Predicting age from 100,000 one week-long 100Hz wrist accelerometer records of physical activity

Abstract

AbstractPhysical activity improves quality of life, physical health and mental health, and is also an important protective factor against highly prevalent age-related diseases such as cardiovascular diseases, diabetes, cancer and mental health. With age, physical activity tends to decrease, leading down a vicious cycle that increases vulnerability to disease in the elderly. In the following, we trained neural network architectures to predict age from 115,456 one week-long 100Hz wrist accelerometer recordings from the UK Biobank (R-Squared=63.5±2.4%; root mean squared error=4.7±0.1 years). We achieved this performance by preprocessing the raw data as 2,271 scalar features, 113 time series and four images. We also considered the raw signal at different time scales (weekly activity patterns vs. gait). We then defined accelerated aging for a participant as being predicted older than one’s actual age and aimed to characterize these participants. We performed a genome wide association on the accelerated aging phenotypes to estimate its heritability (h_g2=12.3±0.9%) and identified nine single nucleotide polymorphisms in seven genes associated with it (e.g HIST1H4L, involved in chromatin organization). Similarly, we identified biomarkers (e.g blood pressure), clinical phenotypes (e.g chest pain), diseases (e.g hypertension), environmental (e.g smoking) and socioeconomic (e.g income and education) variables associated with accelerated aging. We conclude that physical activity-derived biological age is a complex phenotype with both genetic and non-genetic factors.