Measurement-grade video for computer vision in neurology: an international consensus framework for acquisition and reporting

Measurement-grade video for computer vision in neurology: an international consensus framework for acquisition and reporting | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Method Article Measurement-grade video for computer vision in neurology: an international consensus framework for acquisition and reporting Jane Alty, Gianluca Amprimo, Jung Hwan Shin, Max Wuehr, Michal Novotny, and 18 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9437116/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background : Neurological motor assessment relies on semi-quantitative, scale-based measures with inherent clinimetric limitations that disproportionately affect clinical trials aimed at early disease stages, where sensitivity is critical. Computer vision (CV) enables objective, scalable quantification of motor signs from standard video, yet clinical translation remains limited. A central bottleneck lies in the dependence of these methods on visual input: variability in video acquisition propagates into algorithmic outputs, constraining reproducibility and generalizability, while inconsistent reporting obscures a major source of variance. These limitations necessitate a clear definition of measurement-grade video as a prerequisite for developing reliable and scalable computer vision-based biomarkers in neurology. Methods : We conducted an international, multidisciplinary two-round modified Delphi process (n=23/20 experts across five continents) to define requirements for measurement-grade video in neurological motor assessment. In round one, items were elicited through predominantly open-ended queries on acquisition practices, failure modes, and outcome priorities. In round two, 16 acquisition and 21 metadata items were evaluated using a tiered scheme, with consensus pre-specified at ≥70% endorsement. In parallel, four domain-specific workgroups synthesized literature and real-world experience into domain-specific extensions. Results : Acquisition variability emerged as a principal barrier to reliable measurement across domains, with disagreement reflecting differences in enforceability rather than relevance. Three acquisition items reached mandatory consensus: continuous body-region coverage, protocolized video setup, and standardized task instructions. For metadata, task script, patient characteristics, video frame rate, and patient demographics were deemed mandatory. Feasibility varied by domain, highest for hand movements and lowest for eye movements. Most panelists identified underrepresentation of patient diversity and insufficient reference standards as key limitations in current studies. Notably, none of the mandatory elements requires additional hardware or substantial setup. Conclusion : This consensus-defined framework establishes measurement-grade video as a prerequisite for reliable and scalable computer vision-based movement analysis, providing an immediate, infrastructure-independent foundation for digital biomarker development in neurology. Neurology Computational Neuroscience Artificial Intelligence and Machine Learning Neurology Digital Biomarkers Computer Vision Movement Analysis Videography Figures Figure 1 Figure 2 Figure 3 Introduction The practice in Neurology has long relied on expert observation of clinical signs. The 19th-century Charcotian tenet of “ seeing without a preconceived notion whatsoever ” still shapes today’s ritualized examination, with standardized tasks such as finger tapping, gaze pursuit and timed gait 1 . Yet, these rich, dynamic behaviors are ultimately compressed into categorical or ordinal scales, that capture only a fraction of their complexity. The resulting measurement resolution limits objectivity, blunts sensitivity to change and may constrain both clinical trials and translational progress at the very moment when quantitative measurement tools are most needed 2,3 . Progress in neurological metrology has come less from discovering new signs than from improving how we elicit and record them, from the reflex hammer to neurophysiology. Video has been central throughout this evolution, serving as a foundation for documentation, education, and a source of trial endpoints 4,5 . Although what clinicians assess has been standardized through instruments scales such as the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) 6 , how clinical video is acquired remains variable with no universal acquisition standard 4 . Until recently, this gap had limited practical impact, as quantitative video analysis was not readily feasible. This has changed. Computer vision (CV) now enables objective, high-resolution quantification of neurological signs from ordinary video without specialized hardware. Recent systematic reviews encompassing more than 140 studies and thousands of patients demonstrate technical accuracy and clinical validity across neurological disorders and tasks 7–9 . However, while video-based measurement in Neurology has been shown to be feasible and promising, one important remaining problem is how it should be implemented reliably, and at scale. A major bottleneck has been identified: the method of creating and reporting measurement-grade video 7–9 . As a consequence, promising results can hardly be transferred across sites or datasets, largely because video setup directly affects algorithm performance and the resulting measurements 7–9 . For instance, tremor analyses fail when hands leave the frame 10,11 , eye tracking degrades with poor contrast 12 , and gait metrics can shift with camera geometry 13,14 (exemplified in Figure 1). The underlying knowledge of what constitutes adequate acquisition exists, but has remained distributed across research groups, domains, and unpublished practice, not being systematically codified into shared standards that would enable scalable development. Addressing this gap, we develop an international consensus-based framework for acquiring and reporting measurement-grade clinical video, optimized for computer vision analysis. In a two-round consensus process with 23 experts across five continents, and across domain-specific workgroups for eye movements, gait and posture, upper limb function, facial analysis, and head–neck movement, we translate real-world CV experience into a tiered framework for acquisition and reporting. Crucially, the framework requires no new infrastructure or substantial additional effort and provides an implementation-ready resource to transform clinical video into a measurement-grade substrate for digital biomarker discovery. Methods Ethics The study protocol was assessed and exempt from approval by the ethics committee at the University Hospital Ulm, Germany. Participants Panelists were selected using purposive sampling to ensure representation of domain-specific expertise across key areas of video-based neurological assessment including hand/upper limb, gait, face, posture/balance, eye movements, and head/neck. Experts were identified based on their track record of relevant publications and projects, methodological contributions to video, clinical scale, computer vision and/ or digital biomarker research, and/ or recognized clinical expertise in movement disorders, neuro-otology, or related fields. Selection aimed to achieve diversity across domains, disciplines, and geographical regions. Potential panelists were identified through prior collaborations, literature review, and field recognition, and were invited directly via email. The final panel comprised 23 participants (13 scientists, 8 clinician-scientists, and 2 clinicians), representing both research-focused (n= 13) and combined clinical–research settings (n= 9) across five global regions (see Supplementary Figure 1 for details). One invited expert was unable to participate due to institutional constraints. Delphi Procedures We conducted a modified two-round Delphi process integrating expert knowledge elicitation, thematic synthesis, and quantitative consensus formation 15 . Round 1 (n= 23) focused on exploration and thematic synthesis. Open-ended prompts and 5-point Likert scale statements (with 1 indicating lowest, and 5 highest agreement) 16 were administered via an online survey to elicit practical experience, common failure modes, and key decision points across movement tasks and domains. Responses were qualitatively synthesized into thematic statements and preliminary methodological propositions. Round 2 (n= 20) focused on quantification and consensus. Panelists classified each candidate acquisition and metadata item into one of four categories: mandatory (required in every study), optimal (strongly recommended where feasible), contextual (applicable in specific settings or tasks), or not required. Confidence in the resulting framework was rated on a Likert scale. Feasibility was assessed by asking panelists to estimate what proportion of patients in their primary setting could meet all mandatory criteria. Two consensus tiers were pre-specified: “mandatory” (≥70% selected mandatory) and “recommended” (≥70% selected mandatory + optimal). This study was designed and reported in alignment with key elements of established Delphi reporting frameworks (e.g., DELPHISTAR), including structured item elicitation, iterative feedback, and pre-specified consensus criteria 16 . The full question set alongside the responses has been made openly available on Harvard Dataverse (see Data Availability ). Domain-specific consensus recommendations Domain-specific workgroups comprising 5-7 experts were convened across eye movements, gait and posture, hand and upper limb, face/ head/ neck, and posture and balance. Each group completed a loosely structured template covering clinical rationale, task-specific constraints and artifacts, a synthesis of most relevant prior CV approaches, and key failure modes. Based on core priorities that emerged in the first survey, the groups created and consented domain-specific, tiered reference recommendations for acquisition and metadata reporting. Drafts underwent iterative revision before consolidation by domain leads and final synthesis by the lead authors (J. A. & M.U.F.). Domain sections integrated evidence from recent systematic reviews 8,9 alongside practice-based panel knowledge, but no de novo systematic review was conducted. Statistical analysis Rating data were summarized using descriptive statistics, endorsement proportions, and Likert distributions, including means (± SD) for interpretability. Given the ordinal nature of Likert-scale responses, key items are additionally summarized using medians (IQR) and the proportion of ratings ≥4 to reflect agreement 17 . Results Participant demographics 23 experts (8 female, 1 undisclosed) in Europe (n= 12), North America (n= 5), Asia (n= 3), Oceania (n= 2) and South America (n= 1) completed the survey. More than half (13/23) of the panel were scientists, 8 clinician-scientists and 2 clinicians; 13 worked in research-focused settings, 9 in both clinical and research settings. Experience in CV-based video analysis was broad (0-1 years, n= 3; 3-5 years n=8; 6-10 years n=6 and >10 years n=6). Domain expertise clustered in CV analysis of hand/upper limb (78%), gait (61%), face (43%) and posture/balance (43%) with fewer reporting expertise in eye movements (30%) and head/neck (17%, Supplementary Figure 1A) Application priorities for measurement-grade video Panelists prioritized multicenter research cohorts (83%), followed by clinical trials and endpoints (65%), patient self-recording (65%), telemedicine (61%), and routine outpatient clinics (52%, Supplementary Figure 1B). Ground-Truth References standards Sensor-based systems