JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Original Paper Characterization of Anorexia Nervosa on Social Media: Textual, Visual, Relational, Behavioral, and Demographical Analysis Diana Ramírez-Cifuentes1, MSc; Ana Freire1,2, PhD; Ricardo Baeza-Yates1,3, PhD; Nadia Sanz Lamora4, MSc; Aida Álvarez4,5,6, MD, PhD; Alexandre González-Rodríguez4,5,6, MD, PhD; Meritxell Lozano Rochel7, BSc; Roger Llobet Vives7, MSc; Diego Alejandro Velazquez8, MSc; Josep Maria Gonfaus9, PhD; Jordi Gonzàlez8, PhD 1 Department of Information and Communication Technologies, Universitat Pompeu Fabra, Barcelona, Spain 2 UPF Barcelona School of Management, Barcelona, Spain 3 Institute for Experiential AI, Northeastern University, Boston, MA, United States 4 Department of Mental Health, Centro de Investigación Biomédica en Red de Salud Mental, Parc Tauli University Hospital, Sabadell, Spain 5 Parc Tauli Research and Innovation Institute, Sabadell, Spain 6 Autonomous University of Barcelona, Bellaterra, Spain 7 Fundación Instituto de Trastornos Alimentarios, Barcelona, Spain 8 Computer Vision Center, Universitat Autonoma de Barcelona, Bellaterra, Spain 9 Visual Tagging Services, Bellaterra, Spain Corresponding Author: Diana Ramírez-Cifuentes, MSc Department of Information and Communication Technologies Universitat Pompeu Fabra Carrer de Tànger, 122-140 Barcelona, 08018 Spain Phone: 34 643294184 Email: diana.ramirez@upf.edu Related Article: This is a corrected version. See correction statement in: https://www.jmir.org/2021/10/e33447 Abstract Background: Eating disorders are psychological conditions characterized by unhealthy eating habits. Anorexia nervosa (AN) is defined as the belief of being overweight despite being dangerously underweight. The psychological signs involve emotional and behavioral issues. There is evidence that signs and symptoms can manifest on social media, wherein both harmful and beneficial content is shared daily. Objective: This study aims to characterize Spanish-speaking users showing anorexia signs on Twitter through the extraction and inference of behavioral, demographical, relational, and multimodal data. By using the transtheoretical model of health behavior change, we focus on characterizing and comparing users at the different stages of the model for overcoming AN, including treatment and full recovery periods. Methods: We analyzed the writings, posting patterns, social relationships, and images shared by Twitter users who underwent different stages of anorexia nervosa and compared the differences among users going through each stage of the illness and users in the control group (ie, users without AN). We also analyzed the topics of interest of their followees (ie, users followed by study participants). We used a clustering approach to distinguish users at an early phase of the illness (precontemplation) from those that recognize that their behavior is problematic (contemplation) and generated models for the detection of tweets and images related to AN. We considered two types of control users—focused control users, which are those 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 tweeted more frequently at night, with a median sleep time tweets ratio (STTR) of 0.05, than random control users (STTR=0.04) and focused control users (STTR=0.03). Pictures were relevant for the characterization of users. Focused and random control users were characterized by the use of text in their profile pictures. We also found a strong polarization between https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 1 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al focused control users and users in the first stages of the disorder. There was a strong correlation among the shared interests between users with AN and their followees (ρ=0.96). In addition, the interests of recovered users and users in treatment were more highly correlated to those corresponding to the focused control group (ρ=0.87 for both) than those of AN users (ρ=0.67), suggesting a shift in users’ interest during the recovery process. Conclusions: We mapped the signs of AN to social media context. These results support the findings of previous studies that focused on other languages and involved a deep analysis of 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. (J Med Internet Res 2021;23(7):e25925) doi: 10.2196/25925 KEYWORDS social media; Twitter; Spanish; anorexia nervosa; eating disorders; user characterization psychologists in screening [10-12] and treatment proceedings Introduction [6]. Research findings on these topics can also contribute to the Background improvement of the structure and services provided by online social platforms [6], which are a means through which people Mental disorders are psychological syndromes or patterns with mental disorders can find support for their recovery, as associated with distress or disability. Anorexia nervosa (AN) well as they can be used as tools to promote harmful content, is an eating disorder (ED) characterized by the maintenance of which is the case for suicide promoters and pro–eating disorder a very low body weight, with a constant desire for thinness and (pro-ED) communities [6,13]. a misconception about body shape [1]. EDs, such as AN, are related to risk factors including perfectionistic traits [2], and Recent approaches for characterizing AN on social media the existence of comorbid mood disorders such as depression, account for individual posts from users, meaning that the signs of which 33%-50% of anorexia patients experience. Anxiety and symptoms are traced by analyzing the content of a single disorders are common in 50% of these patients [3]. post per user (post level) [6,8,9]. In contrast, there are only a few approaches that address multiple posts of a single user (user The course of AN is highly variable; however, early intervention level) [10,11]. There are no approaches that address the is strongly associated with full recovery. The phases that patients simultaneous analysis of relational, behavioral, demographical, go through as they overcome AN can be mapped to the and multimodal information at the user level, although few transtheoretical model of health behavior change [4], which is studies have combined some of these aspects on AN-related described as an integrative method for understanding how people disorders such as anxiety, depression, and suicidal ideation progress toward adopting and maintaining healthy behaviors. [12,13]. This model identifies the following six stages of change: (1) precontemplation, where the individual does not know that there With regard to the type of information extracted and analyzed is a problem or that a change is required in their life and, thus, about EDs on social media, several state-of-the-art approaches does not seek help; (2) contemplation, where the person are dedicated to the analysis of textual information contained simultaneously considers and rejects the change, while being in posts [6,9,14]. They analyze the topics of interest using topic conscious of the existence of a problem; (3) preparation, where modeling techniques [11], and they also consider the frequency the individual starts to take small steps toward behavioral of the terms used through bag-of-words (BoW) models and change, believing that it can lead to a healthier life; (4) action, n-grams [7]. Other textual elements are also considered, such in which the person has changed their behavior and intends to as the frequency of terms closely related to the illness and other sustain it; (5) maintenance, a stage wherein the person has lexical and syntactic elements [11,14]. In addition, sentiment maintained the behavioral change for a considerable period (>6 analysis tools [15] and methods that use word embeddings, months); and (6) termination, a stage wherein the individual which are vector representations of terms, have been applied has no desire to return to their unhealthy behaviors. It is [7,16]. Most of these features are later used as representations important to state that relapse, which implies returning from suitable for the development of predictive models that are often the action or maintenance stages to an earlier stage, is likely. based on machine learning methods [6,7,16]. Through the This approach has been previously evaluated in patients, analysis of text cues, it has been found that users with depression confirming that the notion of the stages of change is an and those with suicidal ideation make more self-references [13], independent dimension that is very relevant for the treatment whereas proanorexia and prorecovery communities exhibit of EDs [5]. distinctive affective, social, cognitive, and linguistic style markers, as proanorexics express greater negative affect, feelings Automated methods have been designed to detect signs of AN, of social isolation, and self-harm [6]. some of which address the development of early detection approaches [6,7], as it has been proven that the signs and With regard to the analysis of relational information, Wang et symptoms of mental disorders, including AN, can be traced al [17,18] analyzed the interactions between users with EDs to using social media [6,8-13]. The findings of such research have identify the topics shared in web-based conversations. In revealed patterns that can be relevant for the development of particular, they explored the interactions among communities tools to detect harmful content [6] and to assist clinicians and with different stances on EDs, considering both pro-ED and https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 2 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al prorecovery content, and reported that intercommunity social platforms?; (2) Which elements characterize the social interactions are very limited. In the context of similar disorders, network of users that use anorexia-related terms?; (3) Which De Choudhury et al [19] constructed egocentric social graphs are the topics of interest among AN users and their followees?; for postpartum depression detection and found that increased (4) Which linguistic attributes characterize Spanish-speaking social isolation and reduced availability of social capital on users with AN?; (5) Does the proportion of tweets related to Facebook are the best predictors of postpartum depression AN significantly change according to the recovery stage?; (6) detection among mothers. In addition, Colombo et al [20] In a social platform context, can the elements that distinguish studied the connectivity and communication of suicidal users users at the precontemplation stage from users at the on Twitter based on the evaluation of retweets. contemplation stage be identified?; and (7) Are there significant differences between the focused and random control groups? Regarding the analysis of images shared by users, only a few studies analyzing the content of images to determine patterns The main contributions of this work are as follows: that can be linked to AN signs are available in the literature. • We generated a Twitter data set that is annotated by This aspect of analysis is important for studies on EDs, as the psychologists, psychiatrists, and therapists and consists of symptoms of AN, such as extreme weight loss, an obsessive writings of users manifesting AN (including writings interest in excessive exercise, and restricted food intake through corresponding to treatment and of fully recovered users), dieting, can be identified from images [1]. Chancellor et al [21] and two types of control cases (ie, focused and random focused on Instagram, which mainly includes image posts, but control). To the best of our knowledge, this is the first they exclusively studied textual (captions) and relational Spanish data set for the analysis of AN at the user level that elements. Recently, only a few approaches have addressed the considers different stages of the illness toward recovery. actual analysis of image properties to detect mental disorders, • We extracted and inferred several features that consider such as self-harm, depression, and anxiety [12,22,23]. To the multiple elements, as follows: images, texts, relations extent of our knowledge, the work of Chancellor et al [16] is among users, posting patterns, and demographic the only available study on EDs that analyzed images. information. These features were generated to identify Chancellor et al [16] classified moderated pro-ED content and elements that characterize users with AN at different stages found that the identification of deviant pro-ED content through of the illness and recovery. We also determined the elements the analysis of images was possible, reaching a recall of 84%. that distinguish these users from the two types of control This approach mainly focused on the detection of harmful cases. images but did not include the analysis of interpretable elements • We established a deep learning–based method to identify and patterns that might characterize AN-related content on tweets related to AN using individual tweets. images. • We trained a predictive model using the images of users There are also a few studies that consider the stages that patients with AN and control users to detect whether differences go through during the course of the illness. De Choudhury [6] between these groups can be identified on the basis of visual and Wang et al [17] addressed the detection and characterization properties. of prorecovery communities, which can be distinguished from • We further explored the social network of users with AN pro-ED communities. However, both of these communities did through the detection of communities and the analysis of not exclusively include people with the illness. Chancellor et topics of interest of the different types of users, along with al [24] determined the content and participation measures that those of their followees. could predict the likelihood of recovery of these types of users • We applied a clustering approach to distinguish users at the on Tumblr. They found that only half of the analyzed users were precontemplation stage from those at the contemplation likely to exhibit signs of recovery after a 4-year period, stage using the transtheoretical model. demonstrating the importance of further consideration of the stages that lead to recovery, thus supporting the transtheoretical Methods model. Data Collection Regarding the analysis of content generated by Spanish speakers with EDs, we identified two approaches that address writings We chose Twitter as our main data source, as it has been in this language [9,25]. Both approaches are focused on the previously used for studying mental disorders, including EDs, analysis and detection of users with AN at the post level, with on social media [14,27]. Twitter is suitable as it allows the a main focus on textual-based features. Furthermore, these collection of a set of chronologically organized posts in Spanish approaches do not address any further recovery stages. Recent and provides metadata relevant for analyzing the relational and works have also addressed other mental conditions such as behavioral aspects of users. depression [26] and suicidal ideation [13]. We selected keywords and popular hashtags commonly used Objectives by ED communities, phrases likely to be used by people undergoing treatment, and terms used by recovered users. These This paper addresses the characterization of EDs, in particular keywords and phrases were manually collected and classified AN, in social media. We aim to answer the following research from multiple sources in Spanish and English, including proana questions (RQs): (1) Which are the visual, textual, behavioral, blogs, academic publications, and documents made available and demographical elements that characterize and distinguish users at the early, treatment, and full recovery stages of AN on https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 3 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al by the Spanish Association Against Anorexia and Bulimia symptoms of AN in their texts, which includes users at both the [14,28,29]. precontemplation and contemplation stages according to the transtheoretical model; (2) a focused control group in which, In addition, we conducted a survey among volunteers who have similar to our prior work [13], we included users that did not recovered from AN. The phrases and keywords collected were manifest signs of anorexia but use terms related to the disorder evaluated and filtered by clinicians that were asked to agree on in their writings; (3) treatment users that explicitly stated that choosing up to 30 keywords or phrases in Spanish that would they have been diagnosed with AN and are in treatment; (4) lead to reach posts from users with AN, including proana, peso recovered users who claim they have recovered from AN; and objetivo (objective weight), perder peso (lose weight), IMC (5) doubtful cases in which clinicians were not sure about any (body mass index BMI), sibutramina (sibutramine), mi anorexia of the prior categories. A total of 5 annotators participated in (my anorexia), ana y mia (ana and mia). We collected 114,627 the labeling process: 3 psychologists and 2 psychiatrists. These public tweets containing the search phrases, considering a period annotators collaborated closely with organizations specializing in the past, from December 21, 2017, to December 21, 2018. in the treatment of EDs. The final label for a user’s set of tweets At the same time, a sample of up to 10,000 tweets from the was assigned if at least three annotators agreed on the assigned same search period was collected for each user. label. For cases where an agreement was not met, the users’ To protect the privacy and identity of these users, generic tweets were categorized as doubtful cases. identifiers were assigned to each of them and their posts. In From this first classification approach, a total of 195 users were addition, we removed any personal information from the users’ classified as users with AN, 283 as focused control users, 29 as descriptions and tweet texts such as usernames, names (proper under treatment users, 18 as recovered users, and 119 as doubtful nouns), email addresses, URLs, password combinations, location cases. We performed an interannotator agreement analysis and names, and all numbers. The extracted metadata elements and obtained a Light κ coefficient of 0.4751 (P<.001), which is the texts passed through a strict transformation process to build and result of the averaged Cohen κ values calculated between each exclusively store vector representations of the features of pair of annotators. This approach was chosen over Fleiss κ, as interest, guaranteeing the analysis of fully anonymized all annotators evaluated every sample. The values obtained information. suggest a moderate agreement among annotators. Annotation Process In addition to the focused control group, we included another For labeling purposes, we filtered and only considered users control group consisting of 223 randomly selected users called with at least three different tweets containing the selected random control users. These users did not necessarily use terms keywords for each category. Among all categories, 645 users related to AN and were selected using Twitter’s Sample Tweets met this criterion. Before the submission of the text samples to application programming interface (API) [30], which provides the annotators, the sample of tweets' texts selected for annotation a set of random tweets, from which we could obtain the users were anonymized and translated to English to avoid the for our sample. By analyzing the description and a sample of reidentification of users based on their writings. their tweets, we ensured that no users with AN were part of this We defined five independent groups of users: (1) AN users that group. The anonymization and translation process applied for manifest the first stages of the disorder and describe signs and the other groups was followed for this case. Table 1 summarizes the user groups defined initially for the study. