Вы находитесь на странице: 1из 11

European Journal of Operational Research 206 (2010) 642–652

Contents lists available at ScienceDirect

European Journal of Operational Research


journal homepage: www.elsevier.com/locate/ejor

Innovative Applications of O.R.

Scheduling elective surgery under uncertainty and downstream capacity constraints


Daiki Min, Yuehwern Yih *
School of Industrial Engineering, Purdue University, 315 N. Grant Street, West Lafayette, IN 47907-2023, United States

a r t i c l e i n f o a b s t r a c t

Article history: The objective of this study is to generate an optimal surgery schedule of elective surgery patients with
Received 29 October 2008 uncertain surgery operations, which includes uncertainty in surgery durations and the availability of
Accepted 9 March 2010 downstream resources such as surgical intensive care unit (SICU) over multi-periods. The stochastic opti-
Available online 20 March 2010
mization is adapted and the sample average approximation (SAA) method is proposed for obtaining an
optimal surgery schedule with respect to minimizing the total cost of patient costs and overtime costs.
Keywords: A computational experiment is presented to evaluate the performance of the proposed method.
Surgery scheduling problem
Ó 2010 Elsevier B.V. All rights reserved.
Downstream resource constraint
Stochastic programming

1. Introduction

Surgery is one of the most important functions in hospitals and it generates revenue and admissions to hospitals. The operating cost of a
surgery department is the one of the largest hospital cost category, approximately one-third of the total cost (Macario et al., 1995). Surgery
is thus the area with the highest potential for cost savings. While surgery is the largest cost center, it also accounts for approximately two-
third of hospital revenues (Jackson, 2002). Therefore, small improvements in efficiency could translate into significant savings and benefits
to the patient as well as the hospital. For these reasons, managing the surgical resources effectively in order to reduce costs and increase
revenues is one of areas that draw considerable attention from the healthcare community.
The problem of modeling and optimizing surgery operations has been documented in the literature, which can be categorized into
capacity planning problem, block scheduling problem, surgery scheduling problem and surgery sequencing problem (Magerlein and Mar-
tin, 1978; Ogulata and Erol, 2003; Testi et al., 2007). This study focuses on scheduling elective surgery patients over a planning horizon. The
decision of scheduling elective surgery patients is to determine whether an elective patient should be scheduled and, if so, to determine
when the patient should be scheduled. There are two challenges in this problem: capacity constraints of downstream resources such as
surgical intensive care unit (SICU) beds or ward beds and the uncertainty in surgery operations.
The elective surgery schedule will attempt to admit as many patients as possible while satisfying resource constraints (e.g. minimizing
overtime works) in order to maximize the quality of care (e.g. minimizing patient waiting time). With regard to resource constraints for
scheduling elective surgeries, the consideration of operating room (OR) capacity alone does not yield good schedules. Capacity shortage
of downstream resources will keep patients from moving forward and it will significantly deteriorate OR utilization. For example, when
there are not enough SICU beds to accept all incoming patients, some patients have to remain in OR or should find other compatible re-
sources with additional costs. Jonnalagadda et al. (2005) show that 15% of the total cancellation is caused by the lack of an available recov-
ery room bed in the hospital they investigated. Sobolev et al. (2005) also show that patients’ length of stay (LOS) in intensive care unit (ICU)
and the ICU availability affect a surgery schedule. Therefore, it is important to consider downstream resource availability in addition to OR
capacity.
The significance of downstream resources has been considered in the literature. Hsu et al. (2003) investigate a surgery sequencing prob-
lem of which objective is to minimize the number of PACU (Post Anesthesia Care Unit) nurses in a single day setting. They show that, in
daily surgery operations, PACU is critical for determining the sequence and start time of a surgery. However, when scheduling over multi-
ple days, the capacity in SICU or wards should also be included to prevent congestion in the system. Similarly Gupta (2007) proposes a
dynamic programming approach to a problem of elective surgery booking control. Critical downstream resource is considered in his study,
but they do not model multi-period demand and the problem is left as an open research problem. Guinet and Chaabane (2003) investigate
an elective surgery scheduling problem where a patient follows the sequence of OR, recovery room (PACU), ICU and regular ward. Capacity

* Corresponding author. Tel.: +1 765 494 0826; fax: +1 765 494 1299.
E-mail address: yih@purdue.edu (Y. Yih).

0377-2217/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2010.03.014
D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652 643

