Enhancing explainability and performance of the depression detection model on social media utilizing feature...

Area of Science:

• Computational Linguistics
• Mental Health Informatics
• Artificial Intelligence

Background:

• Depression significantly impacts global populations, with traditional diagnosis methods being time-consuming.
• Social media offers a rich data source for mental health analysis, but current deep learning models lack explainability.
• Existing depression detection models often focus on word-level sentiment, limiting deeper insights.

Purpose of the Study:

• To enhance the explainability and performance of depression detection models.
• To explore the utility of feature engineering and Large Language Models (LLMs) for depression detection.
• To provide more interpretable insights into depression characteristics from social media data.

Main Methods:

• Engineered features capturing emotional status over time (12 features per post).
• Utilized various word embedding and sequence models with an attention mechanism to identify key posts.
• Integrated emotional status and classification results into a fine-tuned LLM for explainability.

Main Results:

• The fine-tuned LLM successfully related emotional status features to specific depression characteristics.
• The proposed method demonstrated improved explainability compared to traditional approaches.
• Experimental results indicate enhanced performance in depression detection.

Conclusions:

• Fine-tuning LLMs with mental health data improves the explainability of depression detection models.
• The integration of temporal emotional status and LLMs offers a promising avenue for more interpretable mental health analysis.
• This research contributes to more efficient and understandable depression diagnosis using social media.