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 4 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 1. User groups defined for the study (N=748). Group Description Collection method Users, n (%) ANa • Users most likely at early and advanced stages of AN AN-related keywords using the Twitter Search APIb 195 (26.1) that do not seem to be on treatment • This group corresponds to the precontemplation and contemplation phases of the transtheoretical model. Treatment • Users most likely in treatment of AN AN- and treatment-related keywords using the 29 (3.9) • This group corresponds to the preparation, action, and Twitter Search API maintenance phases of the transtheoretical model. Recovered • Users that seem to be recovered from AN and have AN- and recovery-related keywords using the Twitter 18 (2.4) reached the termination stage Search API • They should specify not having a relapse in a long period (≥4 years). Focused control • Control users that use vocabulary related to AN, such AN-related keywords using the Twitter Search API 283 (37.8) as psychologists, news accounts, or medical centers. Random control • Twitter users randomly selected with no signs or Sample tweets using the Twitter Search API 223 (29.8) symptoms of AN a AN: anorexia nervosa. b API: application programming interface. We also tested a clustering approach to distinguish cases at the Characterizing Users With AN on Social Media precontemplation stage from cases at the contemplation stage This work is dedicated to the characterization of AN considering (RQ6). People in the precontemplation phase (also known as the multiple stages through which people on the recovery path the honeymoon phase) are characterized by being enthusiastic go through (RQ1), including the analysis of the differences about their weight loss and the social support they receive. This between the two control groups analyzed: focused and random is a phase in which people believe they are in control of their control (RQ7). Our goal is to identify the elements and patterns behavior, and therefore, they are in denial of their unhealthy that distinguish people at each phase based on the analysis of condition. This behavior differs from the contemplation stage, multiple perspectives that are usually considered by clinicians in which bad habits and symptoms remain, but as the illness for screening and treatment purposes. The perspectives analyzed progresses, the person notices that there is something wrong include (1) the way Spanish-speaking users express themselves and feels that they are losing control of the situation. People at through written posts (content shared) that can lead to the this stage are constantly depressed and may even experience a exploration of linguistic dimensions, affective processes and decrease in their cognitive capabilities [31]. Taking into account emotions, personal concerns, topics of interest, vocabulary that these aspects, we consider the detection task relevant, as people implies the existence of risk factors associated with the illness, at the contemplation stage are more willing to seek help, and and vocabulary related to signs and symptoms of AN (RQ4); therefore, a potential nonintrusive intervention through social (2) the social network of users, with a focus on the interests media, such as social recommendation, could be successful shared with their followees and the detection of communities having users at this stage as a target. However, such potential (RQ2 and RQ3); (3) the users’ behavior, with the analysis of applications must be further analyzed by considering all the posting patterns in different periods; (4) demographic aspects privacy and ethical aspects involved. such as gender and age ranges; and (5) visual elements that include the analysis of the characteristics of the profile picture Experimental Setup of users, along with the images shared on posts. Data Set Description In addition to the analysis of prior perspectives, for a further The data set built for our analysis consists of a set of features exploration of textual cues, we built and evaluated classifiers generated and inferred based on the text, images, and metadata for the detection of individual tweets related to AN. These tools of the annotated users’ tweets. Table 2 provides relevant can be useful for the detection of users with AN, as we assume information regarding each group in our data collection. For that a larger number of tweets related to anorexia will be found each user, we considered the content from their profiles (tweets) in the profiles of users living with the disorder. Through this during a year (from December 21, 2017, to December 21, 2018). analysis, we also wanted to analyze whether the proportion of Some of the initially labeled users were not further considered tweets related to AN changes significantly according to the in the data set, as they had published less than five tweets during recovery stage (RQ5). In addition, these tools can later be useful the data collection period, which we considered as not for automatically filtering writings that are exclusively related informative enough for our analysis purposes. A total of 694 to the condition. They can also be used for screening and users were part of our final data set, which contained data annotation purposes [13]. collected from 2,133,110 tweets, including 405,909 images. It https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 5 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al is important to recall that retweets were not considered within the scope of our analysis. Table 2. Data set description. Descriptive item User group Anorexia nervosa Treatment Recovered Focused control Random control Users (n=694), n (%) 171 (24.6) 27 (3.9) 18 (2.6) 271 (39) 207 (29.8) Tweets collected (n=2,133,110), n 434,615 (20.4) 8317 (0.4) 52,578 (2.5) 1,109,861 (52) 447,739 (21) (%) Number of tweets collected per user 1239 1748 2036.5 2608 873 (n=2,133,110), median Tweet length (number of words; 14.00 14.00 13.50 12.50 19.00 n=2,133,110), median Images (n=405,909), n (%) 40,142 (9.9) 6584 (1.6) 4202 (1) 298,488 (73.5) 56,493 (13.9) k-means classifier for the detection of users at the Comparative Analysis and Evaluation Measures precontemplation and contemplation stages, as we did not have As part of our data collection process, we extracted, calculated, data annotated for this purpose. Our predictive approaches were and inferred some features for performing the analyses required evaluated using machine learning evaluation measures, such as to answer our RQs. For this purpose, we considered network precision, recall, and F1 score. The details of these approaches clustering and visualization algorithms (RQ2 and RQ3); prebuilt are described in the Proportion of AN-Related Tweets, Detection machine learning models for sentiment analysis; and age range of Precontemplation and Contemplation Phases, and Pictures and gender detection tools, including models for the detection Shared sections. of objects in images (RQ1, RQ4, RQ5, RQ6, and RQ7). We also used external sources with lexicons to detect emotions, Analyzed Perspectives and Features topics of interest, risk factors, and affective processes lexicons Overview (RQ1, RQ4, RQ6, and RQ7). The description of the extracted In this section, we describe the perspectives and specific features features and the calculation or inference processes are detailed analyzed for the characterization of AN across its stages. The in the Analyzed Perspectives and Features section. Phyton perspectives and categories of the features considered within (Python Software Foundation) was used as our main each perspective are studied by considering their relation to programming language for data extraction and processing elements that clinicians address for screening and treatment procedures. purposes [1,2] and their use for the characterization of related To perform a comparison of the groups analyzed, we considered mental health issues such as anxiety [12], depression [23,26], some approaches for hypothesis testing. To do so, we first and suicidal ideation [13,20]. verified that our numerical features did not follow a normal Content Shared and Interests distribution and that there was no homogeneity of variance for most of them. Therefore, we considered nonparametric tests. Overview First, we applied a Kruskal-Wallis test [32] for multiple This perspective addresses RQ1, RQ3, RQ4, RQ5, and RQ7 independent samples, and after finding features with significant through an analysis of the textual content shared by users in differences among the groups, we proceeded to perform the their tweets. It considers linguistic and psychological aspects Mann-Whitney U [33] test to check for differences between through six categories. Some of these categories were based on pairs of groups of interest. As we considered some categorical a classification given by the Linguistic Inquiry and Word Count elements as well, such as age groups, we transformed them into (LIWC) 2007 Spanish dictionary [37,38], which categorizes Boolean representations to perform a two-sided proportion z words into psychologically meaningful categories. The test among the groups with these feature types, which is a test remaining categories were defined by considering psychological equivalent to the proportions chi-square test [34]. aspects related to EDs, which were defined under the supervision Our work addressed three predictive tasks solved using machine of clinicians. The categories analyzed were as follows: linguistic learning: one to build a classifier that identifies tweets related dimensions (24 features) [37,38], affective processes and to AN (RQ5), another to distinguish users in the emotions (29 features) [37,38], personal concerns and biological precontemplation phase from users in the contemplation phase processes (12 features) [37,38], vocabulary related to risk factors (RQ6), and the last one to detect individual images related to (10 features) [1,2,13], anorexia-related vocabulary (9 features) AN (RQ1). We generated deep learning models based on [14], and topics of interest to the users (200 topics). From this convolutional neural networks (CNNs) [35] and statistical perspective, we aim to explore the linguistic factors and shared models such as logistic regression (LR) [36] for the detection interests that characterize each group. In addition, we map of texts and images related to anorexia, taking into account these certain elements that are relevant for specialists when identifying as supervised tasks for which we have annotated data to learn signs and symptoms of AN, such as the existence of risk factors, from. Later, we consider a clustering approach given by a https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 6 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al the use of terms related to the illness, and the expression of Within these aspects, we find terms related to religion; work; emotions and personal concerns of the users. leisure; money; and biological processes, such as body, ingestion, and health. These aspects are relevant to our study, Linguistic Dimensions as personal background aspects are considered to be relevant These features were extracted using a LIWC dictionary [37,38]. for the development of EDs [2]. Biological processes are also The features belonging to this category address the use of relevant to address, as these concerns are representative of grammatical and syntactical elements such as pronouns, verbs, patients with anorexia [1]. It is important to specify that we only adverbs, prepositions, and articles, taking into account the address the use of general terms such as God, doctor, or office, different verbal times and pronoun types. These are relevant for instance, we do not do a further exploration regarding elements for our analysis because they can give us insight into personal religious beliefs, professions, or health information the elements or people that users are drawing their attention to. within the text. These types of elements have been analyzed in prior work where other mental disorders such as depression are analyzed; for Vocabulary of Risk Factors instance, Rude et al [39] in their work on depression found that These features correspond to lexicons consisting of up to people who are experiencing physical or emotional pain tend 3-grams phrases that were mapped to ED risk factors, such as to draw their attention to themselves and therefore use more terms referring to self-injuries; suicidal ideation references; first-person singular pronouns. For our approach, after self-loathing terms; words that express disdain, substance abuse, concatenating all the tweets’ texts of a user into one long text, lack of social support, family issues; and vocabulary, which we counted the frequency of words belonging to each of the refers to discrimination- or abuse-related topics [2,13,42]. The categories considered from the dictionary, which was normalized terms and phrases selected for these categories were based on by the size (in number of terms) of the resulting concatenation. manually mapping common terms and phrases related to In addition to the LIWC elements, we also considered the anorexia and the terms used in a related task for the detection median tweet length (number of terms) as a feature. With the of suicidal ideation [13]. analysis of these features, we also check for the elements that Anorexia-Related Vocabulary characterize the linguistic attributes of Spanish speakers living with this disorder so that they can be compared with prior studies On the basis of the work of Arseniev et al [14], we used the in English (RQ4). categories of terms related to AN and its symptoms. Subsequently, we translated them into Spanish. We also kept Affective Processes and Emotions some of the terms in English, as they are also used by These features are extracted using the LIWC dictionary and Spanish-speaking users. In addition to these categories, we EmoLex [40], which is a dictionary that associates words with added names of known laxatives in Spanish [43]. These eight basic emotions, anger, fear, anticipation, trust, surprise, categories refer to topics such as anorexia promotion, body sadness, joy, and disgust, and two sentiments, negative and image, body weight, food and meals, caloric restrictions, positive. This dictionary also provides a file to associate the compensatory behaviors, and exercise. original terms in English with their Spanish translations. Along Topics of Interest with the emotions explored, we addressed cognitive processes, senses, perceptions, and social processes (LIWC). The values The topics of interest of a user are also among the elements that obtained for a given user were calculated in the same way as we analyze from the content perspective (RQ3), as we would the values for the linguistic dimensions. In fact, the same like to know if there is a shift in the main interests of users calculation approach is also used for personal concerns, risk through the recovery path. For this purpose, we consider factors, and anorexia-related vocabulary features. The analysis predefined topics of interest based on categories that are part of this aspect is relevant, as general psychiatric disturbance and of the Empath tool [44], which generates and validates lexical negative emotionality are elements that characterize people categories on demand from a small set