Comparison of gait deviation index (GDI) and gait variability index (GVI) measured by marker-based and markerless motion capture systems in children with cerebral palsy (CP)

Abstract

Background:

The Gait Deviation Index (GDI) is a metric clinicians use to assess overall gait pathology in children with cerebral palsy (CP) by comparing kinematic data to a normative sample. The Gait Variability Index (GVI) is a related metric that quantifies the variability in spatiotemporal variables during gait. Both the GDI and GVI have been verified using marker-based motion capture approaches, but video-based markerless motion capture has not been compared using these tools in children with CP.

Research question:

When considering limb impairment, what are the differences in GDI and GVI scores between Theia3D and marker-based motion capture for children with CP?

Methods:

Fifteen children with CP (GMFCS levels 1–4) and 24 typically developing children aged 6–18 were recruited for this study. Overground walking was performed at a self-selected pace while the pelvis, and lower limb kinematics were simultaneously recorded using both motion capture systems. Differences in GDI and GVI scores when considering the effect of system and limb impairment were analyzed using two-way repeated-measures ANOVAs.

Results:

Differences in GDI scores were observed, showing a 17.6% reduction when measured using Theia3D in the more affected limb (p < 0.05). Additionally, the more affected limb had a lower GDI score (4.3%) than the less affected limb only with Theia3D (p < 0.05). No differences were identified in GVI scores between systems or limb impairment. Differences in kinematic measurements were found in children with CP, including pelvic tilts, hip flexion/extension, hip rotation, and FPA, where RMSDs between the two systems exceeded 10°.

Significance:

We have confirmed that Theia3D is a suitable alternative for obtaining GVI metrics in children with CP and quantified differences in expected GDI scores. These findings can inform researchers and clinicians on measurement differences between marker-based and markerless methods in this clinical population.

1. Introduction

Cerebral palsy (CP) is a neuromuscular disorder caused by non-progressive brain damage in utero or during early development. It is the primary cause of movement disability in children [1], affecting 1.5–4 of every 1,000 live births worldwide [2]. Children with CP experience gait impairments, such as spasticity, muscle weakness, and reduced selective motor control[3]. These sensorimotor impairments limit functional capacity and increase the variability of kinematic and spatiotemporal parameters [4], [5]. Moreover, the complex and abnormal walking of children with CP can make it difficult to perform activities of daily living and participate in social situations.

Three-dimensional clinical gait analysis (3D-CGA) is considered the gold standard for investigating gait impairments in children with CP [6], [7] and is an effective tool for measuring, identifying, and understanding altered movement patterns [8]. 3D-CGA can facilitate treatment and rehabilitation plans and reduce surgical procedures and healthcare costs by using objective measurements to detect changes in joint motion, kinetics [9], [10], [11] and time-distance variables between normal and pathological gait [12]. However, interpreting 3D-CGA data is challenging due to its complexity due to various aspects of gait parameters. To address this, indices have been developed to condense information from 3D-CGA into a single value, aiding easier interpretation and clinical decision-making.

The Gait Deviation Index (GDI) and Gait Variability Index (GVI) are gait indices used to assess gait pathology and mobility impairments in individuals with CP [13], [14], [15]. The GDI was developed using 15 gait features obtained from 3D-CGA kinematics [14]. The GVI quantifies gait variability using nine spatiotemporal parameters [15]. Both indicies are interpreted by comparing a subject’s gait features to non-pathological data from a control group on a scale of 0–100, where every 10-point increase or decrease indicates one standard deviation (1SD) from the mean (e.g., a GDI of 80 is 2 SD from typical scores) [14]. A score of 100 or higher represents TD gait, while a lower score indicates increased gait impairment or variability [15], [16]. The GDI has shown face and construct validity in various populations, including healthy children, children with CP[17], [18], and adults with spastic CP [19]. Its distributional properties are consistent across healthy children and ambulant children with CP [17]. Additionally, the GDI can differentiate between affected and contralateral legs in individuals with hemiplegic CP [14]. Similarly, the GVI has demonstrated sensitivity in detecting changes in gait variability from childhood to adulthood[20] and is considered a valid assessment tool for individuals with mobility impairments [21], [22], [23], [24].

