JMIR Publications

ABSTRACT

Background:

Wearable accelerometers can record minute-level physical activity (PA) repeatedly over time. Longitudinal change is usually summarized as a change in daily activity volume, which discards when during the day activity is gained or lost. A representation of within-person change in the full daily activity curve could provide more informative measures of functional aging.

Objective:

This study aimed to characterize within-person changes in the timing and magnitude of daily PA in older women and to examine whether features of these changes were associated with physical function (PF).

Methods:

We analyzed hip-worn accelerometry from the Objective Physical Activity and Cardiovascular Health (OPACH) baseline assessment and two Women’s Health Initiative Strong and Healthy (WHISH) follow-up assessments in 1,730 women with at least two consecutive valid assessments. Change between smoothed diurnal PA curves (6:00 a.m. to midnight) was modeled as a Riemannian deformation, whose initial momenta represent timing- and magnitude-related changes. Multivariate functional principal component analysis (MFPCA) summarized these changes as participant-level principal component (PC) scores. Linear mixed-effects models related PC scores and deformation energy, a nondirectional measure of pattern change, to RAND-36 PF, adjusting for baseline PF and covariates. A second model replaced PC scores with the change in net area under the activity curve (Δnet-AUC ), a volume summary.

Results:

On average, activity declined across the day from baseline to WHISH wave 1 (baseline–W1) and from WHISH wave 1 to wave 2 (W1–W2), with morning activity shifted later, and activity after about 10:20 a.m. shifted earlier. Fifteen PCs captured at least 90% of total variance in each transition. Of the variance captured by these 15 PCs, the first PC (PC1) accounted for 22.4% in baseline–W1 and 20.8% in W1–W2. Higher PC1 scores reflected residual upward shifts in activity, especially between 10 a.m. and 7 p.m., and was highly correlated with Δnet-AUC (r=0.9). Among 1,157 participants with complete data, PC1 scores were positively associated with PF scores after adjustment for baseline PF and covariates (P<.001). An interquartile increase in PC1 corresponded to a 1.3-point higher PF score (SE = 0.33). Deformation energy was more strongly associated with PF in W1–W2 than in baseline–W1 (interaction β = 387.4, SE = 129.84, P=.003), and this interaction persisted in the model with Δnet-AUC (P=.005).

Conclusions:

Riemannian deformation analysis of repeated wearable accelerometry characterized changes in when and how much older women were active. The dominant mode of change largely tracked activity volume and was associated with PF. Jointly modeling changes in activity timing and magnitude provides a framework for studying functional aging using longitudinal wearable data.