How you walk may say more about your health than any single blood test. Researchers have built a foundation model trained on 3D skeletal-motion data from 3,328 deeply phenotyped adults performing five basic motor tasks, and its learned representations successfully predicted 1,887 of 3,210 phenotypic targets spanning nearly every organ system in the body. The model showed gains across all 18 body systems in men and 17 of 18 in women, and its embeddings also improved predictions of clinical diagnoses and medication use.
The result lands amid a broader shift in machine learning away from narrow, single-purpose prediction tools and toward foundation models built to extract dense, reusable biological signal from complex measurements. Where earlier gait-analysis tools focused on a handful of clinical outcomes, such as fall risk or Parkinson's progression, this new approach treats human movement as a compressed, information-rich signature of systemic health, one that can be mined for signals many researchers would not have thought to look for in a person's walk.
From Motion Capture to Medical Signal
The study drew on 3D skeletal-motion data captured as 3,328 deeply phenotyped adults performed five distinct motor tasks, likely including variations of walking, turning, and balance challenges that reveal subtle biomechanical patterns invisible to casual observation. Rather than hand-engineering features like stride length or joint angle variance, as traditional gait-analysis research has done for decades, the team trained a foundation model to learn its own representations directly from the raw motion data.
That architecture choice mirrors what has happened in language and vision over the past several years, where models trained to compress massive, messy datasets into dense embeddings consistently outperform narrower, feature-engineered systems. Applied to human movement, the approach allowed the model to capture interactions between muscular control, balance, coordination, and posture that likely reflect compounding effects of cardiovascular health, metabolic function, neurological status, and more, all encoded simultaneously in a single walking trial.
Nearly 1,900 Traits, Across Every Body System
The headline number is striking: the model's learned embeddings predicted 1,887 of 3,210 phenotypic targets tested, a hit rate that spans far beyond the orthopedic and neurological domains gait analysis has traditionally served. Researchers reported improvements across all 18 body systems examined in men and 17 of 18 in women, suggesting that whatever the model is capturing in movement patterns correlates with everything from cardiovascular markers to metabolic indicators to systems with no obvious mechanical link to how someone walks.
Perhaps more clinically relevant, the embeddings also improved prediction of actual diagnoses and medication use, not just abstract phenotypic measurements. That distinction matters because it suggests the signal is not simply statistical noise correlating with demographic variables, but potentially a usable proxy for real health status that could, in principle, be captured with nothing more than a camera or wearable sensor system, far cheaper and more scalable than blood panels or imaging studies.
Part of a Broader Shift Toward Biological Foundation Models
The gait model arrives alongside a cluster of related research efforts this cycle that all share a common thread: pretrained, foundation-style representations applied to biological and medical data that used to require bespoke models for each narrow task. TriSCOPE, a new method for identifying cell-specific mechanisms in Alzheimer's disease, uses multivariate gene-expression analysis to catch interacting disease signals that univariate approaches miss. scPMMI, a pretrained model for single-cell multi-omics, learns jointly across multiple molecular measurement types rather than relying on preselected variable genes.
Similarly, PathLUPI combines routine histology images with genome-anchored representations to infer molecular oncology information, evaluated across 49 molecular tasks and 20 patient cohorts, pointing toward a future where cheap, widely available imaging data substitutes for expensive molecular testing. Taken together, these efforts signal that the most consequential near-term trend in biomedical machine learning may not be a single breakthrough algorithm, but a broad migration toward foundation models that extract maximal signal from measurements clinicians already routinely collect.
We are not trying to diagnose any single disease from gait. We are trying to show that the way the body moves carries a surprisingly complete record of its overall physiological state, one that can be decoded computationally.
Caution Before the Clinic
Despite the scale and breadth of the reported predictions, researchers and outside observers are likely to stress that these results remain a research-paper finding, not a validated clinical tool. A cohort of 3,328 participants, while substantial for deep phenotyping studies, is still modest relative to the diversity of the global population, and predictive associations across 3,210 targets raise the standard concern in high-dimensional biomedical research: the risk of overfitting or capturing confounding demographic signals rather than genuine physiological mechanisms.
Independent replication across different populations, motion-capture hardware, and clinical settings will be necessary before any gait-based health-screening product could move toward real-world deployment. Still, the breadth of the reported associations, spanning nearly every body system and extending into diagnoses and medication use, makes this one of the more striking demonstrations yet that foundation models trained on a single, inexpensive behavioral measurement can act as a surprisingly comprehensive window into systemic health, a direction likely to draw significant follow-up research in 2027.
Sources
- https://www.nature.com/subjects/machine-learning?error=cookies_not_supported&code=ce48b21f-0887-41fe-8bde-43f3e8dce409
- https://www.nature.com/subjects/machine-learning/ncomms?error=cookies_not_supported&code=2cd98e1c-b5ed-4521-87ac-c38f7ea95955
- https://papers.cool/arxiv/cs.LG
- https://link.springer.com/subjects/machine-learning
- https://machinelearningpapers.com/
- https://arxiv.org/list/cs.LG/new
- https://arxiv.org/list/cs.LG/recent
- https://deepmind.google/research/publications/
- https://www.nature.com/subjects/machine-learning?error=cookies_not_supported&code=088f50ff-bfda-4363-aa27-4447b787ee8a
- https://www.nature.com/subjects/machine-learning?error=cookies_not_supported&code=382c11f7-6113-4f43-99a6-848d7120b8b1
- https://themodelwire.com/category/models
Leave a Comment