This project uses computational techniques and AI to study mental illnesses on social media, in particular, suicidal tendencies, depression and eating disorders, which, in severe cases, can also lead to self-harm and/or suicidal ideation.

With the expertise of our team’s psychologists and psychiatrists, we automatically and completely anonymously filter social media posts that include words or phrases often used by patients at high risk of suicide, depression or eating disorders. These posts are labelled based on whether or not they match risk profiles. Subsequently, certain features related to language, image or social media activity are extracted, such as posting frequency or the number of followers. This information is the input for AI algorithms that learn which traits are common to risk groups.

We use computational techniques and AI to study mental illnesses on social media

We use different AI techniques: deep learning for image analysis, clustering algorithms to detect communities, and natural-language processing algorithms for text processing, amongst others.

We analyse texts, images and activity, trying to find factors that specialists usually ask their patients about directly at appointments: whether they have social support, their sleep quality, potential risk factors, recurrent worries, etc. These questions are answered in the social media environment by analysing whether a user has a lot of connections, whether they are active during sleep hours, or whether they comment on or show interest in certain topics.

Our project has revealed differentiating traits between the ‘high risk’ groups for each of the cases studied (depression, suicide and eating disorders). Some of these traits are demographic (age range and most common sex in high-risk users), whilst others are related to the users’ interests. For instance, users with eating disorders often show interest in strict weight-loss diets, vegan diets or high-intensity exercise to lose weight fast.

Exactly. AI, along with the huge computing power we have today, allows us to automate complex processes in a very efficient way, speeding up the search for solutions.

Yes. They are usually women, under the age of 39, who lack social support and refer to substance abuse (drugs, alcohol) or topics related to suicide or self-harm, amongst others.

STOP is an international effort, and it already has seven member institutions: Pompeu Fabra University, the Computer Vision Center (CVC) at the Autonomous University of Barcelona (UAB), Parc Taulí University Hospital in Sabadell, Fundació FITA, Fundació Ajuda i Esperança [Help and Hope Foundation], the University of Lyon and Università della Svizzera Italiana [University of Lugano, USI].

There has been broad participation by UPF students. Undergraduate, master’s and doctoral students have helped advance the research, compiling data and training most of the AI algorithms

UPF plays a fundamental role, as we lead the research from here. Additionally, there has been broad participation in the project by students. Numerous undergraduate, master’s and doctoral students have helped advance the research, compiling data and training most of the AI algorithms. The CVC does image analysis, the University of Lyon and USI contribute to the part related to text analysis and gender bias in anorexia nervosa. Parc Taulí Hospital and Fundació FITA bring expertise in mental health and annotate the data. With Fundació Ajuda i Esperança, we recently began working on a joint social media campaign.

Our research revealed differentiating traits between the ‘high-risk’ groups for each of the cases studied (depression, suicide and eating disorders). We used this information to launch a campaign on Instagram and Facebook targeted at users who fit the high-risk profiles. This campaign, conducted in collaboration with Telèfon de l’Esperança and Telèfon de Prevenció del Suïcidi [Suicide Prevention Hotline], consisted of the publication of ads in feeds (static ads) or stories (dynamic stories) presenting 24-hour hotlines to groups with traits similar to those identified by the AI algorithms in at-risk groups.

The campaign ran for 24 days around the 2020-2021 Christmas holidays, reaching 667,655 people, with the highest number of views on 1 January 2021 (90,993 users). The largest number of views was targeted at the eating disorder group (reaching 545,415 people), achieving peak audience on 25 December 2020 and 1 January 2021 (coinciding with the dates on which most family meals take place). The campaign managed to increase the number of calls from social media to Telèfon de l’Esperança by 60%.

Yes. Actually, research is already being conducted along these lines. Studies have looked at whether social media such as Twitter, Reddit or Instagram can be effectively used to detect problems such as bullying or addictions, as the possibility of having an anonymous profile prompts many users to express certain feelings or perform certain actions that they would not do in another context.

We have worked on two lines in relation to eating disorders.

First, in the context of Diana Ramírez-Cifuentes’s doctoral thesis, we characterized users with anorexia nervosa on Twitter, even differentiating between users in different stages of the disease. For example, we studied the traits of users in the pre-contemplation (characterized by euphoria due to the rapid weight loss and social support received), contemplation (the ravages of the disease begin to become visible and episodes of discomfort, anxiety and even depression can start), treatment and recovery stages. This is the first work to analyse this disorder on social media differentiating between its progressive stages, and it was recently published in the Journal of Medical Internet Research. To study how anorexia is expressed on social media, we analysed a set of more than two million tweets posted between December 2017 and December 2018, written in Spanish. More than 70% of the analysed profiles corresponded to women, and almost 60% to young people under the age of 18.

We saw that anorexia nervosa communities are completely isolated and barely have any links to healthy communities, made up of the accounts of healthy users or institutions or associations that help people with eating disorders. Forcing communication between anorexia nervosa and healthy communities would yield major benefits, since bubbles like that are very harmful for such vulnerable groups. But the recommendation systems on platforms such as Twitter continue to suggest that users with anorexia follow accounts very similar to their own, causing that bubble to expand.

Influencers need to assume the responsibility of taking care with every post, as they may unintentionally be harming some of their followers

Second, in the context of the student Aleix Alonso’s bachelor’s degree final project, we studied how Instagram influencers’ profiles can be harmful to followers with eating disorders. We selected 555 Instagram profiles promoting weight loss, extreme exercise, very low-calorie diets, etc. A total of 85.8% were run by influencers. Of these, 64.1% could potentially cause harm if they were followed by people with a high probability of having an eating disorder. This further supports the need for influencers to assume the responsibility of taking care with every post, as they may unintentionally be harming some of their followers.

Both projects were done with completely anonymous posts and the approval of the UPF Ethics Committee.

Any tool can be used for good or bad, whether a knife, a car or artificial intelligence. Yes, it is possible for someone to use AI to send harmful messages to users; however, that would not only be unethical, but also illegal. In fact, carrying out our project responsibly meant processing the data in a completely anonymous way. In other words, it is impossible to identify any of the accounts used to train the developed algorithms based on the collected data. That is why the advertising campaigns for the hotline could not be targeted at specific users, but anonymous groups of users. And the profile was similar to that found by the AI systems. It was thus impossible to control whether certain ads were shown to users who, despite having traits identified with a risk group, did not belong to it. Therefore, especially the ads shown in the feed included the text ‘Someone you know may need it’, to make sure the campaign was not intrusive and also that it reached the largest number of people.

Carrying out our project responsibly meant processing the data in a completely anonymous way. It is impossible to identify any of the accounts use to train the developed algorithms based on the collected data

Undoubtedly, by regulating the use of AI, which is already being done even on a European scale, and which will need to be fine-tuned and adapted to meet the future challenges that may arise in relation to the use of this technology.

The goal of the crowdfunding campaign, which can be accessed through the project’s website, is to raise funds to expand the joint social media campaigns with Telèfon de l’Esperança and to hire a researcher to explore new issues, such as loneliness, which today, especially in the wake of the pandemic, does not affect only the elderly. The crowdfunding campaign has not been as successful as we had hoped. Fortunately, though, we had the help of UPF’s Planetary Wellbeing Initiative, which allowed us to carry out the first social media campaign, as well as pay the necessary fees to ensure that our publications are freely accessible to anyone wishing to read them.