The application of principal component analysis to characterize gait and its association with falls in multiple sclerosis

People with multiple sclerosis (PwMS) demonstrate gait impairments that are related to falls. However, redundancy exists when reporting gait outcomes. This study aimed to develop an MS-specific model of gait and examine differences between fallers and non-fallers. 122 people with relapsing–remitting MS and 45 controls performed 3 timed up-and-go trials wearing inertial sensors. 21 gait parameters were entered into a principal component analysis (PCA). The PCA-derived gait domains were compared between MS fallers (MS-F) and MS non-fallers (MS-NF) and correlated to cognitive, clinical, and quality-of-life outcomes. Six distinct gait domains were identified: pace, rhythm, variability, asymmetry, anterior–posterior dynamic stability, and medial–lateral dynamic stability, explaining 79.15% of gait variance. PwMS exhibited a slower pace, larger variability, and increased medial–lateral trunk motion compared to controls (p < 0.05). The pace and asymmetry domains were significantly worse (i.e., slower and asymmetrical) in MS-F than MS-NF (p < 0.001 and p = 0.03, respectively). Fear of falling, cognitive performance, and functional mobility were associated with a slower gait (p < 0.05). This study identified a six-component, MS-specific gait model, demonstrating that PwMS, particularly fallers, exhibit deficits in pace and asymmetry. Findings may help reduce redundancy when reporting gait outcomes and inform interventions targeting specific gait domains.


Results
Principal component analysis. Twenty-one gait variables were included in the principal component analysis, yielding six orthogonal components and accounting for 79% of gait variance. The components were labeled as pace (24.81% of total variance), rhythm (16.57%), variability (13.02%), asymmetry (9.27%), anteriorposterior (AP) dynamic stability (8%) and mediolateral (ML) dynamic stability (7.47%). A threshold for relevant item loadings was set at 0.50 or greater, and no cross-loadings were observed. The results of the PCA and factor loadings are shown in Fig. 1.
Fallers versus non-fallers analysis. Seventy-nine PwMS had reported no falls in the last six months, while 42 PwMS had reported at least one fall. MS-F walked with significantly reduced pace compared to MS-NF (p < 0.001, Fig. 2c). The MS-F group walked slower with a reduced swing and stride velocity with short strides than the MS-NF group (p < 0.001 for all; Table 1 and Fig. 2b). Turning metrics were altered across MS-NF and MS-F groups, as MS-F turned slower and took more steps (p < 0.001; Table 1 and Fig. 2b) than MS-NF. Knee and shank range of motion was reduced in MS-F relative to MS-NF. The rhythm domain was also impacted, as MS-F walked with a reduced cadence and prolonged gait cycles than MS-NF (p = 0.003 and p = 0.001 respectively; Table 1 and Fig. 2b). However, this gait domain was not statistically different between MS-F and MS-NF (p = 0.69, Fig. 2c). Finally, MS-F walked with significantly greater asymmetry than non-fallers (p = 0.003; Fig. 2c). MS-F exhibited greater stance and swing asymmetries than MS-NF (both p = 0.01; Table 1

Discussion
This study utilized principal component analysis to create a streamlined MS-specific gait model and directly relate distinct gait domains to falls. The association between gait domains and clinical, cognitive, and quality of life characteristics of MS was also investigated. We identified six distinct gait domains: pace, rhythm, variability, asymmetry, AP, and ML dynamic stability, and the pace and asymmetry domains were worse in MS fallers than non-fallers. Furthermore, pace was associated with subjective fear of falling, cognitive performance, and functional mobility.
