Predicting healthy weight status from physical activity and dietary intake: A time-aware data mining pipeline

Area of Science:

• Adolescent health
• Data mining
• Wearable technology

Background:

• Predicting healthy weight status (HWS) is crucial for adolescent well-being.
• Integrating physical activity (PA), dietary intake, and sociodemographic data offers a holistic approach.
• Time-aware analysis of PA patterns can provide deeper insights than static measures.

Purpose of the Study:

• To develop and evaluate TimePAD, a time-aware data mining pipeline for HWS prediction.
• To identify key features influencing HWS in adolescents.
• To integrate accelerometry-based PA, dietary data, and sociodemographic factors.

Main Methods:

• TimePAD utilizes a three-stage pipeline: PA embedding, data integration, and prediction.
• Stage 1 employs a Transformer encoder with Self-Supervised Learning for time-of-day PA embeddings.
• Stage 2 combines PA embeddings with dietary (FFQ) and sociodemographic data; Stage 3 trains models and ranks predictors.

Main Results:

• TimePAD achieved 82.90% accuracy and 67.92% F1-score for HWS prediction, outperforming a baseline.
• Time-of-day light physical activity (LPA) was a consistently important predictor.
• Other significant predictors included Moderate-to-Vigorous Physical Activity (MVPA), sedentary time, Tanner stage, sleep, age, and specific dietary intakes.

Conclusions:

• TimePAD effectively models the temporal structure of PA and integrates it with other data for HWS prediction.
• Findings highlight the significant association of LPA with HWS, warranting further research.
• The pipeline demonstrates potential for understanding and shaping healthy behaviors by contextualizing PA with dietary intake.