JMIR Publications

ABSTRACT

Background:

The recent emergence of wearable devices has made feasible the passive gathering of intensive, longitudinal data from large groups of individuals. This form of data is effective at capturing physiological changes between subjects (interindividual variability) and changes within subjects over time (intraindividual variability). The emergence of longitudinal datasets provides an opportunity to quantify contribution of such longitude to the control of these sources of variability for applications such as responder analysis, where traditional, sparser sampling methods may hinder the categorization of individuals as consistent responders or nonresponders to an intervention.

Objective:

Here, we aim to quantify the gains made in statistical power and effect size among statistical comparisons when controlling for interindividual variability and/or intraindividual variability compared with controlling for neither.

Methods:

We first identified a model condition using the resting heart rate time series data from 46,350 individuals, collected in 2020 as part of the TemPredict study on COVID-19 detection. We found that weekend nights had larger average heart rate amplitudes than weekday nights (similar in amplitude to those observed during symptomatic COVID-19 infection). Weekends repeat within individuals and COVID-19 often did not, and so weekends suggested themselves as a possible stand in for an intervention with an effect size known to be biologically relevant (as defined by reference to COVID-19 effect size). Using weekends as a surrogate intervention, we assessed the impact of longitudinal sampling in controlling both sources of variability on statistical power for detecting a change of this amplitude. We randomly and iteratively sampled heart rate from weekday and weekend nights, controlling for interindividual variability, intraindividual variability, both, or neither.

Results:

Sampling methods accounting for interindividual variability without artificially increasing statistical power via resampling allowed us to identify individuals with consistently different heart rates between weekends and weekdays. These responsive individuals required 50-fold fewer samples to reach statistical significance than the population assessed by random sampling; at the same time this subpopulation of surrogate intervention responders had an effect size 8-fold larger than the nonresponders. Between-subject variability appeared to be a greater source of structured variability than within-subject fluctuations.

Conclusions:

Thus, leveraging longitudinal, within-subject data may substantially aid in identifying distinct subpopulations for clinical applications such as responder analysis and posthoc latent class analysis.