Knowledge Representation and Indexing Using the Unified Medical Language System

Knowledge Representation and Indexing Using the Unified Medical Language System

Kenneth Baclawski*

Joseph “Jay” Cigna*

Mieczyslaw M. Kokar*

Peter Major † Bipin Indurkhya ‡ * Northeastern University† Jarg Corporation‡ Tokyo University of Agriculture and Technology

Purpose

Biomedical Information Searches

Ontologies & the UMLS

Knowledge Representation

Input - Natural Language Processing

Retrieval - Ontologies & Semantic Frameworks

Information Visualization - Keynets

Results of Usability Studies

Problem: Low Quality Search Searching using keyword matching often has high

volume and low precision. Discrete keywords do not represent knowledge. Result of a search are not be arranged in a

semantically relevant way. Examining search results is often tedious. Search results include only textual documents.

Introduction

Introduction

Solution: Ontologies

Model for knowledge extraction/management using a domain-specific vocabulary and theories expressing the meaning of the vocabulary within the community using the vocabulary.

Advantages of Ontologies

Allows semantically correct retrieval based on domain specific criteria.

No limit to the depth of knowledge that can be represented, managed and retrieved.

Multiplicity of information objects retrieved:images, video, sound, etc. as well as text.

Results of a search are grouped by how documents are relevant to the whole query.

The ontology can be updated as new terminology and relationships are introduced.

UMLS

US National Library of Medicine since 1986 Overcomes retrieval problems

– Differences in terminology– Distributed database sources

Develops machine-readable “knowledge sources”

Allow researchers and health professionals to retrieve and integrate electronically available biomedical information.

Free Iteratively refined and expanded from

feedback Maps many different names for the same

concept Grateful Med and PubMed are applications

of the UMLS

Semantic Categories – > 130 semantic categories

Semantic Relationships– “ is a “, “ part of”, “disrupts”

Semantic Concepts (Vocabulary)– > 1,000,000 concepts map to categories

Natural Language Processing using an Ontology

syntactic

semantic

Keynets

A technique for representing information in a visual manner that can be manipulated into meaningful associations for refinement of the knowledge extracted.

Exploits human – computer interactivity inherent in knowledge processing.

Based on Information Visualization Concept (Schneiderman, 1998)

Knowledge Representation using the UMLS and Keynets

Acyclic directed graph. Provides a consistent categorization for all concepts. Shows the important relationships between the

concepts. NLP using the UMLS produces Keynets, a new

search strategy for knowledge processing of biomedical information.

“Fc-receptors on NK cells”

Usability Study

The purpose was to explore the reactions of users to different representations of biomedical information– Keywords: Fc-receptors, cells, NK cells

– Keynet: Sample: n = 11; MD, PhD, Biomedical engineers,

Pharmacologists - individuals who would typically be required to search for biomedical information

UMLS Keywords and Keynets

Survey Format

Three SectionsI. Demographics.

II. 9 semantic differential focused questions.

III. Open ended questions to assess subjects overall impressions of using keynets and information visualization for knowledge representation,

Semantic Differential Question

scale 1-9 , 0 = N/Ae.g. confusing/clear 1 most like first word or “confusing” 9 most like the last word “clear”

confusing clear 1 2 3 4 5 6 7 8 9

Semantic Differential Question

1a How would you rate the Keynet version in its ability to represent the biomedical text given?

confusing clear 1 2 3 4 5 6 7 8 9

1b How would you rate the Keyword version in its ability to represent the biomedical text given?

confusing clear 1 2 3 4 5 6 7 8 9

UMLS Keywords and Keynets

1 3 5 7 9

confusing/clear

terrible/w onderful

frustrating/satisfying

dull/stimulating

diff icult/easy

w eak/pow erful

rigid/f lexible

inconsistent/consistent

ambiguous/precise

Keynet

Keyword

Que

stio

nSurvey Results

n=11

Score

Results of Usability Study

Level of Understanding of Keynets– Remarkably high given short time to complete study,

population diversity, different examples used.– Example – missing relationship detected (7 of 11)

Limit Complexity– Representations should be concise

drilling down only at the user’s request Keywords versus Keynet

– No statistical difference, Keynets are as least as useful as Keywords in representation of biomedical information retrieval.

Summary

A new strategy is suggested for searching and retrieving biomedical information using NLP, the UMLS and Keynet displays of the retrieved results.

Issues of semantic versus syntactic representations for biomedical information retrieval.

Issues relating information visualization for the processing of biomedical information retrieval.

Conclusion

•Consider the computer-human interactivity issues

•A picture is worth a thousand keywords!!

Acknowledgement This project was performed as part of the “Biomedical Science Information Retrieval

and Management” project supported by grant # 1 R43 LM06665-01 from the National Institute of Health (NIH). Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the NIH.

A portion of this study was conducted in part at Jarg Corporation, 332b Second Ave., Waltham, MA 02451-1104.

Travel expenses for this presentation were provided by a grant from the Dept. of Energy.

www.jarg.com

Addendum

Technical information related to Keynetshttp://www.ccs.neu.edu/home/kenb/key/