Inter‐Session Repeatability of a Smartphone‐Based Markerless System to Measure Segment Length and Lower‐Limb Kinematics in Patients With End‐Stage Knee Osteoarthritis Authors: Chunping Ye, Renxin Ding, Junqing Wang, Yong Nie, Kang Li Published: 2026 ABSTRACT ABSTRACTBackgroundWhile kinematic analysis is crucial to characterize multi‐planar functional deficits in end‐stage knee osteoarthritis (OA), traditional laboratories impose severe physical burdens on patients. In‐field markerless systems like OpenCap offer an accessible alternative, yet their repeatability in this population remains underexplored. The objective of this study was to assess the inter‐session repeatability of OpenCap for quantifying segment lengths and lower‐limb kinematics during gait and sit‐to‐stand tasks in end‐stage knee osteoarthritis (OA) patients.MethodsTwenty‐three patients with end‐stage knee OA scheduled for total knee arthroplasty performed gait and sit‐to‐stand tasks while wearing street clothes across two visits. Segment lengths obtained from static calibration were compared between sessions, while waveform repeatability of lower‐limb joint angles was assessed using root mean square deviation (RMSD) and intraclass correlation coefficient (ICC). Additionally, the standard error of measurement (SEM) and minimal detectable change (MDC) were calculated for discrete kinematic parameters.ResultsThe segment lengths exhibited moderate to excellent repeatability between sessions (mean ICC: 0.79). Measurement bias was generally minimal (< 0.5 cm, except for the upper arm), with the overall 95% limits of agreement ranging from −4.51 to 3.98 cm. For gait, the average RMSD and ICC values were 4.0° (SD: 0.9) and 0.82 (SD: 0.07), respectively, whereas for the sit‐to‐stand task, these values were 5.2° (SD: 1.1) and 0.69 (SD: 0.08), respectively. The average SEM and MDC values for gait were 2.9° and 8.0°, and for sit‐to‐stand, 3.0° and 8.4°, respectively.ConclusionsOpenCap demonstrates moderate to good inter‐session repeatability for static anatomical measurements and gait kinematics in patients with end‐stage knee OA, showing promise as a low‐burden clinical assessment tool. However, the higher variability observed during the sit‐to‐stand task necessitates cautious, task‐specific interpretation, underscoring the need for further improvements before expanding its use across broader functional assessments. Introduction Knee osteoarthritis (OA) is a prevalent degenerative joint disease that significantly impairs mobility and quality of life in affected individuals. Accurate assessment of lower‐limb kinematics is critical not only for understanding the biomechanical alterations associated with knee OA but also for developing targeted rehabilitation strategies and monitoring disease progression [1, 2, 3, 4]. Fully characterizing the complex, multi‐planar kinematic deviations and dynamic compensations of end‐stage knee OA necessitates comprehensive 3D kinematic analysis [4, 5]. For patients awaiting total knee arthroplasty (TKA), establishing this 3D baseline provides critical dynamic metrics that surpass static radiographs in capturing complex kinematic behaviors, thereby enhancing clinical assessments and surgical planning [6]. Traditionally, motion capture systems employing skin‐mounted markers have been the gold standard for quantifying these 3D joint kinematics. However, these systems are often expensive, require specialized laboratory settings, and can be cumbersome for routine clinical use. In recent years, rapid advancements in computer vision and deep learning technologies have facilitated the emergence of markerless motion capture systems. While established commercial markerless systems (e.g., Theia) utilize specialized multi‐camera setups to deliver high‐fidelity kinematics, their high costs and extensive hardware requirements limit their portability and widespread clinical use [7]. To address these limitations, lightweight systems leveraging conventional devices (e.g., smartphones) have emerged as cost‐effective, portable alternatives for biomechanical assessments [8, 9, 10]. Among these emerging tools, OpenCap, a freely available platform, has drawn increasing attention due to its streamlined workflow, which significantly reduces setup time and lessens the testing and physical burden on participants (e.g., eliminating the need for skin‐mounted markers, wearing minimal clothing, or traveling to specialized laboratories) [8]. Recent investigations have demonstrated that the OpenCap system provides kinematic accuracy roughly comparable to other markerless solutions in healthy populations, with the strongest agreement typically observed in sagittal plane motions [8, 9, 