They trained their model to label the emotions expressed in users' posts and map the emotional transitions between different posts, so a post could be labeled "joy," "anger," "sadness," "fear," "no emotion," or a combination of these. The map is a matrix that would show how likely it was that a user went from any one state to another, such as from anger to a neutral state of no emotion.
Different emotional disorders have their own signature patterns of emotional transitions. By creating an emotional "fingerprint" for a user and comparing it to established signatures of emotional disorders, the model can detect them. To validate their results, they tested it on posts that were not used during training and show that the model accurately predicts which users may or may not have one of these disorders.