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

cs.LGAug 30, 2026

Learning Human Health and Diseases from 24-hour Wrist Movement

Authors: Yong Wang, Dylan McGagh, Katya Broomberg, Zizheng Zhang, Jonathan Carter, Junayed Naushad, Laura Brocklebank, Yang Sun,

Organizations: Department of Psychiatry, University of Oxford, Oxford, UK. · Pioneer Centre for SMARTbiomed, Oxford, UK. · Big Data Institute, University of Oxford, Oxford, UK. · Nuffield Department of Orthopaedics, Rheumatology and Musculoskeletal Sciences, University of Oxford, Oxford, UK. · Oxford University Hospitals, Oxford, UK. · Nuffield Department of Population Health, University of Oxford, Oxford, UK. · Department of Engineering Science, University of Oxford, Oxford, UK. · Department of Computer Science, University of Oxford, Oxford, UK. · Department of Statistics, University of Oxford, Oxford, UK. · Department of Epidemiology and Biostatistics, School of Public Health, Peking University Health Science Center, Beijing, China. · Peking University Center for Public Health and Epidemic Preparedness and Response, Beijing, China. · Key Laboratory of Epidemiology of Major Diseases (Peking University), Ministry of Education, Beijing, China. · Department of Behavioural Science and Health, University College London, London, UK. · Institute of Health Informatics, University College London, London, UK. · Interdisciplinary Transformation University (ITU), Linz, Austria. · NIHR Biomedical Research Centre at University College London Hospitals (UCLH) NHS Foundation Trust, London, UK. · National and Kapodistrian University of Athens, Athens, Greece. · Department of Applied Health Sciences, School of Health Sciences, College of Medicine and Health, University of Birmingham, Birmingham, UK. · National Institute for Health and Care Research (NIHR) Biomedical Research Centre: Birmingham, University Hospitals Birmingham NHS Foundation Trust and University of Birmingham, Birmingham, UK.

Abstract

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.

Figures & tables

Characteristic	UKB $n=\text{94,089}$	CKB $n=\text{20,626}$	ELSA $n=\text{3,320}$	NHANES $n=\text{4,605}$
Study design
Device	Axivity AX3	Axivity AX3	Axivity AX3	ActiGraph GT3X+
Wrist placement	Dominant wrist	Dominant wrist	Dominant wrist	Non-dominant wrist
Country	UK	China	UK	USA
Demographics
Age (years), median [P25–P75]	63.6 [56.4–68.7]	65.5 [57.8–71.7]	69.0 [61.5–75.9]	59.0 [51.0–67.0]

Table 1: Characteristics of the study populations. Data are shown for eligible participants from four population-based cohorts: UKB, CKB, ELSA and NHANES.

Figure 1: Study overview. a , Conceptual framework of Sensori. Sensori captures patterns in raw tri-axial wrist movement across multiple temporal scales over a 24-hour period and integrates them into a day-level health representation for downstream tasks. b , Sensori architecture and self-supervised pretraining. Sensori takes 24 hours of raw 10-Hz wrist movement as input and generates a day-level embedding using a multiscale deep neural network. Sensori is pretrained using masked reconstruction at the five-minute level and contrastive learning at the day level. c , Study cohorts and evaluation design. Left, participant partitioning for pretraining and downstream evaluation. Right, cohorts used for each downstream evaluation task.

Figure 2: Sensori representations capture diverse behavioural traits across temporal scales and populations. a , Linear-probe performance for one-minute activity recognition across four external benchmark datasets. Sensori is compared with handcrafted features, general-purpose time-series foundation models (MOMENT and Chronos-2) and domain-specific models (Bio-PM and Harnet). Bars and error bars indicate mean Cohen’s $\kappa$ and s.d. across five participant-wise cross-validation folds, respectively. b , Uniform manifold approximation and projection (UMAP) plots of minute-level Sensori embeddings under structured (PAMAP2) and free-living (CAPTURE-24) conditions. Each point represents a one-minute window coloured by its activity label; CAPTURE-24 is randomly subsampled to 5,000 windows for visualisation. c , Cross-cohort prediction of day-level device-measured behavioural traits and device statistics. Linear probes are fitted on the UKB training set and evaluated without refitting in the UKB test set, CKB test set, ELSA and NHANES. The demographic baseline uses age, sex and BMI, whereas the embedding model uses Sensori 24-hour embeddings. Cell values show Pearson’s $r$ between predicted and target values; cell colour represents values on a scale from 0 to 1, and grey cells indicate unavailable traits. $n$ , number of participants.

