Biomedical Engineering Reference
In-Depth Information
discharge diagnoses are typically not available
without substantial delay, usually limiting their
utility for a surveillance system designed to mini-
mize reporting time. Additionally, while chief-
complaint description may be variable in its
accuracy, it remains the most critical data element.
While case studies exist for many diseases, no
one can know how a weaponized, possibly genet-
ically engineered disease may propagate through
a community and what primary symptoms it may
cause to the majority of its victims. Capturing
the main symptoms that brought a person to
the ED is very important for the purpose of
trying to link cases. However, studies have shown
that the disparity between chief complaint and
final diagnosis is relatively minor and does not
nullify the timeliness advantage of chief complaint
descriptions [15].
3.9 Data Analysis: Algorithms for
Aberration Detection
Aberration detection algorithms have long been
used to aid decision-making in the hospital surveil-
lance process. This process includes: (1) Case
definition, (2) Data conditioning (tabulation,
adjustment), (3) Analysis, (4) Reporting, (5) Inves-
tigation if required, and (6) Consequence manage-
ment if required.
A review article on infection control by Smyth
and Emmerson [19] discusses the practical aspects
of the implementation of this process. Detection
algorithms are more than statistical hypothesis
tests; the case definition and data conditioning
steps have key roles in determining their utility.
The role of these algorithms is to automate the
analysis in a uniform, replicable fashion with effec-
tive detection performance, measured by sensi-
tivity, specificity and timeliness. The input to an
anomaly detection algorithm is typically a time
series of counts, proportions, or other functions of
a monitored quantity. Examples of such quanti-
ties are risk-adjusted surgical site infection rates
or counts of diagnoses of ILI. The time series
are obtained by aggregating data at some regular
time interval. Older surveillance methods [20,21]
used longer aggregation time intervals of weeks
or months to reduce data noisiness and to get
more regular time series distributions. The threat
of bioterrorism is increasing the emphasis on rapid
alerting and shortening these intervals to days and
hours, gradually pushing toward real-time detec-
tion [22]. The case definition and data conditioning
steps thus determine the inputs, approach, and
adaptive tuning of the algorithmic methods.
3.8 Data Acquisition and Presentation
Within a single hospital enterprise, there are few
data acquisition considerations. All desired data
should be stored within institutional databases and
the task of creating a surveillance dataset should be
only a matter of data base administration. However,
when creating systems that consolidate data from
outside of a single institution it is necessary to
consider how information will be transferred from
the data owners to end-users. Automation of the
process is critical for the success of any syndromic
surveillance system as the goals for these systems
are rapid detection and access to critical infor-
mation during emergency situations. Automated,
electronic data acquisition and transmission may
be conducted using a variety of protocols such as
Health Level Seven (HL-7), Simple Mail Transfer
Protocol (SMTP) or File Transfer Protocol (FTP)
to name but a few. When considering the type of
acquisition and transmission routines to implement
it is important to simplify the process as much as
possible for the data provider. If providing data
is cumbersome to the organization then timeliness
and completeness of data reporting is very likely
to be compromised.
3.10 Control Chart Usage
Hospital surveillance algorithms have made
widespread use of industrial process control chart
methods [1]. Control charts provide a graphical,
understandable approach for automated anomaly
detection with a foundation in probability theory
[23]. A process is deemed
in control
if the corre-
sponding monitored quantity varies within confi-
dence limits expected from an underlying statistical
