Comparison of markerless and marker-based motion capture of gait kinematics in individuals with cerebral palsy and chronic stroke: A case study series

Background Three-dimensional (3D) motion analysis is an advanced tool used to quantify movement patterns in adults with chronic stroke and children with cerebral palsy. However, gold-standard marker-based systems have limitations for implementation in clinical settings. Markerless motion capture using Theia3D may provide a more accessible and clinically feasible alternative, but its accuracy is unknown in clinical populations. The purpose of this study was to quantify kinematic differences between marker-based and markerless motion capture systems in individuals with gait impairments. Methods Three adults with chronic stroke and three children with cerebral palsy completed overground walking trials while marker-based and markerless motion capture data were synchronously recorded. Time-series waveforms of 3D ankle, knee, hip, and trunk angles were stride normalized and compared. Root mean squared error, maximum peak, minimum peak, and range of motion were used to assess discrete point differences. Pearson’s correlation and coefficient of multiple correlation were computed to assess similarity between the time series joint angle waveforms from both systems. Results This study demonstrates that markerless motion capture using Theia3D produces good agreement with marker-based in the measurement of gait kinematics at most joints and anatomical planes in individuals with chronic stroke and cerebral palsy. Conclusions This is the first investigation to study the feasibility of Theia3D markerless motion capture for use in chronic stroke and cerebral palsy gait analysis. Our results indicate that markerless motion capture may be an acceptable tool to measure gait kinematics in clinical populations to provide clinicians with objective movement assessment data.


Background
Three-dimensional (3D) motion capture of human movement is a commonly used tool by biomechanists and clinicians to evaluate gait and mobility disorders.(1) Motion capture can objectively quantify segmental movement, joint range of motion, and spatiotemporal parameters. (2) Research in individuals post-stroke has extensively used 3D marker-based systems for measuring gait kinematics to better understand pathological movement patterns and inform rehabilitation strategies. (3)(4)(5)(6) This is because stroke is a leading cause of long-term disability, is widely heterogeneous, and survivors often have signi cant mobility impairment. (7) , (8) Similarly, clinical gait analyses use 3D kinematic outcomes as a key evaluation tool to guide surgical treatment and rehabilitation in children with cerebral palsy (CP). (9) Because of the unique presentations between heterogeneous patients, quantitative evaluation of movement patterns using 3D kinematics is critical for developing therapeutic interventions and evaluating function following corrective surgery. (9) Marker-based motion capture poses several practical limitations that hinder the adoption of such techniques in clinical settings.
As a result, there is limited accessibility to clinicians treating individuals post-stroke or with CP. (1,9) Access to laboratories, nancial cost and space requirements for equipment, and the need for highly trained personnel to physically palpate anatomical landmarks and operate equipment are noted issues.(1) When using marker-based motion capture systems, re ective markers are placed at anatomical landmarks, which requires minimal or tightly tting attire and signi cant time of the participant for preparation and system calibration. Marker-based preparations can be challenging for clinical populations with limited endurance and sensory sensitivities common in those with stroke or CP. (10)(11)(12) Precise palpation of anatomical landmarks for marker placement is critical for accurate kinematic analysis; however, children with CP can present with cognitive disorders that make physical palpation and following instructions di cult for marker-based motion capture. (11)(12)(13) Due to these limitations, clinicians often rely on observational gait analysis scales and physical examination, which are more subjective. (1) Markerless motion capture techniques have emerged as an alternative method for quantifying movement patterns that overcome many constraints of traditional marker-based systems. Theia3D (Theia Markerless, Inc, Kingston, ON), a commercially available  and sophisticated markerless motion capture technology, uses deep learning algorithms to identify human anatomical features and estimate 3D position. (14) Because markerless systems use multiple 2D video cameras to track human movement patterns, they do not need re ective markers placed on the skin. This substantially reduces the burden of marker placement and clothing restrictions on participants, removing practical barriers to kinematic analysis for clinicians. This technology can be used with relatively low-cost 2D video cameras compared to infrared cameras needed for marker-based systems and does not require computer programming knowledge, or data processing time with some manufactures, by the clinical team. Therefore, markerless technology could increase the accessibility of quantitative gait analysis to clinical populations by removing several practical barriers of marker-based systems. (15,16) Previous work has highlighted the accuracy and reliability of Theia3D's markerless technology in healthy populations, but veri cation has not yet been performed in clinical populations that present with gait and anatomical abnormalities. Investigations comparing an earlier version of this markerless software and marker-based motion capture from 30 healthy adults reported joint center estimates below 2.5 cm error for all major body joints (hip 3.6 cm).(16) In addition, lower limb segment angle errors were below 5.5º for all limbs (except transverse plane) with sagittal plane knee angles at 3.2º error in healthy adult gait. (15)(16)(17) Theia3D, has also proven reliable for measuring gait kinematics and spatiotemporal parameters from 55 healthy adults. (15)(16)(17) However, the accuracy of these systems in clinical populations, such as individuals post-stroke and children with CP, must be established for this technology to be translated to clinical settings and used to guide clinical treatment and rehabilitation.
The purpose of this study was to assess the accuracy of Theia3D markerless motion capture compared to a marker-based approach in the gait of those with CP and chronic stroke. We hypothesized that the kinematic analysis using Theia3D markerless motion capture would produce clinically acceptable results compared to gold standard marker-based motion capture. Speci cally, we examined joint angles of the ankle, knee, hip, and trunk-pelvis angles in the sagittal, frontal, and transverse planes during overground walking at self-selected walking speed.

