Explainable Text-Based Depression and Suicide Risk Prediction from Social Media Using Deep Learning and Graph Neural...

Area of Science:

• Computational linguistics
• Artificial Intelligence
• Social media analytics

Background:

• Increasing mental health concerns on social media necessitate reliable detection methods.
• Existing computational approaches often lack explainability and efficiency.
• Social media text provides valuable data for mental health surveillance.

Purpose of the Study:

• To propose an interpretable framework for detecting post- and community-level mental health risk on social media.
• To develop a computational method that is reliable, explainable, and efficient.
• To enhance mental health surveillance using social media data.

Main Methods:

• Combined Secretary Bird Optimization (SBO) for feature selection.
• Utilized a BERT-CNN model for post-level analysis.
• Employed a Graph Neural Network (GraphSAGE) for community-level analysis based on semantic similarity and author relations.
• Applied SHAP and LIME for interpretability and uncertainty analysis.

Main Results:

• Achieved 93.1% accuracy, 0.91 F1-score, and 0.944 ROC-AUC on eRisk and CLPsych datasets.
• SBO reduced features by ~38%, improving generalization.
• Graph-based model enhanced representations by capturing relational dependencies.

Conclusions:

• The proposed framework offers an explainable and robust method for detecting mental health risks from text.
• Graph-based representations facilitate community-level analyses.
• Interpretability and uncertainty estimation support human-in-the-loop decision-making.