A large language model framework for knowledge graph construction of randomized controlled trials for evidence synthesis and querying
Rumjot Kaur, Satwinder Singh
IISE Transactions on Healthcare Systems Engineering
Problems Identified (4)
PICO structured representation gap: Research has automated PICO element extraction but has given limited attention to structured representations of PICO elements.
Scarce irreproducible PICO KGs: Existing PICO-based knowledge graphs are scarce, domain-specific, and difficult to reproduce.
PICO structured representation gap: Research has automated PICO element extraction but has given limited attention to structured representations of PICO elements.
Scarce irreproducible PICO KGs: Existing PICO-based knowledge graphs are scarce, domain-specific, and difficult to reproduce.
Proposed Solutions (4)
EBM-KG reusable PICO knowledge graph: The study presents EBM-KG as a reusable generic PICO-based knowledge graph constructed from the EBM-NLP dataset.
LLaMA-based KG construction framework: The proposed semi-automated framework uses a fine-tuned LLaMA model for entity semantic normalization and instantiates the graph in Neo4j with Cypher queries.
EBM-KG reusable PICO knowledge graph: The study presents EBM-KG as a reusable generic PICO-based knowledge graph constructed from the EBM-NLP dataset.
LLaMA-based KG construction framework: The proposed semi-automated framework uses a fine-tuned LLaMA model for entity semantic normalization and instantiates the graph in Neo4j with Cypher queries.
Results (3)
Large PICO KG resource:
PICO extraction integration performance:
Clinically relevant question answering:
Research Domain
Evidence-based medicine knowledge graphs for randomized controlled trials