Cross Domain Features for Subject-Independent Stress Detection

Stress detection will be useful for addressing physical and mental health being measures. Electrocardiogram (ECG) and Electrodermal activity (EDA) are two popular physiological modalities used to develop different machine learning and deep learning based models. To learn and capture the commonalities jointly, we propose a modality invariant (MI) domain. The traditional way of capturing characteristic features from individual modalities is investigated as modality specific (MS) domain. Deep Models are trained using these domains, and decision-level fusion is used to predict the label of the given test subject. To investigate the generalization capabilities, we applied the proposed framework on two benchmark datasets – ASCER-TAIN and CLAS. We observed that the MI and MS channels complement each other in both datasets. Results also showed that the proposed framework outperformed the state-of-the-art performance by 11-13% (absolute).