of ICU beds is considered in their study, but they assume that a patient occupies an ICU bed for one day instead of multiple days. Practically,
LOS in ICU can be longer than a day. Most recently Pham and Klinkert (2008) and Fei et al. (2008) consider the surgical scheduling problem
in their study, but these papers are limited to deterministic operation and/or recovery times in PACU and SICU for all patients.
Scheduling surgery becomes challenging when considering the uncertainty in surgery operations. Surgery operations have case-depen-
dent durations and there is often a large variation between scheduled durations and actual durations. After surgery in an OR, LOS in a SICU
is also uncertain as well. Emergency surgery is another important source that introduces more uncertainty to the problem.
To address the issue of uncertainty, stochastic optimization recently starts to emerge in the surgery scheduling problems. An early work
is by Gerchak et al. (1996), which applies a stochastic dynamic programming method to generate an optimal policy that determines how
many elective surgeries to be admitted at the start of each day. Hans et al. (2008) introduce sufficient planned slacks to surgery durations
for hedging uncertain surgery durations. Finally a stochastic mixed integer programming model has been proposed for the surgery sched-
uling problem (Lamiri and Xie, 2006; Lamiri et al., 2008a,b). However, surgery durations of all elective cases are assumed to be known and
deterministic. Emergency demand is only the uncertain factor in their model. Uncertain surgery durations may lead the solution based on
deterministic durations to infeasible. Denton et al. (2009) formulate the surgery scheduling problem for assigning surgeries on a given day
of surgery as a two-stage stochastic linear programming. While considering different problem settings, Denton and Gupta (2003) and Den-
ton et al. (2007) propose stochastic programming models for a surgery sequencing problem within each day. L-shape method and sample
average approximation (SAA) algorithm are used for their models.
The objective of this paper is to investigate a stochastic surgery scheduling problem while considering downstream capacity constraints
(i.e. SICU beds). The contributions of this study are as follows: We formulate the problem for scheduling elective patients under SICU bed
constraints as a stochastic mixed integer programming model. All of surgery durations, LOS in SICU and new demand are assumed to be
random with known distributions. A sampling based approach (Sample Average Approximation; SAA) is applied to solve the proposed
model and numerical experiments show that the SAA provides a good solution within a reasonable computation time.
The rest of this paper is structured as follows: in Section 2, we present the proposed optimization model. Within the same section, the
structure of the model is investigated, aiming at validating the solution approach (i.e. SAA). In Section 3, the solution approach is intro-
duced with detailed description of its procedure. Numerical experiments are presented in Section 4. Moreover, a simulation study is con-
ducted to evaluate how the proposed stochastic model is significant in comparison with the corresponding deterministic model. Finally,
concluding remarks are described in Section 5.

2. The model for scheduling elective surgery

2.1. Problem description

Let I be a set of patients waiting surgery and let B be a set of available surgical blocks within an arbitrary planning horizon T. First, we
associate a binary variable xib to the admission decision. For each patient i 2 I and each surgical block b 2 B, let assignment variable

1 if a patient i 2 I is assigned to a surgical block b 2 B
xib ¼
0 otherwise:
A pseudo surgical block B0 2 B is defined with infinite block capacity and a day T 0 2 T is defined that corresponds to the pseudo surgical
block B0 . A patient whose scheduled block is the pseudo surgical block will not get surgery during the planning horizon. That is, xiB0 ¼ 1
means that a patient i is not admitted during the present planning horizon, and should be stay on the wait list. Here let CQ ib be the cost
when a patient i is assigned to a surgical block b. Thus, we can make a reasonable assumption that CQ ib < CQ iB0 for each patient i 2 I and for
all surgical blocks b 2 B n fB0 g. This assumption means patient waiting cost is strictly increasing in waiting time.
While a surgery is completed within few hours, a SICU bed can be occupied more than one day. So, we let yit be a binary variable that
represents a patient i 2 I occupying a SICU bed at day t 2 T.

1 if a patient i 2 I occupies a SICU bed at day t 2 T
yit ¼
0 otherwise:
The randomness in surgery operations is denoted by a secnario n. A scenario defines the vector of outcomes of three random variables;
surgery durations, LOS in SICU and block capacity. By the assumption of independence of random variables, a scenario has support
P
N 2 RjIj  ZjIj  ZjBj . Let /ðnÞ be the corresponding probability of scenario n 2 N, and n2N /ðnÞ ¼ 1.
For each patient i 2 I and each scenario n 2 N, the surgery duration random variable and LOS random variable are represented by W ni and
n
di . C nb represents the capacity of surgical block b 2 B under the scenario n 2 N. For each surgical block b 2 B n fB0 g; C nb is defined as an effec-
tive capacity, which is obtained by subtracting emergency demand or turnaround time between two consecutive surgeries (e.g. cleaning or
setup) from the planned block capacity. That is, the random emergency demands and turnaround time are presented implicitly. If a hospital
separates emergency demands from elective patients and ignores turnaround time, then C nb is deterministic. Here, for each scenario n 2 N,
the decision of yit depends on the randomness and is represented by ynit .
Patients can be assigned over a block capacity with overtimeP costs. For realizations
þ of surgery durations under scenario n 2 N, let onb be
n n n
the total overtime work of a surgical block b 2 B. Thus, ob ¼ i2I W i xib  C b , where xþ ¼ maxf0; xg. The total expected overtime cost is
formulated as
X X
/ðnÞCOb onb ; where COb is an unit overtime cost ðe:g: $=minuteÞ:
n2N b2BnfB0 g

Practically, surgery duration W ni depends on a surgery type. For explicit representation, the random variable W ni can be redefined as W nij ,
where j indexes J surgery types. However, this paper uses W ni instead of W nij for simplification, because the presentation of W nip is mean-
ingless in terms of problem complexity and the algorithm employed in this paper. It is assumed that W ni is identical for the same surgical
service group and implies a surgery type dependent duration.
644 D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652

It should be mentioned that the probability distribution of /ðnÞ is discrete, and the scenario set N is finite. While the actual surgery dura-
tion is continuous, the assumption of discrete durations is reasonable in the sense of scheduling, because a surgery is scheduled in a dis-
crete time interval (e.g. 30 minutes instead of 29.5 or 31.5 minutes). However, a decision of how to determine the time interval is made by
the trade-offs between solution quality and computational efficiency, which is beyond the scope of this paper. In this study, the time inter-
val is arbitrary decided according to the actual surgery data and numerical experiments show that an effective surgery schedule is obtained
within a reasonable time.
Finally, in order to give an explicit representation of a surgical block schedule, for each patient i 2 I, let BðiÞ be the set of surgical blocks in
which a patient i can be scheduled. A surgical block schedule, which is assumed to be given in this paper, is expressed as follows:

