Step Counts Outperform Smartwatch Intensity Algorithms in Health Phenotyping Study

When researchers in Singapore asked nearly a thousand healthy adults to wear a consumer fitness tracker for a week, they uncovered an uncomfortable truth about the devices strapped to millions of wrists worldwide. The same person could be labeled a couch potato by one way of reading the tracker’s data and a dedicated athlete by another, and only one of those labels lined up with what blood tests, body composition scans, and cardiac magnetic resonance imaging actually showed. In a study published in Frontiers in Digital Health, a team from the National Heart Centre Singapore, the National University of Singapore, Duke-NUS Medical School, and Nanyang Technological University compared two ways of turning wearable data into physical activity profiles: one built directly from raw accelerometry outputs such as steps, distance, and floors climbed, and one built from the manufacturer’s own derived metrics, including active minutes and calorie estimates. The verdict was striking. The two approaches agreed on barely half of the participants, and the direct, volume-based picture of activity was far more consistent with objective biology.

The investigation drew on the SingHEART cohort, a population-based study designed to understand the determinants of cardiovascular health among healthy people in Singapore. Volunteers aged 21 to 69 years were recruited through print advertising by the National Heart Centre Singapore, with ethics approval from the SingHealth Centralized Institutional Review Board and written informed consent from every participant. People with known cardiovascular disease, previous stroke, or diabetes mellitus were excluded, and those with a history of hypertension could take part only if they were not on more than one antihypertensive medication. The final analytical dataset comprised 984 adults with a mean age of 45.6 years, of whom 48.3 percent were male and 91.7 percent were ethnic Chinese. Each participant wore a Fitbit Charge HR wrist-worn accelerometer continuously for up to eight consecutive days, removing it only to charge, and the analysis required at least three valid wear days. The mean number of valid monitoring days was 7.1.

The researchers then built two parallel classification systems from the same raw device records. The volume-based approach used direct accelerometry outputs: mean daily steps, the standard deviation and mean average difference of daily steps, and equivalent statistics for distance walked and floors climbed. The intensity-based approach relied on the manufacturer’s algorithmic outputs: mean, standard deviation, and mean average difference of very active and fairly active minutes per day, alongside activity calories and total caloric expenditure. Both metric sets were standardized to z-scores and fed into agglomerative hierarchical clustering with Ward’s minimum variance linkage and Euclidean distance, a method that tends to produce compact, well-separated clusters. Dendrogram inspection, scree plots, and phenotypic interpretability initially identified five clusters for each scheme. To boost statistical power for downstream regression models, the two high-volatility clusters in each solution were merged, yielding four phenotypes for comparison: Sedentary, Average Movers, Regular Exercisers, and High-Variability Movers.

The resulting profiles were physiologically sensible on their own terms. Volume-based Sedentary participants averaged 7,183 steps per day with the lowest day-to-day variability, Average Movers logged 8,635 steps and 6.0 kilometers, Regular Exercisers sustained 12,089 steps, 8.6 kilometers, and 14.0 floors climbed with low variability, and High-Variability Movers reached 13,992 steps, 10.1 kilometers, and 18.6 floors with marked day-to-day volatility. But when the two classification schemes were laid side by side, the agreement collapsed. The adjusted Rand index was just 0.18 in both the five-cluster and four-group configurations, the adjusted mutual information was 0.14, and Cramér’s V was 0.36. Only 46.2 percent of participants, 454 of 983, landed in the same phenotypic category under both systems. Among people the direct accelerometry data called Sedentary, 35.6 percent were reclassified into active categories by the device-derived algorithm. Conversely, 76.1 percent of those the accelerometry labeled High-Variability Movers were demoted into less active categories by the intensity metrics.

To determine which classification better reflected real biology, the team turned to serum metabolomics. Fasting blood samples were analyzed with high-throughput proton nuclear magnetic resonance spectroscopy, quantifying 250 biomarkers spanning 14 lipoprotein subclasses, cholesterol fractions, triglycerides, phospholipids, apolipoproteins, amino acids, glycolysis metabolites, ketone bodies, and the systemic inflammatory marker Glycoprotein Acetyls, or GlycA. Permutational multivariate analysis of variance, adjusted for age, sex, and assay batch, showed that volume-based phenotypes explained a significant share of metabolomic variance, with an omnibus R-squared of 0.93 percent and P less than .001. When active participants, meaning High-Variability Movers plus Regular Exercisers, were contrasted with Sedentary individuals, 106 of the 250 biomarkers differed significantly at a false discovery rate q below 0.05. The pattern read like a textbook of exercise physiology: higher HDL subfractions with a mean family effect of plus 0.19 standard deviations, lower triglyceride-rich VLDL subfractions at minus 0.21, lower LDL subfractions at minus 0.18, and a substantial drop in GlycA at minus 0.36.

The leading individual biomarkers reinforced that coherence. Total HDL cholesterol showed a standardized effect of plus 0.343, HDL phospholipids plus 0.351, medium HDL particle concentration plus 0.351, and apolipoprotein A1 plus 0.341, all at q below .001. The intensity-based classification told a very different story. Although its omnibus metabolomic test was nominally significant at P equal to .003, the direct contrast between device-defined active and sedentary groups showed no significant separation at all, with an R-squared of 0.23 percent, P equal to .29, and zero of 250 biomarkers reaching significance. The most striking intensity-based result was a mixed profile: device-derived High-Variability Movers, compared with Average Movers, showed lower IDL and LDL concentrations but simultaneously elevated GlycA inflammation and increased VLDL particles. In sensitivity analyses adjusting for smoking and body mass index, the volume-based active-versus-sedentary contrast remained significant with 48 biomarkers, while the intensity-based contrast did not, with only 10 biomarkers and P equal to .14.

