Not All Step Counters Are Equal — Here's Which Wearable Algorithm Actually Works

Daily step counts have become one of the most widely used proxies for physical activity in epidemiologic research and consumer health tracking, with multiple large cohort studies linking higher step counts to reduced risk of cardiovascular disease, metabolic disorders, cancer, and all-cause mortality. Yet the accuracy of the algorithms that convert raw wrist accelerometer signals into step counts has rarely been rigorously tested against a true gold standard in naturalistic, free-living conditions. This study directly addresses that gap.

The research team recruited 20 adults (mean age 36.1 ± 14.7 years; 50% female) from San Luis Obispo, CA, between July 2019 and March 2020. Each participant wore a thigh-worn activPAL3 and a wrist-worn ActiGraph GT3X+ simultaneously for seven days. A subset completed two separate three-hour sessions in which they were continuously filmed with a GoPro Hero 5 camera. Five trained coders annotated the footage for posture, activity type, intensity, and step counts — with inter-rater reliability exceeding ICC 0.9 in all categories. This yielded 35 matched observation sessions across 20 participants.

Six step-counting methods were evaluated: the activPAL's proprietary VANE/CREA algorithm, plus five methods applied to ActiGraph data — the proprietary ActiLife, and four open-source algorithms (Oak, SDT, Verisense, and stepcount). Among all wrist-based methods, the OxWearables 'stepcount' algorithm — a hybrid machine-learning plus peak-detection model trained on annotated free-living data — performed dramatically better than its competitors. Its mean absolute percent error (MAPE) was just 17.1% compared to direct observation, versus 40.5% for ActiLife, 41.7% for Verisense, 62.4% for Oak, and a staggering 231.5% for SDT. Both activPAL and stepcount were statistically equivalent to direct observation at the 15% equivalence threshold (two one-sided test, p < 0.05), while no other wrist algorithm achieved even the 15% level.

Variance explained (marginal R²) ranged from 0.64 (SDT) to 0.90 (Oak) at the session level, though high R² did not necessarily indicate low bias — a critical methodological distinction. Over the full seven-day wear period, activPAL and stepcount produced highly concordant estimates (R² = 0.87, MAPE = 12.8%), suggesting these two approaches are interchangeable for population-level step count research and could enable pooling of data across studies that used different device placements.

Activity-type stratification revealed a consistent pattern: all algorithms performed best during sustained walking or running (the intended use case), but accuracy degraded substantially — and variably — during biking, modified walking (e.g., carrying loads, ascending stairs), and mixed movements such as pushing a stroller, food preparation, or computer work. The activPAL's VANE algorithm notably misclassified cycling as stepping at the epoch level, though its daily summary software corrected this. SDT's catastrophic overcount appears driven by false positives during non-locomotor upper-limb movements, a known vulnerability of threshold-based wrist algorithms.

For the longevity research community, the implications are significant. Studies that have used SDT, Oak, or ActiLife to quantify daily steps may have substantially mismeasured exposure, potentially attenuating or distorting dose-response relationships between steps and health outcomes. Researchers designing new wearable studies or reanalyzing existing datasets should strongly consider the stepcount algorithm or, where device placement allows, the activPAL as the preferred step-counting method.