You are on page 1of 12

IMA Journal of Management Mathematics (2009) 20, 39−49

doi:10.1093/imaman/dpn007
Advance Access publication on June 27, 2008

Predicting overflow in an emergency department

L. AU†
Department of Mathematics and Statistics, The University of Melbourne,
Victoria, Australia
G. B. B YRNES
Centre for Molecular, Environmental, Genetic and Analytic Epidemiology, Department of
Public Health, The University of Melbourne, Victoria, Australia
C. A. BAIN
Directorate Office, Western and Central Melbourne Integrated Cancer Service,
Victoria, Australia
M. FACKRELL
Department of Mathematics and Statistics, The University of Melbourne,
Victoria, Australia
C. B RAND
Clinical Epidemiology and Health Service Evaluation Unit, Melbourne Health,
Victoria, Australia
D. A. C AMPBELL
Department of Medicine, Southern Clinical School, Monash University,
Victoria, Australia
AND

P. G. TAYLOR
Department of Mathematics and Statistics, The University of Melbourne,
Victoria, Australia
[Received on 9 October 2007; accepted on 4 February 2008]

Ambulance bypass occurs when the emergency department (ED) of a hospital becomes so busy that
ambulances are requested to take their patients elsewhere, except in life-threatening cases. It is a major
concern for hospitals in Victoria, Australia, and throughout most of the western world, not only from
the point of view of patient safety but also financially—hospitals lose substantial performance bonuses
if they go on ambulance bypass too often in a given period. We show that the main cause of ambulance
bypass is the inability to move patients from the ED to a ward. In order to predict the onset of ambulance
bypass, the ED is modelled as a queue for treatment followed by a queue for a ward bed. The queues are
assumed to behave as inhomogeneous Poisson arrival processes. We calculate the probability of reaching
some designated capacity C within time t, given the current time and number of patients waiting.

Keywords: access block; ambulance bypass; continuous-time Markov chain; emergency department;
Laplace transform.
† Email: l.au@ms.unimelb.edu.au


c The authors 2008. Published by Oxford University Press on behalf of the Institute of Mathematics and its Applications. All rights reserved.
40 L. AU ET AL.

1. Introduction
Large, modern hospitals provide a complex array of services. They are faced with an increasing demand
for emergency and elective patient access in an environment of budgetary constraints and performance
reviews. Unfortunately, patients need to queue to obtain hospital services.
In order to optimize the use of available resources, it is vital that the behaviour of these queues is
understood. In particular, it is important to recognize where additional resources will provide maximum
benefit and to provide forewarning of circumstances in which facilities may become congested to an
unacceptable extent.
In this paper, we are concerned with the queue for access to treatment in the emergency department
(ED) of a large metropolitan hospital (the Royal Melbourne Hospital, Victoria) and subsequently the
queue formed for admission to a ward. In particular, with regard to these queues, we are interested in
• ambulance bypass (or diversion): where the ED is judged to be so busy that ambulances are diverted
to other hospitals except in urgent cases and
• access block: where patients in the ED requiring inpatient care are unable to gain access to appro-
priate hospital beds within a reasonable time frame (for further details, see http://www.acem.org.au/
media/policies and guidelines/standard terminology.pdf).
These are not unrelated since access block is the major factor influencing ambulance bypass (see
Fatovich et al., 2005). It is also a safety issue and is associated with increased mortality (see Pham
et al., 2006; Richardson, 2006; Sprivulis et al., 2006; Hwang et al., 2006, and Schull et al., 2004).
Our data indicate that this is the principal cause of congestion in the ED; see Hoot et al. (2007) and
Jones et al. (2006) for a discussion on ED crowding. We present a model to describe patient flow in the
ED, focussing on the queue formed when patients are waiting for admission to a ward.
In Section 2, we describe in detail the queues in the ED and define the key terms. In Section 3, we
describe our mathematical model for the ward admission queue. Section 4 contains a discussion on the
validation of the model. The paper concludes with Section 5.

