Queue Healthcare
Queue Healthcare
DOI 10.1007/s10916-010-9499-7
ORIGINAL PAPER
Received: 9 February 2010 / Accepted: 12 April 2010 / Published online: 8 May 2010
# Springer Science+Business Media, LLC 2010
Abstract Patient queues are prevalent in healthcare and service system and short customer wait time. In healthcare,
wait time is one measure of access to care. We illustrate patient wait time is one measure of access to care and
Queueing Theory—an analytical tool that has provided excessive waiting may lead to poorer health outcomes.
many insights to service providers when designing new Hence, the imperative to reduce it to a clinically acceptable
service systems and managing existing ones. This estab- level before treatment should begin.
lished theory helps us to quantify the appropriate service Queues exist everywhere but some are not very visible.
capacity to meet the patient demand, balancing system A few examples in the hospital are described. As patients
utilization and the patient’s wait time. It considers four key are prioritized and stabilized in the Emergency Department
factors that affect the patient’s wait time: average patient (ED), patients of lower acuity wait longer until their first
demand, average service rate and the variation in both. We consult. When new patients of higher acuity arrive later,
illustrate four basic insights that will be useful for managers they may push them behind and the average wait time may
and doctors who manage healthcare delivery systems, at be affected. Small surges in ED patient arrivals can lead to
hospital or department level. Two examples from local several patients breaching the maximum wait time limit for
hospitals are shown where we have used queueing models admission, due to the fixed capacity of inpatient beds. In
to estimate the service capacity and analyze the impact of the outpatient setting, the queue or appointment wait list of
capacity configurations, while considering the inherent subsidized and private patients seeking their first consult is
variation in healthcare. both an indicator of the demand and has implications on the
allocation of resources for service differentiation. In an
Keywords Queueing theory . Simulation . Healthcare imaging department, diagnostic tests may not be read one at
applications . Endoscopy planning . Infection control a time as they arrive, but may be batched together into a
queue to reduce the ‘set up’ cost of specialists’ time, but this
has to be balanced with the wait time for the test result. Also,
Introduction every batching exercise may have a downstream impact on
those service providers who receive the batched results in
Queues are commonplace in an urban setting. We partici- intervals. The batching effect is also seen in ward rounds that
pate in two roles: either as a quality conscious and cost are done once a day fixing the decision cycle to 24 h.
aware service provider or as the person waiting in line with The purpose of this paper is to highlight the benefits of
some frustration, be it in traffic, counters or surgery. But the queueing analysis to healthcare managers and doctors. It is the
manager has to find a balance between high utilisation of a authors’ observation that the formidable mathematics in
queueing models does not make them accessible to the
resource allocators who need them most. So they have been
avoided and the intuition behind the quantitative relationships
R. K. Palvannan (*) : K. L. Teow
is developed. Interested readers will find the derivations in
National Healthcare Group,
Singapore, Singapore books on queueing theory like [1]. This paper focuses on the
e-mail: [email protected] factors that affect the patient’s wait time and provide basic
542 J Med Syst (2012) 36:541–547
Seeing an ED patient ED registration rate Triage, consult throughput Waiting for triage, consult etc.
Admitting an ED patient Bed request rate Discharge rate Waiting for admission
Servicing an outpatient appointment Call up rate Patients seen per day Appointment wait list
Emergency Operation Surgery indication rate Specialty surgery rate Waiting for surgery excluding preparation time
(Poisson is intuitive for demand) or duration (negative tracing to the right of the x-axis) if the average patient
exponential is intuitive for service). demand has increased or the system capacity has been
reduced. With this a few observations follow.
Firstly, as the system gets busier, the patient waits
Insights from queueing theory longer, and the waiting gets inexorable as the system is
close to congestion. An increase from 80% to 90% has a
This section highlights fundamental insights of the rela- larger impact on patient wait time than an increase from
tionship between patient demand, service rate and waiting 70% to 80%. The relationship is subtle and non-linear and
time when there is significant variability. Though these can can be counter intuitive as proportional assumptions do not
be written as equations, they are shown graphically to hold at high utilization. If we operate at a high utilisation of
illustrate counter intuitive relationships. 85%, we are sensitive to small changes in demand and
capacity i.e. a small increase in utilisation will cause a large
1. Queues are formed even when capacity is greater than increase in waiting. This is typically faced in a hospital that
demand, when there is variation operates with a high bed occupancy rate (BOR), as a small
surge in bed requests means ‘suddenly’ several patients
Few would argue with the converse—that when demand may have to wait for admission for a very long time [10],
is higher than capacity, there is interminable wait. But even triggering much operational response. On the other hand, a
if we service faster than the rate at which patients are small increase in bed capacity at high BOR—possibly by
coming in, there will still be queues. This happens as using temporary beds reserved for other use—tends to
arrivals and services have variation: they cannot be exactly restore the admission wait time very quickly.
