Transcript Task: Disease Maintenance Summarization

Challenges in Evaluating Natural
Language Processing Systems
for Military Health Records
Carol Friedman, PhD
Columbia University/MedLEE Applications
Technologies
Lawrence Fagan, MD, PhD
Stanford University/MedLEE Applications
Technologies
IEBI Workshop-10/23/07
Outline
• NLP evaluation issues
• Ideal evaluation of NLP output requires
consideration of the context of the
applications
• Catalog of common NLP applications in
biomedicine and the implication for evaluation
IEBI Workshop-10/23/07
Outline
• NLP evaluation issues
• Ideal evaluation of NLP output requires
consideration of the context of the
applications
• Catalog of common NLP applications in
biomedicine and the implication for evaluation
IEBI Workshop-10/23/07
Different Evaluation
Objectives
•
Different NLP communities have
different objectives and traditions
Improvement of:
–
–
–
–
–
Science of NLP
Science of biomedical NLP
Biological research
Clinical research
Clinical care
IEBI Workshop-10/23/07
Evaluation Objectives Determine
• Evaluation design
• NLP requirements
– Type of information needed
• Medical terms with/without modifiers
• Clinical & other external knowledge
– End product
• Codes, facts, yes/no categories
IEBI Workshop-10/23/07
Evaluation to Improve
Clinical Research and Care
Issues to Consider
IEBI Workshop-10/23/07
• Need to start with a concrete clinical
goal
– Detect potential case of tuberculosis in
chest x-ray report for isolation
– Detect positive mammography reports for
follow up
– Find new adverse events to find ways to
avoid them
IEBI Workshop-10/23/07
Type of Task:
Broad vs. Narrow
• Very specific application
– Identify reports of patients who smoke
– Identify x-ray reports positive for pneumonia
• General application
– Data mining & knowledge discovery
– Generate patient problem list
IEBI Workshop-10/23/07
Application Requires
NLP + External Knowledge
• Structural knowledge
– Extract diagnoses from Diagnosis Section of
Discharge Summaries
486 (pneumonia)
• Coding knowledge
for infiltrate in cxr
– ICD-9 coding of x-ray reports for billing
• Clinical knowledge
– Identifying x-ray reports indicating pneumonia
• ~ 38 different combinations of findings & modifiers
IEBI Workshop-10/23/07
NLP Components
• Different steps of process impact results
NLP Components
Clean-up,
recognize text
portions and
boundaries, …
Preprocess
IEBI Workshop-10/23/07
Recognize entities,
relations, generate
codes, …
Extraction
Engine
CXR Findings
opacity
mod: patchy
loc: left lung
........
Clinical logic for
application
Post-process
Pneumonia: possible
NLP Components
Clean-up,
recognize text
portions and
boundaries, …
Preprocess
Recognize entities,
relations, generate
codes, …
Extraction
Engine
CXR Findings
opacity
mod: 5x5cm
loc: left lung
........
IEBI Workshop-10/23/07
Clinical logic for
application
Post-process
Pneumonia: unlikely
Use of Experts
•
•
•
•
•
Need guidelines and examples
How much to train
Inter-annotator agreement & resolution
Borderline cases confound results
Granularity issues
Fever (mod: persistent)
– Comparability
IEBI Workshop-10/23/07
Persistent fever
SNOMED codes:
persistent fever
chronic persistent fever
prolonged fever
fever (mod: persistent)
Document Heterogeneity &
Complexity of Text
• Chief complaints (‘well baby 3 mo’, ‘c/f/h’)
• Discharge summaries, radiology reports
• Reports with structured & unstructured
information
• Telegraphic notes
• Special templates
IEBI Workshop-10/23/07
“Well-Structured” Reports:
Chest Radiology Report
CLINICAL INFORMATION:
F/U.
IMPRESSION:
MODERATE PULMONARY VASCULAR CONGESTION AND
INTERSTITIAL EDEMA SHOWS NO SIGNIFICANT CHANGE FROM
3/25 THROUGH 3/27/95. SIDE HOLE OF THE NG TUBE IS NEAR
THE EG JUNCTION. DEVELOPMENT OF RIGHT BASILAR
ATELECTASIS ON 3/27/95.
