Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Neuroeng Rehabil . 2026 Mar 3;23:120. doi: 10.1186/s12984-026-01919-6 Search in PMC Search in PubMed View in NLM Catalog Add to search AI-based video analysis for the assessment of upper limb function in children with unilateral cerebral palsy: feasibility of remote monitoring Youngsub Hwang Youngsub Hwang 1 Medical Research Institute, Sungkyunkwan University School of Medicine, Suwon, Republic of Korea Find articles by Youngsub Hwang 1, # , Hakje Yoo Hakje Yoo 2 Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea 3 Department of Bio-Mechatronic Engineering, College of Biotechnology and Bioengineering, Sungkyunkwan University, Suwon, Republic of Korea Find articles by Hakje Yoo 2, 3, # , Minkyung Kim Minkyung Kim 2 Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea 4 Department of Digital Health, SAIHST, Sungkyunkwan University, Seoul, Korea Find articles by Minkyung Kim 2, 4 , Myung Jin Chung Myung Jin Chung 2 Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea 5 Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea Find articles by Myung Jin Chung 2, 5, ✉ , Jeong-Yi Kwon Jeong-Yi Kwon 6 Department of Physical and Rehabilitation Medicine, Sungkyunkwan University School of Medicine, Samsung Medical Center, Seoul, Republic of Korea Find articles by Jeong-Yi Kwon 6, ✉ Author information Article notes Copyright and License information 1 Medical Research Institute, Sungkyunkwan University School of Medicine, Suwon, Republic of Korea 2 Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea 3 Department of Bio-Mechatronic Engineering, College of Biotechnology and Bioengineering, Sungkyunkwan University, Suwon, Republic of Korea 4 Department of Digital Health, SAIHST, Sungkyunkwan University, Seoul, Korea 5 Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea 6 Department of Physical and Rehabilitation Medicine, Sungkyunkwan University School of Medicine, Samsung Medical Center, Seoul, Republic of Korea ✉ Corresponding author. # Contributed equally. Received 2025 Jul 24; Accepted 2026 Feb 15; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13067724 PMID: 41776622 Abstract Background Accurate assessment of upper limb function in children with unilateral cerebral palsy (UCP) is essential for clinical decision-making. Although the Melbourne Assessment 2 (MA2) has been widely adopted to achieve this, this method requires substantial evaluator training and lengthy administration times, and is susceptible to inter-rater variability. To address these challenges, we evaluated the feasibility of employing AI-driven automated video analysis to objectively classify the severity of upper limb impairment in pediatric UCP using retrospectively collected MA2 videos captured under realistic, non-standardized clinical conditions. Methods This study retrospectively enrolled 19 children with UCP (mean age 5.50 ± 2.03 years; 10 males, nine females) who underwent repeated MA2 assessments, yielding 644 candidate item-videos. After pre-specified quality control, the final analytic set comprised 616 item-videos. We fine-tuned four index-specific Video Vision Transformer (ViViT) models—one per MA2 index—and evaluated them using subject-wise five-fold cross-validation to predict dichotomized severity (mild vs. severe) for range of motion (ROM), accuracy, fluency, and dexterity. Videos were recorded with handheld devices without standardized protocols, reflecting typical clinical practice. Results The AI model demonstrated a mean area under the curve (AUC) of 0.890, with the highest performance for dexterity (AUC = 0.936), and similarly high performances for ROM (AUC = 0.866), accuracy (AUC = 0.887), and fluency (AUC = 0.869). Item-wise analyses showed that a limited subset of MA2 tasks contributed disproportionately to the correct classification of mild versus severe cases across indices, highlighting the importance of item-level variability when designing automated MA2 assessment protocols. Conclusions This preliminary study showed that AI-driven automated video analysis can classify dichotomized MA2-based upper limb impairment severity using routinely acquired, nonstandardized clinical videos of children with UCP. However, further methodological refinement—including standardized recording guidelines, prediction models that extend beyond binary labels to ordinal or item-level MA2 scoring, and prospective validation in larger and more diverse cohorts—is required to confirm clinical applicability and enhance generalizability. Trial registration Not applicable. Supplementary Information The online version contains supplementary material available at 10.1186/s12984-026-01919-6. Keywords: Cerebral palsy, Upper extremity, Accuracy, Artificial intelligence, Video analytics Background Unilateral cerebral palsy (UCP) is a common pediatric neurological disorder characterized by impaired motor function, primarily affecting a single upper limb [ 1 ], significantly impacting the child’s ability to perform daily activities, and reducing their overall quality of life [ 2 ]. The accurate and reliable assessment of upper limb function is an essential aspect of management, not only for identifying impairment severity, but also for tailoring appropriate rehabilitation strategies and tracking intervention outcomes [ 3 ]. The Melbourne Assessment 2 (MA2) is a standardized tool specifically developed to objectively evaluate unilateral upper limb motor function in children with neurological impairments, including UCP [ 4 ]. Although the MA2 has demonstrated strong psychometric properties and is widely recognized for its validity and reliability [ 5 , 6 ], several limitations for its practical clinical application have been identified. Specifically, the MA2 requires substantial evaluator training, and insufficient training can significantly increase inter-rater variability, negatively impacting scoring consistency and reliability [ 7 ]. Furthermore, the evaluation procedure itself, including rater training and scoring processes, is time-consuming, making it challenging to implement efficiently in routine clinical practice, where time and resources are limited [ 6 ]. Additionally, evidence regarding the psychometric properties of individual MA2 score items and subscales remains insufficient, thus highlighting the need for a comprehensive evaluation of their internal psychometric characteristics [ 6 ]. Given these practical and psychometric challenges, automated AI-driven video analysis has emerged as a promising solution for rehabilitation, offering enhanced