Processing UK Biobank High Resolution Accelerometry Data for Unsupervised Identification of Activity Profiles and...

Accelerometer data obtained with wearable devices over extended periods of time provides objective, valuable information on activity behavior. Building on previous work to derive easy-to-interpret activity parameters - the Activity Types from Long-term Accelerometric Sensor data (ATLAS) index - from such data, we aim to investigate whether this approach is feasible with high-quality, extensive data from the UK Biobank, for identifying activity behavior groups, and if exemplary, clinically relevant parameters differ between these groups. A sample of 6,400 subjects' raw accelerometer data was chosen to be processed for computation of the ATLAS index parameters 'regularity', 'intensity' and 'duration' of moderate-intensity, 15+-minute physical activity events. Subsequently, hierarchical clustering was applied, and differences in HDL cholesterol, BMI and C-Reactive Protein (CRP) lab data levels were evaluated. Clustering yielded five distinct activity clusters, and statistically significant differences in HDL cholesterol, BMI and CRP were found between several clusters. The use of the ATLAS index parameters allows for physical activity group identification from objective accelerometer data. These groups differ in physiologically relevant outcome parameters. More research is necessary to uncover potential causal relationships, e.g., by using causal inference methods.