Automated DNA sequence-based early warning system for the detection of methicillin-resistant Staphylococcus aureus outbreaks.

Automated DNA Sequence-Based Early Warning
System for the Detection of Methicillin-
Resistant Staphylococcus aureus Outbreaks
Alexander Mellmann1, Alexander W. Friedrich1, Nicole Rosenko¨tter2, Jo¨rg Rothga¨nger3, Helge Karch1, Ralf Reintjes2,
Dag Harmsen4*
1 Institute for Hygiene, University Hospital Mu¨nster, Mu¨nster, Germany, 2 Hamburg University of Applied Sciences, Hamburg, Germany, 3 Ridom GmbH, Wu¨rzburg, Germany,
4 Department of Periodontology, University Hospital Mu¨nster, Mu¨nster, Germany
Competing Interests: JR, HK, and
DH have declared a potential conflict
of interest. JR and DH are the
developers of the Ridom StaphType
software mentioned in the
manuscript. The software is
distributed and sold by the company
Ridom GmbH that is partially owned
by them. HK is a scientific advisor of
this company. All other authors have
declared that no competing
interests exist.
Author Contributions: AM, AWF, HK,
and DH designed the study. AM, AWF,
NR, JR, RR, and DH analyzed the data.
AM, AWF, NR, HK, RR, and DH
contributed to writing the paper. HK
supervised the clinical sample and
data collection. AWF headed the ICP
team. AM did the laboratory
identification and strain analysis. JR
and DH developed the Ridom
StaphType software. AMandDHwrote
the main part of the paper. All authors
gave useful comment on the analysis
of data and text of the manuscript.
Academic Editor: Mervyn Singer,
University College London, United
Kingdom
Citation: Mellmann A, Friedrich AW,
Rosenko¨tter N, Rothga¨nger J, Karch
H, et al. (2006) Automated DNA
sequence-based early warning
system for the detection of
methicillin-resistant Staphylococcus
aureus outbreaks. PLoS Med (3)3: e33.
Received: May 12, 2005
Accepted: October 28, 2005
Published: January 10, 2006
DOI:
10.1371/journal.pmed.0030033
Copyright: � 2006 Mellmann et al.
This is an open-access article
distributed under the terms of the
Creative Commons Attribution
License, which permits unrestricted
use, distribution, and reproduction in
any medium, provided the original
author and source are credited.
Abbreviations: CI, confidence
interval; ICP, infection control
professional; MLST, multi-locus
sequence typing; MRSA, methicillin-
resistant Staphylococcus aureus; NPV,
negative predictive value; PFGE,
pulsed-field gel electrophoresis; PPV,
positive predictive value; spa, S.
aureus protein A gene; UHM,
University Hospital Mu¨nster
* To whom correspondence should
be addressed. E-mail: dharmsen@
uni-muenster.de
A B S T R A C T
Background
The detection of methicillin-resistant Staphylococcus aureus (MRSA) usually requires the
implementation of often rigorous infection-control measures. Prompt identification of an MRSA
epidemic is crucial for the control of an outbreak. In this study we evaluated various early
warning algorithms for the detection of an MRSA cluster.
Methods and Findings
Between 1998 and 2003, 557 non-replicate MRSA strains were collected from staff and
patients admitted to a German tertiary-care university hospital. The repeat region of the S.
aureus protein A (spa) gene in each of these strains was sequenced. Using epidemiological and
typing information for the period 1998–2002 as reference data, clusters in 2003 were
determined by temporal-scan test statistics. Various early warning algorithms (frequency,
clonal, and infection control professionals [ICP] alerts) were tested in a prospective analysis for
the year 2003. In addition, a newly implemented automated clonal alert system of the Ridom
StaphType software was evaluated.