objectivity, reduced interrater variability, and improved efficiency in upper-limb function assessments [ 8 ]. Among children with UCP, such approaches may help to overcome the limitations of traditional assessment tools such as the MA2, as supported by prior research. For example, three-dimensional movement analysis (3DMA) applied to children with UCP has demonstrated the benefits of objective, automated scoring over conventional clinical scales, which rely heavily on subjective visual observation [ 9 ]. However, prior UL-3DMA studies in children with UCP have encountered practical constraints that limit clinical scalability because the wide range of UL functions has produced heterogeneous protocols and task sets, reducing comparability and hindering routine deployment [ 9 ]. In adults after stroke, depth-sensing approaches have shown strong correlations between AI-derived scores and the Fugl-Meyer Assessment and have provided quantitative kinematic insights such as movement smoothness [ 10 ]. Additionally, video-based analytical systems, including single-camera setups, have proven to be effective at identifying compensatory movements, thereby offering accurate and cost-effective assessments suitable for both clinical and home settings [ 11 , 12 ]. Although these cohorts differ from children with UCP, the findings support the feasibility of automated scoring and justify pediatric evaluation using routine clinic video recordings. Collectively, these findings underscore the feasibility and utility of AI-driven video analysis as a robust complementary approach to traditional upper-limb assessments. Although prior studies have demonstrated the feasibility of automated video analysis in neurological populations, these studies primarily employed sophisticated laboratory-based 3DMA systems or depth-sensing cameras, thus limiting clinical practicality owing to the associated cost, complexity, and space constraints [ 9 , 12 ]. Specifically, automated video analysis has rarely been applied to MA2, despite its widespread clinical use and significant practical limitations, including its inter-rater variability, intensive evaluator training, and considerable time demand [ 6 , 7 ]. Therefore, investigating an automated AI-driven video analysis approach using routinely-collected non-standardized clinical videos addresses a critical research gap. Establishing initial feasibility could enhance objectivity, reduce clinical workload, and improve the accessibility of standardized upper limb functional assessments, potentially leading to improved clinical decision-making and outcomes in pediatric rehabilitation. In this context, we conducted a preliminary study to examine the feasibility of automating MA2 severity scoring using routinely captured, non-standardized clinic videos recorded with a single monocular RGB camera (i.e., without a depth sensor) in children with UCP. By “non-standardized,” we refer to handheld recordings characterized by variable viewpoints and distance, occasional zoom and therapist occlusion, and multiple takes per session. Establishing the initial feasibility could further lead to reduced clinician burden, enhanced reliability, and increased accessibility to standardized assessment methods, thereby supporting improved clinical decision-making and outcomes in pediatric rehabilitation practice. Methods Participants This study involved a retrospective review of the medical records and corresponding video recordings of 22 children diagnosed with UCP who visited the Samsung Medical Center between 2021 and 2022. The inclusion criteria were as follows: (1) diagnosis of UCP, confirmed by pediatric neurologists; (2) age between 4 and 12 years; and (3) availability of MA2 assessment video(s) from one to three clinical sessions, with repeated sessions (when available) approximately one month apart. Participants were excluded if they had severe cognitive dysfunction, untreated seizures, visual or auditory impairment, or musculoskeletal disorders that could significantly interfere with upper limb assessment. Among the 22 enrolled children, the families of nineteen provided explicit additional consent for their videos to be used specifically for AI model training and validation, and these 19 children constituted the final modeling cohort. In this cohort, the mean age was 5.50 ± 2.03 years, and 10 children (52.6%) were male and 9 (47.4%) were female. Regarding upper limb functional levels, the Manual Ability Classification System (MACS) levels [ 13 ] were distributed as follows: level I ( n = 7; 36.8%), level II ( n = 7; 36.8%), and level III ( n = 5; 26.3%). Seventeen children (89.5%) had right-sided involvement and two (10.5%) had left-sided involvement. The hospital review board approved the study protocol, and informed consent was acquired from the parents or legal guardians of all participating children prior to enrollment. Table 1 summarizes the demographic and clinical characteristics of the nineteen participants included in the model. Table 1. Participant ( n = 19) characteristics Participant Age (years) MACS level Affected side Sex Child 1 5.00 II Right Female Child 2 5.42 III Right Male Child 3 6.50 I Right Male Child 4 5.00 I Right Female Child 5 4.42 I Right Female Child 6 10.58 II Right Male Child 7 9.17 II Right Male Child 8 4.00 I Right Male Child 9 4.50 II Right Female Child 10 4.50 III Right Female Child 11 4.17 II Right Male Child 12 4.33 III Left Male Child 13 4.00 II Right Female Child 14 4.33 I Right Male Child 15 9.58 III Right Female Child 16 5.83 I Right Female Child 17 5.17 I Right Male Child 18 4.00 III Left Female Child 19 4.08 II Right Male Open in a new tab MACS manual ability classification system Measures The MA2 is a standardized assessment tool designed to quantitatively evaluate upper limb function in children with cerebral palsy. MA2 includes four major indices: (1) Range of Motion (ROM), (2) accuracy, (3) fluency, and (4) dexterity. Most MA2 items are scored on a 0–3 ordinal scale, whereas Items 6, 9, and 12 use a 0–4 scale; in all cases, lower scores indicate greater impairment. Given the preliminary nature of this feasibility study we a priori dichotomized the ordinal scores into binary labels using a scale-aware rule: for 0–3 scales, severe (0) = {0,1} and mild (1) = {2,3}; for 0–4 scales (Items 6, 9, 12), severe (0) = {0,1} and mild (1) = {2,3,4} [ 14 ]. This dichotomization scheme was specified before model development and applied consistently across all experiments. Data collection Retrospective clinical data and video recordings from Samsung Medical