Abstract
With wearable, relatively unobtrusive health monitors and smartphone sensors, it is increasingly easy to collect continuously streaming physiological data in a passive mode without placing much burden on participants. At the same time, smartphones provide the ability to survey participants to provide "ground-truth" reporting on psychological states, although this comes at an increased cost in participant burden. In this paper, we examined how analytical approaches from the field of machine learning could allow us to distill the collected physiological data into actionable decision rules about each individual's psychological state, with the eventual goal of identifying important psychological states (e.g., risk moments) without the need for ongoing burdensome active assessment (e.g., self-report). As a first step towards this goal, we compared two methods: (1) a k-nearest neighbor classifier that uses dynamic time warping distance, and (2) a random forests classifier to predict low and high states of affective arousal states based on features extracted using the tsfresh python package. Then, we compared random-forest-based predictive models tailored for the individual with individual-general models. Results showed that the individual-specific model outperformed the general one. Our results support the feasibility of using passively collected wearable data to predict psychological states, suggesting that by relying on both types of data, the active collection can be reduced or eliminated.