OPERATIONS RESEARCH

Vol. 57, No. 2, MarchApril 2009, pp. 261273

issn0030-364X[ eissn1526-5463[ 09[ 5702[ 0261
informs

doi 10.1287/opre.1080.0584
2009 INFORMS
OR PRACTI CE
Efcient Short-Term Allocation and Reallocation
of Patients to Floors of a Hospital
During Demand Surges
Steven Thompson
Robins School of Business, University of Richmond, Richmond, Virginia 23173, [email protected]
Manuel Nunez, Robert Garnkel
School of Business, University of Connecticut, Storrs, Connecticut 06269-1041
{[email protected], [email protected]}
Matthew D. Dean
College of Business, University of New Orleans, New Orleans, Louisiana 70148, [email protected]
Many hospitals face the problem of insufcient capacity to meet demand for inpatient beds, especially during demand
surges. This results in quality degradation of patient care due to large delays from admission time to the hospital until
arrival at a oor. In addition, there is loss of revenue because of the inability to provide service to potential patients.
A solution to the problem is to proactively transfer patients between oors in anticipation of a demand surge. Optimal
reallocation poses an extraordinarily complex problem that can be modeled as a nite-horizon Markov decision process.
Based on the optimization model, a decision-support system has been developed and implemented at Windham Hospital
in Willimantic, Connecticut. Projections from an initial trial period indicate very signicant nancial gains of about 1% of
their total revenue, with no negative impact on any standard quality of care or stafng effectiveness indicators. In addition,
the hospital showed a marked improvement in quality of care because of a resulting decrease of almost 50% in the average
time that an admitted patient has to wait from admission until being transferred to a oor.
Subject classications: hospitals; health care; dynamic programming; Markov decision processes; decision-support
systems.
Area of review: OR Practice.
History: Received July 2005; revisions received May 2006, March 2007, August 2007; accepted December 2007.
Published online in Articles in Advance January 5, 2009.
1. Introduction
Over the past 25 years, United States hospitals have been
subjected to signicant transformations of the operating
landscape. The large-scale penetration of health mainte-
nance organizations in the early 1980s, the Emergency
Medical Treatment and Labor Act in 1986, and Medicare
reform associated with the Balanced Budget Act of 1997,
have limited the ability of hospitals to turn away patients
that are unable to pay for services, and have placed lim-
its on the amount hospitals are able to collect for the ser-
vices they provide. These changes have forced hospitals to
improve operating efciency (Krein and Casey 1998). As a
result, hospitals have aggressively reduced inefciencies by
cutting staff, managing length of stay, and nding innova-
tive ways to reduce incidental costs.
One side effect of these cost-containment efforts has
been a reduction in the number of inpatient beds for which
hospitals maintain staff and to which patients can be admit-
ted. The United States lost a total of 100,000 hospital beds,
including 7,800 intensive care beds between 1990 and 1999
(Kellerman 2001). The loss of inpatient beds has left many
hospitals with minimal surge capacity to handle spikes in
demand. The result is that beds are often in short supply,
and the problem is expected to get worse (Dodge 2001,
Solberg et al. 2003, Wilson 2001).
Emergency department (ED) overcrowding is another
undesirable result of insufcient oor capacity. In a March
2003 report (GAO 2003), the United States General Ac-
counting Ofce found that the factor most commonly asso-
ciated with crowding was the inability to move emergency
patients to inpatient beds, once a decision had been made
to admit them as hospital patients, rather than to treat and
release them. The impact on ED management is clear. As
the number of boarding patients who are waiting for beds
on their assigned oors increases, the resources available
to treat other ED patients are reduced, and eventually the
department ceases to function effectively. Often hospitals
261
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
262 Operations Research 57(2), pp. 261273, 2009 INFORMS
will then attempt to divert ambulances to other facilities
until some of the boarded patients are transferred out of
the ED, yielding a very signicant negative outcome for
patients and the hospital. For patients, the effect of diver-
sion is a longer ambulance ride, which prolongs the delay
in receiving medical treatment. For the hospital, each ambu-
lance that is diverted represents lost revenue. In addition,
when the ED becomes overcrowded, patients with minor
complaints tend to leave the hospital before being seen by
a physician. In these cases, the hospital has missed the
opportunity to provide a service for which payment would
otherwise have been rendered.
An example of a hospital facing the problems delin-
eated above is Windham Community Memorial Hospital
(WCMH), located in Willimantic, Connecticut, and servic-
ing a surrounding community of approximately 100,000
people. Of interest here was that in 2004, when this study
was initiated, WCMH would often begin to experience
capacity-related problems that resulted in patients being
boarded in the ED for the short term, e.g., 46 hours, well
before all inpatient beds were utilized. As a result, delays
were incurred when a patient had to be transferred from
a given oor to make room for a new patient that could
only be assigned to that oor. These last-minute transfers
are often done under duress during critical crunch peri-
ods and are generally undesirable from the standpoints of
patient ow and quality of care.
The bed manager (decision maker) determines the ini-
tial assignments of patients to oors. Another task is to
determine when, and if, it is necessary to transfer patients
from one oor to another, even in noncrunch periods. While
it is generally undesirable to transfer patients after admis-
sion, it was determined that there was a critical need at
WCMH for an optimality-based decision-support system
(DSS) for the bed manager that would allow for preemp-
tive (prior to the occurrence of a demand surge) transfers of
patients between oors, and for the assignment of patients
to oors based partially on capacity considerations. For the
former, in-house patients are transferred for the purpose
of capacity reallocation (proactive transfer), as opposed to
as a last-minute immediate response to make room for
newly admitted patients (reactive transfer). For the latter,
even when beds are available on the ideal (based strictly
on oor specialization) oor, newly admitted patients may
be assigned to feasible alternate oors.
We modeled the problem of nding an optimal capacity
utilization strategy based on patient allocation as a mul-
tidimensional, discrete-time, nite-horizon Markov deci-
sion process. The model has been integrated into a DSS
that has been implemented and, based on an initial trial
period, is projected to result in very signicant nancial
gains of about $600,000 per year, or 1% of total revenue.
No negative impact resulted on any standard quality of
care or stafng effectiveness indicators. In addition, there
was a marked improvement in quality of care because of
a resulting decrease of almost 50% in the average time
that an admitted patient has to wait from admission until
being transferred to a oor. Based on this success, WCMH
decided to create an operations manager position, to be
lled by an individual who will work with the system and
will also identify other opportunities to improve patient
ow and hospital efciency.
1.1. Overview
Many hospitals maintain a myopic strategy of assigning
admitted patients to their ideal oors, based on diagno-
sis, as long as there are available beds on those oors.
Although this strategy works well in many cases, when
capacity is limited it can result in patient ow bottlenecks
that have a number of negative quality of care and nancial
implications.
In hospital settings, the general overall problem is quite
complex. Departures must be considered as well as arrivals,
and the number of oors and patient categories could be
large. Although techniques have been developed for solving
stochastic sequential decision problems, the basic problem
presented here is challenging for a few reasons:
1. There generally are a large number of different patient
categories and oors, depending on the size of the hospital
and the range of diagnoses it can treat.
2. The number of possible actions/policies to consider is
very large.
3. Random events such as patient arrivals and departures
depend on patient category, type of oor, and the current
state of the hospital, and are often nonhomogenous with
respect to time of day, day of week, and season of the year.
4. The amount of time available to reach a decision is
fairly short, i.e., typically less than ve minutes.
In this paper, we develop and implement a solution
methodology that addresses those issues. The remainder
of this paper is organized as follows. First, we dene the
main problem under consideration in 2. In 3, we model
the problem as a nite-horizon Markov decision process
(MDP) and study the computational complexity of nd-
ing an optimal solution to the MDP and conclude that
traditional approaches are not practical for this problem.
In 3.7, we present an original approximation methodol-
ogy that relies on event aggregation to nd an approxi-
mate decision rule solution to the MDP. Section 4 deals
with several issues related to the implementation of our
approximation methodology in general, and we describe the
rolling-horizon algorithm used to implement a DSS based
on the MDP model. Section 5 describes our computational
experience, and 6 presents the details of the implementa-
tion of the DSS at WCMH, including anecdotal experience,
