Transformer-Based Model for the Prediction of Sedentary Behavior Patterns Using Deep Learning models

Abstract

Sedentary behavior continues to be a major health concern, particularly as it correlates with various chronic conditions. While previous studies have focused on utilizing deep learning models, such as stacked LSTMs and CNNs, for predicting sedentary behavior patterns, these approaches face limitations in handling long-range dependencies and providing interpretability in predictions. This research proposes the application of transformer networks, known for their superior ability to capture temporal dependencies through self-attention mechanisms, to predict sedentary behavior more accurately and efficiently. The proposed model builds on previous approaches by integrating enhanced prediction capabilities, reducing error metrics such as MSE, RMSE, and MAPE, and offering improved sensitivity and specificity in classifying sedentary and active periods. Additionally, the attention mechanism offers greater interpretability, enabling the identification of key behavioral patterns and providing actionable insights for health interventions. Experimental results demonstrate an improvement in prediction accuracy, achieving 99.5% accuracy—surpassing previous models—and a 30-40% increase in computational efficiency. The approach is also validated with real-time feedback integration for continuous posture monitoring. This study represents a significant step forward in using deep learning techniques to mitigate sedentary health risks, offering a robust, scalable solution for health monitoring systems in both personal and workplace environments.