A total of 549 of 557 MRSA were typeable using spa sequencing. When analyzed using scan
test statistics, 42 out of 175 MRSA in 2003 formed 13 significant clusters (p , 0.05). These
clusters were used as the ‘‘gold standard’’ to evaluate the various algorithms. Clonal alerts (spa
typing and epidemiological data) were 100% sensitive and 95.2% specific. Frequency
(epidemiological data only) and ICP alerts were 100% and 62.1% sensitive and 47.2% and
97.3% specific, respectively. The difference in specificity between clonal and ICP alerts was not
significant. Both methods exhibited a positive predictive value above 80%.
Conclusions
Rapid MRSA outbreak detection, based on epidemiological and spa typing data, is a suitable
alternative for classical approaches and can assist in the identification of potential sources of
infection.
PLoS Medicine | www.plosmedicine.org March 2006 | Volume 3 | Issue 3 | e330348
PLoSMEDICINE

Introduction
In the United States alone, infections acquired in hospitals
affect 2 million patients, account for half of all major hospital
complications, and result in annual costs of more than $4.5
billion [1]. Staphylococcus aureus is the leading cause of these
nosocomial infections that include a wide range of diseases
such as endocarditis, septicemia, skin infections, soft tissue
infections, and bone infections [2]. Strains resistant to
methicillin, in particular, have become a major concern in
the hospital environment because of the high mortality rate
and the stringent hygienic requirements needed for patients
who are harboring a methicillin-resistant S. aureus (MRSA)
[3,4]. Moreover, since the emergence of strains that are
insensitive or have reduced sensitivity to glycopeptides, there
is a real danger of infections spreading that have even greater
drug resistance [5].
Analysis of laboratory test results and patients’ charts are
the methods usually used to identify outbreaks. However, the
manual review of laboratory test results is time-consuming
and resource-intensive. Electronic analysis of data can help
identify suspicious patterns of disease and antimicrobial
resistance [6], but such sentinel methods are rarely used in
clinical practice. The typing of MRSA isolates, not only from
clinical specimens, but also from surveillance cultures, is
necessary for the elucidation of possible transmission routes.
Because the procedures are slow and laborious, molecular
typing (e.g., pulsed-field gel electrophoresis [PFGE]) is usually
used a posteriori to track the course of nosocomial infections
in an already established outbreak. Furthermore, PFGE
requires great efforts to harmonize protocols and is therefore
only partially successful in generating reproducible results
[7]. In order to improve the speed of typing, DNA sequence-
based approaches, such as the multi-locus sequence typing
(MLST), are becoming more frequently used [8]. However,
MLST is not suitable for routine surveillance of MRSA
because of the high costs involved and the low discriminatory
power compared to PFGE. Frenay et al., who were the first to
use a single-locus sequence typing method for S. aureus,
employed the sequence of the polymorphic region X of the S.
aureus protein A gene (spa) for typing [9]. Since then,
numerous studies evaluated this variable number of tandem
repeat targets as quite suitable for short-term epidemiolog-
ical applications, e.g., [10–13]. Because of the paucity in
software for repeat identification and lack of a consensus in
assigning spa type names, the wide-spread use of the method
was hampered for years until the recent introduction of the
Ridom StaphType software [14]. With this software, the spa
sequences are analyzed automatically and linked to a database
integrated with epidemiological information. A universal
nomenclature is achieved by synchronization with a central
server that assigns new spa types for all users (http://www.
spaserver.ridom.de).
The aim of the study reported here was therefore to
analyze the utility of a spa sequence-based, automatic early
warning algorithm to detect MRSA clusters in hospitals and
to compare this approach with classical surveillance
techniques. We hypothesized that the automated system,
once established, can complement and even replace the
labor-intensive traditional methods used for cluster identi-
fication.
Methods
Setting
Between 1998 and 2003, a total of 557 non-replicate MRSA
isolates were collected at the University Hospital Mu¨nster
(UHM), Germany, a 1,480-bed tertiary-care teaching facility.
In 2003, there were approximately 43,000 annual admissions
to the hospital where the mean length of stay was 9.8 d. The
prevalence of patients with MRSA colonizations and in-
fections was taken as the annual number of persons
harboring MRSA (3100) divided by the total number of
admissions at UHM [15]. The baseline for calculation of the
relative risk was the year 1998.
