By Emil Venere
Wednesday, September 2, 2009
Researchers are developing computer models to comb through
thousands of injury reports in large administrative medical datasets or
insurance claims data to automatically classify them based on specific words or
phrases.
"One goal is to identify the most important causes of
injuries so that efforts could be directed toward reducing the burden of
injuries in society," said Mark Lehto, an associate professor in Purdue University's
School of Industrial Engineering.
The reports, usually filled out by employers, health-care
professionals or claimants themselves, are currently classified by manual
coders hired by users such as the National
Center for Health
Statistics, hospital staff or insurance industry handlers who review thousands
of "injury narratives" included in reports.
"This is obviously very labor-intensive," Lehto
said.
The Purdue engineer and researchers at the Liberty Mutual
Research Institute for Safety in Hopkinton,
Mass., assigned codes to injury
reports from workers' compensation claims using two different models developed
with a technique called "Bayesian methods."
"The predictions were quite good," Lehto said.
"The results were comparable to the human coders. The accuracy is
surprising considering all of the misspellings, run-on words, abbreviations and
inconsistent or missing punctuations seen in these workers' compensation claim
narratives."
An example of an injury-claim narrative included in the
paper is: "HUSB. & SON WERE REARENDED AT RED TRAFFIC LIGHT BY DRUNKEN
DRIVER DRIV-ING AT LEAST 45 MPH INFULL SIZE PICK-UP TRUCK//N."
"Can you imagine reading through 10,000 of these
narratives and trying to interpret what the cause of injury is and assign
different codes?" said Lehto, the 2008 Liberty Mutual Research Institute
for Safety visiting scholar.
Research findings were detailed in a paper published in
August in the journal Injury Prevention. The paper was written by Lehto and
Liberty Mutual research scientists Helen Marucci-Wellman and Helen Corns.
Insurance companies enter, maintain and manage tens of
thousands of claims annually. The study examined approaches for efficient
assignment of each claim using a computer approach with one and two-digit
"event code" categories developed by the U.S. Bureau of Labor
Statistics.
"So now we are trying to take these vast sets of data,
which have been limited in their utility due to the large expense in hiring
manual coders, and we are able to glean important information from the injury
narratives and come up with new knowledge on the potential causes and
prevention of injuries," Lehto said.
The new models might lead to programs that automatically
code reports as they are being filed.
"These models can be easily updated to deal with new
types of accidents they haven't encountered before," Lehto said.
The models calculated the probability that reports would be
classified by human coders in specific categories. One model, called
"naive," reviewed individual words, and the other, called
"fuzzy," looked at sequences of words and phrases in the narratives,
such as "fell off a ladder."
The researchers used a database of 14,000 claim cases, with
11,000 used to develop the models and 3,000 used to test the models.
"It's important to distinguish that we predicted 3,000
cases that were different than the ones used to develop the models," Lehto
said. "These were cases the models hadn't seen before, and the models
accurately predicted how these cases would be classified by human coders."
ABSTRACT
“Bayesian Methods: A Useful Tool for Classifying Injury
Narratives into Cause Groups”
To compare two Bayesian methods (Fuzzy and Naive) for
classifying injury narratives in large administrative databases into event
cause groups, a dataset of 14,000 narratives was randomly extracted from claims
filed with a workers' compensation insurance provider. Two expert coders
assigned one-digit and two-digit Bureau of Labor Statistics (BLS) Occupational
Injury and Illness Classification event codes to each narrative. The narratives
were separated into a training set of 11,000 cases and a prediction set of
3,000 cases. The training set was used to develop two Bayesian classifiers that
assigned BLS codes to narratives. Each model was then evaluated for the
prediction set. Both models performed well and tended to predict one-digit BLS
codes more accurately than two-digit codes. The overall sensitivity of the
Fuzzy method was, respectively, 78% and 64% for one-digit and two-digit codes,
specificity was 93% and 95%, and positive predictive value (PPV) was 78% and
65%. The Naive method showed similar accuracy: a sensitivity of 80% and 70%,
specificity of 96% and 97%, and PPV of 80% and 70%. For large administrative
databases, Bayesian methods show significant promise as a means of classifying
injury narratives into cause groups. Overall, Naive Bayes provided slightly
more accurate predictions than Fuzzy Bayes.
Original
article
SOURCE: Purdue
Univ.