Subject-independent framework for protocol-defined stress task classification using multimodal physiological signals

Prolonged stress can cause cardiovascular disease, depression, and anxiety. Reliable detection is therefore important. Existing studies using multimodal physiological methods have already reported accuracy, up to 99.7% however, two important aspects remain underexplored. First, most studies separate stress only from a neutral baseline, amusement, and different workload levels. Second, subject-independent performance remains limited because most studies evaluate their models using subject-dependent data splits that are prone to data leakage. This proposed study tries to address these gaps. Protocol-defined stress tasks are detected from everyday activities, not from any workload levels. The proposed model, a one-dimensional convolutional neural network (1D-CNN), is evaluated subject independently using leave-one-subject-out (LOSO) cross-validation. This proposed study was conducted on the GalaxyPPG dataset. The dataset includes ECG, PPG, and BVP signals. This enables multimodal fusion for stress detection during everyday activities. Single modalities achieve mean test accuracies of 62.19% for PPG, 67.74% for BVP, and 47.81% for ECG on LOSO. Fusion (ECG+PPG+BVP) achieves the mean test accuracy on LOSO evaluation, 76.47% }14.60%. To the best of the author's knowledge, this modality's combination (ECG+PPG+BVP) has not been explored in the literature for GalaxyPPG dataset.