of seed terms (such as living with EDs [2]. football and tennis to generate a sports category). This tool draws connotations between words and phrases by deep learning In addition to the prior elements, we make use of a sentiment a neural embedding using a corpus with 1.8 billion words. In analysis model, which provides a polarity value for an individual our case, 200 prebuilt topics are considered, including sports, text, in the range of [0,1], from the most negative to the most music, social media, and politics, among others. positive polarity. We used Senti-py [41], trained on Spanish texts from different sources, including Twitter. It is based on a For each user, we calculated the scores for all topics. We take BoW model that goes through an intermediate feature selection into account that the topics of interest of a user are given by the process. To obtain a score per user, we calculated the median interests of their followees, the content they like (given by the polarity scores of all the tweets. The higher the polarity score tweets made by others and marked as favorites), and by the obtained, the more positive the text was expected. content posted by themselves. For each user, we collected (1) a random sample of their own tweets (up to 500 texts), (2) a Personal Concerns and Biological Processes random sample of 200 tweets that they had liked during the This is another perspective addressed through the analysis of same period, and (3) the profile descriptions (biographies) of lexicons (LIWC). The elements explored provide a general up to 200 random followees of the user. These tweets and perspective of the use of terms related to aspects that involve descriptions were relevant enough samples of texts that daily activities and concerns, along with biological aspects. characterized the interests of a user. An individual score with https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 7 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Empath was obtained for each text (tweet or description). Later, where a task-oriented fine-tuning was performed, and we applied the final score for a topic for a given user was calculated by a filter window ({2,3,5} terms). We applied max pooling and averaging the scores obtained by the topic on all the tweets passed the output to a sigmoid layer to generate the final output. considered. It is important to mention that as Empath’s Furthermore, 75% (2649/3532) of the instances were selected categories are in English, we add a translation step before the for training purposes and the remaining 25% (883/3532) for Empath scores’ calculation, using the Googletrans Python API testing. Among the training instances (tweets), 69.98% [45] for this purpose. The amounts of tweets and descriptions (1854/2649) were selected for training the model and 30.01% defined for this approach are also based on the request (795/2649) were considered for validation. limitations of the API of Twitter and the Googletrans API. Social Network Proportion of AN-Related Tweets Overview Our RQ5 analyzes whether the proportion of tweets related to This perspective addresses RQ1, RQ2, and RQ3. Features are AN changes significantly according to the recovery stage, as it extracted taking into account the social network of the user, as is expected that users at the initial stages produce more tweets elements related to this perspective have been proved to be related to their condition. For this purpose, we first built and useful for characterizing other mental conditions such as compared two models to detect, for each user, if each of their depression [19] and suicidal ideation [20]. We analyze some tweets are related to AN. Second, we calculated the median features that characterize the user´s popularity and the support score obtained by the classifier for all user tweets. Finally, we received by other users. These features correspond to the number compared the median values for all users belonging to a group of followers, favorites, and retweets of their posts. We focused to measure the presence of AN tweets in each group. It is on the social network (followees) of users that make use of expected that users with AN have a median value significantly anorexia-related terms (RQ2), as our goal was to detect higher than users in the control and recovered groups. The communities among these types of users. Furthermore, we approach of creating classifiers such as these has previously explored the likelihood of users with AN to follow users living been used to detect users with suicidal ideations [13]. In this with the disorder or anorexia promoters by analyzing the topics case, the model was built exclusively using a BoW model and of interest of their followees. We also explored the differences LR. in their interests and those of the followees of users in treatment For the AN use case, we trained two classifiers to distinguish and the followees of recovered users. tweets of two classes: (1) anorexia related and (2) control. The Measures of Interactions and Engagement instances of the anorexia-related class corresponded to the individual tweets belonging to the users labeled as AN cases These features are extracted from the metadata of the users’ (1766 tweets). Later, an equivalent number of tweets was tweets. These features tell us about the relationships and selected to represent the control class; these tweets were interactions of AN users, which can differ from the interactions randomly extracted using Twitter’s Sample Tweets API [30]. of control users [19]. The features extracted and calculated for The sample of tweets collected was reviewed by annotators to each user are as follows: number of followees, number of discard those that could possibly belong to the anorexia-related followers, total number of favorites given to the posts of other class. users, median number of favorites received by the user, and median number of retweets received by the user. These last two The first classifier was trained over a BoW model with features were calculated by considering the tweets of the sample (1-3)-grams at a term level. For this purpose, we used the of the user profile. Scikit-learn [46] Python library: TfIdfVectorizer to generate a tf.idf representation with (1-3)-grams. We considered Spanish Analysis of Followees and Community Detection stop words [47] and used ekphrasis [48] as a text preprocessing As part of the perspective that analyzes the social network of a tool to replace terms referring to money, hashtags, and user, we explored the network of users that made use of emoticons with generic tags. As we obtained 66,404 features, anorexia-related terms, corresponding to the AN, treatment, we reduced the number of features to 300 using principal recovered, and focused control groups. This was done with the component analysis. We used an LR method and 10-fold cross purpose of identifying characteristics of the network that were validation. capable of distinguishing the groups defined, in particular the AN group and the focused control group, as users representing For the second classifier, a deep learning approach was applied. organizations that provide medical and psychological support The model was defined through a CNN architecture that has could be part of it, and it would be relevant to get an insight been previously applied to text classification tasks [35], into the relationships between both groups. For this purpose, including a similar task for suicide risk assessment on social we extracted a sample of up to 100 followees of each user from media [13,49]. The same preprocessing approach as that used each of these groups (considering Twitter’s API request for the prior model was applied. To train this model, tweets limitations). We built a directed graph where a link between were represented as sequences of terms, and these terms were two nodes was given by a follows relationship, meaning that represented by prelearned word embeddings that were trained users, represented by nodes, are linked to other nodes through over tweets in Spanish [50]. Each tweet was considered as an directed edges where the arrowheads point to the users they instance, and its label (anorexia related or control) corresponded follow. Later, a clustering algorithm was applied to detect to the class assigned to the tweet. For the CNN [35], the communities among these users. We then performed a embedding sequence instances were given as the model input, https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 8 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al comparison between the communities automatically detected as validation subgroups, as explained in Table 3, where we and what we defined as validation groups, which were created describe the general organization of a group. We considered considering the followees of the AN, treatment, recovered, and four main groups and three subgroups per group, where the first focused control groups. These validation groups were defined subgroup always corresponded to the original users labeled. An in such a way that a user was assigned to a validation group instance of a validation group would be Group AN, which is (AN, treatment, recovery, or control) if it was mostly followed composed of three subgroups: G1 composed of the originally by users belonging to the originally labeled groups. This was labeled AN’s users, G2 composed of the users mostly followed done taking into account up to two followees’ levels, denoted by G1 users, and G3 composed mostly of users followed by G2. Table 3. Groups for social network analysis based on users’ labels. Group and subgroup Nodes user type Group Xa Subgroup Gbi Users manually labeled as part of the X group Subgroup Gi+1 Users mostly followed by Gi Subgroup Gi+2 Users mostly followed by Gi+1 a X: anorexia nervosa; focused control; treatment or recovered. b G: group. On the basis of a manual revision of a sample (translated to we address the followees of each user type. As shown in Figure English) of the profile descriptions of users belonging to the 1, for this case, we considered up to 25 followees from a sample communities detected with most nodes, we performed a further of up to 25 users per group analyzed. Then, for each of these analysis of the types of users that were identified as part of each followees, we calculated scores for the Empath topics by community, and we mapped these communities to our considering the descriptions of 25 random followees, a random predefined groups so that we could identify which type of users sample of their own tweets (up to 200 texts), and a random from our groups of interest were part of the communities sample of 200 tweets that they had liked during the same period. detected. For the visualization of the social network, we The score of a topic for each followee of a user is given again considered the Force Atlas 2 [51] algorithm, and for the by the average score obtained from all the texts considered. The detection of communities, we used the method by Louvain [52], score for a topic of a user is given by the median of the scores both implemented on Gephi [53]. of their followees. Once the scores for the samples of users representing each group were obtained, we calculated the median Analysis of Interests Between Users and Their Followees value of each topic using the scores of each user belonging to To address RQ3, as it is our purpose to identify the topics of the group. Later, we performed a comparison between the interest of AN users’ followees, we follow the process applied interests of users of each group (calculated before) and those for the analysis of the topics of interest of users of each group, of their followees. as described in the Topics of Interest section, but in this case, https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 9 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Figure 1. Structure defined for the extraction of the interests of the followees of a given user group. For each labeled user of a group, we analyzed the tweets posted and liked by their followees and the profile description of the followees of the labeled users’ followees. passes between the publication of each tweet, and the ratio of Behavioral Aspects tweets posted during the regular sleep time of the user (sleep Addressing RQ1 and RQ7 from this perspective, we define period tweeting ratio). We also calculated the ratio of tweets features that measure the users’ activity to explore elements posted during the period when the user was usually awake that can link the frequency of use of social platforms and AN. (awake period tweeting ratio). Both of these last features were An instance is the level of activity of users at night, which might calculated based on the sleep time tweets ratio (STTR) and the imply insomnia, a sign that has been linked to related disorders day time tweets ratio (DTTR) inference approach [13]. This such as suicidal ideation [13]. These features were extracted approach was adopted because it addresses the issue of the lack from the metadata of the tweets, and they measured the behavior of information regarding the location of the users, which is of users based on their activity on a daily, weekly, and monthly mostly not shared by them. basis. The features considered are as follows: the number of tweets on weekdays (Monday-Friday) normalized by the number For the calculation of the sleep period tweeting ratio, equivalent of tweets from the whole week (weekly tweets ratio), the number to the STTR (equation 1) [13], a day is divided into eight fixed of tweets on weekend days (Saturday and Sunday) normalized time slots of 3 hours each, and it is assumed that an average by the total number of tweets of the week (weekend tweets user has approximately 6 hours of sleep time. This is precisely ratio), the number of tweets posted by the user within each the period where a fewer number of tweets would be created quarter of the year normalized by the total number of tweets of compared with the rest of the day, so the number of tweets (t) the year period (year quarter tweets ratio), the total number of created within each 3-hour time slot is counted for all the tweets tweets posted during the year, the median time (in seconds) that of the user, and the sum of the number of tweets within each https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 10 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al pair of continuous time slots is calculated. Then, the minimum and objects. As part of the technical features analyzed, we score obtained by all the pairs is selected, so that later this value checked if an image is gray scale, if it is lighter, if it has text, can be normalized according to the total number of tweets of and if it has faces on it. We also analyzed the existence of the user (T). It is assumed that the first and last slots are objects in the pictures. These features are defined through the continuous. The awake period tweeting ratio is given by the use of Python libraries such as imageio [56] for verifying the DTTR, which is equivalent to the difference between 1 and the brightness of an image, PIL [57] for verifying if an image is sleep time ratio: DTTR=1−STTR. gray scale, and the General Recognition AI API from Chooch AI [58], which after taking an image as an input, it outputs the names of elements recognized in the picture, such as texts, clothes, faces, animals, and specific objects. We also detected Demographics emotions expressed on the pictures using the Algorithmia facial We analyzed the demographic characteristics (gender and age emotion recognition API, which implements CNN models [59]. features) of the groups to verify whether these correspond to The models included in the previous APIs were already trained, the actual incidence rates [54] (RQ1 and RQ7). These features so we only ran them over the profile pictures in our data set, are inferred, given the fact that Twitter does not publicly display and no rights over a further use of these images were granted the age and gender of users. We used the approach of Wang et to the tools’ owners. In addition, none of the images were al [55] for demographic inferences. This approach is based on observed by any human annotator, and only the extracted a multimodal deep neural architecture for the joint classification features were stored. of age, gender, and organizational status of social media users. To analyze the existence of objects in the image and represent Their model was trained using data in 32 languages, including a user through these features, we considered a Boolean BoW Spanish. The method analyzes the description of a user and their model. In this model, the names of the objects found were the profile picture. We used the implementation of the method terms considered, and the value for an object was assigned as provided by the authors through a Python library named 1 if it was found on the picture and 0 if it was not found. Given M3-Inference [55]. The tool outputs scores for three gender the sparsity of the model, we only considered objects that categories—male, female, and organization—and four different appeared in at least five images, which led us to retain 20 age ranges. features of this type. Regarding the emotions’ features, we Before using the detection tool on all the users, to increase its assigned to a user (if there are faces on the picture) the emotion performance, and given the fact that the AN, treatment, and with the highest predictive score, and later, to define a score for recovered users are not organizations, we defined that only if a a group, we considered the ratio of users assigned to a given user had a score over 0.70, for the organization class, and if this emotion. The same approach was considered