2. Queues in an ED
In this section, we describe the flow of patients in the ED from the time they arrive until they are either
discharged or admitted to a hospital ward for further treatment. While the process was observed at a
single hospital, we believe that the general features are common to most metropolitan tertiary hospitals
in Australia. The flow of patients can be summarized as follows:
(a) patients arrive, by self-referral, referral by a primary health care provider or other hospital or by
ambulance;
(b) on arrival, patients are assessed by a nurse or medical officer and are assigned to a triage category;
(c) patients either wait until they are treated or, in a small proportion of cases, leave before treatment;
(d) after treatment, patients are either discharged directly from the ED or wait for a ward bed to
become available so they can be admitted;
(e) patients are transferred to a ward.
Figure 1 illustrates the flow of patients in the ED from arrival through ward admission or discharge.
The Australasian Triage Scale has been developed by the Australasian College of Emergency
Medicine, is widely accepted throughout Australasia and is known internationally. (For further details,
see http://www.medeserv.com.au/acem/open/documents/triage.htm.) The triage categories are numbered
PREDICTING OVERFLOW IN AN EMERGENCY DEPARTMENT 41

FIG. 1. Schematic diagram showing patients’ flow in an ED.

from 1 to 5: if a patient is triaged as Category 1, then they are deemed to need immediate attention, while
Category 5 corresponds to a patient having minor symptoms of a low-risk condition. There is a max-
imum period of time allowed for the patient to see a doctor after triage for medical assessment and
treatment, which ranges from immediately to 2 h depending on the triage category. We refer to the
actual time that the patient waits as the waiting time and the queue that forms as the treatment queue.
Occasionally, after a patient has been triaged, he or she chooses to leave the ED rather than wait
for treatment. These patients are known as do-not-wait (DNW) or leave-without-being-seen (LWBS)
patients (see Baker et al., 1991; Rowe et al., 2006, for further details). The reasons for LWBS patients
leaving the ED vary, but include dissatisfaction with the length of time waiting for treatment. These
patients were included in the analysis as they contributed to the congestion experienced in the ED as
measured by queue length.
For each patient triaged, the arrival, treatment and departure times are recorded, see Table 1. Patients
depart from the ED either after assessment and treatment or when they are admitted to a ward. For those
patients needing a ward bed, the bed request time is also recorded. Our analysis is based on the data
collected (arrival, treatment, bed request and departure times from ED) between 1 January 2001 and
18 April 2005 from the Emergency Department Information System. (A computerized system used
to record patient care and movements through the ED. It is commonly used throughout most EDs in
Australia.) A total of 199,480 records appear.
The accuracy of data is an important issue. Although the hospital has addressed this in recent years,
some records were clearly inaccurate. Problems include missing arrival or departure times and implau-
sibly long delays in obtaining treatment or a bed. We excluded from the analysis those records where
the recorded treatment delay was greater than 1 day, the admission delay was greater than 2 days and the
total length of stay in the ED was greater than 3 days. Approximately 5% of all records were discarded
in this way (i.e. 9974 patients).
Generally a quarter (47,763 of 199,480) of ED presentations at this hospital require admission and
enter a second queue to obtain a bed in a ward. We refer to this queue as the bed queue. Here, the time
of arrival in the queue is the bed request time, and the service time is the difference between this time
and the departure time from the ED. The time in this queue is subject to scrutiny by the government,
42 L. AU ET AL.

