Abstract
The purpose of this study was to develop a model capable of predicting stress levels and interpreting the underlying physiological patterns using large-scale, real-life biosignal data. To achieve this, we utilized approximately 137,000 longitudinal measurements voluntarily collected from residents of Sejong Special Self-Governing City over a two-year period (February 2023-December 2024). Based on these data, we constructed a stress prediction framework that integrates both static machine-learning models-such as Random Forest and LightGBM-and time-series deep learning models, including LSTM and Transformer architectures. Model interpretability was further enhanced through SHapley Additive exPlanations (SHAP), which quantified the contribution of key biomarkers, and through visualization of Transformer attention weights to reveal temporal interactions within the biosignal sequences. The central objective of this study was to evaluate how accurately a deep learning model can learn and reproduce stress indices generated by existing heart rate variability (HRV)-based algorithms embedded in K-FDA-approved wearable devices. Accordingly, the ground truth used in this work reflects algorithmic outputs rather than clinically validated assessments such as salivary cortisol or psychological scales. Thus, rather than identifying independent clinical stress markers, the present work focuses on determining whether a Transformer-based model can effectively approximate device-derived physiological stress levels over time, thereby providing a methodological foundation for future applications using clinically validated stress labels. Experimental results demonstrated that the Transformer model achieved approximately 98% classification accuracy across this large dataset, indicating that it successfully captures short-term biosignal fluctuations as well as long-term temporal structure. These findings collectively demonstrate the engineering feasibility of developing a large-scale, wearable-based stress monitoring system.