Gait adaptations to walker-assisted locomotion in elderly: a principal component analysis of spatiotemporal and kinematic parameters

Abstract

This study investigated gait adaptations in elderly individuals by comparing normal walking with walker-assisted gait through a Principal Component Analysis (PCA) of spatiotemporal parameters and sagittal plane kinematics, which are less influenced by intersubject variability. Fourteen unimpaired older adults performed two walking tasks: unassisted gait, and walker-assisted gait with a smart walker designed to enhance stability and support during locomotion. Gait data were acquired using a wearable 3D motion capture system based on inertial measurement units. The results showed that combining joint kinematics and spatiotemporal metrics provides a more comprehensive perspective of gait domains in older adults. In normal walking, distinct components were identified, including pace/variability, asymmetry, rhythm, base of support, and step length, while the incorporation of kinematic data further delineated specific motion patterns of the knee, ankle, and hip joints. Walker-assisted gait exhibited slower pace, increased temporal variability, modifications in the base of support and increased hip flexion throughout the gait cycle. Additionally, this modality exhibited two new gait domains related to swing phase mechanics and step dynamics. These findings form a biomechanical foundation that isolates device-induced adaptations and can be used to tailor rehabilitation protocols and smart walker designs aimed at promoting safe and effective locomotion in older adults.

Introduction

Human gait undergoes significant changes with aging, often resulting in reduced walking speed, increased gait variability, and impaired balance control. These alterations contribute to a higher risk of falls, which are a leading cause of morbidity and mortality among older adults^1,2. Epidemiological data indicate that approximately one-third of individuals over 65 years old experience at least one fall per year, with a substantial proportion leading to fractures, hospitalization, and loss of independence^3,4.

Conventional walking frames are typically prescribed to older adults to assist locomotion and prevent falls. However, despite their widespread use in rehabilitation and daily activities, such walkers can present challenges and have been linked to increased loss of balance and risk of falls due to difficult maneuverability⁵. To address these and other limitations, walkers can be enhanced with actuators and sensor technologies, transforming them into Smart Walkers (SW). These innovations offer personalized locomotion assistance, improving stability, facilitating health monitoring, and enhancing navigation support^6,7.

Although walker users typically present neurological or musculoskeletal impairments, the literature provides limited evidence on the biomechanical impact of conventional and SWs. Researchers have scarcely investigated the biomechanical adaptations associated with walker use, and the literature does not yet report direct comparisons between conventional and SWs⁶. This gap limits our understanding of how different walker designs affect gait patterns and underscores the need for future studies in this field. Early studies conducted with unimpaired older adults are an essential first step in the development of assistive devices because they establish a normative reference that separates device-induced adaptations from pathology-related compensations^7,8. By isolating these effects, clinicians and engineers can later determine whether a walker truly mitigates or merely reshapes pathological gait patterns and can tailor controller parameters or training protocols accordingly.

Gait analysis tools offer the ability to identify key biomechanical markers associated with fall risk and mobility impairments in the geriatric population⁹. Traditional gait studies have primarily relied on spatiotemporal parameters to assess walking stability and efficiency, highlighting the role of gait speed, step variability, base of support, and asymmetry in predicting functional decline^10–12. Unlike simple comparisons of average spatiotemporal or kinematic variables, Principal Component Analysis (PCA) enables the identification of latent domains that emerge from correlations among multiple gait parameters. This multivariate perspective provides a more comprehensive understanding of how different aspects of gait interact, offering insights into motor control strategies that cannot be obtained from univariate analyses alone^13–15. Studies utilizing PCA have consistently demonstrated that step asymmetry, rhythm and gait variability load onto distinct components, reinforcing their independent contributions to gait control^12–17. However, while these measures provide valuable insights, they offer a limited understanding of the underlying joint biomechanics governing gait adaptations in aging populations, since most PCA studies in older adults have focused almost exclusively on spatiotemporal parameters, with little or no inclusion of joint kinematics^14–17.

In recent years, advancements in wearable sensor technology, particularly inertial measurement units (IMUs), have revolutionized gait analysis by enabling the acquisition of kinematic data in real-world environments^18–20. Unlike optical motion capture systems, which are restricted to laboratory settings, IMUs provide a portable and cost-effective means of assessing joint motion during daily activities^21,22. Previous studies have validated the accuracy of IMU-based measurements against optical systems, confirming their suitability for clinical and research applications^23,24. Nevertheless, IMUs remain susceptible to drift, which may compromise precision during long-duration recordings. Despite this limitation, they represent a powerful tool for evaluating the biomechanics of assistive gait modalities, such as walker-assisted ambulation, particularly to study how these devices modify lower limb movement patterns⁸.

In this study, we investigated the hypothesis that smart walker-assisted gait would significantly alter spatiotemporal and kinematic parameters compared to normal walking in unimpaired older adults. To test this, we first conducted pairwise comparisons between conditions. We then employed PCA as an exploratory approach to compare gait domains between normal, unassisted walking and smart walker-assisted gait, aiming to characterize how assistive technology reshapes locomotor patterns and to establish a biomechanical foundation for future investigations in clinical populations such as stroke, Parkinson’s disease, and frailty.

Methods

Participants

Fourteen unimpaired older adults (5 men and 9 women) recruited from the local community through a university extension program that offers recreational activities for older adults and a community-based exercise guidance center. Participants had a mean age of 66.3 (± 3.9) years, a mean weight of 71.43 (± 10.8) kg, and a mean height of 1.66 (± 0.09) m.

Inclusion criteria required participants to be at least 60 years old and able to perform the 10 m walk test with the smart walker, and to present adequate cognitive function (Mini-Mental State Examination score ≥ 24). Exclusion criteria included the presence of comorbidities that impaired walking, smoking, use of thermogenic supplements or medications known to compromise locomotion or cognition.

Participants were fully briefed about the potential risks and discomforts associated with the research procedures. The study was approved by the Research Ethics Committee of the Federal University of Espírito Santo (registration number 6.294.101). All methods were performed in accordance with the guidelines of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to their inclusion in the study.