Surveillance and Infection Control Measures
All new MRSA cases were monitored prospectively by
infection control professionals (ICP) from the day when
MRSA was first identified until hospital discharge. Informa-
tion on each patient was obtained by reviewing medical
records and laboratory data and holding telephone interviews
with the attending physician. Subsequently, the ICP decided if
a transmission event was likely and if further investigation
was necessary. In more detail, the following infection control
measures were implemented: (i) As recommended in the
guidelines of the Robert Koch Institute (Berlin, Germany), all
patients infected or colonized with MRSA were placed in
contact isolation until the time of discharge or until
eradication could be documented in three consecutive sets
of negative surveillance cultures (separated by at least 24 h).
MRSA surveillance cultures included swabs of several body
sites (nose, groin, skin lesions, inguinal, perineal, and axillary
swabs). In the case of infected patients, samples were taken
from the site of infection. (ii) All patients known to have been
previously colonized or infected with MRSA were isolated on
re-admission to UHM and surveillance swabs were obtained.
Negative surveillance cultures were mandatory in order to
terminate contact isolation. (iii) Clinical microbiology labo-
ratory results were monitored daily for the occurrence of
specimens containing MRSA. (iv) spa typing of MRSA isolates,
as performed since 2002, were carried out directly after
detection of a new MRSA isolate. (v) Colonized patients were
treated with nasal mupirocin ointment for 5 d and daily
chlorhexidine body washes were applied. In the case of
patients remaining in the hospital after eradication, weekly
surveillance cultures were recommended over a 4-wk period,
and then at monthly intervals to detect possible re-coloniza-
tion. (vi) To detect MRSA colonization and cross-trans-
mission, surveillance cultures were obtained from
roommates as soon as a new MRSA patient was identified.
(vii) Staff were screened when nosocomial transmissions were
suspected and at intervals as a surveillance method on high-
risk wards. (viii) Hospital staff found MRSA-positive were
suspended from work on the wards until the successful
eradication of MRSA could be documented. (ix) Systematic
surveillance cultures at the time of admission and on a weekly
basis thereafter were begun in 2002 in wards caring for high-
risk patients, e.g., intensive care units [15,16]. Colonization
and infection were defined in accordance with the Centers
for Disease Control and Prevention criteria [17].
Microbiology and Molecular Typing
The strain collection consisted of MRSA from various
clinical sources (e.g., blood cultures and wound infections)
PLoS Medicine | www.plosmedicine.org March 2006 | Volume 3 | Issue 3 | e330349
Automated Detection of MRSA Outbreaks

and included surveillance cultures from patients and staff. Of
all clinical S. aureus isolates, 6.4% exhibited methicillin
resistance in 2003. For species identification, every strain
was tested with API ID 32 Staph (bioMe´rieux, Marci l’Etoile,
France) and for the presence of free coagulase. The presence
of the mecA gene responsible for methicillin resistance was
confirmed using PCR [18]. The sequence of the short
sequence repeat region of the spa gene encoding the S. aureus
protein A was determined in 557 strains [14]. The primers
spa-1113f (59- TAA AGA CGA TCC TTC GGT GAG C �39)
and spa-1514r (59- CAG CAG TAG TGC CGT TTG CT �39)
were used for spa amplification and Taq Cycle sequencing.
DNA sequences were obtained with an ABI Prism 3100 Avant
Genetic Analyzer (Applied Biosystems, Foster City, Califor-
nia, United States) and analyzed with the Ridom StaphType
software version 1.5 beta (Ridom GmbH, Wu¨rzburg, Ger-
many) incorporating the newly added automated early
warning system (‘‘clonal alerts’’) for MRSA cluster detection
[14]. Typability, discriminatory index, and the 95% con-
fidence interval (CI) of the discriminatory index were
calculated using the procedures published previously [19,20].
