This study was conducted in accordance with relevant laws and institutional guidelines. Ethics approval was granted by the Auckland Health Research Ethics Committee (AHREC; reference number AH26160) and the institutional ethics committee of Shanghai Eber Hospital (20231127).

All participating professionals provided informed consent before taking part in the co-design process. For the use of children’s clinical gait video data, informed consent was obtained from their legal guardians under the approved ethics protocols. All clinical video and assessment data used in this study were deidentified before analysis. Participant information, interview transcripts, observation notes, questionnaire responses, and system-generated outputs were stored securely and reported only in deidentified or aggregated form. No identifiable participant information is presented in this manuscript. No financial compensation was provided to study participants.

This study followed FRESCO [22] to iteratively develop and formatively evaluate a clinically oriented VGA decision-support prototype for children with CP. Operationalization of FRESCO in this study comprised five steps, each ending with concrete deliverables that informed the next cycle: (1) establishing a multidisciplinary advisory group to identify clinical needs and define initial functional specifications; (2) developing an initial prototype draft to address these needs from a technical perspective; (3) conducting think-aloud usability evaluations with representative rehabilitation professionals to identify additional requirements and refine the design; (4) testing the prototype in clinical simulation to assess workflow integration and safety; and (5) facilitating a final co-design workshop to converge feedback and generate a release-ready prototype. Because this study focused on a clinician-centered VGA system, the initial design task was to define the clinical outputs, interpretive support, and workflow requirements needed for routine CP gait assessment rather than to select between alternative prototype concepts [29]. Therefore, in this study, baseline clinical and workflow needs were first identified and used to develop a single needs-led prototype (design 1), which was then iteratively refined through subsequent co-design activities. In the text below, we explain how each step of the framework guided the co-design of the prototype gait analysis system to help CP rehabilitation (Figure 1).

Based on the baseline requirements identified in step 1, the first prototype (design 1) was developed by biomechanics researchers using BlazePose. BlazePose adopts a lightweight pose estimation algorithm (PEA), allowing convenient deployment on mobile or CPU-grade devices [30]. An automated rule-based framework was established to process PEA-derived key points from recorded videos, enabling functions like extracting gait cycles, calculating joint angles and spatiotemporal parameters, detecting gait deviations, and generating PDF reports. At this stage, emphasis was placed on implementing core functions, while aspects of usability and clinical integration were deferred to later iterations.

In the next stage, a human factors engineer reviewed design 1 through a usability heuristics evaluation (ie, 10 heuristics by Nielsen) [31] and a context-of-use analysis. This analysis examined factors affecting system integration, including intended users, tasks, physical environment, organizational setting (such as required clinical workflows and staffing roles), and technical constraints. This early expert-based review aimed to identify potential usability issues before involving representative end users. The identified issues and contextual considerations (not fully captured during step 1) were then incorporated into the system requirements to strengthen reliability and clinician confidence. The combined findings informed a refinement of design 2, which, with support from a graphic designer, incorporated usability principles and contextual considerations while retaining the core functional components of design 1.

After developing the prototype, we conducted online think-aloud formative usability sessions with 19 rehabilitation professionals from Shanghai Eber Medical Group in December 2024. These sessions were aimed at examining clinicians’ cognitive processes during interactions with the prototype and identifying usability barriers and their underlying causes among a small group of representative end users.

Participants were recruited from the group’s hospital sites in Shanghai, Qingdao, and Tianjin, using the purposive sampling criteria described in the Composition and Roles of the Advisory Group section. This sample included all 12 clinical members of the multidisciplinary advisory group and 7 additional rehabilitation professionals from the same clinical network. The additional 7 participants were included to broaden formative feedback beyond the core advisory group and to capture views from clinicians working across different hospital sites within the same rehabilitation network.

