AN INTERPRETABLE UNSUPERVISED FRAMEWORK FOR EXERCISE PATTERN DISCOVERY USING WEARABLE SENSOR DATA

Abstract

The rapid growth of wearable sensing technologies has enabled continuous monitoring of physiological signals during physical activity, providing new opportunities for data-driven exercise analysis. This study uses the publicly available Wearable Sports Health Monitoring Dataset, consisting of 500 records collected from athletes during various activities, to identify latent exercise patterns based on physiological responses. An unsupervised clustering framework is proposed to categorize athletes into three exercise groups: short-duration high-intensity activity, long-duration low-intensity activity, and moderate activity. A compact and interpretable set of physiological features—heart rate, step count, systolic blood pressure, and blood oxygen saturation—was selected to reduce model complexity while preserving physiological relevance. Experimental results show that the identified clusters exhibit distinct and realistic physiological profiles, with average heart rate and step count differences of up to 18% and 70%, respectively, between high-intensity and long-duration exercise groups. Furthermore, activity status distributions across clusters remain balanced, indicating that grouping is driven by physiological workload rather than predefined activity labels. Overall, the proposed approach provides an interpretable and label-free framework for exercise pattern discovery using wearable data, with potential applications in personalized training design, performance monitoring, and health assessment