , C. Bennett c a
University of Miami Miller School of Medicine Department of Physical Therapy, Coral Gables, Florida, USA b c
University of Miami, Department of Biomedical Engineering, Coral Gables, Florida, USA
University of Miami, Frost School of Music, Music Engineering Technology, Coral Gables, Florida, USA d
Miami Department of Veterans Affairs Healthcare System, Miami, Florida, USA
_________________________________________________________________________________________ Abstract The use of inertial measurement units (IMUs) for gait analysis has emerged as a tool for clinical applications. Shank gyroscope signals have been utilized to identify heel-strike and toe-off, which serve as the foundation for calculating temporal parameters of gait such as single and double limb support time. Recent publications have shown that toe-off occurs later than predicted by the dual minima method (DMM) suggests, which has been adopted as an IMU-based gait event detection algorithm. In this study, a real-time algorithm, Noise-Zero Crossing (NZC), was developed to accurately compute temporal gait parameters. Our objective was to determine the concurrent validity of temporal gait parameters derived from the NZC algorithm against parameters measured by an instrumented walkway. The accuracy and precision of temporal gait parameters derived using NZC were compared to those derived using the DMM. The results from Bland-Altman Analysis showed that the NZC algorithm had excellent agreement with the instrumented walkway for identifying the temporal gait parameters of Gait Cycle Time (GCT), Single Limb Support (SLS) time, and Double Limb Support (DLS) time. By utilizing the moment of zero shank angular velocity to identify toe-off, the NZC algorithm performed better than the DMM algorithm in measuring SLS and DLS times. Utilizing the NZC algorithm’s gait event detection preserves DLS time which has significant clinical implications for pathologic gait assessment. Keywords: Gait Analysis; Inertial Measurement Unit; Temporal gait parameters; Toe-off identification; Real-time _____________________________________________________________________________________________
1. Background Gait analysis provides information related to diagnosing pathology (Andriacchi et al., 1977), fall risk (Maki, 1997) and mortality (Fritz et al., 2009). Advances in technology, which include optical motion capture systems, instrumented walkways, and forceplates, have become gold standards (Najafi, 2011), but there use is constrained by cost, time, and physical space. Body worn inertial measurement units (IMU) have arisen as a low-cost, portable alternative to these traditional systems and holds the potential for extended, unsupervised data capture. Traditional supervised techniques may influence how subjects walk, provide limited information regarding inter-cycle variability, and capture relatively small numbers of total gait cycles compared to IMU acquisition (Najafi, 2011). Unsupervised data capture requires extended battery life (Najafi, 2011), methods for limiting sensor placement error (Vanegas and Stirling, 2015), and analytic 2
methods for extracting gait parameters from IMU data. A review of the literature shows a fivesensor system attached to shanks, thighs, and sacrum can provide sufficient information for estimation of center of mass (Westerdijk et al., 2012), spatial gait parameters (Aminian et al., 2002; Li et al., 2009; Sijobert et al., 2015), and joint angle dynamics (Seel et al., 2014), as well as temporal gait parameters (Aminian et al., 2002). A Dual Minima Method (DMM) of detecting heel strike (HS) and toe off (TO) gait events using shank angular velocity is a common method of gait event detection that fits with a such a sensor configuration (Aminian et al., 2002; Greene et al., 2010; Lee and Park, 2011; Salarian et al., 2004; Tong and Granat, 1999). Aminian et al. (2002) originally observed the coincidence of the two minima of the shank angular velocity signal with HS and TO events measured by footswitches placed on the heel and first metatarsal. Accuracy between the IMU and footswitch systems was assessed using RMSE of gait event timing and regression of temporal gait parameters such as gait cycle time (GCT), single limb support time (SLS), and double limb support time (DLS). Subsequently, numerous algorithms have been developed to identify these minima (Aminian et al., 2002; Fraccaro et al., 2014; Greene et al., 2010; Lee and Park, 2011) and observe various pathologic populations (Jasiewicz et al., 2006; Salarian et al., 2004). Recent studies have called into question the validity of this DMM (Bötzel et al., 2016; Greene et al., 2010). In a review of the literature, Bötzel et al. found that while algorithms employing the DMM accurately identified HS, they all identified TO earlier than their criterion measures (Bötzel et al., 2016). In their own analysis using a 200Hz IMU system and an optical motion capture system and pressure sensitive insoles, an average TO discrepancy of {84,67,52} ms at gait velocities of {2,4,6} km/h, respectively was found. They ascribe the inconsistency of the previous results to the variable filtering methods and low filtering cutoffs (~3-5 Hz) employed in 3
some of the earlier studies and point to recent studies that corroborate their findings in both the shank (Greene et al., 2010) and foot (Mariani et al., 2013) angular velocities. While these studies agree with using the first minima for HS, they found that TO occurs later in the gait cycle than the DMM predicts. Currently, no comprehensive analysis has been performed on the impact of the inherent TO detection error of the DMM on clinically relevant gait parameters such as GCT, SLS, and DLS. Likewise, no alternate method for accurate detection of TO has been published. Based on the findings of Bötzel et al., this study proposes a novel algorithm for gait phase detection from the pattern of shank angular velocity using a Noise-Zero Crossing method and compares it to Lee and Park’s implementation of the established DMM. The purpose of this study is to: 1) determine the concurrent validity of temporal gait parameters derived from IMU data using a novel gait phase detection algorithm, and 2) determine whether the proposed algorithm demonstrates greater agreement with the criterion measure than the established DMM.