Each online session lasted approximately 60 minutes and was facilitated by the lead author (XG), a rehabilitation engineering researcher involved in prototype development. To reduce facilitator influence, all sessions followed the same task script, semistructured interview guide, and neutral prompting strategy. At the beginning of each session, the facilitator introduced the purpose of the evaluation, explained the think-aloud procedure, and reminded participants that the aim was to evaluate the prototype rather than their clinical performance. Participants were then asked to complete a set of typical clinical tasks using design 2. These tasks included reviewing a consistent CP gait case, opening or uploading the gait video in the prototype, running the video-based analysis, checking the pose-estimation and gait-cycle outputs, reviewing spatiotemporal parameters and joint-angle plots, interpreting automatically identified gait deviations, and reviewing the automatically generated report. During task completion, participants were instructed to verbalize their thoughts continuously while performing these tasks, following the concurrent think-aloud protocol. When participants became silent for more than a few seconds, the facilitator prompted them with neutral cues (eg, “please keep talking”) without directing their responses. All sessions were audio recorded, and the prototype interface was screen recorded to capture interactions. A second researcher (WL) attended the sessions as a nonintervening observer and took field notes on navigation difficulties, hesitation, points of confusion, interpretation difficulties, and workflow-related comments. Immediately after the tasks, participants filled out the System Usability Scale (SUS) questionnaire to rate the usability of the current prototype [32]. Subsequently, the moderator used a semistructured interview guide (see Multimedia Appendix 1) to clarify observations from the think-aloud exercise and to elicit further feedback on system functions, report interpretability, perceived benefits, and workflow compatibility.

The SUS questionnaire was scored using the standard 0‐100 scoring procedure. Overall usability was summarized as mean (SD), and individual item scores were inspected to identify aspects of the prototype that might require further refinement.

Feedback from the think-aloud exercise, semistructured interviews, and field notes was analyzed using a qualitative descriptive approach [33-35]. This approach was selected to provide a direct, low-inference summary of end-user feedback for formative prototype refinement. The lead author reviewed the recordings and field notes and organized participant feedback according to the main domains of the semistructured interview guide. Relevant feedback segments were descriptively labeled and grouped within these domains. Repeated comments were then summarized into perceived benefits and practical refinement categories, including functional expansion, report interpretability, and workflow compatibility. Additional end-user requirements were also identified to extend the baseline specifications defined in step 1. These findings informed subsequent prototype iterations, leading to design 3.

To complement the usability evaluation in step 3, prototype design 3 was tested through in-person clinical simulations at Shanghai Eber Hospital, part of Shanghai Eber Medical Group, in August 2025 (see Figure 1). The aim of the clinical simulation was to assess workflow integration and explore use-related risks associated with prototype use in a realistic routine CP gait-assessment scenario. Rather than using separate risk-specific scenarios, the simulation explored potential risks as they arose during the routine video-based gait-assessment workflow.

Twelve rehabilitation professionals, including 6 physicians and 6 therapists, were recruited from Shanghai Eber Hospital, using the purposive sampling criteria described in the Composition and Roles of the Advisory Group section. Six of the 12 participants were members of the multidisciplinary advisory group and had also participated in the step 3 think-aloud usability evaluation. The remaining 6 were additional clinicians from Shanghai Eber Hospital, 2 of whom had also participated in step 3. Participants represented a range of professional seniority, with clinical experience spanning approximately 3 to 15 years in pediatric rehabilitation.

Simulations took place in their own clinical settings where care is routinely performed in order to maximize ecological validity. All child patients involved in the simulation were routine cases under the care of these clinicians. Parental informed consent was obtained for the use of their video data as part of the clinical simulation and research evaluation. Each clinician completed the same standardized scenario: assessing the gait of a child with CP as part of their routine rehabilitation session. The scenario was performed under 2 conditions: conventional observational gait analysis (OGA) and VGA using the prototype (design 3). To minimize sequence effects, the order of conditions was counterbalanced, with half of the participants (n=6) performing OGA first followed by the prototype, and the other half (n=6) completing the scenario in the reverse order. The inclusion of the conventional OGA condition was not intended to enable a quantitative comparison of diagnostic performance between OGA and VGA. Instead, it served as a familiar reference workflow to contextualize clinicians’ experiences with the prototype and to support the identification of workflow integration issues, interpretability challenges, and potential use-related risks that might not be evident in single-condition testing. In both conditions, participants carried out the full workflow, including video capture, gait analysis, and clinical interpretation of the results. All sessions were audio recorded, and structured field notes were taken to capture observations on workflow integration, system operation, clinical interpretation, and any patient physical safety–relevant issues. Following each pair of simulations, participants took part in a short debriefing discussion to reflect on their experience and highlight potential barriers or risks (see Multimedia Appendix 2 for topic guide questions in this discussion).