such as wearables and inertial measurement units (IMUs) were rated robust ground truth benchmarks by 91% of panelists, followed by clinical rater ensembles (78%). Correlation with a single expert rater, a very common validation approach in published studies, was considered sufficient by only 52%, with 26% rating it as hardly robust (Supplementary Figure 1C). Barriers to cross-site translation Systematic thematic coding of free-text responses (mean 2.6 codes per respondent, 59 codes in total) identified video acquisition heterogeneity as the leading barrier (65%), followed by task and protocol inconsistencies (39%) and absence of standardized software and analysis pipelines (35%). Representative statements were: " Differences in camera placement, framing, lighting, frame rate, and compression introduce systematic bias that cannot be corrected post hoc "; " Small differences in task instructions, duration, or execution style substantially alter extracted features and break cross-site comparability " (Figure 2A). Attitudes towards standardization Panelists strongly agreed that video acquisition should be treated as a protocolized measurement component (mean 4.52 ± 0.73; median (IQR) 5 (4–5); 87% ≥ 4) and that harmonization is currently necessary for multicenter comparability (4.48 ± 0.59; 5 (4–5); 96% ≥ 4). They were cautiously optimistic that future heterogeneity-tolerant CV models may reduce reliance on standardized protocols (3.96 ± 0.88; 4 (4–4); 78% ≥ 4), while calibration requirements showed the greatest variability of opinion (3.74 ± 1.14; 4 (3–5); 65% ≥ 4). Panelists judged acquisition-related variability to be greater than variability arising from CV methods, though this item showed the lowest agreement (3.61 ± 0.89; 4 (3–4); 61% ≥ 4; Figure 2B). Desired consensus outputs and tone of recommendations Almost all panelists prioritized a task-specific or domain-general acquisition protocol (96% and 87%) and metadata reporting guideline (96%) as the most important outputs, with correspondingly high confidence (4.09± 0.85, 4 (4-5); 87% ≥ 4, and 4.00± 0.80, 4 (4-4); 78% ≥ 4, respectively). Most preferred flexible guidance articulating trade-offs (78%), while fewer endorsed strong recommendations where evidence was solid (43%, Figure 2C). Acquisition factors and Metadata priorities In Round 1, strong consensus (≥70% endorsement) was reached for the following acquisition factors: task-specific video setup (96%), camera viewpoint and stability (96% and 87%), camera-to-subject distance (83%), patient positioning (87%), minimization of occlusions (83%), task instructions (83%), and frame rate (83%). Calibration information met moderate consensus (65%); visual clutter (48%) and patient clothing (43%) did not reach consensus (Supplementary Figure 2). For metadata reporting, panelists prioritized a core set: task script/ instructions (96%), video frame rate (96%), criteria for unusable video (96%), spatial resolution (91%), distance/ field of view (91%), CV-related quality metrics (91%), and camera viewpoint (87%). Supplementary items included camera stabilization (74%), calibration approach (65%), privacy level (65%), recording setting (65%), lighting (52%), and camera height (52%). Device models, camera orientation, and video file format were considered lower priority (Supplementary Figure 3). Round 2: Confirmatory consensus These findings informed the second, confirmatory survey, during which panelists categorized candidate items as mandatory, optimal, contextual, or not required. Two acquisition items reached mandatory consensus: explicit video setup defined in the study protocol (90%) and standardized task instructions (85%). Several additional items fell below mandatory threshold but reached recommended consensus (Mandatory + Optimal ≥70%), including minimally viable spatial and temporal resolution (100% and 95%), pre-defined criteria for unusable video (95%), consistent camera viewpoint and stabilization (90% and 85%), minimization and documentation of occlusions (85%), and standardized patient positioning (80%). Background conditions and patient clothing showed greater variability (Figure 3A). Four metadata items reached mandatory consensus: task script/ instructions (90%), disease severity and motor confounders (85%), video frame rate (75%), and patient demographics (75%). Additional items reaching recommended consensus included spatial resolution (100%), criteria for unusable video and quality control approach (95% each), and camera field of view and technical specifications (95% and 90%). Reporting of patient clothing and room setup was more variably endorsed (Figure 3B). Confidence in the resulting framework was high for both acquisition and metadata domains (acquisition: 4.30 ± 0.46, 4 (4–5); 100% ≥ 4; metadata: 4.25 ± 0.62, 4 (4–5), 90% ≥ 4). Overall, 68% of panelists estimated that at least 70% of patients in their primary clinical setting could meet all mandatory acquisition criteria. Role of domain-specific extensions Despite strong support for the domain-general framework, 95% of panelists emphasized that domain-specific extensions are needed or essential. Domain-specific workgroups translated the framework into differentiated protocols across five domains. Please see Supplementary Materials for a detailed report as well as the extended domain-specific acquisition frameworks. Eye movement assessments Eye movement tasks involve subtle, high-frequency dynamics and therefore impose the most stringent technical requirements across domains. High temporal resolution (60–120 Hz) and framing that maximizes face and eye visibility were considered mandatory to avoid aliasing and preserve signal fidelity, particularly for fast saccades and subtle nystagmus. Panelists strongly recommended optimized lighting to enhance pupil–iris contrast and avoidance of occlusions, for example from examiner hands or eyeglasses. 67% of domain experts estimated that ≥70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figure 4). Facial and head movement assessments Assessing facial expression involves measuring both subtle, localized movements and gestalt features across different timescales while head tracking additionally involves complex configurational changes in 3-D space. Accordingly, panelists considered stable lighting, frontal head coverage with a resulting face width of >256 total pixels at a spatial resolution of at least 720p (i.e. 1280 x 720) mandatory. In addition, minimizing facial occlusions to <20% of recording time was endorsed as a mandatory threshold. 85% of domain experts estimated that ≥70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figures 5-6). Hand and upper limb assessments Hand motor tasks usually involve complex configurational changes in 3-D space, challenging monocular video-based analyses. Accordingly, continuous visibility and optimal framing of the hand at closer range (e.g. 30-80cm, depending on optical zoom) using a stabilized camera was considered mandatory. Panelists strongly recommended pilot testing, particularly for underexplored clinical tasks such as reaching or pronation-supination, to anticipate interactions between task, viewpoint and downstream CV algorithm performance. 95% of domain experts estimated that ≥70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figure 7). Gait and posture assessments Gait and posture assessments typically require whole-body tracking across larger spatial scales and often include additional persons in the frame (e.g., for safety during gait or when performing the pull test). Accordingly, continuous full-body visibility throughout the recording and careful control of potential confounding effects from additional persons were considered mandatory. 78% of domain experts estimated that ≥70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figures 8-9). Across all domains, continuous and adequate coverage of the body region of interest emerged as a universal prerequisite for valid measurement and was therefore incorporated as a third mandatory principle into the domain-general acquisition framework (Figure 3). Discussion As the limitations of traditional clinical rating scales become increasingly apparent, particularly in early disease stages, the demand for sensitive, objective, and scalable outcome measures in Neurology continues to grow 2,3,18–21 . Wearable sensors, including inertial measurement units and electromyography, have advanced objective neuromotor assessments 18 , but impose device burden and provide limited access to the richness of human movement (disorders) 22,23 . Video-based computer vision addresses these constraints by enabling contactless, whole-body, multi-joint capture that is retrospectively reanalyzable, requires no instrumentation at point of care, and can scale naturally to remote and decentralized settings 24 . However, a fundamental but underappreciated bottleneck persists: measurement validity depends critically on upstream acquisition conditions 7–9 . Variability in video setup propagates into algorithmic outputs, undermining reproducibility and cross-study comparability. Compounding this challenge, acquisition conditions are inconsistently reported across studies, rendering a major source of measurement variability difficult to identify or control. This variability represents a low-effort, high-yield target for immediate optimization. To address this gap, we establish an international consensus framework for measurement-grade clinical video in Neurology, bridging rapidly advancing computational methods with fragmented real-world acquisition practices. By synthesizing previously uncodified expert knowledge, the framework delineates (i) general acquisition principles applicable across movement domains, (ii) domain-specific constraints that shape measurement validity, and (iii) practical, implementable recommendations for acquisition and reporting. Video standardization of this kind is not without precedent. The integration of video-based eye movement quantification into clinical care and trials was enabled by standardization through specialized recording setups, an approach that has remained robust across successive generations of analytical algorithms 25 . Similarly, in several movement disorders, standardized video protocols have been proposed to reliably elicit and document key clinical signs 4,26 . The present framework extends this logic across neurological movement domains, providing a structured resource that facilitates reproducible clinical measurements with current computational methods 7 . Within the domain-general acquisition protocol, consensus identified two mandatory elements: explicit specification of video setup in the study protocol and standardized task instructions. Continuous, task-appropriate coverage of the body region of interest emerged independently across all domain-specific groups and was incorporated as a third mandatory requirement. For metadata reporting, four items were deemed essential: task script, disease severity and motor confounders, video frame rate, and patient demographics. Importantly, none of these elements require specialized equipment or substantial additional effort, yet acquisition failures are irreversible, and missing metadata critically limits interpretability, reproducibility and reuse. Beyond this mandatory core, consensus extended across nearly all remaining items when mandatory and optimal endorsements were considered together. Where disagreement occurred, it centered largely on the degree of enforcement rather than the relevance of individual factors. The framework captures near-universal agreement on what constitutes measurement-grade video, with its tiered structure reflecting pragmatic implementation constraints rather than