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
TL;DR
Existing wrist-accelerometry research has largely relied on predefined behavioural summaries, leaving the full information in continuous 24-hour movement underused. Sensori addresses this gap with a self-supervised model of raw tri-axial wrist recordings, producing day-level representations that generalize across cohorts and improve disease classification and risk prediction. The findings support 24-hour wrist movement as a scalable source of health information for passive monitoring and disease prediction.
Problem
Existing studies largely reduce continuous wrist movement to predefined behavioural summaries, motivating richer objective measures of health in everyday life.
Method
Sensori uses self-supervised learning to model complete 24-hour recordings of raw tri-axial wrist acceleration and produce general-purpose day-level health representations.
Results
Sensori improved prevalent disease classification for 52 of 102 eligible conditions and 6-year incident disease risk prediction for 26 of 87, with largest gains for neurological and psychiatric disorders.
Takeaways & Limitations
Learned representations of 24-hour wrist movement captured diverse health and disease dimensions and generalized across four population-based cohorts.
Takeaways & Limitations
UKB and CKB may underrepresent frailer people and those with severe disease, multimorbidity or heterogeneous clinical presentations, limiting generalisability across the full health spectrum.
Abstract
from arXiv · showhide
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
Existing wrist-accelerometry studies often reduce continuous movement to predefined summaries, motivating methods that learn clinically meaningful representations from complete 24-hour recordings. Sensori uses self-supervised learning to encode raw tri-axial wrist movement and shows cross-cohort representation of behaviours, health characteristics and disease-related outcomes.
- Introduction: Predefined measures such as activity intensity, step counts and sleep duration capture only a limited portion of continuous, multidimensional behaviour.This limitation motivates modelling 24-hour wrist movement directly rather than relying solely on established behavioural summaries.
- Learning health representations from 24-hour wrist movement: Sensori integrates multiscale signals from full-day raw tri-axial wrist movement into day-level health representations supporting monitoring, function assessment and disease prediction.The model is designed to capture both local patterns, such as tremor and gait asymmetry, and longer-term routines such as inactivity and restless sleep.
- Capturing diverse movement behaviours: Across cohorts, Sensori embeddings predicted device-measured behavioural traits with Pearson’s r ranging from 0.38 to 0.95, versus 0.02 to 0.48 for the demographic baseline.The baseline comprised age, sex and BMI, indicating that the embeddings encoded movement information beyond these variables.
- Encoding key demographic characteristics and multidimensional health axes: Sensori generalized demographic representations across four cohorts, achieving sex-classification AUROCs of 0.996 in UKB, 0.969 in CKB, 0.988 in ELSA and 0.976 in NHANES.Age and BMI were also predictable from the embeddings, with performance exceeding models using only mean and standard deviation of acceleration.
- Encoding key demographic characteristics and multidimensional health axes: Sensori achieved the highest AUROC for all nine cohort–trait combinations when predicting smoking, alcohol drinking and self-rated health across cohorts.For smoking status, AUROCs were 0.833 in UKB, 0.913 in ELSA and 0.844 in NHANES.
- Stability of Sensori representations: Sensori representations were more similar across days within participants than between participants, with median cosine similarities of 0.941 and 0.614, respectively.Adding recording days also improved downstream prediction, although gains were generally largest between one and two days and diminished thereafter.
Discussion
Sensori extends conventional accelerometry by learning general-purpose representations from 24-hour wrist movement that capture health, function and disease-related information across cohorts. Its disease-prediction gains were largest for neurological and psychiatric disorders, but generalisability remains constrained by cohort composition and disease-analysis scope.
- Sensori generalised across four population-based cohorts and captured personal characteristics, physical function, prevalent disease and future disease risk from 24-hour wrist movement alone.
- Sensori could complement wearable health systems that currently rely on predefined behavioural summaries, which may omit information in raw movement signals.
- Sensori produced large prevalent-disease classification gains particularly for neurological and psychiatric conditions, including Parkinson’s disease, multiple sclerosis and essential tremor.
- 26 of 87 conditions showed improved six-year incident risk prediction, with the largest gains for neurological and psychiatric diseases and a median increase in Uno’s C-index of 0.064.
- UKB and CKB may underrepresent frailer individuals and those with severe disease, multimorbidity or heterogeneous clinical presentations because of healthy-volunteer bias.
- Disease analyses were restricted to UKB, and future work should test generalisability in clinically ascertained populations and externally validate findings when longitudinal data accrue.
- Combining wrist movement with other home-based measurements could support more accessible, precise and preventive care at scale.
Methods
The study used raw wrist accelerometry from four cohorts and benchmark datasets to pretrain and evaluate Sensori. The model processes full-day signals with a multiscale neural architecture and self-supervised objectives, then uses fixed embeddings for most downstream analyses.
- Raw tri-axial wrist accelerometer data came from UKB, CKB, ELSA and NHANES, alongside four human activity recognition benchmark datasets.
- Signals were clipped to ±3 g, calibrated to local gravity, low-pass filtered at 5 Hz, downsampled to 10 Hz, and low-quality days were excluded.
- Sensori used a convolutional encoder and transformer backbone to convert 24 hours of 10-Hz tri-axial movement into a day-level representation.
- Masked reconstruction and participant-level contrastive learning were combined to capture within-day temporal structure and stable behavioural characteristics across repeated observations.
- Pretraining used 502,823 person-days from 91,037 individuals, with the 40-million-parameter model trained for 300 epochs.