1 if a surgical block b 2 B is open and available for a patient i 2 I
zib ¼
0 otherwise:
Using the indices, sets, parameters and variables summarized in Table 1, the following mathematical model (1)–(9) describes the stochastic
surgery scheduling problem (SSSP).
XX X X
ðSSSPÞ Min CQ ib xib þ /ðnÞCOb onb ; ð1Þ
i2I b2B n2N b2BnfB0 g
X
xib ¼ 1 8i 2 I; ð2Þ
b2B

xib 6 Z ib 8i 2 I; 8b 2 B n fB0 g; ð3Þ


X n
onb P W i xib  C nb 8b 2 B n fB0 g; 8n 2 N; ð4Þ
i2I
n
ynis P xib s ¼ tðbÞ; . . . ; tðbÞ þ di  1; 8b 2 B n fB0 g; 8i 2 I; 8n 2 N; ð5Þ
X n
yit 6 C ICU
t 8t 2 T n T 0 ; 8n 2 N; ð6Þ
i2I

xib 2 f0; 1g 8i 2 I; 8b 2 B; ð7Þ


ynit 2 f0; 1g 8i 2 I; 8t 2 T; 8n 2 N; ð8Þ
onb 2 Rþ b 2 B n fB0 g; 8n 2 N: ð9Þ
Objective (1) minimizes the total cost that consists of the patient costs and expected overtime costs. In the objective function, implicitly, the
value of CQ ib is a priority score given to a patient waiting for surgery. That is, a patient whose CQ ib is higher than any others should be sched-
uled first when the surgical block capacity is enough. Since a patient’s waiting time depends on his/her priority, the patient cost CQ ib should
be designed with care. To quantify CQ ib , this study follows the concept proposed by Testi et al. (2007). They prioritize patients by priority
scores determined by c  waiting time, where c is the urgency status that is the weighted sum of the numerical values of three clinical criteria
such as disease progression, pain or dysfunction and disability. Each patient is assumed to have an initial cost c, and the cost CQ ib is non-
decreasing over waiting time. After a recommended time given to each patient, the costs will increase significantly and force the patient
to be scheduled first. Readers can refer to MacCormick et al. (2003) and Mullen (2003) for a survey of prioritizing elective patients.
Constraint (2) is a patient assignment constraint, and constraint (3) represents a block schedule related to a patient i. The following
three constraints show capacity constraints of two resources in surgery operations; surgical blocks (i.e. OR) and SICU beds. Constraint
(4) is the capacity constraint of a surgical block, and determines the total amount of overtime work. Constraints (5) and (6) ensure that
patients in SICU will not be over the maximum number of SICU beds at day t. Here tðbÞ is a day t at which a surgical block b is defined.
Recall that the consideration of SICU beds together with surgical blocks is important. The shortage of SICU beds restricts moving patients
from OR, and patients hold an OR until a SICU bed is available (Weinbroum et al., 2003). Consequently, the availability of SICU beds is crit-
ical to decide the admission of elective patients.

Table 1
Notation summary.

Indices and Sets


i2I Index of patient waiting for surgery
b2B Index of available surgical block
t2T Index of day
n2N Index of scenario
Parameters
CQ ib The cost when a patient i is assigned to a surgical block b
W ni Surgery duration for patient i under scenario n
di
n LOS in SICU for patient i under scenario n
C nb The capacity of surgical block b under the scenario n
C ICU The capacity of SICU at day t
t
Z ib 1 if a surgical block b is open and available for a patient i
0 otherwise
Variables
xib 1 if a patient i is assigned to a surgical block b
0 otherwise
yit 1 if a patient i occupies a SICU bed at day t
0 otherwise
ob The total overtime work of a surgical block b
D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652 645

2.2. A closed form model

This paper formulate the model in a closed form based on the research by Kleywegt et al. (2001), in which they provide a model for a
general knapsack problem. Let assume that W ni  Nðli ; r2i Þ for 8i 2 I and C nb  Nðlb ; r2b Þ for 8b 2 B are independent normally distributed
n
random variables. Let di  Uð0; d  Þ for 8i 2 I be independent uniformly distributed random variables. By these assumptions the model SSSP
i
can be written in a closed form.
P
The mean and variance of the total expected overtime of a surgical block b are given by lðxb Þ ¼ i2I li xib  lb and
2 P
rðxb Þ ¼ i2I r2i x2ib þ r 2
b
, respectively. Thus, the expected total overtime of a surgical block b can be written as:
 
2
Eob ¼ lðxb ÞU rðx Þ þ rpðxffiffiffiffi
n lðxb Þ bÞ
2p
exp  2lrðxðxb ÞÞ2 , where U denotes the standard normal cumulative distribution function (CDF). It is easy to show
b b

that the objective function (1) is formulated as


(   !)
XX X lðxb Þ rðxb Þ lðxb Þ2
CQ ib xib þ COb lðxb ÞU þ pffiffiffiffiffiffiffi exp  :
i2I b2B b2BnB0
rðxb Þ 2p 2rðxb Þ2

Let pis be the probability that a patient i occupies a SICU bed s days after surgery. Assuming that dni is not dependent on the surgery day t.
n  Þ,
Since di  Uð0; d i
(
di þ1s  > 0 and
d þ1 if di s ¼ 0; 1; . . . ; di
pis ¼ i

0 otherwise:
Pt P
For a given schedule x, the expected number of patients in SICU at day t follows that j¼0 i2Hj ðxÞ pitj , where Hj ðxÞ # I is a set of patients who
get surgery at day j for a given schedule x. Therefore, SICU capacity constraint (6) is formulated as
X
t X
pitj 6 C ICU
t for 8t:
j¼0 i2Hj ðxÞ

By enumeration, an exact solution can be obtained for a small size problem quickly and accurately. However, if the problem size gets bigger,
the closed form model is intractable. Furthermore, the closed form model is not applicable when the assumptions of normal and uniform
distribution do not hold. Practically it is well known that the distribution of surgery duration is close to a lognormal distribution (May
et al., 2000). While there are computational benefits, due to the limitations of the closed form expression, this paper employs sample average
approximation (SAA) algorithm as a solution strategy. For applying the SAA algorithm, the SSSP is reformulated using recourse model and the
structural properties of the new formulation are investigated.

