Previous research found a strong relation between the users’ psychological disorders and their language use in social media posts in terms of vocabulary selection, emotional expressions, and psychometric attributes. However, although studying the association between psychological disorders and musical preference is considered as rather an old tradition in the clinical analysis of health data, it is not explored through the lens of social media analytics. In this study, we investigate which attributes of the music posted on social media are associated with mental health conditions of Twitter users. We created a large-scale dataset of 1519 Twitter users with six self-reported psychological disorders (depression, bipolar, anxiety, panic, post-traumatic stress disorder, and borderline) and matched with 2480 control users. We then conduct an observational study to investigate the relationship between the users’ psychological disorders and their musical preference by analyzing lyrics of the music tracks that the users shared on Twitter from multiple dimensions including word usage, linguistic style, sentiment and emotion patterns, topical interests and underlying semantics. Our findings reveal descriptive differences on the linguistic and semantic features of music tracks of affected users compared to control individuals and among users from different psychological disorders. Additionally, we build a feature-based and an (explainable) deep learning-based binary classifiers trained on disorder and control users and demonstrate that lyrics of the music tracks of users on Twitter can be considered as complementary information to their published posts to improve the accuracy of the disorder detection task. Overall, we find that the music attributes of users on Twitter allow inferences about their mental health status. © 2023 Elsevier Ltd