Figure 3: Sensori encodes key demographic characteristics and multidimensional health axes. a , Prediction of sex, age and BMI in held-out evaluation sets from UKB, CKB, ELSA and NHANES, shown as receiver operating characteristic curves with AUROC for sex and scatter plots with Pearson’s $r$ for age and BMI. Predictions from day-level Sensori representations are shown in colour, with one colour per cohort, and predictions based on the mean and standard deviation of acceleration (Acc mean/s.d.) are shown in grey. b , AUROC for three harmonised traits—smoking status, alcohol drinking status and self-rated health—across UKB, ELSA and NHANES. These traits were selected for their relevance to health and availability in all three cohorts. Points and error bars indicate AUROC and 95% bootstrap CIs, respectively. c , AUROC for self-reported health axes in UKB and physical function items in ELSA and NHANES. For UKB, linear probes were fitted on the UKB training set and evaluated in the held-out UKB test set; points and error bars indicate AUROC and 95% bootstrap CIs, respectively. For ELSA and NHANES, linear probes were evaluated using within-cohort five-fold cross-validation; points and error bars indicate mean AUROC and s.d. across the five outer folds, respectively. Baseline denotes age, sex and BMI; Acc mean/s.d. denotes the mean and standard deviation of acceleration; Embedding denotes Sensori day-level embeddings. $n$ , number of participants.

Figure 4: Stability across days, performance with multi-day inputs and comparison with conventional device-measured behavioural traits. a , Participant-level cosine similarity of day-level embeddings within and between individuals. Central dots, bars and whiskers represent the median, interquartile range and most extreme values within 1.5

\times

the interquartile range, respectively; violin widths indicate density. b , Linear-probe performance of day-level embeddings using one to six input days at test time. Markers and error bars indicate performance and 95% bootstrap CIs, respectively. The unweighted mean across ten UKB health axes is shown without a CI. c , Gain in linear-probe performance relative to corresponding models using the mean and standard deviation of acceleration (Acc mean/s.d.). In the six task-specific plots, bars show performance gains using device-measured daily traits, hourly traits, their combination and Sensori day-level embeddings. Error bars indicate 95% CIs estimated by paired bootstrap resampling of participants. The rightmost plot shows AUROC gains for each of ten health axes (circles connected by lines) and their unweighted means (diamonds). All analyses were conducted using the held-out UKB test set.

Figure 5: Performance on prevalent disease classification and incident disease risk prediction. a , Prevalent disease classification in UKB, evaluated using AUROC. b , Six-year incident disease risk prediction in UKB, evaluated using Uno’s C-index. For both tasks, three models were compared: clinical covariates alone, clinical covariates plus the mean and standard deviation of acceleration (Acc mean/s.d.), and clinical covariates plus fine-tuned Sensori embeddings. Clinical covariates comprised age, sex, BMI, ethnic background, alcohol consumption and smoking status. Only conditions for which adding Sensori embeddings significantly improved performance over the clinical covariate model are shown; fractions indicate the number of such conditions among all eligible conditions in each disease category. Error bars indicate 95% CIs in the held-out UKB test set.

Wearable Health Monitoring

Wearable health monitoring uses body-worn sensors and AI to track physiological and behavioral signals for continuous, personalized health assessment.

See all 110 papers

Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data
Mar 11, 2026Prithviraj Tarale, Kiet Chu, Abhishek Varghese +4Human Activity Recognition Self-Supervised Time Series Representation Learning
College of Information and Computer Sciences, University of Massachusetts, Amherst, United States · Stevens Institute of Technology, Hoboken, United States · Google Health, Seattle, United States +1
Physical activities enable scalable foundation modelling for broad-spectrum health prediction
Jul 8, 2026Zhenghuang Wu, Yuyao Zhu, Songlin XuWearable Health Monitoring Time-Series Foundation Models
Beihang University, Beijing, China · University of Science and Technology of China, Hefei, China · The Hong Kong Polytechnic University, Hong Kong SAR, China
Towards a General Intelligence and Interface for Wearable Health Data
May 21, 2026Girish Narayanswamy, Maxwell A. Xu, A. Ali Heydari +37Self-Supervised Pre-Training Clinical Outcome Prediction
1Google Research · University of Washington · University of Illinois Urbana-Champaign +2

Mar 11, 2026J/K move · Enter open · S save

Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Prithviraj Tarale, Kiet Chu, Abhishek Varghese, Kai-Chun Liu+3

College of Information and Computer Sciences, University of Massachusetts, Amherst, United States · Stevens Institute of Technology, Hoboken, United States · Google Health, Seattle, United States+1

Human Activity Recognition Self-Supervised Time Series Representation Learning Motion Representation Learning Wearable Sensing

Open paper AskPDF arXiv

Wearable accelerometers enable large-scale health monitoring, yet learning robust human-activity representations has been constrained by scarce labeled data. While self-supervised learning offers a remedy, existing methods treat sensor streams as unstructured time series, overlooking the underlying biological structure of human movement, a factor we argue is critical for effective Human Activity Recognition (HAR). We introduce a novel tokenization strategy grounded in the submovement theory of motor control, which posits that continuous wrist motion is composed of elementary basis functions called submovements. We define our token as the movement segment, a computationally tractable unit of motion composed of a finite sequence of submovements. By pretraining a Transformer encoder via masked reconstruction of these tokens, we shift the learning focus from local waveform morphology to high-level structural and temporal organization. Pretrained on the NHANES corpus (approximately 28k hours; 11k participants), our representations outperform strong wearable SSL baselines across six subject-disjoint HAR benchmarks. Code and pretrained weights are available at https://prithvitarale.github.io/biopm-site/.

Abstract

Wearable Health Monitoring

Explore similar work

Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Physical activities enable scalable foundation modelling for broad-spectrum health prediction

Towards a General Intelligence and Interface for Wearable Health Data