This study investigated the ability of machine learning models to predict depression and other mental health conditions from Twitter data. Using Tweets from 1006 participants who completed questionnaires assessing symptoms of depression and 8 other mental health conditions, the researchers trained an Elastic Net model on depression severity. The model showed modest predictive performance for depression (R²=0.025), and similar or better performance in predicting other mental health problems, indicating non-specific language patterns associated with depression. The study concluded that machine learning analysis of social media data, when trained on well-validated clinical instruments, cannot make meaningful individualised predictions regarding users' mental health.

Sean W. Kelley, Caoimhe Ní Mhaonaigh, Louise Burke, Robert Whelan, Claire M. Gillan