Who measures gait, and with what.
What is actually used in sensor-based gait and physical-activity research. Every entry was verified against a primary source or the manufacturer's official page. Numbers carry their configuration — sampling rate, range, wearing site — because the same device at different settings is a different instrument. Scope is gait analysis only; cardiovascular measurement is out of scope here. Anything that could not be verified is simply not here.
What the field has reached, per task
This is not a leaderboard. Tasks that look alike are different problems once the reference, the cohort or the condition changes, so every figure carries its conditions at equal weight. Three things break comparison: the reference (INDIP's per-step truth, an examiner's stopwatch, a pedometer's daily total), the cohort (for frailty, prevalence moves the AUC on its own), and the condition (supervised versus free living, a single clip versus pooled).
Two studies hold the conditions fixed
The closest thing the field has to a benchmark: conditions fixed, method varied.
| Study | Held fixed | Result |
|---|---|---|
| Kluge 2024 (JMIR Form Res 8:e50035) | the same 83 participants across six disease groups, INDIP reference, ten gait-detection algorithms | the wrist-versus-lower-back gap, quantified |
| Gao 2025 (J Gerontol A 80:glaf034) | the same 1,282 participants, five step counters | 988 to 23,607 steps a day — a 24-fold spread |
1.1Gait detection
| Report | Site | Cohort | Reference | Result |
|---|---|---|---|---|
| Kluge 2024 — best on wrist (Brand 2022) | wrist | 83, six disease groups, free living | INDIP | sensitivity 0.55–0.81 / specificity 0.95–0.98 / gait-duration error 8.9–32.7% |
| Kluge 2024 — best on low back (Iluz 2014) | low back | ibid. | INDIP | sensitivity 0.71–0.91 / specificity 0.96–0.99 / gait-duration error 6.3–23.5% |
| ElderNet (detection), Brand 2024 | wrist | 83 labelled | — | F1 0.829 ± 0.134 |
| This work | wrist | 225 | examiner-labelled walking bouts | F1 0.922 |
The wrist loses to the lower back — 0.16 of sensitivity, on the same data. An arm moves independently of the legs, and that is the cost. Our own 0.922 is supervised structured walking, not the free living the three rows above are scored on.
1.2Step count
| Report | Cohort | Reference | Result |
|---|---|---|---|
| Small 2024 (MSSE 56:1945) | 39 development / 30 external | annotated free-living data | free-living MAPE 12.5% (prior models 65–231%) |
| Watch Walk 2022 | 101, 1,487 walks | instrumented walkway | MAPE 3.0–4.4% (supervised) |
| Gao 2025 (ARIC, five algorithms) | the same 1,282 | method-to-method | 988 – 23,607 steps a day |
| Bai 2021 — Apple Watch 2 | 48, 24 hours | Yamax pedometer | MAPE 20.3% |
| Bai 2021 — Fitbit Charge 2 | ibid. | ibid. | MAPE 17.1% |
| This work (deployment) | 119, 4,179 clips | manual per-step labels | MAE 1.684 steps per 10 s |
| This work (window given) | 111, 216 trials | ibid. | MAE 1.421 steps, contact F1 0.898 |
An order of magnitude separates supervised from free living: 3.0–4.4% on a walkway, 12.5% in the wild, and the latter is the deployment case. The consumer figures are daily totals; nobody has validated those devices step by step.
1.3Walking speed
| Report | Site | Cohort | Reference | Result |
|---|---|---|---|---|
| ElderNet 2026 | wrist | Mobilise-D TVS, 18 | INDIP | MAE 8.82 cm/s, ICC 0.87 |
| Kirk 2024 (Mobilise-D) | low back | 97 | INDIP | lab 6–12 cm/s / free living 9–13 cm/s |
| This work | wrist | 130 | examiner stopwatch | MAE 13.7 cm/s |
The reference matters: per-step truth above, a stopwatch below. The 4 m walk this would replace has a test-retest band of 10.8–14.4 cm/s, so 13.7 sits inside it.
1.4Step and stride length
| Report | Site | Quantity | Cohort | Result |
|---|---|---|---|---|
| Zadka 2024 (npj Digit Med) | low back | step length | 472, >80,000 steps | RMSE 6.08 cm (ICC 0.89) / 4.79 cm over ten steps |
| ElderNet 2026 | wrist | stride length | 18 | MAE 11.61 cm, ICC 0.72 |
| This work | wrist | stride length | 130 | MAE 10.7 cm, r 0.756 |
A stride is two steps, so 6.08 cm and 11.61 cm are not the same measurement — and one is worn at the waist, the other at the wrist.