TABLE 1 Typical patient information from a hospital database of the ED

Arrival time Treatment time Bed request time Departure time Triage category
1/1/2001 00:17 1/1/2001 00:26 1/1/2001 00:59 4
1/1/2001 00:20 1/1/2001 00:25 1/1/2001 01:45 3
1/1/2001 00:27 1/1/2001 00:31 1/1/2001 00:33 1/1/2001 00:48 3
1/1/2001 00:35 1/1/2001 00:39 1/1/2001 03:00 4
1/1/2001 01:10 1/1/2001 01:13 1/1/2001 03:05 4
1/1/2001 01:14 1/1/2001 01:16 1/1/2001 01:33 5
1/1/2001 01:20 1/1/2001 01:23 1/1/2001 03:35 3
1/1/2001 01:25 1/1/2001 01:29 1/1/2001 05:31 1/1/2001 11:01 3
1/1/2001 01:32 1/1/2001 01:35 1/1/2001 03:55 4
1/1/2001 01:34 1/1/2001 01:35 1/1/2001 02:00 4
1/1/2001 01:35 1/1/2001 01:38 1/1/2001 02:00 4
1/1/2001 01:38 1/1/2001 01:47 1/1/2001 02:00 3
1/1/2001 01:47 1/1/2001 01:50 1/1/2001 06:42 3
1/1/2001 01:50 1/1/2001 01:51 1/1/2001 04:56 1/1/2001 15:45 2

with financial penalties (or withdrawal of bonuses) associated with long waits. (For other hospitals’ key
performance indicators, see http://www.health.vic.gov.au/archive/archive2006/hdms/busrule06.pdf.)
Currently, the goal is that no patient should wait longer than 6 h.
A particular concern for EDs is access block, when the supply of beds is insufficient to service the
bed queue. Various definitions and measures of access block can be found in Forero et al. (2004). The
effect of this can lead to an inability to accommodate patients for treatment within the ED, significantly
contributing to the other major concern, ambulance (or hospital) bypass. Analysis of the 283 ambulance
bypass events from 1 February 2003 to 4 April 2005 for the study site suggests that access block is
responsible for 76% (95% confidence interval 71–81%) of all ambulance bypass. Unexpected patient
influx was seen as the cause in 13% of cases, with most (9%) of the remaining instances due to unusual
constraints on resources. Other hospitals have reported a much higher influx rate causing ambulance
bypass, see Fatovich & Hirsch (2003).
Figure 2 shows the empirical distribution of waiting times for the treatment and bed queues for
Mondays. Distributions for the other days of the week were similar. Patients spend far longer in the
bed queue than the treatment queue. The distribution of the number of patients waiting in the treatment
and bed queues for Mondays is shown in Fig. 3. Again, similar features are observed for the other
days of the week. It is apparent that typically there are more patients waiting in the bed queue than in
the treatment queue. At this point, we turn our attention to the bed queue because, given the evidence,
it contributes more to access block (and hence to ambulance bypass) than the treatment queue. (The
formal procedures for hospital to be on ambulance bypass are complicated, see the following web link
for details: http://www.dhs.vic.gov.au/ahs/circular/circ498.htm.)

3. ED capacity prediction model


Here, we present our model for the bed queue and derive an expression for the probability of it reaching
a given number of patients within time t, conditional on the number of patients initially waiting in the
bed queue.
In reality, the decision to initiate ambulance bypass in an ED is complex and is made subjectively
by the ED manager based upon a number of factors such as the size of the current queue, time and day
PREDICTING OVERFLOW IN AN EMERGENCY DEPARTMENT 43

FIG. 2. Distribution of the waiting times in the treatment and bed queues for Mondays.

FIG. 3. Distribution of the number of patients waiting in the treatment and bed queues for Mondays.

of the week, etc. However, it would aid decision making in the period leading up to possible ambulance
bypass events if there could be an accurate prediction of the probability that the bed queue will reach a
specified level C within the time period under consideration. Our analysis shows how to calculate these
probabilities.
At the Royal Melbourne Hospital, the capacity of the bed queue is approximately 20 (i.e. the physical
space in which the bed queue is located can fit 20 patients), so we treated C to be less than 20 in the
ensuing analysis.
44 L. AU ET AL.