for the technical value was higher than the scores for males and females, then features and objects detected. this label would be assigned; otherwise, the maximum value Pictures Shared among the male and female scores was considered. In addition, We explored the pictures shared by users through their if the organization label was assigned to a user, we automatically individual posts to detect AN-related images. For this purpose, assigned a specific age group (classified as an organization) we built two models trained on the images shared by users from for all the users classified as organizations. We evaluated the three of our groups of interest: AN, focused control, and random performance of this approach on a group of manually labeled control. Two binary classification models were trained: (1) an users based on their translated descriptions, where we considered AN versus focused control image detection model and (2) an up to 50 users per group. We obtained a macroaverage accuracy AN versus random control image detection model. To train the of 0.84 for all the gender groups of all the classes and a classifiers, the images of a given user were assigned the same macroaverage accuracy of 0.80 for all the age groups of all the label of the user. The method used to build the model was the classes. one explained in the study by Rodriguez et al [60], which is a Visual Aspects method for the inference of personality using the OCEAN model, which refers to the big five personality traits: openness Overview to experience, conscientiousness, extraversion, agreeableness, In this section, we describe the use of features extracted from and neuroticism. Moreover, this method has been previously the profile pictures of users and from the images of posts shared applied to the detection of suicidal ideation–related images [13]. by users (RQ1 and RQ7). For the first case, we use pretrained The resulting model was applied to all the images of a set of models provided by external sources; for the second case, we users kept for validation purposes, which were not considered train a model on our data set and apply it to the pictures shared in the training process. The output of each model was a score by a set of validation users from our groups of interest. We for each class to predict. For a single user, the score explore visual aspects given that there are physical traits that corresponding to this feature is given by the average score characterize AN [1], and related work has found visual patterns obtained by the classifier for the AN class over all the user’s that characterize similar mental conditions [12,13]. images. For the model generation, a total of 64,615 images were Profile Picture used for training and 15,384 for validation. For the first model (AN vs focused control), 278 users were considered for training We analyzed 32 features extracted from the pictures of users. and 130 for validation purposes. For the second model (AN vs We explored the technical features and the detection of emotions https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 11 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al random control), 240 users were considered for training and From a general perspective, the results suggest the relevance of 106 for validation purposes. the analysis of the content (texts) generated by users, for which more attributes with significant differences among the groups The image classifiers consisted of a CNN [61], and we used were identified. This is consistent with the fact that most of the residual networks [62], which are popular for improving the features map aspects considered by psychologists (eg, emotions, back-propagation step in deep architectures. For both classifiers, risk factors, and the use of vocabulary that indicates harmful we considered a model that was initially trained on ImageNet behaviors). Regardless of this, the findings on the social network [63] and fine-tuned on the target data set (our training images). of users also revealed interesting aspects, such as the evident As in the study by Ramírez-Cifuentes et al [13], we use a polarization between the focused control users and AN users. 101-layer ResneXt [64] that uses grouped convolution, an We also found a high similarity between the interests of AN architecture that consists of convolution groups of size 32 with users and their followees. In addition, images are relevant for a dimensionality of 8 and a fully connected layer at the end, distinguishing AN cases from control cases. Finally, we which is the one that performs the classification between our identified certain elements that distinguish AN users from two defined classes: (1) anorexia related and (2) control image treatment and fully recovered users. (focused or random control depending on the classifier being trained). The training process was performed for eight epochs Content Shared and Interests with stochastic gradient descent with warm restarts [65]. The learning rate was 0.0001, and the weight decay value was 0.001. Overview We used Nesterov with a momentum of 0.99 on two GTX 1080 The results found for this perspective were the most relevant Ti graphic cards. We also used dropout (50%) to avoid for characterizing AN users (RQ1, RQ3, RQ4, and RQ5). This overfitting. perspective explored the textual elements from multiple points of view, including linguistic and psychological factors that were Detection of Precontemplation and Contemplation particularly useful in distinguishing AN users from control Phases groups. These elements were also important for comparing According to the labeling approach, within the AN group, we between our control groups, which were thought to exclusively considered users who were probably experiencing AN and did differ from each other through the use of anorexia-related terms not receive treatment. Among these users, we found those in (RQ7). For the majority of the features analyzed for these the precontemplation phase and those at a more advanced stage perspectives, we calculated their median values for each group that might have been willing to start the recovery process among all its users. P values were also obtained to compare (contemplation phase). To detect the elements that distinguish among the following pairs of groups: AN versus treatment, AN users at the precontemplation stage from those at the versus recovered, AN versus random control, AN versus focused contemplation stage, we used a clustering approach. The features control, and random versus focused control. selected were those for characterizing users based on their Linguistic Dimensions emotions (using the EmoLex features) and that provide an insight on the positivity and negativity of posts as well. We also The results for the 24 linguistic dimension features explored considered the polarity score, given by the Senti-py library, are listed in Table 4. We observed many linguistic features that which classifies a text in the polarity range [0,1]. The use of could distinguish AN users from both the control groups. these features was based on the assumption that people in the Notably, the use of first-person singular verbs, and consequently precontemplation stage are more positive toward their weight first-person singular pronouns, characterized the posts of AN loss and the social support they receive. On the other hand, users, along with a high use of negations and a reduced use of people in the contemplation phase are at a stage where more articles. In addition, there were more features with highly negative emotions are manifested due to frustration and lack of significant differences between the AN group and the focused control over the weight loss process, which leads to isolation control group (features: 22/24, 92%) than between the AN group and the affectation of their cognitive capabilities. We used a and the random control group (features: 15/24, 62%). This can k-means clustering algorithm for the detection of these be explained by the fact that, as shown in our further analysis, communities. Later, we explored the differences among these a high percentage of focused control users were organizations groups using the values of their features. (eg, news sites, nutrition, and medical centers), and their linguistic features were quite distinguishable from those of users Results with personal accounts. This can also be noticed on the elements that distinguish between random and focused control users, as Comparative Results of Each Perspective more personal accounts were part of the random control group. Overview Regarding the differences between the AN group and users in treatment, we observed significant differences in the use of In this section, we present the results obtained for all the features second-person and first-person plural pronouns, which suggests analyzed and explore the differences among the values of these that there might be a change in their attention focus and a higher features for each defined group. We also explored certain level of interaction and inclusion with other people. This pattern elements within each perspective, reporting on the topics of was even more evident among recovered users. interest within each group, the generation of the users’ social network, and the analysis of features generated using classification models. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 12 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 4. Comparative analysis among groups based on linguistic dimensions. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random dian ment, me- ered, me- control, control, treatment, P recovered, P random focused versus fo- dian dian median median valueb valueb control, P control, P cused con- valueb valueb trol, P val- ueb First-person sin- 22.2E-03 20.5E-03 17.9E-03 8.14E-03 6.52E-03 .09 <.001c <.001c <.001c .002d gular verbs First-person plu- 9.79E-04 13.8E-04 20.2E-04 17.6E-04 15.5E-04 .01e <.001c <.001c <.001c .03e ral verbs Second-person 3.01E-03 2.90E-03 4.25E-03 2.38E-03 1.99E-03 .34 .008d .001d <.001c .01e singular verbs Third-person 2.26E-02 2.31E-02 2.50E-02 2.13E-02 1.73E-02 .26 .008d .02e <.001c <.001c verbs Third-person plu- 4.99E-03 4.87E-03 5.36E-03 4.45E-03 4.05E-03 .39 .34 .003d <.001c .08 ral verbs First-person sin- 41.9E-03 41.5E-03 3.0.5E-03 10.1E-03 5.62E-03 .30 <.001c <.001c <.001c <.001c gular pronouns First-person plu- 1.83E-03 2.26E-03 3.75E-03 3.72E-03 3.46E-03 .04e <.001c <.001c <.001c .22 ral pronouns Second-person 9.67E-03 10.6E-03 12.1E-03 8.92E-03 6.52E-03 .16 .003d .24 <.001c <.001c singular pronouns Second-person 2.60E-03 3.38E-03 4.69E-03 2.69E-03 3.12E-03 .004d <.001c .28 <.001c .002d plural pronouns Third-person sin- 3.80E-02 3.93E-02 4.24E-02 4.21E-02 3.91E-02 .05 .005d <.001c .01e .001d gular pronouns Third-person plu- 9.27E-03 9.51E-03 10.6E-03 11.8E-03 10.1E-03 .31 .047e <.001c .06 <.001c ral pronouns Negations 2.66E-02 2.44E-02 2.55E-02 1.98E-02 1.42E-02 .10 .28 <.001c <.001c <.001c Affirmations 7.14E-03 6.46E-03 7.60E-03 7.04E-03 4.50E-03 .33 .24 .19 <.001c <.001c Adverbs 5.24E-02 4.79E-02 4.83E-02 3.81E-02 2.75E-02 .01e .003d <.001c <.001c <.001c Articles 5.67E-02 6.15E-02 6.69E-02 7.06E-02 6.92E-02 .03e <.001c <.001c <.001c .46 Verbs 1.72E-01 1.63E-01 1.71E-01 1.42E-01 1.37E-01 .07 .10 <.001c <.001c .001d Impersonal pro- 10.9E-02 10.3E-02 10.8E-02 8.28E-02 6.64E-02 .14 .21 <.001c <.001c <.001c nouns Personal pro- 8.91E-02 9.00E-02 9.00E-02 6.66E-02 5.61E-02 .35 .26 <.001c <.001c <.001c nouns Total pronouns 1.98E-01 1.94E-01 1.94E-01 1.49E-01 1.22E-01 .22 .18 <.001c <.001c <.001c Prepositions 9.81E-02 9.71E-02 10.1E-02 9.97E-02 10.5E-02 .28 .16 .08 <.001c <.001c Past verb tense 1.84E-02 1.88E-02 1.69E-02 1.39E-02 1.18E-02 .43 .11 <.001c <.001c <.001c Present verb 1.27E-01 1.22E-01 1.22E-01 1.05E-01 1.00E-01 .13 .17 <.001c <.001c .001d tense Future verb tense 4.50E-05 7.50E-05 6.70E-05 12.1E-05 12.8E-05 .15 .42 .01e <.001c .07 Median tweet 14.00 14.00 13.50 12.50 19.00 .45 .43 .001d <.001c <.001c length a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.001. d P<.01. e P<.05. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 13 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Affective Processes and Emotions less negative emotions and more positive emotions than AN The results of these features are described in Table 5. As for users. In fact, the expressions of anxiety of recovered users were the linguistic dimensions, there were significant differences significantly lower than those of AN users. In addition, their between the values of users of the AN and focused control high score on social processes and the highly significant values groups. Negative emotions are found mainly for AN and in comparison with AN users suggest an openness to more treatment users; this can be observed also on the expression of interactions with other people. Finally, the differences between emotions such as sadness, disgust, and anger, which are random and focused control users are mainly observed through significantly higher for AN users than for control users. Within the use of swearing terms, the expression of positive emotions, this same comparison, users with AN use more swearing terms cause and effects, insight, and discrepancies. In this sense, and vocabulary that express anxiety and thoughts on their focused control users seem to be more formal and analytic feelings and perceptions. We observe that there are a few toward things, which meets the characteristics of accounts that attributes with significant differences between AN and treatment represent organizations. For all the groups analyzed, we can users. For joy and positive emotions (LIWC), the scores were observe in Figure 2 a radar chart that expresses the median significantly higher for treatment users, which might reflect an values for the eight basic emotions defined by the wheel of improvement in the mood of people as they recover from AN. emotions by Plutchik [66]. We observed the predominance of Regarding recovered users, we also observed the existence of sadness over all the other emotions in AN users. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 14 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 5. Comparative analysis among groups based on effective processes and emotions. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- dian ment, me- ered, me- control, control, treatment, recovered, random focused sus focused dian dian median median P valueb P valueb control, P control, P control, P valueb valueb valueb Swearing 15.4E-03 14.6E-03 12.7E-03 8.06E-03 4.33E-03 .48 .19 <.001c <.001c <.001c Absolutist terms 5.36E-03 5.03E-03 5.09E-03 4.98E-03 3.43E-03 .28 .49 .046d <.001c <.001c Positive emotions 5.93E-02 6.23E-02 6.50E-02 6.58E-02 5.79E-02 .04d .03d <.001c .14 <.001c (Linguistic Inquiry and Word Count) Negative emotions 6.88E-02 6.50E-02 6.02E-02 4.97E-02 4.44E-02 .24 .01d <.001c <.001c .01d (Linguistic Inquiry and Word Count) Anxiety 11.2E-03 12.2E-03 8.91E-03 6.61E-03 6.42E-03 .14 .004e <.001c <.001c .36 Cognitive process- 2.73E-01 2.62E-01 2.64E-01 2.23E-01 2.19E-01 .09 .25 <.001c <.001c .25 es Cause and effect 1.79E-02 1.72E-02 1.44E-02 1.35E-02 1.62E-02 .13 .03d <.001c .11 <.001c Insight 3.89E-02 3.73E-02 3.72E-02 3.16E-02 3.60E-02 .30 .21 <.001c .003e <.001c Discrepancies 3.99E-02 3.72E-02 3.94E-02 3.03E-02 2.54E-02 .04d .18 <.001c <.001c <.001c Tentative 4.42E-02 4.53E-02 4.33E-02 3.54E-02 3.49E-02 .20 .46 <.001c <.001c .29 Certainty 1.79E-02 1.74E-02 1.99E-02 1.90E-02 1.56E-02 .38 .05 .27 <.001c <.001c Senses and percep- 5.03E-02 5.01E-02 4.89E-02 3.70E-02 3.80E-02 .48 .24 <.001c <.001c .20 tions See 1.23E-02 1.19E-02 1.25E-02 1.06E-02 1.11E-02 .30 .16 .02d .06 .19 Listen 10.6E-03 11.5E-03 13.4E-03 8.87E-03 7.19E-03 .04d .002e <.001c <.001c .03d Feel 16.0E-03 17.0E-03 13.7E-03 8.80E-03 9.30E-03 .42 .01d <.001c <.001c .28 Social processes 1.21E-01 1.21E-01 1.40E-01 1.23E-01 1.14E-01 .33 <.001c .05 .18 .01d References to 5.50E-03 4.39E-03 8.09E-03 5.06E-03 4.69E-03 .10 .02d .15 .003e .04d friends References to fami- 7.97E-03 7.61E-03 9.57E-03 7.93E-03 7.25E-03 .36 .05 .42 .02d .046d ly Joy 1.44E-02 1.55E-02 1.50E-02 1.35E-02 1.38E-02 .02d .29 .04d .02d .42 Trust 1.90E-02 1.92E-02 1.99E-02 2.27E-02 2.25E-02 .09 .09 <.001c <.001c .42 Fear 1.70E-02 1.76E-02 1.48E-02 1.57E-02 1.59E-02 .25 .07 .008e .02d .25 Surprise 9.36E-03 9.90E-03 9.50E-03 9.07E-03 8.75E-03 .33 .49 .06 .003e .24 Sadness 2.43E-02 2.53E-02 2.24E-02 1.85E-02 1.78E-02 .30 .02d <.001c <.001c .24 Disgust 1.58E-02 1.65E-02 1.41E-02 1.27E-02 1.10E-02 .48 .08 <.001c <.001c .003e Anger 1.51E-02 1.64E-02 1.42E-02 1.39E-02 1.23E-02 .11 .15 .004e <.001c .02d Anticipation 1.48E-02 1.63E-02 1.46E-02 1.51E-02 1.59E-02 .03d .50 .48 .01d .02d Polarity median 1.86E-01 1.77E-01 1.87E-01 1.94E-01 1.74E-01 .39 .49 .33 .05 .03d score Positive emotions 3.60E-02 3.73E-02 3.62E-02 3.82E-02 3.97E-02 .07 .32 .046d <.001c .03d EmoLex Negative emotions 3.60E-02 3.72E-02 3.37E-02 3.34E-02 3.02E-02 .42 .07 <.001c <.001c .03d EmoLex https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 15 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.001. d P<.05. e P<.01. Figure 2. Comparative scores for emotions according to the wheel of emotions by Plutchik. Left: anorexia nervosa and control groups. Right: anorexia nervosa, treatment, and recovered groups. comparison of AN and recovered users, where there are very Personal Concerns and Biological Processes significant differences among the same features. Note that the The results obtained for these features are listed in Table 6. We reference to religious aspects is lower for the AN, treatment, observe that most of these features are relevant for distinguishing and recovered users in comparison with random control users. control users from AN users. Control users discuss more about Regarding random and focused control users, there are common concerns such as work, leisure, achievement, money, differences in the scores for the body, ingest, health, and and religion, whereas AN users are more interested in aspects biological process categories, as these are the ones that refer to related to their image, which can be seen by their scores on the signs of the illness. Focused control users are characterized by body, ingest, health, and biological process categories. There their use of AN-related terms, and these findings suggest that was also a significantly higher interest in death, compared with among the focused control users, we can find people and all other categories. For the treatment group, we observed organizations that often address the topic of AN. Among these, significantly lower values for the ingest and biological process we can find foundations, medical centers, nutritionists, and categories, which might be a sign of improvement in their psychologists. We later validated this assumption through a condition compared with AN users. This is more evident in the social network analysis. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 16 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 6. Comparative analysis among groups based on personal concerns and biological processes. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random versus dian ment, me- ered, me- control, control, treatment, P recovered, P random focused focused control, dian dian median median valueb valueb control, P control, P P valueb b b value value Work 3.21E-02 3.50E-02 3.64E-02 4.95E-02 5.25E-02 .05 .01c <.001 d <.001d .09 Achieve- 3.87E-02 4.08E-02 3.92E-02 4.39E-02 4.30E-02 .07 .48 <.001d <.001d .18 ment Leisure 1.70e-02 1.91e-02 1.67e-02 1.99e-02 2.12e-02 .16 .40 <.001d <.001d .27 Home 5.52e-03 5.26e-03 6.38e-03 5.20e-03 5.43e-03 .28 .19 .02c .37 .02c Money 8.22E-03 9.42E-03 11.7E-03 13.8E-03 12.2E-03 .03c .006e <.001d <.001d .003e Religion 2.16E-03 2.67E-03 3.42E-03 5.17E-03 3.38E-03 .10 <.001d <.001d <.001d <.001d Sexual 10.6E-03 11.6E-03 15.6E-03 7.98E-03 7.70E-03 .39 .02c .001e <.001d .22 Death 10.5E-03 8.33E-03 6.50E-03 5.87E-03 6.24E-03 .047c <.001d <.001d <.001d .25 Biological 9.02E-02 6.86E-02 6.41E-02 3.21E-02 5.01E-02 .04c .003e <.001d <.001d <.001d processes Body 2.96E-02 2.45E-02 1.92E-02 1.21E-02 1.55E-02 .07 <.001d <.001d <.001d <.001d Ingest 33.8E-03 17.4E-03 15.3E-03 8.52E-03 11.2E-03 .01c .001e <.001d <.001d <.001d Health 17.7E-03 17.5E-03 17.6E-03 6.83E-03 13.2E-03 .50 .47 <.001d <.001d <.001d a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.05. d P<.001. e P<.01. percentage for recovered users than for AN users and treatment Risk Factors Vocabulary users. We observe that the use of terms related to bullying is For the use of vocabulary related to risk factors, we noticed that higher for treatment and recovered users, which can be explained a large number of features were highly significant for the by the fact that while being on treatment and after recovery, comparison of the AN and control groups. In fact, all the risk patients are more likely to recognize the issues behind their ED. factors considered were significant for distinguishing AN from In general, the scores obtained by all the groups for these random control users, as shown in Table 7. The use of features are very low, as these are issues that do not seem to be suicide-related terms is higher for AN users than for all the other openly addressed often. groups. Hate and self-loathing terms are found in a lower https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 17 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 7. Comparative analysis among groups based on vocabulary related to risk factors. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- dian ment, me- ered, me- control, control, treatment, recovered, P random focused sus focused dian dian median median P valueb valueb control, P control, P control, P valueb valueb valueb Hate 98.4E-05 68.5E-05 33.1E-05 7.60E-05 2.90E-05 .15 <.001c <.001c <.001c .03d Suicide-related 4.20E-05 2.20E-05 0 0 0 .04e .002e <.001c <.001c <.001c terms Self-harm 1.60E-05 4.60E-05 1.00E-05 0 0 .03e .21 <.001c <.001c .02e Work or school 8.80E-05 12.0E-05 6.80E-05 1.60E-05 3.10E-05 .41 .16 <.001c <.001c .12 problems Self-loathing 4.20E-05 1.90E-05 0 0 0 .25 .003d <.001c <.001c .007d Bullying 0 3.00E-06 11.0E-06 0 0 .06 .03e <.001c .02e <.001c Drugs or alcohol 125E-06 143E-06 77.0E-06 6.00E-06 124E-06 .10 .30 <.001c .40 <.001c abuse Lack of social 0 2.00E-06 0 0 0 .25 .29 <.001c <.001c .01e support Relationship is- 5.80E-05 7.00E-05 7.30E-05 0 1.50E-05 .14 .33 <.001c <.001c .001d sues Use of antidepres- 0 0 0 0 0 .36 .34 <.001c <.001c <.001c sants a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.001. d P<.05. e P<.01. recovered and AN users are compared; these users highly differ Anorexia-Related Vocabulary in the use of vocabulary dedicated to the promotion of AN and These features address the use of vocabulary that describes vocabulary that expresses concerns regarding body image, body certain signs and symptoms of AN. The results are presented weight, compensatory behavior, and laxatives references, along in Table 8. All the features are highly significant for with caloric restrictions. AN users showed higher scores on distinguishing AN users from control cases, and they are all these aspects. We also observed that users in the treatment group highly significant for distinguishing random from focused had lower median values for almost all the features considered, control cases. We observed that the scores obtained for the with significant differences in up to four features in comparison focused control cases were higher than the scores obtained for with the AN group. the random control users. This also happens for the case where https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 18 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 8. Comparative analysis among groups based on anorexia-related vocabulary. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- dian ment, me- ered, me- control, control, treatment, recovered, P random focused sus focused dian dian median median P valueb valueb control, P control, P control, P valueb valueb valueb Anorexia promo- 35.4E-04 23.5E-04 13.0E-04 4.99E-04 8.47E-04 .02c <.001d <.001d <.001d <.001d tion Body image 23.5E-04 7.01E-04 4.26E-04 0 1.60E-04 .01c <.001d <.001d <.001d <.001d Body weight 75.7E-05 32.9E-05 9.00E-05 0 8.90E-05 .11 <.001d <.001d <.001d <.001d Food and meals 29.5E-04 21.5E-04 16.8E-04 1.76E-04 5.20E-04 .16 .02c <.001d <.001d <.001d “Eat” as verb 222E-06 100E-06 91.0E-06 0 9.00E-06 .01c .001e <.001d <.001d <.001d Caloric restric- 443E-06 34.0E-06 2.00E-06 0 0 .001e <.001d <.001d <.001d <.001d tion Binge eating 3.10E-05 3.40E-05 0 0 0 .40 .004e <.001d <.001d <.001d Compensatory 9.00E-04 4.88E-04 2.71E-04 0 0 .09 <.001d <.001d <.001d <.001d behavior and lax- atives Exercise 164E-05 91.8E-05 45.2E-05 7.30E-05 43.7E-05 .05 .001e <.001d <.001d <.001d a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.05. d P<.001. e P<.01. significantly different values (P<.05) between the AN group Topics of Interest and the focused control, treatment, and recovered groups. We In this section, we present the results for the exploration of the observe that swearing terms, feminine terms, hate, pain, and topics of interest of users that make use of anorexia-related appearance obtained high scores for AN users, whereas topics terms (RQ3), which include AN, treatment, recovered, and such as economics, college, photography, and work obtained focused control users. We assume that the interests of random high scores for focused control users. We also observed a limited control users are different and depend on the user. This is due interest in topics such as music and art for AN users, in to the fact that we do not consider common interest for these comparison with users in treatment, whereas these users users during the data collection process. (treatment) are less concerned about body and exercise in Table 9 shows the top 20 topics of interest for the groups comparison with users from the AN group. Recovered users are according to the Empath categories. We observe that, apart from also more concerned about general topics such as law, crime, the elements in common among groups, only users of the AN and politics in comparison with AN users. We report on the group refer to topics such as pain, eating, violence, and percentage of topics found with significant differences among suffering. Treatment users have many interests in common with the values for each group (P<.05): AN versus focused control, AN users, but we can also observe other topics of interest such 61% (122/200); AN versus recovered, 40% (80/200); and AN as reading, music, and sports. Similarly, recovered users rank versus treatment, 46.5% (93/200). We also calculated the values topics such as sports and weddings in their list. Focused control for Spearman rank correlation coefficient based on the median users also express interest on different topics, with the highest values obtained for each topic in each group. The following scored topics being health, communication, business, work, pairs of groups were compared: AN versus recovered (ρ=0.87), internet, and sports, which matches with our prior assumptions AN versus treatment (ρ=0.87), AN versus focused control regarding this group. Note that for visualization purposes, the (ρ=0.67), treatment versus focused control (ρ=0.87), recovered actual median values were multiplied by 1000. versus focused control (ρ=0.87), and treatment versus recovered (ρ=0.97). We observe that AN and focused control users are To explore the topics of interest in which the groups differed less interested in similar topics, whereas treatment and recovered the most from the AN group, we performed the Mann-Whitney users’ interests are more correlated with those of the focused U test. Figure 3 shows the top 20 topics with the most control group. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 19 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 9. Top 20 topics of interest (using Empath) among groups that use anorexia-related vocabulary and their median values. Groups and topics Value, median Anorexia nervosa Negative emotion 15.86 Friends 8.24 Speaking 7.67 Positive emotion 7.27 Children 6.41 Pain 6.29 Eating 6.19 Communication 6.13 Optimism 5.93 Family 5.91 Love 5.60 Shame 5.40 Violence 5.21 Party 4.84 Social media 4.70 Suffering 4.25 Home 4.24 Hate 4.08 Childish 4.06 Feminine 3.87 Treatment Negative emotion 10.50 Friends 7.28 Positive emotion 6.95 Speaking 6.74 Social media 6.62 Children 6.55 Communication 6.06 Optimism 5.61 Family 5.23 Party 4.85 Love 4.80 Reading 4.51 Music 4.22 Home 4.10 Internet 4.09 Musical 3.99 Listen 3.94 Wedding 3.79 Violence 3.63 Sports 3.62 https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 20 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Groups and topics Value, median Recovered Negative emotion 10.85 Friends 9.78 Speaking 8.09 Positive emotion 7.60 Communication 7.01 Children 6.66 Family 6.46 Social media 5.62 Home 4.95 Party 4.95 Optimism 4.91 Love 4.91 Eating 4.31 Wedding 3.95 Sports 3.94 Giving 3.94 Violence 3.89 Childish 3.75 Pain 3.75 Affection 3.71 Focused control Health 9.39 Communication 7.41 Business 6.90 Work 5.98 Positive emotion 5.88 Internet 5.83 Negative emotion 5.82 Social media 5.71 Speaking 5.54 Sports 5.44 Messaging 5.03 College 5.01 Eating 4.92 Children 4.86 School 4.84 Family 4.59 Reading 4.41 Party 4.30 Optimism 4.28 Meeting 4.26 https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 21 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Figure 3. Top 20 topics with most significantly different values (P<.05) between the anorexia nervosa group and the focused control, treatment, and recovered groups. The median values for each feature are shown. Proportion of AN-Related Tweets Social Network Regarding RQ5, the results obtained by the LR classifier at the Overview training approach (cross validation) were as follows: F1 Our findings from this perspective (RQ2 and RQ3) provide an score=0.97, precision=0.98, and recall=0.97, whereas the results insight into the users that are part of the focused control group for the deep learning approach averaged after multiple runs over and verify that organizations and specialists are part of it. We the test set were as follows: precision=0.98, recall=0.98, and also observed that the structure of the social network could tell F1 score=0.98. As the second model obtained slightly better us about the different types of accounts that make use of results, it was applied to all the tweets of all the users regardless AN-related terms, in particular, through the use of clustering of the group they belonged to. For each user, the value approaches. Our results also showed a high polarization between considered as a feature was the median score obtained by the focused control users and AN users, which was verified by the classifier on all tweets. We then compared the median values analysis of the topics of interest between the followees of each of each group analyzed. user group, where there were significant differences in 75% We used the Mann-Whitney U test to perform an analysis of (150/200) of the analyzed topics. the median score provided by the classifier to all the users’ Measures of Interactions and Engagement tweets. We applied the classifier to all groups of users. The The results obtained for these features (Table 10) show that median values for each group are the following: AN (0.23), focused control users have a significantly higher median number treatment (0.13), recovered (0.08), random control (0.03), and of followers and followees than AN users. The median number focused control (0.05). The P values for the group comparisons of followers of these users (focused control) shows that these are the following: AN versus recovered (P<.001), AN versus accounts have a higher number of followers than random control treatment (P=.004), AN versus focused control (P<.001), users, which might be an indicator of the popularity of these treatment versus focused control (P<.001), and focused control user types that are more likely to be organizations. We also versus random control (P=.02). We noticed very significant observe that AN users have a reduced number of interactions differences between the AN group and all other groups with other users in comparison with treatment, recovered, and considered. Notably, the median classifier score obtained by random control users (based on the favorites given). In general, AN users was higher than that obtained by users from all other we observe that a reduced number of tweets generated by all classes. Moreover, the median values for the groups decreased user groups are liked or retweeted by other users, probably according to the recovery stage, meaning that the score was because they consume this type of information in a discrete way lower for recovered users than for treatment users. Note that or because they do not generate very popular content. focused control users obtain a higher score than random control users, as focused control users address AN-related topics. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 22 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 10. Comparative analysis among groups based on interaction and engagement measures. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- dian ment, me- ered, me- control, control, treatment, recovered, P random focused sus focused dian dian median median P valueb valueb control, P control, P control, P valueb valueb valueb Number of fol- 621.50 815.00 600.00 540.00 1174.00 .02c .26 .26 <.001d <.001d lowers Number of fol- 286.50 483.50 289.50 492.00 509.00 .02c .18 <.001d <.001d .24 lowees Given favorites 7746.50 10,893.00 23,955.00 10,085.50 4917.00 .04c .004e .02c .02c <.001d Received fa- 0.00 1.00 0.50 0.00 1.00 .11 .22 .001e .001e <.001d vorites Received 0.00 0.00 0.00 0.00 0.00 .04c .29 .07 <.001d <.001d retweets a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.05. d P<.001. e P<.01. nodes in the network (full network diameter) was 15, and the Analysis of Followees and Community Detection average path length was 4.72, which represents the average As explained in the Methods section, we analyzed the structure number of steps it takes to get from one member of the network of the social network of users using AN-related vocabulary to another. The average clustering coefficient was 0.017, which (RQ2). In Table 11, we report on the percentages of nodes implies that most of the nodes were not related. To ease the belonging to each group defined through the approach visualization and interpretation of the results, we applied a previously explained in Table 3. Most of the users considered k-core filter with k=2 to see the maximal subgraph with a were part of the focused control group, followed by AN, minimum degree equivalent to k. The number of nodes displayed recovered, and treatment users. For visualization of these groups, in Figure 4 is 12,680, and the size of the nodes is given by the we used Gephi, as shown in Figure 4. A total of 99,283 nodes page rank score obtained by each node. The graph clearly shows were considered, with each node representing a user. The the polarization between the AN and focused control groups, average number of edges per node (average degree of the graph) with the few treatment and recovery cases displayed in between was 2.57, the shortest distance between the two most distant and closer to the focused control cases. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 23 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 11. Graph information of subgroups according to the type of followers. Validation group and subgroup Nodes user type Node, % ANa Gb1 Labeled as AN 0.05 G2 Mostly followed by G1 10.55 G3 Mostly followed by G2 12.46 Focused control G4 Labeled as focused control 0.16 G5 Mostly followed by G4 46.56 G6 Mostly followed G5 20.57 Treatment G7 Labeled as treatment 0.01 G8 Mostly followed by G7 3.95 G9 Mostly followed G8 0.74 Recovered G10 Labeled as recovered 0.02 G11 Mostly followed by G10 4.63 G12 Mostly followed G11 0.3 a AN: anorexia nervosa. b G: group. Figure 4. Visualization of the social network of the AN, focused control, treatment, and recovered groups according to the types of users they are mostly followed by. Each group is represented by a different color. Groups associated with the same class have similar colors. AN: anorexia nervosa; G: group ID. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 24 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al For further analysis of the full network, we applied a clustering correspond to focused control cases (GC2 and GC3) and are algorithm to detect the communities within it. We found 80 characterized by having users that represent organizations and communities and obtained a modularity value of 0.86. As shown specialists on mental health issues and nutrition centers. We in Table 12, we analyzed the descriptions (biographies) of the also observe a community that corresponds to news and TV users of the 10 communities with the highest node percentages. accounts (GC5), which also characterizes focused control users. We describe the types of users found in each community and We see that the small number of treatment and recovered users identify the types of users from our annotated groups that are are part of different communities that address multiple topics part of each community. Figure 5 shows the network, with the and that the communities that gather users from different groups automatically identified communities highlighted. For are those that involve singers, artists, influencers, and comparison, we used the same structure displayed in Figure 4. leisure-related topics. These results show that users at the It can be seen that the community with the highest number of precontemplation and contemplation stages are isolated from nodes is GC1, which corresponds to the community of users accounts that offer assistance to overcome the illness. In this that are likely to have an ED and users that might be anorexia sense, recommender systems might enforce this behavior of the and bulimia promoters (they correspond to big nodes in the network because they tend to recommend a user to follow graph, ie, higher page rank, meaning more popular nodes). We similar accounts. also observe two other relevant communities that mainly Table 12. Description of the types of users identified in each community with the highest node percentages. Community Description of the community Group users identified Node, % GCa1 Users with eating disorders and anorexia and bulimia promoters Anorexia nervosa 9.65 GC2 Organizations, medical centers, and psychologists Focused control 8.04 GC3 Nutritionists and nutrition centers Focused control 7.37 GC4 Varieties: influencers Focused control 3.88 GC5 News and television Focused control 3.72 GC6 Fans of pop singers AN and recovered 2.69 GC7 Undefined varieties Treatment and focused control 2.54 GC8 Undefined varieties Recovered and AN 2.53 GC9 Comics, anime, and drawing Treatment and focused control 2.31 GC10 Uruguayan community Focused control 2.29 a GC: group community identifier. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 25 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Figure 5. Visualization of the 10 communities with the highest node percentages. Each group corresponds to a community and is represented by a given color. GC: group community. differences among the median values for the following pairs of Analysis of Interests Between Users and Their Followees groups (P<.05): AN followees and recovered followees, 45% Regarding the results for the topics of interest of the users who (90/200); AN followees and focused control followees, 75% make use of anorexia-related terms (RQ3), Table 13 shows the (150/200); AN followees and treatment followees, 48% top 20 topics of interest for our groups’ followees according to (96/200); AN and AN followees, 22% (44/200); and recovered the Empath categories. We observe that negative emotions, and recovered followees, 21% (42/200). From this, we observe eating, pain, death, and violence are among the topics most that AN users and recovered users differ the least in their relevant to AN followees. Regarding the other groups, we cannot interests with their own followees. AN followees and focused observe a pattern that would normally characterize each user control followees show the biggest difference in interests. We type; instead, we observe interest in all types of topics, which also calculated the values for Spearman rank correlation is more evident in focused control users. We can also observe coefficient based on the median values obtained for each topic that topics such as friends, family, children, and parties are in each group. The pairs of groups were as follows: AN and AN relevant for most of the groups. followees (ρ=0.96), treatment and treatment followees (ρ=0.97), For a better comprehension of the results on this topic analysis recovered and recovered followees (ρ=0.96), focused control task, we explored the topics in which certain followee groups and focused control followees (ρ=0.97), AN followees and differ the most. We used the Mann-Whitney U test for this treatment followees (ρ=0.93), AN followees and recovered purpose. Figure 6 shows the top 20 topics, with the most followees (ρ=0.93), AN followees and focused control followees significantly different values (P<.05) between the AN followees (ρ=0.69), treatment followees and focused control followees group and the focused control, treatment, and recovered followee (ρ=0.86), and recovered followees and focused control followees groups. We see a very high value for negative emotion on AN (ρ=0.86). From these results, we can say that for all the groups, followees in comparison with focused control followees. their interests are highly similar to those of their followees; Appearance is also a topic in which AN followees differ from however, the interests of the treatment and recovered followees focused control followees and recovered followees. We also groups are more highly correlated to the focused control group report on the percentage of topics found with significant followees than the AN followees group, indicating a change in interest through the evolution of the disorder. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 26 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 13. Top 20 topics of interest and their Empath median values for the groups that use anorexia-related vocabulary followees. Groups and topics Value, median Anorexia nervosa Negative emotion 10.86 Friends 7.58 Positive emotion 7.30 Speaking 6.50 Communication 5.93 Optimism 5.85 Children 5.80 Social media 5.71 Party 5.62 Love 5.09 Family 4.69 Childish 4.03 Giving 4.00 Eating 3.93 Home 3.83 Pain 3.72 Death 3.72 Wedding 3.70 Violence 3.62 Celebration 3.42 Treatment Negative emotion 7.92 Friends 7.76 Positive emotion 7.02 Communication 6.96 Social media 6.83 Speaking 6.42 Children 5.48 Party 5.43 Optimism 5.22 Family 4.82 Love 4.28 Internet 4.15 Music 4.09 Messaging 4.06 Listen 4.02 Musical 3.92 Reading 3.91 Wedding 3.76 Celebration 3.74 Childish 3.72 https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 27 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Groups and topics Value, median Recovered Negative emotion 9.73 Friends 7.71 Positive emotion 7.23 Communication 7.15 Speaking 6.87 Social media 5.69 Optimism 5.55 Party 5.22 Children 5.20 Family 4.98 Internet 4.33 Giving 4.00 Love 3.96 Messaging 3.83 Reading 3.73 Wedding 3.70 Celebration 3.67 Listen 3.57 Home 3.52 Childish 3.47 Focused control Business 7.29 Communication 7.04 Work 6.80 Positive emotion 6.41 Internet 5.72 Social media 5.36 Party 4.86 Meeting 4.85 Speaking 4.79 Negative emotion 4.66 Leader 4.65 Reading 4.60 School 4.58 Messaging 4.57 Children 4.56 Occupation 4.50 Family 4.49 Optimism 4.35 Government 4.31 Celebration 4.12 https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 28 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Figure 6. The top 20 topics with most significantly different values (P<.05) between the anorexia nervosa followees group and the focused control, recovered, and treatment followees groups. The median values for each feature are shown. in EDs and other associated mental issues, such as depression. Behavioral Aspects Regarding the tweeting periods during the year, we see that The results of the behavioral aspects (RQ1) analyzed (Table between December and February (winter in Europe and summer 14) showed that AN users tweeted more on weekends compared in most countries of South America) users from the AN group with control groups. In addition, the median time between tweets tweeted less than users from all the other groups. However, we was lower for AN users (they tweeted more frequently) in cannot match this finding to a clinical fact related to the seasons comparison with random and focused control users. We also of the year, given the lack of information regarding the users’ observed that the tweeting ratio during sleeping periods was location. Figure 7 shows the comparative box plot of the values significantly higher for AN users than for the control groups. for each group on the sleep period tweeting ratio and the This might indicate some sleep alteration, which is a usual sign weekend tweet count ratio. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 29 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 14. Comparative analysis among groups based on behavioral aspects. Features ANa, me- Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- dian ment, me- ered, me- control, control, treatment, recovered, P random focused sus focused dian dian median median P valueb valueb control, P control, P control, P valueb valueb valueb Working week 0.73 0.73 0.75 0.75 0.75 .37 .15 .03c <.001d .01c tweet count ratio Weekend tweet 0.27 0.27 0.25 0.25 0.25 .37 .15 .03c <.001d .01c count ratio Median time be- 625.25 701.00 1187.50 4063.75 1088.00 .43 .15 <.001d .005e <.001d tween tweets Sleep period 0.05 0.05 0.04 0.04 0.03 .30 .07 <.001d <.001d .001e tweeting ratio Awake period 0.95 0.95 0.96 0.96 0.97 .30 .07 <.001d <.001d .001e tweeting ratio Normalized tweet 0.01 0.16 0.15 0.18 0.16 <.001d .02c <.001d <.001d .01c count per year quarter: Decem- ber to February Normalized tweet 0.01 0.23 0.21 0.23 0.21 .001e .11 <.001d .01c .001e count per year quarter: March to May Normalized tweet 0.27 0.32 0.27 0.25 0.24 .39 .29 .10 .004e .06 count per year quarter: June to August Normalized tweet 0.36 0.24 0.27 0.28 0.31 .002e .20 .001e .32 .002e count per year quarter: Septem- ber to November Number of tweets 7910.00 21,038.50 18,409.00 23,291.50 21,463.00 .004e .12 <.001d <.001d .49 created since the account creation a AN: anorexia nervosa. b P values were analyzed using Mann-Whitney U test. c P<.05. d P<.001. e P<.01. Figure 7. Comparative box plots of the sleep period tweeting ratio (left) and the weekend tweet count ratio (right) of each user group. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 30 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Demographics 15), we see differences between the AN and control groups due Using the age and gender inference tool, we obtained the to the number of female users. We also observed differences percentages of users corresponding to demographic categories between the focused control and random control groups, as there (RQ1) for each group (Figure 8). Most of the AN, treatment, were fewer organizations in the random control group. and recovered users were young women. These results are Regarding age (Table 16), we observe that AN users differ from compatible with the statistics that mention that the incidence the control groups because of the large number of AN users rates for AN are the highest for women aged 15-19 years [54]. aged ≤18 years. We also find differences between the AN and We also observe that a considerable number of users in the recovered groups, as recovered users are normally older than focused control group represent organizations. When comparing AN users. This is consistent with the fact that a full recovery the ratios of users belonging to each gender per group (Table process often takes years, and therefore, users get older as the recovery stages are reached. Figure 8. Composition of the anorexia nervosa, treatment, recovered, and control user groups according to gender and age. Each age and gender subgroup is represented by a color. Table 15. Comparative analysis among groups based on gender. Gender ANa versus treatment, AN versus recovered, P AN versus random con- AN versus focused Random versus focused b b b P value b value trol, P value control, P value control, P valueb Male .15 .14 <.001c .44 <.001c Female .15 .14 .04d <.001c .50 Organization N/Ae N/A <.001c <.001c <.001c a AN: anorexia nervosa. b P values were analyzed using proportions z test. c P<.001. d P<.05. e N/A: not applicable. Table 16. Comparative analysis among groups based on age, including users classified as organizations. Age groups, years ANa versus treatment, AN versus recovered, P AN versus random con- AN versus focused Random versus focused b b b P value b value trol, P value control, P value control, P valueb ≤18 .32 .01c .002d <.001e .78 19-29 .27 .04c .31 <.001e .09 30-39 .80 .06 <.001e .009d .02c ≥40 .50 .59 <.001e <.001e .003d a AN: anorexia nervosa. b P values were analyzed using proportions z test. c P<.05. d P<.01. e P<.001. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 31 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Visual Aspects observed that treatment users expressed more neutral emotions. The visual aspects analyzed address the profile picture of the We also notice that sadness characterizes AN users, which are user and the images shared on their posts (RQ1). We found the only ones with a ratio of users showing such expressions. significant differences between the groups’ profile pictures. On the objects detected (Table 19), there were significant Regarding the technical aspects (Table 17), we can observe that differences in the existence of clothing elements between the focused control users are likely to be distinguished from AN AN and control groups, along with the appearance of hands, users because of the presence of text in their profile pictures. shorts, and accessories, which might suggest that more full-body This also applies to random control users, who tend to use text pictures are shared by AN users, which might consequently as well but in a lower ratio than focused control users. These imply a higher interest in their appearance. We also observe findings can be explained by the use of logos in the profile that there is a high ratio of posters on the control users’ profiles, pictures of the accounts of organizations. In addition, AN users’ which validates our prior assumption about the representation pictures are significantly darker than focused control users’ of organizations. We also observed that few men were identified pictures. In terms of the emotions detected (Table 18), we on pictures of users of the AN group, whereas women were identified on more than half of the AN profile’s pictures. Table 17. Comparative analysis among groups based on technical aspects of profile pictures. Features ANa ratio Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- of users ment ra- ered ratio control ra- control ra- treatment, recovered, P random focused sus focused tio of of users tio of tio of P valueb valueb control, P control, P control, P users users users valueb valueb valueb Is gray scale 0.09 0.13 0.17 0.05 0.03 .78 .57 .31 .06 .43 Is lighter 0.50 0.63 0.50 0.42 0.65 .51 .99 .30 .03c <.001d Has text 0.06 0.13 0.00 0.22 0.34 .51 .53 .008e <.001d .045c Has faces 0.28 0.50 0.33 0.25 0.26 .21 .79 .66 .79 .82 a AN: anorexia nervosa. b P values were analyzed using proportions z test. c P<.05. d P<.001. e P<.01. Table 18. Comparative analysis among groups based on emotions detected from profile pictures. Features ANa ratio Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- of users ment ra- ered ratio control ra- control ra- treatment, recovered, P random focused sus focused tio of of users tio of tio of P valueb valueb control, P control, P control, P users users users valueb valueb valueb Neutral 0.00 0.38 0.00 0.05 0.09 <.001c N/Ad .06 .02e .34 Sad 0.09 0.00 0.00 0.00 0.00 .37 .43 .002f <.001c N/A Fear 0.05 0.00 0.17 0.03 0.02 .53 .23 .61 .20 .43 Surprise 0.02 0.00 0.00 0.00 0.01 .72 .76 .22 .61 .39 Angry 0.09 0.13 0.00 0.06 0.07 .78 .43 .46 .56 .83 Happy 0.03 0.00 0.17 0.08 0.09 .61 .12 .18 .12 .80 Disgust 0.00 0.00 0.00 0.03 0.02 N/A N/A .15 .32 .43 a AN: anorexia nervosa. b P values were analyzed using proportions z test. c P<.001. d N/A: not applicable. e P<.05. f P<.01. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 32 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Table 19. Comparative analysis among groups based on objects detected from profile pictures. Features ANa ratio Treat- Recov- Random Focused AN versus AN versus AN versus AN versus Random ver- of users ment ra- ered ratio control ra- control ra- treatment, recovered, P random focused sus focused tio of of users tio of tio of P valueb valueb control, P control, P control, P users users users valueb valueb valueb Poster 0.00 0.00 0.00 0.09 0.24 N/Ac N/A .01d <.001e .004f Clothing 0.75 1.00 0.50 0.53 0.33 .11 .19 .005f <.001e .003f Person 0.28 0.13 0.00 0.32 0.12 .34 .13 .58 .007f <.001e Man 0.05 0.00 0.00 0.32 0.16 .53 .59 <.001e .03d .003f Dress 0.03 0.00 0.17 0.01 0.02 .61 .12 .34 .74 .47 Boy 0.02 0.00 0.00 0.02 0.01 .72 .76 .81 .61 .40 Tree 0.02 0.00 0.17 0.05 0.03 .72 d .03 .23 .53 .43 Human hand 0.06 0.00 0.00 0.00 0.01 .47 .53 .01d .02d .39 Fashion accesso- 0.08 0.00 0.00 0.02 0.02 .41 .48 .08 .03d .77 ry Flower 0.03 .0.00 0.00 0.01 0.04 .61 .66 .34 .79 .19 Glasses 0.00 0.13 0.00 0.05 0.03 .004f N/A .06 .16 .43 Animal 0.00 0.00 0.00 0.02 0.02 N/A N/A .25 .22 .90 Shorts 0.06 0.13 0.00 0.00 0.00 .51 .53 .01 d .004 e N/A Jeans 0.03 0.00 0.00 0.02 0.01 .61 .66 .68 .21 .40 Human eye 0.06 0.00 0.00 0.01 0.01 .47 .53 .06 .02c .83 Cat 0.00 0.13 0.00 0.01 0.02 .004f N/A .41 .22 .47 Footwear 0.09 0.00 0.00 0.05 0.02 .37 .43 .31 .01d .12 Human nose 0.02 0.00 0.00 0.02 0.00 .72 .76 .81 .16 .10 Girl 0.19 0.25 0.17 0.0 0.03 .67 .90 .006 f <.001 e .43 Woman 0.58 0.75 0.67 0.18 0.22 .35 .67 <.001e <.001e .38 a AN: anorexia nervosa. b P values were analyzed using proportions z test. c N/A: not applicable. d P<.05. e P<.001. f P<.01. Finally, when it comes to the exploration of the pictures posted for random control users was 0.54. We observed lower by users, we developed two classifiers: (1) one trained with the aggregated scores for both control cases, meaning that these images of users from the AN group and users from the focused users share fewer AN-related pictures. These results show that control group and (2) another trained with the images of users the pictures are informative for the detection of users with AN. from the AN group and users from the random control group. Regarding the visual content of the images posted by users with The results showed that there were highly significant differences AN, there are some visual patterns that emerge after training between the AN and control groups (P<.001 for both CNNs, as described previously. These visual patterns were comparisons). The median of the aggregated scores of the first found to be quite characteristic in such posted images because classifier for the AN class (AN vs focused control) for a set of they received the highest scores by the CNN. As expected, most 130 validation users was 0.73, whereas the median value for images correspond to body objectification, that is, selfies, focused control users was 0.36. This means that a higher number extremely thin body parts (mostly legs), and (altered) images of pictures related to AN were found on the posts of AN users. of unrealistic ideals of body size. There are also several images When analyzing the median of the aggregated scores of the of scales showing (rather low) weight numbers. Finally, and second classifier (AN vs random control), on a set of 106 less frequently, we can identify rather healthy food images, validation users, the median value for AN users was 0.78 and mostly salads and fruits. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 33 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Detection of Precontemplation and Contemplation contemplation (0.1694) cluster. In fact, there were significant Phases differences (P<.001) in the values for all features considered in the clustering approach. In addition, the differences between After performing the clustering approach with k-means, we opposite (positive and negative) emotions such as joy and obtained two clusters of users among those labeled as AN sadness are lower for the values of the precontemplation cluster, (RQ6). We analyzed both groups and found strong indicators as the early stages of the illness are characterized by signs of of the groups identified as our contemplation (cluster 1) and the person feeling in control of the weight loss process and precontemplation (cluster 2) groups. This was mainly because being enthusiastic about their progress and the social support the precontemplation cluster was characterized by higher values received. The median values for all features analyzed for each of positivity compared with its negativity, whereas the opposite cluster are displayed in Table 20. Among the AN users, a total occurred for the contemplation cluster (Figure 9). In addition, of 115 users were assigned to the precontemplation cluster, and the median polarity score for the precontemplation cluster 56 users were assigned to the contemplation cluster. (0.1915) was significantly higher (P<.001) than that for the Figure 9. Polarity (calculated using Senti-py) and positivity and negativity (calculated using EmoLex) values for the detected precontemplation and contemplation clusters. Table 20. Median values of the features considered for the contemplation and precontemplation clusters. Emotion Cluster 1 (contemplation), median Cluster 2 (precontemplation), median Joy 15.92 13.85 Sadness 32.88 21.95 Anticipation 17.08 14.13 Surprise 11.49 8.72 Trust 20.36 18.30 Disgust 22.22 13.92 Anger 19.44 13.72 Fear 23.23 14.91 We can observe that the results obtained for this task are approach to address this classification task, which is relevant consistent with the characteristics of people going through each given that users at the contemplation stage are more prone to stage; however, there is no annotated information for the seek help, and therefore, different ways to approach each type validation of the classification performed, and this is an aspect of user may be considered. that needs to be addressed in the future. Ours is the first https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 34 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al high polarization noted among the AN- and focused Discussion control–related communities reinforces the findings of previous Principal Findings studies conducted on networks of English speakers [17,18], which reported limited interactions between pro-ED and This study analyzed the elements that characterize prorecovery communities (RQ4). From a psychological Spanish-speaking users with AN during the stages of their perspective, these findings can be explained by the elements recovery process. We have considered multiple points of view, that characterize people at the precontemplation stage according including linguistic, psychological, relational, behavioral, to the transtheoretical model of health behavior change, where demographical, and visual aspects, for our analysis. We believe people are in denial of their unhealthy conditions and tend to that this is the first approach to analyze both users in treatment feel supported by their equals (pro-ED community members) and those fully recovered in multiple perspectives. Furthermore, [4,5], resulting in a rejection of prorecovery content. Regarding we identified the topics of interest of different types of users the topics of interest of users and their followees (RQ3), we that used AN-related vocabulary, along with those of their found that the interests of AN users and their followees were followees. In this section, we report the main findings according highly correlated (ρ=0.96). We also observed a higher to the RQs addressed. correlation between the treatment followees and focused control With regard to RQ1, we found multiple elements that followee groups (ρ=0.86) and the recovered followees versus characterize and distinguish users with AN at the early, focused control followees groups (ρ=0.86) in comparison with treatment, and full recovery stages. AN users tweeted more the AN followees versus focused control followees groups frequently at night and on weekends than focused control users. (ρ=0.69). These results suggest that more interests are shared These results are consistent with clinical findings that suggest with focused control users as the recovery process advances. that patients with AN often report poor sleep quality and reduced With regard to RQ6, we posed a question regarding the detection sleep time [67]. In addition, the image results indicate that the of users at the precontemplation stage and users at the analysis of visual elements is relevant for the detection of AN contemplation stage. After applying a clustering process and cases and focused control cases. In particular, our findings using emotion indicators, we identified two clusters, where one showed that the features extracted from the content generated was characterized by a predominance of positive emotions over by users are the most relevant for characterizing AN users, the negative ones, suggesting that users at the precontemplation especially those related to linguistic and psychological factors, stage could be the members of this cluster. On the other hand, including terms that describe risk factors and the signs and the second cluster showed values that characterize people at the symptoms of AN (ie, anorexia-related vocabulary). contemplation stage, where negative emotions predominate. With regard to RQ4, we determined the linguistic attributes that Finally, to address RQ7, where we explore the existence of characterize Spanish-speaking users with AN and found that differences among our control groups (random and focused similar to a related work on English texts [6], the high use of control users), we observed that the focused control group had first-person singular pronouns and verbs conjugated with these three times more organizations’ accounts than the random pronouns distinguishes AN users from control and recovered control group. We also noticed that the main differences among cases (P<.001). We have also observed that the AN group is these groups were found in linguistic attributes, especially for characterized by a significantly lower use of articles and higher focused control users that were characterized by the use of a use of impersonal pronouns than control users (P<.001), which, reduced number of swearing terms and more anorexia-related to the best of our knowledge, has not been reported in any vocabulary terms. These findings complement our prior related work in English. assumptions that focused control users were mostly Our findings also reinforced the relevance of textual elements organizations, specialists, and clinicians, corresponding to a through the development of a deep learning model for the prorecovery community [6,17,18]. detection of AN-related tweets, which was constructed to Given our findings, another relevant aspect to discuss is the address RQ5. We explored the change in the ratio of posts indirect and unintended role of social platforms in the promotion related to AN tweeted by users at each stage and found that of harmful content among users with and without EDs. A users at the early stages of AN posted more AN-related tweets. previous study [68] reported significant decreases in caloric In addition, highly significant differences among AN cases, intake among people exposed to pro-ED websites between recovered cases, and control cases (P<.001) were observed. preexposure and postexposure. As recommender systems are There were also very significant differences between the AN designed, users with AN are likely to be recommended to follow and treatment groups (P=.004) in terms of this feature, that is, users similar to them (other AN users), and we believe that this the proportion of tweets related to AN significantly changed indirectly contributes to the reinforcement of unhealthy habits. depending on the recovery stage, indicating the progress in the Tools that are aware of these risks must thus be developed. recovery process of social media users with AN. Limitations Among the relational factors explored (RQ2), we found that the AN and focused control groups could be identified by analyzing The analyses performed in this study present certain limitations the structure of their social network (clustering approach). that are mainly given by the structure of the social platform Among focused control users, there were several organizations analyzed, which does not provide explicit information regarding and specialists for the treatment and prevention of EDs. The elements relevant to our analysis, such as the location, age, gender, or medical records of users. This is the main reason that https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 35 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al has led us to infer information based on the analysis of the users’ Reproducibility and Ethical Concerns posts; therefore, the accuracy of our results is limited to the This work was approved by the ethical review board of Pompeu performance of the tools we have applied. This applies to the Fabra University. To avoid processing and storing personal or demographical features inferred (age groups and gender); the sensitive data, a proper process of data transformation and analysis of elements that are related to the signs and symptoms anonymization was followed. We only stored the extracted of anorexia (terms related to risk factors and AN-related transformed features. Our approach corresponded to an vocabulary); image analysis tools; and the inference of aspects observational study, with findings that implied that the detection that involve the location of the user, such as the weekend tweet of AN on social media is viable; however, the further count ratio and the sleep period tweeting ratio. This last feature development and deployment of detection tools are required. is calculated in a way that overcomes the issue of not knowing Thus, it is necessary to perform a prior and proper risk-benefit the difference in the posting time according to the user’s time assessment accompanied by the analysis of the legal frameworks zone. This aspect is also an issue regarding the tweeting that regulate such activities. It is also important to address the frequency in different periods of the year, as the seasons change potential misusage capabilities of such tools [16,70]. according to the location of users. In this sense, a Spanish variant classifier could be useful to distinguish users from different Regarding the reproducibility of this work, Twitter’s policies locations; however, this would still be insufficient for the cases on the distribution of the data collected through its API should of migrants. be respected. No information that could lead to the identification of the users included in our study will be shared, as we did not We did not measure the biases introduced by the platform in store any personal information [71]. However, the values of the our data set; however, our samples seemed to be representative features calculated are available upon reasonable request and of the reality in terms of age and gender despite most Twitter after a proper evaluation of the use purpose. users being male and middle-aged [69]. The AN group users were mostly female teenagers, consistent with the age group Future Work with the highest incidence rate (15- to 19-year-old females) for We believe that our findings are relevant to the development of AN [54]. We also considered the limitations owing to the predictive models that can assist specialists in the detection of accuracy of the translation of terms to English for the annotation users with AN and that can display indicators for risk factors and use of Empath as a topic detection tool. as well as signs and symptoms that characterize AN. A proper We also considered the limitations that pose the characteristics risk assessment process should be performed before the of users who have a preference for Twitter as a platform and deployment of such models. who choose to make their tweets publicly available, which might Another aspect that we intend to analyze is the area of differ from those that keep their profiles private. It is important recommender systems. According to our findings, AN users to recall that our study is limited to users who make use of mainly followed AN-related content, which reinforces their Twitter; therefore, the analysis of the behavior of users from harmful behaviors. Thus, a recommender system that breaks other social platforms and even of people with AN that do not this bubble and reduces the polarization between AN and have accounts on any social platforms is out of our reach. focused control users can be an interesting path to follow. This is relevant considering that most recommender systems of social platforms are focused on matching users with shared interests, which in this case, might not be appropriate. Acknowledgments This work was supported by the Spanish Ministry of Economy and Competitiveness under the Maria de Maeztu Units of Excellence Program (MDM-2015-0502). The authors acknowledge the funding received from the Spanish project TIN2015-65464-R (Ministerio de Asuntos Económicos y Transformación Digital/Fondo Europeo de Desarrollo Regional). The authors also acknowledge the involvement and support of members of the Fundación Instituto de Trastornos Alimentarios and the Nvidia Corporation, who donated a Tesla K40 GPU and a GTX TITAN Graphics Processing Unit (GPU) used for this research. Conflicts of Interest None declared. References 1. Attia E, Walsh BT. Anorexia nervosa. Am J Psychiatry 2007 Dec;164(12):1805-1810. [doi: 10.1176/appi.ajp.2007.07071151] [Medline: 18056234] 2. Jacobi C, Hayward C, de ZM, Kraemer HC, Agras WS. Coming to terms with risk factors for eating disorders: application of risk terminology and suggestions for a general taxonomy. Psychol Bull 2004 Jan;130(1):19-65. [doi: 10.1037/0033-2909.130.1.19] [Medline: 14717649] https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 36 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al 3. Ulfvebrand S, Birgegård A, Norring C, Högdahl L, von Hausswolff-Juhlin Y. Psychiatric comorbidity in women and men with eating disorders results from a large clinical database. Psychiatry Res 2015 Dec 15;230(2):294-299. [doi: 10.1016/j.psychres.2015.09.008] [Medline: 26416590] 4. Prochaska JO, Velicer WF. The transtheoretical model of health behavior change. Am J Health Promot 1997;12(1):38-48. [doi: 10.4278/0890-1171-12.1.38] [Medline: 10170434] 5. Hasler G, Delsignore A, Milos G, Buddeberg C, Schnyder U. Application of Prochaska's transtheoretical model of change to patients with eating disorders. J Psychosom Res 2004 Jul;57(1):67-72. [doi: 10.1016/S0022-3999(03)00562-2] [Medline: 15256297] 6. De Choudhury M. Anorexia on Tumblr: a characterization study. In: Proceedings of the 5th International Conference on Digital Health 2015. 