11]. Moreover, in individuals with knee OA, it has achieved almost comparable accuracy to marker‐based systems in measuring sagittal plane hip and knee angles [10]. However, beyond measurement accuracy, evaluating the test–retest repeatability of this system is equally crucial. A study conducted on healthy individuals validated the repeatability of OpenCap in assessing kinematics during gait and sit‐to‐stand tasks, reporting repeatability almost comparable to traditional marker‐based systems across multiple variables [12]. However, the findings of this study are limited to healthy populations. Patients with knee OA face unique challenges, including altered gait patterns, pain‐related movement adaptations, and anatomical deformities [4, 13, 14, 15]. Our previous research has indicated that the measurement accuracy of OpenCap against the marker‐based system in individuals with knee OA is significantly reduced compared to that in healthy subjects [10]. Nevertheless, the high accessibility of OpenCap still holds substantial potential for transforming outpatient care. For instance, in routine clinical settings, the measurement of the foot progression angle can directly guide personalized gait retraining to reduce medial compartment loading [16]. Furthermore, evaluating sagittal metrics like peak knee flexion and functional range of motion allows clinicians to identify stiff‐knee gait, customize physical therapy, and monitor post‐surgical recovery [17, 18]. To safely utilize these clinical applications despite the aforementioned accuracy challenges, it is imperative to establish the system's inter‐session repeatability. Clinicians must rely on robust statistical thresholds, such as the minimal detectable change (MDC), to confidently ensure that observed kinematic alterations are attributable to true physiological changes rather than inherent measurement variability. The objective of the present study is to assess the repeatability of OpenCap for segment length measurements in a knee OA population, as well as the repeatability of lower‐limb kinematic test–retest measurements during gait and sit‐to‐stand tasks. Methods Participants Between July and December 2024, a total of 23 patients (5 males and 18 females) with end‐stage knee OA scheduled for TKA surgery were recruited. The mean age of the participants was 66.7 years (SD: 0.9), with an average height of 159.9 cm (SD: 7.4), weight of 62.7 kg (SD: 9.3), and body mass index (BMI) of 24.5 kg/m2 (SD: 3.1). The inclusion criteria were as follows: (1) a definitive radiographic diagnosis of knee OA with a Kellgren and Lawrence (K&L) grade of 3 or higher; (2) scheduled for TKA surgery; and (3) the ability to walk independently and perform sit‐to‐stand movements without assistive devices. Patients were excluded if they (1) had a history of lower‐limb surgery or (2) had other conditions that could potentially affect gait and sit‐to‐stand. The limb scheduled for TKA surgery was defined as the more affected side, while the contralateral limb was considered the less affected side. Among these patients, 14 were classified as K&L grade 3 and 9 as K&L grade 4. Ethics approval was obtained from the local ethics committee (Approval Number: 2024–1088), and informed consent was obtained from all participants. Data Collection The overall test–retest study design, as well as the data processing and reporting framework, were modeled after the methodology established by Horsak et al. [12]. Within 1 month prior to TKA, patients were invited to visit our laboratory twice, with an average interval of 9.1 days (SD: 2.7) between visits. During both visits, participants wore self‐selected clothing and footwear (Figure 1). Female participants were advised in advance to avoid wearing skirts. Furthermore, to minimize the impact of fabric movement on pose estimation, we utilized medical tape to secure and tighten the fabric around the lower limbs if participants wore loose pants. This approach allowed us to control for excessive clothing looseness while maintaining a low‐burden clinical scenario. At each visit, participants were instructed to perform self‐selected speed gait and sit‐to‐stand tasks. Data for these tasks were collected using the OpenCap application (version 1.0; Model Health Inc., USA) and sampled at 120 Hz using two smartphones (iPhone 14 Pro; Apple Inc., Cupertino, CA, USA) mounted on tripods. The smartphones were positioned 1.3 m above the ground, 3 m from the force plates, and angled approximately 35° relative to the direction of walking (Figure 2). To ensure consistent camera placement across all sessions, we utilized specific floor markers to strictly standardize the physical positions of the tripods. During