PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Human Health and Diseases from 24-hour Wrist Movement

Yong Wang, Dylan McGagh, Katya Broomberg, Zizheng Zhang, Jonathan Carter, Junayed Naushad, Laura Brocklebank, Yang Sun, George Nicholson, Dianjianyi Sun, Canqing Yu, Jun Lv, Maxim Barnard, Hubert Lam, Andrew Steptoe, David W. Eyre, Liming Li, Zhengming Chen, Naomi Wray, Spiros Denaxas, Gary S. Collins, Huaidong Du, Aiden Doherty, Hang Yuan

arXiv:2608.29494v1cs.LG

TL;DR

Predefined behavioural summaries leave much of the temporal and contextual information in continuous wrist movement unused, limiting objective health measurement outside clinical settings. Sensori learns day-level representations from raw 24-hour tri-axial wrist acceleration and evaluates them across four cohorts. The representations generalized across populations and improved prevalent and incident disease prediction, particularly for neurological and psychiatric conditions.

  • Problem

    Existing wrist-accelerometer studies largely rely on predefined summaries such as activity intensity, steps and sleep, leaving much of daily movement’s temporal organisation and behavioural context unexplored.

    Source: Introduction · p. 3

  • Method

    Sensori is a self-supervised foundation model that integrates multiscale signals from full-day raw tri-axial wrist movement into day-level health representations.

    Source: Introduction · p. 3, Learning health representations from… · p. 4, §1.2 · p. 28

  • Results

    Sensori improved prevalent disease classification for 52 of 102 eligible conditions and 6-year incident disease-risk prediction for 26 of 87, with the largest gains for neurological and psychiatric disorders.

    Source: Disease prediction · p. 8, Discussion · p. 9

  • Takeaways & Limitations

    Learned representations from 24-hour wrist movement captured diverse health and disease dimensions, generalized across cohorts without retraining, and could support scalable passive health monitoring.

    Source: Introduction · p. 3, Discussion · pp. 9–10

    UKB and CKB may underrepresent frailer or more severely ill individuals, and disease analyses were restricted to UKB, limiting established generalisability across the full spectrum of health and disease.

    Source: Discussion · pp. 9–10

Abstract

from arXiv · show

Much of human health and function unfolds beyond the clinic, through the movements of everyday life. Wrist-worn accelerometers capture these movements continuously, yet their rich signals are often reduced to a small set of predefined behavioural summary measures. Here, we present Sensori, a self-supervised foundation model that learns general-purpose health representations directly from 24 hours of raw tri-axial wrist movement. We developed and evaluated the model across four population-based cohorts from the United Kingdom, China and the United States, comprising 122,640 participants contributing 683,617 person-days of free-living recordings. Sensori condensed each day of movement into a representation that captured diverse movement behaviours, demographic characteristics, health axes and physical function. Evaluation in independent cohorts showed that these representations generalised across populations and measurement settings without retraining. When added to common clinical covariates, Sensori significantly improved prevalent disease classification for 52 of 102 eligible conditions (median delta AUROC, 0.060; range, 0.012-0.242) and incident disease risk prediction for 26 of 87 eligible conditions (median delta Uno's C-index, 0.064; range, 0.025-0.172), with the largest gains for neurological and psychiatric disorders. These findings establish 24-hour wrist movement as a rich and scalable source of health information, with the potential to support passive health monitoring and disease prediction at population scale.

Introduction

Sensori addresses the limits of predefined movement summaries by learning day-level health representations directly from raw 24-hour wrist acceleration. Across cohorts and tasks, these representations captured movement, demographic, health and disease information, transferred without refitting, and improved disease prediction.

