Self-Evolving Human-Centered Framework for Explainable Depression Symptom Annotation
Merged summary
TL;DR - A self-evolving, expert-in-the-loop framework uses LLM-assisted labeling plus expert verification to build explainable, DSM-5-TR-aligned datasets for Major Depressive Disorder annotation—addressing the annotation-quality bottleneck in mental health XAI.
- Three-stage pipeline: candidate evidence selection from text, criterion-level DSM-5-TR analysis, and case-level synthesis producing diagnostic and severity labels; explicitly for dataset construction, not clinical diagnosis.
- A dual-memory design (Example Memory + Reflection Memory) internalizes expert feedback to iteratively improve annotations without retraining.
- Exports clinical evidence, reasoning traces, and edit histories for full auditability, targeting transparency and downstream model interpretability.
- A pilot study on expert-reviewed samples reports improved annotation consistency/explainability and reduced manual revision; multi-cycle evaluation of the memory mechanism is left to future work, so quantitative results are limited.
Sources (1)
Self-Evolving Human-Centered Framework for Explainable Depression Symptom Annotation
TL;DR - A self-evolving, expert-in-the-loop framework uses LLM-assisted labeling plus expert verification to build explainable, DSM-5-TR-aligned datasets for Major Depressive Disorder annotation—addressing the annotation-quality bottleneck in mental health XAI.
- Three-stage pipeline: candidate evidence selection from text, criterion-level DSM-5-TR analysis, and case-level synthesis producing diagnostic and severity labels; explicitly for dataset construction, not clinical diagnosis.
- A dual-memory design (Example Memory + Reflection Memory) internalizes expert feedback to iteratively improve annotations without retraining.
- Exports clinical evidence, reasoning traces, and edit histories for full auditability, targeting transparency and downstream model interpretability.
- A pilot study on expert-reviewed samples reports improved annotation consistency/explainability and reduced manual revision; multi-cycle evaluation of the memory mechanism is left to future work, so quantitative results are limited.