JMIR Publications

ABSTRACT

Physical activity (PA) and diet which are key components of “lifestyle” playing a significant role in maintaining a healthy weight status and influence overall health during growth and development. Understanding the interplay between PA and diet is essential, yet the combined influences of habitual PA and dietary patterns on health indicators, such as weight status, remains underexplored in the adolescent population. Few studies use electronic wearable devices to assess PA in conjunction with dietary assessment. Most studies aggregate tracker data into daily or weekly estimates, overlooking valuable information regarding the timing of PA. Meanwhile, combining wearable activity trackers and detailed food questionnaires enables the capture of detailed, time-stamped information, allowing for an exploration of time-based patterns in PA and dietary intake and their impact on weight status. We propose TimePAD (Time-Based Physical Activity and Dietary Intake), a three-stage prediction method that integrates time-based PA features, dietary intake, and sociodemographic data to predict Healthy Weight Status (HWS) categories, namely Healthy Fitness Zone (HFZ) and Need Improvement (NI). Unlike traditional methods, TimePAD utilizes Transformer for Multivariate Time Series and fine-tuned the model with unsupervised learning method to optimize its ability to extract time-based PA features in the first stage. When combined with dietary intake data in the second stage, these features improve HWS predictions in the final stage. This approach allows for identifying the importance of the related factors influencing HWS, thus providing deeper insights into the complex interplay between PA, diet and health outcomes. TimePAD was tested on a dataset of 206 adolescents, with their PA was continuously recorded for 7 days using a wrist activity tracker. The data set also contained a food frequency dietary record as well as participants characteristics. TimePAD outperformed a baseline ARIMA-based approach, achieving accuracy (82.90%) and F1-score (67.92%), showing the value of temporal patterns in PA and dietary behaviors analysis. Notably, it found that light PA (LPA) was consistently ranked as the most important feature in predicting HWS across all experiments. Other key features for HWS prediction included time-based Moderate-to-Vigorous Activity (MVPA), and time-based Sedentary (SED) features, Tanner stage, total weekly sleeping time, age (in months), weekly fruit intake, weekly vegetable and legumes intake, sugar-sweetened beverages intake. The findings suggest that LPA is likely to have significant impact on HWS, calling for further attention and investigation to better understand the role of LPA in overall health outcomes. This illustrates the potential benefits of TimePAD in advancing the understanding of the complex interplay between PA and dietary patterns in shaping healthy behaviors.