Retrospective Temporal-Scan Test Statistics
To evaluate the various early warning algorithms, we
performed scan test statistics using the epidemiological and
typing information from 1998 to 2002 as historical data to
determine MRSA clusters in 2003 [21,22]. Temporal-scan
statistics evaluates whether an apparent cluster of disease is
unlikely to occur by chance alone. Thereby, the test
determines a likelihood p-value for an observed number of
cases appearing in a window of fixed width as the window is
moved along the time axis studied (2003). Observed and
expected cases, the latter calculated using the historical data
(1998–2002), were compared with a null hypothesis that states
cases occur at random, evaluated against the alternative
hypothesis that states cases cluster in certain time periods. In
this evaluation, a Poisson distribution was assumed because a
positive MRSA finding is a rare and irregular event. Clusters of
two or more infected/colonized patients or colonized staff on
the same ward or wards in close contact (e.g., interdisciplinary
intensive care units) occurring within a 2-wk window and
harboring the same MRSA isolate according to the spa typing,
were identified as significant at the 5% level. These statistically
confirmed clusters were then used as the ‘‘gold standard’’ for
comparing the various alert mechanisms. Non-significant
clusters were considered to be sporadic occurrences.
Early Warning Algorithms
Every MRSA isolate obtained in 2003 was examined in a
prospective analysis by applying descriptive epidemiologic
parameters such as time, place, and person. When two or
more MRSA isolates were detected within a 2-wk window on
the same ward or on wards having close contact, the resulting
alert was regarded as a ‘‘frequency alert’’ and allocated to a
‘‘frequency cluster.’’ If MRSA isolates also shared an identical
spa type, the allocation to ‘‘clonal alerts’’ and associated
‘‘clonal clusters’’ was triggered. An ICP, which is a panel
consisting of two physicians and four infection control nurses
who meet weekly and hold additional meetings when an
outbreak occurs, rate the findings as ‘‘ICP alerts’’ and ‘‘ICP
clusters,’’ respectively. When feasible, the area of surveillance
is widened and an investigation initiated. The ICP uses
microbial data and data from patients’ charts to reach their
decisions but are blind to the occurrence of an outbreak on
the basis of spa typing results.
Statistical Analysis
Sensitivity, specificity, positive and negative predictive
values (PPV, NPV), and pre-test probability were determined
as described by Sackett et al. [23]. The pre-test probability is
defined as the proportion with the target disorder (MRSA
cluster) in the population at risk (MRSA positive) at a specific
time interval. Two-tailed, 95% CIs were calculated to assess
sensitivity, specificity, PPV, and NPV using a normal
approximation for the pertinent (binomial) distribution.
The chi-square distribution, with one degree of freedom,
was used to determine the significance of the differences in
these parameters.
Results
Table 1 summarizes the important epidemiological indica-
tors for MRSA at the UHM. The overall prevalence of MRSA
cases was 0.17 per 100 admissions and the relative risk of
acquiring MRSA increased 4-fold during the study period.
The annual number of patients with MRSA bacteremia
reached a peak in 2003 with six patients. The average turn-
around time for spa typing under routine laboratory
conditions was 2.4 d. Of the 557 MRSA isolates tested, 549
(98.6%) could be typed using spa sequencing, and the eight
strains, which could not be typed, were excluded from the
analysis. A total of 79 different spa types were identified in
samples collected for the period 1998–2003. The discrim-
Table 1. Important Epidemiological Indicators for MRSA at the UHM (1998–2003)
Year Number of
MRSA
Prevalence Relative Risk of
MRSA Acquisition
(95% CI)
Number of
Patients with
MRSA Bacteremia
Number of
spa Types
DI (95% CI)/
% Typability
1998 51 0.11 reference 1 16 0.879 (0.829–0.929)/100
1999 82 0.17 1.6 (1.1–2.3) 0 19 0.895 (0.861–0.928)/96.3
2000 56 0.12 1.1 (0.8–1.7) 2 22 0.873 (0.804–0.942)/100
2001 55 0.13 1.1 (0.8–1.7) 1 28 0.950 (0.926–0.973)/98.2
2002 134 0.31 2.8 (2.0–3.9) 1 31 0.897 (0.870–0.925)/100
2003 179 0.42 3.8 (2.8–5.3) 6 35 0.859 (0.824–0.894)/97.8
1998–2003 557 0.17 - 11 79 0.918 (0.908–0.929)/98.6
DI, discriminatory index.