the setup, data collection, and processing in OpenCap, we adhered to best‐practice guidelines described by Uhlrich et al. [8] and those available on OpenCap's website. Based on recommendations from our previous study [10], participants were instructed to walk along a 7‐m‐long walkway facing the cameras. The sit‐to‐stand task was performed on the force plates, with the chair and both feet positioned on separate plates. During this task, participants were required to cross their arms over the chest and face the camera while completing the movement. A 540 mm × 405 mm chessboard (each square measuring 45 mm × 45 mm) was used to calibrate the OpenCap system. For musculoskeletal model scaling, participants assumed a neutral pose as closely as possible despite end stage knee OA joint deformities, allowing us to utilize the standard neutral scaling procedure in OpenCap. Following the collection of static calibration data, each participant completed four trials of the gait task and four trials of the sit‐to‐stand task. The version of OpenCap used in this study was from July to December 2024. Clothing and footwear worn by participants during the test and retest sessions. This figure illustrates the uncontrolled, self‐selected street clothing conditions under which the markerless system was evaluated. This highlights the real‐world ecological validity and low‐burden nature of the experimental setup. Experimental setup of the smartphone‐based markerless motion capture system. (A) Schematic diagram illustrating the layout of the walkway and the configuration of the two smartphone cameras relative to the walking direction (blue arrow). (B) Photograph of the laboratory environment demonstrating a participant performing the overground walking task. The red boxes highlight the placement of the smartphones mounted on tripods at the end of the walkway. Data Processing Following data collection, OpenCap automatically uploaded the data to the cloud for processing, utilizing the data augmenter model v0.3 and the default musculoskeletal model. Prior to extracting individual cycles, the kinematic data were filtered using a fourth order 12 Hz low pass Butterworth filter [8]. For gait, we employed a kinematic‐based gait event detection algorithm to identify gait events based on marker coordinates [12, 19]. For the sit‐to‐stand task, the start and end of the movement were determined using hip flexion/extension angle [20]. Specifically, the sit‐to‐stand initiation was defined as the point when the hip flexion/extension angle exceeded ±3 standard deviations from the initial static position (calculated by averaging the first 20 data points). The sit‐to‐stand termination was defined as the point when the hip flexion/extension angle returned within ±3 standard deviations of the final position (calculated by averaging the last 10 data points). From each trial, a single gait cycle or sit‐to‐stand cycle was extracted and time‐normalized to 100%. For each participant, the joint angles from all trials of the more affected and less affected sides were averaged separately for each task, resulting in a session‐level mean joint angle for each side in each task. Data Analysis To evaluate the repeatability of static calibration, we measured the segment lengths of the trunk, pelvis, upper arm, forearm, thigh, shank, and foot from each participant's static trial using the key features extracted directly by the pose estimator. Segment lengths were defined as the distances between the proximal and distal ends of each segment. For both the upper and lower limbs, segments were further categorized into the more affected and less affected sides. Pearson's correlation coefficient (R), intraclass correlation coefficient (ICC A,1), and Bland–Altman plots with limits of agreement (LOA) were used to assess the correlation and agreement of segment lengths between the two sessions. To assess the repeatability of gait cycle and sit to stand cycle waveforms between the two sessions, root mean square deviation (RMSD) and ICC were calculated for both sides. Specifically, for each participant, the RMSD and ICC were first computed between the session level mean time series of the two visits across all normalized time points. This approach of utilizing trial averages aligns with recent recommendations by Horsak et al. [21], who demonstrated that averaging effectively mitigates the higher inter‐trial variability inherent to the OpenCap system. These individual metrics were then averaged across all participants to report the group level values. Paired‐sample t‐tests or Wilcoxon signed‐rank tests were used to assess the differences in RMSD and ICC between the more affected and less affected sides. ICC