Center (Seoul, Korea) were used in this study. Children diagnosed with UCP underwent routine upper limb functional evaluation as part of clinical practice between 2021 and 2022. Participants completed the MA2 assessments on one to three occasions approximately one month apart, resulting a source cohort of 22 children. After re-confirmation of consent for secondary analysis, three children declined, leaving 19 participants and 46 session-level videos for processing. Each session-level video was segmented into 14 MA2 item-specific clips (“item-videos”), producing 644 item-videos. After prespecified quality-control screening, which excluded videos with an obscure target, severe camera shake, evaluator occlusion, or excessive zoom-in (28 of 644 videos, 4.3%), the final analyzed dataset comprised 616 item-videos that were used for subject-wise five-fold cross-validation (Additional files 1 and 2). All the MA2 assessments were performed by a single experienced pediatric occupational therapist. However, due to clinical constraints, no standardized operating procedures (SOP) for video recording were implemented. The videos were captured by a second examiner or research assistant using handheld devices (smartphones or tablets) not following any standardized protocols regarding camera positioning, angle, distance, or environmental lighting conditions. This methodological choice was intentionally chosen to reflect real-world clinical conditions, in which standardized filming protocols are often impractical or unavailable. During the assessment, the children performed all tasks defined in MA2 under direct instruction from an evaluating occupational therapist. The therapist observed the task performance directly and manually scored each of the four indices (ROM, accuracy, fluency, and dexterity) according to the standard MA2 scoring criteria. Simultaneously, these task performances were video recorded for subsequent analyses. Data management Video recordings were retrospectively reviewed to confirm sufficient video quality and completeness for all MA2 tasks. From the original dataset, 616 video recordings from nineteen participants who provided explicit additional consent were selected for the development and validation of the AI-driven automated video analysis model. The video files were anonymized prior to analysis; masks were worn during recording, the periorbital region was blurred, and all file names and internal metadata were removed or replaced before encrypted transfer to the engineering team. Clinical data, were de-identified and stored separately from the videos, with access restricted to authorized personnel. MA2 videos were recorded such that each test was repeated multiple times within a single video file, resulting in multiple repetitions embedded in a single video. In this study, each repetition was treated as a separate training unit and divided into fixed-length video clips of 32 frames, which is the input size required for the Video Vision Transformer (ViViT) model [ 15 ]. The choice of ViViT was determined by comparative experiments with representative CNN backbones (VGG and ResNet), and the corresponding results are presented in Additional files 3 and 4. The clip segmentation strategy is based on the number of test repetitions for each video. The total number of frames in a video was divided by the number of repetitions to determine the approximate number of frames per task, which were converted into standardized 32-frame clips. Because the length of each repetition varied, a segmentation strategy was applied to ensure consistency across the data samples despite frame length discrepancies. If a repetition contained fewer than 32 frames, the entire sequence was used as a single clip. For repetitions containing 32–150 frames, a single 32-frame clip was generated using uniformly subsampling; the full frame range was divided into 32 equal intervals, and one frame was selected from each interval to ensure inclusion of the first and last frames (for example, a 96-frame repetition yields approximately every third frame). If a repetition contained more than 150 frames, it was split into non-overlapping 150-frame segments, and the same 32-frame uniform subsampling strategy was applied to each segment, generating multiple clips per repetition. This approach minimizes the loss of temporal information while increasing data diversity. As a sensitivity analysis, we also implemented and evaluated an alternative sliding-window clip sampling strategy (Additional file 5); however, because this approach did not yield clearly superior performance and increased data redundancy, all main results reported in this article are based on the single-clip sampling scheme described above (see Additional files 6 and 7). The resulting clips were preprocessed into four-dimensional tensors of shape (32, 3, 224, and 224), saved in NumPy format, and assigned corresponding labels to construct the final metadata for model training. In total, this procedure yielded 7,812 32-frame clips across 616 item-videos from 19 children, which served as the basic training units for model development. The detailed number of clips contributed by each participant for each MA2 item is summarized in Additional file 8, and the distribution of dichotomized labels (mild vs. severe) for each MA2 index, both overall and within each cross-validation fold, is provided in Additional file 9. Development of the AI model To quantitatively analyze movement characteristics and classify severity levels in clinical videos, four independent binary classifiers were developed, one for each MA2 task, using a pretrained Video Vision Transformer (ViViT-B/16 × 2, factorised encoder, pretrained on Kinetics-400). This ViViT variant applies factorised attention, separating temporal and spatial attention to effectively capture motion and spatial features within the examination. Each clinical examination within a video was represented as a 32-frame clip, reshaped into a tensor of size (32, 3, 224, 224), and processed by the ViViT backbone. The backbone produced a 400-dimensional feature vector obtained from its Kinetics-400 pretraining. For each MA2 task, this feature vector was processed by a task-specific classification layer, which generated a single logit for the clip. A sigmoid activation was subsequently applied to convert the logit into a scalar probability reflecting the severity of impairment for that task (Fig. 1 ). Fig. 1. Open in a new tab The VIVIT Severity classification model based on video vision transformer using MA2 videos. Each MA2 evaluation video is segmented into 32-frame clips, which are processed by an index-specific ViViT model to produce clip-level probabilities of severe impairment for a given MA2 index (accuracy, dexterity, fluency, or ROM). Clip-level probabilities are then averaged within each item-video and thresholded at 0.5 to yield a binary item-level prediction (0 = mild, 1 = severe) for that index For each patient i , the MA2 evaluation video comprises frames. The video was segmented into clips, each of which contained 32 frames. The k th clip from the j th video was defined as follows (Eq. 1 ): 1 The ViViT model generates a feature representation for each clip , and subsequently applies a sigmoid function to output the binary classification probabilities for the four MA2 assessment tasks (Eq. 2 ). 