Understanding Patterns of Anorexia Manifestations in Social Media Data with Deep Learning

Eating disorders are a growing problem especially among young people, yet they have been under-studied in computational research compared to other mental health disorders such as depression. Computational methods have a great potential to aid with the automatic detection of mental health problems, but state-of-the-art machine learning methods based on neural networks are notoriously difficult to interpret, which is a crucial problem for applications in the mental health domain. We propose leveraging the power of deep learning models for automatically detecting signs of anorexia based on social media data, while at the same time focusing on interpreting their behavior. We train a hierarchical attention network to detect people with anorexia and use its internal encodings to discover different clusters of anorexia symptoms. We interpret the identified patterns from multiple perspectives, including emotion expression, psycho-linguistic features and personality traits, and we offer novel hypotheses to interpret our findings from a psycho-social perspective. Some interesting findings are patterns of word usage in some users with anorexia which show that they feel less as being part of a group compared to control cases, as well as that they have abandoned explanatory activity as a result of a greater feeling of helplessness and fear.


Introduction and Previous Work
Anorexia nervosa (AN) is a type of eating disorder that leads to multiple psychiatric and somatic complications and constitutes a major public health concern. It involves a restriction of energy intake in relation to needs, leading to significantly low body weight in relation to age, sex, developmental course and physical health. It includes among its typical symptomatology an intense fear of gaining weight or becoming fat and a distortion of one's body image (APA, 2014).
The incidence of AN, like that of other eating disorders (ED), has increased in recent decades. In a systematic literature review for the 2000-2018 period (Galmiche et al., 2019), the reported weighted means of lifetime ED (proportion of EDs at any point in life) were 8.4% (3.3-18.6%) for women and 2.2% (0.8-6.5%) for men. The authors also report that the weighted means of point ED prevalence increased over the study period from 3.5% for the 2000-2006 period to 7.8% for the [2013][2014][2015][2016][2017][2018] period. This highlights a real challenge for public health and healthcare providers.
In an attempt to understand the psychosocial origins of anorexia nervosa, some studies have investigated how body image is shaped in people suffering from this mental disorder (Giordani, 2006(Giordani, , 2009Giacomozzi and da Silva Bousfield, 2011). From early research in social psychology we already know that body image, in an existential context, is the revelation of an identity that the subject constructs in the frame of concrete social relations (Goffman, 1963). From a sociological perspective, some research proposes to understand bodies attending the interaction with social forces (Turner, 2008). From anthropology, new uses of bodies (tattoos, piercings, etc) support Le Breton's idea about the study of body in modernity as an unfinished material, as "a place of self-presentation" (Le Breton, 2011). This body of research could be applied to the study of anorexia nervosa without forgetting the enormous symbolism of the act of eating. It is well established that eating with others (Dunbar, 2017) and eating the same food as the others is a major symbol of social integration (Harris, 1971;Young et al., 1971).
Mental health disorders in general, as a very significant public health matter (World Health Organization, 2012), have received attention in previous research in computational studies as well. The majority of research has focused on the study of depression (De Choudhury et al., 2013;Eich-staedt et al., 2018;Abd Yusof et al., 2017;Yazdavar et al., 2017), but other mental illnesses have also been studied, including generalized anxiety disorder (Shen and Rudzicz, 2017), schizophrenia , post-traumatic stress disorder (Coppersmith et al., 2014, risks of suicide (O'dea et al., 2015), and self-harm (Losada et al., 2019;Yang et al., 2016).
