Transcript Natural Language Processing for Biosurveillance

Development of ConText Tools in Python
Brian E. Chapman, PhD, Glenn
Dayton, Wendy W. Chapman, PhD
Division of Biomedical Informatics
Caveats and Apologies
• I’m not a linguist, computational or otherwise,
nor am I a grammarian, or multilingual, or
particularly well spoken in my native English
• In fact I’m a medical physicist who has drifted
into imaging informatics with an emphasis on
the image part of imaging informatics
• Which is a long way of saying that I got into
this field because of a specific problem
Motivation
• Received NIH funding for computer-aided
detection of pulmonary embolism in CT
pulmonary angiography (CTPA)
• How to identify appropriate cases from
clinical PACS?
Case Identification Approach #1
• Talk to an honest broker
– Who was obviously overworked
– Who used procedure codes from RIS to
identify potential cases
– Who then read the dictated report
– Who then classified the case
– Who nearly fainted when I told her I needed
hundreds of positive cases
– Who then quickly asked, “Do you have a lot of
money?”
Case Identification Approach #2
• Honest broker’s task is perfect for NegEx
– Use procedure codes to identify reports in
MARS repository at University of Pittsburgh
– Use NegEx to classify reports as +/- for PE
– Within minutes find hundreds of cases
– Very happy honest broker
What if you wanted to answer
more questions?
•
•
•
•
Disease uncertainty
Disease temporality
Image quality
Can we a priori specify all of these?
peFinder
• Application to characterize CTPA reports
– Presence or absence of PE
– Temporal state of positive PE
– Uncertainty of disease state
– Technical quality of the exam
For Review: NegEx
scope
Clinical condition:
Negation:
Cough
Negated
Patient denies cough but complains of headache.
No change in the patient’s chest pain.
trigger term
pseudo-trigger
term
termination
term
Python Implementations
• What Drove My Organic Design
– What existed in NegEx
• GUI program written in Tcl/Tk
• Lots of enumerated trigger terms
– What I wanted
• I wanted a package that could be used to build a
variety of accurate applications
• I wanted it to be easy for others to use
• I am an engineer and so lazy
– Generalize relationships
– Replace exhaustive enumeration of trigger terms with
regular expressions
pyConText: Basic Framework
• Item Objects: 4-tuple containing Lexical
and Domain Knowledge
– Literal (label): “pulmonary embolism”
– Category/Concept
– Regular expression
• r‘’‘(pulmonary )(artery )?(embol[a-z]+)’’‘
– Rule
• Directional influence of item in sentence
• Category interaction?
pyConText: Basic Framework
• Item Objects parse sentence to create Tag
Objects within sentences
– Tag Objects interact/modify each other
• Targets
• Modifiers
• Conjunctions
• Prune to eliminate subset tag objects
• Directional Graph represents relationships
Did I Meet My Objectives?
• Accurate
– Yes: JBI 2011
• Modular
– Yes: package in pypi
• Easy for others to use
– Depends on your definition of others
– Wilson, et al. Journal of Pathology Informatics
– Gentili and Chapman RSNA
Did I Meet My Objectives?
• Easy for others to use (continued)
– Can any application relying on user to provide
regular expressions be defined as easy?
Current and Future Work
• Web and GUI applications
– Django
– Django with Twisted for desktop port
Current and Future Work
• Improved Knowledge Representation
– Separating linguistic and domain knowledge
– Integration with external knowledge bases
• Use graphs to further reduce enumeration
of items
– No/definite/evidence of/pulmonary embolism
• Thanks for the invitation
• Looking forward to
– Learning and
– Working and
– Skateboarding
• For the next three weeks