Inferring Mood Instability on Social Media by Leveraging Ecological Momentary Assessments

Active and passive sensing technologies are providing powerful mechanisms to track, model, and understand a range of health behaviors and well-being states. Despite yielding rich, dense and high fidelity data, current sensing technologies often require highly engineered study designs and persistent participant compliance, making them difficult to scale to large populations and to data acquisition tasks spanning extended time periods. This paper situates social media as a new passive, unobtrusive sensing technology. We propose a semi-supervised machine learning framework to combine small samples of data gathered through active sensing, with large-scale social media data to infer mood instability (MI) in individuals. Starting from a theoretically-grounded measure of MI obtained from mobile ecological momentary assessments (EMAs), we show that our model is able to infer MI in a large population of Twitter users with 96% accuracy and F-1 score. Additionally, we show that, our model predicts self-identifying Twitter users with bipolar and borderline personality disorder to exhibit twice the likelihood of high MI, compared to that in a suitable control. We discuss the implications and the potential for integrating complementary sensing capabilities to address complex research challenges in precision medicine.