Predicting daily physical activity in a lifestyle intervention program

The growing number of people adopting a sedentary lifestyle these days creates a serious need for effective physical activity promotion programs. Often, these programs monitor activity, provide feedback about activity and offer coaching to increase activity. Some programs rely on a human coach who creates an activity goal that is tailored to the characteristics of a participant. Throughout the program, the coach motivates the participant to reach his personal goal or adapt the goal, if needed. Both the timing and the content of the coaching are important for the coaching. Insights on the near future state on, for instance, behaviour and motivation of a participant can be helpful to realize an effective proactive coaching style that is personalized in terms of timing and content. As a first step towards providing these insights to a coach, this chapter discusses results of a study on predicting daily physical activity level (PAL) data from past data of participants in a lifestyle intervention program. A mobile body-worn activity monitor with a built-in triaxial accelerometer was used to record PAL data of a participant for a period of 13 weeks. Predicting future PAL data for all days in a given period was done by employing autoregressive integrated moving average (ARIMA) models on the PAL data from days in the period before. By using a newly proposed categorized-ARIMA (CARIMA) prediction method, we achieved a large reduction in computation time without a significant loss in prediction accuracy in comparison with traditional ARIMA models. In CARIMA, PAL data are categorized as stationary, trend or seasonal data by assessing their autocorrelation functions. Then, an ARIMA model that is most appropriate to these three categories is automatically selected based on an objective penalty function criterion. The results show that our CARIMA method performs well in terms of PAL prediction accuracy (~9% mean absolute percentage error), model parsimony and robustness