Identifying self-disclosed anxiety on Twitter: A natural language processing approach

Background: Text analyses of social media posts are a promising source of mental health information. This study used natural language processing to explore distinct language patterns on Twitter related to self-reported anxiety diagnosis. Methods: A total of 233.000 tweets made by 605 users (300 reporting anxiety diagnosis and 305 not) over six months were comparatively analysed, considering user behavior, Linguistic Inquiry Word Count (LIWC), and sentiment analysis. Twitter users with a self-disclosed diagnosis of anxiety were classified as ‘anxious’ to facilitate group comparisons. Results: Supervised machine learning models showed a high prediction accuracy (Naïve Bayes 81.1 %, Random Forests 79.8 %, and LASSO-regression 79.4 %) in identifying Twitter users’ self-disclosed diagnosis of anxiety. Additionally, a Latent Profile Analysis (LPA) identified four profiles characterized by high sentiment (31 % anxious participants), low sentiment (68 % anxious), self-immersed (80 % anxious), and normative behavior (38 % anxious). Conclusion: The digital footprint of self-disclosed anxiety on Twitter posts presented a high frequency of words conveying either negative sentiment, a low frequency of positive sentiment, a reduced frequency of posting, and lengthier texts. These distinct patterns enabled highly accurate prediction of anxiety diagnosis. On this basis, appropriately resourced, awareness raising, online mental health campaigns are advocated.