values were interpreted as follows: excellent repeatability (ICC > 0.90), good repeatability (ICC = 0.75–0.90), moderate repeatability (ICC = 0.50–0.75), and poor repeatability (ICC < 0.50) [22]. For discrete kinematic parameters, including joint range of motion and peak values within the gait or sit‐to‐stand cycle, the standard error of measurement (SEM) was calculated between the two sessions. SEM was derived from the square root of the mean square error term in a repeated‐measures analysis of variances [23]. Minimal detectable change (MDC) for these discrete kinematic parameters was then computed using the following formula: MDC=SEM×1.96×2 where SEM represents the standard error of measurement, 1.96 corresponds to the 𝑧‐score for a 95% confidence interval, and the factor 2 is included to account for the extra uncertainty arising from the use of difference scores between two measurement time points [24, 25]. Results Repeatability of Static Segment Lengths The majority of segment lengths in the static calibration model demonstrated moderate to excellent repeatability between the two measurement sessions, with an average ICC of 0.79, ranging from 0.60 (foot) to 0.91 (forearm) (Table 1). The mean Pearson's R was 0.80, ranging from 0.59 (foot) to 0.91 (forearm) (Table 1). Except for the upper arm, the measurement bias between the two sessions for all other segments was below 0.5 cm. Additionally, the 95% LOA across all segments ranged from −4.51 to 3.98 cm (Table 1 and Figure 3). TABLE 1Summary metrics for segment length between test and retest sessions.Segment lengthsLimbsMean (SD) difference p(*)Lower LOAUpper LOAPearson's RICC (A, 1)ICC LBICC UBTrunk (cm)−0.41 (2.1)0.359−4.513.700.740.740.480.88Pelvis width (cm)−0.05 (0.85)0.766−1.721.610.880.890.750.95Upper Arm (cm)MAS0.51 (1.16)0.047−1.772.780.820.800.560.91LAS0.68 (1.35)0.024−1.963.320.780.750.450.89Forearm (cm)MAS−0.12 (1.00)0.567−2.091.850.910.910.790.96LAS−0.19 (0.92)0.308−2.001.560.900.900.770.96Thigh (cm)MAS0.33 (1.72)0.359−3.033.710.800.800.590.91LAS0.30 (1.88)0.457−3.383.980.790.800.580.91Shank (cm)MAS−0.06 (1.67)0.876−3.363.250.870.870.730.94LAS−0.09 (1.89)0.830−3.793.620.810.810.600.92Foot (cm)MAS−0.10 (1.50)0.752−3.042.840.590.600.240.81LAS−0.11 (1.44)0.728−2.942.720.620.620.290.82Mean (cm)0.02 (0.33)0.800.79Abbreviations: LAS: less affected side; LB: lower bound; LOA: limits of agreement; ICC: intraclass correlation coefficient; MAS: more affected side; UB: upper bound. Summary metrics for segment length between test and retest sessions. Abbreviations: LAS: less affected side; LB: lower bound; LOA: limits of agreement; ICC: intraclass correlation coefficient; MAS: more affected side; UB: upper bound. Bland–Altman plots assessing the inter‐session repeatability of segment lengths. The x‐axis represents the mean of the segment lengths from the test and retest sessions, while the y‐axis represents the measurement difference between the two sessions. Solid black lines indicate the measurement bias (mean difference), and dashed red lines define the 95% limits of agreement (±1.96 SD). LAS: Less affected side; MAS: More affected side; SD: Standard deviation. Kinematic Repeatability During Dynamic Tasks Overall, the joint angle waveforms for both gait and sit‐to‐stand exhibited a relatively high degree of similarity between the two measurement sessions (Figures 4 and 5). The mean RMSD values calculated across all joint angles between the test and retest sessions were 4.0° (more affected side: 4.2°, less affected side: 3.7°) for gait and 5.2° (more affected side: 5.3°, less affected side: 5.1°) for sit‐to‐stand (Table 2). The average ICC values across all joint angles were 0.82 (more affected side: 0.80, less affected side: 0.84) for gait and 0.69 (more affected side: 0.68, less affected side: 0.69) for sit‐to‐stand (Table 2). Across all measured joint angles, individual RMSD values ranged from 0.9° to 10.0° (more affected side) and 1.0° to 11.1° (less affected side) during gait, and from 0.8° to 17.2° (more affected side) and 0.7° to 15.8° (less affected side) during the sit‐to‐stand task (Table 2). Correspondingly, individual ICCs exhibited joint‐ and plane‐dependent variation: sagittal hip and knee kinematics maintained high repeatability (ICCs ≥ 0.79), whereas extreme fluctuations in other degrees of freedom drove global minimums down to 0.02 (gait) and 0.04 (sit‐to‐stand) (Table 2). During gait, the more affected side demonstrated significantly higher mean RMSD and lower mean ICC values compared to the less affected side. However, no significant differences were observed between the more affected and less affected sides during the sit‐to‐stand task (Table 2). Lower‐limb gait