Sequential Activity Patterns and Outcome-Specific, Real-Time, and Target Group-Specific Feedback: The SPORT Algorithm

Purpose: Physical activity (PA) is crucial for health, but there is insufficient evidence about PA patterns and their operationalization. The authors developed two algorithms (SPORTconstant and SPORTlinear) to quantify PA patterns and check whether pattern information yields additional explained variance (compared with a compositional data approach [CoDA]). Methods: To measure PA, 397 (218 females) adolescents with a mean age of 12.4 (SD = 0.6) years wore an ActiGraph on their lower back for 1 week. The SPORT algorithms are based on a running value, each day starting with 0 and minutely adapting depending on the behavior being performed. The authors used linear regression models with a behavior-dependent constant (SPORTconstant) and a function of time-in-bout (SPORTlinear) as predictors and body mass index z scores (BMIz) and fat mass percentages (%FM) as exemplary outcomes. For generalizability, the models were validated using five-fold cross-validation where data were split up in five groups, and each of them was a test data set in one of five iterations. Results: The CoDA and the SPORTconstant models explained low variance in BMIz (2% and 1%) and low to moderate variance in %FM (both 5%). The variance being explained by the SPORTlinear models was 6% (BMIz) and 9% (%FM), which was significantly more than the CoDA models (p < .001) according to likelihood ratio tests. Conclusion: Among this group of adolescents, SPORTlinear explained more variance of BMIz and %FM than CoDA. These results suggest a way to enable research about PA patterns. Future research should apply the SPORTlinear algorithm in other target groups and with other health outcomes.