Instruments

The materials used in this study included a commercial wearable 3D motion capture system (MVN Awinda, Movella, USA) and the UFES vWalker, a SW developed by our research group.

1) 3D Motion Capture System: The MVN Awinda system was used to capture lower limb gait data from the participants. Seven wireless Inertial Measurement Units (IMUs) were placed on reference sites of the user’s body: one on the pelvis and one sensor for each segment of the legs, with one for the right leg and one for the left leg (upper leg, lower leg, and foot). The pelvic sensor was positioned at the anterior superior iliac spine. For the upper leg, the sensor was placed on the middle of the lateral thigh. The sensor for the lower leg was positioned on the lateral side of the lower leg, aligned with the fibula, above the lateral malleolus. A sensor was also placed on each foot. To enhance the accuracy of kinematic parameters, the IMU data were integrated with 6 anthropometric measurements, including total height, foot length, hip height and width, knee height, and ankle height. These data were processed using the manufacturer’s software (MVN Analyze—Version 2024.1, MVN Awinda, Movella, NV, USA) to extract comprehensive 3D gait information. Beyond its portability, the MVN Awinda system has been previously validated against optical motion capture systems, which are considered the gold standard for gait analysis^23,24.

2) UFES vWalker: The robotic device is equipped with a range of sensors and actuators to enhance stability, navigation, and motion control²⁵. The system includes two triaxial force sensors (MTA400, FUTEK, USA), a leg-tracking Light Detection and Ranging (LiDAR) sensor (RPLIDAR A3, SLAMTEC, China), an environmental LiDAR sensor (URG-04LX, Hokuyo, Japan), an inertial measurement unit (IMU) (BNO055 9-DOF, BOSCH, Germany), two rotary encoders (H1 Series, US Digital, USA), two motorized wheels (MB03024 30 RPM, TEK8, Brazil), and two caster wheels. Figure 1 illustrates the UFES vWalker, along with its sensors and the 3D motion capture system used in this study.

Fig. 1.

UFES vWalker, its sensors and the 3D motion capture system.

The walker operates based on an admittance control strategy, as proposed by Jimenez et al.²⁶, mapping the user’s forward forces into linear velocity, as described by Eqs. 1 and 2.

1

2

In this approach, the forward forces exerted by the user’s left and right arms, F_FL and F_FR respectively, are combined into a resultant force F as shown in Eq. 1.

This force is then used as input in Eq. 2, where it is processed with a virtual mass (m), a damping parameter (d) and the device’s acceleration (a) to generate a linear velocity (v) command that ensures smooth and controlled movement.

For safety purposes, the downward force detected by the force sensors is used to confirm a secure positioning of the user over the device. If this force is below 5% of the user’s body weight, the walker remains stationary, preventing unintended movement. This force threshold was implemented as a safety measure to detect reduced contact between the user and the device, which may indicate loss of balance. In such cases, stopping the walker prevents it from moving away from the user. Given its low maximum speed (0.4 m/s) and stable base, the risk of tripping or hindrance from the stationary device is minimal.

The leg-tracking LiDAR sensor also serves as a safety mechanism by tracking the user’s legs using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. This system continuously measures the distance between the user’s legs and the UFES vWalker²⁷. If this distance exceeds 0.75 m, the device remains stationary, allowing the users to approach or reposition themselves before proceeding.

To warrant accurate motion tracking, odometry was performed by integrating data from the IMU and rotary encoders. Finally, the combination of motorized and caster wheels enables both propulsion and stabilization, ensuring a safe and controlled navigation experience for users.

With this combination of sensing and actuation systems, the UFES vWalker delivers personalized locomotion assistance. Beyond its hardware architecture, the system is capable of dynamically adjusting its control parameters to the user’s context. This adaptability has been demonstrated in previous studies through different strategies: a “following in front” mode that maintains distance and orientation to support natural gait interaction²⁸; admittance-based approaches offering varying levels of guidance, from full autonomy to shared control²⁹; and, more recently, a high-assistance strategy that enabled independent walking in a patient with severe spinocerebellar ataxia³⁰. Together, these studies demonstrate the versatility of the UFES vWalker in modulating its dynamic behavior across both experimental and clinical contexts.

Experimental procedures

All participants underwent a preliminary walker-assisted walking session one day before the formal data collection. This familiarization session allowed them to become familiar with the technology, provide an opportunity to ask questions, and enable the collection of anthropometric measurements for precise sensor placement and optimal adjustment of the walker’s support height⁸.

The official experimental procedure consisted of two distinct walking tasks during the experiments: No Device (ND) and Smart Walker Locomotion (SWL). In the ND task, participants walked 10 m in a straight line without assistance, while in the SWL task, they walked the same 10 m assisted by the UFES vWalker. The walking tasks were performed twice for each condition, beginning with the ND task, followed by the SWL task. The two trials provided enough gait cycles for the reliable extraction of spatiotemporal and kinematic parameters. Throughout both tasks, participants wore the motion capture system to record gait data.

This 10-meter walk test is a widely accepted standard in clinical gait analysis for assessing functional capacity, walking speed, and mobility³¹. By standardizing the analyzed distance in both tasks, the experimental protocol enabled a direct comparison between ND and SWL while reducing variability associated with adaptation, start-up, and stopping. As a result, the extracted gait data more accurately reflects typical walking patterns, providing clear insights into the potential impact of the UFES vWalker on participants’ locomotion.

Gait parameters

This paper analyzed two groups of gait variables across both walking tasks. The first group consisted of the following spatiotemporal parameters:

1. Step Length (m): The distance covered by each step during the tasks.

2. Step Width (m): The horizontal distance between the feet during each step during the tasks.

3. Step Time (s): The time taken to complete each step during the tasks.

4. Stance Time (s): The duration of the stance phase for each step during the tasks.

5. Swing Time (s): The average duration of the swing phase for each step during the tasks.

6. Double Support (DS) Time (s): The time during which both feet are in contact with the ground during each gait cycle in the tasks.