scientific uncertainty. This prioritization is directly supported by recent empirical work in Parkinson’s disease: Irani et al. systematically varied acquisition parameters across 315 finger tapping videos and found that frame rate and hand-to-frame ratio were the primary determinants of pose estimation failure and scoring error, consistent with the panel's placement of body-region coverage and frame rate in the mandatory tier 27 . The resulting flexibility will likely enhance ecological validity and adoption. Accordingly, most panelists estimated that at least 70% of patients in routine clinical settings can meet mandatory criteria. Importantly, standardization need not impose a trade-off; rather, it can enhance conventional clinical assessment, while preserving flexibility for future computational analyses of the same recordings: for instance, a standardized frontal recording of finger tapping enables both comparable multi-rater MDS-UPDRS annotation and CV-based kinematic feature extraction from the same capture. In addition, 95% of panelists endorsed the need for task-specific extensions, resulting in domain-specific protocols provided alongside this general framework. Across domains, measurement validity is determined by the interaction of algorithm, task design, and video acquisition, which must be optimized in concert. Limited sets of clinically informative tasks can be sufficient to capture multiple movement phenotypes of interest, reducing protocol complexity without sacrificing diagnostic yield 28 . The present framework addresses acquisition as an immediate and tractable intervention point, delineating general and domain-specific principles that stabilize downstream measurements to enable reproducible, clinically meaningful outputs across computational approaches. Together, these elements establish measurement-grade video as a prerequisite for reliable and scalable video-based digital biomarkers. Three additional findings merit consideration. First, correlation with a single expert rater, one of the most commonly used reference standards 7–9 , was considered sufficient by only 52% of panelists, with 26% rating it minimally robust. This highlights a growing disconnect between prevailing validation norms and expert expectations for measurement rigor. Second, the majority (85%) of panelists reported that existing studies underrepresent patient diversity, with direct implications for generalizability and equity. To address this, the proposed reporting guideline explicitly includes demographic metadata to enable auditing and contextual interpretation. Third, estimated feasibility varied markedly by domain: hand movements showed the highest, and eye movement the lowest, reflecting stringent technical requirements, while gait uniquely introduces multi-person constraints. These findings reinforce the need for domain-specific extensions rather than a one-size-fits-all standard. Although heterogeneity-tolerant computer vision models remain a long-term objective, they require datasets of a scale and diversity not yet available for most neurological conditions. Standardizing acquisition therefore accelerates systematic evaluation of CV methods, and ultimately their clinical translation. In this sense, the framework extends a core principle of Neurology into the digital era: that rigorous observation underpins reliable inference. As Neurology transitions toward AI-enabled measurement, decentralized trials, and learning health systems, such disciplined acquisition can form the foundation for trustworthy digital biomarker development of human movement. Finally, the principles of measurement-grade video defined here are likely to extend to other domains of video-based movement analysis, including musculoskeletal, or psychiatric conditions 29,30 . Several limitations should be acknowledged. First, the framework is based on expert consensus rather than prospective validation and thus reflects current best practices rather than empirically optimized protocols. Nonetheless, emerging empirical evidence is largely consistent with the panel’s prioritization 27 . By establishing a shared baseline, this framework provides the necessary foundation for further systematic evaluation of task–algorithm–acquisition interactions, enabling iterative, evidence-based refinement of the protocols themselves. Second, representation across domains was not uniform, and certain areas (e.g., head and neck movements) require further systematic investigation. Third, feasibility estimates are context-dependent and may vary across clinical cohorts and environments. Future work should therefore include empiric evaluation of acquisition protocols, and benchmarking across environments, video and wearable modalities to determine where each method or multimodal combination yields the greatest measurement fidelity. Declarations Acknowledgements MUF is supported by the Manfred and Ursula Müller-Stiftung (NeuroTech Innovationspreis 2024), the Jung Stiftung für Wissenschaft und Forschung (Karrierefoerderpreis 2024), and the joint Scientific Attending Physician program of the Medical Faculty and University Hospital Ulm. Conflicts of interest JA reports consulting fees from Lilly, Abbvie, Stada, paid editorial position for Nature Portfolio journals, royalties from CRC Press, all unrelated to the present work. MUF reports consulting fees from CereGate GmbH and Tovly LLC and paid editorial position at Nature Portfolio Journals, unrelated to the present work. JR reports consulting fees from Roche, unrelated to the present work. HH is funded by the Dutch Research Council (NWO) under the grant TTW-Veni 21148, the funder did not play a role in the current work. SW reports honoraria from Mepha and Neurolite, unrelated to the present work. The remaining authors have no disclosures. Data and code availability The anonymized survey data can be found on Harvard Dataverse: https://doi.org/10.7910/DVN/GSBZ5C. References Petrova, N. L., Edmonds, M. E. & Papanas, N. Jean-Martin Charcot: 200 years after his birth, still a paragon in the diabetic foot. Int. J. Low. Extrem. Wounds 24 , 751–753 (2025). Hobart, J. C., Cano, S. J., Zajicek, J. P. & Thompson, A. J. Rating scales as outcome measures for clinical trials in neurology: problems, solutions, and recommendations. Lancet Neurol. 6 , 1094–1105 (2007). Krainc, D. et al. Shifting the trajectory of therapeutic development for neurological and psychiatric disorders. Sci. Transl. Med. 15 , eadg4775 (2023). Rosso, M., Chu, D. & Barton, B. The Influence of Videography in the Foundation and Future of Movement Disorders. Mov. Disord. Clin. Pract. 13 , 359–368 (2026). Jeanjean, A. & Aubert, G. Moving pictures of Parkinson’s disease. The Lancet 378 , 1773–1774 (2011). Martinez-Martin, P. et al. Expanded and independent validation of the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS). J. Neurol. 260 , 228–236 (2013). Friedrich, M. U., Relton, S., Wong, D. & Alty, J. Computer Vision in Clinical Neurology: A Review. JAMA Neurol. https://doi.org/10.1001/jamaneurol.2024.5326 (2025) doi:10.1001/jamaneurol.2024.5326. Martínez-García-Peña, R., Koens, L. H., Azzopardi, G. & Tijssen, M. A. J. Video-Based Data-Driven Models for Diagnosing Movement Disorders: Review and Future Directions. Mov. Disord. Off. J. Mov. Disord. Soc. https://doi.org/10.1002/mds.30327 (2025) doi:10.1002/mds.30327. Pecoraro, P. M., Marsili, L., Espay, A. J., Bologna, M. & di Biase, L. Computer Vision Technologies in Movement Disorders: A Systematic Review. Mov. Disord. Clin. Pract. https://doi.org/10.1002/mdc3.70123 (2025) doi:10.1002/mdc3.70123. Friedrich, M. U. et al. Validation and application of computer vision algorithms for video-based tremor analysis. Npj Digit. Med. 7 , 1–13 (2024). Filippou, V. et al. Comparison of monocular video-based methods for measuring the amplitude of hand tremor. Comput. Biol. Med. 206 , 111615 (2026). Friedrich, M. U. et al. Smartphone video nystagmography using convolutional neural networks: ConVNG. J. Neurol. 270 , 2518–2530 (2023). Stenum, J., Hsu, M. M., Pantelyat, A. Y. & Roemmich, R. T. Clinical gait analysis using video-based pose estimation: Multiple perspectives, clinical populations, and measuring change. PLOS Digit. Health 3 , e0000467 (2024). Boborzi, L., Bertram, J., Schniepp, R., Decker, J. & Wuehr, M. Clinical Whole-Body Gait Characterization Using a Single RGB-D Sensor. Sensors 25 , (2025). de Meyrick, J. The Delphi method and health research. Health Educ. 103 , 7–16 (2003). Niederberger, M. et al. Delphi studies in social and health sciences—Recommendations for an interdisciplinary standardized reporting (DELPHISTAR). Results of a Delphi study. PLOS ONE 19 , e0304651 (2024). Sullivan, G. M. & Artino, A. R. Analyzing and Interpreting Data From Likert-Type Scales. J. Grad. Med. Educ. 5 , 541–542 (2013). van den Brandhof, E. L. et al. Explainable machine learning for movement disorders - Classification of tremor and myoclonus. Comput. Biol. Med. 192 , 110180 (2025). Hampel, H. et al. The foundation and architecture of precision medicine in neurology and psychiatry. Trends Neurosci. 46 , 176–198 (2023). Erickson, C. M., Wexler, A. & Largent, E. A. Digital Biomarkers for Neurodegenerative Disease. JAMA Neurol. 82 , 5–6 (2025). Gupta, A. S. Digital Phenotyping in Clinical Neurology. Semin. Neurol. 42 , 48–59 (2022). Ma, Y., Zhang, Y., Li, R., Cheng, W. & Wu, F. The experience and perception of wearable devices in Parkinson’s disease patients: a systematic review and meta-synthesis of qualitative studies. J. Neurol. 272 , 350 (2025). Cai, H., Hu, J., Zhao, C. & Lin, J. Wearable devices in neurological disorders: a narrative review of status quo and perspectives. Ann. Transl. Med. 13 , 46–46 (2025). Alty, J. et al. The TAS Test project: a prospective longitudinal validation of new online motor-cognitive tests to detect preclinical Alzheimer’s disease and estimate 5-year risks of cognitive decline and dementia. BMC Neurol. 22 , 266 (2022). Leigh, R. J. & Zee, D. S. The Neurology of Eye Movements . The Neurology of Eye Movements (Oxford University Press, 2015). Ganos, C., Stanley, M. P. H. & Lang, A. E. On the practice of video documentation and representation in movement disorders. Brain awag005 (2026) doi:10.1093/brain/awag005. Irani, A et al. Impact of Video Quality on Video-based Digital Assessment of Parkinson’s Disease Digital Biomarkers. Digit. Biomark. in press ,. Sciacca, G. et al. Next move in movement disorders (NEMO): the best clinical tasks for the visibility of essential tremor, dystonia, cortical myoclonus and myoclonus-dystonia. Parkinsonism Relat. Disord. 