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

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Pediatric Oncology

Background:

• Clinical decision support systems (CDSS) are crucial for managing complex pediatric hematologic malignancies.
• Existing CDSS often lack the ability to integrate diverse data types or rapidly adapt to evolving medical knowledge.
• Large language models (LLMs) show promise but may struggle with evidence-dependent queries and require extensive retraining for updates.

Purpose of the Study:

• To develop and evaluate a multimodal, knowledge graph-guided retrieval-augmented generation (RAG) framework for clinical decision support in pediatric acute leukemia.
• To assess the performance of the RAG system against a leading LLM (GPT-4.5) using expert and LLM-based evaluations.
• To determine the feasibility and added value of integrating multimodal data and knowledge graphs into a RAG system for specialized medical domains.

Main Methods:

• Authoritative pediatric hematology-oncology textbooks were processed into text, tables, and figures.
• A multimodal LLM converted visual and tabular data into structured textual descriptions.
• A biomedical knowledge graph was constructed using LightRAG with specific embeddings (gpt-oss-20b, Qwen3).
• The RAG system and GPT-4.5 generated responses to 10 clinical questions, evaluated by 9 medical experts and 2 LLM evaluators.

Main Results:

• The knowledge graph contained 10,062 nodes and 15,876 edges.
• In blind expert evaluations, the RAG system was preferred over GPT-4.5 in 47.8% of comparisons (vs. 35.6%), with significantly higher completeness scores (p=0.016).
• RAG demonstrated a significant advantage in defining ETP-ALL immunophenotype (p=0.016).
• LLM-based evaluations consistently favored RAG, with Claude Sonnet 4.5 preferring RAG in 6/10 questions and Gemini 3 in 7-9/10 questions.

Conclusions:

• Multimodal, graph-based RAG is a feasible approach for clinical decision support in pediatric leukemia.
• The RAG system offers complementary strengths to foundation LLMs, particularly for evidence-dependent queries.
• RAG's ability to incorporate updated guidelines without retraining is a significant advantage in rapidly evolving medical fields.
• Further validation concerning privacy and regulatory compliance is necessary before clinical implementation.