For anorexia, there are very few studies approaching the problem from a computational perspective. To our knowledge, the only publicly available social media dataset dedicated to anorexia is the eRisk dataset (Losada et al., 2019). The winners of eRisk's shared task on anorexia detection (Mohammadi et al., 2019) used a hierarchical attention network and obtain a state-of-the-art F1 score of 0.71. In (Cohan et al., 2018) the authors introduce a dataset annotated for multiple mental disorders including anorexia. Another study (Amini and Kosseim) on the explainability of anorexia detection models analyzes attention weights of a neural network to show that attention at the user level correlates with the importance of individual texts for classification performance.
Explainability of machine learning models, especially in the field of mental health, is a very important issue. In practice, models based on neural networks are vastly successful for most NLP applications, even though they have been only briefly explored in existing computational studies on mental disorders. Nevertheless, neural networks are notoriously difficult to interpret. While there is increasing interest in the field of explainability of machine learning models including in NLP (Gilpin et al., 2018), there are fewer such studies for mental health disorder detection.
In the name of transparency, it is essential for any automatic system that can assist with mental health disorder detection to make its decision-making process understandable. Especially in the medical domain, using black-box systems can be dangerous for patients (Zucco et al., 2018;Holzinger et al., 2017). Moreover, recently, the need of explanatory systems is required by regulations like the General Data Protection Regulation (GDPR) adopted by the European Union.
While many quantitative studies in the computational analysis of mental health use features such as lexicons (Trotzek et al., 2017;De Choudhury et al., 2014) to study the manifestations of mental disorders in user-generated data, these models are very limited computationally in comparison to deep learning models. The behavior of powerful classifiers modelling complex patterns in the data has the potential to help uncover manifestations of the disease that are potentially difficult to observe with the naked eye, and thus be useful not only as tools for the detection of disorders, but also as analysis instruments for generating insights and potentially assisting clinicians in the diagnosis process.
In our study, we propose using deep learning as a tool to aid with a deeper investigation of anorexia manifestations in social media texts. We train a hierarchical attention network to classify people with anorexia against control cases based on their social media activity. This architecture has been shown to provide good results for anorexia detection, and additionally includes in the model a series of features that encode different levels of the language (style, emotions, topics etc). To our knowledge this has not been done in previous work, and allows us the advantage of a more interpretable model. We interpret the predictions of the network as well as its hidden layers as a way to identify different patterns of anorexia symptoms in social media users, which we analyze in view of the different features, and offer hypotheses on the different patterns observed from a psychological perspective. Thus, we aim to answer the following research questions: RQ1. Is it possible to leverage complex deep learning models and their encoding power in order to identify different patterns of anorexia symptoms in social media texts? RQ2. Can we characterize the differences between different groups of people with anorexia based on psycho-linguistic and emotion features, and measures of personality traits? RQ3. Could we identify some features to explore the hypothesis that anorexia nervosa is a way to express some degree of conflict with one's own group?

