Deep Multimodal Fusion for Subject-Independent Stress Detection

This paper explores the influence of convolutional layer in deep multimodal fusion (intermediate fusion) for the detection of subject-independent stress using physiological signals including Electrocardiogram (ECG) and Electrodermal Activity (EDA). We compare performance of deep multimodal fusion models based on Convolutional Neural Network (CNN) by combining modals at different levels of the network. On two benchmark datasets ASCERTAIN and CLAS, the proposed approach is validated and the results shown to be higher for intermediate fusion on convolutional layer.