Examining the Role of Mood Patterns in Predicting Self-reported Depressive Symptoms

Researchers have explored automatic screening models as a quick way to identify potential risks of developing depressive symptoms. Most existing models often use a person’s mood as reflected on social media at a single point in time as one of the predictive variables. In this paper, we study changes in mood over a period of one year using a mood profile, which explicitly models the changes of mood, and transitions between moods reflected on social media text. We used a subset of the "MyPersonality" Facebook data set that comprises users who have consented to and completed an assessment of depressive symptoms. The subset consists of 93,378 Facebook posts from 781 users. We observed less evidence of mood fluctuation expressed in social media text from those with low symptom measures compared to others with high symptom scores. Next, we leveraged a daily mood representation in Hidden Markov Models to determine its associations with subsequent self-reported symptoms. We found that individuals who have specific mood patterns are highly likely to have reported high depressive symptoms. However, not all of the high symptoms individuals necessarily displayed this characteristic, which indicates presence of potential subgroups driving these findings. Finally, we leveraged multiple mood representations to characterize levels of depression symptoms with a logistic regression model. Our findings support the claim that derived mood from social media text can be used as a proxy of real-life mood to infer depressive symptoms in the current sample. Combining the mood representations with other proxy signals can potentially advance the current automatic screening technology for research.