Multimodal Wearable-Based Stress Detection Using Classical Machine Learning and Deep Temporal Architectures

Stress is a major public health concern associated with cardiovascular, metabolic, and psychological disorders. Wearable biosensors provide a non-invasive means of continuous stress monitoring; however, robust, generalizable detection remains challenging due to signal variability, temporal dynamics, and inter-subject differences. This study presents a comprehensive framework for automatic stress detection using multimodal physiological signals from the WESAD dataset. We systematically evaluate classical machine learning models based on engineered features alongside modern deep temporal architectures operating directly on raw sequences. The analysis considers multiple signal representations (raw, clean, phasic), temporal window lengths (16–60s), and synchronization frequencies (4 Hz and 64 Hz). Results demonstrate that multimodal fusion of electrodermal activity (EDA), blood volume pulse (BVP), and skin temperature significantly improves classification performance compared to EDA-only modeling. Ensemble-based methods, particularly Random Forest and LGBM, achieved an F1-score of up to 95.65% at 64 Hz, highlighting the importance of synchronized multimodal data and high temporal resolution. Deep temporal models, including TimeMixer and TimesNet, achieved competitive performance using only EDA input (maximum F1-score = 77%), confirming their ability to learn meaningful temporal representations without manual feature engineering.