Bayesian Network Based Triage Successfully Predict
Sarmad Sadeghi, MD, Curtis Kennedy, MD, Craig W. Johnson, Ph. D.
University of Texas Health Science Center at Houston
A recently developed Bayesian network
based triage system was evaluated against an
emergency room (ER) physician’s triage assessment
to determine its classification performance. Binary
sion analysis revealed that the Bayesian
network’s classification significantly predicted actual
emergency room disposition, whereas the ER
physician’s classification did not. These findings
support expansion of the pilot project to obtain more
Bayesian networks used for clinical
prediction are becoming more common
. We are
assessing the utility of such a tool to predict ER
disposition in a population of patients who
an emergency department with a chief complai
of abdominal pain.
Our data set is from a retrospective review
of the historical elements (excluding physical
95 consecutive ER patient
records where the chief complaint was abdominal
pain. 5 records were excluded from
because of missing disposition.
binary logistic regression on the data to determine the
accuracy of the Bayesian network (BN) and the ER
physician (MD). We included “go to ER” (serious)
as our predictor for both BN and physicia
n:. We used
the actual ER disposition as the dependent measure:
admit versus discharge or referral. There were
insufficient cases in the physician “call 911” category
for use in any analysis.
42 of the 90 patients were admitted.
BN classification was significantly correlated with
admission (p = 0.014) while physician classification
was not (although it approached significance, at p=
0.053). Table 1 summarizes binary logistic regression
models, with their associated significanc
e and odds
ratios. Tables 2 and 3 are classification accuracy
tables for BN and physician, respectively, and
contain sensitivity, specificity, and positive (PPV)
and negative (NPV) predictive values.
Table 1: Binary logistic regression results:
(OR) and 95%
, coefficient (B)
Table 2: Bayesian network classification table.
81%, Specificity = 44%, PPV = 56%,
NPV = 72%.
Table 3: Physician classification table. Sensitivity =
64%, Specificity = 56%, PPV = 56%, NPV = 64%.
In general, the Bayesian network is
more conservative than the ER physician (it assigns a
higher level of disposition). Given the potential
application of serving as an initial triage system for
people with medical concerns, we feel the negative
n specificity is acceptable since it is safer to
err on the side of caution.
The utility of this application is likely to be
triaging potential patients that are still at home. Since
the subjects are from a population that has already
presented to th
e ER, these results cannot yet be
generalized to the population at large.
Based on this preliminary analysis, we feel further
research is justified and necessary to better describe
the scope of potential benefits that the Bayesian
network triage syste
m may provide.
1. P.J.F. Lucas, H. Boot and B.G. Taal.
theoretic network approach to treatment
management and prognosis.
11 (1998) 321