2. Methods 2.1 Description of Shank Angular Velocity Waveform Gyroscope-derived shank angular velocity during gait has a well-defined, quasi-periodic waveform that consists of a region of positive angular velocity corresponding to advancing the shank relative to the knee and a region of negative angular velocity corresponding to advancing the shank relative to the ground (Fig. 1). In the middle of the negative angular velocity region exists a maximum corresponding to MidStance (MSt) (Hanlon et al., 2009) and in the middle of the positive angular velocity region exists a maximum corresponding to MidSwing (MSw).
4
The DMM posits that the two troughs of the shank angular velocity signal can be used to identify HS and TO. Recent findings have quantified the inaccuracy of the DMM TO identification in the context of the laboratory setting, while validating its HS identification by analyzing each step using a post-processing routine and very high IMU data acquisition sampling rates (Bötzel et al., 2016). Similarly, the reviewed methods require complex filtering techniques to reliably identify HS and TO gait events for the computation of temporal gait parameters. We propose an algorithm using a Noise-Zero Crossing (NZC) method and assess its ability to estimate temporal gait parameters from data acquired under the constraints of Bluetooth Low Energy (BLE).
2.2 Noise-Zero Crossing Algorithm NZC is based in the inverted pendulum model of gait (Aminian et al., 2002; Laudanski et al., 2011), which posits that the leg alternates between advancing as a pendulum during swing phase (pivot at the hip) and as an inverted pendulum during stance phase (pivot at the foot). Under this model, HS is identified by the noise spike in the trough following MSw (HS_NZC) and TO is identified by the moment of zero velocity following the trough before MSw (TO_NZC) as in Figure 1.
2.3 Heel-Strike (NZC) The trough after MSw has been associated with a region of increased noise that has been observed across studies involving healthy and pathologic populations (Aminian et al., 2002; Bötzel et al., 2016; Greene et al., 2010; Salarian et al., 2004) in which wavelet transforms and
5
filtering techniques were used to reduce noise to reliably identify the trough. NZC proposes that the universality of this noise can be exploited to identify HS without filtering.
2.4 Toe-Off (NZC) Bötzel et al. note that TO occurs after the second trough in the shank angular velocity signal, but prior to forward velocity. They explain their findings with the physiologic argument that the minimum of the second trough physically represents the point at which the shank begins to decelerate. They argue that TO should occur sometime after that deceleration begins, but before the limb begins to advance about the opposite pivot, represented by the inception of positive angular velocity. We posit that TO can thus be estimated as the zero-crossing after MSt, which is the precise moment of zero shank angular velocity.
2.5 Noise-Zero Crossing Algorithm Implementation The NZC algorithm is outlined in Fig. 2. NZC waits for the first zero-crossing after the onset of walking. If the slope is negative, the algorithm waits for the onset of noise, defined as the first sample where the slope of the signal becomes positive, checks whether that sample is the minimum swing period from the last transition point (τୗ , and stores that point as a swing-tostance heel-strike event (SWtoST). If the slope is positive, the algorithm checks whether the current sample is at least the minimum stance period from the last transition point (τୗ ) and stores this point as a stance to swing toe-off event (STtoSW). Transition points consist of STtoSW and SWtoST events. Minimum swing and stance periods were defined a priori using knowledge of gait phase knowledge and manually verified for accuracy. For each sample, gait
6
phase is assigned as stance if the most recent transition point was SWtoST or swing if it was STtoSW. GCT, SLS, and DLS time intervals are computed using the gait phase from both limbs.
2.6 The Dual Minima Method (Lee and Park) Lee and Park (2011) developed an algorithm (LP) that exploits the lag introduced by low pass filtering to identify HS and TO according to the DMM. It iteratively identifies windows on the each side of the MSw peak in which the minima putatively corresponding to HS and TO lie (Lee and Park, 2011). Their algorithm was chosen for analysis because, like NZC, it is compatible with real-time gait analysis.