managerial insights, and an analysis of the impact on the
hospital.
1.2. Literature Review
The related problems of hospital and ED crowding have
been addressed from clinical and managerial viewpoints
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
Operations Research 57(2), pp. 261273, 2009 INFORMS 263
by health care researchers. Medical researchers have found
that diverting ambulances signicantly lessens the availabil-
ity of ambulances for patients in need of medical treatment
(Eckstein and Chan 2004). This diversion of ambulances
has been found to be primarily due to holding admitted
patients in the ED (Schull et al. 2003). The quality of care
and nancial impact of holding patients in the ED have
been found to be signicant for patients with chest pain
(Bayley et al. 2005) and those in need of thrombolytic ther-
apy (Schull et al. 2004).
Efforts to alleviate the ED crowding problem have
largely focused on reducing the amount of time required
to provide emergency services. Examples include bedside
registration (Parker 2004) and applications of simulation
and queueing theory to identify and eliminate bottle-
necks within the ED (Litvak et al. 2001). Other strate-
gies designed to move admitted patients out of the ED as
quickly as possible include faxing patient condition reports
to the oors (Caissie 2004) and admitting patients to hall-
ways (Derlet and Richards 2000). These efforts, while
achieving some success in ensuring that patients in need
of emergency medical treatment are able to receive it, have
had little effect on the waiting time before moving admitted
patients from the ED to the oors.
The OR/MS literature contains a long history of research
devoted to solving assignment/allocation problems in the
health care and emergency services settings. For a sum-
mary of applications of OR/MS to problems found in the
emergency services setting, see Green and Kolesar (2004).
In addition, a history of the applications of OR tech-
niques to problems in the health care industry can be found
in Flagle (2002). The application of OR techniques to
health care problems also received considerable attention in
Brandeau et al. (2004). Nevertheless, none of these models
can be directly used in our situation due to characteristics
of the problem that are formalized later in the paper.
There has also been a substantial amount of work address-
ing the fundamental problem of assigning different jobs
to different stations when the number of jobs (patients
in our case) and/or the amount of resources (beds) avail-
able to complete each job is unknown. In general, when the
randomness arises from either the number of jobs or the
amount of resources available, but not both, the problem
is typically modeled as a stochastic generalized assignment
problem (SGAP), e.g., see Albareda and Fernandez (2000)
and Mine et al. (1983). To the best of our knowledge, there
are no known exact results or algorithms for the SGAP.
When both the number of jobs and the amount of
resources available are random, queueing theory is typically
used. Queueing theory has been extensively researched and
applied in health care scenarios such as clinics (Cox et al.
1985), hospital appointment systems (Jackson et al. 1964),
and operating room stafng (Tucker et al. 1999). In these
cases, the stochastic processes and the problem structures
were amenable to established queueing models. Speci-
cally, the arrival of all patients was assumed to follow a
Poisson process, all patients were treated as identical in
terms of service time, and no distinctions were made in
terms of the clinical requirements of the patients. Recent
work on the scheduling of emergent and elective surgeries
(Gerchak et al. 1996) and emergent and elective radiology
cases (Green et al. 2006) are related to the problem we
address in that multiple classes of demand are considered.
Despite some similarities, our problem incorporates many
other characteristics that render inappropriate the modeling
assumptions made in those references. It is pointed out by
Green and Nguyen (2001) that to understand the impact
on patient service, more sophisticated methodologies are
needed to support decisions that involve bed capacity and
organization. The goal of this paper is to develop a method
to improve the efciency of capacity allocation.
2. Problem Denition
The number of oors in a hospital is very small compared
to the possible number of characteristics of the patients
that it serves. There exists a standard grouping terminol-
ogy called diagnostic related group (DRG) that is based
on diagnosis upon admission. However, there are over 500
possible DRGs, so that using them as the basis for opti-
mization models results in inordinately large problems.
Instead, we partition the set of DRGs into categories
according to three parameters: (1) the set of oors that
can treat the patients; (2) the ideal (to be expanded on
later in the paper) oor for the DRG; and (3) the expected
lengths of stay of the patients. Then, within a category
DRGs must be identical in the rst two parameters, and
very similar in the third. In the WCMH implementation,
we used 12 patient categories because that was the mini-
mum number that satised the above criteria. All patients
are considered to have arrived to the system as soon as they
are admitted to the hospital. Once admitted, patients wait
to be assigned and then transported to oors.
In our model, the system state at any point in time is
represented, for each patient category, by the number wait-
ing and the number being cared for on each oor. The
patient interarrival times to the system, and the number of
patient departures during a given time interval from the
hospital, are random and dependent on the patient category,
oor assigned, and the time dimension. The parameters of
the arrival and departure distributions are based on actual
historical data at the hospital rather than following any pre-
determined distributions. Patient interarrival times are inde-
pendent of the number of departures.
The bed manager observes the state of the hospital after
a xed-length time interval (period), records the number of
patient arrivals and departures during the period, and makes
assignment and transfer decisions. Because of capacity con-
straints, some admitted patients may not be assignable at
that time, and would have to wait until the next period
before assignment to any oor. The problem is to deter-
mine the best patient assignment and transfer decision in
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
264 Operations Research 57(2), pp. 261273, 2009 INFORMS
terms of immediate assignment rewards and transfer costs
incurred from making the decision, and in terms of long
term future expected rewards and costs.
2.1. Staff and Quality of Care
In any hospital, one measure of quality of care is based on
predened maximum ratios of patients-to-staff and patients-
to-licensed-staff. Staff arrivals and departures exhibit much
less variability than those associated with patients and are
assumed known in advance. Staff members (e.g., registered
nurse, nurses aide, etc.) can always be deterministically
assigned so as to satisfy these ratios, which is possible
because the number and type of staff of the hospital are
already set up to correspond to the bed capacity on each
of the oors. In the short term, it is always true that safe
stafng can be met, even if a manager or supervisor must
step in. Furthermore, as a general rule the hiring strategy of
a hospital is to obtain enough staff to run a oor at maxi-
mum capacity if needed. In the case of a long-term inability
to maintain full stafng levels (injuries, vacations, etc.), the
ratios can be satised by adjusting the capacities of the
oors. For example, if the bed manager knew that for sev-
eral weeks she would not be able to staff more than four
nurses on a oor where the ratio is one nurse for every two
patients, she could temporarily reduce the capacity of the
oor to eight beds from its present level during that period.
Hence, as long as patients are on the oors, proper care
will be provided and we do not explicitly consider staff
allocation as part of our model.
3. Model Formulation and Notation
We model the allocation problem as a nite-horizon, dis-
crete-time, stationary MDP. Time is split into xed-length
intervals (periods). The state of the process is observed at
the beginning of a period, and a decision is chosen from
a nite set of possible decisions. An immediate cost is
incurred depending on the state and decision, which deter-
mine the transition probabilities for the next state. That
state is realized at the end of the period, the process state
is updated, and the procedure repeats.
3.1. Notation Summary
We give the essential notation for various sections of the
paper:
Parameters
c
]
: maximum capacity of oor ];
E : index set of oors;
Q: index set of patient categories;
E
i
: index set of feasible oors for patients of category i
(E
i
E );
o
i]
: reward from assigning a category i patient to oor ];
l
i]k
: cost of a category i patient transfer from oor ] to
oor k.
State Space
m: number of time periods in the process;
M: index set of the time periods (M =]1, . . . , m]);
t: time period index (t M);
x
i]
: number of category i patients on oor ];
x
i0
: number of category i admitted patients not yet on any
oor;
X: matrix of x
i]
values;
S: state of the process (S :=|X, t]).
Random Variables
g
it
: number of category i patient arrivals during period t;
J
i]t
: number of category i patient departures from oor ]
during period t;
G: matrix of g
it
variables;
D: array of J
i]t
variables.
Decision Variables
,
i0]
: number of category i nonassigned patients to be
assigned to oor ];
,
i]k
: number of category i patients to be transferred
from oor ] to oor k;
Y: array of decision variables;
(S): set of feasible decisions for a given state S;
C(Y): total cost associated with decision Y.
Objective Values
V
n
(S): minimum expected n-stage cost in state S;

