Personalized Stress Detection from Chest-Worn Sensors by Leveraging Machine Learning

Stress detection is crucial for individual well-being. We investigated stress detection using machine learning (ML) from chest-worn sensor data. We used the WESAD dataset that recorded physiological data via a chest-worn sensor when 15 subjects were in stress and non-stress. We empirically evaluated different combinations of window sizes and ML models for stress detection. Our analysis shows XGBoost achieved the highest accuracy of 99.71%, F1-score 99.57%, and area under the curve (AUC) 99.61% in detecting population-level stress with an optimal window size of 500 ms. We also developed personalized stress detection models using the same window size. Subject-level results show stress can be more accurately detected (accuracy 99.95%, F1-score 99.90%, and AUC 99.87%) with personalized models. Our findings also demonstrate a 2% improvement in stress detection compared to prior studies. This work could prove beneficial for quick stress detection and understanding the cause of stress.