Physical activity monitoring using wearable devices based on machine learning algorithms.

Wearable accelerometers provide an important source for continuous physical activity monitoring, but high-frequency signals often suffer from missing labels, temporal discontinuity, and inter-individual variability, limiting their direct use in behavioral health analysis. This study proposes an integrated framework for wrist-worn tri-axial accelerometer data. The dataset includes 100 participants under free-living conditions, each monitored for approximately 24-27 h at 100 Hz. The framework consists of three stages: (1) data quality processing with label continuity reconstruction, (2) metabolic equivalent (MET) estimation using an XGBoost regression model based on time- and frequency-domain features, and (3) behavioral interpretation using the predicted MET sequence to derive sleep-stage boundary segmentation and sedentary-event alerts. XGBoost achieved the best performance among the evaluated models (MAPE = 0.2895, MSE = 0.6032). The method effectively captures temporal activity patterns, enabling sleep-stage transition identification and sedentary-event detection (MET < 1.6 for 30 min). The proposed framework provides an interpretable mapping from accelerometer signals to behavioral-health indicators, offering a practical approach for long-term wearable-based monitoring and personalized health management.