Learning health representations from 24-hour wrist movement

  • Sensori models complete 24-hour raw tri-axial wrist acceleration to produce day-level representations that integrate movement signals across temporal scales.The architecture progressively pools minute- and five-minute embeddings before generating a 768-dimensional daily representation, while masked reconstruction and day-level contrastive learning capture contextual and stable individual patterns.Source: Learning health representations from… · p. 4

Introduction

  • Sensori was evaluated across 122,640 participants and 683,617 person-days from four population-based cohorts in the United Kingdom, China and United States.The cohorts differed in study periods, accelerometry protocols, demographics and observed movement behaviours; disease analyses were restricted to UKB because harmonised phenotypes and longitudinal follow-up were required.Source: Learning health representations from… · pp. 4–5, Disease prediction · pp. 7–8

Capturing diverse movement behaviours

  • Sensori embeddings outperformed demographic baselines for device-measured movement traits across cohorts, with Pearson’s r ranging from 0.38 to 0.95 versus 0.02 to 0.48.Sleep traits were not evaluated in NHANES because its idle sleep mode suppressed the low-movement information needed to derive them reliably.Source: Capturing diverse movement behaviours · pp. 5–6

Encoding key demographic characteristics and multidimensional health axes

  • Sensori generalized across cohorts for demographic and health characteristics, achieving sex-classification AUROCs of 0.969–0.996 and the highest AUROC for all nine cohort–trait combinations.It also predicted physical-function items beyond mobility, including preparing meals and attending social events, indicating information about everyday functional expression.Source: Encoding key demographic… · pp. 6–7

Stability of Sensori representations

  • Within-participant day-level embeddings were more similar than between-participant embeddings, with median cosine similarities of 0.941 within participants and 0.614 between participants.Aggregating additional days improved downstream prediction, with the largest gains generally occurring between one and two days.Source: Stability of Sensori representations · p. 7

Disease prediction

  • Adding Sensori to clinical covariates significantly improved prevalent disease classification for 52 of 102 conditions and 6-year incident-risk prediction for 26 of 87 conditions.The largest improvements were concentrated in neurological and psychiatric disorders; conventional daily and hourly behavioural traits were less predictive than Sensori embeddings across health axes.Source: Stability of Sensori representations · p. 7, Disease prediction · p. 8

Discussion

Sensori shows that representations learned from 24-hour wrist movement capture diverse health information and generalise across four population-based cohorts. The study also identifies important population and disease-prediction boundaries, including healthy-volunteer bias, UKB-only disease analyses, possible prodromal disease effects and the absence of multimodal sensor data.

Contribution

  • Sensori captured personal characteristics, physical function, prevalent disease and future disease risk across four cohorts, supporting scalable objective assessment from wrist movement.The authors describe this as extending conventional behavioural-trait research toward machine-learned representations of raw movement.Source: Discussion · p. 9

Disease classification

  • Sensori produced large prevalent-disease classification gains, particularly for neurological and psychiatric conditions including Parkinson’s disease, multiple sclerosis and essential tremor.The authors suggest the representation could help identify conditions for which wearable movement may provide sensitive measures before disease-specific biomarkers are developed.Source: Discussion · p. 9

Risk prediction

  • Sensori improved six-year incident disease risk prediction for 26 of 87 conditions, with the largest gains for neurological and psychiatric diseases.The authors note that gains in other categories were more modest and compare the result cautiously with proteomic profiling.Source: Discussion · p. 9

Limitations

  • UKB and CKB may underrepresent frailer people and those with severe disease, multimorbidity or heterogeneous clinical presentations, limiting generalisability across the full disease spectrum.The authors call for evaluation in clinically ascertained populations.Source: Discussion · pp. 9–10
  • Disease analyses were restricted to UKB, and incident associations may reflect undiagnosed or prodromal disease rather than risk before onset.The authors also note that multimodal wearable measurements were not consistently available and could enrich future representations.Source: Discussion · p. 10

Implications

  • The authors conclude that appropriately combined wrist-movement monitoring could support more accessible, precise and preventive care at scale.This conclusion is framed as potential rather than an established clinical effect.Source: Discussion · p. 10

Methods

The study combines harmonised multi-cohort accelerometry, multiscale Sensori modelling and self-supervised pretraining, then evaluates fixed representations with linear probes and disease-specific fine-tuning. Evaluation protocols preserve participant-level separation and distinguish cross-cohort health tasks from UKB disease prediction.

