Characterization of Anorexia Nervosa on Social Media: Textual, Visual, Relational, Behavioral, and Demographical Analysis (Preprint)

Abstract

BACKGROUND Eating disorders are psychological conditions characterized by unhealthy eating habits. Anorexia Nervosa (AN) is defined by the thought of being overweight despite being dangerously underweight. Psychological signs involve emotional and behavioral issues. There is evidence that signs and symptoms can be manifested on social media, where both harmful and beneficial content is shared daily. OBJECTIVE The aim of this work is to characterize Spanish speaking users with Anorexia signs on Twitter through the extraction and inference of behavioral, demographical, relational, and multi-modal data. This analysis is focused on characterizing and comparing users at different stages of the process to overcome the illness, including treatment and full recovery periods considering the Transtheoretical Model of Health Behavior Change (TTM). METHODS We analyze tweets published by users going through different stages of Anorexia. Users are characterized through their writings, posting patterns, relations, and images. We analyze the differences among users going through each stage of the illness and control users (users not suffering from AN). We also analyze the topics of interest of their followees (users followed by them). We perform a clustering approach to distinguish users at an early phase of the illness (precontemplation) from users that recognize that their behavior is problematic (contemplation); and generate models dedicated to the detection of tweets and images related to AN. We consider two types of control users: focused control users that use terms related to anorexia; and random control users. RESULTS We found significant differences between users at each stage of the recovery process (P<.001) and control groups. Users with AN tend to tweet more at night, with a median sleep period tweeting ratio of 0.05 in comparison to random control users (0.04) and focused control users (0.03). Pictures are relevant for the characterization of users. Focused and random control users are characterized by the usage of text on their profile pictures. We also found a strong polarization between focused control users, and users at the first stages of the disorder. There was a strong correlation (Spearman’s coefficient) among the shared interest between users with AN and their followees (0.96). Also, the interests of recovered users and users in treatment were more highly correlated to those corresponding to the focused control group (0.87 for both) in comparison to AN’s users (0.67), suggesting a shift on users’ interest during the recovery process. CONCLUSIONS We have mapped signs of Anorexia Nervosa to the Social media context. These results enforce the findings of related work on other languages and involve a deep analysis on the topics of interest of users at each phase of the disorder. The features and patterns identified provide a basis for the development of detection tools and recommender systems.

Enhanced Word Embedding Variations for the Detection of Substance Abuse and Mental Health Issues on Social Media Writings

“…For the comparison of the groups analyzed (DEP, ED, SUI, and ALC), as in [20], we apply non-parametric tests after verifying that our features do not follow a normal distribution and that there is no homogeneity of variance for most of them. We first verified that there were features with significant differences among all the groups using Kruskal-Wallis' test [21].…”

“…We analyzed 5 different perspectives as defined in the Linguistic Inquiry and Word Count (LIWC) 2007 dictionary [23], which categorizes words in psychologically meaningful perspectives. We also consider other perspectives that are defined through related work addressing different mental conditions [9,20]. The description of these perspectives and the results obtained for our comparative analysis are described as follows.…”

“…These users (suicide risk) are the ones that express more negative emotions. Users with eating disorders are the most positive ones compared to the other groups within the conditions analyzed, and this can be due to the fact that eating disorders such as Anorexia and Bulimia can be characterized by the Transtheoretical Model of Health Behavior Change (TTM), where patients at the pre-contemplations stage are enthusiastic about their weight loss, and the social support they receive [20,25]. The groups having the least difference between each other are the Depression and Alcoholism groups as shown by the non-significant results of the tests for several categories (disgust, joy, surprise, etc.).…”

“…Fig. 2 shows only the top 20 Empath topics having the most significantly different values (P<.05) [20] between each pair of classes compared, including the mental conditions and control classes. The mean value for each class compared for each topic is shown.…”