2.3. Stochastic programming with recourse

The problem SSSP can be formulated using the following recourse model.
XX
ðSSSP1Þ Min CQ ib xib þ E½Qðx; nÞ;
i2I b2B
X
xib ¼ 1 8i 2 I;
b2B

xib 6 Z ib 8i 2 I; 8b 2 B n fB0 g;
xib 2 f0; 1g 8i 2 I; 8b 2 B;
where E½Q ðx; nÞ is the recourse function, and
X
ðSSSP2Þ Q ðx; nÞ ¼ min COb onb ;
b2B
X
onb P W ni xib  C nb 8b 2 B n B0 ; 8n 2 N;
i2I
n
ynis P xib s ¼ tðbÞ; . . . ; tðbÞ þ di  1; 8b 2 B n B0 ; 8i 2 I; 8n 2 N ;
X n
yit 6 C ICU
t 8t 2 T n T 0 ; 8n 2 N ;
i2I

ynit 2 f0; 1g 8i 2 I; 8t 2 T; 8n 2 N;
onb 2 Rþ b 2 B n B0 ; 8n 2 N:
There are two well-known difficulties related to E½Q ðx; nÞ. First, it is very hard to evaluate the value of E½Q ðx; nÞ because of the huge
random data vector n 2 N. It involves solving a huge number of similar integer programs. The second difficulty is that both Q ðx; nÞ and
E½Q ðx; nÞ are non-convex and discontinuous (Birge and Louveaux, 1997). From these two difficulties, it is difficult to define the objective
function explicitly, and only an implicit approximation is available. Thus, the model is very difficult to solve. Finally, the following struc-
tural property of the SSSP2 is proposed.

Proposition 1. The matrix defining the feasible region of ynit in SSSP2 is totally unimodular (TU).

Proof. Let A be the matrix defining the feasible region of ynit in SSSP2. aij denotes the coefficient of A in row i and column j. Then, it is easy to
show that
646 D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652

(i) aij 2 fþ1; 1; 0g for all i; j.

The structure of the matrix A is


0 1
1
B 1 C
B C
B C
B 1 C
B C
A¼B C:
B 1 C
B .. C
B C
@ .A
1 1 1 1 

From the structure of the matrix A, the following two properties hold.
P
(ii) For each j, i jaij j ¼ 2.
P
(iii) For each j, i2A aij ¼ 0.

(i), (ii) and (iii) satisfy the Proposition 3.2 in Wolsey (1998), and the matrix A is TU. h
According to the Proposition 1, the linear relaxation of ynit in SSSP2 will have integer optimal solutions whenever the right-hand side is
integral (Wolsey, 1998, p. 40). Since xib is binary and C ICU
t is discrete integer, ynit is relaxed as a continuous variable between 0 and 1.

Proposition 2. The function Q ðx; nÞ holds the following three properties: (i) For all n 2 N the function Q ðx; nÞ is piecewise linear and convex. (ii)
The probability distribution /ðnÞ has a finite support, i.e. the set N is finite, and (iii) the solution set of x is polyhedral.

Proof. The proofs of (ii) and (iii) are trivial by the assumptions that the probability distribution of /ðnÞ is discrete and the scenario set N is
finite. Furthermore, for a finite set of random vectors, the number of linear constraints in the model is finite. Therefore, the solution set is
polyhedral. (ii) and (iii) are proven.
For each n 2 N; onb has two possible values. That is,
8 P
<0
> if W ni xib  C nb 6 0
n i2I
Ob ¼ P n ð10Þ
> n
: W t xib  C b otherwise:
i2I

Eq. (10) shows that onb is piecewise linear and convex, and so is Q ðx; nÞ because Q ðx; nÞ is linear combination of onb . h

3. Solution strategy; sample average approximation

By the Proposition 2 and Shapiro et al. (2002), the optimal solution of the SAA problem provides an exact optimal solution of the true
SSSP (i.e. model (1)–(9)) with probability one (w.p. 1) for a sample size N that is large enough. Moreover, Shapiro and Homem-de-Mello
(2001) show that the probability of providing an exact optimal solution of the true problem approaches one exponentially fast as N tends
to infinity. These results imply that a good approximate solution can be obtained with a relatively small sample size.
The following mathematical model describes the SAA problem of the SSSP with sample size N.

