1

Open Health Natural Language Processing Consortium

(OHNLP)

Mayo Clinic:Guergana Savova, Ph.D.

James Masanzclinicalnlp@mayo.edu

IBM Watson Research:Anni Coden, Ph.D.Michael Tanenblatt

mednlp@us.ibm.com

2

Overview

• OHNLP? Oh, NLP?

• Demo of a clinical OHNLP system (cTAKES)

• Demo of a medical OHNLP system (MedKAT) with extensions to pathology (/P)

• How can I adapt the system to my data?

• Lively discussion: how can I get involved, OHNLP future steps…

3

Open Health Natural Language Processing Consortium

• www.ohnlp.org (part of caBIG Vocabulary Knowledge Center web presence)

• Goal• Foster an open-source collaborative community around

clinical NLP that can deliver best-of-breed annotators, leverage the dynamic features of UIMA flow-control, and establish the infrastructure for clinical NLP.

• Two open source releases as part of OHNLP• Mayo’s pipeline for processing clinical notes (cTAKES)

• IBM’s pipeline for processing medical notes (MedKAT) and pathology reports (MedKAT/P)

4

Other non-OHNLP clinical NLP Systems

• Proprietary• medLEE (Columbia University)• Topaz (University of Pittsburgh)• Vanderbilt University• caTIES (University of Pittsburgh)• MPLUS/Onyx (University of Utah)• VA Hospital system

• Open Source• i2b2 HITEx (Health Information Text Extraction)

5

Clinical example:clinical Text Analysis and

Knowledge Extraction System (cTAKES)

Presenters:Guergana Savova

James Masanz

6

Overview• cTAKES

• Developed at Mayo Clinic

• Goals:

• Phenotype extraction

• Generic – to be used for a variety of retrievals and use cases

• Expandable – at the information model level and methods

• Modular

• Cutting edge technologies – best methods combining existing practices and novel research with rapid technology transfer

• Best software practices (80M+ notes)

• Commitment to both R and D in R&D

7

cTAKES: Components

• Clinical narrative as a sublanguage

• Core components• Sentence boundary detection (OpenNLP technology)

• Tokenization (rule-based)

• Morphologic normalization (NLM’s LVG)

• POS tagging (OpenNLP technology)

• Shallow parsing (OpenNLP technology)

• Named Entity Recognition• Dictionary mapping (lookup algorithm)

• Machine learning (MAWUI)

• Negation and context identification (NegEx)

8

Output Example: Disorder Object

• “No evidence of unstable angina.”

• Disorder• Text: unstable angina

• Associated code: SNOMED 4557003

• Named entity type: disease/disorder

• Status: current

• Negation: true

9

Methods

• Preliminary results:

• Savova, Guergana; Kipper-Schuler, Karin; Buntrock, James and Chute, Christopher. 2008. UIMA-based clinical information extraction system. LREC 2008: Towards enhanced interoperability for large HLT systems: UIMA for NLP.

• Manuscript with detailed system description and evaluation under review (JAMIA)

10

cTAKES demo

11

Medical example:Medical Knowledge Analysis System

MedKAT and MedKAT/P

Presenters:Anni Coden

Michael Tanenblatt

12

Overview

• MedKAT and MedKAT/P• Developed at IBM

• Goal:

• Identification of concepts and their attributes based on a standard or proprietary terminology/ontology

• /P adaptation to pathology reports – relation extraction

• Modular, Generic, Expandable

• Terminology, Conceptual Model

• Easy adaptation to specific corpus and conventions

• Integration into institutional system

• Ongoing commitment to Research and Development

13

Core Components

• Document structure

• Syntactic tools (tokenization ... Shallow parsing)

• Concept identification

• Negation

• Relationship extraction

Extracted data F-scoreAnatomic site 0.95Histology 0.98Size 1.00Date 1.00Grade 0.98Gross Desc 0.80Lymph Nodes 0.81Primary Tumor 0.82Metastatic Tumor 0.65

14

Document Structure

16

15

Document Structure

17

16

Document Structure

18

17

Output

18

Cancer Disease Knowledge

Representation Model

19

Demos

• Query by Model / Cancer

• Detailed view of annotations in Document Analyzer

• http://domino.research.ibm.com/comm/research_projects.nsf/pages/medicalinformatics.index.html

20

Adaptation

Presenters:Anni Coden

Michael Tanenblatt

21

Adaptation

• Sentence breaks

• Text case

• Part of speech tags

• Shallow parser

• Dictionary lookup

• Document structure

22

Sentence Breaks

23

Sentence Breaks

• Some solutions:• Use annotator to re-break sentences• Retrain tagger

24

Case/Part of Speech Tags

25

Case/Part of Speech Tags

• Some solutions:• Retrain tagger• Use UIMA annotator to create a “true

case” view

26

Part of Speech Tags

27

Part of Speech Tags

• Some solutions:• Retrain tagger• Use dictionary lookup to modify

incorrect tags• Create rule-based annotator to

modify incorrect tags

28

Shallow Parser

29

Shallow Parser

31

30

Shallow Parser

32

31

Dictionary Lookup

• Dictionary entries can be added, changed, deleted

• Dictionary entry attributes can be added, changed, deleted

• Search parameters can be modified

• Post processing filters

• Tokenization of text and dictionary should be the same

32

Document Structure

• Plain text or XML (e.g., CDA)

• Processes specific document section types (e.g., diagnosis)

• Detection of formatting (e.g. bullets)

• Detection of relations between sections

• Making implicit conventions explicit (e.g. meaning of title)

Discussion: Future of OHNLP.ORG

• Provided seed annotators and tools

• Goal: growing community• Annotators, tools• Methodologies• Gold standards

• Common type system for plug-and-play

• What are the hurdles?

34

Hands-on Customization

35

MedKAT

• Dictionary adaptation

• Concept identification parameters

• Document structure detection

36

cTAKES

• Negation window

• Lookup window

• Dictionary modifications

37

Questions?