The following is based on the work of Ramakrishnan (2002) (see also Ramakrishnan et al., 2005).
The bed queue was modelled as a finite-state, continuous-time Markov chain (see, e.g. Kleinrock, 1975).
We are interested in the probability that the number of patients in the bed queue of reaching C between
now and time t, which we consider as the event that triggers ambulance bypass. That is, ambulance
bypass is deemed to occur when n = C. Hence, we need to consider the queue only on the state space
S = {0, 1, 2, . . . , C}. State n corresponds to the current number of patients in the bed queue. Patients
arrive to the bed queue according to a Poisson process with rate λ, and whenever n beds are occupied a
departure occurs from the bed queue to a ward with rate μ(n).
The arrival and service rates have been observed to vary considerably with the time of the week and
the time of the year, so we have used spline regression to obtain estimates λ̃(t) and μ̃(n, t) of these rates
as continuous functions of time. Unfortunately, we do not have a method for directly using time-varying
rates in our Markov model, so instead we have used a 6-h moving average of these rates to parametrize
our 6-h ahead predictions. Hence, when making predictions for the 6-h window starting from time t, we
have used
Z Z
1 t+6 1 t+6
λ= λ̃(τ )dτ, μ(n) = μ̃(n, τ )dτ.
6 t 6 t

To assess whether this assumption materially affects predictions of access block, we tested predictions
against empirical data in Section 4.

3.1 The Markov chain model


The infinitesimal generator for the Markov chain describing the bed queue is
 
−λ λ 0 ∙∙∙ 0 0
 
μ(1) −(λ + μ(1)) λ ∙∙∙ 0 0 
 
 
 0 μ(2) −(λ + μ(2)) ∙∙∙ 0 0 
 
 
A= .
 .. .. .. .. .. .. 
 . . . . . . 
 
 
 0 0 0 ∙ ∙ ∙ −(λ + μ(C − 1)) λ 
 
0 0 0 ∙∙∙ μ(C) −μ(C)

Define pn (t) to be the probability of moving from state n to C in the time interval [0, t]. That is, the
probability that the ED will go on ambulance bypass by time t, given that n beds are currently occupied.
Let pn (t|x) be the probability of moving from state n to C in the time interval [0, t] given that the first
transition occurs at time x. Thus,

 0, 0 6 n 6 C − 1, x > t,





 p1 (t − x), n = 0, x 6 t,
pn (t|x) =
 λ μ(n)

 pn+1 (t − x) + p (t − x), 1 6 n 6 C − 1, x 6 t,


 λ + μ(n) λ + μ(n) n−1

1, n = C.
PREDICTING OVERFLOW IN AN EMERGENCY DEPARTMENT 45

We can recover the marginal probability pn (t) by integrating against f n , the probability density
function for x given a current occupancy n. Under the Markov assumption, x is exponentially distributed
with mean 1/λ, 1/(λ + μ(n)) and 1/μ(C) when n = 0, 1 6 n 6 C − 1 and n = C, respectively. We
have the following three cases:
1. n = 0,
Z ∞ Z t
p0 (t) = p0 (t|x) f 0 (x)dx = p1 (t − x)λ e−λx dx; (3.1)
0 0

2. 1 6 n 6 C − 1,
Z ∞
pn (t) = pn (t|x) f n (x)dx
0
Z t
= (λpn+1 (t − x) + μ(n) pn−1 (t − x))e−(λ+μ(n))x dx; (3.2)
0

3. n = C,
Z ∞ Z ∞
pC (t) = pC (t|x) f C (x)dx = μ(C)e−μ(C)x dx = 1. (3.3)
0 0
R
Taking the Laplace transforms Pen (s) = ∞ e−st pn (t)dt for all <(s) > 0 (i.e. real part of s) and
0
using the convolution theorem, we obtain the system of linear equations

e0 (s) = λ e
P P1 (s), (3.4)
s+λ
λ μ(n)
en (s) =
P en+1 (s) +
P en−1 (s),
P 1 6 n 6 C − 1, (3.5)
s + λ + μ(n) s + λ + μ(n)

eC (s) = 1 .
P (3.6)
s

For any s with <(s) > 0, the system of equations (3.4–3.6) can easily be solved for P e1 (s),
e0 (s), P
eC (s).
..., P
en (s), which is most readily done
Finally, to obtain pn (t), we need to invert the Laplace transform P
numerically. We chose to use the Euler method due to Abate & Whitt (1995) because, first, it is easy
to implement and efficient and, second, it works particularly well with functions f (t) with | f (t)| 6 1,
pn (t) being such a function. Equations (13–15) in Abate and Whitt give rise to

M  
X M
pn (t) ≈ 2−M s N + j (t), (3.7)
j
j=0

where
A    A l   
e2 A e2 X A + 2π ik
sl (t) = e
< Pn + k e
(−1) < Pn , for N 6 l 6 N + M, (3.8)
2t 2t t 2t
k=1
46 L. AU ET AL.