X
I X
1 XN X
Min CQ ib xib þ COb onb ; ð11Þ
i¼1 b2B
N n¼1 b2BnfB0 g
X
xib ¼ 1 8i 2 I; 8b 2 B n fB0 g; ð12Þ
b2B

xib 6 Z ib 8i 2 I; 8b 2 B n fB0 g; ð13Þ


X n
onb P W i xib  C nb 8b 2 B n fB0 g; n ¼ 1; . . . ; N; ð14Þ
i2I
n
ynis P xib s ¼ tðbÞ; . . . ; tðbÞ þ di  1; 8b 2 B n fB0 g; 8i 2 I; n ¼ 1; . . . ; N; ð15Þ
X
ynit 6 C ICU
t 8t 2 T n T 0 ; n ¼ 1; . . . ; N; ð16Þ
i2I

xib 2 f0; 1g 8i 2 I; 8b 2 B; ð17Þ


ynit 2 ð0; 1Þ 8i 2 I; 8t 2 T; n ¼ 1; . . . ; N; ð18Þ
onb 2 Rþ b 2 B n fB0 g; n ¼ 1; . . . ; N: ð19Þ

Kleywegt et al. (2001) and Ahmed and Shapiro (2002) provide a general SAA algorithm for the type of stochastic discrete optimization prob-
lem. The general SAA algorithm is implemented with some modifications to fit the problem studied in this paper.
The procedure of the algorithm is as follows:

Step 1. For m = 1, . . . , M, do Steps 1.1 through 1.3.


Step 1.1. Generate N samples.
D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652 647

Step 1.2. Solve the SAA problem (i.e. model (11)–(19)), and let ^zm ^m
N and xN be the corresponding optimal objective value and an optimal
solution.
Step 1.3. Generate N 0 independent random samples. Evaluate the true objective function with a feasible solution  x and a correspond-
ing estimate of variance. Estimated true objective value g xÞ and variance S2g^ 0 ðxÞ can be obtained by the following equations:
^N0 ð
N
XX N0
1 X X
g^N0 ðxÞ ¼ CQ ib xib þ 0 COb onb ;
i2I b2B
N n¼1 b2BnB0
2 32
XN0 XX X
1 4
S2g^ 0 ðxÞ ¼ 0 0 CQ ib xib þ COb ob  g^N0 ðxÞ5 :
n
N N ðN  1Þ n¼1 i2I b2B b2BnB0

2
Step 2. Compute zM
N and a corresponding estimate of variance SzM by (20) and (21). N

1 X
M
zM
N ¼
^zm
N; ð20Þ
M m¼1

1 XM
 m 2
S2zM ¼ ^z  zM : ð21Þ
N ðM  1ÞM m¼1 N N

2 2
xm
Step 3. For each solution ^ ^ 0  M
N ; m ¼ 1; . . . ; M, estimate the optimality gap by g N ðxÞ  zN , along with an estimated variance of Sg^N0 ð
xÞ þ SzM .
N
Finally choose one of the M candidate solutions.

In the algorithm, zM ^ 0  


N and g N ðxÞprovide a statistical lower bound (LB) and upper bound (UB) respectively (Mak et al., 1999). Let z denote
 M ^ M ^ 
the optimal value of the true problem. zN is an unbiased estimator of E½zN , and it is known zN ¼ E½zN  6 z under mild conditions (ii) and (iii)
of the Proposition 2. The g^N0 ðxÞ is an unbiased estimator of the true objective value of a suboptimal decision 
x, which is a feasible, but maybe
a suboptimal solution for the original problem. Hence, g^N0 ð xÞ P z holds.
Note that Step 1.3., which is for obtaining statistical upper bound, does not require to solve a mixed integer programming (MIP) problem.
For a given feasible solution  x, the variable onb is determined by (10) instantaneously. This advantageous problem structure allows to use a
0
quite large sample size N without significant increase of computing time.
In Step 1.3., it is also important to obtain a feasible solution  x for a better upper bound. There may be several methods for obtaining a
‘good’ feasible solution  x. One good example is that a deterministic problem with expected value parameters, called the expected value
problem (EVP), provides a feasible solution. Another possible method is solving a problem without stochastic constraints. This method
is used for a two-stage stochastic problem, and the solution obtained by the first-stage problem can be used for evaluating the statistical
upper bound. In this study, preliminary experiments show that the SAA solution provides better upper bound than solutions obtained by
other methods stated previously. Thus, another SAA solution is generated with sample size N to obtain a feasible solution  x in Step 1.3.
Readers refer to Kleywegt et al. (2001) and Ahmed and Shapiro (2002) for detailed description of the SAA algorithm.

4. Numerical experiments

4.1. Example data

This section describes a test problem that is adopted from a regional hospital with some reasonable modifications. It is assumed that
there are nine surgical service groups with 10 available operating rooms and 32 surgical blocks (64% block fill rate). The empty surgical
blocks reflect operating rooms not in use because of other reasons like resource constraints. ORs are assigned to surgical service groups
by an arbitrarily defined surgical block schedule, which is shown in Fig. 1.
A surgical group consists of multiple surgeons having multiple surgical blocks within a planning horizon. A patient will be scheduled in
one of the available blocks. For example, if an ENT surgeon is allowed to use surgical blocks on Monday (operating room 1), Wednesday

Fig. 1. Example surgical block schedule.


648 D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652

Table 2
Statistical information on the distributions of surgery durations.

