In the future, perhaps much closer than we imagine, early detection of symptoms of depression or anxiety disorder may come from social networks. Researchers from the University of São Paulo (USP) are developing an artificial intelligence model capable of analyzing the language used in these media, especially in Brazilian Twitter, to detect possible individuals with a tendency to develop these diseases. All based on behavior, linguistic or otherwise, on the internet.
Mental health disorders have been pointed out by the World Health Organization (WHO) as a growing concern in the world. It is estimated that 11.3% of Brazilians have already been diagnosed with depression.
The project was initiated in the form of a pilot study in September 2019, and began to receive financial support from the Fundação de Amparo à Pesquisa do Estado de São Paulo as of April 2022. The first conclusions of the research “ SeptemberBR: a social media corpus for depression and anxiety disorder prediction ” were published in the magazine “Language Resources and Evaluation”. The name is a tribute to the Yellow September movement – a suicide prevention campaign held annually.
The research team is formed by two doctoral students and two master’s students from EACH/USP, all with training in computing and working in the field of Artificial Intelligence (AI), or more specifically in the field of Natural Language Processing (NLP). . In addition to these, there are two collaborating researchers, one from the NLP area and one from the psychology area.
For the research, 47 million public tweets, from 19,000 Twitter users, were initially analyzed anonymously. Of these, 3,900 were diagnosed with a mental disorder. The Covid pandemic led to a 25% increase in anxiety and depression among collected texts.
Both the individuals’ tweets and their social relationships with friends, followers and interactions with other users count for the project, explains Ivandré Paraboni, professor at the School of Arts, Sciences and Humanities (EACH-USP) and corresponding author of the study.
According to him, the computational models use examples of individuals with diagnosis and other random ones on Twitter (control group) to distinguish linguistic and non-linguistic patterns that are difficult for human interpretation. “However, there are some patterns that are quite evident and easy to recognize, and that even reinforce the findings of the medical field, such as the more pronounced use of first person pronouns (I, me, with me), which are normally accepted as a possible linguistic indicator of depression”, explains Paraboni.
A high incidence was also found among depressive users of the use of the little heart symbol, the emoji of affection, “which perhaps is not yet characterized in psychology”, says Paraboni.
He also observes that, in addition to linguistic patterns, there are also results that demonstrate that the accounts that an individual follows on Twitter (for example, a certain channel about mental health), his/her friends and the users with whom he/she chats on the network may also be indicative of your mental health status.
For the researcher, dealing with textual data from a social network is always quite challenging. “In addition to the more obvious observation that Twitter users do not always write in a conventional way, detecting mental health disorders is complex because we are looking for a normally rare phenomenon that appears very sparsely in an individual’s timelines. And sometimes it doesn’t even show up,” he argues.
Another problem, according to him, is the large volume of data to be processed, both in terms of the number of tweets and the extent of the networks of friends and followers. “In addition, it is important to note that even a random individual, selected as a member of the research control group, may turn out to have a mental health disorder that we are unaware of (and that he himself may not be aware of). This can represent a certain noise for the AI model that is trying to ‘learn’ to distinguish one thing from the other”, he details.
But he considers that, in general, the methods used tend to be robust in the face of these difficulties, which are gradually being overcome.
“These models already have a certain ability to detect depression and anxiety based on an individual’s texts and social relationships, but they need to be improved. The second year of the project will then focus on improving these techniques in order to obtain even more precise results”, says Paraboni.
AI model is not intended to replace medical diagnosis
Paraboni considers that this type of automatic network analysis is not intended to diagnose or even less replace the efforts and knowledge of the medical field for this purpose. “Our project is about an initiative that could be used in the future, for example, to help in the early identification of certain divergent behaviors of the general population, and that could eventually be taken into account by a social network user to decide whether to , for example, whether or not it would be a case of seeking medical advice,” he says.
Another possible application, according to him, would be the creation of some kind of tool that would allow the behavior of users on the social network to be monitored by their parents to signal situations that may require some kind of attention.
Brazil is one of the most connected countries in the world
Brazilians spend 46 hours a month on social networks on average, according to a survey by Comscore, released in March. The country is the third in the world that most consumes social networks, behind only India and Indonesia.
The networks most accessed by Brazilians are YouTube (96.4%), Facebook (85.1%) and Instagram (81.4%). TikTok, Kwai and Twitter follow suit. Twitter, after being bought last year by billionaire Elon Musk, has worried users and experts alike. The site underwent rule changes and started to charge for some services, such as the verification seal.
--