A foundation model for human movement

Learning health from a day in motion.

Sensori learns general-purpose health representations directly from 24 hours of raw wrist movement—preserving the rhythms, transitions and context that conventional summaries leave behind.

Paper & resources Explore the tutorials

00:0024:00

24 hours one continuous representation

122,640participants

683,617person-days

24 hoursper representation

10 Hzraw tri-axial signal

Main results

A day of movement reaches further.

Sensori moves from pre-defined movement signals towards one representation evaluated across behaviour, populations and health outcomes.

01 / Classification

Prevalent disease

Outcomes with reported added value

52260

Schalkamp et al.1 condition WBM30 of 47 outcomes Inertia-15 disease tasks SensorFM34 of 35 tasks Sensori52 of 102 conditions

8 h24 h7 d

Represented recording span · log scale

Sensori improved 52 of 102 conditions beyond clinical covariates. Prior studies use different outcomes and comparators.

02 / Prediction

Incident disease

Outcomes with reported added value

26130

Schalkamp et al.1 condition Sensori26 of 87 conditions

8 h24 h7 d

Represented recording span · log scale

Sensori improved six-year risk prediction for 26 of 87 conditions.

Sensori: paired-bootstrap 95% CI excludes zero

Selected wearable studies: study-specific comparator

Sources: WBM; Inertia-1; SensorFM; Schalkamp et al.

Evidence index

One signal, tested from minutes to years.

Five baselines

Across model families

Sensori was evaluated against handcrafted features, general-purpose time-series models and domain-specific movement models.

6activity 5steps 7sleep

18 behavioural traits

A day keeps its structure

Frozen day embeddings retained six physical-activity, five step and seven sleep measures.

United KingdomUKB · ELSA ChinaCKB United StatesNHANES

Four population cohorts

Across three countries

Health-related information transferred across UKB, CKB, ELSA and NHANES without refitting.

52/102classification

26/87six-year risk

Disease outcomes

Beyond clinical covariates

Sensori added predictive information for prevalent classification and six-year incident risk prediction in UKB.

Selected results

A closer look at the signal.

Three views of the same representation: local activities, population-level characteristics and one disease example.

01 / Minute scale

Activity recognition

Sensori Cohen’s κRank among six approaches

PAMAP2 1st0.852 ± 0.077

RealWorld 1st0.824 ± 0.026

WISDM 2nd0.807 ± 0.081

CAPTURE-24 2nd0.826 ± 0.012

Mean Cohen’s κ ± s.d. across five participant-wise cross-validation folds. Sensori ranked first on PAMAP2 and RealWorld and second to Harnet on WISDM and CAPTURE-24.

02 / Day scale

Across cohorts

CohortSexAUROCAgePearson’s rBMIPearson’s r

UKB0.9960.8370.733

CKB0.9690.7650.511

ELSA0.9880.8460.710

NHANES0.9760.6770.596

Linear probes were fitted in UKB and evaluated without refitting in held-out UKB, CKB, ELSA and NHANES. Metric families are separated rather than encoded on one shared colour scale.

03 / Health outcome

Multiple sclerosis

A high-signal example of prevalent-disease classification in the held-out UKB test set.

This is a population-level classification result, not an individual diagnostic product.

00.080.160.240.32

Clinical baselinereference

Clinical + Sensori+0.242

ΔAUROC 0.242 · paired-bootstrap 95% CI, 0.174–0.308

The result is the beginning.

Read the ideas behind Sensori, then explore the manuscript, code and practical guides as they become available.

Read the white paper Paper & resources