2 The MA2 evaluation consists of 14 distinct assessment items, and each item was rated across four tasks such as range of ROM, Accuracy, Fluency, and Dexterity. For binary classification, all task scores were converted into two categories by applying a common threshold in which values of 1 or below were considered mild corresponding to label 0, whereas values above 1 were considered severe corresponding to label 1. The predicted class for each task was determined by applying a threshold of 0.5 to the sigmoid output, as expressed in Eq. 3 . 3 Performance analyses To evaluate the performance and validity of the AI-based automated MA2 scoring models, a series of quantitative analyses was conducted. Model training and validation used a subject-wise five-fold cross-validation procedure applied separately to each MA2 index, and data were split based on individual participants to ensure that samples from the same subject were not simultaneously included in both the training and validation sets. No separate external test set was reserved; instead, the entire dataset was used within this cross-validation framework. The binary classification performance of each of the four MA2 indices—ROM, Accuracy, Fluency, and Dexterity—was evaluated using its corresponding single-output index-specific model and assessed with standard metrics, including accuracy, precision, sensitivity (recall), specificity, F1-score, and the area under the receiver operating characteristic curve (AUC). All metrics were computed at the item-video level: for each item-video, we averaged the clip-level probabilities produced by the model (Eq. 2 ) across all clips belonging to that item-video to obtain a single item-level probability, and then applied a fixed threshold of 0.5 to these item-level probabilities, consistent with the decision rule in Eq. 3 . The ROC curves and AUCs in Fig. 2 were derived from these item-level predictions. All results were reported as the mean ± standard deviation based on the five folds. A fixed classification threshold of 0.5 was applied for all evaluations. Fig. 2. Open in a new tab Task-specific ROC curves for video-based classification of UCP severity. a accuracy, b dexterity, c fluency, and d ROM. Receiver operating characteristic (ROC) curves for AI-based classification performance across four MA2 functional indices using 5-fold cross-validation. For each index, clip-level probabilities were averaged within each item-video to obtain a single prediction, and fold-wise ROC curves with the corresponding AUC values are shown at the item-video level Cohen’s kappa coefficients (κ) were calculated for each of the four MA2 indices to assess inter-rater reliability between the AI-based predictions and manual scoring by clinical therapists. In addition, a task-level item contribution analysis was conducted to identify the influence of individual MA2 items on model predictions. For each task, the top items contributing to true positive (TP) and true negative (TN) classifications were quantified. Representative TP and TN cases were further examined to understand how input characteristics influenced the model predictions, particularly in terms of video framing, limb visibility, and therapist occlusion. In addition, for qualitative interpretability of the ViViT-based models, we generated gradient-weighted class activation maps (Grad-CAM) from correctly classified MA2 item-videos to visualize which spatiotemporal regions contributed most to the model predictions (Additional file 10). All model training, inference, and statistical analyses were conducted on an Ubuntu-based workstation equipped with an NVIDIA RTX A6000 GPU using the PyTorch deep-learning framework. Results Evaluation of model performance across MA2 indices The AI-driven automated video analysis system, based on a pretrained ViViT, achieved an overall mean area under the receiver operating characteristic curve (AUC) of 0.890 across the four MA2 indices (ROM, accuracy, fluency, and dexterity) (Fig. 2 ). Specifically, the AI model exhibited the highest classification performance for the dexterity index, with an AUC of 0.936, demonstrating its strong potential to accurately classify impairment severity. The accuracy, fluency, and ROM indices also showed high performance (with AUCs of 0.887, 0.869, and 0.866 respectively). To further evaluate the robustness of the model, the sensitivity and specificity were assessed at a classification threshold of 0.5. Among the four tasks, dexterity consistently exhibited the most balanced and reliable performance, with a sensitivity of 78.8% and specificity of 92.2%, indicating effective classification of both mild and severe cases.The accuracy and ROM indices also demonstrated reasonable trade-offs (accuracy: sensitivity 79.5%, specificity 83.9%; ROM: sensitivity 74.7%, specificity 81.6%), whereas the fluency index showed slightly lower discrimination (sensitivity 70.9%, specificity 81.5%, F1-score 0.665). These results suggest that the model could extract meaningful severity-related features across multiple functional domains, particularly when tasks involve structured and distinguishable motor patterns (Table 2 ). Table 2. Performance of video-based UCP severity classification across tasks Accuracy Precision Sensitivity Specificity F1-score AUC Accuracy 0.821 ± 0.024 0.750 ± 0.069 0.795 ± 0.055 0.839 ± 0.046 0.769 ± 0.036 0.887 ± 0.029 Dexterity 0.884 ± 0.038 0.796 ± 0.103 0.788 ± 0.073 0.922 ± 0.033 0.791 ± 0.083 0.936 ± 0.032 Fluency 0.769 ± 0.061 0.657 ± 0.154 0.709 ± 0.123 0.815 ± 0.090 0.665 ± 0.093 0.869 ± 0.064 ROM 0.784 ± 0.044 0.721 ± 0.107 0.747 ± 0.108 0.816 ± 0.076 0.726 ± 0.066 0.866 ± 0.035 Open in a new tab ROM range of motion In addition, the capture conditions are explicitly described in Table 4 (Item-wise