A novel multi-kernel 1D convolutional neural network for stress recognition from ECG

Stress is an emotional state which although experienced in a subjective way, it shares specific common characteristics. Objective stress recognition has proven to be a complicated issue, due to the number of parameters involved. Thus, the investigation of reliable indices associated with the stress response is of utmost importance. Heart activity may provide useful information towards this goal. Traditional machine learning techniques have been used in the area of emotion recognition but they sometimes present specific limitations. The emergence of Deep Learning (DL) techniques permits the reveal underlying patterns in electrocardiography (ECG) which, otherwise, would not be easily observed. The proposed DL architecture utilizes a variety of kernels per module to compute complex feature maps and enables a multi-level modelling of the unique heart rate variability signature for stress state identification. The proposed methodology using 6-fold cross-validation outperforms single kernel networks achieving classification accuracy up to 99.1%, better overall performance (avg. F1-score 88.1%, avg. accuracy 89.8%) and more consistent behaviour across study's experimental phases.