Scopus Indexed Publications

Paper Details


Title
ProStack-MH: An Explainable Stacked Transformer Approach to Anxiety-Depression Differentiation in Real-World Mental-Health Discussions

Author
, Taskin khaleque,

Email

Abstract

Automatically distinguishing anxiety from depression in user-generated online text is a clinically important yet technically challenging task, because the two conditions share substantial linguistic overlap while requiring different interventions. Existing work either treats mental-health detection as a coarse binary problem (depressed vs. non-depressed) or relies on single-model transformers that struggle with fine-grained differentiation. We propose ProStack-MH (Probabilistic Stacking for Mental-Health Text Classification), an explainable stacking ensemble that addresses this gap. The system 1) constructs a rigorous, leakage-free corpus by removing staff content, deduplicating posts, and partitioning data by unique user IDs; 2) stacks two complementary transformer encoders—roberta-base (125M parameters) and distilroberta-base (82M parameters)—whose class probability vectors are concatenated and passed to a Logistic Regression meta-learner; and 3) explains predictions post-hoc via SHAP token-level attributions and GoEmotions-based affective profiling. Under a strict user-wise test split against six baselines (TF-IDF+LR/SVM, BiLSTM, DistilBERT, MentalBERT, DeBERTa, RoBERTa-base), ProStack-MH achieves 93.38% accuracy, F1 = 0.9302, MCC = 0.8674, and ROC-AUC = 0.98, outperforming all baselines on every metric. An ablation study confirms that the learned meta-learner—not either encoder alone—is the primary performance driver. SHAP analysis and GoEmotions profiling reveal psychologically coherent, class-discriminative linguistic patterns aligned with DSM-5 descriptions of anxiety and depression. This framework provides a reproducible foundation for computational mental-health monitoring and early-intervention research.


Keywords

Journal or Conference Name
IEEE Access

Publication Year
2026

Indexing
scopus