Investigating the comparability of wearable accelerometer methods in the association between physical activity and cardiovascular disease: a cohort study using UK Biobank

The selection of accelerometer processing methods may influence the shape of the dose-response association between wearable-measured physical activity and health outcomes. We aimed to compare the association of stroke and myocardial infarction with Moderate-Vigorous Physical Activity (MVPA) assessed by three accelerometer-generated metrics: Low-pass Filtered Euclidean Norm Minus One (LFENMO), machine-learning, and activity counts. We computed MVPA durations in the UK Biobank accelerometer sub-cohort recruited between 2013 and 2015 in the UK. The outcomes, incident stroke and myocardial infarction, were followed up until December 2022. We used Cox regression and a restricted cubic spline to estimate the dose-response association for each of the three MVPA metrics. There were 90,237 cardiovascular disease-free participants at baseline. We observed 1298 incident strokes and 2031 myocardial infarctions. For stroke, a linear decrease in hazard ratio was observed with machine-learning, but not with LFENMO and activity counts. For myocardial infarction, machine-learning and LFENMO showed a curvilinear decrease in hazard ratios, whereas activity counts showed a linear decrease. The dose-response associations between MVPA and cardiovascular disease varied markedly across the three accelerometer-derived MVPA metrics. Research using a single accelerometer metric may caution about the interpretation of the association. • Researchers use several methods to convert wearable data into activity measures. • We evaluated the associations between cardiovascular disease and these methods. • Machine learning-derived activity showed a distinct dose-response association. • Future research benefits from the use of multiple measurement methods.