138 , 107963 (2025). Ruth, P. S. et al. Video-Based Biomechanical Analysis Captures Disease-Specific Movement Signatures of Different Neuromuscular Diseases. NEJM AI 2 , AIoa2401137 (2025). Alagapan, S. et al. Cingulate dynamics track depression recovery with deep brain stimulation. Nature 622 , 130–138 (2023). Additional Declarations The authors declare no competing interests. Supplementary Files CVConsensusFrameworkSI.docx SFigure01.png Panel characteristics, application priorities, and ground truth robustness. (A) Geographic distribution of panelists, years of computer vision (CV) experience, and domain expertise (multiple selections permitted; percentages do not sum to 100%). (B) Priority ratings for five clinical application contexts; percentage at right = proportion rating the context as highest or high priority (top-2-boxes). Perceived robustness of three ground truth reference methods for neurological movement assessment; percentage at right = proportion rating the method as very or somewhat robust (top-2-box). All items are rated on a 5-point Likert scale. IMU, inertial measurement units. SFigure02AcquisitionFactors.png Importance ratings for video acquisition factors from Round 1 survey (n = 23). Stacked horizontal bars showing 5-point Likert rating distributions for 13 domain-general video acquisition factors. Items are grouped by consensus tier and sorted by endorsement rate (top-2-box: very important + somewhat important) descending within each tier. Consensus was defined as ≥70% endorsement; items below this threshold are shown in the no consensus tier. Percentage at right= top-2-box endorsement rate. Items were carried forward to the round 2-tiered confirmation survey (Block A). SFigure03MetadataReporting.png Importance ratings of metadata reporting items from Round 1 survey (n = 23). Stacked horizontal bars showing 5-point Likert rating distributions for 17 candidate metadata reporting items. Items are grouped by consensus tier and sorted by endorsement rate (top-2-box: very important + somewhat important) descending within each tier. Consensus was defined as ≥70% endorsement; items below this threshold are shown in the no consensus tier. Percentage at right= top-2-box endorsement rate. Items were carried forward to the round 2-tiered confirmation survey (Block B). CV, computer vision Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9437116","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Method Article","associatedPublications":[],"authors":[{"id":624251651,"identity":"9bbeafad-0d4b-48eb-a09f-169464fd8cc7","order_by":0,"name":"Jane Alty","email":"","orcid":"","institution":"University of Tasmania, Hobart, Australia","correspondingAuthor":false,"prefix":"","firstName":"Jane","middleName":"","lastName":"Alty","suffix":""},{"id":624251653,"identity":"86a05673-153a-4325-902a-609526691d3d","order_by":1,"name":"Gianluca Amprimo","email":"","orcid":"","institution":"Politecnico di Torino, Turin, Italy","correspondingAuthor":false,"prefix":"","firstName":"Gianluca","middleName":"","lastName":"Amprimo","suffix":""},{"id":624251654,"identity":"75bcef30-0fe8-423e-8674-b35e65b94296","order_by":2,"name":"Jung Hwan Shin","email":"","orcid":"","institution":"Seoul National University Hospital, Seoul, South Korea","correspondingAuthor":false,"prefix":"","firstName":"Jung","middleName":"Hwan","lastName":"Shin","suffix":""},{"id":624251655,"identity":"d0014fb4-f88b-4d5a-ad5f-1ed81739b7cd","order_by":3,"name":"Max Wuehr","email":"","orcid":"","institution":"German Center for Vertigo and Balance Disorders (DSGZ), LMU University Hospital, Munich, Germany","correspondingAuthor":false,"prefix":"","firstName":"Max","middleName":"","lastName":"Wuehr","suffix":""},{"id":624255039,"identity":"405e5cc3-b273-48af-8e60-eaec9342e788","order_by":4,"name":"Michal Novotny","email":"","orcid":"","institution":"Czech Technical University in Prague, Prague, Czech Republic","correspondingAuthor":false,"prefix":"","firstName":"Michal","middleName":"","lastName":"Novotny","suffix":""},{"id":624255040,"identity":"688fe863-69c8-4dad-ae9c-315bcae09481","order_by":5,"name":"Helga Haberfehlner","email":"","orcid":"","institution":"Department of Rehabilitation Medicine, Amsterdam Movement Sciences Amsterdam University Medical Center, Amsterdam, Netherlands","correspondingAuthor":false,"prefix":"","firstName":"Helga","middleName":"","lastName":"Haberfehlner","suffix":""},{"id":624255041,"identity":"28c2212a-8ebd-4dba-a8be-4eb9c0325b30","order_by":6,"name":"Claudia Ferraris","email":"","orcid":"","institution":"Institute of Electronics, Computer and Telecommunication Engineering (IEIIT) - National Research Council (CNR), Turin, Italy","correspondingAuthor":false,"prefix":"","firstName":"Claudia","middleName":"","lastName":"Ferraris","suffix":""},{"id":624255042,"identity":"299ee29c-4533-4c1b-b683-87ce60f94f34","order_by":7,"name":"Renjie Li","email":"","orcid":"","institution":"University of Tasmania, Hobart, Australia","correspondingAuthor":false,"prefix":"","firstName":"Renjie","middleName":"","lastName":"Li","suffix":""},{"id":624256588,"identity":"0c034543-4d93-4678-a43b-dd09dd04838c","order_by":8,"name":"Marcelo Merello","email":"","orcid":"","institution":"University of Buenos Aires, CONICET, Buenos Aires, Argentina","correspondingAuthor":false,"prefix":"","firstName":"Marcelo","middleName":"","lastName":"Merello","suffix":""},{"id":624256589,"identity":"d84772f1-5f9b-4c38-90bf-9191cbbb6b19","order_by":9,"name":"Marina de Koning-Tijssen","email":"","orcid":"","institution":"Expertise Center Movement Disorders, Department of Neurology, University of Groningen, Groningen, Netherlands","correspondingAuthor":false,"prefix":"","firstName":"Marina","middleName":"","lastName":"de Koning-Tijssen","suffix":""},{"id":624256590,"identity":"75d0e9fb-543d-4c7c-ad10-4b8be4c2ec4f","order_by":10,"name":"Samuel Relton","email":"","orcid":"","institution":"Leeds Institute of Health Sciences, University of Leeds, Leeds, United Kingdom","correspondingAuthor":false,"prefix":"","firstName":"Samuel","middleName":"","lastName":"Relton","suffix":""},{"id":624256591,"identity":"fd363638-9629-4927-9147-c009000a5ab3","order_by":11,"name":"David Wong","email":"","orcid":"","institution":"Leeds Institute of Health Sciences, University of Leeds, Leeds, United Kingdom","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Wong","suffix":""},{"id":624256592,"identity":"9374684a-b231-4bcc-a244-c3faad17d2d9","order_by":12,"name":"Johannes Taeger","email":"","orcid":"","institution":"HNO Praxis Taeger, Muenchen, Germany","correspondingAuthor":false,"prefix":"","firstName":"Johannes","middleName":"","lastName":"Taeger","suffix":""},{"id":624256593,"identity":"d095b5a5-33fe-40d0-b076-7bd60c350054","order_by":13,"name":"Andreas Zwergal","email":"","orcid":"","institution":"German Center for Vertigo and Balance Disorders (DSGZ), LMU University Hospital, Munich, Germany","correspondingAuthor":false,"prefix":"","firstName":"Andreas","middleName":"","lastName":"Zwergal","suffix":""},{"id":624256594,"identity":"53885bd1-1e61-4d9e-b757-aee4314f5a8b","order_by":14,"name":"Sebastian Walther","email":"","orcid":"","institution":"Department of Psychiatry, Psychosomatics, and Psychotherapy, University Hospital Würzburg, Würzburg, Germany","correspondingAuthor":false,"prefix":"","firstName":"Sebastian","middleName":"","lastName":"Walther","suffix":""},{"id":624257400,"identity":"f8094335-f3c8-4e08-9fd6-4b4d5a52227f","order_by":15,"name":"Jochen Weishaupt","email":"","orcid":"","institution":"Department of Neurology, University Hospital Ulm, Ulm, Germany","correspondingAuthor":false,"prefix":"","firstName":"Jochen","middleName":"","lastName":"Weishaupt","suffix":""},{"id":624257401,"identity":"c259389e-ef96-41b3-9bb1-893d1e14a40d","order_by":16,"name":"Karl Georg Häusler","email":"","orcid":"","institution":"Department of Neurology, University Hospital Ulm, Ulm, Germany","correspondingAuthor":false,"prefix":"","firstName":"Karl","middleName":"Georg","lastName":"Häusler","suffix":""},{"id":624257402,"identity":"f4241d18-de79-4ef2-9cd0-6506db78359e","order_by":17,"name":"Jan Rusz","email":"","orcid":"","institution":"Czech Technical University in Prague, Prague, Czech Republic","correspondingAuthor":false,"prefix":"","firstName":"Jan","middleName":"","lastName":"Rusz","suffix":""},{"id":624257403,"identity":"920150c5-8960-46d6-8214-dbd0686d77aa","order_by":18,"name":"Babak Taati","email":"","orcid":"","institution":"KITE Research Institute, University Health Network, Toronto Western Hospital, Toronto, ON, Canada","correspondingAuthor":false,"prefix":"","firstName":"Babak","middleName":"","lastName":"Taati","suffix":""},{"id":624257744,"identity":"4b0773da-88cc-4326-a65b-2b00bb3403dd","order_by":19,"name":"Ryan Roemmich","email":"","orcid":"","institution":"Center for Movement Studies, Kennedy Krieger Institute, Baltimore, MD, USA","correspondingAuthor":false,"prefix":"","firstName":"Ryan","middleName":"","lastName":"Roemmich","suffix":""},{"id":624257745,"identity":"b8a5e448-5fba-4c0f-adad-ba24528de530","order_by":20,"name":"Martin McKeown","email":"","orcid":"","institution":"Pacific Parkinson's Research Centre and Division of Neurology, Djavad Mowafaghian Centre for Brain Health, University of British Columbia and Vancouver Coastal Health, Vancouver, BC, Canada","correspondingAuthor":false,"prefix":"","firstName":"Martin","middleName":"","lastName":"McKeown","suffix":""},{"id":624257746,"identity":"ebf8ab8c-128c-4ecc-8f29-ac5d1bc61c44","order_by":21,"name":"Anoopum S. Gupta","email":"","orcid":"","institution":"Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA","correspondingAuthor":false,"prefix":"","firstName":"Anoopum","middleName":"S.","lastName":"Gupta","suffix":""},{"id":624257747,"identity":"dd981e92-145c-409e-8ac8-15439f12df9b","order_by":22,"name":"Maximilian U. Friedrich","email":"data:image/png;base64,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","orcid":"","institution":"Department of Neurology, University Hospital Ulm, Ulm, Germany","correspondingAuthor":true,"prefix":"","firstName":"Maximilian","middleName":"U.","lastName":"Friedrich","suffix":""}],"badges":[],"createdAt":"2026-04-16 10:44:15","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9437116/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9437116/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107488731,"identity":"089905f4-105b-4c2c-a985-1f7f0785009b","added_by":"auto","created_at":"2026-04-22 02:45:41","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":16711331,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExemplary\u003c/strong\u003e \u003cstrong\u003eIllustration of how video\u003c/strong\u003e \u003cstrong\u003eacquisition quality relates to the validity of video-based movement measurement\u003c/strong\u003e. Standardized, measurement-grade video acquisition enables stable computer vision tracking and forms the basis for reproducible movement features (top). In contrast, uncontrolled recording conditions degrade tracking already at the representation stage, introducing uncertainty in keypoint localization (e.g., jitter, misplacement, or dropout) that propagates through downstream computations and results in compounded errors in derived movement features (bottom).\u003c/p\u003e","description":"","filename":"Figure01.png","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/10e8776f1338f02357374bc3.png"},{"id":107486993,"identity":"60d12826-300d-493d-8df8-045cf7944e54","added_by":"auto","created_at":"2026-04-22 02:39:30","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2208357,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAcquisition variability as a primary barrier to scalable video-based measurement.\u003c/strong\u003e (A) Acquisition heterogeneity and protocol inconsistency are identified as leading causes of failed cross-site translation in clinical video studies. (B) Experts show strong agreement that video acquisition constitutes a protocolized measurement component and that harmonization is necessary for multicenter comparability. (C) Accordingly, structured outputs, particularly standardized metadata reporting and task-specific protocols, are prioritized as key requirements for the field.