Data from the clinical simulations were analyzed using the same qualitative descriptive approach. The analysis was organized around the simulation objectives, including workflow integration, system stability, use-related physical safety, clinical interpretation risks, and prototype refinement needs. Field notes, recordings, and debriefing comments were reviewed by task stage, including video capture, system processing, report review, and clinical interpretation. Relevant observations and comments were extracted and mapped to the corresponding objective domains. Observed delays, system interruptions, error messages, technical-support needs, use errors, manual corrections, safety considerations during video capture, and interpretation-related concerns were summarized descriptively. The findings were then narratively summarized and used to guide subsequent prototype refinement, leading to design 4.

A co-design workshop was conducted with the advisory group and key stakeholders (ie, service leads and IT) to consolidate step 4 findings and prepare a release-candidate system. The workshop targeted consensus on three outputs: (1) the final prototype specification (report layout, parameter set and defaults, interaction wording, frozen algorithm/rule set, etc), (2) a standardized clinical workflow (roles, inputs/outputs, timing, and exception handling), and (3) an implementation guideline covering data capture and storage procedures, system operation guidelines, integration requirements, and training materials.

Sessions followed a progressive, item-by-item confirmation process. Each item represented a key component of the system specification, workflow, or implementation plan that had been refined during step 4. These items covered elements, such as report layout and terminology, parameter defaults, workflow roles and timing, data storage procedures, and user training materials. For each item, the facilitator presented the relevant evidence and proposed solution. Members then discussed and annotated the proposal within their respective disciplinary subgroups before reconvening in a plenary session to address any remaining disagreements or points requiring clarification. Agreement was recorded in a decision log, and unresolved items were iteratively refined until reaching ≥80% assent or resolved by chair adjudication in case of ties. This structured process ensured balanced input across clinicians, engineers, and service leads.

The advisory group discussions, supported by a review of existing systems, identified a baseline set of clinical needs and corresponding system requirements for a VGA system in children with CP (Table 1). These centered on three domains: (1) system content, (2) system operation and workflow integration, and (3) clinical usability and support. Collectively, these baseline needs provided the foundation for developing the first prototype (design 1).

The baseline clinical needs and system requirements identified in step 1 guided the development of the first draft prototype (design 1). Emphasis at this stage was placed on implementing system core functions, including visualization of pose recognition, automated and manual gait cycle segmentation, calculation of spatiotemporal parameters, estimation of joint angles, rule-based detection of selected gait deviations, and automated generation of a summary PDF report. The system architecture and algorithmic choices were made based on clinical feasibility and interpretability. MediaPipe BlazePose was selected for pose estimation due to its lightweight implementation, sufficient accuracy for sagittal-plane analysis on standard CPUs [36], and ability to provide foot key-point estimates that enable the assessment of foot and ankle abnormalities. Spatiotemporal and joint angle calculations followed standard biomechanical definitions to ensure clinical relevance, while deviation detection rules were derived from published CP gait literature and refined with input from experienced clinicians.

A subsequent heuristic evaluation and context-of-use analysis of design 1 identified additional usability and contextual requirements beyond those captured in step 1 (Table 1). The heuristic review (based on the 10 principles by Nielsen [31]) revealed several issues related to system safety and transparency, such as the need for clear feedback after each operation and warnings when errors occurred. In parallel, the context-of-use analysis (key components summarized in Multimedia Appendix 3) highlighted practical considerations for real-world clinical deployment, including the range of intended users, routine clinical environments, typical patient presentations, and technical constraints. These findings informed the refinement of design 1 (design 2; see Multimedia Appendix 4), which incorporated visual hierarchy through standardized font size, color, and layout adjustments, along with real-time visual feedback and warning prompts for error prevention. To enhance scalability, the system was also deployed as a web-based interface accessible through a simple URL, enabling easy use without local installation. With these refinements, design 2 emerged as an initial draft prototype that was technically stable and ready for subsequent end-user evaluations. Examples of these refinements are provided in Multimedia Appendix 3.