7. Step Velocity (m/s): The velocity of each step during the tasks.

The second set of variables focused on the kinematic angles of the hip, knee, and ankle joints in the lower limbs. Since sagittal plane data exhibits lower intersubject variability³², they are particularly well-suited for detecting differences between ND and SWL⁸. Although selecting a subset of kinematic parameters may lead to some loss of information, our primary objective was to focus on variables that reflected the most important events of the gait cycle. The joint angles were extracted at key instances of the cycle, primarily at the extreme points of each joint’s movement in the sagittal plane. The methodology used for extracting these parameters follows the approach proposed by Benedetti et al.³² A detailed description of these parameters is provided in Table 1.

Table 1.

Kinematic parameters of sagittal angles in degrees at lower limb joints.

HIP JOINT

H1: flexion at heel strike H2: maximum flexion at loading response. H3: maximum extension in stance phase. H4: flexion at toe-off. H5: maximum flexion in swing phase.

KNEE JOINT

K1: flexion at heel strike K2: maximum flexion at loading response. K3: maximum extension in stance phase. K4: flexion at toe-off. K5: maximum flexion in swing phase.

ANKLE JOINT

A1: dorsiflexion at heel strike A2: maximum plantar dorsiflex. at loading response. A3: maximum dorsiflexion in stance phase. A4: dorsiflexion at toe-off. A5: maximum dorsiflexion in swing phase.

Both groups of variables were acquired from the motion capture system. Using the developer software, where foot contact with the ground was marked, each gait cycle was isolated, and the presented group of variables were extracted. All variables were obtained from both the left and right sides of the participant’s gait and analyzed for mean, variability, and asymmetry, following Eqs. 3, 4, and 5, respectively¹⁷.

3

4

5

Statistical analysis

We first compared spatiotemporal and kinematic parameters between the No Device (ND) and Smart Walker Locomotion (SWL) conditions. This analysis was intended as a preliminary step to describe differences between ND and SWL prior to PCA modeling. For each variable, we assessed normality using the Shapiro–Wilk test and visual inspection of histograms. The distribution of the variables in each data set was assessed using the Shapiro-Wilk test and visual inspection of histograms. Normality was acceptable for all parameters in the ND gait data, but in the SWL data sets, the distribution of temporal gait variability and asymmetry measures were skewed. We compared spatiotemporal and kinematic variables within subjects across the ND and SWL conditions. When normality assumptions were met, we applied paired-sample t-tests; otherwise, we used Wilcoxon signed-rank tests. To avoid excessive loss of statistical power while addressing multiplicity, we grouped related gait variables into six biomechanical families (spatiotemporal means, spatiotemporal variability, asymmetry, hip, knee, and ankle kinematics) and applied the Benjamin-Hochberg False Discovery Rate (FDR) correction within each family, considering variables with FDR-adjusted p-values (q-values) ≤ 0.05 as statistically significant.

PCA is designed to reduce dimensionality by leveraging the mutual information between variables, thereby projecting the original multidimensional space onto a smaller set of orthogonal axes that capture the maximum variance. This is achieved through the eigenvalue–eigenvector decomposition of the correlation matrix, which identifies the optimal directions for data projection. These principal axes (components) are uncorrelated and represent independent patterns within the dataset, allowing for a more concise yet comprehensive description of the underlying data structure³³. Several references provide a complete mathematical description of this procedure^34,35.

The data was organized into four sets: two representing ND parameters and two representing SWL parameters. For each group, one data set included only spatiotemporal variables, while the other combined spatiotemporal and joint kinematic parameters. This hierarchical modeling approach was designed to explore the influence of spatiotemporal parameters on the PC solution and to assess how the inclusion of kinematic variables modifies this structure. Additionally, the PC model based on normal spatiotemporal parameters allows comparisons with previous studies that investigated PCA of elderly gait.

To determine whether sufficient correlation patterns justified PC modeling, the correlation matrix for each data set (encompassing all variables) was analyzed using Bartlett’s test. A variable was included in the PCA only if it showed at least one correlation of magnitude | r | ≥ 0.30 with another gait parameter. This threshold ensured that every retained variable contributed a meaningful portion of shared variance to the model¹⁷. Because PCA is based on Pearson correlation coefficients, whose validity depends on near-normal variable distributions, the SWL dataset was pre-processed by applying a logarithmic transform to the variability parameters and a square-root transform to the asymmetry parameters, thereby improving normality before component extraction³⁶.

Four PC models were generated, one for each data set, with varimax rotation applied to improve interpretability. The number of components to retain was determined using the Kaiser criterion (eigenvalues greater than 1) and by visually inspecting the scree plots^36,37. To assess the stability of the solution, we also tested models with one additional and one fewer component than suggested by the scree plot^17,36. The optimal solution was selected based on the following criteria: (1) all gait parameters should load onto at least one component, (2) minimal cross-loadings, and (3) the model should account for a substantial portion of the total variance¹². Loadings above 0.5 or below − 0.5 were considered significant contributors to component interpretation and labeling^{12,13,15–17}. In cases of significant cross-loadings, the highest loading determined the component to which the gait variable was assigned¹⁷.

Results

Gait characteristics in normal and walker-assisted gait

Participants exhibited significantly higher velocity, step length, and step width, along with lower mean step time, during ND compared to SWL. Variability parameters were significantly greater in walker-assisted gait, except for step velocity and step width. No significant differences were found in asymmetry measures. All hip and ankle kinematic parameters were significantly greater in walker-assisted gait, except for maximum dorsiflexion at swing phase (A5), which was higher during ND, and maximum dorsiflexion at stance phase (A3), which showed no significant difference. Knee joint parameters were significantly greater during ND (Table 2).

Table 2.

Spatiotemporal and kinematic parameters of ND and SWL.

