Toward explainable AI (XAI) for mental health detection based on language behavior

Area of Science:

• Computational linguistics
• Artificial Intelligence
• Mental Health Informatics

Background:

• Deep learning models prioritize predictive accuracy over interpretability in mental health prediction.
• Lack of transparency in AI decision-making is critical in healthcare applications.
• Explainable AI (XAI) is needed for trustworthy psychiatric diagnosis and prediction.

Purpose of the Study:

• To systematically investigate XAI approaches for automated mental disorder detection from social media language.
• To evaluate the trade-off between accuracy and interpretability in predictive mental health models.
• To identify informative language features for specific mental health conditions.

Main Methods:

• Built BiLSTM models using human-interpretable features (syntactic, lexical, readability, cohesion, stylistics, topic, sentiment/emotion).
• Compared BiLSTM models against a black-box transformer model adapted for mental health.
• Enhanced transformer interpretability using multi-task fusion learning (emotion, personality traits) and explanation techniques (LIME, AGRAD).

Main Results:

• Extensive experiments evaluated accuracy-interpretability balance across models.
• Feature ablation identified key linguistic features for specific mental health conditions.
• XAI techniques (LIME, AGRAD) revealed word categories influencing transformer predictions.

Conclusions:

• XAI approaches offer a transparent alternative to black-box models for mental health prediction.
• Combining interpretable features with advanced AI techniques enhances model understanding.
• This research provides a framework for developing explainable AI in mental health informatics.