- For analyses other than disease prediction, the pretrained model was held fixed and linear probes were fitted to its embeddings across predictor sets.
- Activity-recognition performance was estimated with participant-wise five-fold cross-validation so each participant’s data remained in one fold.
- Participant-level embeddings were formed by L2-normalising and averaging valid wear-day embeddings, with preprocessing parameters estimated from training data.
Competing interests
The disclosures report funding, consulting, presentation and data-collection support for A.D.; all other authors declare no competing interests.
- A.D. received grants and support from the Wellcome Trust, Novo Nordisk, Swiss Re, Health Data Research UK, Google and the British Heart Foundation Centre of Research Excellence.
- A.D. accepted consulting fees from the University of Wisconsin and Harvard University, received conference support, and received a Swiss Re donation for CKB accelerometer data collection.
- All other authors declared no competing interests.
Figures
The figures present Sensori’s architecture, cross-population behavioural and health evaluations, multi-day stability analyses, and disease-prediction comparisons.
- Figure 1: Figure 1 shows how raw 24-hour wrist movement becomes a day-level health representation and outlines self-supervised pretraining and cohort evaluation.
- Figure 2: Figure 2 evaluates minute-level activity recognition, embedding structure, and cross-cohort prediction of device-measured behavioural traits and statistics.
- Figure 3: Figure 3 examines prediction of demographic characteristics, harmonised health traits, self-reported health axes, and physical-function items across cohorts.
- Figure 4: Figure 4 compares within- and between-person embedding similarity, one- to six-day input performance, and gains over conventional accelerometry summaries.
- Figure 5: Figure 5 compares clinical covariates alone, covariates plus acceleration summaries, and covariates plus fine-tuned Sensori embeddings for prevalent and six-year incident disease outcomes.
1 Sensori model development
Sensori combines convolutional encoding, pooling and transformer modelling to represent a full day of wrist movement. Its pretraining uses complementary objectives to capture contextual movement structure and participant-level invariants, with ablations guiding the final design.
- Model architecture: Sensori transforms raw 24-hour movement into minute- and five-minute-level embeddings before modelling dependencies across the full day.The convolutional encoder produces minute-level embeddings, mean pooling produces 288 five-minute embeddings, and the transformer contextualises them.
- Pretraining objectives: Masked reconstruction trains the transformer to identify true masked embeddings among distractors from the same recording.The objective uses cosine similarity and an InfoNCE-style loss over masked positions.
- Pretraining objectives: Instance contrastive learning compares matched behavioural states across days to isolate individual-specific invariants despite environmental and life-event confounding.The method contrasts states such as walking with walking and sleeping with sleeping rather than arbitrary windows.
- Pretraining objectives: Day contrastive learning treats different days from the same participant as positives and days from other participants as negatives to capture day-level invariants.The objective was intended to capture stable participant characteristics across repeated observations.
- Pretraining experiments: Demographic-aware contrastive grouping did not yield a noticeable performance improvement in early experiments.The tested grouping used characteristics including age, sex and BMI.
- Pretraining experiments: The masked-plus-day-contrastive model with delayed start achieved the top result on three of four tasks, while the final design used 40M parameters and relative positional embeddings.Performance increased monotonically with training epochs; limited days per participant may have reduced day contrastive effectiveness.
2 Study populations and benchmark datasets
The study draws on population cohorts from the United Kingdom, China and the United States, alongside curated human activity-recognition benchmark datasets. Supplementary materials document cohort variables and benchmark subject, class and duration composition.
- Population cohorts: The population-based datasets comprise UKB, CKB, ELSA and NHANES, with cohorts spanning the United Kingdom, China and the United States.Supplementary inclusion figures cover UKB, CKB, ELSA and NHANES.
- Cohort variables: Supplementary cohort tables define the harmonised variables used for UKB, CKB, ELSA and NHANES analyses.Unmapped non-response categories were treated as missing where specified.
- Benchmark datasets: The benchmark dataset overview reports subject and class counts and analysed durations for the curated subsets rather than complete source datasets.The benchmarks include PAMAP2, RealWorld, WISDM and CAPTURE-24.
- Benchmark datasets: Benchmark class composition is reported as one-minute-window counts and within-dataset percentages for each activity class.These distributions are provided in Supplementary Table 9.
3 Evaluations
Evaluations test Sensori across activity recognition, behavioural traits, demographic characteristics, health axes and disease prediction. The analyses use participant-wise validation, cross-cohort testing and comparisons with handcrafted, general-purpose and domain-specific baselines.
- Human activity recognition: The handcrafted benchmark baseline contains 22 window-level features derived from axis statistics, correlations, acceleration magnitude and spectral characteristics.The features include spectral entropy and the frequencies of the two highest local power-spectral-density peaks.
- Evaluation protocol: All benchmark methods use participant-wise five-fold cross-validation with class-balanced L2-penalised logistic-regression linear probes.Regularisation was selected by participant-wise inner cross-validation.
- Behavioural traits: Sensori embeddings were evaluated without refitting across held-out UKB, CKB, ELSA and NHANES for physical-activity, step and sleep traits.The comparison baseline used age, sex and BMI.
- Disease prediction: Disease prediction used supervised fine-tuning with frozen convolutional feature extractors and compared disease-specific and pretrained shared-organ strategies.The shared-organ approach was designed to mitigate extreme class imbalance by sharing supervision across related diseases.