henhao
Sign in
BioengineeringNEJM AI
Peer-reviewed journal article

Video-based biomechanical analysis captures disease-specific movement signatures of different neuromuscular diseases

Parker S. RuthiD, Scott D. UhlrichiD, Constance de MontsiD, Antoine Motte dit FalisseiD, Julie MucciniiD, Sydney C. CovitziD, Shelby Vogt-DomkeiD, John DayiD, Tina DuongiD, Scott L. DelpiD

Abstract

BackgroundAssessing human movement is essential for diagnosing and monitoring movement-related conditions like neuromuscular disorders. Timed function tests (TFTs) are among the most widespread assessments due to their speed and simplicity, but they cannot capture disease-specific movement patterns. Conversely, biomechanical analysis can produce sensitive disease-specific biomarkers but is traditionally confined to laboratory settings. Recent advances in smartphone video-based biomechanical analysis enable quantification of 3D movement with the ease and speed required for clinical settings. However, the potential of this technology to offer more sensitive assessments of human function than TFTs remains untested. MethodsTo compare video-based analysis against TFTs, we collected an observational dataset from 129 individuals: 28 with facioscapulohumeral muscular dystrophy, 58 with myotonic dystrophy, and 43 controls with no diagnosed neuromuscular condition. We used OpenCap, a free open-source software tool, to capture smartphone video-based biomechanics of nine different movements in a median time of 16 minutes per participant. From these recordings we extracted 34 interpretable movement features. Using these features, we evaluated the ability of video-based biomechanics to reproduce four TFTs (10-meter walk, 10-meter run, timed up-and-go, and 5-time sit-to-stand) while capturing additional disease-specific signatures of movement. ResultsVideo-based biomechanical analysis reproduced all four TFTs (r > 0.98) with similar test-retest reliability. In addition, video metrics outperformed TFTs at disease classification (p = 0.021). Unlike TFTs, video-based biomechanical analysis identified disease-specific signatures of movement such as differences in gait kinematics that are not evident in TFTs. ConclusionVideo-based biomechanical analysis can complement existing functional movement assessments by capturing more sensitive, disease-specific outcomes from human movement. This technology enables digital health solutions for assessing and monitoring motor function, complementing traditional clinical outcome measures to enhance care, management, and clinical trial design for movement-related conditions. DescriptionThis study demonstrates that smartphone video-based biomechanical analysis can accurately replicate traditional timed function tests (TFTs) -- commonly used to diagnose and assess movement-related conditions like neuromuscular disorders -- while also capturing disease-specific movement patterns that TFTs fail to detect. By enabling sensitive and interpretable assessments in clinical settings, this approach offers a scalable and objective tool to enhance diagnosis, monitoring, and clinical trial design for neuromuscular disorders.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
16
FWCI
19.27
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (29 Jul 2025) · CC BY 4.0

Smartphone-video biomechanics (OpenCap) reproduced four timed function tests (r > 0.98) and separated facioscapulohumeral from myotonic dystrophy better than test times (balanced accuracy 82% vs 50%; 129 participants).

Timed function tests (TFTs) are widely used but reduce movement to a single time, while laboratory motion capture is sensitive but expensive and slow, and facioscapulohumeral muscular dystrophy (FSHD) and myotonic dystrophy (DM) lack approved digital outcome measures. The study tests whether smartphone-video biomechanics (OpenCap) can reproduce TFTs and add disease-specific information; the authors note that its potential to offer more sensitive assessments than TFTs had not been tested.

Key results

  1. ShownVideo-derived surrogates of the four TFTs predicted stopwatch times closely, with errors below published minimum clinically important differences (MCIDs) wherever one was available.

    Leave-one-subject-out linear regression: Pearson r > 0.98 for all four tests. Mean absolute error (MAE) vs MCID: 10-m walk 0.31 s vs 0.69 s, timed up-and-go 0.22 s vs 0.92 s (DM MCIDs); 10-m run 0.24 s vs 2.3 s (Duchenne MCID); none exists for 5-time sit-to-stand. Test-retest reliability (13 DM participants) was similar to human-timed TFTs; 70% of other video features had at least moderate reliability. · Video-based reproduction of TFTs; Analysis of video-based reproduction of TFTs

  2. ShownFor FSHD vs DM classification, a linear support vector classifier on video features from all nine activities outperformed classifiers built on TFT times alone or on TFT times plus Brooke upper-extremity scores.

    Balanced accuracy 82% (video) vs 50% (TFT times, p = 0.001) and 72% (TFT times plus Brooke scores, p = 0.021); McNemar test, leave-one-subject-out cross-validation. TFT times from single tasks scored 50% (chance), whereas video features from running alone reached 68% and from walking alone 66%. · Disease classification; Analysis of disease classification

  3. SuggestedWalking and running kinematics differed between FSHD and DM even though gait speed, the quantity a TFT captures, did not.