The current model of gait in PwMS exhibits subtle differences with PCA analyses performed in other populations. The most recent analyses describe a conceptual model of gait consisting of five domains performed in healthy older adults and people with Parkinson's Disease (PD) 8,9,11 . The pace, rhythm, asymmetry, and variability   www.nature.com/scientificreports/ domains are consistent with prior models. Indeed, many of the loadings are compatible with older adults with the pace domain encompassing stride velocity and length, the component containing the cadence and temporal parameters 9 . The current MS cohort also demonstrated a similar divergence from the healthy older adult gait model seen in PD patients. Specifically, compared to neurotypical adults, both PD and MS models show the gait variability domain to explain more of the total variance than asymmetry than controls. All but two variability measures (frontal and sagittal trunk ROM) were loaded on the variability domain, diverging from the gait model observed in older adults in which variability was dispersed between domains 8 . This may suggest that gait variability is a more salient component of pathological gait. The additional gait variables in the current study allowed for the expansion of previous models of gait. A postural control domain has been identified in previous analyses and is characterized by step width and step length asymmetry 8,9,11 . However, additional measures of dynamic postural control are needed. The inclusion of trunk kinematics during gait resulted in expanding the previously reported postural control domain 8,9,11 into two directionally dependent dynamic stability components (AP and ML). Combined, these domains explain approximately 15% of the total gait variance, greater than the ~ 8% explained by the postural control domains in older adults 8 and PD 11 . The inclusion of trunk kinematics in gait models is warranted, particularly in PwMS, who exhibit a larger trunk range of motion and variability in the ML direction 19 .
Metrics of turning were also a novel inclusion in the current model. Deterioration of balance and coordination have been related to turns and other postural transitions, and the examination of gait beyond straight-line walking is necessary 17,20 . Surprisingly, the inclusion of these parameters did not elicit a turning-specific domain of gait. Instead, both the number of turns and peak turn velocity loaded onto the pace domain. The fact that turning data was averaged from three iTUG trials (each contains a single turn) may have limited the turning metrics' predictive power. Future studies analyses should consider incorporating more turns performed during longer duration walking tests.
Within our cohort of MS participants, we identified disparities in the pace, asymmetry, and, to a smaller degree, rhythm domains of gait between fallers and non-fallers. Specifically, PwMS, who had reported at least one fall in the previous six months, walked significantly slower than those who had not fallen. In fact, all seven gait variables that loaded on the pace domain were statistically different between the two groups ( Table 1; Fig. 2). The finding that the pace domain (as defined by the PCA in the MS group) was related to falls reinforces the clinical utility and importance of speed as a measure of walking and lower extremity function in MS. In our study, pace accounted for the most significant variance in gait (24.81%) and (consistent with previous work 21 ) discriminated between fallers and non-fallers. Such as finding is significant as walking speed, often measured during the timed 25-foot walk test, is a simple and commonly used assessment both in the clinic and in pharmacological and rehabilitation trials 22 . Therefore, at a time when the clinical evaluation of gait is advancing with technology, the low-tech quantification of pace (e.g., a stopwatch) may provide an effective and inexpensive approach to measure gait, suitable for clinical settings.
In addition to walking slower, fallers also walked with greater stance and swing time asymmetries. Recent investigations suggest that asymmetric gait is a robust measure characterizing the gait in PwMS and related to disease severity and falls 23,24 . Notably, asymmetric gait is not necessarily an uncoordinated gait 25 . Therefore, future studies should include coordination measures such as the phase coordination index 25 , as this measure may also be a relevant outcome for PwMS 23 . Identifying pace and asymmetry as two domains of gait associated with falls is important in assisting with the early prediction of fall risk in MS 26 . It may also support the development of targeted early interventions to decrease the risk of injurious falls in this population.
Surprisingly, variability measures were not statistically significantly different between fallers and non-fallers (p = 0.07). Greater variability has been shown to be predictive of fall status in MS 27 . It is possible that the ability of variability to detect between-group differences was masked due to the short walking assessment. Also, variability is more pronounced and associated with fall risk in moderately-severely impaired PwMS (EDSS score range 4-5) 28 . The MS cohort in this study was less impaired (median EDSS 2.0), which may contribute to the lack of difference in variability metrics between fallers and non-fallers. Finally, despite the lack of difference in gait variability across fallers and non-fallers, the variability domain was related to falls efficacy, underscoring the importance of variability outcomes for PwMS.
