The validity of smartphones and smartwatches to measure step metrics in free-living conditions

Background Consumer devices (i.e., smartphones and smartwatches) have the potential to monitor and promote physical activity (PA) at a population level via step metrics. However, concerns remain about the validity of step metrics derived from consumer devices. Purpose To assess the validity of consumer devices for measuring step metrics against research-grade accelerometers in free-living conditions. Methods We recruited adults from a general population cohort and a cohort of individuals with cardiovascular disease (CVD). Participants wore a research-grade thigh-worn accelerometer for 8 consecutive days. We concurrently collected participants’ step count data from their smartphone or smartwatch. We calculated daily step count, peak 1- and 30-min cadence (average steps/min recorded for the 1 and 30 highest minutes, resp.) and minutes at moderate-to-vigorous cadence (MVPA) (>100 steps/min). We determined the mean bias with 95% limits of agreement (LoA) and intraclass correlation coefficients (ICC) with 95% confidence intervals (CI). We assessed equivalence by testing whether the 90% CI of the mean difference between consumer devices and accelerometers lay within 10% of the accelerometer mean. We performed stratified analyses for consumer device type and CVD status. Results We included 1,306 measurement days among 215 adults (age: 65±10 years; 38% female; BMI: 25±4 kg∙m-2; 53% with CVD), of whom 48% used a smartphone and 52% a smartwatch. Consumer devices underestimated step count (10,414±8,360 vs. 12,608±8,294 steps; bias -2,194 (LoA -9,041; 4,653) steps), peak 1-min cadence (105±26 vs. 119±15 steps/min; bias -14 (LoA -57; 29) steps/min), peak 30-min cadence (86±33 vs. 101±22 steps/min; bias -15 (LoA -55; 25) steps/min), and minutes of MVPA (35±57 vs. 45±60 min; bias -10 (LoA -44; 24) min) compared to accelerometers. ICC was good for step count (0.88 (95% CI 0.72-0.93)) and excellent for MVPA (0.94 (95% CI 0.88-0.97)), but low for peak 1-min cadence (0.38 (95% CI 0.15-0.54)) and moderate for peak 30-min cadence (0.62 (95% CI 0.27-0.78)). The 90% CI of the mean difference between consumer devices and accelerometers did not fall within 10% of the accelerometer mean for any step metric, indicating non-equivalence. Smartwatches had less underestimation in step count (bias -1,234 (LoA -7,853; 5,386) vs. -3,218 (LoA -9,740; 3,304) steps) than smartphones. Individuals without CVD had less underestimation in step count (bias -1,777 (LoA -8,510; 4,946) vs. -2,596 (LoA -9,467; 4,275) steps) than individuals with CVD. Conclusion Consumer devices systematically underestimate step metrics in free-living conditions, with the highest validity for step count estimates from smartwatches. Our findings suggest that differences in step count and MVPA between consumer devices and accelerometers are consistent, meaning that these biases can be accounted for when using consumer devices in PA programs.