Wrist accelerometers capture continuous signals reflecting physical activity, sleep, and cardiorespiratory motion, but their relationship to future disease across the clinical phenome has been incompletely characterised. This study used two frozen self-supervised foundation models to encode one week of wrist accelerometer recordings from UK Biobank participants and evaluated whether those embeddings predict time-to-event outcomes across many disease categories.
The primary objective was to assess whether a compact, model-derived representation of a single week of movement can serve as a scalable, low-cost predictor of future health outcomes when combined with basic demographics.
The analysis used wrist accelerometry data from 97,696 UK Biobank participants. The raw one-week recordings were passed through two pretrained self-supervised models (referred to as foundation models) to obtain participant-level embeddings. These embeddings, together with participant age and sex, were used to train a multilabel survival model to predict 390 distinct outcomes.
An independent held-out test set of 5,253 participants was used to evaluate primary prognostic performance. A full-cohort out-of-fold analysis was also performed for higher-powered, outcome-specific investigation.
In the held-out sample (n = 5,253), the multilabel survival model produced a mean concordance of 0.688 across outcomes, indicating a robust association between the accelerometry-derived embeddings and subsequent time-to-event outcomes. This demonstrates that a single week of wrist movement encodes meaningful prognostic information when summarised by foundation-model embeddings and combined with age and sex.
Analysis of the model's latent structure found that a single component accounted for 76% of the variance in predicted risk across outcomes. This dominant component was associated with an increased overall disease burden and with mortality, consistent with a shared axis of physiological vulnerability that is reflected in movement patterns.
Although the shared latent axis captured a large proportion of cross-outcome risk variance, disease-specific risk scores still added predictive discrimination. For 85 of 101 well-powered outcomes, disease-specific scores improved discrimination beyond the shared component, indicating that embeddings contain both a general health signal and outcome-specific signatures useful for refining risk predictions.
A higher-powered full-cohort out-of-fold analysis identified prodromal signatures for neurodegenerative disorders. The strongest finding was for Parkinson's disease, where a five-year time-dependent AUROC of 0.90 was reported based on 428 incident cases. The Parkinson's signal showed limited attenuation after applying lead-time washout, suggesting that movement-derived signatures may detect prodromal or early disease-related changes rather than only reflecting short-term prediagnostic symptoms.
Feature-attribution analyses indicated that daytime movement metrics contributed most to the model's predictive performance across outcomes. Sleep-related features and genetic information contributed selectively to prediction for particular outcomes rather than broadly driving model performance. These findings suggest modality-specific contributions within accelerometry-derived embeddings: daytime activity is a dominant predictor for many outcomes, while sleep and genetics add value in a targeted fashion.
The study was conducted under UK Biobank Application Number 62249. In accordance with UK Biobank policy, individual-level derived data including trained model weights and participant-level embeddings are treated as equivalent to individual-level data and cannot be shared outside the UK Biobank environment; aggregate data underlying figures are provided with the manuscript. Code to reproduce analyses will be made publicly available upon publication.
Ethical approval was provided by the North West Multi-centre Research Ethics Committee of the National Health Service. The authors declare no competing interests. The work is reported as a medRxiv preprint; it has not undergone peer review and should not be used to guide clinical practice.
Limitations explicitly noted or implied by the source include the preprint status of the report and the data-sharing constraints imposed by UK Biobank policy. The source does not report additional implementation details such as specific model architectures, training hyperparameters, or calibration metrics beyond the summary performance statistics described above.
The findings establish that one week of wrist accelerometer data, encoded through frozen self-supervised foundation models and combined with simple demographic covariates, can provide a scalable representation of future health risk across many outcomes. The presence of both a dominant shared risk axis and disease-specific signals supports use cases ranging from population-level risk stratification to targeted surveillance for conditions with prodromal movement changes such as Parkinson's disease.
However, translation to clinical practice will require peer review, external replication, transparent sharing of model details where permissible, evaluation of generalisability to other cohorts and devices, and study of how such predictions would be implemented ethically and effectively in clinical or public health settings.