Comparison of Gait Parameters Collected Across Two Commercially Available Gait Systems in Older Adults

Abstract

Background/Objectives: Research examining mobility in older adults often utilizes spatiotemporal gait parameters obtained from computerized walkway systems like GAITRite (Franklin, NY, USA) and Zenometrics, LLC (Peekskill, NY, USA). However, such parameters can vary across these widely used software applications due to differences in algorithms and post-processing techniques, making it potentially unsuitable to pool parameters acquired from different applications. We aim to determine whether gait parameters obtained from GAITRite and processed using two software systems can be reliably pooled for use in studies with older adults. Methods: Baseline gait data from 193 older adults aged 64–94 years old were examined. The gait protocol consisted of normal walking (NW) and walk while talking (WWT) conditions in which participants were instructed to walk on computerized walkways containing embedded pressure sensors. The resulting walks were processed using both software applications to attain eight gait parameters recorded from the right foot (velocity, stride length, double support percentage, cadence, swing time, stance time, stride length standard deviation, and swing time standard deviation). Linear regressions adjusted for age and gender between GAITRite (version 4.7) and their respective PKMAS (ProtoKinetics Movement Analysis Software version 6.0; Zenometrics, LLC.) gait variables were run to determine agreement between variables across the two pieces of software. Results: Adjusted linear regression models revealed that gait parameters across software systems were significantly correlated in both the NW (β ranged from 0.87 to 1.02, p ≤ 0.01) and WWT conditions (β ranged from 0.94 to 1.01, p ≤ 0.01). Swing time variability in the NW condition showed a lower correlation (β = 0.87). Furthermore, intercepts for all parameters except for the double support percentage in the NW condition (intercept = 2.63, p ≤ 0.01) and WWT condition (intercept = 2.75, p = 0.02) and stance time in the WWT condition (intercept = 0.05, p = 0.04) were not significantly different from 0. Conclusions: The results provide support that commonly examined gait parameters from GAITRite and PKMAS can be pooled and analyzed for use in studies with older adults. However, caution should be taken when pooling swing time variability and double support percentage data.

1. Introduction

Research utilizing quantitative gait metrics obtained from normal walking (NW) protocols have revealed quantitative gait dysfunction as predictors of cognitive decline [1,2,3]. Complex walking paradigms that may yield more variable walking patterns, including the walk while talking (WWT) condition, have shown that gait performance predicts falls and dementia [4,5,6]. These studies used gait data acquired on computerized walkways with embedded sensors that reliably record measures like position coordinates, activation times, and activation pressure levels. Programs from GAITRite and PKMAS (ProtoKinetics Movement Analysis Software; Zenometrics, LLC.) utilize these measures to calculate parameters such as velocity and cadence [7,8,9].

In a study comparing gait parameters recorded on GAITRite and Zenometrics walkways, the authors reported high concurrent validity for spatial parameters but less so for temporal gait parameters [10]. The authors processed GAITRite walkway data using GAITRite software and Zenometrics walkway data using PKMAS software. One explanation for the lower concurrent validity in temporal gait parameters between the two walkway systems is the different algorithms used by each software for referencing spatial locations, time frames, and calculating gait parameters. For example, as described in the manual, PKMAS includes an additional scan time of 120 Hz (0.0083 s) in its stance time calculations, which causes stance times from PKMAS to be 0.0083 s longer and swing times from PKMAS to be 0.0083 s shorter than those calculated by GAITRite.

A study comparing GAITRite and PKMAS software found that some gait parameters, like gait speed, cadence, and stride time variability can be used interchangeably [11]. However, this study was limited in that walks were manually reprocessed to contain the same footfalls so that sources of variations were solely due to variations in algorithms—a method which could arguably force walks to appear more consistent across acquisition platforms. Reprocessing quantitative gait data collected from large longitudinal cohort studies using alternative software applications poses concerns regarding time and resources, as well as logistical complications. It is also of interest to compare the two pieces of software as more researchers begin to use and validate the PKMAS software for gait analysis [12,13].

