The Brief
Health & Medicine 4 min read

How You Walk Knows More Than Your Doctor Does

NAVION

Share

The way a person moves through a room carries information. Not just about muscles and joints, but about the cardiovascular system, metabolic health, neurological function, and dozens of other biological processes happening simultaneously beneath the surface. For decades, medicine has known this in principle. A 2009 white paper by Fritz and Lusardi made the case formally, proposing that walking speed should be considered a “sixth vital sign,” alongside blood pressure, heart rate, and the rest. What followed was substantial research into gait assessment. Most of it, however, stayed narrow. A 2020 systematic review of more than 2,700 instrumented gait studies found that over 99% addressed only technical or condition-specific questions. The broader picture was left unexplored.

A study published in Nature Medicine in November 2026 changes that framing significantly.

What 3,328 People Walking in Front of a Camera Revealed

The research, led by Gabet and colleagues, built what they call a gait foundation model. The input was three-dimensional skeletal motion data, captured by a depth camera while participants performed five different motor tasks. The dataset came from 3,328 deeply phenotyped adults, meaning individuals whose health had been characterized in exceptional detail across a wide range of biological and clinical dimensions. The cohort was drawn from the Human Phenotype Project, a large-scale research effort described in a 2025 Nature Medicine paper by Reicher and colleagues.

The model did not simply measure walking speed or stride length. It generated what researchers call embeddings: compressed numerical representations of how each person moved, encoding patterns across the full complexity of three-dimensional motion. These embeddings were then tested against a vast library of health outcomes.

The results were striking. Out of 3,210 phenotypic targets tested, the gait embeddings significantly predicted 1,887 of them. That is more than half of all outcomes examined, spanning an enormous range of biological territory.

Across Every Body System, Not Just the Legs

This is what most coverage of gait research misses. The predictive power was not confined to musculoskeletal or neurological outcomes, which would be the intuitive expectation. The model provided predictive gains across all 18 body systems examined in men, and 17 out of 18 in women. It also improved prediction of clinical diagnoses and medication use.

The technical architecture behind this draws on a method called masked autoencoding, a self-supervised learning approach established in a 2022 paper by He and colleagues at the IEEE/CVF Conference on Computer Vision and Pattern Recognition. The original technique was designed for image data. The gait research team adapted it to work with three-dimensional spatiotemporal skeletal motion, meaning the model learned to understand movement patterns by training on the data itself, without requiring manual labels for every health outcome.

A parallel foundation model for gait, called GaitDynamics, was published in Nature Biomedical Engineering in 2026 by Tan and colleagues, with a focus on biomechanical estimation in clinical settings. The Gabet study takes a different angle: rather than reconstructing biomechanics, it asks what movement patterns can predict about health across the entire body.

Why This Matters Beyond the Clinic

The deeper significance here is about what kind of data carries health information, and how much of it has been left unread.

A depth camera is not an MRI. It does not require a hospital, a specialist, or a lengthy appointment. The data it captures, three-dimensional skeletal motion during ordinary movement tasks, is the kind of signal that could, in principle, be collected in a wide range of settings. The research does not make claims about deployment or clinical readiness. But the underlying finding, that movement encodes multisystem health information at this scale, opens a question worth sitting with: how much diagnostic signal is already present in the way people move through their daily lives, and how much of it has simply never been read?

There is also a methodological point worth noting. The foundation model approach means the system was not built to predict one specific condition. It was trained to represent movement in a general way, and that general representation turned out to be predictive across an unexpectedly wide range of outcomes. This is the same logic behind large language models: learn a rich general representation, and it transfers to many specific tasks. Applied to the body in motion, the results suggest that movement is a richer information source than medicine has historically treated it as.

For healthcare systems, this matters because it points toward a category of assessment that is non-invasive, scalable, and potentially continuous. For individuals, it reframes what a routine health check might eventually look like. For researchers, it raises the question of what other behavioral or physical signals carry similarly dense biological information, waiting for a model capable of reading them.

In Short

A gait foundation model trained on 3D skeletal motion data from 3,328 adults significantly predicted 1,887 out of 3,210 health outcomes, spanning all 18 body systems in men and 17 out of 18 in women. The finding is not that walking predicts orthopedic problems. It is that the way a person moves encodes information about nearly every system in the body, and deep learning can now begin to read it.

Based on reporting from Nature Medicine.

Written by

NAVION