Researchers from Dartmouth have created an artificial intelligence model to identify mental disorders based on discussions on Reddit.
A distinctive feature of the new model is the emphasis on emotions, and not on the specific content of the analyzed texts of social networks. In a paper presented at the 20th International Conference on Web Analytics and Intelligent Technologies, the researchers said that this approach works regardless of the topics of the messages discussed.
There are many reasons why people do not seek help for mental disorders: stigmatization, high cost and lack of access to therapists. There is also a tendency to minimize the signs of mental disorders or mix them with stress, says Xiaobo Guo, co-author of the paper. According to him, such digital screening technologies can provide additional motivation for contacting a doctor.
In the study, scientists trained artificial intelligence to detect three types of mental disorders based on messages written by users of social networks. These are common emotional disorders: deep depression, bipolar affective disorder and anxiety disorder – which are characterized by distinct emotional patterns. They examined data from users who reported having one of these disorders, and users without any known mental disorders.
Scientists trained AI by analyzing messages on the social network Reddit for several reasons. Firstly, because it is a network where people can exchange information, and it has many active users (more than 430 million, according to the study) discussing a wide range of topics. Posts and comments are publicly available, and researchers can collect data starting in 2011.
Various emotional disorders have their own characteristic patterns of emotional transitions. By creating an emotional “fingerprint” of the user and comparing it with the established signs of emotional disorders, the model can detect deviations. To confirm their results, they tested it on messages that were not used during training, and showed that the model accurately predicts which users may have one of these disorders and which may not.
The “training” took place in several stages. Researchers are not the first to be interested in analyzing emotions on social networks. So they started by using existing datasets before “feeding” their AI with posts from Reddit. For each category of disorders, they found 1997 users who claimed that they had a mental disorder. They also found 1997 users for the test group who rejected the absence of mental problems. 70% of these users’ publications were used for AI training, 15% – for validation procedures, and 15% – for real model testing. The researchers trained their model to label emotions expressed in users’ messages and display emotional transitions between different messages so that the message could be labeled as “joy”, “anger”, “sadness”, “fear”, “lack of emotions” or a combination of them. The map is a matrix showing how likely it is that the user will move from one state to another, for example, from anger to a neutral state of absence of emotions.
Therefore, the model they developed focuses on transitions, creating an “emotional imprint” associated with the user, which can be compared with “typical” signatures corresponding to emotional disorders. By testing this model on publications that had not previously been used to train AI, the researchers found that AI was able to accurately determine the presence or absence of an emotional personality disorder.
The researchers hope that the presented work can be used for the prevention of mental disorders. In their article, they make strong arguments in favor of a more thoughtful study of models based on the analysis of social network data.