Infrastructure for Semantic Expansion and Curation of the RadLex Ontology

1
Infrastructure for Semantic Expansion and
Curation of the RadLex Ontology

• Rebecca Hazen Alexander van Esbroeck
• Northwestern University
• Dr. David Channin, Mentor

2
Background

• RadLex - Radiology Lexicon
• Reduce variation and improve clarity in radiology
  reports
• 11,962 terms over 12 categories

3
Establishing the need

• Missing many terms
• Imaging Observations
• Imaging Observation Characteristics
• Committee dependent development process
• Manual, time consuming, expensive
• Larger lexicons are harder to manage
• Difficult to sustain

4
Proposed Solution

• Develop an automatic term extraction system
• Focusing on Imaging Observation and
  Characteristics
• Accelerate the expansion of RadLex
• Decrease the demands on committees
• Propose lists of strong candidates for inclusion
• Reduce development costs

5
Processing System Description

• Collect free full-text articles from medical
  journals
• Identify new terms using LexEVS and NLP
  techniques
• Create ranked lists of imaging observations and
  characteristics

6
Processing System Overview
LexEVS
Concepts/Relationships
Article Text
Article Finder
Candidate Term Identification
Data/Annotations
Ranked Lists of Imaging Observations Characteris
tics
Context Processing
7
LexEVS

• LexEVS was developed by NCI, NIH, caBIG, Mayo
  Clinic
• Designed to fulfill a community need for
  standards in storing, accessing, managing and
  distributing controlled vocabularies
• Combination of LexBIG, LexGrid, EVS
• Programmable interfaces for accessing and
  distributing controlled vocabularies
• Provides a common API

8
UIMA Architecture

• Framework for processing large collections of
  documents
• Processing modules can be connected into
  pipelines

9
Article Finder

• Locates and retrieves scientific articles
• Searches PubMed
• Returns free full-text, English, HTML articles.
• Removes tags and extracts the article text

10
Articles Processed

• 1,128 Documents
• ImagingCTMRPETX-rayUSangiographytomography
  findings Title

11
Candidate Phrase Identification

• Identifies a list of candidate phrases from the
  articles
• Tokenizer
• Part-of-speech Tagger
• Linguistic Filter
• Extracts sequences of words matching a specific
  pattern
• Increased renal enhancement
• -ed verb, adj, noun

12
LexEVS Annotator

• Use LexEVS to access vocabularies
• RadLex 2.0 NCI Thesaurus HL7 CTCAE
• Determine if phrases exist in RadLex as a single
  concept
• Retrieve vocabulary metadata
• What us that
• Annotate the document
• Build database of annotations
• Develop inclusion/exclusion criteria

13
LexEVS Annotator
14
Context Processing

• Find indicator words that are associated with
  existing RadLex terms
• Assign weights to those words as a function of
  the number of RadLex terms with which they are
  associated.

Focal confluent fibrosis can occur in the
cirrhotic liver as a hepatic mass in
approximately 14 of cases . This fibrosis
is accompanied by atrophy of the affected liver
parenchyma and retraction of the overlying
liver capsule (Figure 9 ).
15
Context Processing

• Use those indicator words to identify new
  phrases
• Score new phrases as a function of the strength
  of their association with the indicator words.

Less extensive findings included interlobular
septal thickening. Interlobular
septal thickening was seen in 32 patients
(89). A luminal mass was considered to be
present if there was a soft-tissue mass in the
lumen that arose from the
bowel wall.
16
Phrase Ranking

• Calculate a termhood1 value for each phrase
• Termhood is based on a combination of
• Nesting
• Context Scores
• Length
• Orthography
• Stop List

1 termhood refers to the likelihood that a
candidate is a real term 2
17
Term Splitting

• Phrases typically consist of an observation
  accompanied by one or more characteristics of
  that observation
• Term splitting splits phrases into component
  characteristics and observations
• Based on frequency ratios
• Makes two new ranked lists

Candidate Term mediastinal soft tissue
infiltration

• mediastinal
• soft tissue
• infiltration

18
Results

• imaging observations
• imaging observation characteristics
• precision
• Precision is defined as .

19
Conclusions

• LexEVS is a powerful tool for exploiting a
  variety of controlled vocabularies
• Automatic term extraction can identify new
  imaging observations and observation
  characteristics
• Adjusting context and processing can lead to
  other kinds of terms
• Broader searches for articles will lead to larger
  collections of terms

20
Future Work

• Use syntactic structure to improve extraction
• Automatic identification of relationships
• Infrastructure for distributed editing
• Semantic Wiki

21
Selected References

• 1. Langlotz CP. RadLex a new method for indexing
  online educational materials. Radiographics. 2006
  Nov-Dec26(6)1595-7.
• 2. Frantzi K, Ananiadou S, Mima H. Automatic
  recognition of multi-word terms the
  C-value/NC-value method. International Journal on
  Digital Libraries 2000 3(2)115-130.
• 3. Baneyx A, Charlet J, Jaulent M. Building an
  ontology of pulmonary diseases with natural
  language processing tools using textual corpora.
  International Journal of Medical Informatics 2007
  76(2-3) 208-215.
• 4. Zhou L, Tao Y, Cimino J, Chen E, Liu H,
  Lussier Y, Hripcsak G, Friedman C. Terminology
  model discovery using natural language processing
  and visualization techniques. Journal of
  Biomedical Informatics. 2006 39(6)626-636.
• 5. Church K, Hanks P. Word association norms,
  mutual information, and lexicography.
  Computational linguistics 1990 16(1)22-29.
• 6. Snow R, Jurafsky D, Ng A. Learning syntactic
  patterns for automatic hypernym discovery.
  Advances in Neural Information Processing Systems
  2005 171297-1304.

22
Example Query

• NCI MetaThesaurus,
• Cronkhite-Canada Syndrome,
• exactMatch

23
Structure
24
References

• Retrieve Candidate Terms
• Query LexEVS
• Selecting Scheme, Search Algorithm, Restrictions
• Add Tags
• Coding Scheme, Concept Code, etc
• Store in database

25
UIMA Pipeline Lexicons

• LexEVS Annotator
• Marks existing observations and characteristics,
  as well as anatomic parts, treatments, etc.
• Can be used for context information later on.
• Filters out existing terms.

Mock-up of Annotation Results
26
Context Detection

• Looks for the identified context words
  before/after candidate terms.
• Calculates a context value for each term based on
  the frequency of context terms and their weights.

Sample Context Terms
27
Term Ranking Results
28
Room for Improvement

• Improve precision of candidate term selection
• Use context-term groups as classifiers.
• Segment the article and identify term-rich
  areas
• Normalize the context and inverse document
  frequency distributions.
• Context phrases
• Improve linguistic filter and use more POS
  information
• Stop list additions