V
n
(S): approximation based on simulation;

V
n
(S): approximation based on simulation and simplied
state space.
3.2. Process State
We consider a decision process with a nite time horizon
divided into m time periods of constant length. The tran-
sition probabilities of the embedded Markov chain might
vary from time period to time period, but they cycle with
respect to m. That is, the transition probability from state A
to state B is the same during periods t, t m, t 2m,
etc. Hence, we have a stationary MDP where the period
index t is part of the state denition. To capture the intra-
day, intraweek, and seasonal uctuations in patient arrival
and departure rates, we must allow for this m-period time
frame to be very large, perhaps spanning an entire year.
The state of the process then depends on the number of
patients on the oors, the number of patients admitted but
not yet on any oor, and the time. Thus, the state of the
process is S :=|X, t].
3.3. Constraints on Decision Variables
As indicated earlier, there will generally be restrictions on
the oors to which different categories of patients can be
assigned. Among feasible oors it is generally more desir-
able to assign a patient of a given category to one oor
than another. The preference ranking of multiple feasible
oors for a given patient category is achieved by setting
assignment costs, a topic that is discussed in 3.4. All oors
other than the unique lowest cost (ideal) oor are called
alternate.
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
Operations Research 57(2), pp. 261273, 2009 INFORMS 265
For a state S = |X, t], decision variables in Y must
satisfy:
Patients Waiting:
_
]E
,
i0]
x
i0
, i Q. (1)
Floor Capacity:
_
iQ
_
x
ik
,
i0k

_
]E
,
i]k

_
]E
,
ik]
_
c
k
, k E . (2)
Nonnegativity:
,
i0k
0, ,
i]k
0, i Q, ], k E . (3)
Patient Restrictions:
,
i0k
=0, ,
i]k
=0, i Q, ] E , k E E
i
. (4)
The set of decision rules, formalized in 4.3, satisfying
(1)(4) for a state S, are called feasible, and are denoted
by (S).
3.4. Stage Cost
Once a feasible decision Y has been made at the beginning
of a period, there are a number of resulting expected costs
and rewards. First, there is an expected positive reward o
i]
to the hospital of assigning a category i patient to oor ].
For this work, o
i]
is the expected nancial gain from treat-
ing a patient of category i scaled to reect the desirability
of oor ]. The nonnegative parameter l
i]k
represents the
intrinsic undesirability of the transfer of a category i patient
from oor ] to oor k. Then, the stage cost associated
with Y is
C(Y) :=
_
iQ
_
kE
_
o
ik
,
i0k

_
]E
l
i]k
,
i]k
_
. (5)
3.5. Transition Probabilities
For every patient category, there are two possible random
eventsnew arrivals and departures from oors. Both event
types depend on the time period. It is natural to assume
that entries of G := |g
it
] are statistically independent of
each other. Similarly, the entries of D :=|J
i]t
] are assumed
statistically independent of each other and of the entries
of G, but are clearly dependent on the state matrix X.
Let S :=|X, t] be the current state, Y (S) the chosen
decision array, and

S := |

X, t] the state after arrivals and

departures occur. State

S is updated as follows:
x
i0
=x
i0
g
it

_
kE
,
i0k
, i Q, t M, (6)
x
i]
=x
i]
J
i]t
,
i0]

_
kE
(,
ik]
,
i]k
),
i Q, ] E , t M, (7)
t =
_
_
_
1 if t =m,
t 1 if t -m.
(8)
The transition probability from S to

S, given deci-
sion Y, is
P
S

S
(Y) =

iQ

_
g
it
= x
i0
x
i0

_
kE
,
i0k
_

iQ

]E

_
J
i]t
= x
i]
x
i]
,
i0]

_
kE
(,
ik]
,
i]k
)
_
.
(9)
Here the values of g
it
and J
i]t
in (9) are the possible
numbers of arrivals and departures that respectively must
have occurred to permit the transition from the state S to
the state

S, given that the decision Y has been made.
3.6. Objective Function
Our objective is to nd a decision rule that minimizes the
sum of the immediate cost and the subsequent expected
costs that result from the future evolution of the process.
Let V
n
(S) be the minimum expected cost for an n-stage
problem (evolving for n time periods) that starts in state S.
We have
V
1
(S) = min
Y(S)
C(Y), (10)
V
n
(S)= min
Y(S)
_
C(Y)
_

S
P
S

S
(Y)V
n1
(

S)
_
, n>1. (11)
With respect to the complexity of solving V
n
(S) for a
given state S = |X, t], note that for n = 1 the computa-
tion of V
1
(S) reduces to solving a single-commodity mini-
mum cost ow problem without directed cycles of negative
cost and can be readily solved in polynomial time on the
sizes of Q and E as a linear program subject to con-
straints (1)(4). On the other hand, the complexity of com-
puting V
n
(S) for n >1 depends on the size of the decision
space (S) and the number of states

S in (11) which in
some cases can be innite or nite but with exponential
growth in the size of Q and the number of stages. Hence, as
n grows, the exact computation of V
n
(S) becomes impracti-
cal for traditional methods such as dynamic programming.
In the online supplement, we elaborate on the complex-
ity issues. An electronic companion to this paper is avail-
able as part of the online version that can be found at
http://or.journal.informs.org/.
3.7. Approximation Methodology
To approximate V
n
(S), we use random sampling to estimate
the expectation term
_

S
P
S

S
(Y)V
n1
(

S).
Concretely, let (G
1
, D
1
), . . . , (G

, D

) be a sample of i.i.d.
random variables for a given state S. Then, we approximate
V
n
(S) using

V
n
(S), where

V
1
(S) := min
Y(S)
C(Y) =V
1
(S), (12)

V
n
(S) := min
Y(S)
_
C(Y)
1

_
i=1

V
n1
(

S
i
)
_
, n >1, (13)
and

S
1
, . . . ,

S

are the corresponding updated states.

Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
266 Operations Research 57(2), pp. 261273, 2009 INFORMS
Before we study the convergence of

V
n
(S), we intro-
duce some additional notation. Consider the following
denitions:
A
i
:=max
]E
]o
i]
], i Q, (14)
g
it
:=E|g
it
], i Q, t M, (15)
j
it
:=E|max]g
it
, g
it
]], i Q, t M, (16)
u
2
it
:=Var|max]g
it
, g
it
]], i Q, t M. (17)
Let
s(t) :=1 (t 1)mod m for all t M
denote the successor function modulo m. Let s
0
(t) := t,
t M, denote the identity function on M. For t > 0, we
introduce the notation
s
t
(t) =s(s
t1
(t)).
Using this notation, we state the following bounds on
V
n
(S). (Proofs are found in the online supplement.)
Proposition 1. The following bounds hold for all states
S =|X, t] and n 1:

_
iQ
A
i
x
i0

_
iQ
n2
_
t=0
A
i
g
is
t
(t)
V
n
(S) 0.
A similar argument to the proof of Proposition 1 can be
used to show analogous bounds for

V
n
(S) in the following
proposition.
Proposition 2. The following bounds hold for all states
S =|X, t] and n 1:

_
iQ
A
i
x
i0

_
iQ
n2
_
t=0

_
I=1
1

t1
A
i
g
I
is
t
(t)

V
n
(S) 0,
where g
I
is
t
(t)
, I =1, . . . , , are the realizations of the ran-
dom variables in each of the samples corresponding to each
period t, s(t), . . . , s
n2
(t), and each category i Q.
Using Proposition 1, we establish the following result
concerning the convergence of

V
n
(S) to V
n
(S) as the sam-
ple size increases.
Proposition 3. Let e > 0, 0 - o - 1, and S =|X, t] be a
given state. Then,
][V
n
(S)

V
n
(S)[ -e] 1 o (18)
for all

1
oe
2
__
_
iQ
A
i
x
i0

_
iQ
n1
_
t=1
A
i
g
is
t
(t)

_
iQ
A
i
j
it
_
2

_
iQ
A
2
i
u
2
it
_
. (19)
One important implication of Proposition 3 is that the
sample size needed to achieve good approximations is inde-
pendent of the decisions that are made. The result also
implies that

V
n
(S) converges in probability to V
n
(S) as
~ and so,

V
n
(S) is a consistent estimator of V
n
(S).
On the other hand, the term 1,o in the bound (19) that
originates from using Chebyshevs inequality in the proof
of Proposition 3 yields very large values of the sample size
to obtain reasonable approximations. The bound can be
improved by using the central limit theorem because for
large we can easily modify the proof of the proposition
to obtain an approximate bound

z
2
o,2
e
2
__
_
iQ
A
i
x
i0

_
iQ
n1
_
t=1
A
i
g
is
t
(t)

_
iQ
A
i
j
it
_
2

_
iQ
A
2
i
u
2
it
_
,
where z
o,2
satises the equation 4(z) = 1 o,2 and
4 is the distribution function of a standard normal random
variable.
4. DSS Specication
We provide details about how the solution algorithm
described above was incorporated into a DSS. In particu-
lar, we indicate how to simplify the decision space under
the conditions of a typical hospital, how to implement a
rolling-horizon algorithm to provide timely recommenda-
tions to the bed manager, and provide details on the compu-
tation of probabilities for the random variables considered
in the hospital model.
4.1. Probability Computation
The number g
it
of category i patient arrivals for a given
state S =|X, t] is assumed to have an aggregated Poisson
distribution with parameter \
it
. Note that \
it
is the average
number of category i patients who arrive during the time
interval between periods t and s(t), where as before s(t) =
1(t 1)mod m. To compute \
it
, we partition the interval
between periods t and s(t) into q subintervals. We assume
that the arrivals during each subinterval follow a Poisson
distribution with rate \
(])
it
for ] =1, . . . , q. Hence,
\
it
=
q
_
k=1
\
(])
it
.
The use of such an aggregated Poisson distribution for
patient arrivals ts the observed behavior at WCMH very
well.
For departure variables J
i]t
given state S, we assume that
each variable follows a binomial distribution with param-
eter pair (
i]t
, x
i]
). The parameter
i]t
denotes the proba-
bility of a category i patient departure from oor ] during
period t. To compute the parameter
i]t
, we use a method
similar to that used for the arrivals. We partition the interval
between periods t and s(t) into q subintervals. We denote
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
Operations Research 57(2), pp. 261273, 2009 INFORMS 267
by
(k)
i]t
the probability of one departure during the kth sub-
interval, and then we set

i]t
:=1
q

k=1
(1
(k)
i]t
).
All of the parameters are readily estimated from histori-
cal data and/or forecasting methods. Moreover, note that in
this case
g
it
=\
it
,
j
it
2\
it
,
u
2
it
\
it
(3\
it
1),
for all i Q and t M. Hence, based on bound (19) from
Proposition 3, we obtain the following corollary.
Corollary 4. Let e > 0, 0 - o - 1, and S =|X, t] be a
given state. Then,
][V
n
(S)

V
n
(S)[ -e] 1 o (20)
for all

1
oe
2
__
_
iQ
A
i
x
i0

_
iQ
n1
_
t=1
A
i
\
is
t
(t)
2
_
iQ
A
i
\
it
_
2

_
iQ
A
2
i
\
it
(3\
it
1)
_
. (21)
Bound (21) can be easily computed from the data and
problem parameters corresponding to the WCMH case. As
remarked before, bound (21) can be improved to

z
2
o,2
e
2
__
_
iQ
A
i
x
i0

_
iQ
n1
_
t=1
A
i
\
is
t
(t)
2
_
iQ
A
i
\
it
_
2

_
iQ
A
2
i
\
it
(3\
it
1)
_
for large . For example, using this bound with the
data provided by WCMH, we can show that the rela-
tive error between V
2
(S) and

V
2
(S) is no more than 34%
with approximately 99% probability for a sample size of
=500. Also, for the same sample size, the relative error
between V
2
(S) and

V
2
(S) is no more than 17% with
approximately 80% probability.
4.2. Real-Time DSS
The real-time DSS operates by solving the approximation

V
2
(S) to the dynamic programming formulation (10)(11)
on a rolling-horizon basis, according to the following
algorithm. Given state S and sample size 1, we generate
a random sample (G
1
, D
1
), . . . , (G

, D

). Then, we choose
any Y

such that
Y

argmin
Y(S)
_
C(Y)
1

_
i=1

V
1
(

S
i
)
_
.
The decision is applied in real time to the current state S.
Next, the decision maker waits for an interval of time dur-
ing which realizations of G and D are observed. At the end
of the interval, the state is further updated by adding the
changes that occurred corresponding to (G, D) to yield a
new state

S. Then, the algorithm is reapplied to the updated
state

S.
In our computations, the interval between periods t and
s(t) is eight hours long. In computing the distributions for
G and D, we partition the eight-hour interval into q =32
subintervals of 15 minutes each. Using our real-time algo-
rithm every 15 minutes, a decision Y is computed and
applied. Time intervals of 15 minutes were chosen to strike
a balance between providing the DSS enough time to arrive
at a good recommendation, while avoiding long periods
of computational time that would introduce another bottle-
neck to the patient ow process. In addition, eight hours is
a sufciently long planning horizon to allow for a reason-
able probability of random events occurring, while short
enough to keep the process under control.
4.3. Further Simplication Using Decision Rules
In using

