Evaluating upper limb movement smoothness metrics derived from markerless 3D motion tracking in children with neurological disorders Authors: Nikolas Hesse, Jan Lieber, Hubertus J. A. van Hedel Published: 2026 ABSTRACT BackgroundClinical assessment of movement smoothness during reaching in children is often based on subjective, ordinal-scaled therapist ratings. Markerless 3D motion tracking using an RGB-D camera offers a promising avenue for more objective and fine-grained smoothness evaluation. This study investigated the use of a new markerless RGB-D-based tracking method, SMPLify-KiDS, for the quantitative assessment of upper limb movement smoothness in a heterogeneous population of children with neurological impairments.MethodsThirty-eight children with neurological diagnoses performed three reaching tasks from the Melbourne Assessment 2 (MA2) MA2 with their more affected arm while being recorded with an RGB-D camera. Upper limb kinematics were extracted using SMPLify-KiDS, and two widely used smoothness metrics—the log dimensionless jerk (LDJ) and the modified spectral arc length (SPARC)—were computed. The same movement sequences were independently scored by two therapists according to MA2 guidelines. Correlations between quantitative smoothness metrics and therapist scores were evaluated using Spearman’s rank correlation coefficient, and smoothness metric differences between ordinal therapist score levels were analyzed using non-parametric statistical tests for individual items and averaged values.ResultsQuantitative smoothness metrics showed moderate to strong correlations with therapist scores per item (for LDJ: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.666 \le \rho \le 0.836$$\end{document}; for SPARC: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.556 \le \rho \le 0.674$$\end{document}), and very strong to excellent correlations for averaged values (for LDJ: ρ =0.843, for SPARC: ρ =0.755). Both metrics revealed statistically significant differences between ordinal therapist score levels, and LDJ strongly correlated with task duration.ConclusionsSmoothness metrics derived from markerless RGB-D motion tracking are able to discriminate between clinically assigned smoothness levels in children with neurological impairments. LDJ showed closer alignment with therapist scores than SPARC, potentially due to its relationship with movement duration—a factor that also appears relevant to clinical perception of smoothness. These findings support the integration of markerless motion capture into clinical assessments to provide objective, quantitative, and fine-grained information on movement quality, complementing established therapist-based evaluations.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12984-026-02020-8. Background Children with neurological disorders, including cerebral palsy [1], traumatic brain injury [2], or stroke [3] frequently experience upper limb impairments. Common motor symptoms include muscle weakness, slow and uncoordinated movements, reduced selectivity, spasticity, and dyskinesia [4]. These impairments often restrict self-care and other activities of daily living [5, 6]. As a result, the rehabilitation goals of these children frequently involve improving hand and arm use in daily activities, such as handling objects or performing self-care activities [7]. Movement smoothness is an important characteristic of goal-directed movements and a universal feature of healthy, skilled movement [8]. For this reason, smoothness is often used to characterize the quality of movement during tasks like reaching [9], and as an indicator of recovery [10, 11]. Studies have shown that smoothness of reaching movements improved in children with congenital and/or acquired brain lesions undergoing technology-assisted upper limb therapy [12, 13]. Additionally, even 6- to 24-month-old infants with cerebral palsy demonstrated improved reaching smoothness following a multi-component upper extremity intervention combining constraint of the less affected side, bimanual play, and graduated reaching tasks, despite smoothness not being explicitly trained [14]. During rehabilitation, movement smoothness is often evaluated through clinical assessments, e.g., as part of the Melbourne Assessment 2 [15], where smoothness is rated on an ordinal scale from zero to three. However, such assessments rely on the subjective perception of therapists and the scoring is typically based on a relatively coarse, ordinal scale, which means that patients must make substantial progress to improve their score. While it seems easy to distinguish between perfect smoothness and clearly impaired smoothness, visually classifying subtle differences in smoothness is particularly difficult [16]. To enable a more objective, quantitative, and fine-grained