FIG. 4. The probability of reaching capacity C = 20 on Monday from 12 noon to 6 pm when n = 10, 15, 16, 17, 18, 19 and 20.


and i = −1. Abate and Whitt recommended choosing A = 18.4, M = 11 and N = 15 to ensure that
the approximation error is small. Furthermore, (3.7) and (3.8) can simply be expressed as
A    A M N +k     
2 XX M
pn (t) ≈
e2 en
< P
A e
+ M en A + 2π i j
(−1) j < P . (3.9)
2t 2t 2 t k 2t
k=0 j=1

In Fig. 4, we plot pn (t) for n = 10, 15, 16, 17, 18 and 19, from 12 noon to 6 pm on Monday
assuming that the capacity of the bed queue is C = 20 beds. Clearly, p20 (t) = 1 since the ED is on
ambulance bypass at t = 0. As pn (t) is the cumulative probability of ambulance bypass prior to time t,
we see that pn (t) increases monotonically for a fixed n. Also as expected, pn (t) increases with n for a
fixed t.

4. Model validation
Our model estimates the probability that the queue reaches level C in the time interval [0, t]. A key
assumption in the analysis was that the arrival and service rates are constant over the interval [0, t].
Our estimates for the data indicate that in fact the parameters vary over a timescale of hours, which
is of the same order as the time over which we are predicting ambulance bypass, violating the model
assumptions.
Pragmatically, the model may still be useful in predicting the overflow probability. To test this we
compared the model predictions with observations of a separate sample of data. The data used to validate
the model consisted of 48,092 observations spanning the period from 19 April 2005 to 26 September
2006, which is outside the time period used to estimate the model parameters λ and μ(n).
At the beginning of each 6-h block in the validation data set, we used the model to estimate the
probability that the bed queue would reach C within the block. Note that the total number of blocks
available for validation varied depending on the threshold C selected, since in some cases the observed
initial occupancy n exceeds C. The blocks were then grouped into 10 equally sized bins (each containing
Ni blocks) according to their estimated probability of exceeding C. We then calculated the mean pˉ i and
variance Vi of the probabilities in each bin.
PREDICTING OVERFLOW IN AN EMERGENCY DEPARTMENT 47

TABLE 2 Comparison of the observed and expected frequencies of ambulance bypass in 6-h blocks
assuming capacity C = 20

Bin 1 2 3 4 5 6 7 8 9 10
Number of blocks Ni 182 182 182 182 182 182 182 182 182 182
Observed hits Oi 0 0 1 2 4 8 13 25 55 110
Expected hits E i 0.01 0.10 0.38 1.03 2.36 4.92 9.77 19.21 42.45 106.70
(Oi − E i )2 /var(Oi ) 0.008 0.105 1.04 0.91 1.16 1.98 1.13 1.95 4.84 0.25
p-value = 0.16.

For each bin i, we recorded the number of times Oi that the capacity C was actually reached. This
was compared to the expected number E i = Ni pˉ i using a chi-squared goodness-of-fit test adapted from
Hosmer & Lemeshow (1980, p.1047). If there were any cells with less than five observations per bin,
then the bins were amalgamated for the purposes of the chi-squared test. For instance, in Table 2 the
first five bins were grouped together to give six degrees of freedom in a chi-squared goodness-of-fit test.
The goodness-of-fit statistic was
n
X (Oi − E i )2
G2 = ,
var(Oi )
i=1

with var(Oi ) estimated by


Ni pˉ i (1 − pˉi ) − Ni Vi

to allow for overdispersion relative to the binomial model, due to the spread of probabilities in each bin
and where n was the number of bins in the test. Under the null assumption of a correctly fitted model,
G 2 ∼ χn2 , as there were no parameters estimated from the test data.
The fit of the model for C = 20 is summarized in Table 2, with G 2 = 9.21 ( p = 0.16). The results
for other choices of capacity from 16 to 19 gave similar results, with a p-value between 0.16 and 0.82
(see Tables 3–6).

