1) The document proposes a method called "subconscious crowdsourcing" to collect social media data from users with mental disorders in order to build predictive models for mental disorder detection. 2) It extracts linguistic and behavioral features from Twitter data like emotion transitions and social interactions to train models for bipolar disorder and borderline personality disorder classification. 3) Experimental results show the TF-IDF model achieves the best performance in 10-fold cross validation tests and addresses the selection bias problem of prior work by detecting actual disorder sufferers rather than just those discussing mental health topics.