    Walking: FSHD showed shorter stride length (p = 0.037) and higher swing-phase ankle height (p = 0.008) than DM with no significant speed difference; running stride length also differed (p = 0.044) while running speed did not. Exploratory two-sided Kolmogorov-Smirnov tests, Bonferroni-corrected across 34 features; the authors link ankle height to foot drop in FSHD. · Disease-specific movement signatures; Disease classification; Analysis of disease-specific movement signatures

  4. ShownA continuous video metric of overhead arm raising agrees with the ordinal Brooke Upper Extremity Scale but varies widely within each Brooke grade and differs between diseases.

    Metric: larger of the left/right ratios of shoulder abduction to elbow flexion angle. Spearman r = 0.80 against Brooke grades (ninety participants graded 6, 23 graded 5, 21 graded 4), with wide spread of the metric within a grade; FSHD vs DM difference p < 0.001. Reachable area and isolated elbow flexion angle also differed between diseases. · Disease-specific movement signatures; Figure legends

  5. SuggestedFeature attributions point to upper-limb features for detecting FSHD and lower-limb features for detecting DM, in line with the known clinical pattern of each disease.

    Mean absolute SHAP values from linear support vector classifiers (each disease vs controls): top FSHD features came from shoulder and elbow range-of-motion tasks, top DM features from running, walking and jumping. Interpretation rests on agreement with known clinical presentation; no accuracies for these two classifiers are given in the text. · Disease-specific feature importance; Analysis of feature importance

Methods, data and reproducibility

Nine movements were recorded with two or three smartphones via OpenCap (HRNet pose estimation, OpenSim inverse kinematics) while evaluators stopwatch-timed four TFTs. Leave-one-subject-out linear regressions predicted stopwatch times from video surrogates; 34 clinician-informed kinematic features fed linear support vector classifiers, Kolmogorov-Smirnov tests and SHAP.

  • Observational cohort: OpenCap smartphone-video kinematics of nine movements plus stopwatch TFTs, recruited at a high-throughput gym session, support groups/conferences and clinic visits · human · n = 129 individuals; 1,290 OpenCap recordings · 10.5281/zenodo.13788592
  • Participants with facioscapulohumeral muscular dystrophy (FSHD) · human · n = 28
  • Participants with myotonic dystrophy (DM) · human · n = 58 (49 DM1, 9 DM2)
  • Controls with no diagnosed neuromuscular condition · human · n = 43
  • Test-retest subset: DM participants who repeated data collection the next day · human · n = 13
Methods & tools
OpenCap, HRNet, OpenSim Inverse Kinematics, linear support vector classifier, SHAP (Shapley additive explanations), Kolmogorov-Smirnov test, McNemar test, intraclass correlation (two-way mixed effects), Brooke Upper Extremity Scale, timed function tests (10-meter walk, 10-meter run, timed up-and-go, 5-time sit-to-stand)
Code
https://github.com/stanfordnmbl/opencap-fshd-dm-analysis
Data & accessions
https://doi.org/10.5281/zenodo.13788592
Experimental conditions
Kinematics were computed automatically on OpenCap's cloud servers; recording the nine activities took a median of 16 minutes per participant. No hardware or analysis runtime is stated.
Reproducibility gaps
No software versions or environment files are stated. Participants consented to optional sharing of identifiable video, and the text does not say how much raw video the Zenodo release contains.
Limitations (6)
  • Whether the advantage of video analysis carries over to neuromuscular diseases other than FSHD and DM has not been tested. · stated by the authors
  • The researchers who engineered features also collected the data and were not blinded to all participants' diagnoses; the authors judge overfitting risk minimal because clinicians viewed fewer than 30 of the 1,290 recordings. · stated by the authors
  • The convenience sample has few female participants with FSHD, which affects classification for that group; a classifier trained and tested on male participants only reached balanced accuracy 83%. · stated by the authors
  • The analysis is cross-sectional; some video metrics showed low test-retest reliability, and links to patient-reported outcomes, MCIDs, and sensitivity to progression or treatment remain to be established. · stated by the authors
  • The multi-camera setup makes at-home and continuous monitoring harder; the authors see video as suited to periodic assessments in controlled settings rather than long-term tracking. · stated by the authors
  • Reliability rests on 13 participants with DM retested the next day (no FSHD or control retests), and all classifiers were evaluated only by leave-one-subject-out cross-validation within this single cohort, with no external test set. · noted in this summary

Who should read:Read this if you design smartphone-video or markerless motion outcome measures for neuromuscular disease studies, or want an open dataset pairing 3D kinematics with timed function tests in FSHD and myotonic dystrophy.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISCovitz '26Kanzler '20Uhlrich '22Avilés-Carr '26Cornman '21Francis '26
65% semantic matchBioengineering
Sydney C. Covitz, Parker S. Ruth, Shelby Vogt-Domke+14 more·bioRxiv·Oct 2, 2026