Reproduce “DepressLLM: Interpretable domain-adapted language model for depression detection from real-world narratives” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sehwan Moon, Aram Lee, Jeong Eun Kim, Hee-Ju Kang, Il-Seon Shin, Sung-Wan Kim, Jae-Min Kim, Min Jhon and 1 more

DepressLLM, an interpretable, domain-adapted language model, accurately detects depression from real-world narratives by leveraging a novel Score-guided Token Probability Summation (SToPS) module for reliable confidence estimates and natural-language explanations. The model achieved an AUC of 0.920 on the DAIC-WOZ dataset and its high-confidence predictions often aligned with expert psychiatric judgment more closely than self-reported PHQ-9 scores.

Key insights

  • Training on high-quality, clinically validated narrative data (like TREND-P) significantly improves an LLM's ability to identify nuanced linguistic markers of depression.
  • Providing interpretable predictions alongside reliable confidence scores (via SToPS) is critical for clinical trust and can help overcome limitations of self-reported mental health assessments.
  • The model's ability to predict a continuous range of PHQ-9 scores, rather than binary labels, enhances its sensitivity to varying degrees of depressive severity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.