A gait foundation model reads signals across 18 body systems from one walk

A Nature Medicine study recorded 3D skeletal motion from 3,328 adults performing five motor tasks with one depth camera and trained a foundation model on the resulting gait embeddings. It outperformed engineered gait features for age, BMI and visceral fat and added information for phenotypes, diagnoses and medication use across 18 body systems. Leg motion carried more metabolic and frailty information, while the torso encoded sleep and lifestyle signals. This is cross-sectional deep-phenotyping research without prospective clinical validation, so a walk cannot yet be treated as a disease diagnosis.