V
n
(S) to approximate V
n
(S), one advantage is that
the term

_
]=1

V
n1
(

S
]
)
is easier to compute when the number of scenarios

S
]
is
not very large. However, the computational complexity of
the algorithm will depend on the size of the decision space
(S). Using (1), there are at least

iQ
_
x
i0
[E E
i
[
[E E
i
[
_
feasible solutions to consider, which could be large even
for small x
i0
.
In a typical hospital like WCMH, the following can be
assumed to be true:
1. If oor ]

E
i
is ideal for category i patients, then o
i]

is signicantly larger than o

I]
for any category I patient
for which oor ]

is alternate.
2. Under the same conditions, o
i]
is also signicantly
larger than the cost l
I]

k
of transferring out a category I
patient for which oor ]

is alternate.
Under those assumptions, for each i Q there exists a
unique index ]
i
E
i
such that
A
i
=o
i]
i
,
where A
i
is as dened in (14). Floor ]
i
is the ideal oor
for category i patients. In addition, we can partition the
set of categories into subsets Q
]
for all ] E . Each subset
Q
]
corresponds to the categories for which oor ] is ideal;
that is,
Q
]
:=]i Q: ]
i
=]]. (22)
Using this partition, we are able to prove the following
statement.
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
268 Operations Research 57(2), pp. 261273, 2009 INFORMS
Proposition 5. Let o be such that 0 -o -1 and
0
]
:=(1 (1 o)
1,([E [)
)
1
max
tM
_
_
iQ
]
g
it
_
for all ] E . If state S =|X, t] is such that
c
]

_
iQ
x
i]

_
iQ
]
x
i0
(n 1)0
]
(23)
for all ] E , then with probability at least
(1 o)
(
n1
1),(1)
the approximation

V
n
(S) achieves its lower bound from
Proposition 2 and the greedy solution Y given by ,
i0]
i
:=
x
i0
, ,
i0]
=0 for ] ,=]
i
, and ,
i]k
=0 is optimal for

V
n
(S).
In other words, if all the oors have enough excess
capacity for a given state S (condition (23)), then it is opti-
mal to assign all patients currently waiting to their ideal
oors. The factor (1 (1 o)
1,([E [)
)
1
in the denition of
0
]
grows fast as o approaches zero. To improve the result,
note that in the DSS the variable

iQ
]
g
it
has Poisson
distributions with parameter

iQ
]
\
it
. When the Poisson
parameter is large (like in the WCMH case), the central
limit theorem can be used to approximate the Poisson dis-
tribution with a Normal distribution with mean and vari-
ance equal to the Poisson parameter. Hence, we can obtain
a similar result to Proposition 5 by using
0
]
=max
tM
_
_
iQ
]
g
it
z
p,2
_
_
iQ
]
g
it
_
1,2
_
, (24)
where p = 1 (1 o)
1,([E [)
. For example, for the case
n =2, 1o =0.99, and using a sample size of =500 we
have z
p,2
4. Hence, if the residual capacity of oor ] is
greater than the corresponding 0
]
for all ] E , the greedy
solution is optimal for

V
2
(S) with probability at least 0.99.
Note that the numbers 0
]
do not depend on the state S.
For a given oor, we use the term primary patients to
refer to those patients for whom the oor is ideal, while
all other patients for whom the oor is feasible are labeled
secondary. Hospital regulations establish that a number of
beds on each oor should be reserved for primary patients
(so-called crash beds).
Therefore, the decision space can be reduced to only
those decisions in (S) that obey the above rules. More
precisely, we dene a decision rule as a function ] assign-
ing one and only one decision ] (R) := Y (S) for a
given rule-parameter R, where R := |r
]I
] is an [E [ 3
matrix such that r
]1
r
]2
r
]3
for all ] E . Ideally, we
would choose the r
]3
values to be the 0
]
values from Propo-
sition 5 or from (24), but in practice those values are too
large or imprecise. Hence, we let the r
]3
vary as part of a
decision rule. The process of assigning a decision Y for a
given matrix R is done in three consecutive steps: proactive
transfer; primary assignment; and secondary assignment.
In proactive transfer, we transfer secondary patients from
a oor ] in which there are fewer than r
]2
beds available to
another oor in which there are more than r
]3
beds avail-
able. When there is more than one transfer oor option, the
oor with the lowest transferring cost is chosen. The objec-
tive is to increase the capacity in each oor for potential
or currently waiting primary patients. In primary assign-
ment, patients currently waiting are assigned to their cor-
responding ideal oors as long as oor capacities permit.
If there are oor conicts among primary patients, those
with the highest assignment reward have priority. Finally,
in secondary assignment, all remaining waiting patients
are assigned to oors with at least r
]1
beds of remaining
capacity, where we again resolve oor-conict assignments
according to the highest assignment reward. The details of
the assignment algorithms for each of the three steps are
given in the online supplement.
When r
]2
=0, no proactive transfers occur out of oor ].
Further, when r
]1
= 0, no bed reservation occurs for pri-
mary patients on oor ]. The current decision rule used at
WCMH corresponds to R = 0, that is, it includes neither
proactive transfers nor bed reservations.
In practice, we only consider a nite set (S) of rule-
parameter matrices and the following approximation:

V
2
(S) := min
R(S)
_
C(] (R))
1

_
]=1
V
1
(

S
]
)
_
.
It is not difcult to show that Proposition 3 still holds if
we replace

V by

V.
Accordingly, we obtain the following revised real-time
algorithm.
Algorithm 1. Given state S and sample size , execute
the following sequence:
min ~,
Generate a random sample (G
1
, D
1
), . . . , (G

, D

),
for all R (S) do
s 0,
for ] =1, . . . , do
Update S using ] (R) and (G
]
, D
]
) to yield

S
]
,
s s V
1
(

S
]
),
end for,
if C(] (R)) s, -min then
min C(] (R)) s,,

RR,
end if,
end for,
return

R.
4.4. Approximation Algorithm Performance
To compare the performance of the solutions obtained
from our approximation algorithm to the optimal solutions,
we designed a simulation in a simplied hospital setting.
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
Operations Research 57(2), pp. 261273, 2009 INFORMS 269
Because obtaining optimal solutions for problem instances
of a realistic size is very time-consuming, we were restricted
to exploring relatively small problem instances. In this case,
we consider a hospital dened by the following parameters:
E = ]1, 2, 3]; Q = ]1, 2, 3]; E
1
= ]1, 2, 3]; E
2
= ]1, 2, 3];
E
3
=]3]; c
]
=3, ] E ; l
i]k
=200, i Q, k E
i
. We set the
assignment rewards to
|o
i]
] =
_
_
_
_
1,000 900 900
900 1,000 900
1,000
_

_
.
We assumed arrival rates of three, two, and one patients
per day for the Poisson processes associated with each
of the three categories of patients, respectively. Hence,
because a 24-hour day has 96 15-minute intervals, we use
\
1t
= 3,96, \
2t
= 2,96, and \
3t
= 1,96 for all t. Finally,
we assumed that the length-of-stay for all patient categories
has a Poisson distribution with average of one day per
patient. Therefore,
i]t
=1 exp(1,96) for all i, ], t.
We randomly generated 30 initial states and 46 decision
rule matrices R with various degrees of proactivity lev-
els. The details of these decision rules are provided in the
online supplement. For each initial state S, we computed
optimal values V
2
(S), V
3
(S), and V
4
(S) and correspond-
ing optimal decision rules by using dynamic program-
ming. We also computed

V
2
(S),

V
3
(S), and

V
4
(S) and
corresponding decision rules using our heuristic. Table 1
shows the average relative error between V
n
(S) and

