Transforming Annotated Clinical Narratives into Pruned Interoperable Knowledge Graphs with SNOMED CT

Area of Science:

• Medical Informatics
• Natural Language Processing
• Knowledge Representation

Background:

• Clinical narratives present challenges in processing due to unstructured text, abbreviations, and jargon, hindering semantic interoperability.
• Knowledge graphs (KGs) offer a structured representation, but require optimization for clinical data.

Purpose of the Study:

• To develop and evaluate a context-aware pruning method for clinical knowledge graphs derived from SNOMED CT.
• To improve the semantic interoperability and reduce the complexity of clinical KGs while retaining essential information.

Main Methods:

• Constructed KGs from the n2c2 2019 dataset using SNOMED CT and UMLS.
• Merged subgraphs, removed redundancies, and enforced interoperability via mappings to ICD-10-CM, RxNorm, and LOINC.
• Applied context-aware pruning to remove non-interoperable nodes and evaluated expert-validated pruning quality.

Main Results:

• Radical pruning enhanced interoperability but led to loss of contextual detail.
• Context-aware pruning successfully preserved clinically meaningful structures.
• The resulting KGs were compact, semantically rich, and suitable for downstream NLP tasks.

Conclusions:

• Context-aware pruning is an effective strategy for creating interoperable and informative clinical KGs from unstructured narratives.
• This approach facilitates focused representations for clinical NLP applications.
• Future research will explore hybrid pruning techniques and direct clinical NLP applications.