analysis of reaching smoothness, researchers have started supplementing clinical assessments with marker-based 3D motion analysis (3DMA) systems [11, 17–19]. For marker-based 3DMA, reflective markers are attached to anatomical landmarks on the body and tracked by multiple cameras in 3D to precisely capture a person’s movements. However, 3DMA systems are expensive and require dedicated space, and preparing and performing measurements is elaborate for both patients and therapists [20]. Technological advances have led to the development of markerless, video-based motion tracking methods [21, 22]. While multi-camera systems have shown promising accuracy [21, 23], they require space and effort for setup and calibration. Tracking methods using a single camera, e.g., from a smartphone, can be applied anywhere with little effort, but their accuracy is limited and not high enough for most medical applications [21]. Red–Green–Blue-Depth (RGB-D) cameras capture synchronized color and depth images, with the depth channel providing 3D information about the scene. Such devices, e.g., the Azure Kinect Developer Kit (AKDK), are inexpensive and portable, offering the potential to provide high-quality motion tracking with a single camera. However, we showed in previous studies that the official Azure Kinect Body Tracking SDK, which estimates 3D joint positions from RGB-D images, does not consistently deliver accurate tracking results [24] and smoothness measurements [25] for children and young adults. For this reason, we developed a method for tracking 3D full-body movements of children using a single RGB-D camera, termed SMPLify-KiDS, which was validated against a clinical reference 3DMA system in typically developing children and young adults [24]. We showed that the tracking method is accurate, and that smoothness measurements during reaching highly correlate with the 3DMA system [24, 25]. The code is publicly available at https://github.com/nh236/smplify-kids. However, the quantitative smoothness metrics based on SMPLify-KiDS have only been evaluated with respect to a 3DMA system in healthy children. In this study, we evaluated the extracted smoothness values against the therapist assessment scores in a group of children with neurological impairments. Various metrics have been proposed and evaluated to quantify smoothness in movement kinematics [9, 19, 26]. Among these, the log dimensionless jerk (LDJ) [27] and the modified spectral arc length (SPARC) [9] are considered the most valid and promising. While LDJ measures the rate of changes in acceleration over time, SPARC analyzes the frequency components of the movement speed profile, assuming that smooth movements are characterized by low-frequency components while less smooth, more fragmented movements contain a higher proportion of high-frequency components. The metrics have been thoroughly investigated, and both SPARC and LDJ were found to have excellent reliability and low measurement error [19], although LDJ showed a stronger bias towards task duration and a greater sensitivity to noise than SPARC [9]. While most studies were conducted on simulated data [9], healthy subjects [11, 26], and older stroke survivors [19, 28], only one study included children with hemiplegia [29]. None of these studies compared smoothness measures with therapist assessment of the same task from which the smoothness values were derived; instead, smoothness values were related to (sum) scores from different clinical assessments. In this study, we investigated the use of a new markerless RGB-D camera-based tracking method, SMPLify-KiDS, for the quantitative assessment of movement smoothness during upper limb reaching tasks in a heterogeneous population of children with neurological impairments. Our main objectives were to evaluate (i) if the quantitative smoothness metrics LDJ and SPARC correlated at least moderately with therapist scores on the item level as well as for averaged scores, (ii) which of the two metrics better resembled therapists’ perception, (iii) if the metrics differed significantly between therapist-rated MA2 scores (scale from one to three), and (iv) if the smoothness metrics correlated with task duration, as has been previously reported in literature. Methods In this study, we compared therapist assessment of reaching smoothness to quantitative metrics computed from wrist movements captured with a single RGB-D camera and the SMPLify-KiDS tracking method. Figure 1 presents an overview of the standard clinical assessment pipeline (top) and the proposed quantitative smoothness measurement (bottom).Fig. 1Overview of classical assessment (top, blue) and the proposed quantitative assessment based on an RGB-D camera (bottom, green) Overview of classical assessment (top, blue) and the proposed quantitative