5. Summary
Ambulance bypass, the consequence of ED demand exhausting available inpatient resources, is a serious
medical management and financial problem for major hospitals all over the world. We have presented
a model that can aid in the prediction of ambulance bypass occurring within the short time in one
hospital. The data needed for this prediction are the current ED occupancy, the maximum ED capacity
and detailed estimates of rates of bed requests for admission to the hospital wards and allocation of
hospital beds from historical data.
Although some key mathematical assumptions of the model are violated (constant rates of arrival
and service), the model (based on a goodness-of-fit test) has performed well when validated against
independent, real data for the same hospital. We believe that this model could be a valuable tool for
hospital management by providing a quantitative assessment of the risk of ambulance bypass. This
could be used to anticipate the need for additional resources, such as emergency rostering of additional
staff, or to engage procedures to free inpatient beds well in advance of ambulance bypass conditions
actually occurring.
48 L. AU ET AL.

TABLE 3 Comparison of the observed and expected frequencies of ambulance bypass in 6-h blocks
assuming capacity C = 19

Bin 1 2 3 4 5 6 7 8 9 10
Number of blocks Ni 173 173 173 173 173 173 173 173 173 173
Observed hits Oi 0 0 1 3 5 10 20 22 47 100
Expected hits E i 0.02 0.23 0.76 1.89 4.07 7.61 14.62 26.20 49.24 102.57
(Oi − E i )2 /var(Oi ) 0.02 0.23 0.08 0.66 0.22 0.79 2.17 0.80 0.14 0.17
p-value = 0.79.

TABLE 4 Comparison of the observed and expected frequencies of ambulance bypass in 6-h blocks
assuming capacity C = 18

Bin 1 2 3 4 5 6 7 8 9 10
Number of blocks Ni 164 164 164 164 164 164 164 164 164 164
Observed hits Oi 0 1 2 2 10 15 25 43 63 100
Expected hits E i 0.06 0.49 1.48 3.42 6.86 12 20.93 36.36 60.65 105.23
(Oi − E i )2 /var(Oi ) 0.06 0.53 0.18 0.60 1.50 0.81 0.91 1.57 0.15 0.78
p-value = 0.65.

TABLE 5 Comparison of the observed and expected frequencies of ambulance bypass in 6-h blocks
assuming capacity C = 17

Bin 1 2 3 4 5 6 7 8 9 10
Number of blocks Ni 153 153 153 153 153 153 153 153 153 153
Observed hits Oi 0 1 3 8 9 23 20 36 70 96
Expected hits E i 0.14 0.98 2.69 5.68 10.64 17.50 28.30 44.60 71.22 107
(Oi − E i )2 /var(Oi ) 0.14 0 0.04 0.99 0.27 1.96 3 2.36 0.04 3.96
p-value = 0.17.

TABLE 6 Comparison of the observed and expected frequencies of ambulance bypass in 6-h blocks
assuming capacity C = 16

Bin 1 2 3 4 5 6 7 8 9 10
Number of blocks Ni 139 139 139 139 139 139 139 139 139 139
Observed hits Oi 1 3 5 10 13 20 33 50 70 103
Expected hits E i 0.30 1.76 4.44 8.35 14.69 23.03 34.38 50.22 72.10 105.37
(Oi − E i )2 /var(Oi ) 1.64 0.88 0.07 0.35 0.22 0.48 0.07 0 0.13 0.23
p-value = 0.82.
PREDICTING OVERFLOW IN AN EMERGENCY DEPARTMENT 49

Acknowledgements
A special thanks goes to Associate Professor Marcus Kennedy for the use of his hospital ambulance
bypass data.