kinematics during the test (Session 1) and retest (Session 2) sessions. The columns denote the more affected side (MAS) and less affected side (LAS). The rows display multi‐planar joint angles for the hip (sagittal, frontal, and transverse planes), followed by sagittal plane kinematics for the knee (flexion and extension) and ankle (dorsiflexion and plantarflexion). Solid lines represent the mean waveform averaged across all participants, while shaded areas indicate ±1 standard deviation from the mean. Group‐level mean root mean square deviation (RMSD) and intraclass correlation coefficient (ICC) values are provided within each corresponding subplot to summarize waveform repeatability. The x‐axis represents one full gait cycle (0%–100%), defined from the initial heel strike (0%) to the subsequent ipsilateral heel strike (100%). Lower‐limb sit‐to‐stand kinematics during the test (Session 1) and retest (Session 2) sessions. The columns denote the more affected side (MAS) and less affected side (LAS). The rows display multi‐planar joint angles for the hip (sagittal, frontal, and transverse planes), followed by sagittal plane kinematics for the knee (flexion and extension) and ankle (dorsiflexion and plantarflexion). Solid lines represent the mean waveform averaged across all participants, while shaded areas indicate ±1 standard deviation from the mean. Group‐level mean root mean square deviation (RMSD) and intraclass correlation coefficient (ICC) values are provided within each corresponding subplot to summarize waveform repeatability. The x‐axis represents the entire sit‐to‐stand‐to‐sit movement cycle (0%–100%), where 0% corresponds to the initiation of the sit‐to‐stand phase, and 100% corresponds to the termination of the stand‐to‐sit phase. TABLE 2Root mean square deviation (RMSD) and interclass correlation coefficient (ICC) between test and retest sessions.LimbsGaitSit‐to‐standRMSD (°)RMSD range (°)ICCICC rangeRMSD (°)RMSD range (°)ICCICC rangeHip flex/extensionMAS4.5 (1.6)2.2–7.40.95 (0.04)0.86–0.999.0 (3.2)3.7–17.20.94 (0.04)0.84–0.99LAS4.3 (2.0)1.4–9.00.96 (0.03)0.89–1.008.8 (3.2)4.7–15.80.94 (0.05)0.79–0.98Hip add/abductionMAS2.3 (0.9)0.9–5.30.86 (0.14)0.46–0.992.8 (1.5)0.8–7.80.46 (0.27)0.04–0.89LAS2.2 (0.8)1.2–5.10.89 (0.12)0.49–0.982.4 (0.9)0.7–4.70.50 (0.24)0.1–0.84Hip int/external rotationMAS4.1 (1.6) a 1.5–8.60.47 (0.24) a 0.02–0.903.0 (1.1)1.5–6.80.54 (0.28)0.06–0.91LAS2.9 (0.8)1.0–4.70.60 (0.19)0.15–0.963.3 (1.2)1.8–5.70.47 (0.27)0.05–0.94Knee flex/extensionMAS3.8 (1.8)1.3–9.60.97 (0.03)0.85–1.007.1 (3.1)3.2–13.60.97 (0.05)0.93–1.00LAS4.0 (2.1)1.7–11.10.97 (0.03)0.85–1.006.8 (3.1)2.3–13.50.97 (0.05)0.91–1.00Ankle dorsi/plantarflexionMAS4.8 (1.9)2.2–10.00.80 (0.13)0.51–0.984.4 (1.7)2.3–9.10.69 (0.17)0.28–0.92LAS4.2 (1.2)2.0–6.20.80 (0.11)0.57–0.964.0 (1.7)2.0–7.80.70 (0.24)0.09–0.95Ankle in/eversionMAS5.9 (2.6)2.3–10.00.74 (0.18) a 0.16–0.945.7 (3.0)1.9–11.30.50 (0.25)0.13–0.92LAS4.9 (2.2)1.8–10.30.82 (0.07)0.38–0.995.0 (1.9)2.2–9.40.55 (0.30)0.06–0.95Mean totalMAS4.2 (0.9) a 0.9–10.00.80 (0.08) a 0.02–1.005.3 (1.2)0.8–17.20.68 (0.09)0.06–1.00LAS3.7 (1.1)1.0–11.10.84 (0.07)0.15–1.005.1 (1.2)0.7–15.80.69 (0.11)0.05–1.00Mean4.0 (0.9)0.9–11.10.82 (0.07)0.02–1.005.2 (1.1)0.7–17.20.69 (0.08)0.05–1.00Abbreviations: LAS: less affected side; MAS: more affected side. a Indicates significant difference between the more affected side and the less affected side. Root mean square deviation (RMSD) and interclass correlation coefficient (ICC) between test and retest sessions. Abbreviations: LAS: less affected side; MAS: more affected side. Indicates significant difference between the more affected side and the less affected side. Measurement Error and Minimal Detectable Change During gait, the mean SEM and MDC for all joint angles were 2.9° (more affected side: 3.2°, less affected side: 2.6°) and 8.0° (more affected side: 8.8°, less affected side: 7.1°), respectively (Table 3). For the sit‐to‐stand task, the mean SEM and MDC for all joint angles were 3.0° (more affected side: 3.2°, less affected side: 2.8°) and 8.4° (more affected side: 8.9°, less affected side: 7.9°), respectively (Table 3). The MDC ranged from 4.3° to 14.5°, with the highest value observed in the sagittal hip range of motion during the sit‐to‐stand task (Table 3). TABLE 3Standard error of measurement (SEM) and minimal detectable change (MDC) between the test and retest sessions.LimbsGaitSit‐to‐standSEMMDCSEMMDCSagittal hip RoM (°)MAS2.97.95.214.5LAS2.67.35.013.8Frontal hip RoM (°)MAS2.77.42.05.6LAS2.36.32.05.6Transverse hip RoM (°)MAS2.77.43.18.7LAS1.85.02.56.8Sagittal knee RoM (°)MAS3.610.12.87.9LAS2.67.32.87.9Sagittal ankle RoM (°)MAS4.713.02.67.2LAS3.49.32.56.9Frontal ankle