A hybrid deep temporal architecture for multimodal stress prediction

Stress is a major health concern that motivates reliable and objective monitoring using multimodal physiological and behavioral information. This study presents a hybrid framework for continuous stress regression integrating facial video, Blood Volume Pulse (BVP), and Electrodermal Activity (EDA) with Principal Component Analysis (PCA), Bidirectional Long Short-Term Memory (BiLSTM), Elephant Herding Optimization (EHO), and Extreme Gradient Boosting (XGBoost). The UBFC-Phys cohort comprised 56 subjects; 45 subjects formed the development cohort and 11 fixed subjects were reserved for independent testing. To reduce information leakage, the independent test subjects were excluded from PCA fitting, BiLSTM training, EHO hyperparameter selection, and XGBoost fitting. Within the 45-subject development cohort, five-fold GroupKFold cross-validation was performed by subject, with preprocessing and representation learning refitted within each fold. The strict five-fold evaluation yielded MAE 0.2397 ± 0.0478, RMSE 0.3019 ± 0.0629, and R² -0.5031 ± 0.2763. On the fixed 11-subject independent test cohort, the model obtained MAE 0.2458, MSE 0.1046, RMSE 0.3234, and R² -0.6987. Manual and scikit-learn calculations of independent-test R² agreed exactly. These results show that the integrated pipeline is computationally feasible but does not yet generalize better than a mean-prediction baseline to unseen subjects under the strict protocol. The principal contribution is therefore a leakage-controlled multimodal evaluation and an empirical characterization of the remaining cross-subject generalization challenge. AMIGOS is retained only as an exploratory compatibility assessment because equivalent continuous stress targets were not validated.