predicted in time and therefore cannot be synchronized. At Secondly, to achieve a low wait time, the system must
times, the service rate is higher than the demand, and operate at a very low utilisation by investing in slack
momentarily the service provider will have some free time. capacity and therefore have a high cost structure. Different
This is less likely in a busy service. As the unused capacity response systems operate at different utilisation levels to
cannot be stored and recovered, the average utilization is cater to their patients’ urgency. An emergency response
less than 100%. On the other hand, most of the time, the system (e.g. ambulance) may have to operate at about 20%
capacity is less than the demand and the service providers and a relatively urgent admitting unit (e.g. ICU) at 75%.
are busy seeing one patient after another while the new The graph also shows that when we operate at a low
ones are waiting. The result is a situation where though the
20
average capacity was higher than the average demand
giving a high but not 100% utilisation, the patients have to 18
12
10
2. Short wait time demands low system utilization and
8
high cost structure.
6
48
utilization, it takes a large increase in capacity to reduce the
44
wait time targets, as we have diminishing returns to capacity 40
20-bed
40-bed
increase. 36
100-bed
32
typical of a system with higher variation compared with Fig. 4 Impact of size and partitioning of service capacity on average
Curve B. For the same utilisation level we see a longer wait time
patient wait time. Line C shows an ideal setting with no
variation, which is typical of a synchronized automated bed unit, as the smaller bed unit will face bed shortages
arrival and service mechanism where congestion happens much more frequently.
only when demand exceeds capacity. Similarly this principle of scale applies when a 100 bed unit
This figure suggests a guide to decide the appropriate is partitioned into smaller units for some reason. If the patients
operating utilisation level for a service. We start with the are homogeneous—having the same arrival and service
clinical decision that defines the acceptable average patient patterns—and the service is split into two units there may be
wait time, marking a point on the y-axis. Next we look at a significant impact on patient wait time. So a unit of 100 beds
sample historical data to estimate the variation in arrival will have a shorter wait time than when split into five 20 bed
and service rates—this will help to map out the suitable units, although the total capacity is the same. Splitting a
operating curve. Reading off the x-axis, we can identify the service is a common intervention that we consider and it
region of operating utilisation and not far to its right will be would be good to analyze the impact on patient wait time.
its congestion point. Putting it altogether, the expression below shows how
the four factors affect the patient’s average wait time. This
4. Size and partitioning impacts patient wait time. is an approximation for a system with high utilization as
exact equations are not known [3]. The expression shows
It is known that a larger service unit has economies of that the average wait time is proportional to the square of
scale as the fixed overhead is spread thinner, supporting an the variability in patient arrival or demand (σx) and the
argument for expansion, until some threshold is reached service duration (σc) but lessened by the number of service
[11]. Additionally the larger the pool of resources, the better providers (c). And the higher the arrival rate (λ) or lower
is the absorption of variation in arrival and service duration. the service rate (μ), the longer is the wait time.
Figure 4 illustrates that a 100 bed unit is able to operate at a
s2
higher utilisation for the same patient wait time than a 20 l s 2x þ c2c
Patient0 s average wait time
l
20
2 1 cm
18
16
14
Case examples
Average wait time
12
We will move on to see how the application of simple
10
A B C analytical queueing models, has helped to develop insights
8
into capacity sizing and impact analysis in hospital settings.
More variation
6 They were analyzed using QTSPlus® from [1], an Excel
4 add-in tool for queueing analysis.