assessment based on an RGB-D camera (bottom, green) Participants Children who had a neurological diagnosis, were able to follow instructions, and could perform at least parts of the MA2 without support were included. Participants and parents were informed verbally and in writing. All had to agree verbally before participating. Parents and participants aged 14 years and older were also required to provide written informed consent. Participants could withdraw from the study at any time. We collected information on age, sex, diagnosis, severity using the Manual Ability Classification System (MACS), the more affected side, and the dominant side (hand). Children who met the inclusion criteria and received treatment at the Swiss Children’s Rehab between July 2020 and March 2022 were recruited as a convenience sample. As this study can be considered a psychometric study evaluating the construct validity of the two smoothness metrics, we aimed to recruit at least 30 participants, as international recommendations consider studies with fewer than 30 participants as inadequate [30]. This study was approved by the ethics committee of the canton of Zurich (BASEC-Nr. PB_2016-01843). Reaching items The MA2 is a commonly used assessment of upper limb function in children and evaluates upper limb movement quality across four sub-scales: range of motion, accuracy, dexterity, and movement fluency.1 The 14 items of MA2 include reaching and grasping movements, as well as items inspired by activities of daily living. As the purpose of the test is to measure the quality of upper limb movements rather than the level of cognition in following verbal commands, repeated verbal instruction and demonstrations, as well as repeated attempts, are allowed. This is particularly relevant when testing children. For each item, the scoring criteria define how therapists should rate the patients’ performance, typically on a scale from zero (not able to complete the task) to three (successfully completed the task without violating any scoring criteria). In this study, we focused on the items for which movement smoothness was assessed. To maintain a simple and clinically feasible recording setup, we chose to keep the camera at a fixed position in front of the children throughout the assessment. As a result, certain MA2 items that evaluate smoothness could not be reliably analyzed due to occlusions. These included item 1 (reaching forward to a target, with the hand occluded by the target), item 10 (reaching from the forehead to the neck, with occlusion by the head), and item 11 (palm to bottom, with occlusion by the table). Although item 8 (manipulation of a cube) evaluates smoothness, it focuses on relatively unconstrained, fine-finger movements, which should not be represented by standard smoothness metrics, as these are best suited for point-to-point reaching movements [9]. Therefore, instead of using all seven items of the MA2 that are scored for smoothness, we selected the following three:Item 2: Reach sideways to elevated positionItem 13: Reach to opposite shoulderItem 14: Reach hand to mouth Item 2: Reach sideways to elevated position Item 2: Reach sideways to elevated position Item 13: Reach to opposite shoulder Item 13: Reach to opposite shoulder Item 14: Reach hand to mouth Item 14: Reach hand to mouth Clinical scoring The MA2 guidelines for therapist scoring state that: ”Fluency is scored by observing for the presence of jerkiness, tremor and/or reduced speed of movement”. Scores are given the following way [15]: 3Smoothly coordinated reach with no jerkiness, tremor or reduced speed of movement observed. The action of reaching is easily achieved.2Either subtle or barely noticeable jerkiness, tremor or slowness of movement; or jerkiness or tremor throughout less than half the reach movement. No observed increased effort required to perform the reaching action.1Marked jerkiness, tremor or slowness of movement through greater than half the movement range or reach is achieved with increased effort.0Unable to achieve the task due to excessive jerkiness or tremor on movement, preventing required contactTwo occupational therapists, one with 1 year and one with 8 years of experience with MA2, independently scored the MA2 recordings, i.e., they were blinded to each other’s scores. In cases of divergent scores, videos were discussed and re-evaluated until consensus was found. 3Smoothly coordinated reach with no jerkiness, tremor or reduced speed of movement observed. The action of reaching is easily achieved. Smoothly coordinated reach with no jerkiness, tremor or reduced speed of movement observed. The action of reaching is easily achieved. 