2.7 Subjects The study protocol was reviewed and approved by the Institutional Review Board at University of Miami. Written informed consent was obtained from all subjects prior to enrollment. Eligible subjects were 18-69 years old and healthy, without neuromusculoskeletal limitations or gait deviation as observed by a licensed physical therapist. All subjects wore standardized footwear (NewBalance420).
2.8 Data Acquisition IMU data were collected using a wireless system developed to transmit data to an iPad Air 2 via BLE and process unconstrained gait kinematics in real-time utilizing a stable, synchronized, 5 Multi-Axial Profile Recorder (MAPR) IMU configuration operating at 50 Hz. MAPRs were attached to thighs and shanks as described by Kim et al. (2016a) with an additional sacral sensor, as described by Kim et al. (2016b). A hand-held button that was synchronized with the MAPR system was used to flag interval times. To reduce errors inherent in sensor placement 7
variability (Vanegas and Stirling, 2015) and slippage, placement was standardized using a virtual alignment algorithm and soft-tissue artifacts (Peters et al., 2010) minimized by securing sensors to an elastic open-patella knee brace (Össur America) with sewn-in straps (2.54 cm wide) in anatomic positions as by Kim et al. (2016a). Data from an alignment trial were processed using a Gauss Newton minimizer (Seel et al., 2012) to virtually align the sensors with the knee joint axis (Fig. 3), which is preferable to manual alignment because it standardizes sensor placement across subjects regardless of anatomical differences while allowing sensors to be attached at a location with minimal soft tissue artifact.
2.9 Instrumented Walkway (Criterion Measure) Vertical ground reaction force (GRF) was captured on a custom-built Matscan System (2.57m long, sensel resolution 0.8382 cm2, Type 3150L sensors; Tekscan) (Kim et al, 2016b) positioned in the center of a 12.25 m walkway with a sampling rate of 50 Hz. The MatScan was chosen as the criterion measure because of its high accuracy (Giacomozzi, 2010), and moderate to good reliability (ICC .44-.95) for measuring plantar pressures during gait (Zammit et al., 2010; Hafer et al., 2013), and ability to capture multiple steps per pass. Temporal parameters captured by the MatScan walkway system were manually determined from force-time curves with HS and TO as the points when the force-time curve passed above and below the noise floor, respectively (Adderson et al., 2007; Mizrahi and Daily, 2012; Zeni et al., 2008). GCT was defined as the duration from HS to subsequent HS of the same foot. HS and TO delimited stance time for each foot, with SLS as the duration when only one limb was in stance and DLS when both limbs were in stance.
8
2.10 Data Collection With the MAPR system donned, subjects performed an alignment trial consisting of a sitstand activity followed by 10 steps. Subjects then walked at a self-selected walking speed (SSWS) in alternating directions along a walkway through three demarcated zones; acceleration and deceleration zones of 2.3 m and a central constant velocity zone of 7.65 m (Bohannon, 1997) with the MatScan placed in the center. Intervals were flagged on the MAPR system when the subject’s trunk broke the plane of each zone. To introduce variability in the recorded measures, subjects were instructed to walk at a perceived 50% of their SSWS (Slow). Each subject performed successive trials until they had ten passes with at least one complete stride recorded by the MatScan for each speed condition.
2.11 Data Analysis IMU data captured outside the constant velocity zone was discarded. Since literature agrees that the DMM is valid for HS, NZC accuracy in HS detection was assessed through stride-bystride concurrence comparison to LP. Agreement of temporal gait parameters between each algorithm and the criterion measure was examined using Bland-Altman analysis (Bland and Altman, 1999; Cole et al., 2014; Hanlon and Anderson, 2009). Reported reproducibility coefficient (RPC) was calculated as 1.96 times standard deviation. Concurrence of HS and accuracy of SLS and DLS times provides sufficient information to determine the accuracy of TO estimation, as shown in Equations 1-2. ்ை ுௌ ௌ௧
(1)
ௌ௧ 2∆ௌ ∆ௌௌ (2) Kendall’s tau was computed to determine heteroscedasticity with a cutoff of τ >0.1 (Brehm et al., 2012). Although IMU systems are indirect, noisy measures of gait phenomena, they capture 9
more steps per pass than traditional gait analysis techniques. To account for the discrepancy in observation number between the MAPR and MatScan systems, the average value per pass of each temporal gait parameter observed by each system was used for analysis. The strength of IMU systems is their ability to collect clean data from large numbers of steps, while traditional techniques are unable to capture more than a handful of consecutive strides per “clean pass” and can require multiple passes before a “clean pass” is captured. IMU-based systems are less precise than traditional techniques because they are noisy and indirect measures of gait. This noise can be addressed using windowing, averaging, and outlier-rejection techniques to obtain more accurate “representative” estimates of temporal gait parameters without altering the mean estimate of the parameters. Thus, the average accuracy of NZC and LP, as opposed to the strideby-stride precision, is a better method for comparing each IMU system to the instrumented walkway. Coefficient of variation (CoV) analysis was performed on temporal gait parameters derived from SSWS trials for each patient and Wilcoxon sign-ranked test was used to determine significant differences between each IMU algorithm and the instrumented walkway. CoV was defined as the standard deviation over the mean scaled by 100 (Hausdorff et al., 1998; Riva et al., 2014). All data analysis was completed in MATLAB 2015a (MathWorks Inc, Natick, MA).