To quantify the GDI and GVI, kinematics and spatio-temporal gait parameters are typically acquired using a marker-based(MB) approach. This method uses infrared cameras and reflective markers on the subject’s skin to detect 3D marker position with sub-millimeter accuracy [25], [26]. Segmental position and joint angles are calculated from these data; however, soft tissue artifacts and manual marker placement can introduce errors between (or within) research centers [27], [28], [29], [30], [31]. Additionally, MB methods require a controlled environment that may affect natural movement [25], highly trained personnel [32], and time-consuming marker placement, which can be challenging for cognitively-delayed or touch-sensitive clinical populations [33].

Video-based markerless (ML) motion capture technology has emerged as an alternative to MB systems and addresses several limitations while introducing others. Markerless technology eliminates the need for markers, reducing tissue artifacts and minimizing data collection and processing time. It offers the advantage of capturing unencumbered movement and can be used more easily outside the laboratory [34]. Theia3D (Theia Markerless Inc., Kingston, ON) is a commercially available deep learning-based software designed for ML motion capture. It quantifies 3D pose estimation using multiple video cameras [35]. The latest version, Theia 2023, has increased the number of anatomical key points for pose estimation from 51 to 124, leading to tracking improvements in healthy subjects [36]. A prior Theia3D version had reliability and validity similar to MB systems in healthy adults, with slightly better inter-session consistency [34], [37]. Recent studies in pediatric patients [38] indicate comparable kinematic measurements between Theia3D and MB system, with most root mean square deviations (RMSD) below 6°. However, patients with foot issues or those using assistive devices may exhibit slightly higher (RMSD < 8°). While the current iteration of Theia3D can be reasonably compared to 3D-CGA kinematic waveforms, the impact of these kinematic differences on summary metrics used in clinical practice remains uncertain.

The primary aims of this study were to quantify differences in measuring GDI and GVI scores from lower limb kinematics in children with CP amongst two motion capture approaches. This study is an additional step in assessing the validity of Theia3D’s technology for assessing children with CP . We hypothesized that there will be no significant differences in GDI scores between ML and MB motion capture systems in children with CP. We also hypothesized there will be no significant differences in GVI scores between ML and MB motion capture systems in children with CP. A secondary aim was to quantify kinematic differences between moction capture approaches to assist with interpreting our results and provide a reference for future users of video-based ML technology.

2. Materials and Methods

2.1. Participants

This study involved 15 children with CP and 24 TD children (Table 1). Ethical approval was obtained from our local Institutional Review Board, and informed consent was obtained from parents, with child assent also obtained. Inclusion criteria for children with CP included ages 6–18, Gross Motor Function Classification System (GMFCS) [39] levels 1–4, adequate cognition, and no recent Botulinum toxin type A or surgical interventions. Typically developing children were included if they matched the age range and had no recent lower limb issues requiring hospitalization. Exclusion criteria for both groups were lower limb joint replacement or amputation, and current pregnancy.

Table 1:

Participant demographic, anthropometric, and clinical information.

	TD (N = 24)	CP (N = 15)
Demographics
	Mean ± SD
Age (years)	10.5 ± 4.1	13.4 ± 3.7
Weight (kg)	40.9 ± 15.7	41.1 ± 18.9
Height (cm)	146.9 ± 19.1	142.8 ± 16.6
Gender (Male: Female)	13 (54%):11 (46%)	7 (47%):8 (53%)
Clinical Information
Affected Side (Left:Right)	-	7:8
GMFCS	N (%)
Level 1	-	4 (26.7%)
Level 2	-	6 (40.0%)
Level 3	-	4 (26.7%)
Level 4	-	1 (6.7%)
Type of cerebral palsy	N (%)
Hemiplegia	-	3 (20.0%)
Diplegia	-	8 (53.3%)
Not specified	-	4 (26.7%)
Assistance/assistive device used for ambulation	N (%)
Wheel Walker	-	3 (20.0%)
Ankle Foot Orthosis (AFO)	-	4 (26.7%)
Supra Malleolar Orthosis (SMO)	-	1 (6.7%)
None	-	7 (53.3%)

TD is typically developing; CP is cerebral palsy; and GMFCS is gross motor function classification system.