Only the pace domain was significantly associated with quality-of-life outcomes. Specifically, and accounting for other domains and possible confounding variables, increased pace was related to lower fear of falling, increased functional mobility, and increased cognitive performance. These findings are partially consistent with previous work. Upwards of 60% of PwMS report a fear of falling, and this outcome has been linked to gait impairments, including lower walking speed and increased stride time variability, and shorter stride lengths 29,30 . A recent meta-analysis showed cognition also to be associated with gait speed in older adults 31 . Specifically, when cognition and gait were portioned into domains, pace was associated with attention and executive function 32 . We observed that increases in pace were related to Stroop interference score improvements, a measure of executive function and attention 33,34 . Significantly, executive function was estimated to correspond to 5-to 10-year deterioration in gait 35 . Together, this work further and specifically links pace with clinically relevant outcomes in PwMS. It further provides preliminary evidence that interventions impacting pace could also, directly or indirectly, improve clinical outcomes such as fear of falling and executive function and attention.
The inclusion of gait data quantified using body-worn wireless inertial sensors is novel to this PCA, as previous studies have utilized some form of a gait mat 8,11 . Inertial sensors enable continuous measurement throughout the entire walking assessment instead of only collecting data while walking on the mat. The use of body-worn sensors enables the measurement of trunk angles, lower limb kinematics, and turning characteristics during gait, in addition to spatiotemporal gait measures. However, spatial data is difficult to capture via body-worn inertial sensors, reducing the ability to capture characteristics such as step length and width via this approach. www.nature.com/scientificreports/ There are several limitations to this study. The gait assessment was performed appending three trials of the iTUG protocol. Ideally, gait data, and particularly variability data, would be collected during a continuous walk of a longer duration. The body-worn inertial sensors provide stride characteristics and not step data, so we could not measure step length and step width data (including variability and asymmetry). Also, our MS cohort was restricted to those with relapse-remitting MS and demonstrated relatively minimal impairment (median EDSS 2.0). Therefore, our findings' generalizability is limited to PwMS to MS patients with RMSS and mild severity. Our study accounted for only retrospective falls. It is possible that some participants misestimated the number of falls they experienced in the previous six months, especially given the prevalence of cognitive dysfunction in this population 14 . Future studies should incorporate prospective reporting of falls to avoid such erroneous recall and, when possible, integrate feedback from a spouse or caregiver. In addition, some components of the PCA model may be under-specified (2-items) but were retained as they could discriminate between MS-F and MS-NF. The assessment and interpretation of cognitive function were limited to the domains evaluated in the Stroop Color Word Test. Using the scoring method proposed by Golden & Freshwater 1978 36 , the interference score generated describes the ability to inhibit cognitive interference and fronto-executive functioning 37 . Future studies should examine the association between specific domains and a broader range of cognitive domains such as information processing speed, visuospatial memory, and working memory, which are related to gait in PwMS 38,39 . Finally, while this is a critical first step in creating a more streamlined gait model in MS, PCA analysis only explains communal variance. Therefore, the possibility that variables that are loaded together may not necessarily represent the underlying construct. Future research should conduct confirmatory PCAs in this population to determine the robustness of the domains and constructs identified in the current model.

Conclusion
This study identified an MS-specific model of gait consisting of six distinct domains. Of these domains, pace and asymmetry were significantly different between MS fallers and non-fallers, and increased pace lowered the likelihood of being an MS-F. The pace domain was also associated with functional mobility and fear of falling. Establishing a more streamlined gait model in PwMS may reduce redundancy in gait outcome reporting for future studies, improve the characterization of clinically observable function and their deficits (i.e., pace, rhythm, timing), and ultimately advance the clinical evaluation and rehabilitation of gait via a standardized multi-domain assessment. Further, identifying which domains are relevant for important outcomes such as falls and fear of falling may assist in the early prediction of fall risk in MS and support earlier interventions to reduce the risk of injurious falls. Future studies should expand analyses to represent a more heterogeneous sample across the MS severity spectrum, and track gait domains' progression and their relation to falls longitudinally. Follow-up work should also investigate the potential imaging markers associated with the distinct gait domains and their association with fall risk. Such investigation would provide insight into the neural correlates of domain-specific gait deficits in MS and may facilitate more targeted rehabilitation.

Participants.