2Either subtle or barely noticeable jerkiness, tremor or slowness of movement; or jerkiness or tremor throughout less than half the reach movement. No observed increased effort required to perform the reaching action. Either subtle or barely noticeable jerkiness, tremor or slowness of movement; or jerkiness or tremor throughout less than half the reach movement. No observed increased effort required to perform the reaching action. 1Marked jerkiness, tremor or slowness of movement through greater than half the movement range or reach is achieved with increased effort. Marked jerkiness, tremor or slowness of movement through greater than half the movement range or reach is achieved with increased effort. 0Unable to achieve the task due to excessive jerkiness or tremor on movement, preventing required contact Unable to achieve the task due to excessive jerkiness or tremor on movement, preventing required contact We excluded samples with a score of zero because this indicates that the child or adolescent was unable to perform the task, and the computation of smoothness values for point-to-point reaching movements only makes sense if the target point is reached. Data acquisition The RGB-D camera (Azure Kinect Developer Kit) was mounted on a tripod, connected to a laptop, and placed in front of a table where the children were seated, so that the full range of motion of the children’s upper body and upper limbs was in view. When the therapist and the child were ready, the recording began. Depth images were recorded at a resolution of 640x576 pixels and colour images at 1920x1080 pixels, both at 30 frames per second. All items of the MA2 were conducted as described in the clinical assessment, i.e., with an additional person filming the child with a hand-held camera according to the assessment instructions, which was used for clinical scoring. After the assessment was completed, the recording was stopped. Children wore their regular clothes. Setting up the laptop and camera took less than 5 min. Data processing The complete assessment was recorded with the RGB-D camera from onset to end in a single recording, with the camera in a fixed position to minimize effort and distraction for the therapist and patient. Instead of starting and stopping the recording for each item and attempt, the recording was segmented afterwards. For this, a therapist watched the recorded video and annotated the exact same attempts that were scored in the clinical assessment by pressing a button at the onset and end of each attempt. This segmentation can not be easily automated; e.g., when a child performs a task incorrectly and has to repeat it several times, it is up to the therapist to decide which of the repetitions should be assessed. The remaining processing was automated and did not require manual intervention. As displayed in Fig. 1, the segmented RGB-D video was processed with the SMPLify-KiDS method. The method takes a sequence of color and depth images as input, and outputs a virtual 3D body, the SMPL-H model [31], consisting of the body surface and an underlying skeleton, i.e., joint positions and angles. We adapted the original SMPLify-KiDS method [24] by restricting the fitting to the upper body, since the children were sitting at a table that occluded their lower body. We extracted 3D wrist joint positions over time, computed wrist speed, and then applied a 6th-order Butterworth filter with a cut-off frequency of 5 Hz. Smoothness metrics We selected two metrics from the literature to quantify movement smoothness: the log dimensionless jerk (LDJ) [27] and the modified spectral arc length (SPARC) [9]. Jerk is defined as the rate of change of acceleration, and the LDJ is the natural logarithm of the dimensionless integrated squared jerk [27]:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{LDJ} \triangleq -\ln \left( \frac{(t_2 - t_1)^3}{v_{peak}^2} * \int _{t1}^{t2}\left| \frac{d^2v(t)}{dt^2}\right| ^2 dt\right) \end{aligned}$$\end{document}with speed profile v(t), peak speed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{peak}$$\end{document} and time window \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_1$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_2$$\end{document}. Unlike other jerk-related metrics, LDJ’s dimensionless design makes it independent of movement duration and amplitude [27]. While LDJ measures smoothness in the temporal domain, SPARC operates in the frequency domain. It decomposes the speed profile into lower and higher frequencies using the Fourier transform, and calculates the arc length of the Fourier magnitude spectrum [9]:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{SPARC}&\triangleq -\int _0^{\omega _c}\left[ \left( \frac{1}{\omega _c}\right) ^2 + \left( \frac{d\hat{V}(\omega )}{d\omega }\right) ^2 \right] ^{\frac{1}{2}} d\omega ; \nonumber \\ \hat{V}(\omega )&= \frac{V(\omega )}{V(0)} \end{aligned}$$\end{document}where V(ω ) is the Fourier