Multi-Sensor Wearable-Based Sleep Stage Classification Using Federated Learning for Enhanced Privacy

Abstract

Sleep is an essential aspect of human health, and Obstructive Sleep Apnea (OSA) is a serious condition that is frequently unidentified because traditional Polysomnography (PSG) is costly and available only to a small number of patients. To address these limitations, this study develops and implements a Federated learning ensemble technique for privacy-preserving, sleep stage analysis with wearable devices. The framework adapts Random Forest classifiers for local client-side training and Logistic Regression for global aggregation of models, with a high level of classification accuracy and user data privacy. A principal approach in this methodology is the fusion of Blood Volume Pulse (BVP), Electrodermal Activity (EDA), heart rate, and triaxial accelerometry data for the assessment of sleep dynamics. The proposed approach highlights the possibility of developing low-cost, privacy-preserving health monitoring systems for multiple applications that often face obstacles such as data heterogeneity and resource limitations. This work provides support in progressing wearable sleep technologies and helps in the creation of effective, reliable applications in the healthcare industry.