Clinical simulations were conducted in real clinical settings to evaluate workflow integration, safety, and potential use-related risks associated with the prototype (design 3). The main findings are summarized in Table 3.

One key discussion point was the system’s practicality in real clinical workflows. Clinicians confirmed that the prototype could be operated independently without technical support, which streamlined the gait assessment process by reducing manual notetaking and interpretation time. The system functioned stably throughout all simulations, and no safety incidents occurred. Nonetheless, given the inherent fall risk for patients with balance difficulties during gait assessment, clinicians emphasized the need for adaptive safety guidelines and flexible supervision protocols during video capture. This flexibility is necessary to allow for immediate adjustments, such as micro-adjustments to the environment or the therapist’s position, ensuring patient safety across varied clinical settings. Another major point concerned the interpretability of automatically generated reports. Participants noted occasional differences in how deviation labels were understood and highlighted the risk of overreliance on automated outputs.

The simulation findings collectively informed the optimization of the prototype, leading to design 4, which incorporated combined automatic-manual segmentation, a clearer and more intuitive report layout, and a concise user guide with safety instructions to support standardized operation (see Multimedia Appendix 4).

The co-design workshop consolidated the findings from the clinical simulation and resulted in consensus across the multidisciplinary advisory group on the final system specification, standardized workflow, and implementation guideline.

The group confirmed the final prototype (design 5; see Multimedia Appendix 5), which retained all validated analytical functions while incorporating refinements to report layout, terminology, and visual hierarchy for improved clarity and interpretability. Agreement was also reached on a standardized clinical workflow detailing roles, inputs, and timing to ensure consistent use across settings. Figure 2 illustrates how clinicians use the VGA system during routine consultations, from video capture and automated analysis to report generation and interpretation. In parallel, a comprehensive implementation guideline was developed, covering data capture and storage procedures, integration with existing documentation systems, and concise training materials for clinical staff (see Multimedia Appendix 6). The workshop achieved high agreement among participants (≥80%) on all key items, completing the co-design phase and preparing the system for large-scale evaluation.

As the system is designed for Chinese-speaking clinicians, the operational interface of the deployed system is presented in Chinese. For this manuscript, the interface screenshots in Multimedia Appendix 5 have been translated into English and are provided for illustrative display only.

This study applied a structured co-design framework (FRESCO) to develop and iteratively refine a VGA system tailored for the clinical assessment of children with CP. Beyond the development of the prototype itself, the study also examined how a structured co-design process can support the translation of a research-based VGA system into a clinician-centered decision-support tool. In doing so, it makes three design-oriented contributions: (1) establishing baseline clinical and workflow requirements before prototype refinement; (2) translating VGA-derived kinematic parameters into clinician-centered and decision-relevant outputs; and (3) incorporating multidisciplinary feedback and clinical workflow simulation into iterative refinement of usability, safety, and implementation fit. Through 5 progressive steps, the system evolved from a conceptual prototype into a release-candidate decision-support tool (final prototype) that is technically stable and ready for large-scale clinical evaluation. This formative process demonstrated how multidisciplinary collaboration among rehabilitation clinicians, engineers, and biomechanics experts can translate clinical requirements for gait assessment into an interpretable and workflow-compatible digital solution.

Several findings were directly supported by the current formative data. Across the context-of-use analysis, think-aloud usability evaluation, and clinical simulation, clinicians emphasized the importance of workflow fit, low operational burden, interpretable outputs, and retaining professional control over final clinical interpretation. Building on these findings, we derived two data-informed design principles that may be transferable to other clinician-centered digital health systems: (1) supporting workflow-compatible implementation by minimizing technical and operational barriers, and (2) prioritizing clinical autonomy and actionability. As these principles were developed from a single-site formative co-design study, they are intended to provide practical reference points for future system development rather than definitive implementation rules.