Parameters	No Device		Smart Walker Locomotion		Raw p-values	q-values (BH-FDR)c
Mean	Mean (SD)	Min; Max	Mean (SD)	Min; Max
Step VEL (m/s)	1.19 (0.11)	0.96; 1.37	0.33 (0.05)	0.21; 0.38	< 0.001 b	= 0.0007
Step length (m)	0.66 (0.05)	0.59; 0.77	0.34 (0.06)	0.23; 0.43	< 0.001 a	= 0.0007
Step time (s)	0.56 (0.05)	0.49; 0.65	1.06 (0.22)	0.75; 1.48	< 0.001 a	= 0.0007
Step width (m)	0.14 (0.04)	0.08; 0.19	0.09 (0.05)	0.02; 0.17	< 0.001 b	= 0.0007
Step ST time(s)	0.65 (0.06)	0.57; 0.77	1.49 (0.34)	1.03; 2.11	< 0.001 a	= 0.0007
Step SW time(s)	0.46 (0.04)	0.42; 0.56	0.62 (0.15)	0.35; 0.98	< 0.001 b	= 0.0007
DS time (s)	0.20 (0.04)	0.14; 0.26	0.86 (0.28)	0.52; 1.37	< 0.001 a	= 0.0007
Variability
Step VEL (m/s)	0.06 (0.02)	0.03; 0.11	0.04 (0.02)	0.02; 0.08	< 0.03 a	= 0.023
Step length (m)	0.03 (0.01)	0.01; 0.04	0.04 (0.02)	0.02; 0.07	= 0.104b	= 0.104
Step time (s)	0.02 (0.01)	0.01; 0.03	0.15 (0.13)	0.05; 0.54	< 0.001 b	= 0.0012
Step width (m)	0.03 (0.01)	0.01; 0.05	0.02 (0.01)	0.01; 0.03	< 0.01 a	= 0.009
Step ST time(s)	0.02 (0.01)	0.01; 0.04	0.24 (0.22)	0.06; 0.79	< 0.001 b	= 0.0012
Step SW time	0.02(0.004)	0.01; 0.02	0.09 (0.11)	0.03; 0.43	< 0.001 b	= 0.0012
DS time (s)	0.02 (0.01)	0.01; 0.03	0.23 (0.22)	0.06; 0.80	< 0.001 b	= 0.0012
Asymmetry
Step length (m)	0.02 (0.04)	-0.05; 0.09	0.03 (0.06)	-0.06; 0.14	= 0.43	= 0.82
Step time (s)	0.001 (0.02)	-0.04; 0.03	0.006 (0.10)	-0.17; 0.18	= 0.82	= 0.82
Step SW time(s)	0.001 (0.01)	-0.01; 0.01	0.03 (0.14)	-0.41; 0.23	= 0.39	= 0.82
Step ST time (s)	0.001 (0.01)	-0.02; 0.01	0.02 (0.11)	-0.11; 0.35	= 0.71	= 0.82
Hip (°)
H1	28.38 (2.43)	25.08; 34.54	50.43 (5.46)	43.18; 61.93	< 0.001 a	= 0.0007
H2	22.75 (2.95)	18.47; 28.69	44.95 (5.55)	35.18; 57.23	< 0.001 a	= 0.0007
H3	-7.98 (3.12)	-14.11; -3.3	18.12 (7.41)	8.23; 31.66	< 0.001 b	= 0.0007
H4	-2.59 (3.21)	-7.94; 3.53	24.22 (8.55)	12.95; 38.57	< 0.001 b	= 0.0007
H5	32.91 (2.72)	29.26; 36.28	52.73 (6.01)	44.10; 64.09	< 0.001 a	= 0.0007
Knee (°)
K1	3.21 (4.79)	-3.37; 15.17	23.29 (15.9)	-1.68; 53.45	< 0.001 a	= 0.013
K2	11.45 (5.49)	-0.01; 18.64	21.14 (13.2)	-0.12; 44.89	= 0.01 a	= 0.04
K3	1.65 (3.22)	-3.78; 7.16	5.21 (6.65)	-3.55; 20.59	= 0.04 a	= 0.04
K4	33.27 (3.87)	23.89; 37.83	28.43 (4.70)	18.93; 38.51	< 0.01 a	= 0.011
K5	62.94 (3.88)	55.52; 67.11	51.00 (6.58)	39.51; 61.32	< 0.001 b	= 0.002
Ankle (°)
A1	-0.55 (3.17)	-7.29; 4.89	8.15 (7.62)	-2.64; 19.90	< 0.01 a	= 0.008
A2	-7.96 (2.11)	-12.4; -5.1	-1.23 (7.22)	-12.53; 9.58	< 0.01 a	= 0.008
A3	15.01 (2.37)	11.25; 20.80	17.16 (3.56)	11.56; 24.86	= 0.12a	= 0.12
A4	-4.25 (5.10)	-11.68; 6.06	1.87 (7.27)	-12.59; 11.2	< 0.01 a	= 0.008
A5	-17.44 (5.10)	-25.89; -5.2	-6.02 (7.43)	-16.31; 4.70	< 0.01 b	= 0.008

Gait parameters are non-transformed; VEL: Velocity; DS: Double Support; ST: Stance; SW: Swing;

^a Comparison by paired-samples t-test.

^b Comparison by Wilcoxon Signed-Rank test;

^c p-value adjustment (q-values) by the Benjamin-Hochberg False Discovery Rate (BH-FDR).

Principal component models of normal gait

All 18 spatiotemporal parameters had at least one correlation above 0.30 (positive or negative) and were included in the first model (PCA-ND). The optimal PCA solution comprised five components, explaining 86% of the total variance. These components were labeled as follows: Pace/Variability (30% of variance), Asymmetry (19%), Rhythm (18%), Base of Support (10%), and Step Length (9%). The parameters with the highest loadings for each component were step velocity variability (0.96; Pace/Variability), swing time asymmetry (0.94; Asymmetry), mean swing time (0.91; Rhythm), mean step width (0.87; Base of Support), and step length asymmetry (0.80; Step Length). All parameters had loadings ≥ 0.50 (positive or negative) in at least one component, and 15 variables showed communalities ≥ 0.80 (Table 3).