Classification Experiments
In order to explore the proposed hypotheses, we start by building a deep learning model in order to automatically classify texts belonging to users with anorexia and control cases.

Dataset
eRisk Reddit dataset on anorexia. The eRisk CLEF lab 1 is focused on the early prediction of mental disorder risk from social media data, focused on disorders such as depression (Losada et al., 2018), anorexia and self-harm tendencies (Losada et al., 2019(Losada et al., , 2020. Data is collected from Reddit posts and comments selected from specific relevant sub-reddits. Users suffering from a mental disorder are annotated by automatically detecting self-stated diagnoses. Control cases are selected from participants in the same sub-reddits (having similar interests), thus making sure the gap between healthy and diagnosed users is not trivially detectable. A long history of posts are collected for the users included in the dataset, up to years prior to the diagnosis. The dataset on anorexia, released as part of eRisk 2019 (Losada et al., 2019), contains 823,754 posts collected from 1,287 users, of which 10.4% are anorexic users.

Model and Features
We choose a hierarchical attention network (HAN) as our model: a deep neural network with a hierarchical structure, including multiple features encoded with LSTM layers and two levels of attention. HANs have previously shown to be successful for the detection of anorexia in social media. Mohammadi et al., 2019).
The HAN is made up of two components: a post-level encoder, which produces a representation of a post, and a user-level encoder, which generates a representation of a user's post history. The post-level encoder and the user-level encoder are modelled as LSTMs. The word sequences encoded using embeddings initialized with GloVe pre-trained embeddings (Pennington et al., 2014) and passed to the LSTM are then concatenated with the other features to form the hierarchical post encoding. The obtained representation is passed to the user-encoder LSTM, which is connected to the output layer. We use the train/test split provided by the shared task organizers, done at the user level, making sure users occurring in one subset don't occur in the other. Since individual posts are too short to be accurately classified, we construct our datapoints through concatenating groups of 50 posts, sorted chronologically. We use a weighted loss function to compensate for the class imbalance. The architecture of the model is shown in Figure 1.
We represent social media texts using features that capture different levels of the language (semantic, stylistic, emotions etc.) and train the model to predict anorexia risk for each user. Content features. We include a general representation of text content by transforming each text into word sequences, represented as embeddings. Style features. The usage pattern of function words is known to be reflective of an author's style, at an unconscious level (Mosteller and Wallace, 1963). As stylistic features, we extract from each text a numerical vector representing function words frequencies as bag-of-words, which are passed through an additional dense layer of 20 units. We complement function word distribution features with other syntactical features extracted from the LIWC lexicon, as described below. LIWC features. The LIWC lexicon (Pennebaker et al., 2001) has been widely used in computational linguistics as well as some clinical studies for analyzing how suffering from mental disorders manifests in an author's writings. LIWC is a lexicon mapping words of the English vocabulary to 64 lexico-syntactic features of different kinds, with high quality associations curated by human experts, capturing different levels of language: including style (through syntactic categories), emotions (through affect categories) and topics (such as money, health or religion). Emotions and sentiment. We dedicate a few features to representing emotional content in our texts, since the emotional state of a user is known to be highly correlated with her mental health. Aside from the sentiment and emotion categories in the LIWC lexicon, we include a second lexicon: the NRC emotion lexicon (Mohammad and Turney, 2013), which is dedicated exclusively to emotion representation, with categories corresponding to a wider and a more fine-grained selection of emotions, containing the 8 Plutchik's emotions (Plutchik, 1984), as well as positive/negative sentiment categories: anger, anticipation, disgust, fear, joy, sadness, surprise, trust. We represent LIWC and NRC features by computing for each category the proportion of words in the input text which are associated with that category. These are concatenated with the other features to form the post-level encodings, which are then stacked and passed to the final user-level LSTM which is connected to the output layer.

Classification Results
The results obtained with our neural network for the detection of anorexic users are shown in Table 1. As performance metrics we compute the F1-score of the positive class and the area under the ROC curve (AUC), which is more robust in the case of data imbalance. We compare the results of our model with baselines such as a logistic regression model with bag-of-word features, and transformer-based models including RoBERTa (Liu et al., 2019) and AlBERT (Lan et al., 2019) with word sequences as features.
Our HAN model achieves the best results in terms of AUC. In the following sections, we explore this model in more depth in order to explain its behavior and leverage it to discover insights on linguistic patterns associated with anorexia.

Explaining Predictions
In this section we present different analyses meant to uncover insights into how the model arrives at its predictions, first looking at the attention weights and abstract internal representations of the data in the layers of the neural network, and secondly providing several feature-focused analyses, using emotions and LIWC categories, as well as personality markers.

Attention Analysis
Attention is a mechanism frequently used in recurrent neural networks in order to weigh the parts of the input sequence differently according to their importance for prediction. Attention weights are learned by the network, and thus can be used as a means to interpreting its decisions. In our models we include recurrence at the user level, along with an attention layer, which can thus be used to infer the weight placed on each part in a user's post history by the neural network.   Figure 3, we plot the attention weights for the user-level attention layer for each of the classifiers trained on the three datasets. For this experiment, we train the neural network on one datapoint for each user, so as to ensure attention weights consistently correspond to the same part of the post history for each training example. The plot shows a general increasing importance for users suffering from anorexia: posts in the end of a user's history are more heavily weighted. This is an interesting finding, since intuition, supported by findings such as those presented in the previous sections related to emotion evolution, would suggest a user's activity on social media becomes increasingly indicative of their mental state as time goes by.
Recent studies (Jain and Wallace, 2019; Serrano and Smith, 2019; Wiegreffe and Pinter, 2019) have questioned whether attention mechanisms necessarily help with the interpretability of neural network predictions. We further explore additional techniques in order to interpret the representations learned by the model.

