Taming the Emergency Department Chief Complaint:
Transforming Unstructured Data Elements for Secondary Uses

Stephanie W. Haas

School of Information and Library Science

UNC-CH

The Challenge

How to use text created for one purpose for new applications

reuse, repurposing, recycling, refactoring
different context of use
different audience with different knowledge, expectations
different requirements (e.g., precision, standardization, expressiveness)

What if the original text wasn't that easy to use in the first place?

Cast of Characters

Debbie Travers and Anna Waller, Dept of Emergency Medicine, UNC-CH
Emergency Medicine Text Processor (EMT-P)
North Carolina Emergency Department Database (NCEDD) and staff
North Carolina Disease Event Tracking and Epidemiologic Collection Tool (NC Detect) and staff

Emergency Department Patient Records

Primary uses

gather and record initial information (symptoms, situation, registration)
guide initial patient care
record additional findings, treatment, diagnoses

Secondary uses

clinical and other research
quality control
administration
syndromic surveillance (early warning)

Characteristics of Records

Data Elements for Emergency Department Systems (DEEDS), (recommended in 1997)
Some elements are standardized

final diagnosis uses ICD-9 codes
temperature usually in oC

Others are not

chief complaint (CC), triage notes, physician and nurse notes

Survey of EDs
NC & Seattle/King County WA

Travers, Waller, Haas, Lober & Beard, (2003)
Is ED data ready for use in bioterrorism surveillance systems?

availability, timeliness, sources, standards
NC 1999, n=96
NC 2003, n=80
Seattle/King County 2003, n=17

Available in Electronic Form?

Other results

Availability of diagnosis: 24 hrs 1 month
Only 2 respondents knew more than a little or nothing about DEEDS
Wide variety of information systems in use

local specialization is common
different systems in hospital and ED
rapid change in IT

Not easy to aggregate or merge data from different EDs

Chief Complaint (CC)

Recorded at triage (2-5 minute encounter)
Reason patient came to the ED
In NC 2003 survey

51% electronic, more since then
1-4 fields, 12 characters "paragraphs"
free text, locally- and/or vendor-developed pick list, ICD-9-CM
patient's exact words, paraphrase, codes

Characteristics of CC Pick Lists

Travers, Haas, Waller, Reeder, & Spicer, (2005)
Vary in size, specificity, restrictions on use

Size: 68 terms 1,680 terms
Specificity: pain vs. abd pain or neck pain
Restrictions on use:
no additions, triage notes used for additional detail
free text additions or substitutions for CC

Unknown sources of lists

Characteristics of CC Free Text
(Travers, Travers & Haas)

Telegraphic, semantically dense

cough/diarrhea/congestion
cough/abdominal pain/inhaled toxic fumes
abd pain- loss of appetite
gun shot wound knee

Multiple ways of expressing a concept

chst pn, cp, chert pn, chest/arm pain, chest pain

Ambiguous expressions

Variable punctuation, spacing, capitalization

h/a/rt shoulder & left hip pain x2d
tvfellonhim
stomachulcer

Coordinate structures and other use of slash

hip/thigh pain
tingling feet/hands
nausea neck/arm pain/see spots

Abbreviations, initialisms, truncations, misspellings

ha, h/a, headache
ETOH withdrawal
pian
should, shoulde, shldr, shoulder
n/v/d

Indecipherable

tacky
RBLD
5320407
other
feels weird!
fuzzy shin

Goals

Standardization (of entry terms or post-entry mapping routines)
Translate from immediate clinical use for individual patient to aggregated data for secondary uses
Extract CC concepts

clean, normalize and map to UMLS

Interpret modifiers, qualifiers, negation, temporal relationships, other relationships (e.g., co-morbidity)

EMT-P (Travers and others)

Clean, normalize, and map to UMLS concepts
Identify concepts needed to express CC
Prepare CCs for aggregation
Improve data quality of CCs for secondary uses
Identify ED concepts missing from or inappropriately defined in UMLS

jail clearance, MVC

Round 1

expand acronyms, abbreviations, truncations, correct common misspellings

Round 2

handle a variety of punctuation, expand coordinate structures, segment into terms

Round 3

delete unnecessary modifiers, qualifiers, temporal clauses

Provide more accurate picture of reasons people come to ED

Validity

Not just matching a UMLS concept, but matching correctly
Sample of matches validated by clinicians

equivalent: a-fib, atrial fibrillation
related: cramps, cramps(muscle)
no match: arrest, law enforcement arrest

No agreement: vag bleed, vag bleeding, bleeding of vagina
Influence of context in interpretation (i.e., w/o patient) concern for secondary uses?

Next The Triage Note

Additional information to supplement CC
Shares many characteristics of CC

PT AMB TO TRIAGE AT 1440 C/ORASH ALL OVER BODY ONSET THURSDAY AM, RED RAISED SPOTS TO NECK, PT REPORTS ITCHI G NO PAIN, ALSO HA X 1 WEEK DENIES INSECT BITES OTHER C/O OR FEVER

Length varies from phrases to "sentences"
Form varies from rough notes to auto-complete templates producing well-formed text
More negation, temporal relationships, other discourse-level relationships
Essentially no limit on topic, therefore structure or vocabulary

contrast to lab reports (Friedman)

Immediate goals

Extract concepts as with CC
Correct interpretation of negation

denies fever
no sob, fvr
Chapman et al., 2001

Extract temporal relationships
Extract additional relationships

Concepts of Time

past, present, future
span, point, duration, frequency
relative time (yesterday), absolute time (2/28/06, 3:30 p.m.)
vagueness and uncertainty
events

sequential, simultaneous, overlapping
cause, result, unrelated

report (when someone told someone else)

Now

Thursday AM

1 Week Ago

headache

rash onset

triage

Syndromic Surveillance

Close to real-time identification of potential outbreak or occurrence of health problems of interest

gastrointestinal severe, fever/rash illness
Follow-up by epidemiologists, clinicians, public health officials

Limited by

availability and readiness of data
sensitivity and specificity of classification of data into syndrome definitions

North Carolina Emergency Department Database (NCEDD)

Collects ED data daily from 63% of NC EDs
Collects up to 22 data elements, including CC and triage note
Aggregated data (graphs and reports) available for use by hospitals, researchers, public health officials
Uses EMT-P to pre-process CC

North Carolina Disease Event Tracking and Epidemiologic Collection Tool
(NC DETECT)

Collects data daily from NCEDD, Carolinas Poison Center, NC Pre-hospital Medical Information System (PreMIS), NC State College of Veterinary Medicine Laboratories.
Coming soon: Piedmont Wildlife Center

Classifies data into syndromes
Creates daily reports for public health officials and other approved users
Syndromic definitions:

Map CDC definitions into DB queries
ED CC pre-processed with EMT-P
Other data in various forms DB queries can be exceedingly messy!
Sensitivity and specificity are of concern

Merging Data from Multiple Sources

CC and triage note (collected at same time and place, serve same purpose)

recognize agreement, paraphrase
recognize and fill gaps for complete description of symptoms, situation
recognize disagreement -- which element should be believed?

Then add different types of data elements from various sources

Remaining Issues

Controlled vocabulary for CC?

CC symposium

Extracting additional information

time, situation (e.g., circumstances of injury), co-morbidity, history, other relationships

Merging information from disparate sources
Reconciling primary and secondary stakeholders' needs

Stephanie W. Haas

stephani@ils.unc.edu

EMT-P

http://www.med.unc.edu/wrkunits/2depts/emergmed/EMTP/index.html

NCEDD

http://www.ncedd.org/index.html

NC DETECT

http://www.ncdetect.org/