3. Results 3.1 Subjects and Data Acquisition Eleven subjects participated in the study (5 females) with a mean age of 32.4±6.9yrs (20.345.9yrs), height of 173.4±11.4cm, and weight of 75.9±23.2kg. The IMU system captured roughly 6 times as many measurements of GCT and SLS and 3 times as many measurements of DLS as the MatScan (Table 1). 10
3.2 Temporal Gait Parameters CoV was calculated for each temporal gait parameter across subjects under the SSWS condition (Table 2). The Wilcoxon sign-ranked test found no significant difference between NZC and the criterion measure for GCT and DLS. Both NZC and LP were found to have a significant difference from the criterion measure for the CoV of SLS. NZC showed good concurrence with LP for HS event timing with a mean difference of -1.5 samples (~30ms) and 72% of gait events within ±1 sample (±20ms) of simultaneity across gait speeds (Table 3). Bland Altman analysis demonstrated the validity of gait intervals of SLS, DLS, and GCT for each algorithm in comparison to the instrumented walkway (Fig. 4; Table 4). Performance of NZC and LP was almost identical for GCT and had the highest agreement between the algorithms and the criterion measure (mean difference ~10ms). The NZC algorithm had excellent agreement with the criterion measure for SLS and DLS, as noted by the mean difference near zero and narrow limits of agreement (LoA). The LP algorithm had poor agreement with the criterion measure for SLS (mean difference 170ms; LoA [50ms,280ms]) and DLS (mean difference 170ms; LoA [-280ms,-66ms]). Kendall’s tau showed that all Bland-Altman analyses were homoscedastic except for DLS via LP, where τ =.57, p<0.001.
4. Discussion The results show that while both NZC and LP yield accurate and precise estimates of GCT, only NZC accurately estimates SLS and DLS times. NZC identifies HS concurrently with LP, validating our earlier hypothesis that mean difference in SLS and DLS times from the instrumented walkway reflects the accuracy of TO event detection. Using Bland-Altman results and Eq. 2, LP identifies TO 170 ms early, while NZC identifies TO only 27 ms (~1 sample) late, 11
demonstrating that the zero-crossing can be accurately used to practically identify TO. Likewise, the underestimation of DLS by LP (Table 2) supports the findings of Bötzel et al. that the DMM identifies TO early.
4.1 Temporal Gait Parameters Regression, as well as correlation coefficients, coefficients of determination, and comparison of means, have been shown to be inappropriate statistical methods for measuring concurrent validity between two systems of measurement. Nonetheless, in a systematic review by Zaki et al. (2012), 10% of reviewed articles only utilized these inappropriate statistical tests for assessing agreement between medical devices. Bland-Altman analysis, however, was developed to overcome the limitations of regression analysis (Bland and Altman, 1999). While many studies have used regression to show concurrent validity of GCT, stance, and swing time (Aminian et al, 2002; Aminian et al, 2004;Trojaniello et al., 2014), a distinct absence of regression or BlandAltman plots for sub-stance gait phases has been noted. Bland-Altman analysis demonstrated that while NZC showed good concurrence with the instrumented walkway for determining DLS and SLS times, LP shows poor concurrence. NZC was accurate within roughly one sample for DLS (-23 ms) and SLS (19 ms) and half a sample for GCT (9.7 ms) (Fig. 4). The LoA for NZC are within ± 4 samples (55 ms, 73 ms, 81 ms) for DLS, SLS, and GCT. However, GCT calculated from HS using both LP and NZC were determined with equal accuracy and precision (Fig. 4). Since GCT depends only on the identification of a single gait event type (HS), GCT CoV functions as a measure of regularity of that identification process (Marcar et al., 2014). Likewise, since SLS and DLS are delimited by both HS and TO, they represent the regularity of HS and TO. While GCT CoV from the NZC algorithm was not found to be significantly different than 12
that of the instrumented walkway, the GCT CoV of the LP algorithm was significantly larger. This suggests that the HS identification process of LP may be less reproducible and stable than that of NZC and may explain why 28% of HS events identified by the two IMU algorithms were more than 1 sample different. This difference was greater in the Slow condition (40%) than SSWS (12.5%). The increased CoV and LoA of SLS and DLS (Fig. 4) provide further evidence that LP, not NZC, had issues reproducibly identifying gait events. The increased variance in the Slow condition compared to the SSWS condition for LP corroborates that this effect was larger at slow speeds. Manual inspection of NZC and LP identification on the angular velocity trace showed that LP sometimes had difficulty identifying the center of the trough because of noise peaks, a problem which was exacerbated at slower speeds.