Table 3.

PCA of Spatiotemporal parameters of ND.

Labels	Parameters	PC1	PC2	PC3	PC4	PC5	Com.
Pace/Variability	Step velocity variability	0.96					0.93
	Stance Time variability	0.93					0.94
	Step Time Variability	0.90					0.88
	DS time variability	0.82					0.89
	Step Length variability	0.71					0.71
	Mean step velocity	-0.70					0.86
Asymmetry	Swing time asymmetry		0.94				0.91
	Stance time asymmetry		0.93				0.93
	Step time asymmetry		-0.92				0.88
	Step width variability		0.77				0.61
Rhythm	Mean swing time			0.91			0.92
	Mean step time			0.89			0.98
	Mean stance time			0.64			0.92
	Mean DS time			0.54			0.89
	Swing time variability			0.58			0.91
Base of support	Mean step width				0.87		0.80
Step length	Step length asymmetry					0.80	0.68
	Mean step length					0.59	0.90
Proportion of variance (%)		0.30	0.19	0.18	0.10	0.09
Cumulative variance		0.30	0.49	0.67	0.77	0.86

The second PC model (PCA-NDK) included 33 parameters (spatiotemporal and kinematics), with the best solution comprising seven components that explained 91% of the total variance. These components were labeled Knee Motion (20% of variance), Variability (19%), Asymmetry (13%), Ankle Motion (11%), Hip Motion (11%), Rhythm (11%), and Base of Support (6%). The parameters with the highest loadings in each component were K3 (0.93; Knee Motion), step velocity variability (0.93; Variability), swing time asymmetry (0.93; Asymmetry), A4 (0.88; Ankle Motion), H4 (0.87; Hip Motion), mean swing time (0.93; Rhythm), and mean step width (0.93; Base of Support). All variables loaded onto at least one component, and the communalities were consistently high, with most exceeding 0.85. The only exceptions were step length asymmetry and step width variability, which showed communalities below 0.70 (Table 4).

Table 4.

PCA of Spatiotemporal and kinematic parameters of ND.

Label	Parameter	PC1	PC2	PC3	PC4	PC5	PC6	PC7	Com
Knee kinematics/ Peak swing	K3	0.93							0.92
	K2	0.88							0.91
	K1	0.88							0.93
	A2	0.85							0.96
	K4	0.82							0.90
	K5	0.71							0.86
	A5	0.64							0.93
	H5	0.57							0.92
Pace/ Variability	Step velocity variability		0.93						0.93
	DS time variability		0.85						0.93
	Step length variability		0.84						0.89
	Stance time variability		0.83						0.97
	Step time variability		0.74						0.84
	Mean step velocity		-0.66						0.95
	H1		0.51						0.85
Asymmetry	Swing time asymmetry			0.93					0.93
	Step time asymmetry			0.93					0.93
	Stance time asymmetry			-0.93					0.92
	Step width variability			0.74					0.90
Ankle kinematics	A4				0.88				0.77
	A1				0.69				0.98
	A3				0.63				0.83
Hip kinematics	H4					0.87			0.82
	H3					0.84			0.95
	Step length asymmetry					-0.66			0.88
	H2					0.58			0.75
Rhythm	Mean swing time						0.93		0.92
	Mean step time						0.85		0.95
	Mean step length						0.60		0.98
	Swing time variability						0.58		0.96
	Mean stance time						0.57		0.79
	Mean DS time						0.51		0.99
Base of support	Mean step width							0.93	0.90
Proportion of variance		0.20	0.19	0.13	0.11	0.11	0.11	0.06
Cumulative variance		0.20	0.39	0.52	0.63	0.74	0.85	0.91

PCA-NDK revealed a clear separation between spatiotemporal and kinematic parameters. The spatiotemporal portion retained its original component structure, while the kinematic variables formed three additional components, each corresponding to a specific lower limb joint. Two interactions were observed between kinematic and spatiotemporal variables: H1 loaded positively onto the Pace/Variability component, and step length asymmetry loaded negatively onto the Hip kinematics component.

Principal component models of walker-assisted gait

The spatiotemporal PC model for walker-assisted gait (PCA-SWL) included 18 variables, yielding the best solution with five components that explained 92% of the total variance (Table 5). These components, structured differently from ND, were labeled as follows: Pace/Variability (34% of variance), Rhythm (21%), Swing Time (15%), Base of Support (12%), and Step Asymmetry (9%). The variables with the highest loadings were double support time variability (0.99; Pace/Variability), mean swing time (0.94; Rhythm), swing time asymmetry (-0.98; Swing Time), mean step length (0.64; Base of Support), and step time asymmetry (-0.89; Step Asymmetry). All parameters showed high communalities, with values ranging from 0.80 to 0.99.

Table 5.

PCA of Spatiotemporal parameters of SWL.

Labels	Parameters	PC1	PC2	PC3	PC4	PC5	Com
Pace/ Variability	DS time variability	0.99					0.99
	Step time variability	0.99					0.99
	Stance time variability	0.99					0.99
	Step velocity variability	0.88					0.92
	Mean DS time	0.74					0.85
	Mean step velocity	-0.68					0.99
Rhythm	Mean swing time		0.94				0.94
	Stance time asymmetry		0.87				0.91
	Mean step time		0.68				0.87
	Mean stance time		0.59				0.98
Swing phase dynamics	Swing time asymmetry			-0.98			0.96
	Step length variability			0.62			0.98
	Swing time variability			0.56			0.85
Base of Support	Mean step length				0.64		0.88
	Step width variability				-0.90		0.85
	Mean step width				-0.68		0.84
Asymmetry	Step time asymmetry					-0.89	0.92
	Step length asymmetry					0.67	0.80
Cumulative variance		0.34	0.21	0.15	0.12	0.09
Proportion of variance (%)		0.34	0.55	0.70	0.82	0.92

