Generalized Arousal Prediction through Machine Learning

Arousal manifests across the spectrum of human activities, which is a key contributor to human behavior. The right level of arousal is an intriguing part of optimal performance with balancing healthy physical and mental conditions. For this reason, the study of arousal is of immense importance in human-machine interactions and human-human interactions. With the advent of ubiquitous unobtrusive physiological sensing, continuous measurement of arousal in the lab and in the wild became feasible. In this work, we investigate the question if there are underlying universal features of arousal across the problem space that would allow a unifying treatment. We define the problem space along two dimensions: the type of subject activity (cognitive and dexterous) vs. the realism of the activity (controlled and naturalistic). We use datasets from major studies in each of the four cells arising from this decomposition problem. We experiment with machine learning training configurations and feature architectures. Our ultimate goal is to identify the minimum type of training that leads to maximum universal performance for arousal detection. Our effort comes to fill a gap in the literature, which has been treating arousal in the context of specific study paradigms rather than generally.