2
0
Example 1—Capacity sizing to meet target patient wait time
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Utilisation
The endoscopy centre of Tan Tock Seng Hospital, Singapore
Fig. 3 Impact of variability in demand and service on average wait time considered expanding the number of endoscopy theatres from
J Med Syst (2012) 36:541–547 545
Endoscopic operation suites Table 3 Patient arrival and length of stay data by cluster
Table 2 Percentage of patients who have to wait for a recovery bed Table 4 Impact of cohorting at cluster and hospital level on
Prob of not Demand = # of theatres * % req beds * Recovery time / Ops duration admission wait time
finding a bed 5 6 7 8 9 10 11 12 13 14
22 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.1% 0.3% 0.7% 1.2%
Input parameters Additional bed capacity required (%)
21 0.0% 0.0% 0.0% 0.0% 0.0% 0.1% 0.2% 0.6% 1.1% 2.0%
20 0.0% 0.0% 0.0% 0.0% 0.1% 0.2% 0.5% 1.0% 1.8% 3.0%
# of Recovery beds
19 0.0% 0.0% 0.0% 0.0% 0.1% 0.4% 0.8% 1.6% 2.8% 4.4% Cohorted Cohorted Scenario 1 Cluster Scenario 2
18 0.0% 0.0% 0.0% 0.1% 0.3% 0.7% 1.5% 2.7% 4.3% 6.3%
17 0.0% 0.0% 0.1% 0.2% 0.6% 1.3% 2.5% 4.1% 6.2% 8.6% group (%) patients level cohorting Hospital level
16 0.0% 0.0% 0.1% 0.5% 1.1% 2.2% 3.9% 6.0% 8.6% 11.5% ALOS (days) cohorting
15 0.0% 0.1% 0.3% 0.9% 2.0% 3.6% 5.9% 8.6% 11.6% 14.8%
14 0.0% 0.2% 0.7% 1.7% 3.4% 5.7% 8.5% 11.7% 15.1% 18.6%
13 0.1% 0.5% 1.4% 3.1% 5.4% 8.4% 11.9% 15.5% 19.2% 22.8% 5% 8 5.4% 1.6%
12 0.3% 1.1% 2.7% 5.1% 8.3% 12.0% 15.9% 19.9% 23.7% 27.5%
11 0.8% 2.3% 4.8% 8.1% 12.1% 16.3% 20.6% 24.8% 28.7% 32.4% 10% 12 7.7% 2.1%
10 1.8% 4.3% 7.9% 12.2% 16.8% 21.5% 26.0% 30.2% 34.1% 37.7%
546 J Med Syst (2012) 36:541–547
separated from those who are not. This is expected to tative approximation. When evaluating or planning a new
improve the adherence to hand hygiene protocol of health- service system we start by estimating the wait time and
care providers. But it is also felt that the more hard queue, using the queueing models. It is then followed by a
partitions are imposed within a homogeneous ward, the discrete event simulation which handles greater complexity.
greater will be the chances of smaller ring fenced wards The queueing analysis guides a subsequent simulation
being full and therefore increases patient’s admission wait analysis by assessing the feasibility of resource require-
time. We will see how we can quantify the impact using the ments and screening the alternatives of a service design,
partitioning concepts we had looked at earlier. thereby reducing the exploration space when simulating.
Table 3 shows the statistics of the ED patients who were Since they require aggregate data they do not impose on
admitted. It shows the weekday bed request rate and length high resolution data requirements at an early stage of
of stay of patients in the hospital by clusters (groupings of engaging the managers. But the queueing models, at the
specialties e.g. oncology, orthopedics). There is significant benefit of being tractable, have limitations. Classical
variation in demand and service which is expected as these models do not handle the time variation in the arrival rate
come from unscheduled emergency patients. The hospital’s over the course of the day or day of the week and plan for
average length of stay (ALOS) was 5 days and the average peak load only. Much information is lost when the
bed wait of the patient in the ED was 2 h. The classical waiting probability distributions are not Poisson and complex
model with Poisson arrival and service rate with multiple routing logic in a queueing network is not addressed.
servers was used to estimate the effect of partitioning. It was While the academic community has developed the
analyzed using the M/M/c module in QTSPlus®. theory of queueing models to sophistication, the authors
Moving forward, we want to see the impact of partitioning believe that they have not been exploited by the resource
every cluster into 2 groups: one for the clean patients and one managers in healthcare. Much queueing development has
for those who are colonized. We looked at 2 scenarios of been done for performance modeling of computer networks
cohorting: one at a cluster level and one at a hospital level. It is and some of these may be applicable to the vast queueing
assumed that the ALOS of the cohorted patients is in the range networks in healthcare at hospital and health system level.