The spatiotemporal and kinematic PC model for walker-assisted gait (PCA-SWLK) included 33 variables, yielding the best solution with six components that explained 91% of the total variance (Table 6). These components were labeled as follows: Knee and Ankle Stance Kinematics/Rhythm/Pace (25% of variance), Hip and Knee Swing Kinematics (20%), Variability (17%), Step Dynamics (13%), Swing Time (8%), and Base of Support (8%). The variables with the highest loadings were K1 (0.96; Knee and Ankle Stance Kinematics/Rhythm/Pace), H5 (0.94; Hip and Knee Swing Kinematics), step time variability (0.95; Variability), mean swing time (0.92; Step Dynamics), swing time asymmetry (-0.96; Swing Time), and step width variability (0.86; Base of Support). Communalities were high across almost all variables, with most exceeding 0.80. Step time asymmetry was the only parameter that showed communality below 0.70.

Table 6.

PCA of Spatiotemporal and kinematic parameters of SWL.

Labels	Parameters	PC1	PC2	PC3	PC4	PC5	PC6	Com
Knee & Ankle stance kinematics/ Rhythm/ Pace	K1	0.96						0.98
	K2	0.95						0.98
	A1	0.92						0.89
	A3	0.87						0.86
	A2	0.84						0.92
	Mean DS time	-0.75						0.96
	Mean stance time	-0.71						0.97
	K3	0.68						0.97
	Mean step time	-0.65						0.97
	A5	0.59						0.82
	Mean step velocity	0.56						0.95
Hip kinematics / Knee Swing kinematics	H5		0.94					0.95
	H1		0.92					0.92
	H3		0.88					0.97
	H2		0.88					0.84
	H4		0.83					0.98
	K4		-0.73					0.84
	Step length asymmetry		-0.65					0.78
	K5		-0.62					0.85
Variability	Step time variability			0.95				0.99
	Stance time variability			0.95				0.99
	DS time variability			0.95				0.98
	Step velocity variability			0.94				0.97
	Step length variability			0.61				0.82
Step dynamics	Mean swing time				0.92			0.90
	Stance time asymmetry				0.91			0.93
	A4				-0.72			0.89
	Mean step length				0.65			0.91
	Step time asymmetry				-0.64			0.69
Swing time	Swing time asymmetry					-0.96		0.96
	Swing time variability					0.61		0.85
Base of support	Step width variability						0.86	0.85
	Mean step width						0.58	0.83
Proportion of variance		0.25	0.20	0.17	0.13	0.08	0.08
Cumulative variance		0.25	0.45	0.62	0.75	0.83	0.91

Discussion

This study addressed the differences in gait domains of unimpaired elderly subjects between ND and SWL. We had hypothesized that walker-assisted gait would alter spatiotemporal and kinematic parameters compared to normal walking, while PCA would provide an exploratory framework to identify latent gait domains. The initial pairwise comparisons between ND and SWL indicated that SWL was characterized by slower pace, shorter and narrower steps, increased temporal parameters and altered joint kinematics, most notably at the hip joint, that sustained increased flexion posture throughout the gait cycle when compared to ND. This slower pace and shorter steps are a result of the UFES vWalker’s internal speed limit of 0.4 m/s, which is a design characteristic of rehabilitation robots in general aimed at providing a more controlled walking pattern for rehabilitation^38–40. Furthermore, the narrower steps are considered an expected outcome with the use of assistive devices, as a wider step width is a common mechanism used to increase walking stability. The increased stability provided by the device’s forearm support likely allowed participants to reduce their step width³⁰.

The increased hip flexion likely resulted from the hunched posture adopted to grip the smart walker, which positions the trunk and pelvis in a more flexed alignment relative to the device. Such postural adjustments may redistribute loads across the lower limb joints, alter the dynamics of the stance and swing phases, and influence the emergence of novel principal components when kinematic variables are considered. PCA was performed hierarchically first using only spatiotemporal variables, followed by a second round that incorporated sagittal kinematic parameters of the lower limb joints. The results demonstrated that gait variables correlated differently between the two walking modalities, forming unique patterns that resulted in distinct principal components, particularly when joint kinematics were included in the models.

Principal component analysis of the spatiotemporal parameters

Both ND and SWL yielded five principal components, with SWL explaining slightly more variance than ND (92% vs. 86%). The initial PCA-ND included 18 spatiotemporal variables. The first component was characterized by the interaction between step velocity and variability, which loaded inversely onto the component. This finding indicates a negative correlation between these parameters, supporting the findings of previous studies^10,12,41. It was hypothesized that age-related changes in variability measures are largely influenced by walking speed, as slower walking may disrupt the temporal automaticity of gait, leading to reduced step-to-step consistency¹⁰. The decline in automatic stepping mechanism contributes to an increased risk of falls due to poor foot placement or insufficient postural control⁴². In PCA-SWL, the first component was also associated with Pace/variability and exhibited similar loading patterns to those observed in PCA-ND. Temporal variability parameters, particularly double support time variability, demonstrated the strongest correlation, reflecting the significantly higher step time and phase time values observed in this modality of assisted gait. However, we hypothesize that these patterns may result from the interaction between the user and the device during walking, rather than an intrinsic decline in gait performance.

PCA studies on normal walking have shown that step velocity and mean step length often load together within the same component^14–17, suggesting that increasing step length is the preferred strategy for enhancing step velocity during normal gait of older adults. In contrast, our study found that step length loaded alongside step asymmetry onto a distinct component in normal walking, which appeared to reflect a feature of the postural control domain¹². This finding suggests that the Step length component observed in our analysis may reflect variations in the sagittal portion of the subject’s base of support.