A convenience sample of 122 people with relapse-remitting MS and 45 age-matched controls was recruited (Table 3). Participants were recruited via the MS clinic at the University of Kansas Medical Center. Exclusion criteria were: (1) an inability to give consent, (2) an Expanded Disability Status Scale Score > 5.5 or the use of an assistive device, (3) any musculoskeletal or orthopedic impairments that would affect balance or mobility and, (4) any neurological disorder other than MS. All participants provided written informed consent before participation. The study protocol was approved by the ethics board at the University of Kansas Medical Center and was conducted in accordance with the Declaration of Helsinki.  Quantitative gait assessment. Protocol. Gait data were computed from the instrumented timed up and go (iTUG) assessment, a reliable and valid measure of gait and turning 6,47 . Participants were instructed to stand up from the chair, walk just past a line placed 7 m straight ahead at a comfortable pace, turn around, walk back, and sit down. The iTUG was extended from the traditional 3 m to enable the computation of gait cycle data. Participants completed three iTUG trials.
Data analysis. Gait data were collected using Opal wireless inertial sensors (128 Hz). Six sensors placed on the feet, wrists, chest, and lumbar region of the lower back were utilized. Spatiotemporal gait outcomes were determined from inertial measurements and foot positions during the gait cycle measured by the sensors 48 . Specifically, Mobility Lab software (Version 2) (Opal Sensors, APDM Inc., Portland, OR) was used to stream and export gait metrics automatically 48 . This is a reliable and valid system for quantifying gait and mobility dysfunction 49 . The gait cycle's temporal characteristics were computed relative to gait cycle duration, defined as the duration from the foot's initial contact to the next initial contact of the same foot. Steps during gait and turns were detected using the shanks' two sensors 47 . Gyroscopes on the trunk and lumbar sensor detected turns 47 .

Gait characteristics for principal component analysis.
To ensure an adequate number of gait cycles, data from each participant's three iTUG trials were appended. After appending data, the median (range) of gait cycle observations was 20 (12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26) for controls, 21  for MS non-fallers, and 22  for MS fallers. Mean spatiotemporal parameters, averaged across the right and left limbs, kinematic measures, turning parameters, variability metrics, and asymmetry measures were then assessed. Table 2 outlines gait parameters, and definitions for all gait outcomes are presented in supplementary table 1. The variability of gait outcomes was defined as the coefficient of variation (CoV) computed as standard deviation/mean. Several previous studies have used short walks (i.e., up to 10 m) to report variability in kinematic and spatiotemporal gait parameters 5,50 . Asymmetry across the lower limbs was calculated as gait asymmetry [%] = 100 * |ln (Right Limb/Left Limb)|. The number of steps in each turn was measured by taking the average of the three iTUG trials.
Choice of outcomes to include in the model was made to ensure a breadth of spatial and temporal outcomes while limiting redundancy. Specific outcome choice was informed by gait deficits highlighted in MS in systematic reviews and meta-analyses 3,5 , and previous factor analyses 8,9,11,12 . Further, we sought to expand on previous principal component analyses by including metrics of turning and trunk range of motion as a marker of dynamic stability given their importance for walking performance and falls 6,18,51,52 . Variability was defined as the coefficient of variation (CoV) to remain consistent with previous reports 5 .
Statistical analysis. The data analysis consisted of three parts: (1) a PCA on spatiotemporal gait parameters in MS participants, (2) the comparison and association of gait domains between MS-fallers (MS-F) and non-fallers (MS-NF) using independent t-tests and binary logistic regression, and (3) multiple linear regression analyses of cognitive, clinical, and quality of life characteristics to gait domains. Although not a primary aim of this report, gait parameters were also compared between PwMS and controls. Data were checked for normality using the Shapiro-Wilks test and the visual inspection of histograms. The linearity of variables and normal distribution of residuals was examined via the visual inspection of scatterplots and Q-Q plots for the multiple regression analysis. Homoscedasticity was assessed by visually inspecting a scatterplot of the residuals and predicted values.
A PCA that uses the communal variance of included gait parameters was performed to identify a more parsimonious representation of the latent construct gait in PwMS. Mean gait data from healthy control participants were used for comparative and reference purposes. Therefore, a PCA analysis was not performed with this data. Components were derived using varimax rotation to produce orthogonal partitions. Kaiser's score and Cattell's Scree plot were examined to identify the number of components to extract. Cross-loadings were also examined.