Personalized versus Generic Mood Prediction Models in Bipolar Disorder

Abstract

A number of studies have been investigating the use of mobile phone sensing to predict mood in unipolar (depression) and bipolar disorder. However, most of these studies included a small number of people making it difficult to understand the feasibility of this method in practice. This paper reports on mood prediction from a large (N=129) sample of bipolar disorder patients. We achieved prediction accuracies of 89% and 58% in personalized and generic models respectively. Moreover, we shed light on the "cold-start" problem in practice and we show that the accuracy depends on the labeling strategy of euthymic states. The paper discusses the results, the difference between personalized and generic models, and the use of mobile phones in mental health treatment in practice.