V
n
(S),
as well as the percentage of times that the rule chosen by
the heuristic coincided with the optimal decision rule for
n =2, 3, 4. In addition, we compared the performance of
the prior rule used by the hospital to the optimal solution
for each of the states in our sample. We found that the
approximation method signicantly outperforms the prior
WCMH rule as the number of stages increases. Note that
for two stages, the prior WCMH rule was always optimal
because it involved only one decision period, so a myopic
approach is optimal by denition. It is also interesting to
note that the heuristic performs better as n increases. This
pattern is due to the fact that for relatively low values
of n, the predictive ability of the heuristic is very robust.
We expect that at some value of n that trend will not persist.
However, because the exponential growth of the state space
makes computing optimal solutions prohibitive for n 5,
Table 1. Comparing performance.
Relative Worst-case
error (%) error (%) Optimal (%)
Number of Prior Prior Prior
stages (n) Heuristic rule Heuristic rule Heuristic rule
2 5.5 0.0 34.9 0.0 53.3 100.0
3 1.1 10.0 13.6 26.9 86.7 53.3
4 0.6 14.2 10.4 40.1 90.0 40.0
it is not possible to determine that cutoff point. For exam-
ple, solving for the optimal policy took 30 minutes for the
four-stage problem but it took 74 hours to analyze a single
decision rule for the ve-stage problem. That extrapolates
into 370 days of simulation time to evaluate the ve-stage
problem. By contrast, our heuristic is able to determine a
recommended action for problems of this size in less than
ve seconds and for realistic problems in less than two
minutes.
5. Computational Experiments
We conducted several computational experiments to ne-
tune some of the parameters eventually used in the imple-
mentation of the DSS at WCMH.
5.1. Tractable Decision Rule Space Determination
Because the number of potential decision rules is very
large, we conducted more computational experiments to
identify a tractable number of decision rules to be used in
the WCMH DSS. The challenge was to identify a small
set of decision rules that yield good performance results.
A total of 46 decision rule matrices R were considered, and
each simulation depicted a hypothetical one-month period
for WCMH (=96 30 = 2,880 time periods). Based on
historical data on arrival rates for patients of each category,
we simulated a total of 455 patient arrivals. We computed
14 simulation runs, all starting from the same initial state.
Details of the simulation design are provided in the online
supplement.
5.1.1. Simulation Results. For each simulation run,
we used the rolling-horizon Algorithm 1, computing

V
2,880
(S) in each iteration, and updating states according
to the decision rule

R returned by the algorithm and the
simulated arrivals and departures of patients.
The following variables were recorded:
1. number of transfers;
2. number of last-minute transfers (reactive transfers);
3. number of patients assigned to their ideal oor;
4. number of patients assigned to an alternate oor;
5. number of times each decision rule was invoked;
6. number of periods waiting before assignment.
The objective of this analysis was to successively add
decision rules and identify a point of diminishing returns
beyond which the addition of decision rules resulted in little
or no improvement in results.
Table 2 shows that the initial addition of decision rules
makes a dramatic difference in reducing the number of
reactive transfers. However, beyond four decision rules, the
incremental improvements diminish. In other words, a rel-
atively small set of decision rules can achieve performance
similar to that of a larger set. By studying the most fre-
quently used decision rules, we learned that as long as a
varied set was used, we obtain good results.
Figure 1 shows the average delay before making a
bed assignment for different levels of capacity utilization.
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
270 Operations Research 57(2), pp. 261273, 2009 INFORMS
Table 2. Simulation results.
Simulation Decision Total Reactive Assigned to Assigned to
run rules transfers transfers ideal an alternate
1 1 73 73 402 53
2 2 80 51 382 73
3 4 33 12 363 92
4 6 33 12 363 92
5 10 34 12 364 91
6 14 32 11 367 88
7 18 31 11 367 88
8 22 30 11 367 88
9 26 30 10 367 88
10 30 31 10 369 86
11 34 29 9 369 86
12 38 29 10 370 85
13 42 28 9 371 84
14 46 28 8 371 84
In Figure 1, we compare simulation runs 1, 2, 3, and 14
with a passive strategy in which no assigned patient is
transferred between oors to make room for newly arrived
patients. Simulation runs 413 were omitted because the
similarity in performance resulted in a very cluttered graph.
Figure 1 shows that the reactive decision rule used by
WCMH results in signicantly better performance than the
passive decision rule. However, the large number of last-
minute transfers, most of which occur at higher levels of
capacity utilization, result in prolonged waiting times dur-
ing critical crunch periods. This is illustrated in Table 2,
which shows that all 73 of the transfers were reactive
transfers. By contrast, the inclusion of additional decision
rules in runs 214 shows an important pattern in reduced
delays prior to making a bed assignment and in the num-
ber of reactive transfers. In both cases, there is a dramatic
improvement in performance that quickly levels off. The
last signicant performance improvement is found at Run 3,
where four decision rules are included. For Run 3, the time
delay before a bed assignment can be made remains stable
Figure 1. Waiting time as a function of occupancy.
Occupancy rate (%)
T
i
m
e

