Wearable sensor models predicted physical activity adherence with 84.6% accuracy and health-risk behaviors with up to 91% F1-score in university students.
Do wearable sensor data predictive models improve the accuracy of predicting physical activity adherence and health risk behaviors compared to traditional survey methods in university students?
Machine learning models utilizing wearable sensor data significantly improve the prediction of physical activity adherence and cardiovascular risk behaviors in university students compared to traditional surveys.
Absolute Event Rate: 0% vs 0%
Abstract Background Prolonged sedentary lifestyles among college students have become a major public health crisis. Over 60% of students fail to meet WHO-recommended physical activity levels, resulting in obesity (13.3% prevalence rate), metabolic syndrome, and increased risks of cardiovascular diseases (CVD) such as mental health issues. Traditional survey methods struggle to capture real-time behavioral patterns, while wearable sensors combined with machine learning algorithms provide a new paradigm for personalized CVD risk monitoring. This technology continuously collects physiological data including heart rate variability, activity intensity, and sleep quality. Through predictive analysis, it identifies early CVD risk markers, enabling the identification of high-risk students before irreversible damage occurs. Purpose Predictive models (Random Forest and LSTM) were developed to classify physical activity adherence (consistent, intermittent, inactive) and identify correlations with health-risk behaviors (sedentary bouts, irregular sleep, Affecting cardiovascular diseases). Model performance was evaluated via 5-fold cross-validation. Methods A total of 312 university students participated in this 12-week study, wearing ActiGraph GT9X tri-axial accelerometers (30Hz sampling) and Polar H10 heart rate monitors to collect multimodal sensor data (accelerometer, gyroscope, heart rate variability). Data preprocessing included noise filtering, normalization, and feature engineering (e.g., step count variance, circadian rhythm metrics). Results The Random Forest model achieved 84.6% accuracy (F1-score=0.82) in predicting activity adherence, with step count variance (22.3%), heart rate recovery rate (18.7%), and sleep quality index (15.4%) as key predictors. LSTM models demonstrated superior performance in temporal pattern recognition, with an F1-score of 0.87 for identifying health-risk behaviors and 0.91 for exercise persistence forecasting. Cluster analysis revealed four distinct risk profiles, with 68% of inactive students exhibiting co-occurring sleep deprivation and poor dietary habits. The framework improved prediction accuracy by 31.4% compared to traditional survey-based methods. Conclusions The "Wearable Sensor Data Predictive Model for University Students' Physical Activity Adherence and Health Risk Behaviors on Cardiovascular Diseases" integrates Random Forest and LSTM models to analyze wearable sensor data, focusing on physical activity adherence and CVD risk behaviors in university students. Key biomarkers include step count variance, heart rate recovery, and sleep quality. Risk clustering enables differentiated interventions, enhancing early CVD risk detection. Implementation could optimize university health programs through proactive outreach and resource allocation, improving long-term student health outcomes and academic performance. Future work will focus on federated learning to strengthen data privacy while maintaining predictive accuracy across diverse student populations, addressing the urgent need for personalized CVD prevention in sedentary young adults.
Luo et al. (Thu,) reported a other. Wearable sensor models predicted physical activity adherence with 84.6% accuracy and health-risk behaviors with up to 91% F1-score in university students.