Gait is increasingly recognized as a vital sign, but it is usually treated as a symptom of specific conditions rather than a signal of overall health. A new study in Nature Medicine takes a different view.
Researchers built a gait foundation model from 3D skeletal motion, recorded during five simple motor tasks by 3,328 adults in the Human Phenotype Project. Classic gait analysis relies on hand-picked measures such as stride length and cadence. This model instead learns embeddings: sets of numbers that describe how all of the body's joints move together over time.
Key findings
- Learned embeddings outperformed engineered gait features at predicting age (r = 0.69), body mass index (r = 0.91) and visceral fat (r = 0.83).
- Embeddings significantly predicted 1,887 of 3,210 health measures tested.
- Compared with age, BMI, visceral fat and height alone, gait improved prediction in all 18 body systems in men and 17 of 18 in women, as well as clinical diagnoses and medication use.
- Leg dynamics drove metabolic and frailty predictions; torso dynamics encoded sleep and lifestyle.
Our part
BioPilot co-founder Assaf Rotem is a co-author of the study, together with Prof. Eran Segal's lab at the Weizmann Institute of Science and the team at Newton Tech, whose Movement Lab captured the motion data.
After setting the baseline — the power of canonical gait measures from the 3D motion data to predict medical conditions — it was exciting to find that the foundation model’s learned embeddings are more informative!
Why it matters
This is a classic BioPilot problem: raw, high-dimensional biological signals turned into learned representations that predict health. Foundation models don't just measure what we already know to look for. They find structure we didn't know was there.
Motion and wearables data is one of our core areas. If you're working on digital biomarkers, clinical-trial endpoints or movement data of your own, we'd like to hear about it.
