The study team recorded three-dimensional skeletal motion using a single-depth camera as participants performed five motor tasks. The cohort included 3,328 deeply phenotyped adults, and investigators sought to treat gait not merely as a symptom but as a systemic biomarker reflecting multisystem health. Instead of relying only on conventional engineered features—measurements such as cadence, stride length and gait speed—the authors trained a foundation model that translates raw joint trajectories into compact numerical representations, or embeddings, that capture how all body joints move together over time.
The pipeline combined standardized motor-task recordings with an embedding model trained to describe time-varying joint relationships. The paper provides figures illustrating the data-acquisition and training workflow, along with embedding spaces that map to biological traits. The authors positioned this approach as a route to preserve fine-grained movement detail that engineered features may miss.
Data indicated that learned embeddings predicted continuous biological traits with high fidelity. The authors reported Pearson correlations of r = 0.69 for age prediction, r = 0.91 for body mass index and r = 0.83 for visceral adipose tissue area, with embeddings outperforming engineered features on these targets. Short, repeatable movement patterns therefore encoded robust signals about body composition and chronological age in this dataset.
The study team contrasted embedding-based models with feature-based models in multiple predictive tasks. Embeddings returned superior performance across several metrics and endpoints shown in the paper’s figures. These results suggest that embeddings synthesize coordinated joint dynamics into representations that better reflect underlying physiology than isolated engineered measures.
Investigators tested the embeddings against a broad set of phenotypes: 3,210 targets spanning 18 body systems. Embeddings significantly predicted 1,887 of those phenotypes, according to the authors. To assess added value, the team compared embedding models to a baseline that included age, body mass index, visceral adipose tissue and height. Embeddings improved prediction across all 18 systems in males and in 17 of 18 systems in females.
Beyond continuous traits, the paper reports improved detection of clinical diagnoses and medication use when embeddings were added to the baseline predictors. Figures in the manuscript summarize system-level gains, and selected phenotype classes—metabolic markers, frailty indices and sleep-related traits—illustrate how movement-derived embeddings can contribute to risk stratification and phenotype inference in a routine clinical setting.
The authors performed anatomical ablation analyses to localize predictive information within the body. Removing or isolating segments of the skeleton allowed the team to test which regions contributed most to specific predictions. Leg dynamics dominated predictions related to metabolism and frailty, the authors reported. Torso motion contained stronger signals for sleep and lifestyle phenotypes.
These region-specific findings imply that different clinical domains map to distinct movement signatures. The ablation experiments offer a path toward interpretability: clinicians and researchers can examine which anatomical subsystems drive particular phenotype predictions and thereby connect model outputs to plausible biomechanical mechanisms.
Access to the dataset is restricted because of privacy and institutional review board requirements. The cohort data belong to the Human Phenotype Project (HPP) and are available to bona fide researchers at https://humanphenotypeproject.org/data-access after completion of a non-commercial agreement initiated via info@pheno.ai; the source states that requests will be reviewed within two months and data will be provided in a secure cloud environment. The publication includes source data and references the HPP for controlled access.
Code used to train and evaluate the gait foundation model is publicly available. The authors provided a GitHub repository (https://github.com/AdamGabet/GaitPredict) and a Zenodo archive, as reported in the paper.
Limitations and open questions recorded in the source include data access constraints driven by privacy and ethical review board conditions. The paper does not report additional implementation details in this summary; researchers interested in reproducing or extending the work should consult the provided code repository and the HPP access procedures to obtain the full dataset and methodological specifics.