User Embeddings
We continue explaining the model's behavior by analyzing the internal representations of the neural network. We can regard the final layer of the trained neural network as the most compressed representation of the input examples, which is, in terms of our trained model, the optimal representa- tion for distinguishing between control cases and those suffering from a disorder. Thus, the final layer (the output of the 32-dimensional user-level LSTM) can be interpreted as a 32-dimensional embedding for the input points, corresponding to the users to be classified.
We analyze the output of the user embedding layer by reducing it to 2 dimensions using principal component analysis (PCA) and visualizing it in 2D space with a kernel density estimate (KDE) plot to show the distribution of scores across the 2 dimensions ( Figure 4). We make sure to train the PCA model on a balanced set of positive and negative users, then we extract 2D representations for all users in the test set. By looking at these representations, we can gain insight into the separability of the classes, from the perspective of the trained model, and better understand where it encounters difficulties in separating between the datapoints belonging to different classes.
We notice that, even in two dimensions, the groups of people with anorexia and control cases appear as separate clusters in the user embedding space, suggesting that the encodings generated by the model while training for its objective are powerful enough to separate the two groups.
We notice an interesting bimodal aspect in the distribution of positive users, which appear clustered into two distinct groups, while control cases' representations seem to be more compact. One of the clusters of people with anorexia is more clearly distinct from control cases in this space (ANO1), while the other cluster shows a higher overlap with control cases, presumably leading to false negatives in the model's predictions (ANO2). We suggest that the two observed clusters might represent two different groups of people with anorexia and of patterns of symptoms of anorexia, which the model is able to capture during the training process.
The problem of false negative predictions in (a) Cluster of users with anorexia ANO1.
(b) Cluster of users with anorexia ANO2.
(c) Cluster of control cases. particular is important to study, since false negatives can lead to missing cases of people with high risk of anorexia, and, left undetected, the disorder might further develop. In the following sections we take a deeper dive into the three identified clusters and attempt to interpret from different perspectives the differences among the clusters of people with anorexia and control cases.

Clustering Method
We begin the analysis of the different patterns of anorexia manifestations by attempting to automatically identify the clusters of people with anorexia in the user embedding space. We perform k-means clustering on the user embeddings for users with anorexia in the test set, and cluster the user embeddings into 2 groups. We obtain 31 users in one cluster which we denote as ANO1, and 43 in the second, denoted as ANO2.
As a first analysis, we take a look at the vocabulary of the three groups. Figure 5 illustrates word clouds corresponding to the vocabulary used by the 3 groups. As we can see, in both groups where users with anorexia are present, the word people has a greater presence suggesting it would be interesting to go deeper into analyzing the dynamics related to social relations among users with anorexia. In order to move in this direction, we will use different strategies that try to describe in more depth these three clusters and to identify the most common narratives in these clusters.

Emotions and Psycho-linguistic Features across Clusters
We analyze the three groups from the perspective of usage of emotion words and psycho-linguistic  for words in each category, separately for the texts in each cluster. We then identify features where there are statistically significant differences among the groups (using a t-test).
We identify separately features which show significant differences in usage across all three clusters, or just between users with anorexia and control cases. In general, we observe a pattern of ANO2 having intermediate values between users with anorexia and control cases, for most of the significant features (36 out of 75 categories show statistically significant differences among all three clusters, and for 30 of them the values for ANO2 are situated between ANO1 and control cases).
To obtain a deeper interpretation of the observed differences we select features which are most relevant to anorexia symptoms (see Table 2) and the features which refer to negative emotions. Interestingly, these features also show a distinct pattern from an error analysis perspective: if we select those categories which have lower values in misclassified examples in a statistically significant way, we obtain these four LIWC categories: anx, health, bio and ingest.

