One week of wrist movement to predict future disease: what two accelerometry foundation models really read in the UK Biobank

A Stanford and Denmark team (with James Zou and Emmanuel Mignot) encodes one week of wrist movement from 97,696 UK Biobank participants using two frozen foundation models, then trains a survival model to predict 390 future diseases at once. Mean concordance reaches 0.688, one "how much you move" axis explains 76% of predicted risk, and a prodromal Parkinson's disease signature stands out clearly (five-year AUROC 0.90 on 428 cases). But everything rests on a single non-representative cohort, with no external validation, and the encoders were pretrained on those same data.

The context

Wrist accelerometers — the sensor that, in a smartwatch, measures arm acceleration along three axes — have become ubiquitous. Over a week of wear they record physical activity, sleep and fine oscillations related to breathing. We have known for a decade that low activity predicts mortality; we know less about what one week of movement says about the risk of developing this or that specific disease, across the whole "phenome" (the full range of an individual's conditions). Prior work on the UK Biobank — the large British cohort of 500,000 volunteers — mostly extracted activity summaries (minutes of moderate activity, sedentary time) or trained one model per disease. The question in this preprint: can a raw week of signal be turned into a reusable representation that predicts hundreds of diseases simultaneously?

The method

Two "foundation models" — neural networks pretrained on large amounts of unlabelled data, then reusable for many tasks — are put to work, both previously developed by the same team and left frozen (their weights are not re-tuned here). They were trained by self-supervision (the model learns to reconstruct a masked portion of the signal with no human label) on about 109,000 UK Biobank recordings. The first, named HA (Human Activity), is a transformer masked autoencoder — the architecture that processes a sequence by weighting the relative importance of its elements — that splits the signal into 10-second windows and reconstructs a spectrogram (the decomposition of movement into frequencies); it targets daytime behaviours. The second, AcceleRest, works on 30-second windows with a respiratory-amplitude objective, targeting sleep and the cardiorespiratory signal. Each emits a 256-dimensional vector representation (embedding) per window.

The data come from the UK Biobank accelerometry sub-study: an Axivity AX3 worn on the dominant wrist for seven days, sampled at 100 Hz. After quality control, 97,696 participants (56% women, mean age 62, standard deviation 8) are kept, split into 87,652 for training, 4,772 for validation and 5,272 for test, with no participant overlap. Hospital diagnoses (ICD-10 codes) and causes of death are grouped into "phecodes"; those affecting at least 0.30% of the cohort (389) plus all-cause mortality form the 390 outcomes to predict. On the daily representations — condensed then concatenated into a 1,536-dimensional vector — a survival model (which predicts not a label but the time until a disease appears) is trained: a transformer sliding over three-day windows, with a multilabel Cox loss (the reference model in survival analysis, extended here to 390 diseases at once). Age and sex are added, and four models with different random seeds are averaged. Performance is read as concordance (C-index): the probability, between 0.5 and 1, that the model assigns a higher risk to the patient who falls ill first.

The results

On the 5,253 test participants, mean concordance is 0.688 (95% CI 0.679–0.697) across the 388 evaluable outcomes — modest on its face, but above chance for every category. All-cause mortality reaches 0.767. The most striking result is structural: a single axis, found by seeking the principal direction of predicted risks (after removing the effect of age and sex), explains 76% of the variance in risk, and almost every disease projects positively onto it (97% positive loadings). In other words, most of what the wrist captures is a general "health/vigour" gradient — the more you move, the lower your risk of nearly everything. The authors call it SHARC and show it correlates mainly with movement intensity (95th percentile of acceleration, correlation −0.52) and BMI (0.48).

But that is not all. Beyond this shared axis, a disease-specific score further improves discrimination for 85 of the 101 outcomes well supplied with cases. The clearest gains concern neurodegenerative diseases. In test, Parkinson's disease reaches a concordance of 0.91 (on only 25 incident cases) and Alzheimer's 0.85 (15 cases). Across the full cohort (a higher-power "out-of-fold" analysis, 428 Parkinson's cases, median time to diagnosis 4.9 years), the time-dependent AUROC — the area under the ROC curve computed at a given horizon — is 0.90 at five years (CI 0.873–0.923) and 0.87 at seven years. Crucial for ruling out reverse causation: after "lead-time washout" (excluding cases diagnosed within three years of wear, to check the model is not already reading established motor symptoms), the seven-year AUROC only drops from 0.872 to 0.849. The signal thus appears prodromal — preceding diagnosis — rather than a mere reflection of disease already present. The model even beats a published bespoke Parkinson's model (mean prodromal precision 0.18 versus 0.07).

