Development and Validation of a Machine Learning Wrist-worn Step Detection Algorithm with Deployment in the UK Biobank

Abstract

Background: Step count is an intuitive measure of physical activity frequently quantified in a range of health-related studies; however, accurate quantification of step count can be difficult in the free-living environment, with step counting error routinely above 20% in both consumer and research-grade wrist-worn devices. This study aims to describe the development and validation of step count derived from a wrist-worn accelerometer and to assess its association with cardiovascular and all-cause mortality in a large prospective cohort study.

Methods: We developed and externally validated a hybrid step detection model that involves self-supervised machine learning, trained on a new ground truth annotated, free-living step count dataset (OxWalk, n=39, aged 19-81) and tested against other open-source step counting algorithms. This model was applied to ascertain daily step counts from raw wrist-worn accelerometer data of 75,493 UK Biobank participants without a prior history of cardiovascular disease (CVD) or cancer. Cox regression was used to obtain hazard ratios and 95% confidence intervals for the association of daily step count with fatal CVD and all-cause mortality after adjustment for potential confounders.

Findings: The novel step algorithm demonstrated a mean absolute percent error of 12.5% in free-living validation, detecting 98.7% of true steps and substantially outperforming other recent wrist-worn, open-source algorithms. Our data are indicative of an inverse dose-response association, where, for example, taking 6,596 to 8,474 steps per day was associated with a 39% [24-52%] and 27% [16-36%] lower risk of fatal CVD and all-cause mortality, respectively, compared to those taking fewer steps each day.

Interpretation: An accurate measure of step count was ascertained using a machine learning pipeline that demonstrates state-of-the-art accuracy in internal and external validation. The expected associations with CVD and all-cause mortality indicate excellent face validity. This algorithm can be used widely for other studies that have utilised wrist-worn accelerometers and an open-source pipeline is provided to facilitate implementation.

Figure 1:. Schematic of the process for generating step count from 30 seconds of raw triaxial accelerometer data using a hybrid self-supervised learning (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: baseline acceleration threshold model, Centre: Verisense algorithm, and Right: the novel hybrid self-supervised learning model.

Figure 3:. Adjusted estimated marginal mean (95% confidence interval) daily step count according to self-reported overall health status, hospital data derived chronic disease status, and select diagnoses for 75,493 UK Biobank participants.

Mean daily step counts are adjusted for age and sex.

Figure 4:. (Top) Forest plots for all-cause mortality and cardiovascular disease mortality associations with quintiles of daily step count, (Bottom) continuous daily step count for 75,493 UK Biobank participants.

Hazard ratios (HR) and 95% confidence intervals 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% confidence interval of the association of continuously modelled median daily step count. Vertical bars along the step axis indicate distribution of participant daily step counts.