Gait foundation model predicts health measures across 18 body systems from one camera

Researchers publishing in Nature Medicine on 6 November 2026 have built a foundation model that reads three-dimensional skeletal motion from a single depth camera and uses it to predict health measures across 18 body systems¹. The model was trained on recordings of 3,328 adults, each filmed during 5 motor tasks and each deeply phenotyped¹.

What the study found

The approach replaces the hand-calculated gait measures clinicians usually rely on, such as cadence and stride length, with learned embeddings: numerical descriptions of how all the joints in the body move together over time¹. The authors argue that engineered features can miss fine detail in how a person moves¹.

On direct comparison, the embeddings beat engineered features at predicting age (Pearson's r = 0.69), body mass index (r = 0.91) and visceral adipose tissue area (r = 0.83)¹. Across a wide screen, they significantly predicted 1,887 of 3,210 phenotypes tested, from measures such as grip strength to sleep quality¹.

The more demanding test was whether movement added anything beyond basic body measurements. Against a baseline model built only on age, body mass index, visceral adipose tissue and height, the gait embeddings improved prediction in all 18 systems for men and in 17 of 18 for women¹. They also improved prediction of clinical diagnoses and of medication use¹.

Ablation analyses, which strip out parts of the body to see what drives each prediction, showed that leg movement carried most of the signal for metabolic and frailty measures, while torso movement carried sleep and lifestyle information¹.

What it means for the market

The authors conclude that gait behaves as an independent signal spanning many body systems, and they call for gait analysis to be built into routine clinical assessment as a "scalable, passive vital sign"¹. For health systems and device makers, the relevant detail is the hardware: a single depth camera, rather than a gait laboratory or body-worn sensors, was enough to capture the input¹. That points to a low-cost capture step that could sit in a clinic corridor or a care setting, although the paper as summarised does not test any such deployment.

For AI vendors, the study is another case of the foundation-model pattern moving beyond text and images into physiological signals, with one learned representation serving many downstream predictions instead of one model per condition¹. What remains unknown from the published summary is how the model performs in populations outside the training cohort, how it would be regulated as a clinical tool, and what it would cost to run at scale.

Caveats

The results come from a single cohort of 3,328 adults, and the published summary does not report external validation in a separate population¹. Many of the reported figures are correlations with existing measurements rather than outcomes such as hospital admission or death¹. Improvements over the baseline are reported by body system, and the summary does not give the size of those gains for each system¹. The authors frame routine clinical use as something their findings motivate, not something they have shown¹.