Using Large Language Models to Detect Depression From User-Generated Diary Text Data as a Novel Approach in Digital...

Area of Science:

• Digital Health
• Computational Psychiatry
• Natural Language Processing

Background:

• Depression poses significant global challenges, impacting productivity and increasing disability.
• Current depression screening tools lack objectivity and accuracy.
• Emerging objective indicators include image analysis, biomarkers, and ecological momentary assessments (EMAs).

Purpose of the Study:

• To detect depression using user-generated diary text via a large language model (LLM).
• To validate semistructured diary text data as a valuable EMA source for depression assessment.

Main Methods:

• Participants completed depression (PHQ) and suicide risk (BSI) assessments.
• A 2-week diary writing period was implemented.
• Performance of LLMs (GPT-3.5, GPT-4) was evaluated using fine-tuning, zero-shot, and chain-of-thought prompting on diary text.

Main Results:

• GPT-3.5 fine-tuning achieved high accuracy (0.902) and specificity (0.955).
• GPT-3.5 without fine-tuning showed the highest balanced accuracy (0.844) with strong recall (0.929).
• Both GPT-3.5 and GPT-4 demonstrated effective depression risk recognition.

Conclusions:

• User-generated text data holds significant clinical potential for depression detection.
• LLMs show promise in analyzing qualitative digital expression for mental health.
• Future research should integrate qualitative digital expression with quantitative measures like activity data.