Graph Attention-Based Curriculum Learning for Mental Healthcare Classification.

Area of Science:

• Computational linguistics
• Mental health informatics
• Machine learning for healthcare

Background:

• Depression diagnosis often relies on subjective assessments.
• Online user-generated content offers a rich source for objective depression indicators.
• Previous methods primarily focused on word analysis in personal statements.

Purpose of the Study:

• To develop and evaluate a Graph Attention Network (GAT) model for classifying depression from online media.
• To improve depression detection accuracy using advanced machine learning techniques.
• To enhance model interpretability by visualizing word contributions to depressive symptoms.

Main Methods:

• Utilized a Graph Attention Network (GAT) model with masked self-attention layers.
• Extended an emotion lexicon using hypernyms for improved feature representation.
• Employed word embeddings to illustrate the contribution of activated words to specific symptoms.
• Implemented a graph-based curriculum learning strategy to handle complex training data.

Main Results:

• Achieved significant improvements in depression classification performance.
• Observed an increase in Receiver Operating Characteristic (ROC) performance metrics.
• Demonstrated enhanced model performance through lexicon expansion and increased vocabulary.
• Validated qualitative agreement from psychiatrists regarding word contribution embeddings.

Conclusions:

• The proposed GAT model effectively classifies depression from online user-generated data.
• Lexicon extension and graph-based curriculum learning are effective strategies for enhancing model performance.
• The embedding visualization technique provides interpretable insights into model predictions for clinical relevance.