RoM (°)MAS4.011.14.712.8LAS2.87.73.39.2Peak hip extension (°)MAS2.67.22.46.6LAS2.87.72.46.6Peak hip flexion (°)MAS2.77.44.713.1LAS2.87.84.311.8Peak hip abduction (°)MAS1.74.71.85.1LAS1.64.31.74.8Peak hip adduction (°)MAS1.95.42.77.5LAS1.85.11.95.4Peak knee flexion (°)MAS3.610.12.98.2LAS3.28.82.87.7Peak ankle dorsiflexion (°)MAS4.813.43.710.3LAS2.98.03.08.2Peak ankle plantarflexion (°)MAS3.39.12.77.5LAS2.98.22.77.4Mean total (°)MAS3.28.83.28.9LAS2.67.12.87.9Mean2.98.03.08.4Abbreviations: LAS: less affected side; MAS: more affected side; RoM: range of motion. Standard error of measurement (SEM) and minimal detectable change (MDC) between the test and retest sessions. Abbreviations: LAS: less affected side; MAS: more affected side; RoM: range of motion. Discussion This study evaluated the inter‐session repeatability of OpenCap, a smartphone‐based markerless system, for quantifying segment lengths and lower‐limb kinematics during gait and sit‐to‐stand tasks in patients with knee OA. Our main findings indicate moderate to excellent repeatability for static segment lengths. During dynamic tasks, lower‐limb kinematics demonstrated good overall repeatability for gait and moderate repeatability for the sit‐to‐stand task, with sagittal hip and knee motions showing the highest reliability. Furthermore, the more affected limb exhibited significantly higher measurement variability during gait. Finally, we established MDC thresholds averaging approximately 8° across joints. Overall, these findings support OpenCap's utility for low‐burden functional assessments, provided that task‐specific, joint‐specific, and side‐specific variations are considered. Repeatability of Static Segment Lengths During the static calibration, segment length measurements overall demonstrated good repeatability, with an average ICC of 0.79 and an average Pearson's R of 0.80. These values are slightly lower than those reported by Horsak et al. [12] in healthy populations (ICC: 0.87; Pearson's R: 0.88). One potential explanation for this discrepancy is the difference in clothing conditions: our participants wore street clothes, whereas those in the study by Horsak et al. wore minimal clothing. Additionally, our study involved patients with advanced knee OA, some of whom exhibited severe knee flexion and varus deformities. These factors may pose challenges for OpenCap in accurately identifying joint centers. Despite these challenges, most segments demonstrated minimal mean measurement bias (less than 0.5 cm) between sessions. However, the wide 95% limits of agreement (spanning from −4.51 to 3.98 cm) reveal substantial per‐subject variability. Individual segment length errors of up to 4.5 cm can negatively impact the precision of subsequent inverse kinematics solvers. Therefore, while the system's automated scaling provides consistent anatomical measurements on average, it requires cautious interpretation when applied to single patients. Notably, the foot segment exhibited much lower repeatability (ICC: 0.60; Pearson's R: 0.59) compared to the findings of Horsak et al. (ICC: 0.89; Pearson's R: 0.89) [12], possibly due to our participants wearing street shoes rather than being tested barefoot. Kinematic Repeatability During Dynamic Tasks Beyond static segment lengths, reliable 3D kinematic analysis is essential for end‐stage knee OA patients to fully characterize their complex, multi‐planar deformities and dynamic compensations [4, 5, 6]. Given the clinical necessity of these dynamic metrics, our evaluation of the lower‐limb kinematic waveforms revealed moderate‐to‐good inter‐session repeatability at the group level. Specifically, during gait, the lower‐limb joint angle waveforms exhibited good repeatability across testing sessions, with an average ICC of 0.82. This finding aligns with results obtained using a commercial markerless system (i.e., Theia) in healthy older adults, which reported an average ICC of 0.83 for lower‐limb joint angles [26]. These comparable values suggest that OpenCap's repeatability in measuring lower‐limb joint angles during gait in an older population is comparable to that of established commercial systems. In contrast, the sit‐to‐stand task exhibited moderate repeatability, with an average ICC of 0.69, which is lower than that observed during gait. This reduction in repeatability may be attributable to the greater inherent variability of the sit‐to‐stand movement compared to gait [27]. Crucially, the observed joint‐ and plane‐dependent variation in individual ICCs highlights the current limitations of markerless pose estimation. While sagittal hip and knee kinematics remained robust, the tracking of non‐sagittal movements and distal ankle joints appears