Calibrating the Classifier: Siamese Neural Network Architecture for End-to-End Arousal Recognition from ECG

Affective analysis of physiological signals enables emotion recognition in mobile wearable devices. In this paper, we present a deep learning framework for arousal recognition from ECG (electrocardiogram) signals. Specifically, we design an end-to-end convolutional and recurrent neural network architecture to (i) extract features from ECG; (ii) analyse time-domain variation patterns; and (iii) non-linearly relate those to the user’s arousal level. The key novelty is our use of a shared-parameter siamese architecture to implement user-specific feature calibration. At each forward and backward pass, we concatenate to the input a user-dependent template that is processed by an identical copy of the network. The siamese architecture makes feature calibration an integral part of the training process, allowing modelling of general dependencies between the user’s ECG at rest and those during emotion elicitation. On leave-one-user-out cross validation, the proposed architecture obtains +21.5% score increase compared to state-of-the-art techniques. Comparison with alternative network architectures demonstrates the effectiveness of the siamese network in achieving user-specific feature calibration.