Surgical group Mean (minute) Standard deviation (minute) Skewness Observations (# of surgeries)
ENT 74 37 1.60 788
OBGYN 86 40 0.96 342
ORTHO 107 44 0.86 859
NEURO 160 77 0.54 186
GEN 93 49 1.88 817
OPHTH 38 19 0.36 110
VASCULAR 120 61 0.71 303
CARDIAC 240 103 0.33 90
UROLOGY 64 52 2.11 198

(operating room 10) and Friday (operating room 2), a patient in the waiting list of an ENT surgeon can be assigned to one of the available
surgical blocks. In the meantime, the surgical block can be shared with other surgeons in the same ENT service group. Each surgical block
has an 8-hour capacity per day. The total time of a surgical block required for emergency demands and turnaround time is assumed to be
2 hours, 2.5 hours or 3 hours with an even probability.
In order to obtain the discrete distribution of surgery duration, the actual data of surgery procedures is compiled and analyzed from the
hospital. The number of surgeries performed by the nine surgical service groups is about 3700 for seven months, and it equals, on average,
three patients being operated on in a surgical block. Practically, the surgery duration is dependent on the surgery type and the surgeon as
well as the patient. However, we assume that all patients in the same surgical service group follow identical distribution of surgery dura-
tion. Table 2 summarizes statistical information on the distribution of surgery durations, and Fig. 2 represents the distribution of surgery
durations of the general surgery service group with 30-minute interval duration. Except for a few, most distributions we derive from the
data analysis are consistent with empirical studies by May et al. (2000) and Spangler et al. (2004) that conclude a lognormal distribution
fits in actual surgery durations.
Due to the lack of actual data, LOS in SICU after surgery is generated randomly. Different mean values and discrete distributions are used
for each surgical service group. The average LOS in ICU of ENT, OBGYN, ORTHO, NEURO, GEN, OPHTH, VASCULAR, CARDIAC and UROLOGY
are 0.1 day, 2 days, 1.5 days, 2 days, 0.05 day, 0.05 day, 3.5 days, 2 days and 0.8 day, respectively. Thus, some patients may not stay in SICU
after surgery. We also perform experiments with three different numbers of SICU beds, 5, 10 and 15 when the average SICU demands is
around 10 in the given problem setting.
It is known that a minute of operating room time costs about $13 (Stodd et al., 1998). With the assumption of 100% overtime allowance,
the block overtime cost COb is $780 per 30 minutes (recall that we use 30 minutes interval for planning), and it is equal to $1560/hour.
Patient cost CQ ib is generated based on the overtime cost by the following equation. a is a random value obtained from uniform distribution
U(0.5, 1.0).

a  COb if b 2 B n fB0 g
CQ ib ¼ for 8i 2 I: ð22Þ
a  COb  1:5 if b 2 fB0 g

4.2. Experimental results

In this section we describe experimental results from the test problem of 203 elective patients. In the implementation, we use N = 1, 5,
10, 20, 30, 40, 50, 100, 200 and 500; M = 10; and N0 = 50,000. The SAA algorithm presented in Section 3 is implemented in C++ with ILOG
CPLEX 11.0 for solving MIP problems.

Fig. 2. Distribution of surgery durations of general surgery.


D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652 649

Fig. 3. Convergence of the objective value.

Fig. 3 demonstrates how the objective value of the SAA optimal solution changes as the sample size N increases, and Table 3 summarizes
corresponding results in detail. The results show the convergence of the SAA solutions with exponential rates. For the test problem, a good
solution of 0.5% optimality gap is obtained with sample size 100. Though the original sample space is very large, a good solution can be
obtained with relatively small sample size. The size of sample space (i.e. the size of scenario set NÞ considered in numerical experiments
is calculated byjIjs  jIj‘  jBje , where s; ‘, and e are possible realization of surgery durations, LOS in SICU and emergency demands, respec-
tively. The sample space of the example problem is 4  1034 approximately. Although we do not perform experiments to evaluate the ef-
fects of variable sizes, it is proven by Shapiro and Homem-de-Mello (2001) that the sample size to obtain a good SAA solution is
logarithmical to the variable size. Therefore, the required sample size will increase linearly as the number of patients and surgical blocks
increases.
The dotted line in Fig. 3 indicates the solution obtained by the EVP of the corresponding stochastic problem. The solution of EVP is ob-
n
tained by replacing the random parameters W ni ; C nb and di by their means and then solving the resulting deterministic problem. The result
shows that, compared to the SAA problem, the deterministic problem produced a poor solution. Simulation studies are performed to eval-
uate the value of the stochastic optimization compared to the deterministic EVP. A detailed result is shown in the next section.
Recall that during the experiments, the value of M (number of replications in the procedure) is set to 10. Table 3 shows statistical lower
and upper bounds, and it can be identified that 10 replications are enough to obtain a reasonable confidence interval of the bounds. If the
variances are too large, then the value of M should be increased until satisfying an underline criterion.
SSSP considers two typical surgical resources, surgical block and SICU beds. Here the effects of the SICU capacity (constraints (5) and (6))
are discussed. Table 4 compares test results from three different SICU beds with N = 50, N0 = 50,000 and M = 10. From Table 4 it can be ob-
served that SICU capacity limits the overall performance. The case of five SICU beds schedules 30 patients less than the case of 15 SICU beds.
Consequently the average surgical block utilization deteriorates from 75.39% to 54.80% due to the reduced number of scheduled patients.
Furthermore, it is observed that a short SICU capacity to cover demands for SICU will increase the total cost. This result comes from the
P P
increase of total patient costs obtained by i2I b2B CQ ib xib because of the cost structure of CQ ib presented by (22). While this study does

Table 3
Experimental results.

Sample size N 95% CI of LB 95% CI of UB Optimality gap (%) Computation time (second)
1 (200,486, 200,759) (258,670, 258,737) 22.45 43
5 (212,615, 212,829) (225,531, 225,568) 5.68 58
10 (215,289, 215,701) (226,232, 226,207) 4.75 59
20 (217,624, 217,936) (223,446, 223,479) 2.54 85
30 (218,891, 219,180) (223,993, 224,025) 2.22 112
40 (219,691, 219,968) (223,169, 223,195) 1.50 142
50 (219,866, 220,172) (223,328, 223,357) 1.49 178
100 (220,761, 220,958) (221,857, 221,875) 0.45 420
200 (221,811, 222,028) (222,985, 223,004) 0.44 1318

Table 4
Effects of SICU capacity.