1.5Cadence
| Report | Cohort | Result |
|---|---|---|
| ElderNet 2026 | wrist, 18 | MAE 3.11 steps/min |
| This work | wrist, 33, pooled 100 clips each | MAE 2.13 steps/min, r 0.951 |
Ours is the pooled figure. Per clip it is 8.08 and per subject 6.29 — a different condition from ElderNet's, and not directly comparable.
1.6Step-time variability
| Report | Measured with | Cohort | Result |
|---|---|---|---|
| Hausdorff 2001 | pressure insoles | 52, followed one year | stride-time variability 106 ± 30 ms in fallers versus 49 ± 4 ms |
| This work | wrist | 94 | r 0.633 against the reference, 65% of its own 0.975 reliability |
No wrist-worn figure was found. Nothing in the searched literature estimates this quantity from the wrist and scores it against a reference; the only comparator is the clinical one.
1.7Frailty discrimination
| Report | Site | Cohort | Prevalence | Result |
|---|---|---|---|---|
| Zheng 2025 (correction issued) | — | care-home residents | — | AUC 0.92 |
| Song 2025 (conference abstract) | thigh | 44 | — | AUC 0.81 |
| Park 2021 | chest pendant, 48 h | 259 | 72% | AUC 0.769 (0.60–0.74 by item) |
| UK Biobank 2025 | wrist | 10,156 | — | mortality C 0.73 (Fried phenotype 0.72) |
| Cobo 2023 | wrist | NHANES, 1,279 | — | AUC 0.69 |
| This work | wrist, free living | 213 | 13.6% | binary AUC 0.726 (gait alone) / 0.754 (with demographics), three-stage κw 0.309 |
0.769 at 72% prevalence and 0.726 at 13.6% are not the same measurement. And demographics alone reach 0.758 in this cohort — above Cobo's 0.69. Read any wrist-accelerometer figure by asking whether it was compared against age, sex and body size.
Research-grade raw recorders
Devices that store the sensor stream itself. Because nothing is reduced to step counts on-device, a different algorithm can be applied later — nearly every large cohort runs on these.
| site | Device | Maker | Site | Sensors | Verified configuration | Used by |
|---|---|---|---|---|---|---|
| AX3 | Axivity (UK) | wrist | 3-axis accel | 100 Hz, ±8 g (UK Biobank setting) | UK Biobank accelerometry sub-study (100k), CAPTURE-24 | |
| AX6 | Axivity (UK) | wrist / low back | 6-axis (accel + gyro) | — | an option for Mobilise-D free-living recording | |
| GT9X Link | ActiGraph (US) | wrist | 3-axis accel | 80 Hz | ARIC (n = 1,282, mean age 83.4) | |
| MoveMonitor+ (MM+) | McRoberts (NL) | low back | IMU | — | Mobilise-D technical validation study (TVS) | |
| Opal V2R | APDM / Clario (US) | whole body | accel, gyro, magnetometer, barometer | 128 Hz for 12 h; up to 24 units synchronized per network | Mobility Lab; maker reports use in 700+ peer-reviewed papers | |
| activPAL | PAL Technologies (UK) | thigh | accel | — | Song 2025 (n = 44, 10 days); thigh placement separates sitting, standing and walking well |
The wearing site decides what a device is good at: the low back for walking speed, the thigh for posture, the wrist for wear-time compliance. This project chose the wrist for the third reason.
Reference systems — what grades the wearables
The instruments the wearables are scored against — their precision is the ceiling of every evaluation. Whether a reference resolves individual steps in free living is what divides this field's difficulty.
| site | System | Composition | Reference for | Source |
|---|---|---|---|---|
| INDIP | 2 foot MIMUs, 1 low-back MIMU, 2 pressure insoles, 2 ankle distance sensors | the only free-living reference in Mobilise-D; resolves individual initial contacts | Mobilise-D TVS | |
| Marker-based motion capture | optical | Mobilise-D in-lab reference | ibid. | |
| GAITRite | instrumented walkway (pressure floor) | spatiotemporal parameters on a walkway; the source of the MDC figures | ref. 7 (Youdas 2010, n = 24) | |
| Yamax Digi-Walker SW-200 | mechanical pedometer | the baseline consumer step counts are compared against | ref. 20 (Bai 2021, n = 48) |
A supervised walkway is the easy case; per-step references in free living are as rare as Mobilise-D's INDIP. Consumer-device validation stops at daily totals against a pedometer.
