Ensemble transformer with post-hoc explanations for depression emotion and severity detection

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Mental health support requires scalable and trustworthy AI solutions.
• Social media text contains valuable indicators of depression-related emotions and severity.
• Existing methods struggle with accuracy and interpretability in depression detection.

Purpose of the Study:

• To develop an ensemble transformer framework (DepTformer-XAI-SV) for detecting depression-related emotions and classifying severity in social media text.
• To enhance model accuracy and interpretability using a weighted soft-voting mechanism and LIME (Local Interpretable Model-agnostic Explanations).
• To address class imbalance challenges inherent in mental health datasets.

Main Methods:

• Integration of four transformer models into an ensemble framework.
• Weighted soft-voting mechanism based on validation macro-F1 scores for improved accuracy.
• Utilizing LIME for explainability, highlighting key linguistic features.
• Employing class-weighted cross-entropy, stratified k-fold splits, and minority-aware sampling to handle class imbalance.
• Evaluation on two benchmark datasets: DepressionEmo and merged depression severity detection (MDSD).

Main Results:

• The DepTformer-XAI-SV model achieved macro-F1 scores of 80.44% on DepressionEmo and 79.88% on MDSD.
• The ensemble framework significantly outperformed individual transformer models and traditional methods.
• Demonstrated substantial improvement in detecting minority classes, crucial for imbalanced datasets.
• LIME successfully identified key linguistic features contributing to depression detection.

Conclusions:

• The proposed ensemble transformer framework offers a scalable and accurate solution for depression detection and severity classification in social media.
• The integration of XAI (Explainable AI) techniques enhances trust and interpretability in mental health AI applications.
• Developed a web application for interactive and interpretable inference, facilitating practical application.