We aimed to determine whether parameters calculated by GAITRite and PKMAS for analysis in studies involving older adults without reprocessing could be reliably pooled. We compared the two software systems using raw gait data acquired on the GAITRite walkway. In our study, we did not manually reprocess the raw walks to have the same footfalls across the different pieces of software since differences in cleaning protocols were inherent across programs and could be a potential source of variation. We extended our analyses to include NW and WWT paradigms since the added attentional cognitive stressor during the WWT condition can increase gait variability. We hypothesized that linear regressions between gait parameters acquired from GAITRite and PKMAS platforms would indicate proportional relationships.

2. Materials and Methods

2.1. Study Population

Ashkenazi-Jewish adults, aged 65 years and older, enrolled in the LonGenity study from March 2008 to May 2023 were considered (1150 enrolled participants). The LonGenity study aims to examine genetic factors associated with healthy aging, where participants are defined as either offspring of parents with exceptional longevity (at least one parent lived to 95 years) or offspring of parents with usual survival [14]. During screening, participants were excluded for major cognitive impairment (as determined by a cutoff score of >2 on the AD8 or >8 on the Blessed Mental Status Examination), severe self-reported visual or hearing impairment, or having a sibling in the study.

For the analysis in this paper, we used gaits acquired on the GAITRite walkway from 193 participants after all exclusion criteria were applied, as described in the Section 3.

All participants provided written informed consent at enrollment, which was approved by the institutional review board at the Albert Einstein College of Medicine.

2.2. Gait Collection and Processing

Research assistants collected NW quantitative gait data by instructing participants to walk at their normal pace on a 20-foot computerized walkway (GAITRite, CIR Systems, Havertown, PA, USA). For the WWT condition, participants were instructed to walk at their normal pace while reciting alternating letters of the alphabet. They were instructed to pay equal attention to both their walking and talking [15]

The walkway included four extra feet at each end to account for initial and terminal accelerations. Participants were asked to complete a second trial if they froze or stumbled. Walks were originally processed using GAITRite software. For this study, the raw GAITRite data were subsequently exported and imported into the PKMAS application. A research assistant reviewed all of the walks in PKMAS and removed partial footfalls or walking aid sensors not algorithmically detected. A footfall was considered to be partial if the foot contact did not show a heel, center, and toe. Footfalls that displayed fragmented or irregular foot contacts were subsequently removed since these can lead to the inclusion of inaccurate total footfalls when calculating gait parameters. The walks were then processed in PKMAS with the foot reference as the heel to obtain the gait parameters of interest. Based on our previous work [1,4], the following 8 parameters (for the right foot) were selected: velocity, stride length, double support percentage, cadence, swing time, stance time, stride length variability, and swing time variability.

2.3. Data Analysis

Baseline characteristics were examined using descriptive statistics, and residual plots were inspected graphically to confirm model assumptions. Linear regressions were used to examine agreement across the GAITRite and PKMAS gait variables. In our models, GAITRite parameters were set as predictors while PKMAS parameters were set as outcome variables. We adjusted for age and gender since they have been associated with differences in gait performance [14,16,17]. In the linear regression models, agreement between the two systems was indicated by β coefficients that were not significantly different from 1 and intercepts that were not significantly different from 0. We further assessed the root mean square errors (RMSEs) of the linear regression models to evaluate the goodness of fit.

3. Results

3.1. Study Population

Of the 1150 LonGenity participants, 295 had no gait assessment and 586 were missing raw GAITRite data files. Of the 269 participants with baseline gait assessments, 70 were excluded since they lacked eight footfalls, and 3 were excluded since they used a walking aid. We required walks to contain at least eight footfalls for a reliable calculation of the two variability measures included in the study. Stride length variability and swing time variability for the right foot were calculated by taking the standard deviation of the stride lengths and swing times of the right foot, respectively. For there to be at least three data points in the standard deviation calculations, at least four right footfalls were required, which amounts to eight footfalls overall.

Participants with a diagnosis of Parkinson’s Disease (N = 1) or dementia (N = 2) were also excluded. After exclusions, walks from a total of 193 older adults were included in this comparison study (

Table 1. Demographic and Clinical Characteristics of Study Sample at Baseline ^†.

Variable	N = 193
Age, years	74.17 ± 6.92 (64–94)
Female	138 (71.5)
Education, years	17.80 ± 2.83 (4–25)
Fall in the past year	44 (22.80)
Arthritis	83 (43.01)
Hypertension	60 (31.09)
Stroke	5 (2.59)
Angina	3 (1.55)
Myocardial infarction	10 (5.18)
Cardiac arrythmia	36 (18.65)
Diabetes	11 (5.70)

^† Values are presented as mean ± SD and range for continuous variables, and as counts and % for dichotomous variables.

