[Paper Review] PDD Graph: Bridging Electronic Medical Records and Biomedical Knowledge Graphs via Entity Linking

Meng Wang, Jiaheng Zhang+5 more|arXiv (Cornell University)|Jul 17, 2017

Biomedical Text Mining and Ontologies11 references3 citations

TL;DR

This paper introduces PDD Graph, a large-scale, high-quality heterogeneous knowledge graph that bridges electronic medical records (EMRs) from MIMIC-III with biomedical knowledge graphs (DrugBank and ICD-9) via entity linking. By extracting and linking 58,030 entities and 2.3 million RDF triples—achieving 94% precision and 85% recall in entity linking—the PDD Graph enables advanced clinical discovery, such as treatment recommendation and drug interaction analysis, via a publicly accessible SPARQL endpoint.

ABSTRACT

Electronic medical records contain multi-format electronic medical data that consist of an abundance of medical knowledge. Facing with patient's symptoms, experienced caregivers make right medical decisions based on their professional knowledge that accurately grasps relationships between symptoms, diagnosis and corresponding treatments. In this paper, we aim to capture these relationships by constructing a large and high-quality heterogenous graph linking patients, diseases, and drugs (PDD) in EMRs. Specifically, we propose a novel framework to extract important medical entities from MIMIC-III (Medical Information Mart for Intensive Care III) and automatically link them with the existing biomedical knowledge graphs, including ICD-9 ontology and DrugBank. The PDD graph presented in this paper is accessible on the Web via the SPARQL endpoint, and provides a pathway for medical discovery and applications, such as effective treatment recommendations.

Motivation & Objective

• To address the gap between clinical data in EMRs and biomedical knowledge graphs by integrating patient-level data with structured medical facts.
• To enable the discovery of clinically relevant relationships between drugs, diseases, and patient outcomes that are missing in standalone biomedical KGs.
• To construct a high-quality, open-access PDD graph dataset linking entities from MIMIC-III to DrugBank and ICD-9 ontology for research use.
• To provide a SPARQL endpoint for researchers to explore relationships across EMRs and biomedical knowledge graphs.
• To improve the accuracy and utility of entity linking between EMR text and external knowledge sources through a novel framework.

Proposed method

• Extract medical entities (patients, drugs, diseases) from MIMIC-III using clinical notes, prescriptions, and diagnoses.
• Apply an entity linking model (ENM) to map extracted entities to corresponding entities in DrugBank and ICD-9 ontology using string matching and context-aware disambiguation.
• Represent the resulting knowledge as an RDF graph using the PDD definition, with triples expressing patient-drug and patient-disease relationships.
• Use SPARQL endpoint (Apache Jena Fuseki) to expose the PDD graph for programmatic access and federated querying.
• Evaluate entity linking accuracy via manual validation on 500 samples, achieving 94% precision and 85% recall.
• Handle edge cases such as non-standard drug names (e.g., 'NS') and short identifiers (e.g., 'N') through error analysis and data curation.

Experimental results

Research questions

• RQ1 Can clinical entities from MIMIC-III be accurately linked to existing biomedical knowledge graphs like DrugBank and ICD-9?
• RQ2 To what extent can EMR-derived patient data be integrated with structured biomedical knowledge to reveal clinically meaningful relationships?
• RQ3 How effective is the proposed entity linking model in mapping diverse clinical terminology to standardized biomedical ontologies?
• RQ4 What is the coverage and quality of the resulting PDD graph in terms of entity linking and triple generation?
• RQ5 Can the integrated PDD graph support advanced clinical applications such as treatment recommendation and drug interaction analysis?

Key findings

• The PDD Graph contains 58,030 entities and 2.3 million RDF triples, forming a large-scale, heterogeneous knowledge graph linking patients, drugs, and diseases.
• 83.4% of drug-taken records in MIMIC-III were successfully linked to DrugBank, with 3,449 out of 4,525 drugs matched.
• 99.9% of diagnosed conditions were linked to ICD-9 ontology, with only two unmatched codes ('71970' and 'NULL').
• The entity linking model achieved 94% precision and 85% recall on a manual validation set of 500 samples.
• The precision of linked entities was confirmed at 93% in a separate manual evaluation of 200 randomly selected links.
• Challenges in linking arose primarily from non-standard drug names (e.g., 'NS') and short identifiers (e.g., 'N'), which lack sufficient context for accurate mapping.