Abstract
Large language models (LLMs) deployed for clinical question answering remain vulnerable to hallucination when reasoning over structured medical knowledge. We present CLAP-4-QnA (Clinical AI Pipeline for Question & Answering), an agentic Graph-Retrieval- Augmented Generation (Graph-RAG) system that grounds a 4B-parameter clinical reasoning model (Med-Gemma 1.5 4B IT) in a SNOMED-CT knowledge graph hosted in Neo4j. A pipeline of cooperating agents handles privacy-preserving entity masking (spaCy NER), structured extraction (Qwen 2.5 1.5B Instruct), multidirectional graph retrieval combining SapBERT candidate-node identification with vertical (IS_A) and horizontal (“Power-9”) relation traversal, prompt orchestration into a context-dense “Super Prompt”, clinical reasoning (Med-Gemma 1.5 4B IT), and long-context conversational orchestration (Gemini API), within a 16 GB VRAM budget distributed via a Tailscale/WireGuard mesh. On a 41-case subset of the USMLE-style MedQA benchmark, CLAP-4-QnA achieves Graph Recall 0.85, RAGAS Faithfulness 0.78, RAGAS Relevance 0.63, and SapBERT-cosine Semantic Similarity 0.53. We additionally report a retrieval evaluation on the LabQAR dataset (550 items, 364 unique lab-test concepts) using the cambridgeltl/SapBERT-from-PubMedBERT-fulltext embedding model over the full 364-concept candidate pool: canonical lab-test mentions achieve Recall@1 = 1.00, while colloquial paraphrases achieve SapBERT Recall@1 = 0.375 and Recall@5= 0.708, compared with 0.167 and 0.333 for lexical-similarity matching, a roughly 2–3× improvement that supports semantic over lexical retrieval for laboratory reference-range QA. We discuss the agentic design rationale, deployment constraints, and limitations.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_002.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=xIYC3gxJva
BibTex
@InProceedings{MitPra_CLAP4QnA_MICCAISAT2026,
author = { Mittal, Prateek AND Mishra, Aviral AND Singh, Prabal Pratap AND Chauhan, Joohi},
title = { { CLAP-4-QnA: An Agentic Graph-RAG Pipeline for SNOMED-CT-Grounded Clinical Question Answering } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}