Personality Analysis
As a separate analysis, we try to analyze the different clusters from the perspective of personality types using the Five Factor Model. The Five Factor Model is a process of attributing certain psychological characteristics to an individual according to the so-called 'Big Five' taxonomy that has been developed into a laborious research paradigm initiated by the social psychologist Gordon Allport (Allport, 1937). Allport (1937) formulated The lexical hypothesis proposing that most of the socially relevant and salient personality characteristics have become encoded in the natural language (John et al., 1999). After decades of research, the field is approaching consensus on a general taxonomy of personality traits, the "Big Five" personality dimensions. The five factors are openness to experience (1) conscientiousness (2) agreeableness (3) extraversion (4) and neuroticism (5), as emotional stability. Exploiting this theoretical framework to extract information about users' personality from their posts means identifying such semantic associations and mapping the text around the five factors according to the words referring to them. An effective approach to do this consists in the one proposed by Neuman and Cohen (2014): the evidences of a particular personality trait are summarised into a score, which is calculated as the semantic similarity between the context-free word embedding representations respectively of the text written by the author and of the set of the benchmark adjectives (i.e., the terms empirically observed to be able to encode each of the five personality aspects according to the 'Big Five' framework). In more detail, for each trait, Neuman and Cohen (2014) define a positive and a negative sub-dimension, which correspond respectively to the possession of a sufficient degree of a given factor or, vice versa, the evidence of the exact opposite characteristic. Neuman and Cohen (2014) associate a small series of benchmark adjectives to all the 19 sub-dimensions. In the Appendix we list in full the adjectives that make up the vectors associated with each personality trait.
The set of the benchmark adjectives for the personality traits proposed by Neuman and Cohen has been successfully employed in other tasks such as profiling fake news spreaders (Giachanou et al., 2020). We use a similar approach, and measure personality scores by computing the overlap of the words in the defined vectors for each trait with the words used in each text in our dataset, normalized by text length. Following this approach we found significant differences (p-value <.005) between users with anorexia and control cases in three factors: Agreeableness (+) (-), Extraversion (-), Neuroticism (+).  Table 3: Personality-related words, usage per-mille across the clusters. *** Statistically significant difference across the three clusters ** Statistically significant difference between people suffering from anorexia and control users. * Statistically significant difference between ANO1 and others compare users suffering from anorexia with control cases, but not between the two clusters of people with anorexia. It seems that more introverted personality traits characterize users who suffer from anorexia. The same applies to the factor Agreeableness, in positive and in negative sense, the difference in agreeableness is statistically significant among people with anorexia and control cases. Users suffering from anorexia show more characteristic traits of a pleasant personality and unplesant personality than those who do not suffer from this disorder. The explanation for this difference, in positive and in negative traits, may be related to a greater manifestation of emotions and feelings among people with anorexia. With the factor measuring Neuroticism (+) we find statistically significant differences among the three clusters suggesting that users in ANO1 are at a more severe stage of this mental disorder than ANO2 because they have higher scores in this factor. Neuroticism speaks about emotional instability that leads to frequent experiences of negative emotions and which is said to result from a low ability to handle stress or strong external stimuli.