4.3 Advantages and Future Work Bötzel et al. (2016) argue that the prominent studies supporting the DMM suffered from flawed methodological and signal processing techniques, which resulted in the misidentification of the TO landmark in the shank gyroscope signal. This study supports those findings and further 1) demonstrates the adverse impact these findings have on the calculation of clinically relevant temporal gait parameters from IMUs, 2) proposes the NZC algorithm as a suitable replacement to calculate these parameters in real-time, and 3) validates the accuracy and precision of NZC when compared to an instrumented walkway. Since NZC operates on the unfiltered signal, we may have avoided the pitfalls of previous studies suggested by Bötzel et al. Also, unlike previous studies, our methodology standardized sensor placement across subjects and minimized gaitphase dependent soft-tissue artifact using a virtual alignment protocol.
13
While NZC is as accurate as the instrumented walkway, it does not suffer from the limitations of traditional gait analysis techniques. It can be operated without a clinician present, it minimizes subject fatigue by not requiring multiple passes to get an acceptable or “clean” pass, it is portable and user friendly, it does not require the extensive post-processing time of the instrumented mat (~8 hours per subject), and it captures ~4 times the number of observations per pass (Table 1), which may account for the relatively low CoV, for example in SLS times. Studies critiquing the lack of standardization in gait variability analysis have demonstrated that hundreds of consecutive steps are required to stably measure gait cycle variability using IMU measurements (Riva et al., 2014) or instrumented walkways (Hollman et al., 2010). While Hollman et al note that this kind of data collection is onerous and nearly impossible in the clinical setting using traditional techniques, this study demonstrates the feasibility of collecting and processing vast amounts of accurate temporal gait parameters in real-time. This has a broad impact on future work because accurate estimation of temporal gait parameters is a prerequisite for developing models for the calculation of more complex spatial parameters. Future studies, however, should compare variability measures across measurement systems using the sample sizes recommended by Hollman et al and Riva et al. Since the growing consensus is that the DMM is inaccurate, this study suggests that current methods of spatial gait parameter calculation developed using the DMM should be reevaluated using temporal estimates derived via NZC.
5. Conclusion The NZC method, although an indirect measure of temporal gait parameters, has been shown to be concurrently valid with the direct measure instrumented walkway. NZC produces stable and repeatable measures of gait event landmarks that yield accurate and precise estimates of gait
14
phase intervals that can be acquired in the subject’s everyday life. The ability of NZC to accurately detect GCT variability and DLS interval times allows for clinicians to utilize an IMU system for gait assessment. Additionally, since temporal gait parameters are the foundation for computing higher-level gait features, the accuracy of NZC over the DMM implies a need for the reevaluation of current techniques for the calculation of spatial gait parameters such as stride length, step length, and gait deviation.
Conflict of Interest All authors declare that there is no proprietary, financial, professional or other personal interest of any nature or kind in any product, service or company that could be construed as influencing the position presented in this manuscript.
Acknowledgments The Authors would like to thank the Joint Incentive Fund for supporting this work under the DoD-VA Mobile Device Outcomes-based Rehabilitation Program.