of 8–12 days and that the percentage of cohorted patients is in On the other hand, some simulation analyses may not
the range of 5%–10% as this may vary across the clusters. The precede their computational effort by exploiting queueing
clusters are partitioned and the new wait time is computed models to leverage their analysis.
which is expected to be longer. After that we computed how In summary, inherent variation in patient presentation
much of additional bed capacity one needs to add to restore and treatment course means that patient queues have to be
the bed wait time back to 2 h. The additional bed capacity is managed through appropriate capacity sizing and impact
another way of expressing increase in bed wait time due to analysis. We have seen how queueing theory and the
partitioning. fundamental principles of queueing systems offer practical
The results are shown in Table 4. The first 2 columns are the insights on the factors that impact the patient’s wait time.
2 unknowns: cohorted patient load and cohorted patients’ Patient flow is nevertheless complex and we will have to
ALOS. The next 2 columns show the results of the 2 scenarios: resort to computational simulation studies that require
cohorting at a cluster level and hospital level. So when we extensive data to get a more accurate picture of the
have 5% of cohorted patients and ALOS of 8 days and we workflow and resource bottlenecks. But a preliminary
partition each cluster into a clean and cohorted ward, the queueing analysis can help us to design new and manage
impact on the wait time will be such that, an additional 5.4% existing queueing systems in hospitals to a large extent.
bed capacity is required to restore the original 2 h wait time.
Hospital level cohorting means an impact of only 1.6% of Acknowledgements We thank Ms Grace Chiang (Director Opera-
additional bed capacity, naturally due to the large pooling tions, National University Hospital), Sister Y P Chia and Sister H H Tan
effect. In short, cluster level cohorting has a significant impact (Tan Tock Seng Endoscopy Centre) for permission to use their data.
ranging from 5%–7% while hospital level cohorting is only
about 1%–2%. This was a counter intuitive observation as the
increase in the bed capacity is to absorb the increase in the wait References
time due to partitioning for the same overall patient demand.
1. Gross, D., Shortle, J. F., Thompson, J. M., and Harris, C. M.,
Fundamentals of queueing theory, Wiley series, 2009.
Discussion 2. Halstrøm, H. L., and Jensen, A., The life and works of A.K.
Erlang. By E. Brockmeyer, Transactions of the Danish Academy of
Technical Sciences, 1948 No. 2.
This paper has highlighted the usefulness of analytical 3. Kleinrock, L., Queueing systems: Volume I & II—Theory. Wiley
queueing models in giving qualitative insights and quanti- Interscience, New York, 1975.
J Med Syst (2012) 36:541–547 547
4. Green, L., Queueing analysis in healthcare, in Patient flow: 9. Green, L., How many hospital beds? Inquiry 39:400–412, 2002.
reducing delay in healthcare delivery, Hall, R.W., ed., Springer, Winter 2002/2003.
New York, 281–308, 2006. 10. Bagus, A., Place, M., and Posnett, J. W., Dynamics of bed use in
5. Fomundam, S., and Herrmann, J. W., A survey of queuing accommodating emergency admissions: stochastic simulation
theory applications in healthcare. Technical report, University model. BMJ 319:155–158, 1999.
of Maryland, 2007. 11. Teow, K. L., Practical operations research applications for health-
6. Worthington, D. J., Queueing models or hospital waiting lists. J. care managers. Ann. Acad. Med. Singapore 38:564–566, 2009.
Oper. Res. Soc. 38(5):413–422, 1987. 12. Teow, K. L., Chia, Y.P., Tan, H. H., Zhu, Z., Heng, B. H., A pure
7. Restrepo, M., Henderson, S. G., and Topalogu, H., Erlang loss loss queueing model for endoscopy recovering bed planning.
models for the static deployment of ambulances. Health Care NHG Annual Scientific Congress 2009. 16–17 Oct 2009.
Manage. Sci. 12:67–79, 2009. 13. Palvannan, R. K., Teow, K. L., Zhu, Z., Heng, B. H., Estimating
8. McManus, M. L., Long, M. C., Cooper, A., and Litvak, E., impact of service partitioning with demand variability using
Queueing theory accurately models the need for critical care queueing theory. NHG Annual Scientific Congress 2009. 16–17
resources. Anesthesiology 100(5):1271–1276, 2004. Oct 2009.