Improving Mental Health Classifier Generalization with Pre-diagnosis Data

Recent work has shown that classifiers for depression detection often fail to generalize to new datasets. Most NLP models for this task are built on datasets that use textual reports of a depression diagnosis (e.g., statements on social media) to identify diagnosed users; this approach allows for collection of large-scale datasets, but leads to poor generalization to out-of-domain data. Notably, models tend to capture features that typify direct discussion of mental health rather than more subtle indications of depression symptoms. In this paper, we explore the hypothesis that building classifiers using exclusively social media posts from before a user's diagnosis will lead to less reliance on shortcuts and better generalization. We test our classifiers on a dataset that is based on an external survey rather than textual self-reports, and find that using pre-diagnosis data for training yields improved performance with many types of classifiers.