DOI: 10.1371/journal.pmed.0030033.t001
PLoS Medicine | www.plosmedicine.org March 2006 | Volume 3 | Issue 3 | e330350
Automated Detection of MRSA Outbreaks

inatory power of spa typing was 91.8%. Table 2 shows the
frequency with which the various spa types were isolated at
UHM and nationally. spa types t003, t004, t001, and t032 were
the types most frequently isolated during the study period
and accounted for 52.9% of all cases. In Table 2 the typing
results are also brought into a global epidemiological context
as defined by PFGE and MLST [11,24–26]. The dynamics,
expressed on an annual basis of the epidemic MRSA clones at
UHM and in Germany as a whole, are depicted in Figure 1. In
general, the findings for UHM followed the national trend,
i.e., the number of ‘‘Barnim’’ and ‘‘Rhine Hesse’’ MRSA clones
increased in parallel throughout the study period. As shown
by data for the ‘‘Southern German’’ MRSA clone, the
fluctuation at the regional level was more marked, and this
is probably due to the smaller number of cases.
In 2003, a total of 175 MRSA isolates (154 patients, 21 staff)
could be typed, and these comprised 34 different spa types.
The results of the year 2003 scan test analysis are shown in
Table 3. The encoded name of the clinic/ward was derived
from the location of the first MRSA isolated in a particular
cluster. In total, there were 42 MRSA isolates forming 13
significant clusters and representing seven different spa types,
but five clusters involved spa type t003, a common spa type in
our hospital (29.6% of all MRSA in 2003). The average time-
span of the clusters was 10 d (range 1–31 d) and the number
of isolates in each cluster ranged from 2–11 (mean 3.2). Six
clusters were located on a single ward, whereas seven other
clusters were located at sites distributed throughout a clinic.
In the prospective analysis of all MRSA in 2003 using the
various alert procedures, there were 106 frequency alerts
assignable to 31 frequency clusters (Table 4). A total of 36
clonal alerts, comprising 20 clonal clusters, were triggered by
the early warning system. The ICP called 22 ICP alerts
corresponding to nine ICP clusters, but in only five clusters
(the two largest clusters, numbers five and seven were
included) was the recognition of an existing outbreak and
the need for further investigation correct. The four other
clusters arose from false alerts by the ICP. In Table 4, the alerts
triggered by the various methods are categorized as true or
false alerts using the alerts for the 13 significant ‘‘true’’
clusters. The sensitivity, specificity, PPV, and NPV, and where
appropriate, the 95% CIs for the various alert methods, are
displayed in Table 5. Because of the high number of false-
positive frequency alerts (n ¼ 77), the specificity of the
frequency and clonal methods (47.2% and 95.2%, respec-
tively) differed considerably. The ICP alerts had the highest
specificity, but the number of false-negative alerts (eight of 13
confirmed clusters were missed) led to the lowest sensitivity
(62.1%). Given a pre-test probability of 24%, the PPV (same as
the post-test probability) of the ICP and clonal alerts was
above 80%, whereas the PPV of the frequency alert was only
27.4%. There were no significant differences in specificity and