The first principle, supporting workflow-compatible implementation by minimizing technical and operational barriers, was grounded in findings from the context-of-use analysis and clinical simulation. These findings showed that clinicians needed a system that could fit within routine outpatient workflows without requiring dedicated hardware, complex setup procedures, software installation, or specialized training in motion capture. Unlike traditional marker-based motion capture systems that require specialized laboratory environments, controlled setups, and lengthy postprocessing [6], our system was intentionally designed for use in routine outpatient settings. By relying on routine video recordings without introducing new data acquisition procedures, the system integrates directly into established assessment workflows. To further minimize technical and logistical barriers to adoption, the system can be deployed in a lightweight, web-based format, allowing flexible access via a simple URL [38,39]. This design choice allowed the system to function within the existing physical and technological infrastructure of typical rehabilitation clinics. Moreover, the end-to-end video-to-report pipeline ensures that clinicians can capture, process, and interpret gait data within the time constraints of a standard consultation, thereby supporting seamless integration into clinical workflows.

The second principle, prioritizing clinical autonomy and actionability, was grounded in findings from the think-aloud usability evaluation and clinical simulation. Participants valued clinically meaningful and interpretable outputs, but the simulation also highlighted the need to avoid overreliance on automated outputs and to preserve clinicians’ control over the final interpretation. In contrast to systems that primarily present final predictions without explicit representation of intermediate analytical steps [17,40,41], the proposed system was intentionally designed to support stepwise clinical reasoning rather than automated decision-making. Specifically, system outputs are structured to align with clinicians’ established reasoning processes, progressing from raw kinematic information to identified gait deviations and, where appropriate, to potential contributing biomechanical or neuromuscular factors. This transparent, staged representation enables clinicians to evaluate, contextualize, and, if necessary, challenge system-generated assessments based on their professional expertise and patient-specific knowledge. By providing interpretable and clinically contextualized outputs within a stepwise decision-support pipeline, the system facilitates informed clinical judgment while preserving clinicians’ control over final interpretation and decision-making. In this way, clinical autonomy is supported through meaningful engagement with the reasoning process underlying system outputs. This design perspective aligns with the sociological and informatics accounts of clinical autonomy, which emphasize the need for clinicians to retain interpretive control over diagnostic and evaluative processes through access to transparent and interpretable reasoning pathways, rather than being limited to opaque, outcome-only system outputs that obscure how assessments are derived [42,43].

Furthermore, the study also highlighted how combining co-design with a mixed methods approach helped strengthen both the design process and the evaluation outcomes. Specifically, the integration of iterative qualitative insights with quantitative usability metrics (eg, SUS scores) provided systematic triangulation, enabling the active engagement of clinicians to mutually inform experiential feedback and quantitative results. This ensured design decisions remained closely aligned with complex clinical practice and contributed directly to the system’s high usability and workflow compatibility, leading to more robust and generalizable findings [44,45].

The formative usability evaluation (step 3) confirmed a good usability score (SUS score: mean 75.8, SD 7.5), while detailed item analysis revealed 2 items with notably lower ratings: “I need the support of a technical person to be able to use this system” and “I need to learn a lot of things before I could get going with this system.” These responses suggest that while the prototype was intuitive in operation, participating clinicians and therapists still lacked technical and analytical confidence in interpreting gait data. This observation aligns with previous reports that the main barrier to using quantitative gait analysis technologies in practice is clinicians’ limited data expertise and difficulty interpreting kinematic information [16,46-48]. Recent evidence suggests that providing clinicians with structured reports, visual summaries, and key result highlights can improve interpretability and facilitate clinical adoption of gait technologies [11]. Accordingly, in addition to presenting relevant quantitative parameters and plots, our system reports were designed to place particular emphasis on interpretable outputs. The reports present clinically meaningful information, including normative references, identified gait deviations with severity ratings, and potential contributing factors inferred from the knowledge base. In addition, an implementation guideline was developed to specify user roles in data capture, video processing, report generation, and clinical interpretation. Within this guideline, it is recommended that health care institutions assign technically trained personnel to manage video processing and report generation, as these steps require familiarity with the software workflow and quality-control procedures. This role allocation is intended to support consistent report generation and reduce the operational burden on front-line clinicians, while clinicians retain responsibility for interpreting the full report, including quantitative parameters, visual plots, identified gait deviations, and explanatory information. Structured predeployment training sessions are also included in the implementation guideline to enhance data literacy and ensure the system’s effective and reliable use within routine clinical workflows.