Clinical translation is essential here. An AUROC of 0.90 on a disease affecting 0.44% of the cohort is not a diagnostic test. Flagging the 10% of participants at highest risk captures 77% of future Parkinson's — but those 10% are nearly 9,700 people, the vast majority of whom will never develop the disease. In the prodromal setting, mean precision is 0.18: useful to enrich a trial cohort (61 times the base prevalence), useless to reassure or alarm an individual. That is the difference between triaging a population and diagnosing a person.

What is good

Real scale and breadth. 97,696 participants, 390 diseases predicted by a single model from frozen, reusable embeddings: the marginal cost of adding a disease is zero, and one week of wear is enough. It is a credible demonstration that a raw wrist signal encodes broad health information, well beyond the activity minutes usually extracted.

Honest comparators. Unlike many studies that show only their best metric, the authors measure themselves against the right baselines: age + sex + BMI alone give 0.639, a regression on the UK Biobank's standard activity summaries 0.577, and the wrist alone 0.684 — adding age and sex contributes only +0.005. They go further, showing that age, sex and BMI are themselves reconstructible from the embedding (sex AUROC 0.997, age r 0.852), a rare transparency about the confounders lurking in the representation.

A carefully tested prodromal signal. The three-year washout, internal geographic replication (leave-one-region-out, 0.65–0.67), reported calibration (median slope 1.03) and decision-curve analysis show a team actively trying to break its own result rather than sell it. That the Parkinson's signature survives washout is the paper's most interesting finding.

What is less good

No external validation — and a possible representation leak. Everything rests on the UK Biobank. Yet both encoders were pretrained on the same recordings: only the disease labels are held out, not the representation. The model has never seen another device (the research-grade Axivity AX3 is neither an Apple Watch nor a Fitbit), another population, another coding system. This is precisely the step — external validation — that makes most AI-health models collapse, and it is missing here; the authors concede it explicitly.

A major population bias. The UK Biobank is an archetype of the healthy-volunteer bias: participants more educated, wealthier and healthier than the general population, with few of the very young or the very old. The requirement to wear a sensor for seven days further removes the most affected patients — to the point that, in this cohort, recorded dementia is rarer than Parkinson's, the reverse of reality. Moreover, outcomes come only from hospital and death records: about half of dementias and 39% of strokes are coded only in primary care and thus escape the count, underestimating cases and biasing discrimination.

A single axis that looks like fitness. That 76% of predicted risk hangs on one gradient correlated with movement intensity and BMI invites caution: for the majority of the 390 diseases, the model mostly replays "fit people fall ill less," a useful but unspecific truth, partly confounded with age and body size (which the embedding already encodes). The real novelty — the specific, notably neurodegenerative signatures — rests on sometimes tiny counts (25 Parkinson's cases, 15 Alzheimer's in test), hence wide confidence intervals. Finally, neither weights nor data are shareable (UK Biobank policy) and code is promised only upon publication: immediate reproducibility is nil.

What it changes

For the research community, the idea of frozen, reusable accelerometry embeddings as a phenome-wide substrate is appealing and likely to spread. The obligatory next step is validation on consumer devices and non-British cohorts — without which generalization remains a hypothesis. The prodromal Parkinson's signature deserves prospective testing: as a tool to enrich clinical trials (spotting high-risk populations to test neuroprotectants), not as an individual test.

For clinicians, nothing changes today. No watch "detects Parkinson's" on this basis; what is shown is that one week of movement contains a statistical risk signal, at the scale of a very particular population. The population/individual distinction here is anything but academic.

For patients and the public, two messages. First: part of what your wrist "knows" about your future health is simply your general fitness, age and body size included. Second: an impressive AUROC on a rare disease licenses no individual alarm — of the 10% flagged, almost none will develop the disease.

To go further

The preprint "Dual foundation models for accelerometry predict future health" is available on medRxiv (10.64898/2026.07.24.26358894), by Marcus Dige, Niels R. Lorenzen, Magnus Ruud Kjær, Angus Burns, Poul Jennum, Emmanuel H. During, James Zou, Emmanuel Mignot and Andreas Brink-Kjaer (Stanford Sleep Medicine and Biomedical Data Science; Technical University of Denmark; Rigshospitalet; Icahn School of Medicine at Mount Sinai; Harvard Medical School). Conducted under UK Biobank Application 62249, ethics approval from the NHS North West Multi-centre Research Ethics Committee; no conflict of interest declared; licence CC BY-NC-ND 4.0. Data and weights not shareable (UK Biobank policy), code announced for publication. On medical foundation models, self-supervision and risk prediction, see our decryptages on representational convergence in medical foundation models, on a masked autoencoder's reconstruction error as a mortality predictor, and on genetic prediction of Parkinson's disease.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. The test concordances for Parkinson's disease rest on a very small number of cases (25); the five-year AUROC of 0.90 comes from the out-of-fold analysis on the full cohort (428 cases). English, Spanish and Chinese translations produced with AI assistance and reviewed.