Recording Characteristics). Across items, the recording perspectives included frontal, oblique, lateral, and mixed views; median item-level video durations ranged from 5 to 26.5 s with IQRs reported; items were deliberately repeated during routine recording, typically yielding 2–12 repetitions per item; upper-limb framing was predominantly full for several items (e.g., Items 2 and 10–13 ≥ 95%); and the total number of 32-frame clips used by the model is reported for each item (e.g., Item 8: 1,315; Item 9: 864). Table 4. Item-wise recording characteristics of MA2 videos Item Video n Repetitions per item-video (median [IQR]) Perspective (%) F/O/L/M Duration sec median [IQR] Framing of UL (%) full/partial Clips used† (sum) 1 46 9 [5–10] 0.0 / 4.3 / 95.7 / 0.0 11.5 [10–15] 97.8 / 2.2 641 2 45 9 [6–10] 97.8 / 2.2 / 0.0 / 0.0 11 [8–14] 100.0 / 0.0 529 3 46 8 [6–9] 4.3 / 39.1 / 56.5 / 0.0 9 [5–12] 41.3 / 58.7 533 4 44 3 [2–3] 2.3 / 36.4 / 61.4 / 0.0 11.5 [8.75–18.5] 45.5 / 54.5 559 5 46 8 [5–10] 4.3 / 34.8 / 60.9 / 0.0 8 [7–12] 43.5 / 56.5 424 6 45 7 [5–10] 2.2 / 35.6 / 62.2 / 0.0 8 [6–10] 40.0 / 60.0 453 7 45 8 [5–11] 2.2 / 37.8 / 60.0 / 0.0 6 [5–8] 37.8 / 62.2 304 8 46 3 [2–3] 2.2 / 37.0 / 60.9 / 0.0 26.5 [17–44.75] 45.7 / 54.3 1315 9 40 12 [7–18] 0.0 / 42.5 / 57.5 / 0.0 20.5 [17–26.5] 39.0 / 61.0 864 10 44 5 [4–6] 0.0 / 9.1 / 40.9 / 50.0 12 [8–17] 100.0 / 0.0 612 11 42 4 [3–7] 0.0 / 40.5 / 59.5 / 0.0 11 [7.25–15.75] 95.2 / 4.8 459 12 39 2 [2–4] 46.2 / 48.7 / 5.1 / 0.0 5 [3–7] 100.0 / 0.0 237 13 45 4 [3–6] 60.0 / 40.0 / 0.0 / 0.0 9 [6–13] 100.0 / 0.0 427 14 43 4 [2–5] 0.0 / 39.5 / 60.5 / 0.0 8 [7–14] 95.3 / 4.7 455 Open in a new tab F frontal, O oblique, L lateral, M mixed, UL upper limb † Clips used: Total number of 32-frame clips fed to the model The inter-rater reliability between manual therapist scoring and AI classification was also calculated using Cohen’s kappa (κ). The agreement was found to be moderate-to-good for dexterity (κ = 0.71), moderate for ROM (κ = 0.56), accuracy (κ = 0.58), and fluency (κ = 0.53). Beyond these quantitative metrics, Grad-CAM visualizations indicated that the ViViT models primarily attended to clinically meaningful regions, including the child’s fingers, wrist, elbow, and the task objects, while largely ignoring irrelevant background areas (Fig. 3 ). In fine-motor items, the attention maps were concentrated on fingertip–object contact and in-hand manipulation, whereas in gross reaching items they emphasized proximal joint motion and end-point trajectories of the affected limb. Fig. 3. Open in a new tab Representative Grad-CAM visualizations across MA2 indices. Grad-CAM heatmaps from correctly classified MA2 item-videos show that the index-specific ViViT models primarily attend to distal upper limb segments and task objects rather than background regions. Panels correspond to top-contributing items from the task-level analysis: a Accuracy, true positive – Item 1 (Reach forwards); b Dexterity, true negative – Item 8 (Manipulation); c Fluency, true positive – Item 10 (Reach forehead–back neck); d ROM, true negative – Item 11 (Palm to bottom) Item-wise contribution analysis based on task-level predictions To further investigate the internal mechanisms underlying task-level classification, we analyzed the contribution of each MA2 item in correctly classifying TP and TN predictions. This analysis was conducted by aggregating the results across all five cross validations. The top three contributing items for both the TP and TN outcomes were identified for each MA2 index and are summarized in Table 3 . As shown in Table 3 , the primary items that contributed to true-positive predictions of severe cases differed across tasks. For the Accuracy index, item 1 accounted for the largest proportion of TP classifications (19.0%), followed by Items 9 (18.0%) and 5 (14.2%). For Dexterity, Items 6 (26.0%), 3 (25.6%), and 7 (21.5%) were the greatest contributors. In the Fluency domain, Items 13 (18.2%), 10 (17.8%), and 1 (16.5%) showed the highest TP contributions, whereas for ROM, Items 2 (18.4%), 10 (15.3%), and 1 (13.3%) contributed most strongly. Table 3. Top 3 contributing items to TP and TN predictions by task based on combined results of 5-fold analysis Task TP-Top3 TP-count TP% TN-Top3 TN-count TN% Accuracy Item 1 116 19.0 Item 9 26 52.0 Item 9 110 18.0 Item 5 9 18.0 Item 5 87 14.2 Item 13 8 16.0 Dexterity Item 6 120 26.0 Item 8 271 86.9 Item 3 118 25.6 Item 4 197 8.3 Item 7 99 21.5 Item 3 26 2.9 Fluency Item 13 86 18.2 Item 8 84 57.5 Item 10 84 17.8 Item 10 24 16.4 Item 1 78 16.5 Item 14 20 13.7 ROM Item 2 109 18.4 Item 11 46 49.5 Item 10 91 15.3 Item 10 14 15.1 Item 1 79 13.3 Item 5 14 15.1 Open in a new tab TP True Positive (correct classification of severe cases); TN True Negative (correct classification of mild cases). The values represent the item-wise frequency and proportion of contribution among all correctly predicted cases per task based on the combined results from all five cross-validation folds In contrast, the items contributing to the correct classification of mild cases (TN) exhibited different patterns. For the Accuracy task, Item 9 was the dominant contributor to TN outcomes (52.0%), followed by Items 5 (18.0%) and 13 (16.0%). For Dexterity, Items 8 (53.2%), 9 (38.7%), and 4 (5.1%) contributed most to TN classifications. In the Fluency task, Items 8 (57.5%), 10 (16.4%), and 14 (13.7%) were the leading contributors to TN classifications. For the ROM index, Items 11 (49.5%), 10 (15.1%), and 5 (15.1%) accounted for the largest proportions of correctly classified mild cases. The corresponding task descriptions for each MA2 item are presented in Additional file 11 to facilitate interpretation (Table 4 ). To examine how input video characteristics influenced the model predictions, representative cases were selected and visualized for each prediction outcome category TP, TN, false positive (FP) and false negative (FN) (Fig. 4 ). The TP (a; Item 3) and TN (b; Item 2) cases both featured clear visibility of the patient’s upper limbs and trunk within the camera frame, with the TP example captured from an oblique diagonal viewpoint and the TN example from a stable frontal view, which likely facilitated accurate recognition of movement execution. In contrast, the FP case (c; Item 10) was recorded from behind the child, and the examiner’s body partially overlapped the affected arm, making the child’s actual movement difficult to observe and potentially leading the model to overestimate impairment. The FN case (d; Item 11) involved the palm-to-bottom