Identifiying Different Narratives and Cognitive Styles
In RQ3 we raised the possibility of exploring if some features from LIWC and emotion lexicons allow us to identify among users with anorexia a higher degree of conflict with their own group and some degree of social isolation. As we can see in Table 4, users with anorexia talk less about work, money and leisure (LIWC categories) than control cases. The absence of words from these three categories tells us about a certain  Table 4: Features about everyday activities and social relations, percentage of average usage per cluster. *** Statistically significant difference across the three clusters ** Statistically significant difference between people suffering from anorexia and control users. * Statistically significant difference between ANO1 and others degree of social isolation. These results connect with a pattern that we find in the use of personal pronouns among the different clusters.
As we can observe, users with anorexia employ significantly fewer words under the category we (we, us, our), a clear linguistic marker of a sense of belonging. Users in ANO1 use it significantly less than users in ANO2 and than control cases, suggesting that a higher degree of conflict with one's own group or a higher degree of social isolation may be linked to the more severe manifestations of anorexia. The opposite pattern is found in the use of the first person pronouns that LIWC collects under the category I (I, me, mine): users suffering from anorexia use it significantly more than control cases and cluster ANO1 uses it significantly more than cluster ANO2.
Three other features expressing social processes, family, friends and humans are more present in the narratives of users with anorexia than among control cases. The greater presence of these linguistic categories may indicate a greater centrality of social relations in the identity of these subjects suffering from anorexia. We would need to design new strategies to go deeper into this interpretation, but this greater centrality of social relations categories in people with anorexia could be derived, precisely, from a greater degree of conflict with the social environment.
We also observed in Table 5 some differences among clusters in relation to some LIWC categories related to cognitive and perceptual processes. These differences may indicate the existence of different cognitive styles between users who suffer from anorexia and those who do not.
Cognitive style is a concept used in cognitive psychology to describe the way individuals think, perceive, and remember information (Grigorenko and Sternberg, 1995). Research in psychology suggests that some cognitive styles are more prevalent in some patients suffering from depression and anorexia (Lo et al., 2008;Kaye et al., 1995).
As we can see in Table 5, people with anorexia use in their narratives more words that LIWC classifies as cognitive processes (cogmech) than control cases and the difference is also statistically significant between clusters ANO1 and ANO2. A greater presence of these traits among users with anorexia is indicative of a special effort to reason about reality, which is also a characteristic of conflictual states. We could also consider that this effort to understand involves the subject at a personal level because we see a higher presence of words belonging to the feel category among users with anorexia. Feel is, among the perceptual process in LIWC, the one that involves the subject at a deeper level.
Within cognitive processes we find it interesting to analyze certain features such as certainty, tentative and causation where there are significant differences among clusters. Certainty expresses a rigid or absolute style of thinking and tentative expresses a more flexible or less absolute style of thinking. We find significant differences in these two categories between users with anorexia and control cases. Certainty is more used in narratives from users suffering anorexia and could indicate a major degree of cognitive conflict. We see the opposite pattern with tentative, that just expresses a flexible style. It could be expressing the two sides of the same phenomenon.
With causation we only found statistically significant differences between ANO1 and the other two clusters (similarly to what we found for the expression of fear, as seen in Table 2). Reasoning about causes indicates an effort to understand the world that shows a healthier position of the subjects, and one possible interpretation is that users in ANO1 have abandoned this explanatory activity as a result of a greater feeling of helplessness and feeling more fear.