more susceptible to the severe multi‐planar deformities and compensations typical of end‐stage knee OA [4, 14, 15]. Consequently, while sagittal hip and knee metrics are reliable for clinical monitoring (e.g., tracking improvements in stiff‐knee gait), applying other kinematic outputs for precise single‐patient decisions (e.g., prescribing offloading orthotics based on frontal‐plane kinematics) requires caution. During gait and sit‐to‐stand tasks, the average RMSDs of lower‐limb joint angles were 4.0° and 5.2°, respectively. These values are higher than those reported in previous studies involving healthy individuals, which documented average RMSDs of 2.8° for gait and 3.3° for sit‐to‐stand tasks [12]. This discrepancy may be attributed to the combined effects of differences in attire (i.e., street clothing vs. minimal clothing) and disease‐related factors (i.e., knee osteoarthritis vs. healthy), such as joint deformities and abnormal movement patterns (e.g., stiff‐knee gait) [13, 14, 15]. Crucially, the wide range of individual RMSD values, spanning globally from 0.9° to 11.1° during gait and from 0.7° to 17.2° during sit‐to‐stand across both limbs, highlights significant per‐subject variability that is masked by group‐level averages. Furthermore, although both the current and previous studies observed higher RMSDs during the sit‐to‐stand task compared to gait, the disparity between the two tasks was more pronounced in the present study (1.2° vs. 0.5°). This finding may be related to the increased variability in the sit‐to‐stand movement among patients with knee OA, potentially stemming from quadriceps muscle weakness and subsequent compensatory strategies, such as asymmetric weight‐bearing or an exaggerated forward trunk lean [28, 29]. A notable finding in our study was that, during gait, the more affected side exhibited significantly lower repeatability compared to the less affected side, as evidenced by a significantly higher average RMSD and a lower average ICC. Conversely, in the sit‐to‐stand task, no significant differences in repeatability were observed between the more and less affected sides. This discrepancy may reflect biomechanical adaptations in patients with knee OA in response to specific task demands. In gait, the reduced repeatability on the more affected side could be due to increased movement variability associated with pain, instability, or compensatory strategies. Patients with knee OA often adopt altered gait patterns (e.g., reduced knee flexion, shorter stance times) to offload the more affected joint [4]. These adaptations may vary between sessions due to fatigue or fluctuating pain levels, leading to reduced consistency in kinematic measurements on that side. In contrast, the sit‐to‐stand task is a more constrained and shorter‐duration movement, which may encourage patients to adopt consistent compensatory strategies (such as a forward trunk lean) [30], thereby reducing the differences in repeatability between sides. Minimal Detectable Change and Clinical Considerations In addition to evaluating the repeatability of joint angle waveform, this study also computed the SEM and MDC for discrete kinematic parameters. To the best of our knowledge, this is the first report of MDC values for a markerless system in a patient population. Our results showed that, in patients with knee OA, the average MDC for gait and sit‐to‐stand tasks was 8.0° (range: 4.3°–13.4°) and 8.4° (range: 4.8°–14.5°), respectively. These MDC values provide critical thresholds for discerning true changes beyond measurement error, indicating that only changes exceeding these thresholds can be considered clinically significant [31]. Crucially, our data indicate that these MDC values should be applied in a side‐specific manner, as pathological conditions introduce greater measurement variability on the affected side. For instance, when utilizing OpenCap for post‐TKA rehabilitation monitoring, clinicians evaluating peak knee flexion during gait should apply a strict 10.1° statistical boundary for the more affected limb, compared to 8.8° for the contralateral limb. This precise side‐specific differentiation ensures that elevated pathological noise is not misinterpreted as clinical recovery, effectively preventing the overestimation of treatment efficacy. Our evaluation of the OpenCap system highlights the critical value of out‐of‐the‐lab gait analysis for end‐stage knee OA patients. Due to severe pain and mobility restrictions, requiring this population to travel to specialized facilities and undergo lengthy marker placement is often clinically impractical and physically exhausting. Portable smartphone‐based systems like OpenCap overcome these barriers