magnitude spectrum of speed profile v(t), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{V}\left( \omega \right) $$\end{document} is the normalized magnitude spectrum, normalized with respect to the DC magnitude V(0), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _c$$\end{document} is adaptively selected based on a given threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{V}$$\end{document} and is upper-bound by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _c^{max}$$\end{document}:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \omega _c\ =\min {\left\{ \omega _c^{max},\min {\left\{ \omega ,\hat{V}\left( r\right) <\overline{V}\forall r>\omega \right\} }\right\} } \end{aligned}$$\end{document}Both smoothness metrics are calculated based on wrist speed during (point-to-point) reaching movements. Therefore, the filtered wrist speed served as input for the calculation of LDJ and SPARC, for which we used the code provided by Balasubramanian et al. [9] (https://github.com/siva82kb/SPARC). Due to the minus signs, both LDJ and SPARC result in negative values. Values closer to zero represent smoother movements. Statistics We correlated the smoothness values per item with the therapist item scores using Spearman’s rank correlation ρ . Similarly, after calculating the average smoothness metrics and therapist scores across the three items per participant, we correlated the average smoothness values with the average therapist scores. We interpreted the magnitude of the correlation coefficient as follows: 0−0.25 (no or little relationship), 0.25−0.50 (fair degree), 0.50−0.75 (moderate to strong relationship), 0.75−1.00 (very strong to excellent), and calculated 95% confidence intervals (95%CI) using the bias-corrected and accelerated bootstrapping (BCa) method with 2000 resamples. Additionally, we conducted a subgroup analysis, dividing the participants into groups based on the therapist scores. For each group, we computed the mean smoothness value and standard deviation. We then performed a Shapiro-Wilk test to check for normality. Since data were not normally distributed in some subgroups, we decided to always use the non-parametric Mann–Whitney-U test to test for significant differences between groups. For the pair-wise comparisons between ordinal score subgroups, the significance was corrected for multiple comparisons and α was set to 0.05/3 = 0.0167. Results Thirty-eight children (15 females, 23 males, age range 6–18 years, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varnothing 11.7 \pm 2.9$$\end{document} years) were included in this study. Eleven children were diagnosed with bilateral cerebral palsy, 6 with unilateral cerebral palsy, 5 with stroke, 2 with ataxia, 2 with traumatic brain injury, 2 with astrocytoma, and the remaining 10 with other diagnoses. Thirteen children were categorized with MACS [32] level I (objects are handled easily and successfully), twelve with MACS level II (Handles most objects but with some reduced quality and/or speed), eight with MACS level III (Handles objects with difficulty - the child will need help to prepare and/or modify activities), and three with MACS level IV (Handles a limited selection of easily managed objects and always requires some help from others), two unknown. The overall MA2 scores ranged from 23 to 92, with an average of 73.6 out of a maximum of 92, and the scores of the regular fluency sub-scale ranged from 4 to 21, with an average of 15.6 out of a maximum of 21. All 38 children performed the assessment with both the left and the right arm, but we only evaluated the more affected side. For four children who did not have an explicitly more affected side, we included the non-dominant side. Some children were unable to complete all items, resulting in 110 evaluated sequences (instead of 114). Agreement with therapist scores Smoothness values (LDJ and SPARC) and therapist scores for all items are visualized in Fig. 2. Samples with therapist scores 3 (green) and 1 (red) are relatively well separated, while score 2 samples (yellow) overlap with the other groups.Fig. 2Smoothness values per item. Each dot in a plot represents one patient. X-axis: therapist score, Y-axis: smoothness value. *denotes significant differences between score groups (p<0.0167, corrected for multiple comparisons) Smoothness values per item. Each dot in a plot represents one patient. X-axis: therapist score, Y-axis: smoothness value. *denotes significant differences between score groups (p<0.0167, corrected for multiple comparisons) Results showed consistently higher correlations with therapist scores for LDJ compared to SPARC, at ρ values between 0.644 and 0.836 for LDJ and between 0.556 and 0.674 for SPARC, with all p values < 0.001 (Fig. 2). LDJ had the highest correlation value in the