SICU beds Average surgical block utilization (%) Average number of scheduled patients Average number of patients in ICU Average objective value
5 54.80 89 2.98 220,715
10 64.58 106 7.10 215,201
15 75.39 119 11.56 211,034
650 D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652

not provide the optimal level of SICU capacity, a decision maker is given a guideline for determining the number of SICU beds by exper-
iments. In particular, the problem that relaxes the SICU constraints ((15) and (16)) will provide a baseline solution.

4.3. Evaluation of stochastic solutions

Simulation experiments are performed to compare two solutions which are obtained from the stochastic optimization problem (SSSP)
and the deterministic expected value problem (EVP). The comparison aims at evaluating the benefit of a stochastic model over its compu-
tational efforts. From the simulation results, we claim that SSPP provides a better solution than EVP in terms of the robustness to the ran-
domness in surgery operations. Second, the EVP can be considered as the current clinical practice to schedule surgery patients, which
assign patients using fixed surgery durations into the surgical block until the total surgery duration does not exceed the planned block
capacity. Hence, the results from the simulation study show the significance of the solution quality.
The test problem presented in Section 4.1 is used for this simulation study with the following additional considerations. First, determin-
istic demands are considered, so the number of elective patients waiting for admission is constant in every simulation run. The number of
patients who have undergone surgeries is equal to the number of new patients during the planning horizon. Second, the costs of patients
who are not scheduled in the present planning horizon and cancelled due to overtime or lack of SICU beds will increase significantly. Dur-
ing the next planning horizon, the probability of admission of patients carried over from the previous planning horizon will increase be-
cause of the increase of patient cost. If scheduled surgeries are too many to be performed within given capacity, all patients over the
capacity are cancelled. In the simulation, 2 hours overtime is applied for the threshold of a cancellation. Finally, a in (22) is obtained from
uniform distribution U(0.5, 1.5). The summary of the simulation procedure is given in Fig. 4. Arbitrarily, the run length is determined as

Fig. 4. Simulation procedure.

Fig. 5. Simulation results: total costs.


D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652 651

Table 5
Summary of the performance measures obtained from simulation.

P < 0.001 Average objective Average overtime (hour/block Average undertime (hr/block Average scheduled patients/ Average cancellation/
value week) week) week week
SSPP 268,904 2.09 2.86 103 0.67
EVP 325,382 3.07 2.16 105 1.96

100 weeks, and the warm-up period is set up as 20 weeks. Thirty replications are performed for statistical evaluations. The simulation is
implemented in C++.
Five performance measures are evaluated: total cost, average overtime, average undertime (i.e. utilization), total number of scheduled
patients and cancellations. In Fig. 5, the total costs for 100 weeks are presented. The dotted line and the solid line represent the total cost
(objective function value) of the EVP and the SSSP, respectively. It is shown that the EVP has higher total cost than the SSSP, and about 17%
of cost reduction will be achieved by the stochastic optimization.
Other performance measures, which are presented in Table 5, indicate that the SSSP outperforms the EVP. The SSSP schedules almost the
same number of patients, and the overtime per surgical block is reduced by one hour on average. Additionally, significant improvement of
cancellations with less overtime work is identified. In the case of the EVP, average two patients are cancelled every week. However, less
than one patient is cancelled under the schedule obtained from the SSSP. In practice, a high rate of cancellations due to an ill schedule will
deteriorate service quality in terms of patient satisfaction and operation costs.
Since the initial SSSP excludes undertime costs, the EVP has less average undertime. Generally, under-utilizations reduce service pro-
viders’ profit, but the SSSP schedules almost the same number of patients as the EVP. So, profit loss is not expected in this simulation result.
Furthermore, when considering the same number of scheduled patients, more average undertime may provide flexibility to surgery oper-
ations without loss of profit.

5. Conclusions

This study proposes a stochastic optimization model for elective surgery scheduling with considering SICU capacity constraints. SAA
algorithm is employed to solve the problem and numerical experiments demonstrate the convergence of statistical bounds with moderate
sample size for a given test problem. Finally, a simulation study is conducted to show that SSSP outperforms EVP.
There are several topics for further research. First the patient cost CQ ib is described in Section 2.1, but an explicit structure of the cost is
not presented in this paper. In literature, the cost parameter CQ ib implies social, medical, economical and even psychological content. Fur-
thermore, the structure could be non-decreasing, non-increasing or independent of waiting time, because each patient may have different
initial condition, speed of disease progress, and either be making a recovery from a disease or relapsing. Hence, it is very difficult to define
an explicit structure and quantify the value of CQ ib . More research is needed to define it clearly.
Other areas of interest include coming up with methods for computational efficiency. We have not considered variance reduction tech-
niques, but this technique will certainly improve the performance. In addition, the problem structure is not fully utilized in this study. By
relaxing Eqs. (2), (5) and (6), the initial problem can be decomposed into several subproblems of each surgical block b. Since the sample size
and computation time are at least linear to block size, the decomposition method will improve the computational efficiency.
We assume that demands are deterministic. If we release this assumption, the initial SSPP is altered into a dynamic model. A stochastic
dynamic programming method can be employed for the new problem.