by enabling low‐burden biomechanical evaluations directly in routine clinical settings. This accessibility facilitates longitudinal monitoring (such as evaluating the efficacy of a pre‐habilitation exercise program) and broadens the clinical reach of advanced movement analysis without exacerbating patient discomfort. Limitations This study has several limitations. First, the investigation exclusively enrolled patients with end‐stage knee OA scheduled for TKA. However, end‐stage patients typically present severe functional limitations and joint deformities [4, 13, 14, 15], and those in the early to middle stages with better functional and milder joint deformities might achieve higher repeatability than those in this study. Furthermore, daily postural variations during static calibration (e.g., due to fluctuating pain levels) could alter model scaling and reduce repeatability. Future studies may consider utilizing OpenCap's “any pose” scaling feature to mitigate this issue. Additionally, this study assessed repeatability under street clothing conditions (all participants wore long pants and either short‐ or long‐sleeved shirts). Although we utilized medical tape to secure excessively loose pants, our findings may not directly generalize to scenarios involving minimal clothing. Nonetheless, we believe that street clothing best represents the clinical setting in which OpenCap is likely to be used, as testing under these conditions minimizes both testing time and patient burden, thereby facilitating the application of OpenCap in large‐scale field studies. We calculated segment lengths using raw pose estimator keypoints rather than the scaled OpenSim model. While this isolates the algorithm's repeatability, future studies may evaluate the final OpenSim model parameters to account for biomechanical scaling constraints. Although OpenCap's repeatability in healthy populations is established [12], the lack of a concurrent control group precludes direct comparisons of measurement variability between healthy and pathological cohorts. Future studies should include healthy controls to quantify these population‐specific differences. Finally, our sample size (N = 23) restricts our ability to perform gender stratification (18 females, 5 males). Although this distribution reflects knee OA epidemiology, gender specific anatomical features may inherently increase biomechanical variability. Future large‐scale studies with gender stratification are required to disentangle these effects. Conclusion This study demonstrates that OpenCap exhibits moderate to good inter‐session repeatability for both static anatomical measurements and dynamic lower‐limb kinematics in patients with end‐stage knee OA. Inter‐session errors in lower‐limb joint angles during gait were generally below the clinically relevant threshold of 5 degrees, whereas errors for certain joint angles during the sit‐to‐stand task exceeded this threshold. These findings provide support for the potential utility of OpenCap in assessing gait function within clinical settings following interventions or rehabilitation for patients with end‐stage knee OA. However, caution is warranted when interpreting results from other tasks, such as sit‐to‐stand. Furthermore, attention should be paid to side‐specific differences when utilizing MDC values for data interpretation. Author Contributions Chunping Ye: conceptualization, methodology, investigation, validation, formal analysis, data curation, visualization, writing – original draft, writing – review and editing. Renxin Ding: conceptualization, formal analysis, software, validation, writing – original draft, writing – review and editing. Junqing Wang: conceptualization, data curation, writing – review and editing, formal analysis, visualization, writing – original draft. Yong Nie: writing – review and editing, project administration, supervision, resources, funding acquisition, conceptualization. Kang Li: writing – review and editing, resources, project administration, supervision, funding acquisition. Funding This work was supported by Advanced Materials ‐ National Science and Technology Major Project (No. 2025ZD0619901), National Key Research and Development Program of China (2023YFB4606700), 1.3.5 project for disciplines of excellence, West China Hospital, Sichuan University (ZYYC21004), 1.3.5 projects for Artificial Intelligence West China Hospital, Sichuan University (ZYAI24038), Sichuan Provincial International Science and Technology Innovation Cooperation (2025YFHZ0110), 1.3.5 project for disciplines of excellence‐Clinical Research Fund, West China Hospital, Sichuan University (2023HXFH024). Conflicts of Interest The authors declare no conflicts of interest.