Mental Health Disorder Detection in Social Media Using Deep Learning

Abstract

Depression, anxiety, and post-traumatic stress disorder (PTSD) are among the fastest-growing health concerns worldwide, and social media has become an important source of early behavioural cues for these conditions. Existing detection systems, however, are limited in three respects: they usually frame the task as binary classification (disordered versus not), rely on a single platform's data, and seldom combine sequential language modelling with deep contextual semantics within one architecture. This work addresses that gap through an ensemble that pairs a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model with a Bidirectional Long Short-Term Memory (Bi-LSTM) network enhanced with self-attention. Each branch produces its own class probabilities, which are then combined using a weighted soft-voting rule (w1 = 0.55 for BERT, w2 = 0.45 for Bi-LSTM+Attention) tuned through grid search on a held-out validation split. The system is evaluated on two benchmarks, the CLPsych 2015 Twitter dataset and the Reddit Mental Health dataset, which together span over 62,000 labelled posts across depression, anxiety, PTSD, and control categories. The Reddit dataset shows only mild class imbalance, handled with class-weighted cross-entropy, whereas CLPsych 2015 skews toward low-risk posts. On both datasets combined, the ensemble achieves 94.4% accuracy, 93.8% precision, 94.1% recall, and 93.9% F1-score, surpassing standalone BERT, Bi-LSTM with attention, a conventional LSTM, and a Random Forest baseline. These results indicate that combining a transformer's contextual strength with a recurrent, attention-guided representation of sequence order provides a measurable advantage for multi-class mental-health detection in everyday social media text.