u
n
t
i
l

b
e
d

a
s
s
i
g
n
m
e
n
t
75 77 79 81 83 85 87 89 91 93 95 97 99
0:00
0:30
1:00
1:30
2:00
2:30
3:00
Passive
Run 1
Run 2
Run 3
Run 14
Table 3. Patient categories at WCMH.
Avg. length
Category Description of stay o
i
($)
1 Short stay general medicine 2.51 7,464
2 Long stay general medicine 4.86 11,616
3 Short stay general surgery 2.54 11,075
4 Long stay general surgery 8.42 27,653
5 Pediatrics 2.24 4,396
6 Short stay telemetry 1.94 8,268
7 Long stay telemetry 4.71 14,095
8 Clean general surgery 3.81 24,381
9 Short stay critical care 4.47 12,716
10 Long stay critical care 16.29 61,323
11 Surgical gynecology 3.05 11,894
12 Obstetrics/special gynecology 2.42 4,639
and low, even when capacity utilization is high. In addi-
tion, the number of reactive transfers is dramatically less
than the number incurred using the hospitals prior rule
(12 compared to 73). Run 14, not surprisingly, yields the
best results in terms of time to oor assignment and number
of reactive transfers. However, the incremental improve-
ment is small and, more signicantly, the amount of time
needed to analyze all 46 decision rules is prohibitively long.
6. DSS at WCMH
Based on the simulation conducted in 5, the DSS was
encoded with the decision rules depicted in simulation
Run 3 (see the online supplement). Table 3 shows the
patient categories that were identied at WCMH based on
oor assignment and length of stay for the different patient
populations the hospital serves. Table 3 also shows the
expected reimbursement for each category, again based on
historical averages.
Prior to implementing the DSS, hospital administrators
were asked to participate in an exercise to determine the
appropriate values of the o
i]
and l
i]k
. The values for the o
i]
were initially set to the expected reimbursement of a patient
of category i depicted in Table 3, and then scaled down to
reect the desirability of assigning a patient of that category
to the different oors. For example, o
1, 1
=7,464 because
Floor 1 is the oor that specializes in cardiology and gen-
eral medicine. Floor 2 specializes in general surgery and
pediatrics (although the staff are still capable of providing
health care service for a general medical patient) and was
therefore considered an alternate oor with o
1, 2
=7,164.
Determining the values of the l
i]k
was also done interac-
tively. Because determining the values of the l
i]k
does not
require iterating through multiple time periods, it was pos-
sible to represent one instance of (10)(11) in a Microsoft
Excel spreadsheet which was then solved using the stan-
dard Excel Solver engine. The initial values of all l
i]k
were
set to 100. Then, a baseline hypothetical state was created
that was similar to an average day in terms of census,
stafng, and patient mix at WCMH. Nursing administrators
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
Operations Research 57(2), pp. 261273, 2009 INFORMS 271
were then asked to evaluate the patient assignment/transfer
decisions that were obtained by solving the patient allo-
cation problem in response to different events. The values
of the l
i]k
were modied, based on feedback from nurs-
ing administrators, until decisions consistent with hospi-
tal strategy were obtained. The process of determining the
appropriate values for the l
i]k
took approximately one hour
and involved evaluating 100 different events and associated
responses. Microsoft Excel proved to be a valuable tool
because the graphical layout improved the ability of the
nursing administrators to understand and critique the events
and associated responses. A sensitivity analysis regarding
the impact of cost parameters on DSS recommendations is
presented in the online supplement.
The DSS was programmed using version 1.4.2 of the
Java 2 Platform, Standard Edition (J2SE) and all optimiza-
tion problems were solved using lp_solve v5.1. The appli-
cation was run on a desktop PC with a 3 GHz Pentium 4
processor and 512 MB of RAM.
The 18-day trial period began January 17, 2005. The
DSS included the decision rule set identied in the simu-
lation described in 5. During the trial period, a total of
292 patients were admitted, an average of 16.22 admis-
sions per day, which is roughly 10% higher than the
14.79 admissions per day average obtained from histori-
cal data. This increase in the number of admissions per
day was associated with a greater than average number of
medical patients, and is consistent with expected seasonal
uctuations.
In total, 83% of the actions recommended by the DSS
(269 out of 324) were followed. The DSS allowed the
user to enter the reason why a given recommendation was
not taken. This information was provided for 47 of the 55
recommendations that were not taken. A total of 20 rec-
ommendations were not followed because, based on some
preliminary information on patients currently being evalu-
ated in the ED, the bed manager was concerned that fol-
lowing the recommendation would complicate the ability
of the hospital to take an admission that was imminent
but not ofcial in the sense that no admission orders
had been written. Concerns related to future stafng were
the cause of rejecting 11 recommendations and represented
cases where the bed manager was hesitant to bring a oor
to full capacity, when the ability to staff the oor properly
1624 hours in the future was in question. The remain-
ing 16 instances were related to: gender compatibility (4);
staff too busy (6); and a range of clinical concerns (6).
We measured the impact on the number of last-minute
transfers using the DSS, and also the actual time needed
to transfer each patient admitted from the ED to the oor.
To determine the impact on last-minute transfers, a real-
time simulation was conducted using the hospitals pre-
vious, reactive assignment decision rule, and the status
updates obtained during the trial period. This enabled us
to compare the number of last-minute transfers and pre-
emptive transfers occurring during the trial period to the
Table 4. Impact on patient transfers.
Type of transfer DSS Prior rule
Preemptive 21 0
Last minute 7 69
Total 28 69
number that would have resulted during the trial period had
the hospital used its prior decision rule. Table 4 shows the
results (note that under the prior decision rule, all in-house
transfers were last-minute transfers).
Overall, the total number of in-house transfers was less
using the DSS (28 versus 69 using the prior rule), and the
reduction in the number of last-minute transfers was even
more pronounced (7 versus 69 using the prior rule). The
signicant reduction in the number of last-minute transfers
came at a cost. The number of patients assigned to alternate
oors by the DSS was higher than under the hospitals prior
assignment rule.
In terms of the number of patients assigned to alternate
oors during the trial period, a total of 105 of this type
of assignments were made using the DSS, as opposed to
only 56 using the prior rule. These results together with
the results from Table 4 show that ensuring that beds were
available when needed was achieved primarily through
the assignment of patients to alternate oors as opposed
to transfers. This is a signicant positive nding because
in-house transfers, even if done preemptively, are time-
consuming and disrupt the continuity of care for the patient
that is transferred.
Baseline data was obtained from the hospital to deter-
mine whether the reduction in last-minute transfers was
associated with the expected reduction in the average wait-
ing time before transferring patients from the ED to the
oors. The hospital provided data on waiting time in the
ED collected at different times over the last half of 2004.
The baseline data only contained observations made when
beds were available in the hospital. The important implica-
tion is that these patients were not being delayed because
no beds were available in the hospital (which does hap-
pen), but because the available beds were not on the oors
to which these patients could be assigned. This baseline
data showed very little monthly uctuation. The Anderson-
Darling test (Stephens 1974) was performed on the baseline
transfer time data provided by the hospital. The results,
with test statistic A
2
=0.31, did not refute the null hypoth-
esis that the data come from a lognormal distribution. The
observed decrease in transfer time during the trial period,
depicted in Figure 2, was signicant at o =0.001.
The impact of the number of available beds on wait-
ing times was also studied. The reason for this analysis is
that presumably when each oor has several beds avail-
able, achieving short waits should not be difcult. How-
ever, as the number of available beds decreases, following
decision rules that ensure that remaining beds are allocated
in a manner that reects expected arrival and discharge
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
272 Operations Research 57(2), pp. 261273, 2009 INFORMS
Figure 2. Impact on transfer time.
Mean Median
E
l
a
p
s
e
d