Study design
The aim of this study was to assess the accuracy of Theia3D markerless motion capture compared to a marker-based approach in CP and chronic stroke gait. This study was a case series; therefore, convenience sampling was used. A case series approach was used to demonstrate the ability of Theia3D to appropriately capture kinematics among very different clinical presentations of those with CP and chronic stroke. See Table 1  Marker-based and markerless data were time-normalized and processed using Visual3D (v2021.03.2, C-Motion Inc., Germantown, MD). Marker-based data were ltered using a 4th-order dual-pass Butterworth lter with a lowpass cutoff of 6 Hz.(18) Segmental de nitions used for marker-based anatomical model creation are described in Appendix B. For markerless data, Visual3D automatically creates all segments upon loading Theia3D data. Theia3D software uses a customized Visual3D anatomical model with pre-de ned joint de nitions that are not modi ed by an end-user, therefore, note that there are inherent differences in joint mobility constraints between anatomical models. The Visual3D marker-based and markerless anatomical models were used to calculate lower limb joint angles identical to prior comparisons of these approaches with healthy adults.(16) Joint angles were decomposed using a Cardan sequence X (lateral), Y (anterior), and Z (vertical). Joint angle time-series waveforms are reported between two consecutive heel strikes as % Stride determined by the algorithm described by Zeni et al. (2008). (19) Following joint angle calculations, data were exported to Matlab (R2022a, The Mathworks, Inc., Natick, MA). Each joint angle time series was normalized by removing the best straight-line linear trend from the time series, using the 'detrend' function. The following discrete outcome variables were computed from joint angle data for both marker-based and markerless methods: root mean square (RMS), maximum peak (MAX), minimum peak (MIN), and range between the maximum and minimum peaks (RNG). The average of all trials was computed for each participant and used in statistical analysis.

Statistical Analysis
All data were normally distributed according to Kolmogorov-Smirnov tests. Therefore mean, standard deviation (SD), 95% con dence interval (95% CI), and paired t-tests were used for comparisons between marker-based and markerless methods for noted discrete outcome variables. Cohen's effect size and statistical power were calculated for each comparison using an a priori = 0.05. Root mean square error (RMSE), Pearson's coe cient of correlation (r), and coe cient of multiple correlation (CMC) were computed for time series waveform comparisons. All statistical analyses were performed in Matlab (R2022a, The Mathworks, Inc., Natick, MA).

Results
The purpose of this study was to compare the kinematics of children with CP and individuals with chronic stroke during overground walking measured with marker-based and markerless motion capture systems using Theia3D. Few signi cant differences were observed between the two systems, suggesting this markerless technology could be a viable alternative for use in clinical settings. When possible, we discuss our comparisons in the context of Minimal Clinically Important Difference (MCID) and Minimal Detectable Change (MDC). Values at or below MCID or MDC suggest that any difference between the two systems is not great enough to affect clinical decision-making.
Root mean square values of joint angle variables had only two signi cantly different outcomes between marker and markerless systems ( Table 2). Differences between the mean maximum joint angle measured by marker-based and markerless systems were minimal and below 4.5º (Appendix D). Differences between the mean minimum joint angle were also minimal and below 5º (Appendix E). Waveform comparisons from CMC and Pearson correlations had general trends of higher correlations in the sagittal plane joint angles than frontal and transverse (Appendix F).
At the trunk, mean differences of maximum and minimum joint angles between were less than 1º and not statistically signi cant amongst all planes. Discrete RMSE were between 2-4º for all planes, which resulted in large normalized RMSEs, particularly in the α sagittal and frontal planes ( Table 2). The total range of motion was also similar between the two systems with no statistical differences reported in our clinical populations.
At the hip, no signi cant differences were observed between systems in any anatomical plane for maximum joint angle of either limb or the minimum joint angle for the more affected limb. Mean differences in the sagittal plane ( exion/extension) were below an MDC of 4.69º and 4.01º shown for hip exion and extension in stroke during stance and swing, respectively.(20) Transverse plane (rotation) of the more affected limb had a signi cant difference of 3.5º (Cohen's d = 1.26, p = 0.017) for the minimum joint angle. The RMS for frontal plane motion of the hip was different between the two systems − 1.2º (Cohen's d = 1.42, p = 0.026). However, RMSE for the more affected side was within 2.72º, and RMSE for the less affected side was smaller than that of 2.6º in healthy adults,(16) indicating that these results in neurological populations are near or within typical bounds of those in healthy adults.
There were no signi cant differences between the two systems at the knee, on either the more or less affected leg, in any of the three planes. The mean difference in minimum and maximum joint angles were less than the MCID for both knee sagittal plane range of motion of 8.48º for the affected side and 6.81º for the unaffected side(21) as well as the 6.43º and 5.25º MDC of knee exion and extension during stance and swing in chronic stroke gait, respectively. (20) The frontal plane knee angle showed insigni cant difference, consistent with previous literature,(16) again demonstrating that the difference between the two systems in neurological populations is similar to that in healthy adults.
At the ankle, the mean difference in the sagittal plane for both the more affected and less affected limbs was well within MDC of 2.05º and 3.95º shown for ankle dorsi-and plantar exion in stroke in stance and swing for chronic stroke gait, respectively. (20) The RMS in the frontal plane for the less affected limb was signi cant (-1.3º [Cohen's d = 1.04] p = 0.043). However, the RMSEs for the sagittal, frontal, and transverse planes were all less than those of healthy adults found in previous work (6.7º, 8.0º, and 11.6º, respectively(16)). This indicates our results in neurological populations are within the bounds of those of healthy adults.
This subject was a Gross Motor Function Classi cation System (GMFCS) level 1, which implies minor issues with ambulation.
Waveforms are normalized to 100% of a gait cycle de ned as initial contact to initial contact. Shaded regions represent one standard deviation about the solid line which represent ensemble means.
This subject was 6.55 years post-stroke and used a right ankle foot orthosis and two-wheel walker. Waveforms are normalized to 100% of a gait cycle de ned as initial contact to initial contact. Shaded regions represent one standard deviation about the solid line which represent ensemble means.