In previous PCA studies, step width and step width variability consistently loaded together onto the Base of Support component of ND, showing a negative correlation^14,15. In contrast, our findings revealed that the Base of Support during ND was defined solely by mean step width. Step width variability, however, loaded onto the Asymmetry component alongside temporal asymmetry parameters, potentially reflecting a compensatory strategy to address asymmetries in stance and swing time. Greater step width variability in older adults may indicate impaired dynamic balance control during walking in an attempt to control the center of mass within their base of support¹⁰.

For walker-assisted gait, the Base of Support component was defined by three variables: mean step length on the positive side, and mean step width and step width variability on the negative side. This relationship appeared to capture both the sagittal and mediolateral dimensions of the participants’ base of support during assisted walking. Interestingly, contrary to previous studies^14,15, mean step width and step width variability were positively correlated in this context. This feature may be attributed to the narrower mean step width and reduced step width variability observed during walker-assisted gait. We believe this pattern likely resulted from the partial transfer of bodyweight support to the forearm supports, reducing the need for a wider mediolateral base of support to balance the subject’s center of mass. Patients with impaired balance control, often observed in neurodegenerative conditions, may benefit from motor control rehabilitation strategies that leverage the specific properties of the walker to enhance stability.

The Rhythm and Asymmetry components of PCA-ND were clearly defined, comprising temporal and asymmetric measures, respectively, consistent with previous findings^{12,13,15–17}. Both components accounted for nearly the same proportion of variance. In PCA-SWL, these component’s pattern differed considerably. The Rhythm domain remained well-defined, encompassing the expected temporal variables and stance time asymmetry. Given the significant increase in stance time during walker-assisted gait, its asymmetry influenced the rhythm of walking. Similarly, the Asymmetry component was clearly defined but included only two variables and accounted for the smallest proportion of variance. This finding suggests that step asymmetry has a reduced impact on the participant’s ability to walk, indicating that forward walking can be maintained even in the presence of asymmetrical steps. We hypothesize that this is due to the supportive properties of the walker, which could be leveraged to facilitate symmetry retraining, enhance independence, and boost motivation during rehabilitation.

Finally, we identified a novel component (Swing Dynamics) associated with the swing phase in PCA-SWL, which accounted for 15% of the variance. The structure of this component revealed a negative correlation between swing time asymmetry, swing time variability, and step length variability. This suggests a trade-off between swing time asymmetry and variability that influences step length variability. This component may have emerged from the combined effects of the partial bodyweight support provided by the walker and the participant’s inclined posture relative to the device. This finding may have significant implications for the prolonged use of walkers in gait rehabilitation, as the dynamics of the stance and swing phases are trained differently compared to unassisted gait. It remains unclear how this scenario might affect the patient’s weaning from the walker and their ability to achieve independent ambulation. One potential approach to address this issue is to gradually reduce hip flexion and the percentage of body weight transferred to the walker by adjusting the height of the forearm supports.

Principal component analysis of spatiotemporal and kinematic parameters

The PCA of gait in older adults, both with normal and pathological patterns, has traditionally focused on spatiotemporal parameters. This emphasis stems largely from the widespread use of the GAITRite^® instrumented walkway, which provides extensive spatiotemporal data but lacks the capability to capture lower limb joint motion. To overcome this limitation and gain deeper insights into the relationship between spatiotemporal and kinematic variables, we incorporated discrete sagittal plane parameters directly extracted from joint motion waveforms corresponding to specific gait events. This approach allowed us to explore how these events are represented in PCA through the correlations between variables and components. To our knowledge, this is the first study to investigate such interactions.

The PCA-NDK identified seven distinct domains, preserving the overall structure observed in spatiotemporal analysis with minor modifications. Additionally, kinematic variables from the knee, ankle, and hip joints formed three novel components, each corresponding to the dominant motion of a specific joint. Compared to PCA-ND, which extracted five components explaining 86% of the total variance, PCA-NDK required seven components to explain 91% of the variance. These results show that incorporating kinematic variables increased the number of extracted domains and slightly raised the proportion of explained variance, thereby providing a more detailed representation of gait component organization in elderly individuals.

Previous studies have demonstrated that the Central Nervous System employs a limited number of muscle groups co-activation (synergies) to generate a broad range of motor outputs, enabling stable locomotion across diverse environments⁴³. Five primary synergies, or modules, have been identified, each contributing to key sagittal-plane biomechanical functions such as bodyweight support, forward propulsion, and leg swing^44–46. The structure of the components identified through PCA, derived from correlations among spatiotemporal and kinematic variables, may reflect the outputs of such synergy-based strategies. While our study did not include EMG data, this interpretation suggests that future work combining PCA with electromyographic analysis could help to directly link biomechanical components with underlying neuromuscular control modules.

The first principal component derived from the PCA-NDK was primarily associated with kinematic variables that highlight the importance of knee mobility throughout the gait cycle, ankle motion during early stance and peak swing, and peak hip flexion during the swing phase. Given that reduced lower limb joint mobility is a critical concern in older adults and an independent risk factor for morbidity, disability, and mortality^47,48, this component underscores the necessity of preserving joint freedom of motion, particularly at the knee. In addition, components related to ankle and hip kinematics were primarily associated with variables describing joint motion during the stance phase and collectively accounted for 22% of the variance in the data. This finding reinforces the role of proximal-to-distal joint mobility in facilitating essential biomechanical subtasks such as leg deceleration, bodyweight support, and propulsion during toe-off^43,45.

The structure of the spatiotemporal domains remained largely consistent with that observed in PCA-ND, based solely on spatiotemporal parameters. However, variables that originally loaded onto the previous Step Length component were redistributed across the Hip Kinematics and Rhythm domains. Our results indicate that incorporating sagittal plane kinematics led to the reorganization of the least significant component of the spatiotemporal PCA while reinforcing the robustness of the Pace/Variability, Asymmetry, Rhythm, and Base of Support domains, as observed in previous models^12,15–17. This suggests that integrating kinematic data in the PCA provides a deeper and more detailed understanding of the biomechanics of ND.