movement; because the child’s clothing obscured the distal upper limb, the model may have underestimated the true movement limitation, resulting in misclassification. Fig. 4. Open in a new tab Representative cases of true positive (TP; a ), true negative (TN; b ), false positive (FP; c ), and false negative (FN; d ) predictions. Representative video frames illustrating true positive (TP; a ), true negative (TN; b ), false positive (FP; c ), and false negative (FN; d ) classification outcomes Discussion This preliminary study examined the feasibility of using an AI-driven automated video analysis approach to evaluate upper limb function in children with UCP using MA2. Despite the retrospective and nonstandardized clinical recording conditions, our findings indicate that MA2-based automated video analysis can reliably provide clinically meaningful severity classification across all four indices in real-world settings. These findings support the potential application of such systems as a complementary tool to manual scoring in pediatric rehabilitation. This study demonstrated that a fine-tuned pretrained ViViT model could robustly classify the dichotomized upper limb impairment severity from retrospective clinical videos of children with UCP, achieving a mean AUC of 0.890 across all four MA2 indices. Although the tasks, scoring scales, and target populations differ, these results fall within the performance ranges reported in prior upper-limb video–based deep learning studies in adults post-stroke. For example, in adult stroke cohorts, a transformer-based model achieved 89% classification accuracy in automated scoring of the Action Research Arm Test (ARAT), a clinical assessment focusing on grasping and object manipulation [ 16 ]. Similarly, the HRTR single-stage transformer achieved state-of-the-art performance in segmenting fine-grained upper limb actions during stroke rehabilitation [ 17 ], and Shi et al. reported that a CNN–LSTM model using Mediapipe-extracted keypoints classified Fugl–Meyer–based upper limb movements with 97.5% accuracy [ 18 ]. Although this study differs from previous works in population (pediatric UCP vs. adult stroke), assessment tool (MA2 vs. ARAT/FMA), and outcome definition (dichotomized severity vs. multi-class scores or action labels), the evidence collectively supports the suitability of transformer-based and related architectures for analyzing functional upper limb movement across heterogeneous clinical contexts, while highlighting the need for further refinement of task design, labeling schemes, and dataset scale before direct cross-study comparisons. In this study, we deliberately framed MA2 severity estimation as a binary classification problem (mild vs. severe). Although this approach reduces the detailed 0–3/0–4 MA2 scoring system to a single decision boundary, the AUCs achieved across all four indices (0.866–0.936) indicate that this coarse-grained approach still captures clinically meaningful differences between milder and more impaired upper-limb function. In the context of a modest, real-world pediatric dataset, dichotomization provides a pragmatic initial approach that reduces label noise and class imbalance, consistent with many early machine-learning studies in neurorehabilitation that similarly adopted coarse binary or categorical formulations [ 19 , 20 ]. Collectively, these findings suggest that the current model should be interpreted as a feasibility baseline for binary risk stratification using routine clinical videos, while highlighting the need for future studies with larger, more standardized pediatric cohorts to extend toward ordinal or regression-based MA2 scoring and richer movement representations capable of capturing subtle within-class differences relevant for goal-setting and longitudinal monitoring. Analysis of item-wise contributors revealed an asymmetry in how the model arrived at severe (TP) versus mild (TN) classifications. For all four indices, TP contributions were relatively evenly distributed across multiple items, whereas TN contributions were dominated by a single “high-yield” task in each domain (e.g., accuracy: Item 9; dexterity: Item 8; fluency: Item 8; ROM: Item 11). Thus, the model “spread” its evidence for severe impairment across multiple tasks, but often relied on one or two canonical items to confirm that function was mild. Several factors likely underlie this pattern. First, TP counts aggregate across all severe cases, which are more heterogeneous impairment patterns (e.g., different children may “fail” different tasks), causing the model to draw on a broader set of items to detect severity. Second, TN percentages are calculated only within ground-truth mild cases. For certain items, such as Item 8 in the Dexterity index, the absolute number of mildly impaired clips is small; however, when these clips are present and clearly intact, they contribute disproportionately to TN classifications. Overall, these observations suggest that the network uses a more diffuse item set to recognize severe impairment while relying on a small number of highly informative tasks to confirm mild status. This exploratory finding should be interpreted cautiously but aligns conceptually with prior work in adults after stroke, which shows that reduced subsets of upper-limb items from the ARAT capture most of the information in the full battery [ 21 ]. This pattern may inform the design of shortened, AI-oriented MA2 item subsets in future studies. Analysis of representative prediction cases further highlighted the influence of input video characteristics on the model’s predictions. Specifically, FP and FN cases often resulted from misleading visual cues, such as waiting periods before task initiation or therapist interventions that obscure the child’s movements, rather than from actual motor performance deficits. These observations indicate that AI-based video analysis depends not only on functional severity but also on visual clarity, task framing, and the precise segmentation of task segments [ 22 , 23 ]. In children with CP, neurodevelopmental comorbidities such as attention-deficit/hyperactivity disorder (ADHD) are highly prevalent, thus contributing to difficulties in sustaining attention during structured task performance [ 24 ]. Under such conditions, practical constraints, including limited attention spans, required therapist involvement, and continuous video recordings that capture off-task behaviors, can introduce noise into the video data, thereby complicating automated analyses. These factors likely account for a substantial