TL;DR
This paper leverages deep learning, specifically a hierarchical attention network, to automatically detect signs of anorexia in social media data. Beyond detection, it delves into interpreting the model's findings to uncover distinct patterns of anorexia symptoms, offering novel psycho-social hypotheses.
What Problem Does This Paper Address?
Anorexia nervosa (AN) is a serious mental health disorder with increasing prevalence, yet it remains under-studied computationally compared to other conditions like depression. While deep learning models show promise for detecting mental health issues, their "black-box" nature poses a significant challenge, especially in the sensitive domain of mental health where interpretability is crucial for trust and clinical application. This paper aims to bridge this gap by not only detecting AN from social media but also by providing interpretable insights into the underlying patterns and user behaviors associated with the disorder.
Methods
The core of this research is a Hierarchical Attention Network (HAN). This deep learning architecture is designed to process sequential data like text, with two levels of attention: one at the word level within a post, and another at the post level within a user's history. The HAN is trained to classify users as either having anorexia or being in a control group.
Beyond the standard text processing, the model incorporates several feature types:
- Content Features: Standard word embeddings.
- Style Features: Bag-of-words representations of function words, capturing linguistic style.
- LIWC Features: Lexicon-based features (Linguistic Inquiry and Word Count) that categorize text based on psychological constructs like emotions, cognitive processes, and social interactions.
- Emotion Features: Specifically, the NRC emotion lexicon is used to capture a fine-grained range of emotions.
To interpret the model's predictions, the authors employ:
- Attention Analysis: Visualizing attention weights to understand which words and posts are most influential for classification.
- User Embeddings: Reducing the model's internal representations of users into a 2D space (using PCA) to visualize clusters and identify distinct user groups.
- Feature-Focused Analysis: Examining the usage of specific LIWC and emotion categories across identified clusters.
- Personality Analysis: Employing the Five Factor Model (Big Five) to analyze personality traits associated with different user clusters.
Datasets
The study utilizes the eRisk Reddit dataset on anorexia. This dataset, collected from specific sub-reddits, contains over 800,000 posts from over 1,200 users, with approximately 10.4% identified as anorexic. The data was collected up to years prior to diagnosis, allowing for early detection analysis. Control cases were selected from users with similar interests to minimize confounding factors. The dataset was split at the user level for training and testing.
Key Insights and Ideas
- Distinct Anorexia Sub-clusters: The user embeddings reveal that individuals with anorexia do not form a single monolithic group. Instead, they cluster into at least two distinct groups (ANO1 and ANO2), suggesting different symptom presentations or stages of the disorder.
- Social Isolation and Group Conflict: Users with anorexia tend to use fewer "we" pronouns, indicating a potential sense of social isolation or conflict with their group. Conversely, they use more first-person "I" pronouns.
- Body and Health Focus: LIWC features related to "ingest," "bio," "health," and "body" are significantly more prevalent in anorexic users' language, reflecting the central role of these themes in the disorder.
- Emotional Landscape: Anorexic users exhibit higher usage of negative emotion words like "anx," "negemo," "disgust," and "fear," particularly in the more severe ANO1 cluster.
- Personality Traits: Users with anorexia show statistically significant differences in personality traits, particularly lower agreeableness and extraversion, and higher neuroticism, with ANO1 users exhibiting more pronounced neuroticism.
- Cognitive Styles: Anorexic users show a tendency towards more rigid cognitive styles ("certainty") and less flexible ones ("tentative"), suggesting a potential effort to reason about reality that is characteristic of conflictual states.
Key Results
- The Hierarchical Attention Network (HAN) achieved the best AUC score among the tested models (including RoBERTa and AIBERT), demonstrating its effectiveness in classifying anorexic users.
- Attention analysis showed that posts later in a user's history are more heavily weighted, suggesting that a user's social media activity becomes more indicative of their mental state over time.
- LIWC feature analysis identified significant differences in the usage of categories like "ingest," "health," "body," and various emotion terms across the three groups (ANO1, ANO2, Control). For example, "ingest" was used 1.15% by ANO1, 0.82% by ANO2, and only 0.38% by controls.
- Personality analysis revealed statistically significant differences in Agreeableness (-), Extraversion (-), and Neuroticism (+) between anorexic users and controls.
- Cognitive style analysis showed significant differences in "certainty" and "tentative" usage, with ANO1 users showing higher "certainty" and lower "tentative" compared to ANO2 and controls.
Limitations and Future Work
- False Negatives: The paper acknowledges the problem of false negatives, particularly with the ANO2 cluster showing overlap with control cases, which could lead to missed diagnoses.
- Interpretability of Attention: The authors note that recent studies question the inherent interpretability of attention mechanisms, prompting the exploration of additional techniques.
- More Semantically Rich Representations: Future work could benefit from using more advanced word embeddings for features beyond simple bag-of-words.
- Temporal Dynamics: Incorporating a temporal dimension could reveal how users transition between symptom clusters over time.
- Connection to Other Disorders: Exploring the link between anorexia and other mental health conditions (e.g., suicide attempts) is a promising avenue.
- Clinical Application: The authors suggest that identifying patient clusters could aid in risk stratification and provide insights for clinical discussions, particularly regarding the concept of anorexia as a "communicative disorder."
Summary updated