\u003c/p\u003e","description":"","filename":"Figure02.png","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/a0f7f6d7d0efc812636af841.png"},{"id":107487670,"identity":"2df7ff9c-a815-4b92-b4b6-b5f3319ffa9a","added_by":"auto","created_at":"2026-04-22 02:42:35","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2982374,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConsensus framework for measurement-grade video in Neurology: \u003c/strong\u003eacquisition protocol (Block A) and metadata reporting guideline (Block B). Items are grouped by consensus tier: Mandatory (M ≥70%), Recommended (Mandatory + Optimal (O) ≥70%, M \u0026lt;70%), and Contextual (M+O \u0026lt;70%), and sorted by combined Mandatory + Optimal endorsement rate (M+O%) descending within each tier. Stacked bars show the full rating distribution across four response categories (Mandatory, Optimal, Contextual, Not required); the percentage at right represents M+O% throughout. Bold items reached Mandatory consensus. A1* denotes a universal acquisition principle, continuous adequate coverage of the body region of interest, derived independently from all four domain-specific task groups and endorsed at 100% cross-group agreement; it is listed first as a foundational requirement applicable across all movement domains. n = 20 for all items. FOV, field of view; QC, quality control.\u003c/p\u003e","description":"","filename":"Figure03.png","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/432cb1208440412f96f01f5d.png"},{"id":107490014,"identity":"3c85c1a6-2599-4088-a70b-7bdd1a53491b","added_by":"auto","created_at":"2026-04-22 02:49:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":18507016,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/1d735772-8654-4165-ba66-5377d0c2f1fd.pdf"},{"id":107363691,"identity":"64db6225-05cf-49d6-85d7-9aa048c86f5d","added_by":"auto","created_at":"2026-04-20 19:06:42","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":880227,"visible":true,"origin":"","legend":"","description":"","filename":"CVConsensusFrameworkSI.docx","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/8997bec34162b3dabdc14f71.docx"},{"id":107486808,"identity":"a10dd95a-aade-4358-8030-f4cbf0ddefa4","added_by":"auto","created_at":"2026-04-22 02:38:59","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":442073,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePanel characteristics, application priorities, and ground truth robustness\u003c/strong\u003e. (A) Geographic distribution of panelists, years of computer vision (CV) experience, and domain expertise (multiple selections permitted; percentages do not sum to 100%). (B) Priority ratings for five clinical application contexts; percentage at right = proportion rating the context as highest or high priority (top-2-boxes). Perceived robustness of three ground truth reference methods for neurological movement assessment; percentage at right = proportion rating the method as very or somewhat robust (top-2-box). All items are rated on a 5-point Likert scale. IMU, inertial measurement units.\u003c/p\u003e","description":"","filename":"SFigure01.png","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/8ae12a9276a9ea7225f98529.png"},{"id":107363693,"identity":"a8e6d0fe-100e-4edd-8c11-f4a7dae6fbcb","added_by":"auto","created_at":"2026-04-20 19:06:42","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":500820,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eImportance ratings for video acquisition factors\u003c/strong\u003e \u003cstrong\u003efrom Round 1 survey (n = 23\u003c/strong\u003e). Stacked horizontal bars showing 5-point Likert rating distributions for 13 domain-general video acquisition factors. Items are grouped by consensus tier and sorted by endorsement rate (top-2-box: very important + somewhat important) descending within each tier. Consensus was defined as ≥70% endorsement; items below this threshold are shown in the no consensus tier. Percentage at right= top-2-box endorsement rate. Items were carried forward to the round 2-tiered confirmation survey (Block A).\u003c/p\u003e","description":"","filename":"SFigure02AcquisitionFactors.png","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/8a4fc553138176659eef1793.png"},{"id":107363696,"identity":"48bce8d1-e94c-48c5-9828-67824a389bd4","added_by":"auto","created_at":"2026-04-20 19:06:42","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":468981,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eImportance ratings of metadata reporting items from Round 1 survey (n = 23). \u003c/strong\u003eStacked horizontal bars showing 5-point Likert rating distributions for 17 candidate metadata reporting items. Items are grouped by consensus tier and sorted by endorsement rate (top-2-box: very important + somewhat important) descending within each tier. Consensus was defined as ≥70% endorsement; items below this threshold are shown in the no consensus tier. Percentage at right= top-2-box endorsement rate. Items were carried forward to the round 2-tiered confirmation survey (Block B). CV, computer vision\u003c/p\u003e","description":"","filename":"SFigure03MetadataReporting.png","url":"https://assets-eu.researchsquare.com/files/rs-9437116/v1/ad401a0894b6ace69a5c0216.png"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eMeasurement-grade video for computer vision in neurology: an international consensus framework for acquisition and reporting\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe practice in Neurology has long relied on expert observation of clinical signs. The 19th-century Charcotian tenet of “\u003cem\u003eseeing without a preconceived notion whatsoever\u003c/em\u003e” still shapes today’s ritualized examination, with standardized tasks such as finger tapping, gaze pursuit and timed gait\u003csup\u003e1\u003c/sup\u003e. Yet, these rich, dynamic behaviors are ultimately compressed into categorical or ordinal scales, that capture only a fraction of their complexity. The resulting measurement resolution limits objectivity, blunts sensitivity to change and may constrain both clinical trials and translational progress at the very moment when quantitative measurement tools are most needed\u003csup\u003e2,3\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eProgress in neurological metrology has come less from discovering new signs than from improving how we elicit and record them, from the reflex hammer to neurophysiology. Video has been central throughout this evolution, serving as a foundation for documentation, education, and a source of trial endpoints\u003csup\u003e4,5\u003c/sup\u003e. Although \u003cem\u003ewhat\u003c/em\u003e clinicians assess has been standardized through instruments scales such as the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS)\u003csup\u003e6\u003c/sup\u003e, \u003cem\u003ehow\u003c/em\u003e clinical video is acquired remains variable with no universal acquisition standard\u003csup\u003e4\u003c/sup\u003e. Until recently, this gap had limited practical impact, as quantitative video analysis was not readily feasible.\u003c/p\u003e\n\u003cp\u003eThis has changed. Computer vision (CV) now enables objective, high-resolution quantification of neurological signs from ordinary video without specialized hardware. Recent systematic reviews encompassing more than 140 studies and thousands of patients demonstrate technical accuracy and clinical validity across neurological disorders and tasks\u003csup\u003e7–9\u003c/sup\u003e. However, while video-based measurement in Neurology has been shown to be feasible and promising, one important remaining problem is how it should be implemented reliably, and at scale.\u003c/p\u003e\n\u003cp\u003eA major bottleneck has been identified: the method of creating and reporting measurement-grade video\u003csup\u003e7–9\u003c/sup\u003e. As a consequence, promising results can hardly be transferred across sites or datasets, largely because video setup directly affects algorithm performance and the resulting measurements\u003csup\u003e7–9\u003c/sup\u003e. For instance, tremor analyses fail when hands leave the frame\u003csup\u003e10,11\u003c/sup\u003e, eye tracking degrades with poor contrast\u003csup\u003e12\u003c/sup\u003e, and gait metrics can shift with camera geometry\u003csup\u003e13,14\u003c/sup\u003e (exemplified in Figure 1). The underlying knowledge of what constitutes adequate acquisition exists, but has remained distributed across research groups, domains, and unpublished practice, not being systematically codified into shared standards that would enable scalable development.\u003c/p\u003e\n\u003cp\u003eAddressing this gap, we develop an international consensus-based framework for acquiring and reporting measurement-grade clinical video, optimized for computer vision analysis. In a two-round consensus process with 23 experts across five continents, and across domain-specific workgroups for eye movements, gait and posture, upper limb function, facial analysis, and head–neck movement, we translate real-world CV experience into a tiered framework for acquisition and reporting. Crucially, the framework requires no new infrastructure or substantial additional effort and provides an implementation-ready resource to transform clinical video into a measurement-grade substrate for digital biomarker discovery.\u003c/p\u003e"},{"header":"Methods","content":"\u003ch2\u003eEthics\u003c/h2\u003e\n\u003cp\u003eThe study protocol was assessed and exempt from approval by the ethics committee at the University Hospital Ulm, Germany.\u003c/p\u003e\n\u003ch2\u003eParticipants\u003c/h2\u003e\n\u003cp\u003ePanelists were selected using purposive sampling to ensure representation of domain-specific expertise across key areas of video-based neurological assessment including hand/upper limb, gait, face, posture/balance, eye movements, and head/neck. Experts were identified based on their track record of relevant publications and projects, methodological contributions to video, clinical scale, computer vision and/ or digital biomarker research, and/ or recognized clinical expertise in movement disorders, neuro-otology, or related fields. Selection aimed to achieve diversity across domains, disciplines, and geographical regions. Potential panelists were identified through prior collaborations, literature review, and field recognition, and were invited directly via email. The final panel comprised 23 participants (13 scientists, 8 clinician-scientists, and 2 clinicians), representing both research-focused (n= 13) and combined clinical–research settings (n= 9) across five global regions (see Supplementary Figure 1 for details). One invited expert was unable to participate due to institutional constraints.\u003c/p\u003e\n\u003ch2\u003eDelphi Procedures\u003c/h2\u003e\n\u003cp\u003eWe conducted a modified two-round Delphi process integrating expert knowledge elicitation, thematic synthesis, and quantitative consensus formation\u003csup\u003e15\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eRound 1\u003c/u\u003e (n= 23) focused on exploration and thematic synthesis. Open-ended prompts and 5-point Likert scale statements (with 1 indicating lowest, and 5 highest agreement)\u003csup\u003e16\u003c/sup\u003e were administered via an online survey to elicit practical experience, common failure modes, and key decision points across movement tasks and domains. Responses were qualitatively synthesized into thematic statements and preliminary methodological propositions.\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eRound 2\u003c/u\u003e (n= 20) focused on quantification and consensus. Panelists classified each candidate acquisition and metadata item into one of four categories: mandatory (required in every study), optimal (strongly recommended where feasible), contextual (applicable in specific settings or tasks), or not required. Confidence in the resulting framework was rated on a Likert scale. Feasibility was assessed by asking panelists to estimate what proportion of patients in their primary setting could meet all mandatory criteria. Two consensus tiers were pre-specified: “mandatory” (≥70% selected mandatory) and “recommended” (≥70% selected mandatory + optimal). This study was designed and reported in alignment with key elements of established Delphi reporting frameworks (e.g., DELPHISTAR), including structured item elicitation, iterative feedback, and pre-specified consensus criteria\u003csup\u003e16\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe full question set alongside the responses has been made openly available on Harvard Dataverse (see \u003cem\u003eData Availability\u003c/em\u003e).