2.2. Experimental Protocol

Data were collected in two laboratory spaces dedicated to gait assessments. In laboratory A, data were gathered from 23 subjects (19 TD/4 CP) using a combination of 10 MB cameras (10 Oqus, Qualisys AB, Gothenburg, SE) and 10 video-based cameras (8 Miqus Hybrid and 2 Miqus Video, Qualisys AB, Gothenburg, SE). In laboratory B, walking data were collected from 16 subjects (5 TD/11 CP) using 12 MB cameras (Kestral 4200, Motion Analysis Corp, Rohnert Park, CA, US) and 10 video-based cameras (Miqus Hybrid, Qualysis AB, Gothenburg, SE). For both laboratories, cameras were positioned around a capture volume (length: 5m, width: 2m, high: 2m). Data were recorded using Qualisys Tracking Manager (QTM) and Cortex (version 8.1, Motion Analysis Corp, Rohnert Park, CA, US) software. Marker and video-based data were synchronously recorded at 70 Hz, with the video data captured at 1080p resolution. Participants were instrumented with 54 reflective markers placed on their trunk, pelvis, and both thighs, shanks, and feet (Appendix A). Coordinate systems for pelvic, thigh, shank, and foot segments were constructed based on these landmarks. A static calibration trial was conducted with participants standing in a T-pose on the ground in the center of the walkway. Subsequently, participants completed five barefoot overground walking trials at their self-selected speed on a 10m flat, obstacle-free walkway. If needed, participants could use their own assistive devices, such as anterior or posterior wheel walkers.

2.3. Data Analysis

Synchronized motion capture data from both systems were processed separately. Marker data were tracked and labeled in QTM and Cortex, while video data were processed using Theia3D (v2023.1.0.3161(patch 4), Theia Markerless Inc., Kingston, ON). Marker data were exported to generate skeletal models in Visual3D (V3D) (version 2023.07.2 Student, C-Motion, USA). A 4^th order bi-directional Butterworth filter with a lowpass cutoff frequency of 6 Hz was applied to filter the marker data [40]. The MB model was created based on default V3D settings for joint constraints. Theia3D settings included Render Smooth IK, 3-DOF Knee, 6-DOF Feet, and a GCVSL Filter Cutoff Frequency of 5 Hz. Time normalization (0–100%) over the gait cycle was determined based on foot-strike to foot-strike events detected using an automated tool in V3D [41]. Lower limb joint kinematic and spatiotemporal gait parameters were calculated using V3D pipelines. Custom MATLAB code (version R2021a, Natick, MA) extracted joint angles at 2% increments throughout the gait cycle to calculate GDI and determine means and standard deviations for each spatiotemporal parameter for subsequent GVI analyses.

Gait Deviation Index (GDI) Calculation

A GDI was derived from four barefoot strides of both limbs of children with CP. For comparing GDI scores, baseline data of both legs from 24 TD children recorded in our laboratory were used. This included pelvic and hip angles (in all three planes), knee flexion-extension, ankle dorsiflexion-plantarflexion, and foot progression angles (FPA) based on established methods [14].

Gait Variability Index (GVI) Calculation

The GVI was calculated by analyzing nine spatiotemporal gait parameters: step length (cm), stride length (cm), step time (s), stride time (s), swing time (s), single support time (s), double support time (s), velocity (cm/s), and standard deviations (SDs) of each parameter referencing TD data from both systems recorded in our laboratory. The computation of GVI values involves using an Excel spreadsheet created by Gouelle et al.[15].

2.4. Statistical analyses

Statistical analyses were performed using RStudio (version 2022.12.0). Descriptive statistics, including means and standard deviations, were used to compare demographic and anthropometric characteristics. Differences in GDI and GVI outputs between the two motion capture systems when considering impairment of limb assessed were determined using a two-way repeated ANOVA (2×2 factorial of system: Theia3D and MB, and limb impairment: less and more affected). Root-mean squared differences (RSMD) are also reported for waveform comparisons between motion capture systems but were not subjected to statistical testing and are provided for informational purposes.