t
i
m
e
0:00
0:30
1:00
1:30
2:00
2:30
2:17
1:19
2:09
1:15
Baseline Trial
patterns, should result in shorter waiting times than fol-
lowing decision rules that do not. The results depicted in
Figure 2 show that aggressive assignment and reallocation
strategies, represented by some of the rules in the decision
rule set, enable the hospital to achieve short waiting times
even when the number of available beds is low. In addition,
inspection of the daily census from each day in the data set
that provided the baseline waiting time of two hours and
17 minutes shows that the total number of available beds
was at least seven.
Figure 3 plots the average and median waiting times
based on the number of available beds at the time admis-
sion orders were written. The trend line shows that, on
average, the waiting time is slightly over one hour. When
the number of available beds decreases to less than four,
the waiting times begin to increase. This suggests that the
approach developed here is effective, even when remaining
capacity is low.
Figure 3. Wait time and bed availability.
Available medical/surgical beds
T
i
m
e
0 1 2 3 4 5 6 7 8 9 10 11 12 13
0:00
0:30
1:00
1:30
2:00
2:30
3:00
3:30
4:00
4:30
Average
Median
6.1. Impact on the Hospital
The reduction in waiting time, almost one hour per patient
on average, has signicant quality of care and nancial
repercussions. The impact on quality of care is obvious.
Admitted patients are moved more quickly to a oor, and
the process of carrying out the admission orders begins
in a more timely fashion. For patients waiting to be seen
by a physician in the ED, the fact that admitted patients
are vacating more quickly means they will have shorter
waits. For WCMH, which averages 280 ED admissions per
month, a one-hour reduction for each of these admissions
translates into a gain of approximately 3,360 additional
bed-hours in the ED per year. Each bed in the ED averages
$258/hour in charges, so if these additional hours are used
to see additional patients, the expected benet to WCMH
is $866,880 per year. During the trial implementation, the
additional bed-hours in the ED were used to see additional
patients, and according to James Papadakos, Chief Finan-
cial Ofcer at WCMH, the anticipated increase in the num-
ber of patients treated and in revenues were observed.
Assigning patients based on the recommendations pro-
vided by the DSS also improved capacity utilization.
Although not as dramatic as the reduction in waiting time,
the resulting improvement has signicant quality of care
and nancial implications. During the 18-day study, by fol-
lowing the recommendations provided by the DSS, WCMH
was able to accept four patients totaling 11 bed-days that
otherwise would have been diverted to a different hospi-
tal. This means the hospital was able to provide services
to more patients in the community as a result of divert-
ing fewer ambulances. This is an important accomplish-
ment because diverting ambulances results in a delay in
the provision of emergency care which is clearly undesir-
able. From a nancial standpoint, based on WCMH data,
the hospital expects to generate $3,043 in charges per
bed-day. An increase of 11 bed-days corresponds to an
expected increase of $33,473 in charges and translates into
an annualized increase of $678,758 in charges. Because
WCMH only collects, on average, 39% of the amount
billed, the total expected increase in charges of $1,545,638
corresponds to an expected increase of $602,798 in actual
revenue. This reects a 1% increase in revenue from oper-
ations, which is not trivial in the not-for-prot health care
setting where operating margins of 1%2% are considered
a success. Furthermore, this increase in revenue is achieved
using existing physical and staff resources, so that very lit-
tle additional cost is incurred. As a result, a large portion
of the $602,798 in additional revenue is retained.
Finally, WCMH, like all hospitals, routinely tracks
patient safety and stafng effectiveness indicators. One of
these indicators, the amount of time needed to transfer a
patient from the ED to the oor, has been the primary focus
of this paper. The other indicators include fall rates, med-
ication errors, hospital-acquired infection rates, restraint
usage, length-of-stay, and patient satisfaction surveys that
are sent to patients postdischarge. A very important nding
Thompson et al.: Efcient Short-Term Allocation and Reallocation of Patients During Demand Surges
Operations Research 57(2), pp. 261273, 2009 INFORMS 273
of a key metric for WCMH was that there was no sta-
tistically signicant change in any of these other indica-
tors. This offers further conrmation that quality of care
for admitted patients is not being compromised by taking
the actions recommended by the DSS. The DSS implemen-
tation was considered a success by WCMH and the hos-
pital has opted to continue using the system as a tool to
aid in patient assignment decisions. The hospital has also
decided to create an operations manager position, an indi-
vidual who will work with the DSS and work to identify
other opportunities to improve patient ow and hospital
efciency. WCMH has continued to use the DSS to aid in
patient allocation decisions. Furthermore, the effectiveness
of optimality-based resource utilization tools has prompted
WCMH to initiate work with us on tools to improve the
utilization of their cardiac diagnostic facilities.
7. Electronic Companion
An electronic companion to this paper is available as part
of the online version that can be found at http://or.journal.
informs.org/.
Acknowledgments
The authors gratefully acknowledge Allison Breault, Vice
President for Patient Care Services at Windham Hospi-
tal, for her time, support, and invaluable contributions to
this research. In addition, the authors thank the anonymous
associate editor and referees for their helpful, construc-
tive comments that greatly improved the quality of this
manuscript.
References
Albareda, M., E. Fernandez. 2000. The stochastic generalized assignment
problem with Bernoulli demands. TOP 8(2) 165190.
Bayley, M. D., J. S. Schwartz, F. S. Shofer. 2005. The nancial burden
of emergency department congestion and hospital crowding for chest
pain patients awaiting admission. Ann. Emergency Medicine 45(2)
110117.
Brandeau, M. L., F. Sainfort, W. P. Pierskalla, eds. 2004. Operations
Research and Health Care: A Handbook of Methods and Applica-
tions. International Series in Operations Research and Management
Science. Springer, New York.
Caissie, L. 2004. Fax patient reports to save time. Patient Flow Weekly
1(3). http://www.hcpro.com.
Cox, T. F., J. F. Birchall, H. Wong. 1985. Optimising the queuing system
for an ear, nose and throat outpatient clinic. J. Appl. Statist. 12(1)
113126.
Derlet, R. W., J. R. Richards. 2000. Overcrowding in the nations emer-
gency departments: Complex causes and disturbing effects. Ann.
Emergency Medicine 35(1) 6368.
Dodge, G. 2001. Californias emergency services and trauma care system.
Testimony to Assembly Committee on Health by Emergency Nurses
Association, California State Council, Des Plaines, IL, February 13.
Eckstein, M., L. S. Chan. 2004. The effect of emergency department
crowding on paramedic ambulance availability. Ann. Emergency
Medicine 43(1) 100105.
Flagle, C. D. 2002. Some origins of operations research in healthcare.
Oper. Res. 50(1) 5260.
General Accounting Ofce (GAO). 2003. Hospital emergency depart-
ments: Crowded conditions vary among hospitals and communi-
ties. United States General Accounting Ofce, GAO-03460, March.
http://www.gao.gov.
Gerchak, Y., D. Gupta, M. Henig. 1996. Reservation planning for elective
surgery under uncertain demand for emergency surgery. Management
Sci. 42(3) 321334.
Green, L., P. Kolesar. 2004. Improving emergency responsiveness with
management science. Management Sci. 50(8) 10011014.
Green, L., V. Nguyen. 2001. Strategies for cutting hospital beds: The
impact on patient service. Health Services Res. 36(2) 421442.
Green, L., S. Savin, B. Wang. 2006. Managing patient service in a diag-
nostic medical facility. Oper. Res. 54(1) 1125.
Jackson, R. R. P., J. D. Welch, J. Fry. 1964. Appointment systems in
hospitals and general practice. Oper. Res. Quart. 15(3) 219232.
Kellerman, A. 2001. Emergency care in California: No emergency?
Health Affairs, Web Exclusives Supplement, March. http://content.
healthaffairs.org/webexclusives.
Krein, S., M. Casey. 1998. Research on managed care organizations in
rural communities. J. Rural Health 14(3) 180199.
Litvak, E., M. C. Long, A. Cooper, M. L. McManus. 2001. Emer-
gency room diversion: Causes and solutions. Academic Emergency
Medicine 8(11) 11081110.
Mine, H., M. Fukushima, K. Ishikawa, I. Sawa. 1983. An algorithm for
the assignment problem with side constraints. Memoirs of the Faculty
of Engineering, XLV, Part 4. Kyoto University, Kyoto, Japan.
Parker, P. 2004. Move care to a higher level with emergency systems.
Nursing Management 35(9) 8286.
Schull, M. J., M. Vermeulen, G. Slaughter. 2004. Emergency department
crowding and thrombolysis delays in acute myocardial infarction.
Ann. Emergency Medicine 44(6) 577585.
Schull, M. J., K. Lazier, M. Vermeulen, S. Mawhinney, L. J. Morrison.
2003. Emergency department contributors to ambulance diversion:
A quantitative analysis. Ann. Emergency Medicine 41(4) 467476.
Solberg, L., B. Asplin, R. Weinick, D. Magid. 2003. Emergency depart-
ment crowding: Consensus development of potential measures. Ann.
Emergency Medicine 42(6) 824834.
Stephens, M. A. 1974. EDF statistics for goodness of t and some com-
parisons. J. Amer. Statist. Assoc. 69(347) 730737.
Tucker, J. B., J. E. Barone, J. Cecere, R. G. Blabey, R. Chan-Kook. 1999.
Using queueing theory to determine operating room stafng needs.
J. Trauma, Injury, Infection, and Critical Care 46(1) 7179.
Wilson, M. E. 2001. Overcrowding and other crises: How can we survive?
J. Emergency Nursing 27(3) 225227.