Early Depression Detection in Social Media: Monitoring of Individual Nighttime Dynamics and Large Language Model...

Area of Science:

• Computational psychiatry
• Digital mental health
• Machine learning for healthcare

Background:

• Depression is a global health challenge requiring timely intervention.
• Existing social media-based depression detection methods lack real-time capabilities and interpretability.
• Current early risk detection (ERD) models overlook temporal activity patterns and rely on inefficient sequence models.

Purpose of the Study:

• To develop an efficient, reliable, and interpretable ERD model for depression.
• To extract temporal activity patterns from posting timestamps to enhance risk detection features.
• To utilize large language models (LLMs) for precise text filtering and depression-related factor analysis.

Main Methods:

• Proposed the Monitoring of Individual Nighttime Dynamics (MIND) model integrating circadian activity dynamics and LLM analysis.
• Transformed posting timestamps into temporal activity patterns to derive sleep-related features.
• Employed LLMs for dynamic text filtering, noise reduction, and identification of latent depression risk factors.

Main Results:

• The MIND model significantly outperformed baseline models on the eRisk2017 dataset in early detection.
• Achieved superior sensitivity, specificity, and accuracy in identifying depression risk.
• Demonstrated interpretable and traceable predictions, supporting clinical treatment.

Conclusions:

• The MIND model effectively combines temporal activity patterns and LLM analysis for enhanced ERD.
• Addresses limitations of existing methods regarding interpretability and early-stage data utilization.
• Offers a novel paradigm for social media data application in ERD, facilitating earlier intervention and reducing public health burden.