Related Experiment Videos
Extracting medical knowledge for a coded problem list vocabulary from the UMLS Knowledge Sources
J W Hales1, K M Schoeffler, D P Kessler
1Division of Medical Informatics, Duke University Medical Center, Durham, North Carolina, USA.
Proceedings. AMIA Symposium
|February 3, 1999
Summary
Extracting relationship knowledge from the Unified Medical Language System (UMLS) for problem list vocabularies is challenging. A multiplicative decline in matching rates hampers integration efforts, limiting the augmentation of existing problem lists with UMLS relationship data.
Area of Science:
- Medical Informatics
- Knowledge Representation
- Computational Linguistics
Background:
- The Unified Medical Language System (UMLS) Knowledge Sources offer extensive medical knowledge.
- Integrating UMLS concept relationships into existing problem list vocabularies is a key challenge in medical informatics.
Purpose of the Study:
- To assess the feasibility of extracting and integrating UMLS concept relationship information into The Medical Record (TMR) problem list vocabulary.
- To quantify the success rate of matching TMR terms to UMLS concepts and their relationships.
Main Methods:
- Matched TMR problem list terms to UMLS Metathesaurus concepts using normalized string matching.
- Identified UMLS concepts with 'parent' relationships to matched TMR concepts.
- Translated identified relationships back to TMR codes for integration.
Main Results:
- 67% of TMR codes matched to UMLS concepts.
- 91% of matched concepts had parent relationships, but only 28% of these relationships mapped back to existing TMR codes.
- Only 19% of the original TMR problem list could be augmented with UMLS relationship information.
Conclusions:
- A multiplicative decline in matching rates occurs at each step of term-concept, concept-relationship, and concept-term matching.
- This decline significantly hampers the extraction of relationship knowledge from UMLS for non-UMLS source vocabularies.
- Efforts to enrich problem lists with UMLS relationship data face substantial challenges due to these matching limitations.