FEDERATED LEARNING FOR STRESS DETECTION USING DISTRIBUTED WEARABLE DATA IN PRIVACY PRESERVING HEALTHCARE SYSTEMS

Abstract

Nowadays, comprehensive, continuous, and reliable monitoring is required in healthcare systems to account for stress-related impacts not only on physical but also on mental health, such as cardiovascular diseases, anxiety, and impaired cognitive function. This research aims to propose a federated learning approach for stress detection while considering the data privacy of the sensor network. FL-SDNet is proposed as a fusion approach combining a CNN and a BiLSTM for feature extraction, integrating spatial and temporal features from multi-modal physiological signals, including heart rate (HR), electrodermal activity (EDA), temperature (TEMP), and motion (MOTION). The proposed system uses a cross-modal attention fusion mechanism that dynamically balances multiple attention signals to boost performance in scenarios with diverse data distributions and introduces an adaptive federated aggregation (AFA) strategy that considers client reliability and data quality. It is fully decentralized and security-preserving, with raw data stored on local devices and only model updates sent and received in encrypted form. The results of extensive experiments with non-IID federated data show that the proposed model achieves 90.2% accuracy, 0.88 F1 score, and 0.92 ROC-AUC in a non-IID federated setting, results comparable to those of the centralized deep learning model. Moreover, it can operate in resource-limited applications (such as wearables) by reducing communication overhead by approximately 40%. A comparative study is conducted between traditional machine learning and deep learning using standard federated learning methods, indicating that the proposed method is effective. An ablation study is conducted to verify the effectiveness of the new component added to the experiment. Overall, the results of this study show that federated learning can achieve high accuracy for stress detection while ensuring data privacy, supporting the potential to exclude local data from global collection while maintaining high accuracy in real-world healthcare applications. This work serves as a valuable asset for the development of next-generation personalized real-time stress-monitoring systems for 'privacy-preserving' healthcare, remote patient monitoring, and intelligent wearables.