Assessment of Scalability and Adaptability in Federated Learning Framework for Robust Driver Stress Monitoring

Driving is an essential yet complex task that demands continuous attention, situational awareness, and the ability to make split-second decisions. External conditions such as traffic congestion, adverse weather, and unpredictable road behavior can contribute to stress, negatively impacting a driver's cognitive functions, reaction time, and overall driving performance. Recently, federated learning has emerged as technique to create and share universal model while maintaining data privacy. This study evaluates the feasibility of Federated Learning (FL) for driver stress detection using physiological signals, offering a privacy-preserving alternative to centralized machine learning models. For this, multi-modal physiological signals are recorded from drivers (N = 20) in varied situations (normal, calling, texting, ADAS on and off) in a controlled environment. Signals, namely ECG, Respiration (Abdomen and Thoracic) and accelerometer, are segmented and applied to the proposed federated learning framework (FL). The proposed framework contains a decentralized server-client model with an improved ResNet model as a base in the local network. Experiments are performed using key model performance metrics, including generalization across unseen subjects and inference efficiency for both subject-dependent and subject-independent learning paradigms. The results indicate that the proposed framework is capable of detecting the stress state. ResNet-based FL models achieve competitive accuracy (ACC = 92.62% for Round count = 10) in subject-independent settings, with improvements as the number of clients and training rounds increases. In addition, FL-based models maintain superior privacy protection and scalability, mitigating concerns associated with sharing raw physiological data. While subject-dependent models show higher initial accuracy, subject-independent FL models yields the highest accuracy (ACC = 97.82%), demonstrating their potential for real-world deployment in vehicular and biomedical stress monitoring applications. For unseen physiological data, subject-independent FL models exhibit strong adaptability, improving accuracy over multiple training rounds and achieving robust generalization across diverse participants. Thus, the proposed framework could be used in real-time and clinical stress monitoring systems.