A Privacy-Preserving Health Monitoring Framework Using Federated Learning on Wearable Sensor Data †

Abstract

Health monitoring systems play a crucial role in every life. In the 21st century, advanced technologies like wearable sensors have emerged and make healthcare better overall. These sensors collect massive amounts of data about our health over time in many dimensions. In this paper, our objective is to develop and evaluate a machine learning-based clinical decision support system using wearable sensor data to accurately classify users’ physiological states and activity contexts. The most accurate and effective model is for identifying wearable sensor-based physiological signal classification. However, there are serious privacy and security issues with sending raw sensor data to centralized computers. We gathered the multivariate physiological and activity data from wearable technology, including smartwatches and fitness trackers, which make up the dataset. Physiological signals, including heart rate, resting heart rate, normalized heart rate, entropy of heart rate variability, and caloric expenditure, are all included in the dataset. Lying, sitting, self-paced walking, and running at different MET(Metabolic Equivalent of Task) levels are examples of activity context labels. To secure our data, we proposed an architecture based on federated learning that helps machine learning model training across several dispersed devices without exchanging raw data. In this study, we used eight classifiers, and these are XGBoost, RF, Extra Trees, LightGBM, CatBoost, Bagging, DT, and GB. It has been observed that XGBoost performs well in comparison to the other classifiers with an accuracy of 0.94, a precision of 0.90, a recall of 0.89, an F1-score of 0.90, and an AUC-ROC of 0.98. This study demonstrates the potential of wearable sensor data, combined with machine learning, for accurately classifying activity and physiological conditions. The ML boosting family, especially XGBoost, exhibited strong generalization across diverse signal inputs and activity contexts. These results suggest that explainable, non-invasive wearable analytics can support early detection and monitoring frameworks in personalized healthcare systems. The proposed federated learning framework effectively combines privacy-aware computation and accurate classification using wearable sensor data.