PPV between clonal and ICP alerts. Frequency alerts were
Table 2. Most Frequent spa Types and Epidemiological Background (PFGE, MLST) at the UHM and in Germany
spa
Type
Cases at
the UHM,
2003 (%)
Cases at
the UHM,
1998–2003 (%)
% Clones
in Germany,
2003/1998–2003
Concordant
MLST STa
Commentb
t003 53 (29.6) 87 (15.6) 24.5/10.1
(together with t002)
ST-5, ST-225
Rhine Hesse MRSA (subclone), EMRSA-3, New York clone, CC5
t032 30 (16.7) 58 (10.4) 29.0/20.1 ST-22 Barnim MRSA (prototype and subclone), EMRSA-15, prototype of ST-22, CC22
t001 19 (10.6) 68 (12.2) 13.8/26.6
(together with t023
and t041)
ST-5, ST-222, ST-228
Southern German MRSA (prototype and subclone), Rhine Hesse MRSA (subclone),
EMRSA-3, New York clone, CC5
t004 14 (7.8) 82 (14.7) 13.8/22.5
([together with t026
and t038)
ST-45
Berlin MRSA (prototype), USA600 ORSA II, USA600 ORSA IV, CC45
t008 10 (5.6) 29 (5.2)
0.1/2.9
(together with t051)
ST-8, ST-247,
ST-250, ST-254
Northern German MRSA (subclone), USA300 ORSA IV (cMRSA in the US), Archaic/
Iberian, ST250 ORSA I, CC8
t038 5 (2.8) 10 (1.8) (see t004) ST45 Berlin MRSA (subclone NRW), CC45
t002 4 (2.2) 13 (2.3) (see t003) ST-5, ST-231 Rhine Hesse MRSA (prototype), EMRSA-3, New York clone, Japan clone, Pediatric,
USA100 ORSA II, USA800 ORSA IV, ST 5 ORSA I, CC5
NT 4 (2.2) 8 (1.4) - - NT
t044 2 (1.1) 8 (1.4) Data not available ST-80 cMRSA (lukS–lukFþ) widely disseminated in Europe
t009 1 (0.6) 22 (3.9) 0.5/4.6 ST-254 Hannover MRSA, EMRSA-10, CC8
t018 1 (0.6) 8 (1.4) Data not available ST-30, ST-36, ST-38 Prototype of ST-36, EMRSA-16, USA200 ORSA II, CC30
t037 0 (0.0) 19 (3.4) 0.0/0.3 ST-239, ST-240, ST-241 Vienna MRSA, Brazilian/Hungarian, ST239 ORSA III, ST240 ORSA III, EMRSA-1, �4,
�7, �9, �11, CC8/239
t051 0 (0.0) 1 (0.2) (see t008) ST-247 Northern German MRSA (prototype and subclone), Archaic/Iberian, ST247 ORSA I,
EMRSA-5, CC8
t041 0 (0.0) 3 (0.5) (see t001) ST-111, ST-228 Southern German MRSA, CC5
t026 0 (0.0) 1 (0.2) (see t004) ST45, ST-47 Berlin MRSA (subclone), CC45
t023 0 (0.0) 0 (see t001) ST-228 Southern German MRSA, CC5
t190 0 (0.0) 0 Data not available ST-8 EMRSA-2, �6, CC8
Other 36 (20.2) 140 (25.1) 18.3/12.9
aPredominant ST in bold.
bGerman clone as defined by PFGE and MLST CC.
ST, MLST sequence type; CC, clonal complex; EMRSA, epidemic MRSA clone; ORSA, oxacillin resistant S. aureus; NT, non-typeable; cMRSA, community acquired MRSA; lukS–lukFþ, two-component leukocidin positive, encodes Panton-Valentine
leukocidin.
DOI: 10.1371/journal.pmed.0030033.t002
PLoS Medicine | www.plosmedicine.org March 2006 | Volume 3 | Issue 3 | e330351
Automated Detection of MRSA Outbreaks

significantly less sensitive (p , 0.001) and less accurate in
making positive predictions than clonal and ICP alerts.
Discussion
In this paper we have presented a new method for
prospective MRSA outbreak surveillance in a hospital that
uses case and molecular typing data. Historically, MRSA
outbreak detection in hospitals has relied on the watchful
eyes of physicians and other health-care workers. However,
the increasing availability of timely electronic surveillance
and molecular typing data raises the possibility of earlier
outbreak detection and intervention if suitable analytic
methods are found.