DESCRIPTION:
A series of portable chest x-rays demonstrate worsening but stable
vascular congestion and interstitial edema from 3/25 through 3/27/95.
The NG tube side hole is seen near the EG junction. A duo- tube is
seen extending into the stomach, but its distal tip is not seen. A
tracheostomy is seen in good position.
……………………………………………………..
IEBI Workshop-10/23/07
Mixed Structure:
Catheterization Report
IEBI Workshop-10/23/07
Poorly Structured Report:
Telegraphic Note
Admit 10/23
71 yo woman h/o DM, HTN, Dilated CM/CHF,
Afib s/p embolic event, chronic diarrhea, admitted
with SOB. CXR pulm edema. Rx’d Lasix.
All: none
Meds Lasix 40mg IVP bid, ASA, Coumadin 5,
Prinivil 10, glucophage 850 bid, glipizide 10 bid,
immodium prn
Hospitalist=Smith PMD=Jones Full Code,
Cx>101
IEBI Workshop-10/23/07
Reducing Potential Bias
NLP developers should avoid
– Designing study
– Being involved in choice or determination
of reference standard
– Correcting bugs
– Changing system
– Performing actual evaluation
IEBI Workshop-10/23/07
Analyzing Results & Errors
• Determine effect of components on performance
–
–
–
–
NLP vs. domain knowledge
Document characteristics/quirks
Frequency of adding/updating clinical terms
Type of NLP task: classification/information
extraction/specialized
– Borderline situations
• Report degree of complexity needed to correct
errors
• Determine if performance is adequate for task
• Report on confidence intervals
IEBI Workshop-10/23/07
Other Issues:
Clinical Environment
• Heterogeneity
– Systems
– Document formats
– Document types
– Clinical Domain
• Working with physicians
• Clinical evaluation tradition
• Workflow issues
IEBI Workshop-10/23/07
Patient Documents
• Lack of access to patient records
– Significant bottleneck for NLP progress
• Difficult to get permission to share from health
care institutions
• Large scale effort needed to establish
scrubbed document sets for development and
evaluation
• Individual efforts beneficial but limited and
scattered
IEBI Workshop-10/23/07
Outline
• NLP Evaluation Issues
• Ideal evaluation of NLP output requires
consideration of the context of the
applications
• Catalog of common NLP applications in
biomedicine and the implication for evaluation
IEBI Workshop-10/23/07
Context-based Evaluation:
Example Record
• Chief Complaint: Asthma re-evaluation.
• Subjective: 8 year-old girl with past history of moderate
persistent asthma while living in Alaska until 2 years
ago
• The primary triggers for her asthma have been viral
colds and irritant exposure, and she had particular
difficulty with the forest fire smoke in central Alaska.
• She also has a history of a low serum IgA. Her last IgA
determination was August 2004, which showed an IgA
level of 29 mg/dl, with the lower limit of normal for a
child her age being 33.
IEBI Workshop-10/23/07
Context-based Evaluation
• Chief Complaint: Asthma re-evaluation.
• Subjective: 8 year-old girl with past history of
moderate persistent asthma while living in Alaska
until 2 years ago
• Tasks: Disease Maintenance Summarization
•
vs. Infectious Disease Reporting
IEBI Workshop-10/23/07
Context-based Evaluation
• Chief Complaint: Asthma re-evaluation.
• Subjective: 8 year-old girl with past history of moderate
persistent asthma while living in Alaska until 2 years
ago
• Tasks: Disease Maintenance Summarization
•
vs. Infectious Disease Reporting
IEBI Workshop-10/23/07
Context-based Evaluation
• Chief Complaint: Asthma re-evaluation.
• …
• The primary triggers for her asthma have been viral
colds and irritant exposure, and she had particular
difficulty with the forest fire smoke in central Alaska.
• …
• Tasks: Disease Maintenance Summarization
•
vs. Infectious Disease Reporting
IEBI Workshop-10/23/07
Context-based Evaluation
• Chief Complaint: Asthma re-evaluation.
• …
• The primary triggers for her asthma have been viral
colds and irritant exposure, and she had particular
difficulty with the forest fire smoke in central Alaska.