References Adderson, J., Parker, K., Macleod, D., Kirby, R., McPhail, C., 2007. Effect of a shock-absorbing pylon on transmission of heel strike forces during the gait of people with unilateral trans-tibial amputations: a pilot study. Prosthet. Orthot. Int. 31, 384–393. doi:10.1080/03093640701254018 Aminian, K., Najafi, B., Büla, C., Leyvraz, P.F., Robert, P., 2002. Spatio-temporal parameters of gait measured by an ambulatory system using miniature gyroscopes. J. Biomech. 35, 689-699. Aminian, K., Trevisan, C., Najafi, B., Dejnabadi, H., Frigo, C., Pavan, E., Telonio, A., Cerati, F., Marinoni, E., Robert, P., 2004. Evaluation of an ambulatory system for gait analysis in hip osteoarthritis and after total hip replacement. Gait Posture 20, 102-107. Andriacchi, T., Ogle, J., Galante, J., 1977. Walking speed as a basis for normal and abnormal gait measurements. J. Biomech. 10, 261-268. 15
Barker, S., Craik, R., Freedman, W., Herrmann, N., Hillstrom, H., 2006. Accuracy, reliability, and validity of a spatiotemporal gait analysis system. Med. Eng. Phys. 28, 460-467. Bland, J.M., Altman, D.G., 1999. Measuring agreement in method comparison studies. Stat. Methods in Med. Res. 8, 135-160. Bohannon, R.W., 1997. Comfortable and maximum walking speed of adults aged 20—79 years: reference values and determinants. Age Ageing 26, 15-19. Bötzel, K., Marti, F.M., Rodríguez, M.Á.C., Plate, A., Vicente, A.O., 2016. Gait recording with inertial sensors - How to Determine Initial and Terminal Contact. J. Biomech. 49, 332–337. Brehm, M.A., Scholtes, V.A., Dallmeijer, A.J., Twisk, J.W., Harlaar, J., 2012. The importance of addressing heteroscedasticity in the reliability analysis of ratio-scaled variables: An example based on walking energy-cost measurements. Dev. Med. Child Neurol. 54, 267–273. Cole, M.H., Van Den Hoorn, W., Kavanagh, J.K., Morrison, S., Hodges, P.W., Smeathers, J.E., Kerr, G.K., 2014. Concurrent validity of accelerations measured using a tri-axial inertial measurement unit while walking on firm, compliant and uneven surfaces. PLoS One 9, 1–12. doi:10.1371/journal.pone.0098395 Fraccaro, P., Walsh, L., Doyle, J., O’Sullivan, D., 2014. Real-world Gyroscope-based Gait Event Detection and Gait Feature Extraction. eTELEMED 2014, Sixth Int. Conf. eHealth, Telemedicine, Soc. Med. 247–252. Fritz, S., Lusardi, M., 2009. White paper: “Walking speed: The sixth vital sign”. J. Geriatr. Phys. Ther. 32, 2-5. Giacomozzi, C., 2010. Appropriateness of plantar pressure measurement devices: A comparative technical assessment. Gait Posture 32, 141-144. Greene, B.R., McGrath, D., O’Neill, R., O’Donovan, K.J., Burns, A., Caulfield, B., 2010. An adaptive gyroscope-based algorithm for temporal gait analysis. Med. Biol. Eng. Comput. 48, 1251-1260. Hafer, J.F., Lenhoff, M.W., Song, J., Jordan, J.M., Hannan, M.T., Hillstrom, H.J., 2013. Reliability of plantar pressure platforms. Gait Posture 38, 544-548. Hanlon, M., Anderson, R., 2009. Real-time gait event detection using wearable sensors. Gait Posture 30, 523-527. Hausdorff, J.M., Cudkowicz, M.E., Firtion, R., Wei, J.Y., Goldberger, A.L., 1998. Gait variability and basal ganglia disorders: Stride-to-stride variations of gait cycle timing in Parkinson’s disease and Huntington's disease. Mov. Disord. 13, 428–437. Hollman, J.H., Childs, K.B., McNeil, M.L., Mueller, A.C., Quilter, C.M., Youdas, J.W., 2010. Number of strides required for reliable measurements of pace, rhythm and variability parameters of gait during normal and dual task walking in older individuals. Gait Posture 32, 23–28. doi:10.1016/j.gaitpost.2010.02.017 Hordacre, B.G., Barr, C., Patritti, B.L., Crotty, M., 2015. Assessing Gait Variability in Transtibial Amputee Fallers Based on Spatial-Temporal Gait Parameters Normalized for Walking Speed. Arch. Phys. Med. Rehabil. 96, 1162–1165. Jasiewicz, J.M., Allum, J.H.J., Middleton, J.W., Barriskill, A., Condie, P., Purcell, B., Li, R.C.T., 2006. Gait event detection using linear accelerometers or angular velocity transducers in able-bodied and spinal-cord injured individuals. Gait Posture 24, 502–509. doi:10.1016/j.gaitpost.2005.12.017 Kim, K.J., Agrawal, V., Gaunaurd, I., Gailey, R., Bennett, C., 2016a. Missing Sample Recovery for Wireless Inertial Sensor-Based Human Movement Acquisition. IEEE Trans. Neural Syst. Rehabil. Eng. 4320, 1–1. Kim, K.J., Lucarevic, J., Bennett, C., Gaunaurd, I., Gailey, R., Agrawal, V., 2016b. Testing the assumption of normality in body sway area calculations during unipedal stance tests with an inertial sensor. 2016 IEEE EMBC. Lee, J.K., Park, E.J., 2011. Quasi real-time gait event detection using shank-attached gyroscopes. Med. Biol. Eng. Comput. 49, 707-712.
