Abstract
We present our work towards automatic monitoring of major depressive disorder at the population-level leveraging social media and natural language processing. In this pilot study, we manually annotated Twitter tweets i.e., whether the tweet conveys clinical evidence of depression or not, and if the tweet is depression-related, whether it conveys low mood, fatigue or loss of energy, or problems with social environment. Our classifiers trained with simple features can automatically distinguish between tweets with clinical evidence of depression or not with promising results, suggesting complete automation is possible.