The clinical simulation (step 4) revealed a potential factor that could influence the system’s practical effectiveness: variability in how clinicians interpret automatically identified gait deviations. Although explanatory notes were provided in the prototype to link each deviation with its possible biomechanical or neuromuscular causes and thereby enhance interpretability, clinical interpretation ultimately depends on the clinician’s professional knowledge and holistic understanding of the patient’s condition. Identifying the potential contributors to gait deviations can only offer supplementary insights for clinicians in terms of subsequent treatment planning and the evaluation of therapeutic outcomes [49]. Previous OGA studies have reported variability in clinicians’ interpretation of the same gait data/outputs [50,51]. Although automated or semiautomated gait assessment systems are designed to standardize outputs, differences in how clinicians interpret these outputs remain a potential source of variability [52]. Therefore, even when the system provides indicators or explanatory cues, its reliability in supporting clinical decision-making may still be constrained if clinicians interpret those cues inconsistently. These findings highlight the importance of supporting clinicians’ understanding of system outputs and underlying reasoning processes to promote consistent interpretation and effective clinical use.

This study has several limitations. First, the sample of participating clinicians was geographically limited to a single clinical group, which may restrict the generalizability of the findings. Second, the system’s reliance on a rule-based algorithm for deviation-cause mapping inherently limits its algorithmic precision and robustness, potentially resulting in a struggle to accurately interpret complex or atypical CP gait patterns. Third, the system’s explanatory and reporting functions were evaluated qualitatively rather than quantitatively. This study also did not formally compare different report formats or levels of explanatory detail, which may influence clinicians’ efficiency, trust, and interpretation consistency. Fourth, children with CP and their caregivers or family members were not included as co-design participants in this study. The main reason for this decision was that the prototype was clinician-centered and was designed primarily as a decision-support tool for rehabilitation professionals, rather than as a patient- or caregiver-facing system. Nevertheless, the perspectives of patients and caregivers should not be overlooked, as they may influence the acceptability, feasibility, and ethical implementation of VGA in routine care, particularly in relation to video data capture, privacy, and communication of assessment results.

Building on these findings, future work should focus on 3 main priority areas. First, multisite trials are essential to evaluate the system’s generalizability, interrater reliability, and impact on clinical decision-making and treatment outcomes. Second, efforts must be directed toward algorithmic and interpretive enhancement: expanding the knowledge base to integrate probabilistic reasoning or machine-learning models could enhance the precision of deviation-cause mapping. To mitigate the critical risk of variability in clinical interpretation, future evaluations should also compare different report formats and levels of explanatory detail to determine how best to balance efficiency, interpretability, clinical justification, and avoidance of overreliance on automated outputs. Finally, in terms of implementation and training frameworks, establishing shared interpretation frameworks is key to maximizing clinical benefit. This requires developing structured training modules focusing on standardized interpretation, embedding clinician feedback loops and uncertainty indicators within the report interface, and incorporating patient and caregiver perspectives to support broader implementation, particularly around acceptability, privacy, and communication of results.

Using the FRESCO co-design framework and iterative collaboration among clinicians, engineers, and biomechanics experts, we progressively refined a clinically oriented VGA system for children with CP into a stable and interpretable decision-support tool that demonstrated good usability and workflow compatibility. The study highlights how clinician-centered, structured co-design can facilitate the creation of digital tools that are both technically feasible and aligned with clinical reasoning. Beyond the immediate context of CP, the transferable design principles identified in this work, emphasizing interpretability, workflow integration, and low-barrier data capture, may inform the development of other clinically relevant digital health systems.