References

Ahmed, S., Shapiro, A., 2002. The sample average approximation method for stochastic programs with integer recourse. Technical Report, School of Industrial and Systems
Engineering, Georgia Institute of Technology.
Birge, J.R., Louveaux, F., 1997. Introduction to Stochastic Programming. Springer, New York.
Denton, B., Gupta, D., 2003. A sequential bounding approach for optimal appointment scheduling. IIE Transactions 35, 1003–1016.
Denton, B., Viapiano, J., Vogl, A., 2007. Optimization of surgery sequencing and scheduling decisions under uncertainty. Health Care Management Science 10 (1), 13–24.
Denton, B., Miller, J.M., Balasubramanian, H.J., Huschka, T.R., 2009. Optimal allocation of surgery blocks to operating room under uncertainty. Operations Research.
Fei, H., Chu, C., Meskens, N., Artiba, A., 2008. Solving surgical cases assignment problem by a branch-and-price approach. International Journal of Production Economics 112,
96–108.
Gerchak, Y., Gupta, D., Henig, M., 1996. Reservation planning for elective surgery under uncertain demand for emergency surgery. Management Science 42 (3), 321–334.
Guinet, A., Chaabane, S., 2003. Operating theatre planning. International Journal of Production Economics 85, 69–81.
Gupta, D., 2007. Surgical suites’s operations management. Production and Operations Management 16 (6), 689–700.
Hans, E., Wullink, G., Houdenhoven, M.V., Kazemier, G., 2008. Robust surgery loading. European Journal of Operational Research 185 (3), 1038–1050.
Hsu, V.N., Matta, R., Lee, C.Y., 2003. Scheduling patients in an ambulatory surgical center. Naval Research Logistics 50 (3), 218–238.
Jackson, R.L., 2002. The business of surgery: managing the OR as a profit center requires more than just IT. It requires a profit-making mindset, too – Operating Room Info
Systems, Health Management Technology.
Jonnalagadda, R., Walrond, E.R., Hariharan, S., Walrond, M., Prasad, C., 2005. Evaluation of the reasons for cancellation and delays of surgical procedures in a developing
country. International Journal of Clinical Practice 59 (6), 716–720.
Kleywegt, A.J., Shapiro, A., Homem-de-Mello, T., 2001. The sample average approximation method for stochastic discrete optimization. SIAM Journal on Optimization 12 (2),
479–502.
Lamiri, M., Xie, X., 2006. Operating rooms planning using Lagrangian relaxation technique. In: Proceeding of the 2006 IEEE International Conference on Automation Science
and Engineering Shanghai, China, October 7–10, 2006.
Lamiri, M., Xiea, X., Dolguia, A., Grimauda, F., 2008a. Stochastic model for operating room planning with elective and emergency demand for surgery. European Journal of
Operational Research 185 (3), 1026–1037.
Lamiri, M., Xie, X., Zhang, S., 2008b. Column generation approach to operating theatre planning with elective and emergency patients. IIE Transactions 40 (9), 838–852.
Macario, A., Vitez, T.S., Dunn, B., McDonald, T., 1995. Where are the costs in peri-operative care? Analysis of hospital costs and charges for inpatient surgical care.
Anesthesiology 83 (6), 1138–1144.
MacCormick, A.D., Collecutt, W.G., Parry, B.R., 2003. Prioritizing patients for elective surgery: a systematic review. ANZ Journal of Surgery 73, 633–642.
Magerlein, J.M., Martin, J.B., 1978. Surgical demand scheduling: a review. Health Services Research 13 (4), 418–433.
Mak, W.K., Mortonand, D.P., Wood, R.K., 1999. Monte Carlo bounding techniques for determining solution quality in stochastic programs. Operations Research Letters 24, 47–
56.
652 D. Min, Y. Yih / European Journal of Operational Research 206 (2010) 642–652

May, J.H., Strum, D.P., Vargas, L.G., 2000. Fitting the Lognormal distribution to surgical procedure times. Decision Sciences 31 (1), 129–148.
Mullen, P.M., 2003. Prioritizing waiting lists: how and why? European Journal of Operational Research 150, 32–45.
Ogulata, S.N., Erol, R., 2003. A hierarchical multiple criteria mathematical programming approach for scheduling general surgery operations in large hospitals. Journal of
Medical Systems 27 (3), 259–270.
Pham, D.N., Klinkert, A., 2008. Surgical case scheduling as a generalized job shop scheduling problem. European Journal Operational Research 185 (3), 1011–1025.
Shapiro, A., Homem-de-Mello, T., 2001. On the rate of convergence of optimal solutions of Monte Carlo approximations of stochastic programs. SIAM Journal on Optimization
11 (1), 70–86.
Shapiro, A., Homem-de-Mello, T., Kim, J., 2002. Conditioning of convex piecewise linear stochastic programs. Mathematical Programming 94, 1–19.
Sobolev, B.G., Brown, P.M., Zelt, D., FitzGerald, M., 2005. Priority waiting lists: is there a clinically ordered queue? Journal of Evaluation in Clinical Practice 11 (4), 408–410.
Spangler, W.E., Strum, D.P., Vargas, L.G., May, J.H., 2004. Estimating procedure times for surgeries by determining location parameters for the Lognormal model. Health Care
Management Science 7, 97–104.
Stodd, K., Ortiz, A., Tenzer, I., 1998. Operating room benchmarking: the Kaiser performance experience. The Performance Journal 2 (4), 5–16.
Testi, A., Tanfani, E., Torre, G., 2007. A three-phase approach for operating theatre schedules. Health Care Management Science 10, 163–172.
Weinbroum, A.A., Ekstein, P., Ezri, T., 2003. Efficiency of the operating room suite. The American Journal of Surgery 185, 244–250.
Wolsey, L.A., 1998. Integer Programming. John Wiley and Sons, New York.

Вам также может понравиться