Leveraging Routine Behavior and Contextually-Filtered Features for Depression Detection among College Students

Abstract

The rate of depression in college students is rising, which is known to increase suicide risk, lower academic performance and double the likelihood of dropping out of school. Existing work on finding relationships between passively sensed behavior and depression, as well as detecting depression, mainly derives relevant unimodal features from a single sensor. However, co-occurrence of values in multiple sensors may provide better features, because such features can describe behavior in context. We present a new method to extract contextually filtered features from passively collected, time-series mobile data via association rule mining. After calculating traditional unimodal features from the data, we extract rules that relate unimodal features to each other using association rule mining. We extract rules from each class separately (e.g., depression vs. non-depression). We introduce a new metric to select a subset of rules that distinguish between the two classes. From these rules, which capture the relationship between multiple unimodal features, we automatically extract contextually filtered features. These features are then fed into a traditional machine learning pipeline to detect the class of interest (in our case, depression), defined by whether a student has a high BDI-II score at the end of the semester. The behavior rules generated by our methods are highly interpretable representations of differences between classes. Our best model uses contextually-filtered features to significantly outperform a standard model that uses only unimodal features, by an average of 9.7% across a variety of metrics. We further verified the generalizability of our approach on a second dataset, and achieved very similar results.