Summary
A bioRxiv preprint reports deep-learning models that distinguished people with 12 rare neuromuscular diseases from healthy controls using movement data derived from smartphone video. The researchers also developed an impairment score that correlated with participants’ functional ratings and run speed.
A bioRxiv preprint describes deep-learning models that use movement data derived from smartphone video to distinguish people with neuromuscular disease from healthy controls. The study included 337 people with 12 rare, genetically defined neuromuscular diseases and 78 age-matched controls. The researchers also developed a score for movement impairment and compared it with participants’ own ratings of their functional abilities and with clinical measures such as run speed.
Turning video into movement measurements
The team used OpenCap, a smartphone application for video-based, markerless motion capture. Instead of attaching markers to a person’s body, this approach derives measurements of movement from video. The researchers used those measurements—kinematics describing body-segment motion over time—to train deep-learning models.
Data were collected during regular visits to neuromuscular clinics, as well as at conferences and community events across the United States. This allowed the researchers to study movement outside a single laboratory setting. The abstract describes the dataset as the largest publicly available resource of body-segment kinematics for this group of diseases.
One model task was to distinguish participants with a neuromuscular disease from healthy controls. The reported area under the receiver operating characteristic curve (AUROC) was 0.97, with a reported range of 0.93–1.00. AUROC summarises how well a model separates two groups across different decision thresholds; it is not, by itself, a measure of accuracy at a particular threshold.
A score linked to functional measures
The researchers also derived a movement-impairment severity score. It correlated with participants’ ratings of their own functional abilities (ρ = -0.81, p < 0.001) and with existing clinical measures of movement, including run speed (ρ = -0.83, p < 0.001). The negative correlations indicate that the score varied inversely with these measures; the abstract does not give further detail on how the score is scaled.
The findings suggest that video-derived movement measurements could provide a quantitative complement to clinical assessments. The authors propose that the approach may help inform screening protocols and clinical-trial endpoints, where consistent measurements of function can be useful.
The reported classification is a distinction between people with diagnosed disease and healthy controls. It is not a test shown here to identify which of the 12 diseases a person has or to distinguish neuromuscular disease from other causes of movement difficulty. The work is a bioRxiv preprint, so its findings are preliminary research rather than a validated clinical diagnostic tool.