Biomedical Research · global
Gait Holds Clues to Whole-Body Health: AI Predicts Metabolic and Frailty Measures from Three-Dimensional Motion
A study involving 3,328 adults used AI to identify multiple health traits from joint movements. The findings offer a research direction for noninvasive screening, but whether predictive performance can translate into clinical benefits remains to be validated.
How the legs alternate and the torso maintains balance during walking may reveal more about health than walking speed alone. A new study in *Nature Medicine* trained an AI model on three-dimensional skeletal motion to explore whether body movements could reveal clues about metabolism, frailty, sleep, and other aspects of health, paving the way for more accessible health assessment tools in the future.
The original paper was published on November 6, and the journal published a research briefing on November 9. The research team used data from 3,328 adults, recording their skeletal movements with a single depth camera as they completed five movement tasks, then comparing those movements with detailed health measurements. This approach requires specific equipment and task procedures; the findings do not yet support assuming that ordinary smartphone videos would produce the same results.
Traditional gait analysis often begins by calculating measures such as cadence and stride length. This model instead learned patterns in how joints move together over time, converting complex movements into numerical features for prediction. The study found that these movement features had statistically significant predictive ability for 1,887 of the 3,210 health traits tested, but that number does not mean the model can diagnose the same number of diseases.
For several specific measures, the Pearson correlation coefficients between model predictions and actual age, body mass index, and visceral fat area were 0.69, 0.91, and 0.83, respectively, outperforming traditional manually designed gait features. Correlation coefficients describe how consistently two variables change together; they cannot be interpreted as diagnostic accuracy, nor can these numbers alone establish whether prediction errors for individuals are acceptable for clinical use.
Movement information was also more than another representation of age or body size. Compared with a baseline model using only age, body mass index, visceral fat, and height, adding the learned movement features improved predictions across 18 body systems in men and 17 of those systems in women. Predictions of clinical diagnoses and medication use also improved. However, identifying differences in diagnoses and medication use within the data is still not equivalent to predicting future disease onset in advance.
The researchers further examined the contributions of different body regions and found that leg dynamics were most important for predictions related to metabolism and frailty, while torso dynamics carried information about sleep and lifestyle. This provides some interpretable clues to the model’s assessments, but associations between body regions and health traits are not sufficient to establish causality.
Before the model can enter screening workflows, its reliability across different populations and real-world care settings must still be confirmed, along with whether follow-up examinations after an abnormal finding can deliver health benefits. The study has made code and statistical source data available, while participant data must be obtained through a controlled research process. This supports further evaluation and replication; whether gait analysis can improve medical decision-making still requires direct clinical evidence.