A hybrid transformer-graph neural networks framework for enhanced physical activity recognition and sedentary behavior analysis

Abstract

Sedentary behavior has been identified as a major risk factor for chronic diseases such as cardiovascular disorders, obesity, and diabetes. The accurate prediction of sedentary health risks is essential for early intervention and personalized healthcare strategies. This study proposes a novel machine learning-based predictive model that leverages transformer-based architectures and graph neural networks to analyze multidimensional behavioral data. Unlike traditional models, our approach incorporates temporal attention mechanisms to capture long-term dependencies in activity patterns and graph-based learning to model complex relationships between physiological and behavioral factors. The study utilizes real-world datasets, including wearable sensor data and self-reported activity logs, to train and validate the models. Experimental results demonstrate that the proposed framework outperforms conventional machine learning techniques such as random forest and XGBoost, achieving superior predictive accuracy and robustness. The findings highlight the potential of advanced machine learning algorithms in assessing sedentary health risks, enabling proactive health management and intervention strategies.