Multimodal Knowledge Graph-Guided RAG-LLM for Clinical Decision Support in Pediatric Leukemia

This study aims to develop and evaluate a multimodal, knowledge graph-guided retrieval-augmented generation (RAG) framework for clinical decision support in pediatric acute leukemia.

Authoritative pediatric hematology-oncology textbooks were decomposed into text, tables, and figures. Visual and tabular elements were converted into structured textual descriptions using a multimodal large language model (LLM). A biomedical knowledge graph was constructed using LightRAG with gpt-oss-20b and Qwen3 embeddings. System performance was evaluated using 10 clinical questions, with responses generated by the RAG system and GPT-4.5. Nine medical experts (4 pediatric hematology-oncology specialists, 3 nurse specialists, and 2 medical students) conducted blind evaluations, complemented by two LLM evaluators (Claude Sonnet 4.5 and Gemini 3).

The knowledge graph comprised 10,062 nodes and 15,876 edges. In expert evaluation, RAG was preferred in 47.8% of 90 paired comparisons versus 35.6% for GPT-4.5, with higher completeness scores (3.84 vs 3.51, p = 0.016). RAG showed significant advantage for ETP-ALL immunophenotype definition (p = 0.016). LLM-based evaluation consistently favored RAG: Claude Sonnet 4.5 preferred RAG in 6 of 10 questions, and Gemini 3 in 9 of 10 (Fast mode) and 7 of 10 (Thinking mode).

Multimodal graph-based RAG is feasible for clinical decision support in pediatric leukemia. RAG showed complementary strengths to foundation model LLMs, providing added value for questions requiring evidence-dependent information. Unlike LLMs with static training knowledge, RAG can incorporate updated guidelines and protocols without model retraining, particularly relevant in rapidly evolving fields. Further validation regarding privacy and regulatory issues is required before clinical deployment.