Self-Supervised Machine Learning to Characterize Step Counts from Wrist-Worn Accelerometers in the UK Biobank

Abstract

Purpose: Step count is an intuitive measure of physical activity frequently quantified in health-related studies; however, accurate step counting is difficult in the free-living environment, with error routinely above 20% in wrist-worn devices against camera-annotated ground truth. This study aimed to describe the development and validation of step count derived from a wrist-worn accelerometer and assess its association with cardiovascular and all-cause mortality in a large prospective cohort.

Methods: We developed and externally validated a self-supervised machine learning step detection model, trained on an open-source and step-annotated free-living dataset. Thirty-nine individuals will free-living ground-truth annotated step counts were used for model development. An open-source dataset with 30 individuals was used for external validation. Epidemiological analysis was performed using 75,263 UK Biobank participants without prevalent cardiovascular disease (CVD) or cancer. Cox regression was used to test the association of daily step count with fatal CVD and all-cause mortality after adjustment for potential confounders.

Results: The algorithm substantially outperformed reference models (free-living mean absolute percent error of 12.5% vs 65%-231%). Our data indicate an inverse dose-response association, where taking 6430-8277 daily steps was associated with 37% (25%-48%) and 28% (20%-35%) lower risk of fatal CVD and all-cause mortality up to 7 yr later, compared with those taking fewer steps each day.

Conclusions: We have developed an open and transparent method that markedly improves the measurement of steps in large-scale wrist-worn accelerometer datasets. The application of this method demonstrated expected associations with CVD and all-cause mortality, indicating excellent face validity. This reinforces public health messaging for increasing physical activity and can help lay the groundwork for the inclusion of target step counts in future public health guidelines.

FIGURE 1

Schematic of the process for generating step count from 30 s of raw triaxial accelerometer data using a hybrid SSL and peak detection step counting model.

FIGURE 2

Bland–Altman plots with dotted 95% limits of agreement for the comparison of step counting models in the (top) OxWalk free-living dataset of 39 adults and (bottom) Clemson laboratory-based dataset of 30 young adults. Left: Ducharme baseline acceleration threshold model (8), centre: Verisense algorithm (25), and right: the novel hybrid SSL model. The Clemson dataset serves as an internal validation for the Verisense algorithm and an external validation for the hybrid SSL model. The OxWalk dataset serves as an external validation for the Verisense algorithm and an internal validation for the Hybrid SSL model. Both datasets serve for external validation of the Ducharme algorithm.

FIGURE 3

Adjusted estimated marginal mean (95% CI) daily step count according to self-reported overall health status, hospital data–derived chronic disease status, and select diagnoses for 75,263 UK Biobank participants. Mean daily step counts are adjusted for age and sex.

FIGURE 4

Top: Forest plots for all-cause mortality and CVD mortality associations with quintiles of daily step count. Bottom: continuous daily step count for 75,263 UK Biobank participants. Hazard ratios (HR) and 95% CI were calculated using age as a timescale, adjusted for sex, ethnicity, education, alcohol intake, smoking status, Townsend deprivation index, processed meat intake, fresh fruit intake, oily fish intake, and added salt intake. HR is above and number of events is plotted below each data point. Spline plot of hazard ratio and 95% CI of the association of continuously modeled median daily step count. Vertical bars along the step axis indicate distribution of participant daily step counts.