• …
• Tasks: Disease Maintenance Summarization
•
vs. Infectious Disease Reporting
IEBI Workshop-10/23/07
Context-based Evaluation
• Chief Complaint: Asthma re-evaluation.
• …
• She also has a history of a low serum IgA. Her last IgA
determination was August 2004, which showed an IgA
level of 29 mg/dl, with the lower limit of normal for a
child her age being 33.
• Task: Disease Maintenance Summarization
•
vs. Infectious Disease Reporting
IEBI Workshop-10/23/07
Context-based Evaluation
• Chief Complaint: Asthma re-evaluation.
• …
• She also has a history of a low serum IgA. Her last IgA
determination was August 2004, which showed an IgA
level of 29 mg/dl, with the lower limit of normal for a
child her age being 33.
• Task: Disease Maintenance Summarization
•
vs. Infectious Disease Reporting
IEBI Workshop-10/23/07
Outline
• NLP evaluation issues
• Ideal evaluation of NLP output requires
consideration of the context of the
applications
• Catalog of common NLP applications in
biomedicine and the implication for
evaluation
IEBI Workshop-10/23/07
Potential NLP Applications
•
•
•
•
•
•
•
Health reporting requirements
Known disease surveillance
Unknown disease surveillance
Recognizing adverse drug reaction
Quality assurance/avoiding clinical errors
Charge capture
Recognizing scientific relations in text databases
IEBI Workshop-10/23/07
Health Reporting Requirements
• Example: Reporting new TB cases
• Task description: Governmental
requirements that certain disease states must
be identified within a period after the original
information (typically diagnosis) is identified.
• Task requirements: Text may be confined to
one or more sections of record. May require
inference to identify disease state. May be
easier to get the “right” answer than other
apps.
IEBI Workshop-10/23/07
Known Disease Surveillance
• Example: Locating Hospital Acquired
(nosocomial) infections
• Task description: Looking at a set of fixed
reports for specific findings or combination of
findings that suggest disease state
• Task requirements: Need to combine free text
with structured text such as lab reports, and
existing codes (e.g., ICD-9 coding on
discharge)
IEBI Workshop-10/23/07
“Unknown” Disease
Surveillance
• Example: Looking for the next “gulf war
syndrome.”
• Task description: By far, the most difficult task
because it is not clear what is being searched
for. Looking for a pattern of signs, symptoms,
lab tests, time course, etc, not explained by
known patterns
• Task requirements: Every concept is
potentially relevant plus need significant
inference to determine novelty of problem.
IEBI Workshop-10/23/07
Recognizing Adverse Drug
Reactions
• Example: Searching for known (and possibly
unknown) side effects of treatments
• Task description: Side effect profiles are
known for many drugs/regimens. Early
recognition of onset of those side effects
important to decreasing morbidity
• Task requirements: Temporal relationship
between treatment and possible side effects
important to glean from narrative.
IEBI Workshop-10/23/07
Quality Assurance/Avoiding
Clinical Errors
• Example: Flagging contra-indicated
treatments due to a drug allergy
• Task description: Extract from narrative
signs/symptoms/lab tests that suggest
unanticipated response to prior treatment.
• Task requirements: combining concepts from
narrative with structured parts of records and
comparing to guidelines/protocols
IEBI Workshop-10/23/07
Charge Capture
• Example: Locating clinic/hospital charges
that have not been otherwise captured
• Task description: Scan narrative for
suggestion of procedures performed or
supplies used that have not been billed
• Task requirements: Inferring actions from
narrative and comparing with billing codes.
Concepts are well defined and can be
enumerated.
IEBI Workshop-10/23/07
Recognizing scientific relations
in text databases
• Example: Finding protein-protein interactions
in pubmed database
• Task description: Scan abstracts to identify
protein names and description of
relationships
• Task requirements: Requires understanding
of naming schemes in biology and ability to
handle naming issues. Inference to identify
correctly the relationship described in the text
IEBI Workshop-10/23/07
Summary
• Overview of evaluation issues
• Key point: evaluation requires
consideration of the context of the
applications
• Catalog of common NLP applications in
biomedicine and the implication for
evaluation
IEBI Workshop-10/23/07