For mental health AI systems, structured annotation with expert-in-the-loop feedback and traceable clinical reasoning produces more reliable, interpretable datasets than traditional labeling—critical for building trustworthy diagnostic support tools.
This paper presents a framework for creating high-quality, explainable depression datasets by combining LLM-assisted annotation with expert feedback. The system produces DSM-5-aligned labels with clinical evidence and reasoning traces, using a dual-memory architecture that learns from expert corrections to improve future annotations without retraining.