Can a Wrist Tracker Measure Gait? ElderNet Gets Closer

Your smartwatch may eventually measure more than steps and heart rate. A 2026 validation study reports that a self-supervised model, ElderNet, can estimate gait speed, cadence, and stride length from a wrist accelerometer—without a treadmill, lab visit, or sensor strapped to the lower back.

That is a meaningful technical result because wrists have long been one of the worst places to estimate gait. They swing, gesture, carry groceries, and generally behave like wrists. Yet the new model produced a mean absolute error of 8.82 cm/s (0.088 m/s) for gait speed and an intraclass correlation coefficient of 0.87 against reference measures in 85 people with gait impairment. It also outperformed comparison models using wrist and lower-back sensors. The catch matters: 0.088 m/s is close to the 0.10 m/s change often treated as clinically meaningful for an individual patient.

So ElderNet looks useful for research and population-level screening. It does not yet justify having an app tell one person that their mobility has meaningfully worsened after a small week-to-week change.

Why estimating gait from the wrist has been hard

Gait sensors work best near the body's center of mass, which is why hip and lower-back devices have traditionally had an advantage. A wrist device must separate walking-related movement from arm swing and every other upper-limb motion that happens during daily life. Comparative work against electronic walkways has documented the remaining accuracy gaps between wearable systems and reference gait measures.

Earlier wrist algorithms often depended on hand-built signal-processing rules. Those rules can work under tidy conditions and then stumble when real life enters the room. ElderNet takes a different route. In the 2026 study by Brand et al., the researchers first pretrained a deep-learning model on unlabeled, free-living wrist-accelerometer data from 819 older adults in the Rush Memory and Aging Project. Self-supervised training lets a model learn patterns in raw data without requiring costly motion-capture labels for every training example.

The model then faced an external gait-estimation test in the Mobilise-D cohort: 85 people with gait impairment recruited across six European centers. Against the study's reference measures, ElderNet's gait-speed error averaged 8.82 cm/s and its ICC was 0.87. It significantly outperformed the lower-back comparison models for gait speed (p<0.001), reversing the usual assumption that a trunk-mounted sensor must beat a wrist device. For cadence and stride length, it also significantly outperformed competing wrist-based approaches (p<0.01).

A better wrist algorithm does not make a wrist a clinical gait lab

The headline result deserves precision. The reported ICC comes from standardized Mobilise-D assessment sessions, not from unrestricted continuous monitoring over months of ordinary life. The study therefore shows that the model can estimate gait well under its validation protocol; it does not establish identical performance when people wear devices inconsistently, move through varied environments, or use different hardware.

Nor does the average error clear the threshold needed for every clinical use. A 0.10 m/s gait-speed change is widely used as a minimal clinically important difference. ElderNet's 0.088 m/s mean absolute error equals 88% of that threshold. An average error is not a guarantee that every estimate misses by that amount, but it sets a practical boundary: a measured change near 0.10 m/s could reflect true change, measurement error, or both.

That makes the model a poor candidate, for now, to determine whether an individual patient responded to rehabilitation or has begun an early decline. It may be far more useful when averaged across groups, used to enrich a clinical-trial cohort, or treated as a screening signal that prompts a proper assessment.

What the disability result can—and cannot—tell us

In the 819-person Rush MAP cohort, ElderNet classified mobility disability with an area under the receiver-operating-characteristic curve of 0.80. The authors report that this exceeded conventional gait-speed and physical-activity summary metrics in that cohort. An AUC of 0.80 means the model had useful discriminatory ability, not that it diagnosed disability with 80% accuracy or replaced a clinician.

The result also lacks a direct comparison with established clinical benchmarks such as the Timed Up and Go or 4-meter walk test. Without that head-to-head evidence, the AUC cannot tell us whether ElderNet improves on standard care. Independent replication is also absent, and Rush MAP was predominantly composed of older white adults from one U.S. metropolitan area. A model pretrained on that population may perform differently across races, body types, socioeconomic settings, diseases, and device models.

Why gait is worth measuring outside the clinic

The appeal of passive gait measurement rests on a reasonable premise: a short walk in a clinic captures only a few minutes of a person's movement. Daily-life walking may reveal something different. In a prospective study, Urbanek et al. (2023) found that free-living gait cadence measured by wrist accelerometer outperformed structured clinic-based mobility assessments for discriminating fall risk in older adults. That finding applies specifically to cadence and fall-risk discrimination; it does not prove that every wrist-derived gait measure outperforms clinic testing for every outcome.

Other studies make the health relevance plausible without validating ElderNet itself. A large observational Watch Walk–UK Biobank study found that real-world wrist-accelerometer walking patterns were associated with regional brain atrophy and white-matter lesion burden in middle-aged and older adults. The study was cross-sectional, so it cannot establish whether altered walking contributes to brain changes, results from them, or tracks another cause.

Gait and functional mobility also carry long-term prognostic weight. In a São Paulo cohort followed for 10 years, Nascimento et al. (2022) found that clinic-based functional mobility performance predicted all-cause and cause-specific mortality in older adults. This supports gait as a consequential marker of health, but it does not show that an ElderNet score predicts mortality. Different device, different measurement setting, different question.

Wrist-derived gait measures may also capture variation outside geriatric frailty. A 2026 cross-sectional study of chronic low back pain associated the condition with remotely collected wrist-accelerometer gait quality and quantity parameters. Association is the operative word: the result cannot show whether pain changed gait, gait contributed to pain, or another factor affected both.

The missing study is the one that follows people forward

ElderNet has not yet been tested as a prospective predictor of falls, dementia, hospitalization, or death. The case for those applications currently joins separate bodies of evidence: ElderNet estimates gait in a validation study; other gait measures associate with clinical outcomes. The chain is sensible, but it remains a chain of inference.

Several design limits reinforce that caution. The 85-person Mobilise-D validation sample was too small to map performance reliably across disease severity, body composition, or device type. The six-site design improves geographic breadth, but it does not solve the precision problem created by a small sample. The Rush MAP training and disability-classification cohort cannot establish performance in populations it scarcely included.

Research on real-world inertial sensors also shows that prediction remains difficult even with other sensor placements and disease-specific models. For example, Ullrich et al. (2023) examined fall-risk prediction using real-world inertial-sensor gait data in Parkinson's disease. That work supports the broader idea that everyday gait contains clinically relevant information; it does not validate a wrist model in the general older population.

What this means if you wear a smartwatch

No consumer smartwatch currently offers ElderNet as a validated clinical feature. The study describes a research method, not a regulatory-cleared product or a tested clinical workflow. It offers no evidence that today's consumer gait scores can diagnose disability, predict dementia, or guide treatment.

The credible conclusion is narrower and more interesting: a wrist accelerometer can now produce gait estimates that compete with, and in this study exceeded, lower-back sensor models in older adults with gait impairment. Self-supervised learning may make that approach easier to scale because it reduces dependence on expensive labeled gait recordings.

Verdict: ElderNet is a real advance in wrist-based gait measurement, with high confidence in its technical validation result. Its best near-term role is research, trial monitoring, and population-level risk stratification. Individual clinical decisions need more precision, broader validation, direct comparisons with standard assessments, and prospective outcome studies.