Nature Medicine Study Demonstrates 3D Skeletal Motion Encodes Multisystem Health Markers Across 1,887 Phenotypes

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  Nature Medicine https://doi.org/10.1038/s41591-026-04724-9
Dr. Adam Gabet et al.

Points

Researchers developed a gait foundation model utilizing three dimensional skeletal motion data captured from three thousand three hundred twenty eight deeply phenotyped adults participating in the Human Phenotype Project.
The artificial intelligence system translated whole body joint coordination patterns into learned embeddings that substantially outperformed conventional engineered metrics in predicting age and visceral fat distribution.
Statistical evaluations established that movement embeddings shared significant associations with one thousand eight hundred eighty seven out of three thousand two hundred ten evaluated health phenotypes spanning eighteen distinct body systems.
Anatomical ablation analyses demonstrated that lower body leg dynamics primarily dominated metabolic and frailty predictions while torso movement patterns specifically encoded sleep and lifestyle related phenotype characteristics.
Authors concluded that comprehensive movement analysis operates as a robust multisystem biosignal establishing markerless gait assessment as a scalable clinical vital sign for future research.

Summary

This study evaluated the capacity of a novel deep learning gait foundation model to predict multisystem health phenotypes from three-dimensional skeletal motion. Published in Nature Medicine by researchers including lead author Adam Gabet and co-author Yuval Brodsky from Newton Tech, the investigation addressed the limitation of conventional functional assessments that reduce human movement to isolated metrics like walking speed. Given that human gait carries complex biomechanical signals, the research sought to determine whether learned movement embeddings could characterize systemic physiological status across diverse clinical domains.

The study analyzed 17,589 movement sequences captured via single-depth cameras during five standardized motor tasks across 3,328 deeply phenotyped adults from the Human Phenotype Project. The foundation model translated whole-body joint dynamics into learned embeddings, which significantly outperformed engineered features in predicting chronological age (r=0.69), body mass index (r=0.91), and visceral adipose tissue area (r=0.83). Across 3,210 tested health phenotypes spanning 18 body systems, the model identified statistically significant associations with 1,887 measures. Relative to baseline models incorporating age, BMI, visceral adipose tissue, and height alone, gait embeddings improved predictive accuracy across all 18 body systems in males and 17 of 18 in females, as well as enhancing clinical diagnoses and medication utilization tracking. Anatomical ablation revealed that lower-body leg dynamics dominated metabolic and frailty predictions, whereas torso dynamics encoded sleep and lifestyle phenotypes.

The authors conclude that human movement functions as an independent, multisystem biosignal capable of reflecting physiological status across nearly every organ system. By demonstrating that markerless 3D skeletal motion capture can quantify metabolic risk, liver health, mental health, and frailty at scale, this foundation model provides a scalable framework for integrating gait analysis into clinical trials and routine healthcare as a passive vital sign.

Link to the article: https://www.nature.com/articles/s41591-026-04724-9

References


Gabet, A., Kohn, S., Lutsker, G., Gelman, S., Godneva, A., Sasson, G., Zulti, A., Krongauz, D., Shaulitch, R., Rotem, A., Doron, O., Brodsky, Y., Weinberger, A., & Segal, E. (2026). A gait foundation model for prediction of multisystem health phenotypes from 3D skeletal motion. Nature Medicine, 1–12. https://doi.org/10.1038/s41591-026-04724-9