A step towards quantifying when an algorithm can and cannot predict an individual's wellbeing.

Researchers are exploring the ability to infer complex signals, such a mental wellbeing, from easily collected smartphone behavioral data. Rather than focusing on improving overall accuracy of such an approach, we seek to understand when we are and are not capable of predicting an individual's wellbeing. In particular, we consider the ability to predict daily wellbeing from smartphone GPS location data as a case study. We hypothesize that user characteristics, such as behavioral variability, level of depression symptoms, and amount of labeled data, are related to improvements in prediction accuracy. Our preliminary results indicate that there may be a relationship between an algorithm's ability to successfully predict an individual's wellbeing reports and the individual's location behavior variability. While further work is needed to improve model accuracy and confirm this relationship in a larger study, our work is a step in the necessary direction of understanding which individuals can be monitored with smartphone data.