Using Federated Learning to Analyze Occupational Stress: Preserving Privacy and Identifying Patterns

Abstract

Occupational stress is a relevant problem in contemporary work environments, affecting employee well-being and organizational performance. Wearable sensing and artificial intelligence can support continuous stress monitoring by analyzing physiological signals, but the centralized collection of these sensitive data raises privacy and security concerns. In this context, this study proposed and evaluated federated learning as a privacy-preserving approach for detecting and monitoring occupational stress. The work combines a rapid literature review with a practical prototype that integrates wearable sensing, a mobile privacy gateway, anonymized communication, authenticated data submission, and local model training. The review discusses the relevance of occupational stress, the role of wearable technologies and artificial intelligence in stress monitoring, and the privacy risks associated with centralized physiological data collection. For the practical implementation, a layered Internet of Medical Things (IoMT) architecture was developed in which ESP32-based wearable devices transmit physiological signals to an owner-authorized mobile application through authenticated Bluetooth Low Energy bonding. The mobile application stores provisioning credentials securely, signs outgoing JSON payloads with HMAC-SHA256, and forwards authenticated requests through the Tor anonymity network to a local training node hosted at a healthcare institution. The stress classification component detects general physiological stress states (amusement, baseline, stress) from the WESAD laboratory protocol, not specific occupational stressors; therefore, the occupational stress framing is the motivating application context, not the validated experimental scope. Within this scope, local model training was validated on the WESAD dataset using a regularized tabular Multilayer Perceptron (MLP), with the physiological data used for classifier training and validation originating exclusively from the WESAD benchmark dataset, not from the prototype sensors; this model achieved a validation macro F1-score of 0.9565 for three-class stress classification. The ESP32-based sensing hardware and mobile privacy gateway were functionally validated for data transmission and authentication. The proposed model was compared with an original MLP baseline and alternative training strategies, using accuracy, precision, recall, macro F1-score, validation loss, and overfitting gap as evaluation criteria.