).

3.2. Linear Regressions

The gait parameter values generated by each piece of software were reasonable based on the previous literature (Supplementary Tables S1 and S2) [1,8,15]. Adjusted linear regressions revealed significant correlations between GAITRite and PKMAS for all gait parameters in the NW and WWT conditions (

Table 2. Age- and Gender-Adjusted Linear Regression for Normal Walking (NW): ProtoKinetics Movement Analysis Software (PKMAS) and GAITRite.

	Root Mean Square Error	Unstandardized Coefficients—β			95% Confidence Interval for β		Unstandardized Coefficients—Intercept
		β	Std. Error	p-Value	Lower Bound	Upper Bound	Intercept	Std. Error	p-Value
Velocity, cm/s	0.71	1.01	0.00	<0.001	1.00	1.02	−0.83	0.75	0.27
Stride length, cm	0.36	1.00	0.00	<0.001	0.99	1.00	0.65	0.48	0.18
Double support %	0.82	1.00	0.02	<0.001	0.97	1.03	2.63	0.71	<0.001
Cadence, steps/min	0.32	1.00	0.00	<0.001	0.99	1.00	−0.21	0.35	0.55
Stance time, s	0.01	1.02	0.01	<0.001	0.99	1.04	0.02	0.01	0.17
Swing time, s	0.00	0.96	0.01	<0.001	0.94	0.98	0.00	0.01	0.63
Stride length variability (SD)	0.67	0.94	0.02	<0.001	0.90	0.98	−0.36	0.52	0.49
Swing time variability (SD)	0.00	0.87	0.03	<0.001	0.81	0.94	0.00	0.00	0.67

and

Table 3. Age- and Gender-Adjusted Linear Regression for the Walk While Talking (WWT) condition: ProtoKinetics Movement Analysis Software (PKMAS) and GAITRite.

	Root Mean Square Error	Unstandardized Coefficients—β			95% Confidence Interval for β		Unstandardized Coefficients—Intercept
		β	Std. Error	p-Value	Lower Bound	Upper Bound	Intercept	Std. Error	p-Value
Velocity, cm/s	0.93	1.01	0.00	<0.001	1.01	1.02	−0.79	0.80	0.32
Stride length, cm	0.35	1.00	0.00	<0.001	0.99	1.00	0.39	0.39	0.33
Double support %	1.35	0.94	0.02	<0.001	0.90	0.98	2.75	1.17	0.02
Cadence, steps/min	0.29	1.00	0.00	<0.001	1.00	1.00	0.16	0.24	0.52
Stance time, s	0.03	0.98	0.01	<0.001	0.97	1.00	0.05	0.02	0.04
Swing time, s	0.01	0.98	0.00	<0.001	0.98	0.99	0.00	0.01	0.87
Stride length variability (SD)	0.63	0.97	0.02	<0.001	0.94	1.01	0.72	0.49	0.15
Swing time variability (SD)	0.01	0.99	0.01	<0.001	0.98	1.01	0.00	0.00	0.81

).

In the NW condition, the 95% confidence intervals for the β coefficient contained 1.00 for velocity, stride length, double support percentage, cadence, and stance time, indicating that the slopes for these variables were not significantly different from 1 (Table 2). Although the β coefficients for swing time and stride length variability were less than 1, the differences were minimal (β = 0.96, 95% CI = (0.94, 0.98), p ≤ 0.01, and β = 0.94, 95% CI = (0.90, 0.98), p ≤ 0.01, respectively). Although swing time variability did not show a proportional relationship with a β coefficient of 1, a significant proportional relationship was still demonstrated (β = 0.87, p ≤ 0.01). Furthermore, the intercepts for the linear regression equations for all parameters except for the double support percentage were not significantly different from 0 (intercept = 2.63, p ≤ 0.01) (Table 2).

