Abstract

Purpose: A data mining approach was applied to establish a multilevel hierarchy predicting physical activity (PA) behavior, and to methodologically identify the correlates of PA behavior.

Methods: Cross-sectional data from the population-based Northern Finland Birth Cohort 1966 study, collected in the most recent follow-up at age 46, were used to create a hierarchy using the chi-square automatic interaction detection (CHAID) decision tree technique for predicting PA behavior. PA behavior is defined as active or inactive based on machine-learned activity profiles, which were previously created through a multidimensional (clustering) approach on continuous accelerometer-measured activity intensities in one week. The input variables (predictors) used for decision tree fitting consisted of individual, demographical, psychological, behavioral, environmental, and physical factors. Using generalized linear mixed models, we also analyzed how factors emerging from the model were associated with three PA metrics, including daily time (minutes per day) in sedentary (SED), light PA (LPA), and moderate-to-vigorous PA (MVPA), to assure the relative importance of methodologically identified factors.

Results: Of the 4582 participants with valid accelerometer data at the latest follow-up, 2701 and 1881 had active and inactive profiles, respectively. We used a total of 168 factors as input variables to classify these two PA behaviors. Out of these 168 factors, the decision tree selected 36 factors of different domains from which 54 subgroups of participants were formed. The emerging factors from the model explained minutes per day in SED, LPA, and/or MVPA, including body fat percentage (SED: B = 26.5, LPA: B = − 16.1, and MVPA: B = − 11.7), normalized heart rate recovery 60 s after exercise (SED: B = -16.1, LPA: B = 9.9, and MVPA: B = 9.6), average weekday total sitting time (SED: B = 34.1, LPA: B = -25.3, and MVPA: B = -5.8), and extravagance score (SED: B = 6.3 and LPA: B = − 3.7).

Conclusions: Using data mining, we established a data-driven model composed of 36 different factors of relative importance from empirical data. This model may be used to identify subgroups for multilevel intervention allocation and design. Additionally, this study methodologically discovered an extensive set of factors that can be a basis for additional hypothesis testing in PA correlates research.