16
Li, Q., Young, M., Naing, V., Donelan, J.M., 2009. Walking speed and slope estimation using shankmounted inertial measurement units. 2009 IEEE Int. Conf. Rehabil. Robot. ICORR 2009 839–844. Maki, B.E., 1997. Gait changes in older adults: predictors of falls or indicators of fear? J. Am. Geriatr. Soc. 45, 313-320. Marcar, V.L., Bridenbaugh, S.A., Kool, J., Niedermann, K., Kressig, R.W., 2014. A simple procedure to synchronize concurrent measurements of gait and brain electrical activity and preliminary results from a pilot measurement involving motor-cognitive dual-tasking in healthy older and young volunteers. J. Neurosci. Methods 228, 46–49. doi:10.1016/j.jneumeth.2014.03.003 Mariani, B., Rouhani, H., Crevoisier, X., Aminian, K., 2013. Quantitative estimation of foot-flat and stance phase of gait using foot-worn inertial sensors. Gait Posture 37, 229–234. doi:10.1016/j.gaitpost.2012.07.012 Mizrahi, J., Daily, D., 2012. Modeling the Foot-Strike Event in Running Fatigue via Mechanical Impedances. Inj. Skelet. Biomech. 153–170. Najafi, B., 2011. Laboratory in a Box: Wearable Sensors and Its Advantages for Gait Analysis, IEEE EMBS. pp. 6507–6510. Peters, A., Galna, B., Sangeux, M., Morris, M., Baker, R., 2010. Quantification of soft tissue artifact in lower limb human motion analysis: a systematic review. Gait Posture 31, 1-8. Riva, F., Bisi, M.C., Stagni, R., 2014. Gait variability and stability measures: Minimum number of strides and within-session reliability. Comput. Biol. Med. 50, 9–13. Salarian, A., Russmann, H., Vingerhoets, F.J.G., Dehollain, C., Blanc, Y., Burkhard, P.R., Aminian, K., 2004. Gait assessment in Parkinson's disease: toward an ambulatory system for long-term monitoring. Biomedical Engineering, IEEE Trans. on Biomed. Eng. 51, 1434-1443. Seel, T., Schauer, T., 2012. Joint Axis and Position Estimation from Inertial Measurement Data by Exploiting Kinematic Constraints, IEEE. Intl. Conf. Contr. pp. 0–4. Seel, T., Raisch, J., Schauer, T., 2014. IMU-based joint angle measurement for gait analysis. Sensors (Basel). 14, 6891–909. doi:10.3390/s140406891 Sijobert, B., Benoussaad, M., Denys, J., Pissard-Gibollet, R., Geny, C., Azevedo Coste, C., 2015. Implementation and Validation of a Stride Length Estimation Algorithm, Using a Single Basic Inertial Sensor on Healthy Subjects and Patients Suffering from Parkinson’s Disease. Health (Irvine. Calif). 7, 704–714. doi:10.4236/health.2015.76084 Tong, K., Granat, M.H., 1999. A practical gait analysis system using gyroscopes. Med. Eng. Phys. 21, 87–94. doi:10.1016/S1350-4533(99)00030-2 Trojaniello, D., Cereatti, a, Mori, L., Pelosin, E., Della Croce, U., 2014. Stride-by-stride gait spatiotemporal parameters estimation from shank-mounted magneto-inertial sensors Application to healthy and hemiparetic gait. XIII Int. Symp. 3D Anal. Hum. Mov. 11–14. Vanegas, M., Stirling, L., 2015. Characterization of inertial measurement unit placement on the human body upon repeated donnings, IEEE Intl. Conf. BSN, Cambridge, MA, pp. 1-6. Westerdijk, M.J., Schepers, H.M., Veltink, P.H., Asseldonk, E.H.F. Van, Buurke, J.H., 2012. Use of Inertial Sensors for Ambulatory Assessment of Center-of-Mass Displacements During Walking. IEEE Trans. Biomed. Eng. 59, 2080–2084. Zaki, R., Bulgiba, A., Ismail, R., Ismail, N.A., 2012. Statistical methods used to test for agreement of medical instruments measuring continuous variables in method comparison studies: A systematic review. PLoS One 7, 1–7. Zammit, G. V, Menz, H.B., Munteanu, S.E., 2010. Reliability of the TekScan MatScan(R) system for the measurement of plantar forces and pressures during barefoot level walking in healthy adults. J. Foot Ankle Res. 3, 11. Zeni, J.A., Richards, J.G., Higginson, J.S., 2008. Two simple methods for determining gait events during treadmill and overground walking using kinematic data. Gait Posture 27, 710–714.