proportion of the remaining false-positive and false-negative classifications, even when the overall AUC remains relatively high. Although evidence specific to pediatric CP remains limited, general video analytics studies have revealed that environmental variables such as lighting variations and visual artifacts can destabilize classification accuracy [ 25 ]. Nevertheless, under structured conditions with visually clear task execution, our model consistently provided reliable severity classification. This supports its potential for clinical implementation, and with standardized protocols, feasible adaptation to home-based monitoring or decision-support applications in pediatric rehabilitation. Despite the limitations inherent to the preliminary study design, including the relatively small sample size, retrospective design, narrow participant age range, and simplified binary severity classification, our findings offer substantial promise for practical implementation. Furthermore, because of the modest sample size, we did not reserve an independent external test set; instead, all data were used within a subject-wise five-fold cross-validation framework, which may limit the generalizability of the reported performance estimates. Unlike prior studies that relied on highly controlled laboratory setups or specialized motion capture systems [ 26 ], the present study utilized handheld, nonstandardized clinical video recordings that reflected the visual quality and variability typical of real-world conditions, similar to those captured by parents or caregivers in home settings. This enhances the ecological validity of our findings and indicates that the proposed AI-based system can function robustly, even under suboptimal recording conditions. Future research should therefore focus on developing standardized recording guidelines, the incorporation of multimodal sensor data, and prospective validation in larger and diverse cohorts to significantly enhance the clinical utility, accuracy, and generalizability of AI-based automated assessments for pediatric upper limb rehabilitation. We initiated a prospective follow-up employing a simplified standardized capture protocol to reduce input variability without increasing clinical burden. This protocol involves two fixed cameras, where feasible, consistent subject-to-camera distance, full upper limbs in frame without zoom, and a stable frame rate (≥ 30 fps). Additionally, three actigraph sensors are placed on the affected limb (dorsum of hand, wrist, and upper arm), sampling at 100 Hz to enable multimodal modeling. We also plan to expand the cohort and enhance data augmentation to improve the robustness and generalizability of the models. Conclusion This preliminary study provides initial evidence that AI-driven automated video analysis can feasibly and objectively classify dichotomized upper limb impairment severity in pediatric UCP populations using MA2, even under non-standardized, real-world clinical recording conditions. Across all four MA2 indices, the fine-tuned ViViT achieved consistently high discrimination (AUCs ≈ 0.87–0.94), supporting its potential utility as a screening and triage tool rather than a full replacement for detailed MA2 scoring. However, the reliance on a binary (mild vs. severe) classification, a modest single-center cohort, and single-camera RGB video without depth or kinematic measurements underscores key limitations and necessitates cautious interpretation. As one of the first investigations to explore automated MA2 scoring without standardized video acquisition, our findings highlight the potential and practical challenges of integrating AI-based tools into routine clinical practice. Future research should therefore focus on developing standardized video recording guidelines, validating multimodal assessment methods that integrate complementary wearable sensors, and conducting larger prospective validation studies to improve clinical validity, reliability, and generalizability. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. (49.7KB, docx) Supplementary Material 2. (27.7KB, docx) Supplementary Material 3. (28KB, docx) Supplementary Material 4. (28.1KB, docx) Supplementary Material 5. (27.9KB, docx) Supplementary Material 6. (27.8KB, docx) Supplementary Material 7. (28.4KB, docx) Supplementary Material 8. (32.4KB, docx) Supplementary Material 9. (32.9KB, docx) Supplementary Material 10. (567.3KB, docx) Supplementary Material 11. (5.5MB, docx) Acknowledgements We acknowledge the families who agreed to participate in this study. Abbreviations AUC Area under the curve FN False negative FP False positive MA2 Melbourne assessment 2 MACS Manual ability classification system ROM Range of motion ROC Receiver operating characteristic SGD Stochastic gradient descent SOP Standard operating procedure TN True negative TP True positive UCP Unilateral cerebral palsy Author contributions Conceptualization: Y.H. and H.Y.; Methodology: Y.H., H.Y., and J.-Y.K.; Data curation: Y.H., M.K., and H.Y.; Formal analysis and investigation: H.Y. and Y.H.; Software and model development: H.Y., M.K., and M.J.C.; Writing—original draft: Y.H. and H.Y.; Writing—review and editing: Y.H., H.Y., J.-Y.K., and M.J.C.; Visualization: H.Y.; Funding acquisition: J.-Y.K.; Resources: J.-Y.K. and M.J.C.; Supervision: J.-Y.K. and M.J.C.Y.H. and H.Y. contributed equally to this work and share first authorship. J.-Y.K. and M.J.C. jointly supervised this work and shared corresponding authorship. Funding This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (No. RS-2023-00278700), and Future Medicine 20 × 30 Project of the Samsung Medical Center (SMO1250071). Data availability Data supporting the findings of this study can be obtained from the corresponding author upon reasonable request and ethical approval. Declarations Ethics approval and consent to participate The study was conducted in accordance with the guidelines of the Declaration of Helsinki and was approved by the Ethics Committee of the Samsung Medical Center. Written informed consent for participation was obtained from all children’s guardians. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Youngsub Hwang and Hakje Yoo have contributed equally to this work. Contributor Information Myung Jin Chung, Email: [email protected]. Jeong-Yi Kwon, Email: [email protected]. References 1. Reid LB, Rose SE, Boyd RN. Rehabilitation and neuroplasticity in children with unilateral cerebral palsy. Nat Reviews Neurol. 2015;11(7):390–400. