An Explainable Retrieval-Augmented Clinical Decision Support Chatbot for Mental Disorder Assessment Using DeepSeek R1 Distill-Llama-8B

Abstract

This study presents a retrieval-augmented clinical decision support chatbot designed to assist mental health assessment through knowledge-grounded conversational reasoning. The system integrates document chunking, semantic embeddings, and vector-based retrieval with the DeepSeek-R1 Distill-Llama-8B model to generate diagnostically relevant responses grounded in structured psychiatric knowledge. Diagnostic text was processed into semantic chunks, embedded using the all-MiniLM-L6-v2 model, and indexed in a ChromaDB vector database to enable efficient context retrieval. Evaluation across multiple simulated clinical scenarios demonstrated strong performance, achieving perfect relevance (1.00), high faithfulness (0.95), and mean cosine similarity of 0.76 between generated responses and reference diagnostic descriptions. The retrieval module also achieved robust results with recall of 1.0, precision of 0.79, and an F1-score of 0.88. These findings demonstrate that retrieval-augmented large language models can provide transparent, contextually grounded, and explainable support for conversational mental health decision systems.

Ähnliche Arbeiten