Refinement of Bio-signal Data from Wearables for Stress Recognition Using Machine Learning and Transparent Artificial Intelligence

This study explores the use of wearable devices for real-time detection of stress and evaluates the impact of meditation audio in alleviating stress following academic activities. It involves the collection of physiological signals specifically Heart Rate Variability (HRV) derived from Interbeat Intervals (IBI), Blood Volume Pulse (BVP), and Electrodermal Activity (EDA) during the Montreal Imaging Stress Task (MIST). To enhance the accuracy of stress classification, the study integrates a Genetic Algorithm with Mutual Information for efficient feature selection by minimizing redundancy. Additionally, Bayesian optimization is employed for fine-tuning the hyperparameters of machine learning models. Experimental results show that combining EDA, BVP, and HRV yields peak classification accuracies of 98.28% for two-level and 97.02% for three-level stress detection using the Gradient Boosting (GB) algorithm. When using only EDA and HRV, the system still performs well, achieving 97.07% and 95.23% accuracy for two- and three-level classifications, respectively. SHAP-based Explainable AI (XAI) analysis further confirms that HRV and EDA are the most influential features in determining stress levels. The study also demonstrates that meditation audio has a measurable calming effect, supporting its potential for stress management. These findings underscore the promise of integrating wearable technologies with machine learning for effective stress monitoring and intervention in academic settings.