Stress is a major health concern, contributing to cardiovascular and mental health issues, as well as reduced productivity. Wearable technologies enable continuous monitoring of vital signs—such as heart rate (HR), respiratory rate (RR), and heart rate variability (HRV)—supporting real-time stress detection. Although machine learning (ML) methods have shown strong predictive capabilities, interpretability remains essential for clinical and personal health use. This study integrates SHapley Additive exPlanations (SHAP) with an XGBoost classifier to enhance transparency in stress prediction from physiological signals. Using data from three wearable datasets, SHAP identified HR, RR intervals, and HRV metrics—particularly the LF/HF ratio—as key predictors. The method also revealed the impact of temporal and individual differences on predictions. These findings highlight the potential of SHAP to deliver both accurate and interpretable stress monitoring, advancing trustworthy AI integration in wearable health systems.