17
Figures and Tables Figure 1. Comparison of Gait Phase events detected using two different gait event detection methods a) Represented by red lines DMM = Dual Minima Method, b) Represented by blue lines NZC = Noise-Zero Crossing method; HS = Heel-Strike, TO = Toe-Off Figure 2. Flowchart of the Noise-Zero Crossing (NZC) Algorithm for calculating gait phase Figure 3. MAPR Alignment Algorithm virtually aligns the orientation of the thigh and shank IMUs. Figure 4. Bland-Altman Plots assessing the performance of the LP and NZC algorithms when compared to the MatScan for each temporal gait parameter. Each point represents the results from a single trial. Red lines represent performance of the LP algorithm, and blue lines represent the NZC algorithm, + = Slow, □= SSWS. Mean Difference is represented by the solid line, and the limits of agreement are represented by the dotted line. GCT=gait cycle time, SLS=single limb support time, DLS= double limb support time.
Table 1. Comparison of Observation Counts of Temporal Gait Parameters Between the MatScan and MAPR Systems Table 2. Average Coefficient of Variation for each measurement system during Self Selected Walking Speed. Table 3. Stride-by-stride analysis of HS concurrence between the NZC and LP algorithms in number of samples (sample = 20ms) for Slow and SSWS conditions, as well as cumulative analysis. Mean difference is average difference between observed HS identification times. RMSE is root mean square error between the observed HS identification times. % ±1 sample is the percent of strides where NZC and LP identified HS within one sample of each other. Table 4. Temporal Gait Analysis Performance Comparison of NZC and LP Algorithms: Agreement with the Criterion Measure. RPC= reproducibility coefficient, LoA = limits of agreement, CV=coefficient of variation.
18
Figguree 1
19
Figure 2
20
Figure 3
21
Figure 4
22
Table 1 Slow Temporal Gait Parameters
SSWS
Cumulative
MatScan
MAPR
MatScan
MAPR
MatScan
MAPR
GCT
213
1169
125
876
338
2045
SLS
232
1169
136
876
368
2045
DLS
338
1169
243
876
581
2045
GCT = Gait Cycle Time, SLS = Single Limb Support Time, DLS = Double Limb Support Time,
Table 2 Coefficient of Variation (%) Temporal Gait Parameters
p values
MatScan
LP
NZC
MatScan vs LP
MatScan vs NZC
GCT
3.50 ± 0.74
4.60 ± 1.40
3.73 ± 1.12
0.0418*
0.6458
SLS
9.20 ± 1.47
6.84 ± 1.95
6.81 ± 1.95
0.0071**
0.0071**
DLS
15.0 ± 3.67
58.33 ± 27.01
16.2 ± 6.66
<0.0001**
0.8955
GCT = Gait Cycle Time, SLS = Single Limb Support Time, DLS = Double Limb Support Time, LP=Lee&Park, NZC = Noise Zero Crossing * = p<0.05; ** = p<.01
Table 3 Slow
SSWS
Cumulative
Mean Difference (samples)
-2.71
-.06
-1.54
RMSE (samples)
5.15
2.29
4.14
59.7%
87.5%
72.0%
±1 Sample (percent)
RMSE=root mean square error, 1 sample = 20ms
23
Table 4 Mean
LoA
RPC
CV
Kendall’s τ
LP
-11
[-93, 71]
82 (6.2%)
3.1%
0.064 (p = 0.17)
NZC
-9.7
[-91, 71]
81 (6.1%)
3.1%
-0.12 (p = 0.0068)
LP
170
[50, 280]
120 (24%)
12%
-0.61 (p < 0.001)
NZC
19
[-54, 91]
73 (18%)
9%
-0.21 (p < 0.001)
LP
-170
[-280, -66]
110 (58%)
30%
0.57 (p < 0.001)
NZC
-23
[-78, .32]
55 (22%)
11%
-0.017 (p = 0.71)
GCT
SLS
DLS GCT = Gait Cycle Time, SLS = Single Limb Support Time, DLS = Double Limb Support Time, LP=Lee & Park, NZC = Noise Zero Crossing, LoA= Bland-Altman Analysis limits of agreement, RPC= Reproducibility Coefficient, CV= coefficient of variation
24