Data sources

  • Four population cohorts and external activity-recognition datasets supplied wrist accelerometry, with UKB and CKB partitioned for training and testing and ELSA, NHANES and benchmarks reserved for external evaluation.Participants wore accelerometers for seven consecutive days in UKB, CKB and NHANES and eight days in ELSA.Source: Data sources · p. 10

Accelerometry preprocessing

  • Signals were clipped, gravity-calibrated, low-pass filtered at 5 Hz, downsampled to 10 Hz and quality-controlled using wear-time, calibration, acceleration and interruption criteria.Days with fewer than 22 hours of wear were excluded.Source: Accelerometry preprocessing · p. 11

Sensori network

  • Sensori uses a convolutional encoder and transformer to convert 24 hours of 10-Hz tri-axial acceleration into a 768-dimensional day-level embedding.Pooling reduces the encoded sequence to 288 five-minute embeddings before transformer processing; the 40-million-parameter model was selected as the best performance–efficiency trade-off.Source: Sensori network · p. 11

Pretraining

  • Pretraining combines five-minute masked reconstruction with participant-level day contrastive learning to capture within-day temporal structure and stable characteristics across repeated observations.Reconstruction preceded contrastive training, and augmentations included signal scaling, channel swapping and random start times.Source: Pretraining · pp. 11–12

Linear probing

  • For non-disease analyses, the pretrained model was frozen and compared fairly with alternative predictors using regularised linear probes and participant-disjoint evaluation.Binary, continuous and survival outcomes used logistic, ridge and Cox models, respectively, with preprocessing parameters estimated only from training partitions.Source: Linear probing · pp. 12–13

Disease prediction

  • Disease analyses used UKB phenotypes, excluded prevalent cases and early events for incident outcomes, and analysed 102 prevalent and 87 incident conditions with supervised fine-tuning.Prevalent disease used AUROC, whereas six-year incident risk used Uno’s C-index; paired bootstrap confidence intervals assessed between-model differences.Source: Disease prediction · pp. 13–14

Competing interests

The disclosures report extensive external funding and consulting relationships for A.D., while all other authors declare no competing interests.

Disclosures

  • A.D. reports support from multiple funders, consulting fees, conference support and a Swiss Re donation for China Kadoorie Biobank accelerometer data collection.The listed funders include Wellcome Trust, Novo Nordisk, Swiss Re, Health Data Research UK, Google and the British Heart Foundation Centre of Research Excellence.Source: Competing interests · pp. 14–15
  • All other authors declare no competing interests.Source: Competing interests · p. 15

Figures

The figures trace Sensori from raw wrist movement and self-supervised learning through representation quality, health-axis transfer, multi-day stability and disease prediction. Together, they show the study’s architecture, evaluation scope and comparison models.

Study overview

  • Figure 1 presents the multiscale framework, 24-hour 10-Hz input, masked reconstruction and day-level contrastive pretraining, plus cohort partitioning and downstream evaluation.Source: Figure 1 · p. 21

Behavioural representations

  • Figure 2 evaluates minute-level activity recognition against handcrafted and foundation-model baselines, visualises embeddings, and reports cross-cohort behavioural-trait correlations against demographic baselines.Source: Figure 2 · p. 22

Health representations

  • Figure 3 reports demographic, harmonised health-trait and physical-function prediction across UKB, CKB, ELSA and NHANES using AUROC or Pearson’s r against baseline predictors.Source: Figure 3 · p. 23

Stability and aggregation

  • Figure 4 compares within- and between-person embedding similarity, performance from one to six input days, and gains over conventional device-measured traits.Source: Figure 4 · p. 24

Disease prediction

  • Figure 5 compares clinical covariates, acceleration summaries and fine-tuned Sensori embeddings for prevalent disease AUROC and six-year incident-risk Uno’s C-index.Only conditions with significant improvement from adding Sensori are shown, with category fractions and held-out UKB confidence intervals.Source: Figure 5 · p. 25

1 Sensori model development

Sensori combines a convolutional encoder and transformer to convert 24-hour wrist movement into day-level representations, using complementary self-supervised objectives selected through ablations.

