Anxiety: artificial intelligence can measure risk – 09/04/2023 – Equilibrium

The language we use on social media can become an indication of how our mental health is going – and the machines would be able to find patterns and early signs of conditions such as anxiety and depression.

This is the premise of a work that is in progress at the School of Arts, Sciences and Humanities of the University of São Paulo (EACH-USP).

Over there, a group of researchers is building an algorithm capable of analyzing Twitter profiles and looking for some clues that suggest psychiatric disorders.

The work, which is in the preliminary stages, has already built a database that was named SeptemberBR – a double homage to the Yellow September, a suicide prevention campaign that takes place every year, and to the month in which the project began.

massive comparisons

Computer scientist Ivandré Paraboni, project coordinator, explains that the database gathers information from 3,900 Twitter users who claim to have been diagnosed with depression or anxiety.

The researchers compiled the networks of connections of these profiles and all the text content shared by them on the social network – which totals about 47 million small texts of up to 280 characters.

All of this material was compared to another group of Twitter users chosen at random, who did not demonstrate that they had undergone a medical evaluation on mental health or were undergoing treatment for psychiatric disorders.

“Of course, in the midst of this universe, there may be individuals who lied or omitted this information. But, as the database is large, we estimate that these false positives or false negatives are few”, ponders Paraboni, who is also a researcher associate of the Center for Artificial Intelligence, an engineering institute maintained by the São Paulo Research Foundation (FAPESP) and IBM Brazil.

This veritable library of posts was anonymized – that is, the researchers deleted references to the identity of the users who wrote the texts (such as first names and usernames), for privacy reasons.

The project’s first task was to “dig” this entire database to remove hashtags (keywords preceded by the # symbol), mentions of other accounts (which appear with an @), non-standard characters, and hyperlinks.

Afterwards, the two groups were compared. In addition to evaluating the texts shared on Twitter, the researchers were also able to analyze the users’ contact network, including the accounts that each of them follows.

Paraboni emphasizes the importance of carrying out an initiative like this in Portuguese.

“There are already other studies of this type carried out abroad, but they primarily analyze the content in English”, he explains.

And, of course, there are a number of cultural and linguistic particularities – the standards that apply in these countries can be completely different from what is common in Brazil, among Portuguese speakers.

“Someone needs to develop this computational infrastructure, so that we have access to these tools adapted to Portuguese”, adds the researcher.

The results

The models found some early patterns, which may indicate a propensity for illnesses such as anxiety and depression.

The first is a higher frequency of postings about oneself observed in the group that claimed to have psychiatric disorders – for example, with the use of verbs and pronouns “I”, “me”, “mim” – in the first person .

Another finding was that these individuals resort a lot to emojis and graphic symbols that symbolize the heart.

In addition, topics such as death, crisis and psychology are also more common in these accounts.

To complete, individuals with anxiety or depression tend to follow other pages and users that deal with the subject – for example, patient groups or the profile of a celebrity who announced a recent diagnosis of one of these disorders.

“It’s important to explain that the patterns found by deep learning models can be literally anything”, points out Paraboni.

“The way a person expresses himself on social networks is not necessarily the same as the way he speaks in real life or in the psychiatrist’s office”, he adds.

That is, it is possible that posts on social media such as Twitter reveal traits and characteristics different from those that appear during a formal evaluation with a physician.

“Most of the patterns we found are abstract and there is no explanation for them in psychological theories”, points out Paraboni.

That is, it may even be relatively easy to speculate the reasons that make someone with anxiety or depression talk more about themselves, not least because this is also observed during contact with a health professional.

However, other aspects and behaviors, such as distributing heart symbols or following beads with the same theme, are not things that will come up so easily during the conversation in the office.

the next steps

Now that the first versions of the September BR artificial intelligence model have already been created, the group of specialists from EACH-USP is starting to plan the next steps of the project.

One of the goals is to expand the database that will be evaluated and refine the deep learning techniques, so that the results improve and the analyzes become more accurate.

Questioned by BBC News Brasil if the goal is to make this tool capable of diagnosing cases of anxiety and depression in the future, Paraboni asks for caution.

“This is one of the most dangerous areas when we think about the use of these new technologies”, he ponders.

“No one wants to be misdiagnosed or, on the contrary, see a condition like depression or anxiety go unnoticed.”

“I prefer to see these databases more as a complement, an aid, or a first indication that the person may have a mental health issue.”

The computer scientist anticipates that the work may serve, in a few years, to alert parents when their child is facing a problem.

“Who knows, maybe this could become a tool that analyzes the social networks of children and adolescents and helps to indicate any behavioral issue that deserves attention and evaluation by a health professional?”, he speculates.

The coming together of artificial intelligence, social networks and mental health could not have come at a more opportune time.

First, machine learning has never been talked about as much as it is now, when tools like GPT Chat reach the public and provoke great discussions in society.

Second, the use of social media continues to rise – and Brazil is the third country with the highest number of active users of these platforms in the world, behind only India and Indonesia.

According to a survey published in March by Comscore, Brazilians have 131 million active accounts on social networks and spend 46 hours (or almost two full days) of the month using YouTube, Instagram, Twitter, TikTok and the like.

To top it off, psychiatric disorders are also on the rise. The World Health Organization (WHO) estimates that depression affects 3.8% of the population (or 280 million people).

The entity also points out that these numbers have grown even more in recent years: there has been a 25% increase in the prevalence of anxiety and depression since the beginning of the Covid-19 pandemic.

This text was published here