Body composition and cardiac imaging sharpened the divergence further. Using a multi-frequency bioelectrical impedance analyzer and cardiac magnetic resonance scanners, the researchers assessed 16 outcomes including body fat percentage, skeletal muscle percentage, muscle-to-fat ratio, and indexed measures of ventricular mass, volumes, stroke volumes, and ejection fractions. Volume-based phenotypes were associated with 11 of the 16 outcomes at false discovery rate q below 0.05, following clear dose-response relationships. Relative to Sedentary participants, volume-based High-Variability Movers showed lower body fat percentage, higher skeletal muscle percentage, and an elevated muscle-to-fat ratio, alongside physiologically expected cardiac adaptation: higher left ventricular mass index, larger left ventricular end-diastolic and stroke volume indexes, and larger right ventricular end-diastolic volume index, all without impairment of ejection fraction. Intensity-based phenotypes, by contrast, reached significance for only 6 outcomes, often in paradoxical directions. Device-derived High-Variability Movers had significantly higher body mass index and waist circumference than Sedentary individuals, and showed reduced left ventricular end-diastolic volume index, reduced left ventricular stroke volume index, and reduced right ventricular end-diastolic volume index, while device-derived Regular Exercisers did show the expected cardiac remodeling.

What could explain an algorithm that labels sedentary people as active and genuinely active people as sedentary? The authors point to how commercial wearables estimate intensity. Many devices combine wrist-worn optical photoplethysmography heart rate data with accelerometry to derive active minutes and energy expenditure, but heart rate measured this way can be driven by factors unrelated to exercise, including psychological stress, elevated ambient temperature, caffeine intake, and autonomic variability. An elevated heart rate without movement may therefore be misread by the algorithm as physical activity, potentially pushing less active or more stressed individuals into higher intensity categories. Commercial algorithms also commonly anchor intensity to age-predicted maximal heart rate rather than measured cardiopulmonary capacity, which varies widely between individuals and can overestimate vigorous activity in some people while underestimating it in the highly trained. Prior validation work has shown weak correlations in METs-based measurements across devices, systematic errors in optical heart rate monitoring for darker skin tones, and misclassification of activity intensity in populations with higher body mass index, raising concerns about both accuracy and equity.

The researchers are careful about the boundaries of their evidence. The cross-sectional design cannot establish causal relationships between activity patterns and the metabolomic, body composition, or cardiac findings, and longitudinal studies with serial metabolomic profiling would be needed for that. The intensity-based results reflect one manufacturer’s algorithms and may not generalize to other brands, although the authors argue this strengthens the case for raw metrics that are more comparable across devices. The absence of a minimum daily wear-time threshold may have inflated variability estimates, wrist-worn accelerometers capture cycling, swimming, and resistance training poorly, and the predominantly Chinese, healthy cohort limits generalizability while preventing analysis of racial disparities in sensor accuracy. The proprietary algorithms themselves were not accessible, two CMR scanner platforms were used without scanner type as a covariate, and larger studies would allow examination of effect modification by demographic factors.

Even with those caveats, the implications for digital health research are direct. Because volume-based activity metrics showed more physiologically consistent associations across circulating metabolomics, body composition, and cardiac structure than metrics produced by proprietary algorithms, the authors argue that researchers should prioritize direct accelerometry data such as steps, distance, and floors climbed. Such metrics are transparent, interpretable without knowledge of black-box code, and more likely to be reproducible across device brands, an increasingly important consideration as wearable data informs clinical decision-making and health policy. The team also calls for standardized, open-source algorithms for deriving activity intensity from accelerometry, developed through collaboration among manufacturers, academic researchers, and regulators, and for validation studies using criterion measures such as doubly labeled water. Future work testing whether volume-based phenotypes predict incident cardiovascular events and mortality would establish whether these transparent digital phenotypes can earn a place in risk stratification.

Subject of Research: Comparison of volume-based versus intensity-based wearable activity metrics for physical activity phenotyping and their biological associations

Article Title: Physical activity phenotypes from volume-based versus intensity-based wearable metrics: discordant classification and divergent metabolomic, body composition, and cardiac magnetic resonance associations

Article References: Goh, S. S. S., Yeo, K. K., Yap, J., Kong, S. C., Lim, W. K., Wong, M. K. F., Ng, E. Y. K., Shen, X., Tan, A. W., & Huang, W. (2026). Physical activity phenotypes from volume-based versus intensity-based wearable metrics: discordant classification and divergent metabolomic, body composition, and cardiac magnetic resonance associations. Frontiers in Digital Health, 8, Article 1982891. https://doi.org/10.3389/fdgth.2026.1982891

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1982891

Keywords: wearable devices, physical activity, accelerometry, metabolomics, cardiac magnetic resonance, body composition, cluster analysis, digital health, Fitbit, METs, SingHEART, validation studies