reach sideways task (item 2), SPARC in the reach to opposite shoulder task (item 13). Subgroup analysis per therapist score The mean values per score group were relatively consistent across items, as plotted in Fig. 2. For LDJ, mean values for a therapist score of 1 ranged from -10.468 to -11.400, for a score of 2 from -7.887 to -8.484, and for a score of 3 from -6.885 to -7.121. Mean SPARC values for a therapist score of 1 ranged from -2.256 to -2.878, for a score of 2 from -1.777 to -1.928, and for a score of 3 from -1.690 to -1.761. For LDJ, differences between score groups (Fig. 2) were statistically significant (p<0.0167, corrected for multiple comparisons) for all items and scores, except for item 14 when comparing score 2 versus score 3. The largest effect sizes were consistently observed between scores 1 and 3, with all comparisons reaching p values below 0.001. Differences between score 1 and score 2 were at similar levels, with a slightly lower p value of 0.004 for item 13. LDJ values for scores 2 and 3 showed lower, but still significant, p values for items 2 (p=0.0014) and 13 (p=0.0165), and no significant difference for item 14 (p=0.08). For SPARC, p values for differences between score 1 and score 3 were consistently below 0.001, similar to LDJ. Differences between scores 1 and 2 were significant for items 2 (p=0.0106) and 13 (p=0.007), but not for item 14 (p=0.031). Comparisons between scores 2 and 3 were only significant for item 13 (p=0.0087). Average scores over all items Correlations for average scores were higher than for single-item scores at very strong to excellent levels for all metrics, with LDJ showing a higher correlation of ρ = 0.843 (p<0.001) than SPARC at ρ = 0.755 (p<0.001). The average LDJ/SPARC scores are plotted against the average therapist scores in Fig. 2g and h. Task duration The correlation between task duration and LDJ was consistently excellent with ρ =-0.933 for item 2, -0.896 for item 13, and -0.884 for item 14, and all p values < 0.001 (Fig. 3a, c, e). Similarly, the correlation between average duration and average LDJ was excellent (ρ =-0.934, p< 0.001, Fig. 3g).Fig. 3Comparison of smoothness values and duration per item. Each dot in a plot represents one patient, color according to therapist score (red: 1, yellow: 2, green: 3). X-axis: duration in seconds, Y-axis: smoothness value Comparison of smoothness values and duration per item. Each dot in a plot represents one patient, color according to therapist score (red: 1, yellow: 2, green: 3). X-axis: duration in seconds, Y-axis: smoothness value For SPARC, correlations with task duration were comparatively lower, but remained within the moderate to strong range, with ρ =-0.668 for item 2, -0.655 for item 13, and -0.695 for item 14, and all p values < 0.001 (Fig. 3b, d, f). The correlation between average duration and average SPARC was excellent (ρ =-0.836, p< 0.001, Fig. 3h). Discussion This study evaluated the use of a 3D markerless motion tracking method for the quantitative assessment of movement smoothness during reaching tasks. When compared with normative values for point-to-point reaching reported in the literature (LDJ: -6; SPARC: -1.6) [9], the values obtained for samples assigned the highest therapist score, i.e., those considered to show no apparent reduction in movement smoothness, were slightly lower, at a mean LDJ value of -7.02, and a mean SPARC value of -1.72. This modest reduction can be attributed to two factors. First, the reported normative reference values were derived using gold-standard marker-based systems, whereas the present measurements were obtained using a markerless approach with inherently lower tracking accuracy. In support of this explanation, SPARC values computed from SMPLify-KiDS data have previously been evaluated in a healthy population of children and young adults [25], yielding an average SPARC value of -1.65 during a reaching task. Second, despite receiving the highest smoothness scores, the data were collected from children with neurological diagnoses. Consequently, samples assigned the highest score by therapists may not exclusively reflect optimal movement smoothness, but may also include movements deemed sufficiently smooth to warrant the maximum score. Our first two objectives were to evaluate whether the quantitative smoothness metrics LDJ and SPARC correlated at least moderately with therapist scores at the item-level and for averaged scores across tasks, and whether LDJ or SPARC correlated better with therapist scores. Indeed, all correlation coefficients were at least of moderate magnitude (i.e., ≥ 0.5). The evaluation of smoothness metrics with respect to therapist scores revealed that per-task correlation values are mostly higher with LDJ than with SPARC. An exception to this overall trend was