Connection to Clinical Research
Trying to identify different groups of patients who claim to suffer from anorexia nervosa or have compatible symptomatology may be a way to develop a better understanding of this complex pathology (Clinton et al., 2004). Research such as that of Viborg et al. (2018) shows a relationship between six clusters of young adolescents suffering from anorexia and higher levels of psychological difficulties and lower levels of body esteem. However, these clusters rely exclusively on symptomatology linked to eating behavior. We think that our results show that a deep learning model applied to social media texts allows us to identify clusters of patients considering more variables than those related to the eating disorder itself.
From these initial results we can provide some insights to clinical discussion. More and more clinicians (Gutiérrez and Carrera, 2021) ask themselves about the intractability of anorexia nervosa, including the disconcerting aspect of the recovery of a significant number of patients not receiving formal treatment (Keski-Rahkonen, 2014). As Gutiérrez and Carrera (2021) state in a recent review of this issue, Bruch's proposal regarding the charactization of typical anorexia in terms of Body Image Disturbances (BID) (Bruch et al., 1974)  Analyzing the language of people suffering from anorexia, we found traces of problems that could guide new approaches indicating the existence of a conflict between the patient and his or her own group: lower use of the first person plural and a greater presence of features expressing social processes. Some research focusing on the discourse analysis of people suffering from anorexia points to the existence of a link between acting on one's own body as a mechanism to take control over their lives (Malson and Ussher, 1996).
Our results may indicate that a possible origin of this need of control could be the social conflict with one's own group and the inability to communicate this conflictual situation. In this sense, Botta and Dumlao (2002) have shown the relationship between parent-child conflict and family communication styles and the development of eating disorders. In addition, Davies et al. (2012) have demonstrated in their experimental research the relevance of a verbal expression of emotions in patients with anorexia nervosa. More research is needed, for instance it may be interesting to revisit one of Bruch's ideas in her seminal work which has not been as  Table 5: Features about cognitive styles (cognitive processes and perceptual processes), percentage of average usage per cluster. *** Statistically significant difference across the three clusters ** Statistically significant difference between people suffering from anorexia and control users. * Statistically significant difference between ANO1 and others successful as her concept of BID: the description of anorexia as "a communicative disorder" which is experienced as a means of taking control over one's body as a pseudosolution to intra-and interpersonal difficulties (Bruch, 1978). Following this idea, we think that deep learning models applied to social media data can open an interesting avenue to explore the language of people suffering from anorexia and provide elements for further clinical discussion.

Ethical Considerations
Powerful machine learning models that can be trained to detect or predict the development of mental health disorders can be very valuable, but any deployment of a tool for mental disorder detection should take into account potential ethical concerns. If such tools are used by third parties (such as employers seeking to filter candidates based on their mental health profile), this could compromise the privacy of the subjects. We suggest that the development of an ethical standard is necessary, and that launching such tools could be accompanied by an ethical statement to constrain its use. Moreover, it is ethically necessary, and recently even required by regulations in some countries (such as countries in the European Union) that artificial intelligence models used in the mental health domain have interpretable behavior. We hope that we were able to take a step forward in this direction, by providing an in-depth explanation of the representations generated by our neural network, and thus facilitating trust in the predictive model.

Conclusions and Future Work
In this study, we have approached the problem of detecting people suffering from anorexia in social media thorugh training a deep learning model, and taken it a step further by explaining the behavior of the model. Based on this, we identified different clusters of users suffering from anorexia with regards to the manifested symptoms, as encoded by the trained model (RQ1). We presented several analyses for interpreting the decisions of the model trained to profile social media users at risk of developing anorexia, going beyond more common techniques such as attention weight analysis, and including hidden layer analysis and error analysis at different levels of the language for better understanding how mental disorders manifest in social media data (RQ2).
We have shown that we can interpret the detected clusters from the point of view of the social behavior of people with anorexia (RQ3), and provided in-depth interpretations of these patterns as a sociopsychological phenomenon. To our knowledge, our approach and findings are novel in the domain of the computational study of anorexia, and encourage us to go deeper into the analysis of these patterns, such as looking into the use of pronouns in relation with emotions, which could help identify more clearly the existence of a social conflict.
From a technical perspective, a more sophisticated analysis of the features included here (LIWC categories, emotions and personality vectors) could be achieved by using more semantically rich representations than bag-of-words. One such approach would be using word embeddings to identify sub-emotions (Aragón et al., 2019) starting from Plutchik's 8 emotions.
Moreover, including a temporal dimension as a variable could also reveal additional insights such as cases of patients with symptoms moving from one cluster to another over time. The possibility to categorize people with anorexia into different groups according to their symptoms might help with identifying those people at a higher risk of more serious developments of their disorder and also could give some inputs for for a more in-depth discussion of symptomatology in clinical forums.
Finally, it would also be interesting to explore the connection between anorexia and other mental disorders or manifestations such as suicide attempts.