Advancing Mental Disorder Detection: Evaluating Transformer and LSTM-Based Models on Social Media

Mental disorders such as suicidal ideation, bipolar disorder, and depression are common, and early detection is crucial for effective intervention. Social media has emerged as a bountiful source of mental illness indicators since many individuals express their psychological state over the internet. This thesis explores social media text usage for identifying mental disorders based on advanced Natural Language Processing (NLP) techniques. The scope spans four principal tasks: early prediction of suicidal ideation; bipolar disorder-related language classification; binary classification distinguishing between individuals with mental disorders and healthy controls; and multi-class diagnosis of six individual conditions (Attention-deficit/hyperactivity disorder (ADHD), anxiety, bipolar disorder, depression, Complex Post-Traumatic Stress Disorder (CPTSD), and schizophrenia) versus a control group. The performance of recent Transformer-based models (BERT, RoBERTa, DistilBERT, ALBERT, ELECTRA) and Long Short-Term Memory (LSTM) networks augmented with attentional and contextual embedding is investigated, evaluating their efficacy across these tasks. Experimental results show that Transformer models consistently display excellent classification performance, whereas LSTM-based models offer competitive accuracy with reduced computational expense when integrated with contextual embedding. Above all, light Transformer versions (e.g., DistilBERT) achieve near that of the large models’ accuracy while being significantly more efficient, indicating their feasibility for deployment in real-time or resource-limited settings. Model robustness regarding identifying general mental disorder is further assessed, and it is concluded that top-performing models exhibit robust performance under various data conditions. The main contributions of this work include the construction of annotated social media corpora, an exhaustive benchmark of state-of-the-art NLP models on mental health prediction, and new findings on models’ efficiency-performance-generalizability trade-offs. These findings advance the development of feasible NLP-based early intervention monitoring tools for mental health.