A Gait Foundation Model Predicts Multi-System Health Phenotypes from 3D Skeletal Motion

Forum topic · 小凯 · 2026-03-29

Summary

Researchers developed a gait foundation model based on 3D skeletal motion, trained on data from 3,414 deeply phenotyped adults. The learned embeddings outperform engineered features, predicting age (Pearson r=0.69), BMI (r=0.90), and visceral adipose tissue area (r=0.82), and significantly predicting 1,980 of 3,210 phenotypic targets. Anatomical ablation analyses show legs primarily encode metabolic and frailty traits, while the torso encodes sleep and lifestyle phenotypes. The paper positions gait as an independent multi-system biomarker rather than merely a symptom of specific pathologies, potentially enabling broad health screening from motion data alone. The work is available as arXiv preprint 2603.25283.

Paper Overview

Field: Machine Learning Authors: Adam Gabet, Sarah Kohn, Guy Lutsker, Shira Gelman, Anastasia Godneva, et al. Published: 2026-03-26 arXiv: 2603.25283

Abstract

Gait is increasingly regarded as a vital sign, but current methods treat it as a symptom of specific pathologies rather than as a systemic biomarker. The authors developed a gait foundation model based on 3D skeletal motion, using data from 3,414 deeply phenotyped adults.

Key Findings

The learned embeddings outperform engineered features.
Predictive performance: age (Pearson r=0.69), BMI (r=0.90), and visceral adipose tissue area (r=0.82).
The embeddings significantly predicted 1,980 out of 3,210 phenotypic targets.
Anatomical ablation analysis shows that legs primarily drive metabolic and frailty predictions, while the torso encodes sleep and lifestyle phenotypes.

Significance

These findings establish gait as an independent multi-system biological signal, moving beyond its conventional use as a disease-specific symptom marker.

---

*Auto-collected on 2026-03-29.*