For the WWT condition, adjusted linear regressions indicated significant direct and proportional relationships between the two pieces of software for all gait parameters (β range: 0.94–1.01, p ≤ 0.01). The 95% confidence intervals for the β coefficient contained 1.00 for all gait parameters except for the double support percentage and swing time (Table 3), while their β coefficients were slightly lower than 1 (β = 0.94, 95% CI = (0.90, 0.98), p ≤ 0.01, and β = 0.98, 95% CI = (0.98, 0.99), p ≤ 0.01, respectively). To add to this, the intercepts for the linear regression equations for all parameters were shown to not be significantly different from 0 except for the double support percentage (intercept = 2.75, p = 0.02) and stance time (intercept = 0.05, p = 0.04) (Table 3).

4. Discussion

Based on the NW condition regression results, the two systems showed agreement on velocity, stride length, cadence, and stance time. Differences were shown in the change rate for swing time and stride length variability, but these differences are acceptable since the slopes are close to 1. The double support percentage showed a significant difference in the intercept but the same change rate. Apart from the swing time variability and double support percentage, six of the eight GAITRite and PKMAS gait variables can be pooled in the NW condition.

For the WWT condition, the two systems reached agreement on stride length, cadence, stride length variability, and swing time variability. The intercept for velocity was slightly less than 1 and the 95% confidence interval for β did not contain 1. Additionally, the slope for swing time was slightly lower than 1. However, both were very close to 1 with intercepts not being significantly different from 0. This indicates strong agreement between the two systems. Stance time had a small significant intercept but an intercept value of 0.05 is minimal, suggesting agreement in this case as well. The double support percentage had a slope that was slightly lower than 1 but a significant intercept of 2.75, indicating a small difference between the systems. Therefore, seven of the eight GAITRite and PKMAS gait variables can be pooled in the WWT condition, excluding the double support percentage.

The β coefficients ranged from 0.94 to 1.02 except for swing time variability in the NW condition. The lower correlation may stem from differences in the software algorithms. Spatial parameters such as velocity and stride length differ in their absolute values since GAITRite defines the reference point as the center of the heel, while PKMAS defines the reference point as the back of the heel. Egerton et al. explained that greater foot angles can affect the location that GAITRite calculates as the heel center [11]. Walks with greater variability in foot angles can cause variability in the heel center and, consequently, the calculation of spatial parameters. Furthermore, spatial parameters in PKMAS are calculated based on the direction of progression for the ipsilateral stride, while GAITRite assumes a constant direction of progression (i.e., parallel to the mat’s length). In linear walks, this should not result in noticeable differences in values across different pieces of software. In nonlinear walks, however, the resulting spatial parameter values may differ since the direction of progression is not constant [18]. We saw this affect the velocity values calculated by the different pieces of software. For our study, this is shown in the velocity differences. PKMAS calculates velocity by dividing the sum of all stride lengths by the sum of all stride times, while GAITRite calculates velocity by dividing the total distance traveled by the ambulation time. Because PKMAS uses stride length in its velocity calculations, which takes into account the instantaneous direction of progression, the PKMAS calculated velocities are expected to be greater if the direction of progression is not straight and aligned with the mat. Based on Supplementary Tables S1 and S2, the velocities obtained from PKMAS were, on average, greater than those from GAITRite by 1 cm/s.

Differences in the temporal gait parameters, such as swing time and stance time, may be due to the additional scan time (120 Hz/0.0083 s) included in PKMAS’s stance time calculation [10]. As was described in the introduction, PKMAS stance times were expected to be 0.0083 s longer and swing times were expected to be shorter than those from GAITRite. The absolute differences in this study’s dataset did not average 0.0083 s, but stance times from PKMAS were, on average, greater than stance times from GAITRite (0.0222 s for NW, 0.0227 s for WWT) and swing times from PKMAS were, on average, shorter than swing times from GAITRite (0.0125 for NW, 0.0159 for WWT) (Supplementary Tables S1 and S2). This also helps explain the non-zero intercept found in the linear regression models for the NW double support percentage. To calculate the double support time in PKMAS, the software takes the sum of the initial and terminal double support times as well as the stance time, so a difference in stance times can affect the final calculated double support time and percentage.

This study is limited to the gait variables examined. However, most of the other gait variables can be derived from the eight gait variables evaluated here. The study cohort was also genetically homogenous, consisting of community-dwelling Ashkenazi-Jewish older adults so the results may not be generalizable to populations with more severe gait dysfunction.