\u003c/p\u003e\n\u003ch2\u003eDomain-specific consensus recommendations\u003c/h2\u003e\n\u003cp\u003eDomain-specific workgroups comprising 5-7 experts were convened across eye movements, gait and posture, hand and upper limb, face/ head/ neck, and posture and balance. Each group completed a loosely structured template covering clinical rationale, task-specific constraints and artifacts, a synthesis of most relevant prior CV approaches, and key failure modes. Based on core priorities that emerged in the first survey, the groups created and consented domain-specific, tiered reference recommendations for acquisition and metadata reporting. Drafts underwent iterative revision before consolidation by domain leads and final synthesis by the lead authors (J. A. \u0026amp; M.U.F.). Domain sections integrated evidence from recent systematic reviews\u003csup\u003e8,9\u003c/sup\u003e alongside practice-based panel knowledge, but no de novo systematic review was conducted.\u003c/p\u003e\n\u003ch2\u003eStatistical analysis\u003c/h2\u003e\n\u003cp\u003eRating data were summarized using descriptive statistics, endorsement proportions, and Likert distributions, including means (± SD) for interpretability. Given the ordinal nature of Likert-scale responses, key items are additionally summarized using medians (IQR) and the proportion of ratings ≥4 to reflect agreement\u003csup\u003e17\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003eParticipant demographics\u003c/h2\u003e\n\u003cp\u003e23 experts (8 female, 1 undisclosed) in Europe (n= 12), North America (n= 5), Asia (n= 3), Oceania (n= 2) and South America (n= 1) completed the survey. More than half (13/23) of the panel were scientists, 8 clinician-scientists and 2 clinicians; 13 worked in research-focused settings, 9 in both clinical and research settings. Experience in CV-based video analysis was broad (0-1 years, n= 3; 3-5 years n=8; 6-10 years n=6 and \u0026gt;10 years n=6). Domain expertise clustered in CV analysis of hand/upper limb (78%), gait (61%), face (43%) and posture/balance (43%) with fewer reporting expertise in eye movements (30%) and head/neck (17%, Supplementary Figure 1A)\u003c/p\u003e\n\u003ch2\u003eApplication priorities for measurement-grade video\u003c/h2\u003e\n\u003cp\u003ePanelists prioritized multicenter research cohorts (83%), followed by clinical trials and endpoints (65%), patient self-recording (65%), telemedicine (61%), and routine outpatient clinics (52%, Supplementary Figure 1B).\u003c/p\u003e\n\u003ch2\u003eGround-Truth References standards\u003c/h2\u003e\n\u003cp\u003eSensor-based systems such as wearables and inertial measurement units (IMUs) were rated robust ground truth benchmarks by 91% of panelists, followed by clinical rater ensembles (78%). Correlation with a single expert rater, a very common validation approach in published studies, was considered sufficient by only 52%, with 26% rating it as hardly robust (Supplementary Figure 1C).\u003c/p\u003e\n\u003ch2\u003eBarriers to cross-site translation\u003c/h2\u003e\n\u003cp\u003eSystematic thematic coding of free-text responses (mean 2.6 codes per respondent, 59 codes in total) identified video acquisition heterogeneity as the leading barrier (65%), followed by task and protocol inconsistencies (39%) and absence of standardized software and analysis pipelines (35%). Representative statements were: \u0026quot;\u003cem\u003eDifferences in camera placement, framing, lighting, frame rate, and compression introduce systematic bias that cannot be corrected post hoc\u003c/em\u003e\u0026quot;; \u0026quot;\u003cem\u003eSmall differences in task instructions, duration, or execution style substantially alter extracted features and break cross-site comparability\u003c/em\u003e\u0026quot; (Figure 2A).\u003c/p\u003e\n\u003ch2\u003eAttitudes towards standardization\u003c/h2\u003e\n\u003cp\u003ePanelists strongly agreed that video acquisition should be treated as a protocolized measurement component (mean 4.52 \u0026plusmn; 0.73; median (IQR) 5 (4\u0026ndash;5); 87% \u0026ge; 4) and that harmonization is currently necessary for multicenter comparability (4.48 \u0026plusmn; 0.59; 5 (4\u0026ndash;5); 96% \u0026ge; 4). They were cautiously optimistic that future heterogeneity-tolerant CV models may reduce reliance on standardized protocols (3.96 \u0026plusmn; 0.88; 4 (4\u0026ndash;4); 78% \u0026ge; 4), while calibration requirements showed the greatest variability of opinion (3.74 \u0026plusmn; 1.14; 4 (3\u0026ndash;5); 65% \u0026ge; 4). Panelists judged acquisition-related variability to be greater than variability arising from CV methods, though this item showed the lowest agreement (3.61 \u0026plusmn; 0.89; 4 (3\u0026ndash;4); 61% \u0026ge; 4; Figure 2B).\u003c/p\u003e\n\u003ch2\u003eDesired consensus outputs and tone of recommendations\u003c/h2\u003e\n\u003cp\u003eAlmost all panelists prioritized a task-specific or domain-general acquisition protocol (96% and 87%) and metadata reporting guideline (96%) as the most important outputs, with correspondingly high confidence (4.09\u0026plusmn; 0.85, 4 (4-5); 87% \u0026ge; 4, and 4.00\u0026plusmn; 0.80, 4 (4-4); 78% \u0026ge; 4, respectively). Most preferred flexible guidance articulating trade-offs (78%), while fewer endorsed strong recommendations where evidence was solid (43%, Figure 2C).\u003c/p\u003e\n\u003ch2\u003eAcquisition factors and Metadata priorities\u003c/h2\u003e\n\u003cp\u003eIn Round 1, strong consensus (\u0026ge;70% endorsement) was reached for the following acquisition factors: task-specific video setup (96%), camera viewpoint and stability (96% and 87%), camera-to-subject distance (83%), patient positioning (87%), minimization of occlusions (83%), task instructions (83%), and frame rate (83%). Calibration information met moderate consensus (65%); visual clutter (48%) and patient clothing (43%) did not reach consensus (Supplementary Figure 2).\u003c/p\u003e\n\u003cp\u003eFor metadata reporting, panelists prioritized a core set: task script/ instructions (96%), video frame rate (96%), criteria for unusable video (96%), spatial resolution (91%), distance/ field of view (91%), CV-related quality metrics (91%), and camera viewpoint (87%). Supplementary items included camera stabilization (74%), calibration approach (65%), privacy level (65%), recording setting (65%), lighting (52%), and camera height (52%). Device models, camera orientation, and video file format were considered lower priority (Supplementary Figure 3).\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eRound 2: Confirmatory consensus\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThese findings informed the second, confirmatory survey, during which panelists categorized candidate items as mandatory, optimal, contextual, or not required.\u003c/p\u003e\n\u003cp\u003eTwo acquisition items reached mandatory consensus: explicit video setup defined in the study protocol (90%) and standardized task instructions (85%). Several additional items fell below mandatory threshold but reached recommended consensus (Mandatory + Optimal \u0026ge;70%), including minimally viable spatial and temporal resolution (100% and 95%), pre-defined criteria for unusable video (95%), consistent camera viewpoint and stabilization (90% and 85%), minimization and documentation of occlusions (85%), and standardized patient positioning (80%). Background conditions and patient clothing showed greater variability (Figure 3A).\u003c/p\u003e\n\u003cp\u003eFour metadata items reached mandatory consensus: task script/ instructions (90%), disease severity and motor confounders (85%), video frame rate (75%), and patient demographics (75%). Additional items reaching recommended consensus included spatial resolution (100%), criteria for unusable video and quality control approach (95% each), and camera field of view and technical specifications (95% and 90%). Reporting of patient clothing and room setup was more variably endorsed (Figure 3B).\u003c/p\u003e\n\u003cp\u003eConfidence in the resulting framework was high for both acquisition and metadata domains (acquisition: 4.30 \u0026plusmn; 0.46, 4 (4\u0026ndash;5); 100% \u0026ge; 4; metadata: 4.25 \u0026plusmn; 0.62, 4 (4\u0026ndash;5), 90% \u0026ge; 4). Overall, 68% of panelists estimated that at least 70% of patients in their primary clinical setting could meet all mandatory acquisition criteria.\u003c/p\u003e\n\u003ch2\u003eRole of domain-specific extensions\u003c/h2\u003e\n\u003cp\u003eDespite strong support for the domain-general framework, 95% of panelists emphasized that domain-specific extensions are needed or essential. Domain-specific workgroups translated the framework into differentiated protocols across five domains. Please see Supplementary Materials for a detailed report as well as the extended domain-specific acquisition frameworks.\u003c/p\u003e\n\u003ch3\u003eEye movement assessments\u003c/h3\u003e\n\u003cp\u003eEye movement tasks involve subtle, high-frequency dynamics and therefore impose the most stringent technical requirements across domains. High temporal resolution (60\u0026ndash;120 Hz) and framing that maximizes face and eye visibility were considered mandatory to avoid aliasing and preserve signal fidelity, particularly for fast saccades and subtle nystagmus. Panelists strongly recommended optimized lighting to enhance pupil\u0026ndash;iris contrast and avoidance of occlusions, for example from examiner hands or eyeglasses. 67% of domain experts estimated that \u0026ge;70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figure 4).\u003c/p\u003e\n\u003ch3\u003eFacial and head movement assessments\u003c/h3\u003e\n\u003cp\u003eAssessing facial expression involves measuring both subtle, localized movements and gestalt features across different timescales while head tracking additionally involves complex configurational changes in 3-D space. Accordingly, panelists considered stable lighting, frontal head coverage with a resulting face width of \u0026gt;256 total pixels at a spatial resolution of at least 720p (i.e. 1280 x 720) mandatory. In addition, minimizing facial occlusions to \u0026lt;20% of recording time was endorsed as a mandatory threshold. 85% of domain experts estimated that \u0026ge;70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figures 5-6).\u003c/p\u003e\n\u003ch3\u003eHand and upper limb assessments\u003c/h3\u003e\n\u003cp\u003eHand motor tasks usually involve complex configurational changes in 3-D space, challenging monocular video-based analyses. Accordingly, continuous visibility and optimal framing of the hand at closer range (e.g. 30-80cm, depending on optical zoom) using a stabilized camera was considered mandatory. Panelists strongly recommended pilot testing, particularly for underexplored clinical tasks such as reaching or pronation-supination, to anticipate interactions between task, viewpoint and downstream CV algorithm performance. 95% of domain experts estimated that \u0026ge;70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figure 7).\u003c/p\u003e\n\u003ch3\u003eGait and posture assessments\u003c/h3\u003e\n\u003cp\u003eGait and posture assessments typically require whole-body tracking across larger spatial scales and often include additional persons in the frame (e.g., for safety during gait or when performing the pull test). Accordingly, continuous full-body visibility throughout the recording and careful control of potential confounding effects from additional persons were considered mandatory. 