1.1 Model architecture

  • Sensori transforms 24-hour raw movement into 288 five-minute embeddings, contextualises them with a transformer, and mean-pools the outputs into a day-level representation.The architecture was implemented in 20M- and 40M-parameter configurations.Source: §1.1 · p. 28

1.2 Pretraining objectives

  • Masked reconstruction, instance contrastive learning and day contrastive learning were evaluated to capture complementary movement information across temporal scales.Masked reconstruction targets contextual recovery of five-minute embeddings, whereas contrastive objectives compare matched behavioural states or different days across participants.Source: §1.2 · pp. 28–29
  • Day contrastive learning treats different days from the same participant as positives and days from other participants as negatives to capture participant-level invariants.The activity-intensity contrastive objective instead matches the same intensity decile across two days and uses other participants as negatives.Source: §1.2 · p. 29

1.3 Pretraining experiments

  • The masked-plus-day-contrastive model with delayed start achieved the top result on three of four tasks, while performance increased monotonically with training epochs.Day contrastive was less effective than expected, possibly because few days were available per participant.Source: §1.3 · p. 30, Table 2 · p. 30
  • The selected design used the 40M model, relative positional embeddings, a 5-minute masking horizon and no weighted masking for masked reconstruction.These choices were informed by the architecture and masking ablations.Source: §1.3 · p. 31, Table 3 · p. 31

2 Study populations and benchmark datasets

The study assembled four population cohorts from the United Kingdom, China and the United States alongside four curated human-activity-recognition datasets for external benchmarking.

2.1 Prospective cohort studies

  • The prospective cohorts comprised UKB, CKB, ELSA and NHANES, with participant inclusion documented separately for each cohort.Variable-definition tables specify cohort-specific demographic and health measures.Source: Figure 2 · p. 32, Figure 3 · p. 33, Figure 4 · p. 34

2.2 Human Activity Recognition

  • The human activity recognition benchmarks included PAMAP2, RealWorld, WISDM and CAPTURE-24, with curated subject, class and duration information reported for analysis subsets.Class composition was reported as counts and within-dataset percentages.Source: Table 8 · p. 39, Table 9 · p. 39

3 Evaluations

Evaluations tested Sensori across activity recognition, behavioural traits, demographics, health axes and disease prediction using participant-wise validation and cross-cohort assessments.

3.1 Human activity recognition benchmarks

  • Human activity recognition used 60-second windows, participant-wise five-fold cross-validation and frozen embeddings from Sensori and five comparison methods.The comparison included handcrafted features, MOMENT, Chronos-2, Bio-PM and Harnet.Source: §3.1 · p. 40

3.2 Evaluating day-level behavioural traits

  • Day-level behavioural traits covered physical activity, steps and sleep, with Sensori probes trained in UKB and evaluated without refitting across UKB, CKB, ELSA and NHANES.The behavioural labels were algorithm-derived from free-living cohort accelerometer data.Source: Table 10 · p. 42, Table 11 · p. 43

3.3 Evaluating key demographic characteristics and multidimensional health axes

  • Supplementary analyses separately examined the ten UKB health axes and compared them with conventional device-measured traits.The analyses matched the main-text evaluations while displaying each health axis individually.Source: Table 12 · p. 44, Figure 7 · p. 45

3.5 Disease prediction

  • Disease prediction fine-tuned downstream Sensori layers while freezing the convolutional extractor and selected shared-organ fine-tuning to address rare-disease imbalance.The shared-organ model reduced fine-tuning computation by approximately 99% compared with 102 disease-specific models.Source: §3.5 · p. 46, Table 14 · p. 50
  • Supplementary disease analyses compared clinical covariates alone with acceleration statistics, conventional behavioural traits and Sensori embeddings for prevalent and incident outcomes.Performance was summarised with AUROC for prevalent classification and Uno’s C-index for incident risk prediction, with 95% confidence intervals in the UKB held-out test set.Source: Figure 8 · p. 51, Figure 9 · p. 52, Figure 10 · p. 53

Related papers

Loading 2608.29494v1…