Funding
The Australian Research Council through Linkage grant (LP0349153); Melbourne Health (LP0349153).

R EFERENCES
A BATE , J. & W HITT, W. (1995) Numerical inversion of Laplace transforms of probability distributions. ORSA J.
Comput., 7, 36–43.
BAKER , D. W., S TEVENS , C. D. & B ROOK , R. H. (1991) Patients who leave a public hospital emergency depart-
ment without being seen by a physician. Causes and consequences. J. Am. Med. Assoc., 266, 1085–1090.
FATOVICH , D. M. & H IRSCH , R. L. (2003) Entry overload, emergency department overcrowding, and ambulance
bypass. Emerg. Med. J., 20, 406–409.
FATOVICH , D. M., NAGREE , Y. & S PRIVULIS , P. (2005) Access block causes emergency department overcrowd-
ing and ambulance diversion in Perth, Western Australia. Emerg. Med. J., 22, 351–354.
F ORERO , R., M OHSIN , M., BAUMAN , A. E., L ERACI , S., YOUNG , L., P HUNG , H. N., H ILLMAN , K. M.,
M CCARTHY, S. M. & H UGELMEYER , C. D. (2004) Access block in NSW hospitals, 1999–2001: does the
definition matter? Med. J. Aust., 180, 67–70.
H OOT, N. R., Z HOU , C., J ONES , I. & A RONSKY, D. (2007) Measuring and forecasting emergency department
crowding in real time. Ann. Emerg. Med., 49, 747–755.
H OSMER , D. W. & L EMESHOW, S. (1980) Goodness of fit tests for the multiple logistic regression model. Comm.
Stat. Theory Methods, 10, 1043–1069.
H WANG , U., R ICHARDSON , L. D., S ONUYI , T. O. & M ORRISON , R. S. (2006) The effect of emergency depart-
ment crowding on the management of pain in older adults with hip fracture. J. Am. Geriatr. Soc., 54, 270–275.
J ONES , S. S., A LLEN , T. L., F LOTTEMESCH , T. J. & W ELCH , S. J. (2006) An independent evaluation of four
quantitative emergency department crowding scales. Acad. Emerg. Med., 13, 1204–1211.
K LEINROCK , L. (1975) Queueing Systems, Volume 1: Theory, New York: Wiley Interscience.
P HAM , J. C., PATEL , R., M ILLIN , M. G., K IRSCH , T. D. & C HANMUGAM , A. (2006). The effects of ambulance
diversion: a comprehensive review. Acad. Emerg. Med., 13, 1220–1227.
R AMAKRISHNAN , M. (2002) Modelling an acute care facility. Honours Thesis, Department of Mathematics and
Statistics, University of Melbourne, Victoria.
R AMAKRISHNAN , M., S IER , D. & TAYLOR , P. G. (2005) A two-time-scale model for hospital patient flow. IMA
J. Manag. Math., 16, 197–215.
R ICHARDSON , D. B. (2006) Increase in patient mortality at 10 days associated with emergency department over-
crowding. Med. J. Aust., 5, 213–216.
ROWE , B. H., C HANNAN , P., B ULLARD , M., B LITZ , S., S AUNDERS , L. D., ROSYCHUK , R. J., L ARI , H.,
C RAIG , W. R. & H OLROYD , B. R. (2006) Characteristics of patients who leave emergency departments
without being seen. Acad. Emerg. Med., 13, 848–852.
S CHULL , M., V ERMEULEN , M., S LAUGHTER , G., M ORRISON , L. & DALY, P. (2004) Emergency department
crowding and thrombolysis delays in acute myocardia infarction. Ann. Emerg. Med., 44, 577–585.
S PRIVULIS , P. C., DA S ILVA , J. A., JACOBS , I. G., F RAZER , A. R. & J ELINEK , G. A. (2006) The association
between hospital overcrowding and mortality among patients admitted via Western Australian emergency
departments. Med. J. Aust., 5, 208–212.

You might also like