Discussion
The mean differences between systems may be challenging to interpret due to differences in approach between the two systems. This general variability in joint angle consistency between the two systems may be due to inconsistencies in the calculation of segment angle in the global coordinate system. Therefore, if there is error in the global segment pose, the error will be ampli ed in joint angle measurement. In addition, while few of our transverse plane (rotation) comparisons were signi cantly different, absolute and normalized RMSE values were consistently high in this plane. This is evident when reviewing comparative waveforms (Figs. 1 and 2) as the markerless waveforms for CP patients showed consistently poor agreement. Several studies have highlighted di culties in accurately measuring tibial torsion and general internal/external rotation of the knee joint from marker-based approaches when compared to computed tomograph and goniometry. (22,23) We speculate that this could be related to foot deformities and smaller anatomical features of the pediatric CP foot proving di cult for the machine learning algorithm to identify.
There are several potential sources of error that could in uence joint angle measurement results in both systems. Marker-based systems are subject to operator error and inaccurate marker placement resulting in incorrect joint center estimation, as well as soft tissue or clothing artifact. These errors may lead to poor representation of anatomical landmarks and subsequent joint center locations, with ampli ed errors when calculating kinematics. (24) Marker placement error may contribute to up to 5º of error in lower limb joint angles.(25) Additionally, movement of soft tissue relative to bone during gait may translate up to 2.5 cm and rotate up to 8º, which introduces participant-speci c errors in joint angle calculations. (26) Markerless joint angle calculations may be affected by the training of the deep learning algorithm used to estimate pose, which is both a strength and weakness of the technology. This method also utilizes a frame-by-frame approach, which may introduce more noise than marker-based systems, but this noise may be overcome with the use of multiple measurements. (15) However, the sensitivity of the markerless motion capture system to participant-speci c characteristics like age, sex, ethnicity, health condition, anatomical deformities, orthosis use, assistive device use, and clothing, as well as environmental factors, such as lighting, have yet to be fully tested. (16) This study is not without limitations. This case study series included 6 total participants -3 children with CP and 3 individuals with chronic stroke -both of which are largely heterogeneous populations. However, we attempted to address this limitation by including participants with varying levels of gait function. Two stroke participants utilized ankle foot orthoses and assistive devices, which may have in uenced marker placement and marker visibility. In addition, the GMFCS III CP participant used a rearwalker

Conclusions
This study demonstrates the accuracy and feasibility of the current Theia3D iteration for use in kinematic gait analyses of those with CP and chronic stroke. Our data show the differences between the two systems primarily fall within the relevant clinical measurement error for gait kinematics. These minor differences across a wide range of ages, heights, and adiposity suggests that Theia3D could be a suitable replacement for marker-based tracking of the trunk segment in clinical applications. This technology may prove a valuable tool in clinical settings where the practicality of a markerless motion capture system is necessary. Further investigations are warranted to determine if markerless techniques are an acceptably accurate method of capturing kinematic data with a variety of clinical populations in cases where the practical bene ts of markerless data correction are warranted.       Figure 1 Time-series joint angle waveforms for a representative child subject with cerebral palsy.

Figure 2
Time-series joint angle waveforms for a representative adult subject with chronic stroke.