Our last model incorporates spatiotemporal parameters and kinematic variables of walker-assisted gait (PCA-SWLK). The output can be categorized into three distinct sections. The first encompasses the interactions between kinematic and spatiotemporal variables, which define the first and second principal components, collectively accounting for nearly half of the total variance. Almost all kinematic parameters in the walker-assisted gait model were concentrated within these two components, suggesting that in walker-assisted gait, pace was primarily regulated by cadence and the early stance kinematics of the knee and ankle joints. Meanwhile, hip kinematics and knee swing parameters, which defined the second domain, appear to influence step length asymmetry.

The second section of the output corresponds to the Variability, Swing Phase Dynamics, and Base of Support. These components retained a structure similar to PCA-SWL. However, the importance of the Variability and Base of Support domains was reduced, and together, they accounted for 25% of the total variance. Notably, despite the inclusion of kinematic parameters, these domains preserved their original structure and robustness, reinforcing previous findings that identified them across different populations^14,15,17.

Finally, a novel component emerged in PCA-SWLK, which we labeled as Step Dynamics. This domain encompassed spatiotemporal variables related to step length, temporal asymmetry, and ankle flexion at the transition between the stance and swing phases (toe-off)³². We hypothesize that this component arose from the unique step profile generated by the interaction between the participant and the walker device at a self-selected speed.

Study limitations and clinical implications

This study presents certain limitations. Research investigating the effects of novel assistive technology devices often relies on small sample sizes, and our study faced the same constraint^49,50. However, previous findings have demonstrated that walker-assisted gait induces significant alterations in joint kinematics, with large effect sizes and statistical methods can achieve sufficient power to detect differences among these parameters even with a small sample size⁸. While PCA is generally regarded as a technique that benefits from larger samples, the required sample size also depends on factors such as the strength of correlations and the number of extracted domains. Specifically, when strong correlations are present, communalities remain consistently high (> 0.60), and a small number of well-defined components emerge, a smaller sample size can still yield reliable results³⁶. All these criteria were met in this study.

SWL involves distinct gait domains compared to ND, underscoring the fact that the UFES vWalker actively reshaped locomotor patterns rather than simply supporting existing ones. These device-induced adaptations, such as altered joint kinematics and temporal dynamics, must be carefully considered when designing future rehabilitation protocols. While such modifications may serve a compensatory role in the short term, prolonged use without proper progression may reinforce gait patterns that diverge from physiological norms. Therefore, gait training with the UFES vWalker should be strategically planned to promote gradual transition toward independent weight-bearing and conventional rehabilitation, especially in patients with recovery potential.

Although the results showed biomechanical changes in unimpaired individuals, it remains unclear how they interact with pathological gait in clinical populations. Future studies should extend this analysis to determine whether the device-induced adaptations compensate for or reinforce pathological motor patterns. For individuals with permanent neurological impairments, future developments should incorporate adaptive control strategies, adjustable support parameters and enriched environments, such as gamified tasks using virtual or augmented reality, encourage active engagement, enhance lower limb function and reduce fall risk.

Another limitation of this study is the absence of a comparison with a conventional walker. Although this would be clinically relevant, such analyses have not yet been systematically reported in the literature. Our results therefore provide a first step by establishing a normative reference for smart walker-assisted gait, while future studies should directly contrast conventional and smart walkers to clarify their distinct biomechanical effects.

Our investigation examined the characteristics of walker-assisted gait by setting the forearm support to a standard height⁵¹ and allowing self-selected walking speeds. Future research should investigate whether increasing the support’s height (and consequently reducing hip flexion angles) alters correlation patterns and principal components in a way that more closely resembles ND. This approach could serve as a strategy for gradual weaning from walker dependence.

Conclusion

This study offers new insights into the interaction between kinematic data and spatiotemporal parameters, as well as the biomechanical adaptations associated with smart walker-assisted gait. Combining joint kinematics and standard spatiotemporal metrics provided a more comprehensive perspective of gait in older adults by capturing additional features that are not fully represented by spatiotemporal parameters alone.

In No Device (ND), PCA identified distinct gait domains, and the inclusion of kinematic data further clarified specific motion patterns of the knee, ankle, and hip joints. Consistent with our initial hypothesis, the direct comparison between ND and Smart Walker Locomotion (SWL) revealed significant alterations in gait performance. SWL was characterized by slower gait, increased temporal variability, and modifications in the base of support, likely influenced by the interaction between the user and the device. Notably, this interaction is marked by an increased hip flexion posture throughout the gait cycle. Together, these features may explain the emergence of unique principal components observed in PCA-SWL (walker-assisted spatiotemporal only) and PCA-SWLK (walker-assisted spatiotemporal + joint kinematics), such as the Swing Phase and Step Dynamics domains.

These findings highlight that future biomechanical studies and rehabilitation protocols involving the UFES vWalker, or similar Smart Walkers (SWs), in clinical populations must account for the device-induced changes. Understanding and addressing these features are important for advancing both rehabilitative efficiency and the design of next-generation SWs.

Author contributions

Conceptualization: A.E., M.L., F.M, A.F; Methodology: A.E, M.L; Investigation: M.L, F.M.; Statistical analysis and modelling: A.E., Writing (original draft preparation): A.E; Writing (review and editing): M.L, F.M, A.F; Funding acquisition and resources: A.F.; Supervision: A.F.

Funding

This work was supported by Fundação de Amparo à Pesquisa e Inovação do Espírito Santo - FAPES (2025-3N254, 2022-C5K3H, 2021-07KJ2), Conselho Nacional de Desenvolvimento Científico e Tecnológico - CNPq (308155/2023-8, 402080/2025-4, 445343/2025-7), Financiadora de Estudos e Projetos - FINEP (0036/21, 2784/20), and Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES (Funding Code 001).

Data availability

Data will be made available on request. To obtain the data, please contact the corresponding author at anselmo.frizera-neto@ufes.br.

Declarations

Competing interests

The authors declare no competing interests.