Machine Learning–Driven Prediction and Pattern Analysis of Sedentary Behavior in Adults for Preventive Health Monitoring

Abstract

In terms of its implication on health in adults, sedentary behavior is widely reported as an associated risk factor of chronic diseases such as cardiovascular disease, obesity, metabolic problems and deterioration of mental health. Most conventional monitoring methods depend on self-reporting mechanisms which are often not precise in terms of temporal representation and susceptible to reporting bias. We present a machine learning based framework for the prediction of and modeling of patterns of adult sedentary behavior utilizing the data obtained from wearable sensors on activity and physiological metrics. Detailed preprocessing and feature engineering was applied to extract measures such as sedentary duration, transitions to and from physical activity, posture distribution, heart rate trends and indications of energy consumption. Five prediction models, including Logistic Regression, Random Forest, Support Vector Machine, Gradient Boosting and a deep learning based Long Short-Term Memory network, are trained and compared. To achieve robust and generalization accuracy, a weighted ensemble combination of Random Forest, Support Vector Machine and stacked LSTM predictions are further proposed. Based on stratified cross-validation experimentations, deep learning models outperformed classical classifiers significantly and stacked LSTM offered the best performance in modeling the temporal characteristic. In sum, the proposed ensemble framework demonstrated top performance in prediction and modeling (97.1% in accuracy, 96.5% in F1-score, 0.99 in AUC). The performance was consistent over various time segments throughout the day through temporal modeling and robust against the presence of noise and its impact in real world wearable applications. It is concluded that combination models offer highly accurate methods for identifying periods of prolonged sedentary behavior and for prediction of future risks.