Consumer wearables, as validated
For consumer devices the number that matters is the independent validation, not the spec sheet. Across 31 studies, step counting on the Apple Watch was validated in only 2 (ref. 21).
| Device | Verified accuracy | Conditions | Source |
|---|---|---|---|
| Apple Watch Series 2 | steps MAPE 20.3% | n = 48, 24 h free living, pedometer reference | ref. 20 |
| Fitbit Charge 2 | steps MAPE 17.1% | ibid. | ref. 20 |
| Fitbit Alta | steps MAPE 20.7% | ibid. | ref. 20 |
| Fitbit Charge / Charge HR | MAPE <25% across 20 studies | systematic review | ref. 21 (Germini 2022) |
Open analysis algorithms
Inseparable from the recorders: the same raw data under a different algorithm gives a different answer — on the same 1,282 subjects, daily step counts ranged 988 to 23,607, a factor of 24 (ref. 19).
| Name | What it is | Source |
|---|---|---|
| ADEPT / Oak / SDT / Verisense / Stepcount | wrist-acceleration step counting, all open source | ref. 19 |
| ssl-wearables (OxWearables) | weights pretrained on 700,000 person-days of UK Biobank | ref. 22 |
| ElderNet | self-supervised learning on free-living wrist data; separate gait-detection and gait-measure models | refs. 15, 16 |
Open datasets
The index — UK Biobank, CAPTURE-24, Mobilise-D TVS, ARIC, NHANES and more — is maintained as its own page.
Open the dataset indexLabs to watch
Groups that hold either the methods or the validation infrastructure of wearable gait measurement. Affiliations and roles were verified on official pages.
OxWearables / Wearables Group
population-scale accelerometry and self-supervised learning; built UK Biobank's activity measures and CAPTURE-24
Center for the Study of Movement, Cognition and Mobility
gait variability, freezing of gait, falls; the origin of variability-based fall prediction, and a co-author of ElderNet
Brain and Movement Research Group (BAM)
regulator-grade validation of real-world gait; coordinates Mobilise-D across 34 institutions in 13 countries
Wearable & Implantable Technology (WIT)
statistical methods and open-source tooling for accelerometry; the ARIC / NHANES step-algorithm comparisons
Falls, Balance and Injury Research Centre / UNSW
wrist-derived gait biomarkers (Watch Walk), validated on 78,822 UK Biobank participants, extended to fall and depression prediction
iCAMP
wearable frailty, falls and the diabetic foot; the prior art on scoring the five frailty items individually from sensors (ref. 10)
Politecnico di Torino, DET
INDIP, Mobilise-D's free-living reference system: foot MIMUs + pressure insoles + distance sensors
Center for Gerontology and Social Science / Frailty Research
the revised J-CHS criteria — the definition behind this project's training labels; also reports daily-gait monitoring versus frailty progression
They divide into method holders (Oxford, Johns Hopkins), validation-infrastructure holders (Newcastle / Mobilise-D, Politecnico di Torino) and clinical-question holders (Tel Aviv, Baylor, NCGG). Three lineages pursue gait from the wrist — Oxford, Tel Aviv and UNSW/NeuRA — and they are the ones this project must compare against directly.
- 7Youdas JW, et al. Responsiveness of 2 procedures for measurement of temporal and spatial gait parameters in older adults. PM&R 2010;2(6):537–543.link
- 10Park C, Mishra R, Golledge J, Najafi B. Digital biomarkers of physical frailty and frailty phenotypes using sensor-based physical activity and machine learning. Sensors 2021;21(16):5289.link
- 15Continuous assessment of daily-living gait using self-supervised learning of wrist-worn accelerometer data. npj Digital Medicine (2026).link
- 16Brand YE, et al. Self-supervised learning of wrist-worn daily living accelerometer data improves the automated detection of gait in older adults. Scientific Reports 2024;14:20854.link
- 19Gao S, et al. Comparing step-counting algorithms for high-resolution wrist accelerometry data in older adults in the ARIC study. J Gerontol A 2025;80(6):glaf034.link
- 20Bai Y, et al. Comprehensive comparison of Apple Watch and Fitbit monitors in a free-living setting. PLoS One 2021;16(5):e0251975.link
- 21Germini F, et al. Accuracy and acceptability of wrist-wearable activity-tracking devices: systematic review. J Med Internet Res 2022;24(1):e30791.link
- 22Yuan H, et al. Self-supervised learning for human activity recognition using 700,000 person-days of wearable data. npj Digital Medicine 2024;7:91.link