observed for item 13 (reach to the opposite shoulder), where SPARC achieved its highest correlation (ρ = 0.674) and LDJ showed its lowest correlation (ρ = 0.666). For this item, task durations for score levels 2 and 3 were shorter and less variable than for Items 2 and 14 (see Fig. 3, c, a, and e), possibly because the larger target area, i.e., the shoulder, required less precision and allowed children to complete the movement more quickly. Given LDJ’s sensitivity to task duration, this may have contributed to reduced discriminability between score levels 2 and 3 for this item. Furthermore, when considering average scores across tasks, the correlations were higher than the single-task correlations and, again, higher for LDJ than for SPARC. Aggregating scores across tasks likely reduces variability and measurement noise, thereby strengthening associations with therapist ratings. Our third objective was to evaluate whether the LDJ and SPARC metrics differed significantly between therapist-rated ordinal MA2 scores. Consistent with the correlation results, LDJ discriminated better between therapist-rated score levels than SPARC, with significant differences observed across most pairwise comparisons. As expected, discrimination was strongest between score levels 1 and 3, as these levels distinctly separate smooth from non-smooth movements. In contrast, comparisons involving the intermediate score level showed greater overlap, reflecting the difficulty of mapping the continuous range of movement smoothness to discrete scores. Overall, these findings suggest that LDJ is better suited for a quantitative evaluation of movement smoothness. This seems to conflict with Balasubramanian et al. [9], who recommend using SPARC over LDJ due to LDJ’s sensitivity to noise. However, their analysis was based on simulated data. In the present study, the noise level of the tracked movements seems sufficiently low for LDJ’s noise sensitivity to have a limited impact. Importantly, because therapist ratings rather than an objective ground truth were used as reference, our findings should be interpreted as indicating that LDJ more closely reflects therapist judgment than SPARC, rather than as definitive evidence of superior smoothness representation. Our fourth objective was to evaluate whether the smoothness metrics correlated with task duration. Previous literature has noted that, although LDJ is dimensionless and therefore does not depend directly on task duration [9, 26, 27], it remains strongly correlated with duration [19, 26]. Our results confirm this relationship, with very strong correlations between task duration and LDJ (ρ around 0.9). While such dependence may appear undesirable, it may in fact align with human perceptual criteria for assessing smoothness. According to the MA2 scoring guidelines, ”fluency is scored by observing for the presence of: jerkiness, tremor and/or reduced speed of movement”. From a biomechanical perspective, deviations from a direct trajectory or repeated accelerations and decelerations naturally increase movement duration. In line with this interpretation, we found moderate to excellent correlations between task duration and therapist scores. Duration also showed moderate to strong correlations with SPARC, though consistently lower than those observed for LDJ and therapist scores. The relationship of duration and LDJ has previously been examined in healthy middle-aged adults [26] and in older adults with post-stroke hemiparesis [19]. The present work extends these findings to children with a broad range of neurological diagnoses. While duration alone showed strong associations with therapist scores, it cannot fully substitute smoothness metrics, as movements may be fast yet irregular or slow yet smooth. Thus, duration and smoothness metrics capture complementary aspects of movement quality. Methodological considerations In contrast to some prior work, our results showed higher correlations for LDJ than for SPARC. This discrepancy may be explained by differences in motion capture quality. For instance, Adams et al. [29] evaluated both SPARC and LDJ, computed using a custom motion-tracking method that builds on the Kinect v2 body-tracking SDK and measurements from a smart toy, during a video game. Smoothness values were correlated with multiple assessments of upper limb function, yielding strong correlations with SPARC and low-to-moderate correlations with LDJ. The authors attributed the lower correlations with LDJ to its sensitivity to measurement noise, similar to Balasubramanian et al. [9]. We hypothesize that the higher correlations for LDJ observed in our study result from the comparatively lower noise levels of kinematics extracted using SMPLify-KiDS , which has previously been shown to be suitable for smoothness estimation [25]. Based on these existing studies, we conclude that selecting a smoothness metric should depend on the noise level of the motion capture method, and that SPARC is preferable for noisier signals. Furthermore, the application of these metrics should be limited to point-to-point reaching movements, which represent the context for which both metrics were originally designed. Under the conditions of this study, LDJ appears to align more closely with therapist assessment, potentially due to its relationship with task duration. Regarding generalizability, the acceptable correlations between smoothness metrics and therapist scores suggest that tracking quality was not substantially compromised by the clinical population, supporting the generalizability of SMPLify-KiDS to patients with neurological diagnoses. The method demonstrates strong performance in controlled clinical settings where the child remains visible and relatively stationary, but is less suitable for unconstrained settings involving free ambulation or movements that cause occlusion. The combination of SMPLify-KiDS with smoothness metrics is expected to generalize across diverse patient populations, as evidenced by the heterogeneity of diagnoses in the current sample, provided that the assessed tasks involve point-to-point reaching movements. However, only three of the seven MA2 fluency subscale items were evaluated. For the remaining four items, the combination of the fixed camera position and the placement of the children behind the table caused relevant body parts to be occluded during task execution, limiting the comprehensiveness of the current evaluation. A key limitation of this study is the use of therapist scores as the reference. These scores are inherently ordinal and subject to inter-rater variability, which precludes perfect correlations with continuous smoothness metrics. To mitigate subjectivity, consensus scores from two experienced MA2 therapists were used. Another challenge is that therapists must map a continuous movement quality into discrete score categories, inevitably producing overlap at score boundaries that limits discriminability. This limitation underscores the need for fine-grained, interval-scaled, and objective assessment methods to complement existing ordinal scales. Future work will involve collecting a normative dataset of typically developing children performing the same MA2 tasks, thereby providing an objective reference for comparison with patient performance. Additionally, we plan to investigate the responsiveness of the smoothness metrics through repeated assessments to evaluate their sensitivity in detecting longitudinal changes during rehabilitation. Conclusion This study investigated the use of quantitative movement smoothness metrics derived from markerless RGB-D-based motion capture to support the clinical assessment of upper limb reaching in children with neurological impairments. Both evaluated smoothness measures, LDJ and SPARC, showed at least moderate correlations with therapist-assigned MA2 smoothness scores at both the task level and when averaged across tasks, demonstrating their general suitability for capturing clinically relevant aspects of movement smoothness. Across all analyses, LDJ consistently exhibited stronger correlations with therapist scores and more clearly differentiated between score levels than SPARC. These results indicate that LDJ more closely reflects therapist judgment of movement smoothness under the conditions examined in this study. We further demonstrated that LDJ is strongly correlated with task duration. Rather than representing a limitation, this relationship appears to align with clinical perception, as reduced movement speed and prolonged execution are explicitly considered in MA2 fluency scoring. Nevertheless, task duration alone cannot fully capture smoothness, which is why smoothness metrics and duration provide complementary information about movement quality. Overall, this work supports the use of markerless motion capture-based smoothness metrics as objective, fine-grained complements to established clinical assessments. Such approaches have the potential to enhance clinical decision-making and longitudinal monitoring by providing quantitative measures of movement quality that extend beyond ordinal therapist ratings. One key advantage of markerless full-body tracking is that it removes the need to predefine marker placement on specific body segments. Instead, the visible body parts, here, the upper body, are consistently tracked, allowing the focus of the analysis to be determined retrospectively. This flexibility supports future extensions of the present work to more comprehensive analyses of movement quality beyond smoothness. Supplementary Information Additional file 1. Additional file 1.