78% of domain experts estimated that \u0026ge;70% of patients would be able to meet mandatory acquisition criteria in routine clinical settings (Supplementary Figures 8-9).\u003c/p\u003e\n\u003cp\u003eAcross all domains, continuous and adequate coverage of the body region of interest emerged as a universal prerequisite for valid measurement and was therefore incorporated as a third mandatory principle into the domain-general acquisition framework (Figure 3).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eAs the limitations of traditional clinical rating scales become increasingly apparent, particularly in early disease stages, the demand for sensitive, objective, and scalable outcome measures in Neurology continues to grow\u003csup\u003e2,3,18–21\u003c/sup\u003e. Wearable sensors, including inertial measurement units and electromyography, have advanced objective neuromotor assessments\u003csup\u003e18\u003c/sup\u003e, but impose device burden and provide limited access to the richness of human movement (disorders)\u003csup\u003e22,23\u003c/sup\u003e. Video-based computer vision addresses these constraints by enabling contactless, whole-body, multi-joint capture that is retrospectively reanalyzable, requires no instrumentation at point of care, and can scale naturally to remote and decentralized settings\u003csup\u003e24\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eHowever, a fundamental but underappreciated bottleneck persists: measurement validity depends critically on upstream acquisition conditions\u003csup\u003e7–9\u003c/sup\u003e. Variability in video setup propagates into algorithmic outputs, undermining reproducibility and cross-study comparability. Compounding this challenge, acquisition conditions are inconsistently reported across studies, rendering a major source of measurement variability difficult to identify or control. This variability represents a low-effort, high-yield target for immediate optimization.\u003c/p\u003e\n\u003cp\u003eTo address this gap, we establish an international consensus framework for measurement-grade clinical video in Neurology, bridging rapidly advancing computational methods with fragmented real-world acquisition practices. By synthesizing previously uncodified expert knowledge, the framework delineates (i) general acquisition principles applicable across movement domains, (ii) domain-specific constraints that shape measurement validity, and (iii) practical, implementable recommendations for acquisition and reporting.\u003c/p\u003e\n\u003cp\u003eVideo standardization of this kind is not without precedent. The integration of video-based eye movement quantification into clinical care and trials was enabled by standardization through specialized recording setups, an approach that has remained robust across successive generations of analytical algorithms\u003csup\u003e25\u003c/sup\u003e. Similarly, in several movement disorders, standardized video protocols have been proposed to reliably elicit and document key clinical signs\u003csup\u003e4,26\u003c/sup\u003e. The present framework extends this logic across neurological movement domains, providing a structured resource that facilitates reproducible clinical measurements with current computational methods\u003csup\u003e7\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eWithin the domain-general acquisition protocol, consensus identified two mandatory elements: explicit specification of video setup in the study protocol and standardized task instructions. Continuous, task-appropriate coverage of the body region of interest emerged independently across all domain-specific groups and was incorporated as a third mandatory requirement. For metadata reporting, four items were deemed essential: task script, disease severity and motor confounders, video frame rate, and patient demographics. Importantly, none of these elements require specialized equipment or substantial additional effort, yet acquisition failures are irreversible, and missing metadata critically limits interpretability, reproducibility and reuse.\u003c/p\u003e\n\u003cp\u003eBeyond this mandatory core, consensus extended across nearly all remaining items when mandatory and optimal endorsements were considered together. Where disagreement occurred, it centered largely on the degree of enforcement rather than the relevance of individual factors. The framework captures near-universal agreement on what constitutes measurement-grade video, with its tiered structure reflecting pragmatic implementation constraints rather than scientific uncertainty. This prioritization is directly supported by recent empirical work in Parkinson’s disease: Irani et al. systematically varied acquisition parameters across 315 finger tapping videos and found that frame rate and hand-to-frame ratio were the primary determinants of pose estimation failure and scoring error, consistent with the panel's placement of body-region coverage and frame rate in the mandatory tier\u003csup\u003e27\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe resulting flexibility will likely enhance ecological validity and adoption. Accordingly, most panelists estimated that at least 70% of patients in routine clinical settings can meet mandatory criteria. Importantly, standardization need not impose a trade-off; rather, it can enhance conventional clinical assessment, while preserving flexibility for future computational analyses of the same recordings: for instance, a standardized frontal recording of finger tapping enables both comparable multi-rater MDS-UPDRS annotation and CV-based kinematic feature extraction from the same capture.\u003c/p\u003e\n\u003cp\u003eIn addition, 95% of panelists endorsed the need for task-specific extensions, resulting in domain-specific protocols provided alongside this general framework. Across domains, measurement validity is determined by the interaction of algorithm, task design, and video acquisition, which must be optimized in concert. Limited sets of clinically informative tasks can be sufficient to capture multiple movement phenotypes of interest, reducing protocol complexity without sacrificing diagnostic yield\u003csup\u003e28\u003c/sup\u003e. The present framework addresses acquisition as an immediate and tractable intervention point, delineating general and domain-specific principles that stabilize downstream measurements to enable reproducible, clinically meaningful outputs across computational approaches.\u003c/p\u003e\n\u003cp\u003eTogether, these elements establish measurement-grade video as a prerequisite for reliable and scalable video-based digital biomarkers.\u003c/p\u003e\n\u003cp\u003eThree additional findings merit consideration. First, correlation with a single expert rater, one of the most commonly used reference standards\u003csup\u003e7–9\u003c/sup\u003e, was considered sufficient by only 52% of panelists, with 26% rating it minimally robust. This highlights a growing disconnect between prevailing validation norms and expert expectations for measurement rigor. Second, the majority (85%) of panelists reported that existing studies underrepresent patient diversity, with direct implications for generalizability and equity. To address this, the proposed reporting guideline explicitly includes demographic metadata to enable auditing and contextual interpretation. Third, estimated feasibility varied markedly by domain: hand movements showed the highest, and eye movement the lowest, reflecting stringent technical requirements, while gait uniquely introduces multi-person constraints. These findings reinforce the need for domain-specific extensions rather than a \u003cem\u003eone-size-fits-all\u003c/em\u003e standard.\u003c/p\u003e\n\u003cp\u003eAlthough heterogeneity-tolerant computer vision models remain a long-term objective, they require datasets of a scale and diversity not yet available for most neurological conditions. Standardizing acquisition therefore accelerates systematic evaluation of CV methods, and ultimately their clinical translation. In this sense, the framework extends a core principle of Neurology into the digital era: that rigorous observation underpins reliable inference. As Neurology transitions toward AI-enabled measurement, decentralized trials, and learning health systems, such disciplined acquisition can form the foundation for trustworthy digital biomarker development of human movement. Finally, the principles of measurement-grade video defined here are likely to extend to other domains of video-based movement analysis, including musculoskeletal, or psychiatric conditions\u003csup\u003e29,30\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eSeveral limitations should be acknowledged. First, the framework is based on expert consensus rather than prospective validation and thus reflects current best practices rather than empirically optimized protocols. Nonetheless, emerging empirical evidence is largely consistent with the panel’s prioritization\u003csup\u003e27\u003c/sup\u003e. By establishing a shared baseline, this framework provides the necessary foundation for further systematic evaluation of task–algorithm–acquisition interactions, enabling iterative, evidence-based refinement of the protocols themselves. Second, representation across domains was not uniform, and certain areas (e.g., head and neck movements) require further systematic investigation. Third, feasibility estimates are context-dependent and may vary across clinical cohorts and environments. Future work should therefore include empiric evaluation of acquisition protocols, and benchmarking across environments, video and wearable modalities to determine where each method or multimodal combination yields the greatest measurement fidelity.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgements\u003c/h2\u003e\n\u003cp\u003eMUF is supported by the Manfred and Ursula Müller-Stiftung (NeuroTech Innovationspreis 2024), the Jung Stiftung für Wissenschaft und Forschung (Karrierefoerderpreis 2024), and the joint Scientific Attending Physician program of the Medical Faculty and University Hospital Ulm.\u003c/p\u003e\n\u003ch2\u003eConflicts of interest\u003c/h2\u003e\n\u003cp\u003eJA reports consulting fees from Lilly, Abbvie, Stada, paid editorial position for Nature Portfolio journals, royalties from CRC Press, all unrelated to the present work. MUF reports consulting fees from CereGate GmbH and Tovly LLC and paid editorial position at Nature Portfolio Journals, unrelated to the present work. JR reports consulting fees from Roche, unrelated to the present work. HH is funded by the Dutch Research Council (NWO) under the grant TTW-Veni 21148, the funder did not play a role in the current work. SW reports honoraria from Mepha and Neurolite, unrelated to the present work. The remaining authors have no disclosures.\u003c/p\u003e\n\u003ch2\u003eData and code availability\u003c/h2\u003e\n\u003cp\u003eThe anonymized survey data can be found on Harvard Dataverse: https://doi.org/10.7910/DVN/GSBZ5C.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePetrova, N. L., Edmonds, M. E. \u0026amp; Papanas, N. Jean-Martin Charcot: 200 years after his birth, still a paragon in the diabetic foot. \u003cem\u003eInt. J. Low. Extrem. Wounds\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 751\u0026ndash;753 (2025).\u003c/li\u003e\n\u003cli\u003eHobart, J. C., Cano, S. J., Zajicek, J. P. \u0026amp; Thompson, A. J. Rating scales as outcome measures for clinical trials in neurology: problems, solutions, and recommendations. \u003cem\u003eLancet Neurol.\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 1094\u0026ndash;1105 (2007).\u003c/li\u003e\n\u003cli\u003eKrainc, D. \u003cem\u003eet al.