Abstract
Background:
Physical activity is a complex modifiable risk factor (RF) for cardiovascular disease (CVD). Current methods to measure physical activity are limited by the use of summary scores such as total metabolic equivalents score (METS).
Hypothesis:
Physical activity patterns derived by a data-driven clustering method are associated with CVD RFs independently of METS.
Methods:
The Northern Manhattan Study is a prospective cohort of older, urban-dwelling, multiethnic, stroke-free individuals. Questionnaires were used to capture multi-dimensions of leisure-time physical activity, which was summarized with METS (total activity minutes х intensity in MET). Participants were grouped into previously defined METS categories (less than 1, greater than 14, and 1-14), and also into clusters by multivariate finite mixture modeling based on activity frequency, duration, energy expenditure, and number of activity types. Bayesian information criterion was used to decide number of clusters. Associations between model-based clusters and 4 RFs (diabetes, hypertension, obesity, high waist circumference) were assessed in the entire cohort and in each METS category; associations between METS and RFs were assessed in each cluster. Chi-squared test was used.
Results:
Physical activity data were available in 3293, with mean age 69 years, 63% women, and 52% Hispanic. Six clusters were identified and labeled I-VI (Table 1). Model-based clusters were associated with all four RFs (all p≤0.01), with clusters V and VI having lower RFs prevalence than the others: the association with obesity prevailed among those with 1≤METS≤14 (p<0.01); and with hypertension among those with METS>14 (p=0.03). METS categories were associated with all four RFs in the entire cohort (all p≤0.04); METS and RFs became no longer significantly associated within clusters.
Conclusions:
A data-driven clustering method for depicting physical activity data is a principled, generalizable approach to form subgroups associated with CVD RFs independently of METS.