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Sakzewski L, Carlon S, Shields N, Ziviani J, Ware RS, Boyd RN. Impact of intensive upper limb rehabilitation on quality of life: a randomized trial in children with unilateral cerebral palsy. Dev Med Child Neurol. 2012;54(5):415–23. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Shishov N, Melzer I, Bar-Haim S. Parameters and measures in assessment of motor learning in neurorehabilitation; a systematic review of the literature. Front Hum Neurosci. 2017;11:82. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Randall M, Johnson L, Reddihough D. The melbourne assessment 2. Melbourne: Royal Children’s Hospital; 1999. [ Google Scholar ] 5. Bourke-Taylor H. Melbourne assessment of unilateral upper limb function: construct validity and correlation with the pediatric evaluation of disability inventory. Dev Med Child Neurol. 2003;45(2):92–6. [ PubMed ] [ Google Scholar ] 6. Gerber CN, Plebani A, Labruyère R. Translation, reliability, and clinical utility of the Melbourne assessment 2. Disabil Rehabil. 2019;41(2):226–34. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Cusick A, Vasquez M, Knowles L, Wallen M. Effect of rater training on reliability of Melbourne Assessment of Unilateral Upper Limb Function scores. Dev Med Child Neurol. 2005;47(1):39–45. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Ahmed T, Thopalli K, Rikakis T, Turaga P, Kelliher A, Huang JB, et al. Automated movement assessment in stroke rehabilitation. Front Neurol. 2021;12:720650. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Mailleux L, Jaspers E, Ortibus E, Simon-Martinez C, Desloovere K, Molenaers G, et al. Clinical assessment and three-dimensional movement analysis: An integrated approach for upper limb evaluation in children with unilateral cerebral palsy. PLoS ONE. 2017;12(7):e0180196. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Kim WS, Cho S, Baek D, Bang H, Paik NJ. Upper extremity functional evaluation by Fugl-Meyer assessment scoring using depth-sensing camera in hemiplegic stroke patients. PLoS ONE. 2016;11(7):e0158640. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Sauerzopf L, Panduro CGC, Luft AR, Kühnis B, Gavagnin E, Unger T, et al. Evaluating inter-and intra-rater reliability in assessing upper limb compensatory movements post-stroke: creating a ground truth through video analysis? J Neuroeng Rehabil. 2024;21(1):217. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Yang C, Kerr A, Stankovic V, Stankovic L, Rowe P, Cheng S. Human upper limb motion analysis for post-stroke impairment assessment using video analytics. IEEE Access. 2016;4:650–9. [ Google Scholar ] 13. Eliasson AC, Krumlinde-Sundholm L, Rösblad B, Beckung E, Arner M, Öhrvall AM, et al. The Manual Ability Classification System (MACS) for children with cerebral palsy: scale development and evidence of validity and reliability. Dev Med Child Neurol. 2006;48(7):549–54. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Randall M, Imms C, Carey LM, Pallant JF. Rasch analysis of The Melbourne Assessment of Unilateral Upper Limb Function. Dev Med Child Neurol. 2014;56(7):665–72. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Arnab A, Dehghani M, Heigold G, Sun C, Lučić M, Schmid C. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision. 2021;pp. 6836–6846. 16. Ahmed T, Rikakis T, Automated ARAT. Scoring Using Multimodal Video Analysis, Multi-View Fusion, and Hierarchical Bayesian Models: A Clinician Study. arXiv preprint arXiv. 2025;2505.01680. 17. Helvaci HI, Huber JP, Bae J, Cheung SCS. HRTR: A Single-stage Transformer for Fine-grained Sub-second Action Segmentation in Stroke Rehabilitation. arXiv preprint arXiv. 2025;2506.02472. 18. Shi L, Wang R, Zhao J, Zhang J, Kuang Z. Detection of rehabilitation training effect of upper limb movement disorder based on MPL-CNN. Sensors. 2024;24(4):1105. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Wallich M, Lai K, Yanushkevich S. Assessing upper limb motor function in the immediate post-stroke period using accelerometry. In: 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE; 2023:215–220. 20. Tozlu C, Edwards D, Boes A, Labar D, Tsagaris KZ, Silverstein J, et al. Machine learning methods predict individual upper-limb motor impairment following therapy in chronic stroke. Neurorehabil Neural Repair. 2020;34(5):428–39. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Burton Q, Lejeune T, Dehem S, Lebrun N, Ajana K, Edwards MG, et al. Performing a shortened version of the Action Research Arm Test in immersive virtual reality to assess post-stroke upper limb activity. J Neuroeng Rehabil. 2022;19(1):133. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Xiao Z, Xia Z, Zheng H, Zhao BY, Jiang J. Towards performance clarity of edge video analytics. In 2021 IEEE/ACM Symposium on Edge Computing (SEC); 2021; pp. 148–164. IEEE. 23. Perumalla C, Kearse L, Peven M, Laufer S, Goll C, Wise B, et al. AI-based video segmentation: procedural steps or basic maneuvers? J Surg Res. 2023;283:500–6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Påhlman M, Gillberg C, Himmelmann K. Autism and attention-deficit/hyperactivity disorder in children with cerebral palsy: high prevalence rates in a population‐based study. Dev Med Child Neurol. 2021;63(3):320–7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Paul S, Rao K, Coviello G, Sankaradas M, Po O, Hu YC et al. Why is the video analytics accuracy fluctuating, and what can we do about it? In: European Conference on Computer Vision; 2022; Cham: Springer Nature Switzerland. pp. 430–448. 26. Valevicius AM, Jun PY, Hebert JS, Vette AH. Use of optical motion capture for the analysis of normative upper body kinematics during functional upper limb tasks: a systematic review. J Electromyogr Kinesiol. 2018;40:1–15. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1. (49.7KB, docx) Supplementary Material 2. (27.7KB, docx) Supplementary Material 3. (28KB, docx) Supplementary Material 4. (28.1KB, docx) Supplementary Material 5. (27.9KB, docx) Supplementary Material 6. (27.8KB, docx) Supplementary Material 7. (28.4KB, docx) Supplementary Material 8. (32.4KB, docx) Supplementary Material 9. (32.9KB, docx) Supplementary Material 10. (567.3KB, docx) Supplementary Material 11. (5.5MB, docx) Data Availability Statement Data supporting the findings of this study can be obtained from the corresponding author upon reasonable request and ethical approval. Articles from Journal of NeuroEngineering and Rehabilitation are provided here courtesy of BMC ACTIONS View on publisher site PDF (1.7 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top