2015 Presented at: DH '15: Digital Health 2015 Conference; May 18-20, 2015; Florence Italy p. 43-50. [doi: 10.1145/2750511.2750515] 7. Losada D, Crestani F, Parapar J. Overview of eRisk - early risk prediction on the internet. In: Bellot P, Trabelsi C, Mothe J, Murtagh F, Nie JY, Soulier L, et al, editors. Experimental IR Meets Multilinguality, Multimodality, and Interaction. Switzerland: Springer; 2018:343-363. 8. Cavazos-Rehg PA, Krauss MJ, Costello SJ, Kaiser N, Cahn ES, Fitzsimmons-Craft EE, et al. "I just want to be skinny.": a content analysis of tweets expressing eating disorder symptoms. PLoS One 2019;14(1):e0207506 [FREE Full text] [doi: 10.1371/journal.pone.0207506] [Medline: 30650072] 9. Prieto VM, Matos S, Álvarez M, Cacheda F, Oliveira JL. Twitter: a good place to detect health conditions. PLoS One 2014;9(1):e86191 [FREE Full text] [doi: 10.1371/journal.pone.0086191] [Medline: 24489699] 10. Amini H, Kosseim L. Towards explainability in using deep learning for the detection of anorexia in social media. In: Natural Language Processing and Information Systems. Switzerland: Springer; 2020:225-235. 11. Ramírez-Cifuentes D, Mayans M, Freire A. Early risk detection of anorexia on social media. In: Bodrunova SS, editor. Internet Science. Switzerland: Springer; 2018:3-14. 12. Guntuku SC, Preotiuc-Pietro D, Eichstaedt JC, Ungar LH. What Twitter profile and posted images reveal about depression and anxiety. In: Proceedings of the International AAAI Conference on Web and Social Media. 2019 Presented at: Thirteenth International AAAI Conference on Web and Social Media (ICWSM-19); June 11-14, 2019; Münich, Germany p. 236-246. 13. Ramírez-Cifuentes D, Freire A, Baeza-Yates R, Puntí J, Medina-Bravo P, Velazquez DA, et al. Detection of suicidal ideation on social media: multimodal, relational, and behavioral analysis. J Med Internet Res 2020 Jul 07;22(7):e17758 [FREE Full text] [doi: 10.2196/17758] [Medline: 32673256] 14. Arseniev-Koehler A, Lee H, McCormick T, Moreno MA. #Proana: Pro-eating disorder socialization on Twitter. J Adolesc Health 2016 Jun;58(6):659-664. [doi: 10.1016/j.jadohealth.2016.02.012] [Medline: 27080731] 15. Spinczyk D, Nabrdalik K, Rojewska K. Computer aided sentiment analysis of anorexia nervosa patients' vocabulary. Biomed Eng Online 2018 Feb 02;17(1):19 [FREE Full text] [doi: 10.1186/s12938-018-0451-2] [Medline: 29394945] 16. Chancellor S, Kalantidis Y, Pater J, De Choudhury M, Shamma D. Multimodal classification of moderated online pro-eating disorder content. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. 2017 Presented at: CHI '17: CHI Conference on Human Factors in Computing Systems; May 6-11, 2017; Denver Colorado USA p. 3213-3226. [doi: 10.1145/3025453.3025985] 17. Wang T, Brede M, Ianni A, Mentzakis E. Social interactions in online eating disorder communities: a network perspective. PLoS One 2018;13(7):e0200800 [FREE Full text] [doi: 10.1371/journal.pone.0200800] [Medline: 30059512] 18. Wang T, Brede M, Ianni A, Mentzakis E. Characterizing dynamic communication in online eating disorder communities: a multiplex network approach. Appl Netw Sci 2019 Apr 17;4(1):1-22. [doi: 10.1007/s41109-019-0125-4] 19. De Choudhury M, Counts S, Horvitz E, Hoff A. Characterizing and predicting postpartum depression from shared Facebook data. In: Proceedings of the 17th ACM Conference on Computer Supported Cooperative Work & Social Computing. 2014 Presented at: CSCW'14: Computer Supported Cooperative Work; February 15-19, 2014; Baltimore Maryland USA p. 626-638. [doi: 10.1145/2531602.2531675] 20. Colombo GB, Burnap P, Hodorog A, Scourfield J. Analysing the connectivity and communication of suicidal users on Twitter. Comput Commun 2016 Jan 01;73(Pt B):291-300 [FREE Full text] [doi: 10.1016/j.comcom.2015.07.018] [Medline: 26973360] 21. Chancellor S, Pater J, Clear T, Gilbert E, De Choudhury M. #thyghgapp: Instagram content moderation and lexical variation in pro-eating disorder communities. In: Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work & Social Computing. 2016 Presented at: CSCW '16: Computer Supported Cooperative Work and Social Computing; February 27-March 2, 2016; San Francisco California USA p. 1201-1213. [doi: 10.1145/2818048.2819963] 22. Scherr S, Arendt F, Frissen T, Oramas J. Detecting intentional self-harm on Instagram: development, testing, and validation of an automatic image-recognition algorithm to discover cutting-related posts. Soc Sci Comput Rev 2019 Mar 18;38(6):673-685. [doi: 10.1177/0894439319836389] 23. Reece AG, Danforth CM. Instagram photos reveal predictive markers of depression. EPJ Data Sci 2017 Aug 8;6(1):15. [doi: 10.1140/epjds/s13688-017-0110-z] 24. Chancellor S, Mitra T, De Choudhury M. Recovery amid pro-anorexia: analysis of recovery in social media. In: Proc SIGCHI Conf Hum Factor Comput Syst. 2016 May Presented at: CHI Conference on Human Factors in Computing Systems; https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 37 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al 2016; New York, NY, USA p. 2111-2123 URL: http://europepmc.org/abstract/MED/28840201 [doi: 10.1145/2858036.2858246] 25. López-Ubeda P, Plaza-Del-Arco F, Diaz-Galiano M, Urena-Lopez L, Martin-Valdivia MT. Detecting anorexia in Spanish tweets. In: Proceedings of Recent Advances in Natural Language Processing. 2019 Presented at: International Conference on Recent Advances in Natural Language Processing (RANLP); Sep 2–4, 2019; Varna, Bulgaria p. 655-663. [doi: 10.26615/978-954-452-056-4_077] 26. Leis A, Ronzano F, Mayer MA, Furlong LI, Sanz F. Detecting signs of depression in tweets in Spanish: behavioral and linguistic analysis. J Med Internet Res 2019 Jun 27;21(6):e14199 [FREE Full text] [doi: 10.2196/14199] [Medline: 31250832] 27. Coppersmith G, Dredze M, Harman C. Quantifying mental health signals in Twitter. In: Proceedings of the Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality. 2014 Presented at: Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality; June 2014; Baltimore, Maryland, USA p. 51-60. [doi: 10.3115/v1/w14-3207] 28. Elrod K, Dykeman C. A corpus linguistic analysis of pro-anorexia public Tumblr posts written in Spanish. PsyArXiv 2019:A. [doi: 10.31234/osf.io/g2qye] 29. Les pagines pro ana I pro mia inunden la xarxa. Iqua - Asociación Contra La Anorexia Y Bulimia. 2011. URL: https:/ /tinyurl.com/5dkxeen7 [accessed 2021-06-24] 30. Sample realtime tweets. Twitter. URL: https://developer.twitter.com/en/docs/tweets/sample-realtime/api-reference/ get-statuses-sample [accessed 2021-06-24] 31. Leichner P. A new treatment approach to eating disorders in youth. B C Med J 2005;47(1):23-27 [FREE Full text] 32. Kruskal WH, Wallis WA. Use of ranks in one-criterion variance analysis. J Am Stat Assoc 1952 Dec;47(260):583. [doi: 10.2307/2280779] 33. Mann HB, Whitney DR. On a test of whether one of two random variables is stochastically larger than the other. Ann Math Statist 1947 Mar;18(1):50-60. [doi: 10.1214/aoms/1177730491] 34. statsmodels.stats.proportion.proportions_ztest. Statsmodels Org. URL: https://www.statsmodels.org/stable/generated/ statsmodels.stats.proportion.proportions_ztest.html [accessed 2021-06-24] 35. Kim Y. Convolutional neural networks for sentence classification. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014 Presented at: Conference on Empirical Methods in Natural Language Processing (EMNLP); October 2014; Doha, Qatar. [doi: 10.3115/v1/d14-1181] 36. Tolles J, Meurer WJ. Logistic regression: relating patient characteristics to outcomes. J Am Med Assoc 2016 Aug 02;316(5):533-534. [doi: 10.1001/jama.2016.7653] [Medline: 27483067] 37. Ramírez-Esparza N, Pennebaker J, García FA, Suriá R. La psicología del uso de las palabras: Un programa de computadora que analiza textos en español. In: Revista Mexicana de Psicología. Distrito Federal, México: Sociedad Mexicana de Psicología A.C; 2007:85-99. 38. Tausczik YR, Pennebaker JW. The psychological meaning of words: LIWC and computerized text analysis methods. J Lang Soc Psychol 2009 Dec 08;29(1):24-54. [doi: 10.1177/0261927X09351676] 39. Rude S, Gortner E, Pennebaker J. Language use of depressed and depression-vulnerable college students. Cogn Emot 2004 Dec;18(8):1121-1133. [doi: 10.1080/02699930441000030] 40. Mohammad S, Turney P. Crowdsourcing a word-emotion association lexicon. Comput Intell 2013;29(3):436-465. [doi: 10.1111/j.1467-8640.2012.00460.x] 41. Hofman E. Senti-py. GitHub, Inc. URL: https://github.com/aylliote/senti-py [accessed 2021-06-24] 42. Culbert KM, Racine SE, Klump KL. Research Review: What we have learned about the causes of eating disorders - a synthesis of sociocultural, psychological, and biological research. J Child Psychol Psychiatry 2015 Nov;56(11):1141-1164. [doi: 10.1111/jcpp.12441] [Medline: 26095891] 43. Divins M. Laxantes. Farmacia Professional. 2004. URL: https://www.elsevier.es/ es-revista-farmacia-profesional-3-pdf-13064573 [accessed 2021-07-06] 44. Fast E, Chen B, Bernstein M. Empath: Understanding topic signals in large-scale text. In: Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. 2016 Presented at: CHI'16: CHI Conference on Human Factors in Computing Systems; May 7-12, 2016; San Jose California USA p. 4647-4657. [doi: 10.1145/2858036.2858535] 45. Han S. googletrans 3.0.0. PyPI Library. 2020. URL: https://pypi.org/project/googletrans/ [accessed 2021-06-24] 46. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: machine learning in python. J Mach Learn Res 2011;12(1):2825-2830. 47. Savand A. Stop-words. PyPI Library. 2018. URL: https://pypi.org/project/stop-words/ [accessed 2021-06-24] 48. Baziotis C, Pelekis N, Doulkeridis C. Datastories at SemEval-2017 task 4: Deep LSTM with attention for message-level and topic-based sentiment analysis. In: Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017). 2017 Presented at: 11th International Workshop on Semantic Evaluation (SemEval-2017); August 2017; Vancouver, Canada p. 747-754. [doi: 10.18653/v1/s17-2126] 49. Shing H, Nair S, Zirikly A, Friedenberg M, Daumé III H, Resnik P. Expert, crowdsourced, and machine assessment of suicide risk via online postings. In: Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 38 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al From Keyboard to Clinic. 2018 Presented at: Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic; June 2018; New Orleans, LA p. 25-36. [doi: 10.18653/v1/w18-0603] 50. Deriu J, Lucchi A, De LV, Severyn A, Müller S, Cieliebak M, et al. Leveraging large amounts of weakly supervised data for multi-language sentiment classification. In: Proceedings of the 26th International Conference on World Wide Web. 2017 Presented at: WWW '17: 26th International World Wide Web Conference; April 3-7, 2017; Perth Australia p. 1045-1052. [doi: 10.1145/3038912.3052611] 51. Jacomy M, Venturini T, Heymann S, Bastian M. ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. PLoS One 2014;9(6):e98679 [FREE Full text] [doi: 10.1371/journal.pone.0098679] [Medline: 24914678] 52. Blondel VD, Guillaume J, Lambiotte R, Lefebvre E. Fast unfolding of communities in large networks. J Stat Mech 2008 Oct 09;2008(10):P10008. [doi: 10.1088/1742-5468/2008/10/P10008] 53. Bastian M, Heymann S, Jacomy M. Gephi: an open source software for exploring and manipulating networks. In: Proceedings of the International AAAI Conference on Web and Social Media. 2009 Presented at: Third International AAAI Conference on Weblogs and Social Media; May 17–20, 2009; San Jose, California USA p. 361-362 URL: https://ojs.aaai.org/index.php/ ICWSM/article/view/13937 54. Hoek HW. Incidence, prevalence and mortality of anorexia nervosa and other eating disorders. Curr Opin Psychiatry 2006 Jul;19(4):389-394. [doi: 10.1097/01.yco.0000228759.95237.78] [Medline: 16721169] 55. Wang Z, Hale S, Adelani D, Grabowicz P, Hartmann T, Flöck F, et al. Demographic inference and representative population estimates from multilingual social media data. In: Proceedings of the 2019 World Wide Web Conference. 2019 Presented at: WWW '19: The Web Conference; May 13-17, 2019; San Francisco CA USA p. 2056-2067. [doi: 10.1145/3308558.3313684] 56. Silvester S, Tanbakuchi A, Müller P, Nunez-Iglesias J, Harfouche M, Klein A, OrganicIrradiation, jackwalker64, rreilink, NiklasRosenstein, Zulko, et al. Imageio/imageio v.0.9.0. Python Library. 2020. URL: https://zenodo.org/record/3931847#. YN427JgzbIU [accessed 2021-07-01] 57. Pillow 8.2.0. PyPI : Python Library. URL: https://pypi.org/project/Pillow/ [accessed 2021-06-24] 58. Computer vision with Chooch AI. Python API. URL: https://chooch.ai/api/#general-recognition-api [accessed 2021-06-24] 59. Levi G, Hassner T. Emotion recognition in the wild via convolutional neural networks and mapped binary patterns. In: Proceedings of the 2015 ACM on International Conference on Multimodal Interaction. 2015 Presented at: ICMI '15: International Conference on Multimodal Interaction; November 9-13, 2015; Seattle Washington USA p. 503-510. [doi: 10.1145/2818346.2830587] 60. Rodriguez P, Gonzàlez J, Gonfaus J, Roca F. Integrating vision and language in social networks for identifying visual patterns of personality traits. Int J Soc Sci Humanity Stud 2019 Feb:6-12. [doi: 10.18178/ijssh.2019.v9.981] 61. Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proc IEEE 1998;86(11):2278-2324. [doi: 10.1109/5.726791] 62. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016 Presented at: IEEE Conference on Computer Vision and Pattern Recognition (CVPR); June 27-30, 2016; Las Vegas, NV, USA. [doi: 10.1109/cvpr.2016.90] 63. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM 2017 May 24;60(6):84-90. [doi: 10.1145/3065386] 64. Xie S, Girshick R, Dollár P, Tu Z, He K. Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017 Presented at: IEEE Conference on Computer Vision and Pattern Recognition (CVPR); July 21-26, 2017; Honolulu, HI, USA p. 1492-1500. [doi: 10.1109/cvpr.2017.634] 65. Loshchilov I, Hutter F. SGDR: Stochastic gradient descent with warm restarts. arXiv: Computer Science - Machine Learning. 2016. URL: https://arxiv.org/abs/1608.03983 [accessed 2021-06-24] 66. Plutchik R. A psychoevolutionary theory of emotions. Soc Sci Inf 2017 Feb 06;21(4-5):529-553. [doi: 10.1177/053901882021004003] 67. Padez-Vieira F, Afonso P. Sleep disturbances in anorexia nervosa. Adv Eating Disorders 2016 May 04;4(2):176-188. [doi: 10.1080/21662630.2016.1175958] 68. Jett S, LaPorte DJ, Wanchisn J. Impact of exposure to pro-eating disorder websites on eating behaviour in college women. Eur Eat Disord Rev 2010;18(5):410-416. [doi: 10.1002/erv.1009] [Medline: 20572210] 69. Jashinsky J, Burton SH, Hanson CL, West J, Giraud-Carrier C, Barnes MD, et al. Tracking suicide risk factors through Twitter in the US. Crisis 2014;35(1):51-59. [doi: 10.1027/0227-5910/a000234] [Medline: 24121153] 70. Guntuku SC, Yaden DB, Kern ML, Ungar LH, Eichstaedt JC. Detecting depression and mental illness on social media: an integrative review. Curr Opin Behav Sci 2017 Dec;18:43-49. [doi: 10.1016/j.cobeha.2017.07.005] 71. Chancellor S, Birnbaum M, Caine E, Silenzio V, De Choudhury M. A taxonomy of ethical tensions in inferring mental health states from social media. In: Proceedings of the Conference on Fairness, Accountability, and Transparency. 2019 Presented at: FAT* '19: Conference on Fairness, Accountability, and Transparency; January 29-31, 2019; Atlanta GA USA p. 79-88. [doi: 10.1145/3287560.3287587] https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 39 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Ramírez-Cifuentes et al Abbreviations AN: anorexia nervosa API: application programming interface BoW: bag-of-words CNN: convolutional neural network DTTR: day time tweets ratio ED: eating disorder LIWC: Linguistic Inquiry and Word Count LR: logistic regression pro-ED: pro-eating disorder RQ: research question STTR: sleep time tweets ratio Edited by R Kukafka; submitted 21.11.20; peer-reviewed by K Usher, M Roncero, Y Hwang; comments to author 10.02.21; revised version received 10.03.21; accepted 04.05.21; published 20.07.21 Please cite as: Ramírez-Cifuentes D, Freire A, Baeza-Yates R, Sanz Lamora N, Álvarez A, González-Rodríguez A, Lozano Rochel M, Llobet Vives R, Velazquez DA, Gonfaus JM, Gonzàlez J Characterization of Anorexia Nervosa on Social Media: Textual, Visual, Relational, Behavioral, and Demographical Analysis J Med Internet Res 2021;23(7):e25925 URL: https://www.jmir.org/2021/7/e25925 doi: 10.2196/25925 PMID: 34283033 ©Diana Ramírez-Cifuentes, Ana Freire, Ricardo Baeza-Yates, Nadia Sanz Lamora, Aida Álvarez, Alexandre González-Rodríguez, Meritxell Lozano Rochel, Roger Llobet Vives, Diego Alejandro Velazquez, Josep Maria Gonfaus, Jordi Gonzàlez. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 20.07.2021. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included. https://www.jmir.org/2021/7/e25925 J Med Internet Res 2021 | vol. 23 | iss. 7 | e25925 | p. 40 (page number not for citation purposes) XSL• FO RenderX