An Inter-domain Study for Arousal Recognition from Physiological Signals

Abstract

Arousal recognition from physiological signals is a task with many challenge remaining, especially when performed in several different domains. However, the need for emotional intelligent machines increases day by day, starting with timely detection and improved management of mental disorders in mobile health, all the way to enhancing user experience in human-computer interaction (HCI). One of the open research questions, which we analyze in this paper, is which machine-learning (ML) methods and which input is most suitable for arousal recognition. We present an inter-domain study for arousal recognition on six different datasets. The datasets are processed and translated into a common spectro-temporal space of R-R intervals and Galvanic Skin Response (GSR) data, from which features are extracted and fed into ML algorithms. We present a comparison between dataset-specific models, “flat” models build on the overall data, and a novel stacking scheme, developed to utilize knowledge from all six datasets. When one model is built for each dataset, it turns out that whether the R-R, GSR, or merged features yield the best results is domain (dataset) dependent. When all datasets are merged into one and used to train and evaluate the models, the stacking scheme improved upon the results of the “flat” models.