3. Results

3.1. Gait Deviation Index (GDI) in CP

There was a main effect of system (F (1, 56) = 4.46, p < 0.05, η_p² = 0.074 ) and limb impairment (F (1, 56) = 4.37, p < 0.05, η_p² = 0.072), with no interaction effects for GDI outcomes. GDI scores calculated from the MB system were consistently higher than those from Theia3D in less-affected and more-affected limbs (Figure 1). Marker-based GDI scores were 74.0 ± 11.1 and 71.0 ± 12.5, compared to Theia3D GDI scores of 70.9 ± 12.8 and 60.4 ± 13.7 for the less-affected and more-affected limbs, respectively.

Figure 1:

Violin plots illustrating GDI distributions in children with CP, grouped by less-affected and more-affected limbs. MB: marker-based, ML: markerless system. From top down, a * indicates a main effect of impairment and system at the p < 0.05 level.

3.2. Gait Variability Index (GVI) in CP

Results indicate no differences in GVI scores between system (F (1, 56) = 1.370, p = 0.247) or limb impairment (F (1, 56) = 1.179, p = 0.282) (Figure 2). GVI scores were higher when measured using Theia3D compared to the MB system in both less-affected and more-affected limbs (Theia3D: 84.67 ± 7.96 vs MB: 81.33 ± 9.02 and Theia3D: 81.55 ± 9.55 vs MB: 78.97 ± 12.12, respectively).

Figure 2:

Violin plots illustrating GVI value distributions in children with CP, grouped by less-affected (Left) and more-affected limbs (Right). MB: marker-based, ML: markerless system.

3.3. Joint kinematic differences between systems

For the secondary aim of this work, the TD group has the smallest offsets, while the more affected limb in children with CP exhibits the largest offset between motion capture systems. In all three planes, the smallest waveform offset was observed in the frontal plane, with RMSD below 5°, except for hip abduction-adduction in children with CP. In the sagittal and transverse planes, the ankle shows the smallest RMSD between motion capture systems in both TD and CP groups, while the hip joint exhibits the most significant offset between motion capture systems in TD children. Additionally, FPA has the highest RMSD between motion capture systems in children with CP (Figures 3 – 4).

Figure 3:

Mean ± 1 SD of joint kinematics used for GDI computation across TD subjects (top), less-affected limbs of children with CP (middle), and more-affected limbs of children with CP (bottom), with waveform RMSD values indicated in each graph.

Figure 4:

RMSD values are reported for TD and children with CP by more-affected and less-affected legs when comparing markerless kinematic waveforms to marker-based data. AB: Abduction, AD: Adduction, DF: Dorsiflexion, PF: Plantarflexion, and FPA: Foot progression angle.

4. Discussion

This study aimed to compare Theia3D markerless technology with MB approaches for determining GDI and GVI scores in children with CP. Our findings indicate that Theia3D yielded lower GDI scores compared to MB data, and GDI scores varied among assessed limbs with Theia3D but not with MB data. However, there were no differences in GVI scores between systems or among assessed limbs. Therefore, caution is advised when interpreting GDI scores in children with CP using different technologies. Conversely, consistent GVI scores suggest that Theia3D and MB systems are comparable in detecting variations in parameters contributing to gait variability across impairment levels.

The GDI scores were consistently lower with Theia3D than the MB approach, showing an 18.0% decrease in the more affected limb (Figure 1). This suggests Theia3D may be more sensitive to impaired gait kinematics than MB approach, particularly in hip rotation. The observed difference suggests challenges in accurately detecting kinematic parameters for GDI calculations (Figure 3 – 4), emphasizing the need for careful interpretation when comparing GDI scores from different technologies. Moreover, differences in joint definitions between systems may lead to undesirable kinematic crosstalk, where conflicting interpretations of joint motion occur due to misaligned coordinate systems. In MB approaches, joint angles are calculated in V3D using markers placed on the subject’s body to identify segmental definitions and anatomical landmarks. In contrast, Theia3D employs a custom skeletal model for analysis in V3D, which cannot be modified by the end-user. These variations in joint definitions result in differences in joint mobility constraints compared to other anatomical models [42], complicating the assessment of motion patterns and undermining interlab consistency for clinical decision-making. Furthermore, kinematic disparities may result from system configuration, joint angle calculation algorithms and inherent limitations [43], [44]. Marker-based approaches are affected by marker placement variations, soft tissue artifacts, and joint position errors [45], [46]. In contrast, Theia3D’s accuracy in measuring kinematics is influenced by factors related to its algorithm, beyond end-user control. Biases like subject characteristics, age, gender, health status, anatomical deformities, and scene lighting warrant further exploration in training data for machine learning and neural networks. Furthermore, given the focus on a clinical pediatric population, our kinematic findings align with a study conducted in pediatric patients with gait impairment, including CP subjects. They subtracted the mean value across the gait cycle (RMSDoffset) for their < 6° statement [47]. However, they used a different version of Theia3D (2021.2.0.1675), which complicates comparison. We strongly advocate reporting the specific version of Theia3D used, as changes can impact calculated kinematics (from our experience in different populations).