Germany belongs to a group ofWestern European countries
with an intermediate level of MRSA (approximately 20% of all
S. aureus diagnosed in laboratories are MRSA positive).
However, the isolation rate has increased significantly in
recent years [27]. Although the MRSA laboratory isolation rate
in UHM of 6.4% in 2003 is still rather low in comparison with
other German hospitals, the relative risk of acquiring MRSA
within this hospital facility rose significantly during the study
period (Table 1). Furthermore, the absolute risk will also
probably rise because of epidemiological pressure and the
rising prevalence of MRSA in Germany as a whole (Table 2 and
Figure 1). It is clear that control of MRSA is a pressing concern
where new concepts are needed, and therefore we studied spa
typing in combination with an automatic early warning
algorithm to detect MRSA clusters at UHM.
We showed that the feasibility and speed with which it was
possible to carry out spa typing was highly satisfactory. The
discriminatory power, however, was lower than previously
reported, probably because only a local strain collection was
analyzed [12,13]. Although not examined by us, the high intra-
and inter-laboratory reproducibility of 100% and the robust-
ness of the method have recently been documented (Aires-de-
Sousa et al., unpublished data). Moreover, there is a high
concordance of results between spa and PFGE, microarray
and MLST [11,12]. The practicability of using spa in short-
term epidemiological studies has been questioned because
differences in PCR amplicon sizes in related strains was
thought to imply instability in the target gene [28]. In the
meantime, however, there has been a plethora of publications
demonstrating the value of spa in the investigation of MRSA
outbreaks, e.g., [10,14]. Moreover, it has recently been shown
that spa data not only contain information on short-term, but
also long-term evolutionary events, as observed in whole
repeat duplications and deletions [12,29]. Because of the
steady fall in the cost of DNA sequencing and an average
hands-on time of only 20 min per sample (determination of
both strands of DNA and processing ten samples in parallel),
this technique is within the capability of even small
laboratories [30].
The present study has compared three early warning
algorithms for the detection of nosocomial MRSA outbreaks
before limited clusters of preventable MRSA transmissions
develop into larger outbreaks. The evaluation of an early
warning system, however, is difficult because there is no
accepted ‘‘gold standard’’ and it is likely that no system will be
completely reliable [31]. Therefore, we chose to combine
epidemiological and molecular typing data with statistical
analysis to provide an objective measure of performance
between the varied approaches. In this approach, by
definition, the sensitivity and NPV of the frequency and
clonal alert methods will always be 100%. This means that the
ICP method will give results that are less sensitive, or at best,
of only equal sensitivity to those of the automated methods.
Infection and colonization with MRSA were given equal
status since both can lead to further transmissions. The
typing data accumulated since 1998 enables the significance
of spa-type clustering with respect to time to be calculated for
all those occasions when there is a suspicion of an outbreak.
By excluding all non-significant clusters, it was possible to
reduce the likelihood that two or more MRSA with the same
spa type, coincidentally isolated on the same or related wards
within the 2-wk window, would be counted as correct. A 2-wk
window is approximately 1.5 times as long as the mean
duration of hospitalization in the UHM hospital. A 4-wk time
window yielded similar results (unpublished data).
An outbreak can be defined as (i) two or more cases of
infection by a common agent that are linked epidemiolog-
ically. However, this definition has usually limited practical
relevance in the identification of outbreaks, because it
presupposes that detailed epidemiological and typing data
are available as soon as the outbreak occurs. Thus, in more
practical terms, an outbreak is often defined operationally as
(ii) an increase in the number of cases above expected levels
[32]. Historical data can be used to calculate a baseline and an
alert is given when the number of cases exceeds a certain
threshold. Early warning systems at national levels are based
on this definition of an outbreak and have already imple-
Figure 1. Annual Dynamics of Epidemic MRSA Clones at the UHM as
Defined by spa Typing (A) and in Germany by PFGE (B) from 1998 to
2003.
DOI: 10.1371/journal.pmed.0030033.g001
PLoS Medicine | www.plosmedicine.org March 2006 | Volume 3 | Issue 3 | e330352
Automated Detection of MRSA Outbreaks