Wearable Gait Sensors Can Spot Parkinson’s in Groups, but Not Yet in Individuals

Parkinson’s disease changes the way people walk long before many other symptoms become impossible to ignore. Steps grow uneven, the rhythm of a stride loses its metronomic regularity, and the left and right legs begin to behave slightly differently. For years, researchers have hoped that inexpensive wearable motion sensors could capture these changes well enough to serve as an objective diagnostic signal, replacing the subjective judgment of a clinician watching a patient cross an examination room. A new systematic review and meta-analysis published in PLOS One by Jiahao Pan and Hui Tang puts that hope to one of its most rigorous tests to date, and the results are both encouraging and sobering in roughly equal measure.

The study focused specifically on commercial inertial measurement unit, or IMU, based gait systems, the consumer and clinical-grade devices that bundle accelerometers, gyroscopes, and sometimes magnetometers into a package small enough to strap onto a shoe, a waistband, or a lower limb. Unlike laboratory-grade optical motion capture, which requires cameras, reflective markers, and a dedicated walkway, commercial IMU systems promise something far more scalable: gait measurement anywhere, at a fraction of the cost. The central question the researchers asked was blunt. When these off-the-shelf systems measure gait variability and gait asymmetry, can the numbers actually tell a person with Parkinson’s disease apart from a healthy older adult?

To answer it, Pan and Tang searched four electronic databases and cross-referenced the results, screening a staggering 1,725 studies. After dual independent selection and data abstraction, the strict inclusion criteria, which required cross-sectional designs using commercial IMU-based systems to measure gait variability or asymmetry in Parkinson’s patients at any disease stage alongside healthy adults, whittled the field down to just nine studies. Those nine studies together represented up to 571 people with Parkinson’s disease and 470 healthy controls, a reminder of how thin the evidence base remains even for a technology that is marketed aggressively around the world.

The statistical machinery behind the analysis is worth understanding, because it explains why the conclusions carry unusual nuance. For every gait variable reported in the included studies, the team calculated Hedges’ g, a standardized mean difference that expresses the gap between the Parkinson’s and control groups in pooled standard deviation units, a metric that corrects for the small-sample bias that plagues simpler effect size estimates. Those individual effect sizes were then pooled using an inverse heterogeneity model, a pooling approach designed to remain robust when the underlying studies differ substantially from one another, which they almost invariably do when sensor placement, walking protocols, and disease severity vary across labs.

The headline result concerns gait variability, the degree to which stride-to-stride timing fluctuates from one step to the next. Across the pooled studies, individuals with Parkinson’s disease showed significantly greater gait variability than healthy controls, with a Hedges’ g of 0.368 and a 95 percent confidence interval of 0.139 to 0.597. Statistical significance was determined conservatively, requiring two-tailed z-alpha values of 0.05 or below alongside confidence intervals that did not cross zero. Heterogeneity was assessed with the Q statistic and the I-squared index, with Q treated as significant at an alpha of 0.10, an intentionally sensitive threshold that acknowledges how easily heterogeneity hides in small clinical samples. On its face, this finding validates a core intuition: the Parkinsonian gait really is measurably less regular than a healthy one, even when the measuring device is a consumer sensor rather than a research-grade instrument.

The analysis then sliced the data by sensor configuration, and here the story becomes more complicated. Studies that used multiple IMUs simultaneously showed significantly greater gait variability in the Parkinson’s group, with a Hedges’ g of 0.397 and a confidence interval of 0.221 to 0.574, whereas studies relying on a single IMU did not reach statistical significance. Vendors of multi-sensor systems might read that as vindication. But the between-subgroup difference itself was not statistically significant, meaning the data cannot confirm that adding sensors genuinely improves discrimination rather than simply shifting estimates within the noise of between-study variation. More damning still, the multi-IMU subgroup’s prediction interval included the null value. A prediction interval estimates where the true effect of a future, similar study would plausibly fall, and when that range spans zero, it signals that any individual study in this literature could plausibly find no difference at all. For a field hoping to move from describing group averages to diagnosing individual patients, that is a serious caution.

Gait asymmetry, the second major class of variable, measures how differently the left and right limbs contribute to walking, a quantity that has long seemed biologically plausible as a Parkinson’s marker given the disease’s frequent asymmetrical onset. Here the meta-analysis found nothing. The pooled effect was a Hedges’ g of 0.180 with a 95 percent confidence interval stretching from minus 0.012 to 0.373, a range that includes zero. In other words, based on the available commercial IMU data, there is no statistically significant overall difference in gait asymmetry between people with Parkinson’s disease and healthy controls. The asymmetry signal that seems so obvious at the bedside apparently dissolves into measurement noise and methodological variation when subjected to rigorous pooling.

The authors also scrutinized the literature for small-study effects, the statistical fingerprint of publication bias in which smaller studies that find dramatic results get published while null findings vanish into file drawers. Their tools were doi plots and the LFK index, graphical and quantitative methods that assess whether effect sizes correlate suspiciously with study precision. This kind of bias assessment matters enormously in a field where commercial interests have a stake in the outcome, because an optimistic literature can manufacture the appearance of diagnostic capability that individual data points do not support. The small-study analysis forms part of the paper’s broader verdict, which the authors state with commendable bluntness: the findings provide preliminary evidence of an average group-level difference in gait variability, but it is hard to establish individual-level discrimination, clinical utility, or the superiority of multi-IMU systems.

That distinction between group-level and individual-level performance is the crux of the entire enterprise, and it is where much of the wearables hype has historically stumbled. A pooled effect size of 0.368 tells us that the average person with Parkinson’s walks with more variable stride timing than the average control, but it says nothing about whether a single patient’s gait print falls outside the range that healthy variability produces. The two distributions overlap substantially, which is exactly what the wide prediction interval implies, and overlapping distributions are the fundamental obstacle to diagnostic classification. A clinician cannot act on a population average; a clinician needs to know whether this patient, on this walk, in this hallway, carries the disease. For that task, the current evidence does not yet clear the bar.

None of this means commercial IMU gait analysis is a dead end. Quite the opposite: the technology has obvious appeal for a disease that affects millions worldwide and for which objective, scalable monitoring could transform both clinical trials and everyday care. What the meta-analysis offers is a calibration of expectations and a map of what the field must do next. Future studies will need larger and better-characterized cohorts, standardized sensor placement and walking protocols to shrink the heterogeneity that currently inflates uncertainty, and validation designs that test individual classification rather than group means. Until then, the worn sensor on a shoe remains a promising research instrument, a tool that can genuinely detect the tremor of irregularity in a population’s stride, but one whose verdict on any single walker must still wait for stronger evidence.

Subject of Research: Meta-analysis of commercial IMU-based gait systems for detecting gait variability and asymmetry differences in Parkinson's disease

Article Title: The ability of the commercial inertial measurement unit-based gait system in determining Parkinson’s disease: A systematic review with meta-analysis of cross-sectional studies

Article References: Pan, J., & Tang, H. (2026). The ability of the commercial inertial measurement unit-based gait system in determining Parkinson’s disease: A systematic review with meta-analysis of cross-sectional studies. PLOS One, 21(10), e0360793. https://doi.org/10.1371/journal.pone.0360793

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360793

Keywords: Parkinson's disease, inertial measurement unit, gait variability, gait asymmetry, wearable sensors, meta-analysis, systematic review, digital health, biomechanics, diagnostics, healthy controls, Hedges' g