Summary
A model trained on 3D movement recordings from 3,328 adults predicted health phenotypes across 18 body systems. Its learned movement representations outperformed conventional gait measurements for several traits, including age and body composition.
Researchers report a gait foundation model that uses three-dimensional skeletal motion to predict health traits across 18 body systems. Developed from movement recordings of 3,328 deeply phenotyped adults, the model learned patterns from five motor tasks captured with a single depth camera. It predicted traits including grip strength and sleep quality, as well as measures of body composition.
What the model learned from movement
Many gait analyses rely on engineered measurements such as cadence and stride length. The researchers instead converted patterns of joint movement over time into learned embeddings: numerical representations of how the body moves as a coordinated whole. They then used those representations to predict health phenotypes.
The embeddings outperformed engineered gait features for predicting age, body mass index (BMI) and visceral adipose tissue area. The reported Pearson correlations between predicted and measured values were 0.69 for age, 0.91 for BMI and 0.83 for visceral adipose tissue area. Pearson’s r describes how closely predictions track measured values; it is not a percentage accuracy score.
Across the 3,210 tested phenotypes, the paper reports statistically significant predictions for 1,887. Compared with a baseline using age, BMI, visceral adipose tissue and height, adding gait embeddings improved predictions in all 18 body systems for males and 17 of 18 for females. The model also improved prediction of clinical diagnoses and medication use relative to that baseline.
The researchers used nested five-fold cross-validation, repeated across 15 random seeds for the reported analyses. This tests predictions on held-out portions of the study data. The team also examined which parts of the skeleton contributed information: leg movement was most influential for metabolic and frailty predictions, while torso movement carried information about sleep and lifestyle phenotypes.
From movement signal to clinical use
The findings suggest that movement may contain information about health beyond conventional gait measurements and basic demographic or body-size measures. A depth camera recording is also a relatively simple way to capture skeletal movement compared with systems that require multiple wearable sensors. The study therefore points to a possible route for measuring gait as a broader health signal, rather than treating it only as an indicator of a particular mobility problem.
The results are predictions from a cohort study, not evidence that the model can diagnose conditions or improve patient outcomes in routine care. Cross-validation evaluates performance within the study data; clinical screening would require testing the model in other populations and settings, and assessing how its predictions should guide care. The paper’s findings support further evaluation of gait as a multisystem biomarker, while leaving its practical clinical role to be established.