In contrast to GDI outcomes, GVI scores showed no difference between motion capture technologies, regardless of limb impairment, with scores below 100 for each limb indicating the capability of both systems to quantify deviations from TD walking patterns [15], [16]. The agreement between systems can be attributed to comparable measurement of spatio-temporal parameters, aided by camera settings such as high resolution (1080p) and adequate coverage of the capture volume. Simplifying the measurement of gross foot motion may enhance machine learning tasks in Theia3D, making it comparable to MB approaches. Our findings support Theia3D’s effectiveness in assessing spatiotemporal gait parameters in healthy young adults during treadmill walking. Hence, we recommend that users can reliably measure spatiotemporal gait parameters in children with CP using at least version v2023.1.0.3161 (patch 4) of Theia3D. In addition, the lack of significant GVI differences among limb assessments in our study may stem from participant characteristics, with over half demonstrating good mobility (GMFCS 1–2) and a majority diagnosed with diplegia (53.3%). Unlike hemiplegia, which typically shows more distinct limb differences, diplegia tends to exhibit less pronounced variations [48]. Previous studies have indicated hemiplegic patients often demonstrate a shorter step time on the affected side during walking in 96% and 87% of cases [49]. Therefore the predominance of diplegia in our sample may have led to the lack of significant variations in observed gait parameters. Additionally, approximately 20% of participants used a wheel walker during walking, which could have positively influenced their mobility and minimized observable differences in gait variability. However, this raises questions about the effectiveness of GVI in identifying impairments in clinical populations with lower gait variability, as seen in conditions like Parkinson’s disease, where subtle variations may not be captured by the index [16], [50], [51].

Our study has limitations, including the small sample size in both groups. This study utilized reference document from Schwartz et al. (2008) [14], but we supplemented it with our own sample of 24 TD children for the control population. This step was performed to control for differences in segmental definitions between laboratories, as our approach calculated foot-progression-angle differently (personal communication). However, the limited number of TD children may compromise our ability to detect subtle differences in calculating GDI and GVI scores. Additionally, the small sample size of children with CP (GMFCS 1–4) is inadequate for categorizing them into different GMFCS levels. Future research should involve a larger and stratified sample based on GMFCS levels, explore the impact of assistive devices such as AFOs on walking [38], and thoroughly analyze the offset in joint angle measurements between the two systems. Nevertheless, researchers and clinicians can enhance their assessment selection by gaining a deeper understanding of the discrepancies in measuring gait indices between these systems through our findings.

5. Conclusion

This study compared Theia3D markerless technology with traditional MB approaches in determining GDI and GVI scores in children with CP , considering limb impairment. Our findings showed that GDI scores were influenced by both the system used and the level of impairment. Theia3D consistently yielded lower GDI scores than the MB system, with more affected limbs scoring lower than less affected ones. These discrepancies may be due to variations in kinematic measurements between the systems, emphasizing the need for cautious interpretation, especially in cases of increased impairment. In contrast, GVI scores were not different between systems or among limbs assessed. The similarity in GVI scores suggests a comparable ability between Theia3D and the MB approach in measuring spatiotemporal gait parameters, regardless of impairment. These insights can inform researchers and clinicians in their selection of assessment tools, ultimately leading to more precise diagnoses and tailored treatment plans for individuals with gait disorders. Future studies may further refine these methodologies, potentially enhancing the accuracy and applicability of markerless motion capture systems in clinical settings.