\u003c/em\u003e Shifting the trajectory of therapeutic development for neurological and psychiatric disorders. \u003cem\u003eSci. Transl. Med.\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, eadg4775 (2023).\u003c/li\u003e\n\u003cli\u003eRosso, M., Chu, D. \u0026amp; Barton, B. The Influence of Videography in the Foundation and Future of Movement Disorders. \u003cem\u003eMov. Disord. Clin. Pract.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 359\u0026ndash;368 (2026).\u003c/li\u003e\n\u003cli\u003eJeanjean, A. \u0026amp; Aubert, G. Moving pictures of Parkinson\u0026rsquo;s disease. \u003cem\u003eThe Lancet\u003c/em\u003e \u003cstrong\u003e378\u003c/strong\u003e, 1773\u0026ndash;1774 (2011).\u003c/li\u003e\n\u003cli\u003eMartinez-Martin, P. \u003cem\u003eet al.\u003c/em\u003e Expanded and independent validation of the Movement Disorder Society-Unified Parkinson\u0026rsquo;s Disease Rating Scale (MDS-UPDRS). \u003cem\u003eJ. Neurol.\u003c/em\u003e \u003cstrong\u003e260\u003c/strong\u003e, 228\u0026ndash;236 (2013).\u003c/li\u003e\n\u003cli\u003eFriedrich, M. U., Relton, S., Wong, D. \u0026amp; Alty, J. Computer Vision in Clinical Neurology: A Review. \u003cem\u003eJAMA Neurol.\u003c/em\u003e https://doi.org/10.1001/jamaneurol.2024.5326 (2025) doi:10.1001/jamaneurol.2024.5326.\u003c/li\u003e\n\u003cli\u003eMart\u0026iacute;nez-Garc\u0026iacute;a-Pe\u0026ntilde;a, R., Koens, L. H., Azzopardi, G. \u0026amp; Tijssen, M. A. J. Video-Based Data-Driven Models for Diagnosing Movement Disorders: Review and Future Directions. \u003cem\u003eMov. Disord. Off. J. Mov. Disord. Soc.\u003c/em\u003e https://doi.org/10.1002/mds.30327 (2025) doi:10.1002/mds.30327.\u003c/li\u003e\n\u003cli\u003ePecoraro, P. M., Marsili, L., Espay, A. J., Bologna, M. \u0026amp; di Biase, L. Computer Vision Technologies in Movement Disorders: A Systematic Review. \u003cem\u003eMov. Disord. Clin. Pract.\u003c/em\u003e https://doi.org/10.1002/mdc3.70123 (2025) doi:10.1002/mdc3.70123.\u003c/li\u003e\n\u003cli\u003eFriedrich, M. U. \u003cem\u003eet al.\u003c/em\u003e Validation and application of computer vision algorithms for video-based tremor analysis. \u003cem\u003eNpj Digit. Med.\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 1\u0026ndash;13 (2024).\u003c/li\u003e\n\u003cli\u003eFilippou, V. \u003cem\u003eet al.\u003c/em\u003e Comparison of monocular video-based methods for measuring the amplitude of hand tremor. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cstrong\u003e206\u003c/strong\u003e, 111615 (2026).\u003c/li\u003e\n\u003cli\u003eFriedrich, M. U. \u003cem\u003eet al.\u003c/em\u003e Smartphone video nystagmography using convolutional neural networks: ConVNG. \u003cem\u003eJ. Neurol.\u003c/em\u003e \u003cstrong\u003e270\u003c/strong\u003e, 2518\u0026ndash;2530 (2023).\u003c/li\u003e\n\u003cli\u003eStenum, J., Hsu, M. M., Pantelyat, A. Y. \u0026amp; Roemmich, R. T. Clinical gait analysis using video-based pose estimation: Multiple perspectives, clinical populations, and measuring change. \u003cem\u003ePLOS Digit. Health\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, e0000467 (2024).\u003c/li\u003e\n\u003cli\u003eBoborzi, L., Bertram, J., Schniepp, R., Decker, J. \u0026amp; Wuehr, M. Clinical Whole-Body Gait Characterization Using a Single RGB-D Sensor. \u003cem\u003eSensors\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, (2025).\u003c/li\u003e\n\u003cli\u003ede Meyrick, J. The Delphi method and health research. \u003cem\u003eHealth Educ.\u003c/em\u003e \u003cstrong\u003e103\u003c/strong\u003e, 7\u0026ndash;16 (2003).\u003c/li\u003e\n\u003cli\u003eNiederberger, M. \u003cem\u003eet al.\u003c/em\u003e Delphi studies in social and health sciences\u0026mdash;Recommendations for an interdisciplinary standardized reporting (DELPHISTAR). Results of a Delphi study. \u003cem\u003ePLOS ONE\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, e0304651 (2024).\u003c/li\u003e\n\u003cli\u003eSullivan, G. M. \u0026amp; Artino, A. R. Analyzing and Interpreting Data From Likert-Type Scales. \u003cem\u003eJ. Grad. Med. Educ.\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 541\u0026ndash;542 (2013).\u003c/li\u003e\n\u003cli\u003evan den Brandhof, E. L. \u003cem\u003eet al.\u003c/em\u003e Explainable machine learning for movement disorders - Classification of tremor and myoclonus. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cstrong\u003e192\u003c/strong\u003e, 110180 (2025).\u003c/li\u003e\n\u003cli\u003eHampel, H. \u003cem\u003eet al.\u003c/em\u003e The foundation and architecture of precision medicine in neurology and psychiatry. \u003cem\u003eTrends Neurosci.\u003c/em\u003e \u003cstrong\u003e46\u003c/strong\u003e, 176\u0026ndash;198 (2023).\u003c/li\u003e\n\u003cli\u003eErickson, C. M., Wexler, A. \u0026amp; Largent, E. A. Digital Biomarkers for Neurodegenerative Disease. \u003cem\u003eJAMA Neurol.\u003c/em\u003e \u003cstrong\u003e82\u003c/strong\u003e, 5\u0026ndash;6 (2025).\u003c/li\u003e\n\u003cli\u003eGupta, A. S. Digital Phenotyping in Clinical Neurology. \u003cem\u003eSemin. Neurol.\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 48\u0026ndash;59 (2022).\u003c/li\u003e\n\u003cli\u003eMa, Y., Zhang, Y., Li, R., Cheng, W. \u0026amp; Wu, F. The experience and perception of wearable devices in Parkinson\u0026rsquo;s disease patients: a systematic review and meta-synthesis of qualitative studies. \u003cem\u003eJ. Neurol.\u003c/em\u003e \u003cstrong\u003e272\u003c/strong\u003e, 350 (2025).\u003c/li\u003e\n\u003cli\u003eCai, H., Hu, J., Zhao, C. \u0026amp; Lin, J. Wearable devices in neurological disorders: a narrative review of status quo and perspectives. \u003cem\u003eAnn. Transl. Med.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 46\u0026ndash;46 (2025).\u003c/li\u003e\n\u003cli\u003eAlty, J. \u003cem\u003eet al.\u003c/em\u003e The TAS Test project: a prospective longitudinal validation of new online motor-cognitive tests to detect preclinical Alzheimer\u0026rsquo;s disease and estimate 5-year risks of cognitive decline and dementia. \u003cem\u003eBMC Neurol.\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 266 (2022).\u003c/li\u003e\n\u003cli\u003eLeigh, R. J. \u0026amp; Zee, D. S. \u003cem\u003eThe Neurology of Eye Movements\u003c/em\u003e. \u003cem\u003eThe Neurology of Eye Movements\u003c/em\u003e (Oxford University Press, 2015).\u003c/li\u003e\n\u003cli\u003eGanos, C., Stanley, M. P. H. \u0026amp; Lang, A. E. On the practice of video documentation and representation in movement disorders. \u003cem\u003eBrain\u003c/em\u003e awag005 (2026) doi:10.1093/brain/awag005.\u003c/li\u003e\n\u003cli\u003eIrani, A \u003cem\u003eet al.\u003c/em\u003e Impact of Video Quality on Video-based Digital Assessment of Parkinson\u0026rsquo;s Disease Digital Biomarkers. \u003cem\u003eDigit. Biomark.\u003c/em\u003e \u003cstrong\u003ein press\u003c/strong\u003e,.\u003c/li\u003e\n\u003cli\u003eSciacca, G. \u003cem\u003eet al.\u003c/em\u003e Next move in movement disorders (NEMO): the best clinical tasks for the visibility of essential tremor, dystonia, cortical myoclonus and myoclonus-dystonia. \u003cem\u003eParkinsonism Relat. Disord.\u003c/em\u003e \u003cstrong\u003e138\u003c/strong\u003e, 107963 (2025).\u003c/li\u003e\n\u003cli\u003eRuth, P. S. \u003cem\u003eet al.\u003c/em\u003e Video-Based Biomechanical Analysis Captures Disease-Specific Movement Signatures of Different Neuromuscular Diseases. \u003cem\u003eNEJM AI\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, AIoa2401137 (2025).\u003c/li\u003e\n\u003cli\u003eAlagapan, S. \u003cem\u003eet al.\u003c/em\u003e Cingulate dynamics track depression recovery with deep brain stimulation. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e622\u003c/strong\u003e, 130\u0026ndash;138 (2023).\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cbr /\u003e \u0026nbsp;\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University Hospital Ulm","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Neurology, Digital Biomarkers, Computer Vision, Movement Analysis, Videography ","lastPublishedDoi":"10.21203/rs.3.rs-9437116/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9437116/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cu\u003eBackground\u003c/u\u003e: Neurological motor assessment relies on semi-quantitative, scale-based measures with inherent clinimetric limitations that disproportionately affect clinical trials aimed at early disease stages, where sensitivity is critical. Computer vision (CV) enables objective, scalable quantification of motor signs from standard video, yet clinical translation remains limited. A central bottleneck lies in the dependence of these methods on visual input: variability in video acquisition propagates into algorithmic outputs, constraining reproducibility and generalizability, while inconsistent reporting obscures a major source of variance. These limitations necessitate a clear definition of measurement-grade video as a prerequisite for developing reliable and scalable computer vision-based biomarkers in neurology.\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eMethods\u003c/u\u003e: We conducted an international, multidisciplinary two-round modified Delphi process (n=23/20 experts across five continents) to define requirements for measurement-grade video in neurological motor assessment. In round one, items were elicited through predominantly open-ended queries on acquisition practices, failure modes, and outcome priorities. In round two, 16 acquisition and 21 metadata items were evaluated using a tiered scheme, with consensus pre-specified at ≥70% endorsement. In parallel, four domain-specific workgroups synthesized literature and real-world experience into domain-specific extensions.\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eResults\u003c/u\u003e: Acquisition variability emerged as a principal barrier to reliable measurement across domains, with disagreement reflecting differences in enforceability rather than relevance. Three acquisition items reached mandatory consensus: continuous body-region coverage, protocolized video setup, and standardized task instructions. For metadata, task script, patient characteristics, video frame rate, and patient demographics were deemed mandatory. Feasibility varied by domain, highest for hand movements and lowest for eye movements. Most panelists identified underrepresentation of patient diversity and insufficient reference standards as key limitations in current studies. Notably, none of the mandatory elements requires additional hardware or substantial setup.\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eConclusion\u003c/u\u003e: This consensus-defined framework establishes measurement-grade video as a prerequisite for reliable and scalable computer vision-based movement analysis, providing an immediate, infrastructure-independent foundation for digital biomarker development in neurology.\u003c/p\u003e","manuscriptTitle":"Measurement-grade video for computer vision in neurology: an international consensus framework for acquisition and reporting","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-20 19:06:38","doi":"10.21203/rs.3.rs-9437116/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5598fa68-8ca6-46ec-8483-17bf6d0dfa58","owner":[],"postedDate":"April 20th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":66442933,"name":"Neurology"},{"id":66442934,"name":"Computational Neuroscience"},{"id":66442935,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-04-20T19:06:38+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-20 19:06:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9437116","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9437116","identity":"rs-9437116","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}