Clustering-based accelerometer measures to model relationships between physical activity and key outcomes

TL;DR

This study evaluates a novel accelerometer data summary derived from unsupervised clustering to model physical activity patterns and their relationship with health outcomes, showing comparable explanatory power to traditional metrics.

Contribution

It introduces a clustering-based summary measure for accelerometer data, highlighting its potential to capture temporal activity patterns in health outcome models.

Findings

Cluster membership explained similar variation as traditional measures.

The new measure captured temporal activity patterns not reflected in traditional metrics.

Challenges in reproducibility and processing of machine learning-derived variables were identified.

Abstract

Accelerometers produce enormous amounts of data. Research that incorporates such data often involves a derived summary metric to describe physical activity. Traditional metrics have often ignored the temporal nature of the data. We build on previous work that applies unsupervised machine learning techniques to describe physical activity patterns over time. Specifically, we evaluate a summary measure of accelerometer data derived from unsupervised clustering in a regression framework through comparisons with other traditional measures: duration of time spent in different activity intensity states, Time Active Mean (TAM), Time Active Variability (TAV), Activity Intensity Mean (AIM), and Activity Intensity Variability (AIV) using data from 268 children participating in the Stanford GOALS trial. The proportion of variation explained by the new measure was comparable to that of traditional measures across regressions of three pre-specified clinical outcomes (waist circumference, fasting insulin levels, and fasting triglyceride levels). For example, cluster membership explained 25%, 11%, and 6% of the variation in waist circumference, fasting insulin levels, and fasting triglyceride levels whereas TAM explained 25%, 10%, and 6% for these same outcomes. Importantly, however, there are challenges when regressing an outcome on a variable derived from unsupervised machine learning techniques, particularly regarding replicability. This includes the processing involved in deriving the variable as well as the machine learning approach itself. While these remain open topics to resolve, our findings demonstrate the promise of a new summary measure that enables addressing questions involving a temporal component that other traditional summary metrics do not reflect.