Accepted Manuscript

A Mixed Integer Programming Approach to the Patient Admission

Scheduling Problem

Leonardo S.L. Bastos , Janaina F. Marchesi , Silvio Hamacher ,

Julia L. Fleck

PII: S0377-2217(18)30749-5
DOI: https://doi.org/10.1016/j.ejor.2018.09.003
Reference: EOR 15346

To appear in: European Journal of Operational Research

Received date: 14 September 2017

Revised date: 28 August 2018
Accepted date: 2 September 2018

Please cite this article as: Leonardo S.L. Bastos , Janaina F. Marchesi , Silvio Hamacher ,
Julia L. Fleck , A Mixed Integer Programming Approach to the Patient Admission Scheduling Problem,
European Journal of Operational Research (2018), doi: https://doi.org/10.1016/j.ejor.2018.09.003

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights

 The static offline operational Patient Admission Scheduling Problem is addressed

 An exact method is proposed to solve a mixed integer programming formulation
 The method is parameter-free and does not require pre-processing of penalties
 New best known solutions are generated for 9 out of 13 benchmark instances
 Proof of optimality of two best known solutions reported in the literature

T
IP
CR
US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

A Mixed Integer Programming Approach to the Patient Admission Scheduling Problem

Leonardo S.L. Bastos1, Janaina F. Marchesi1, Silvio Hamacher1*, Julia L. Fleck1

lslbastos@tecgraf.puc-rio.br; janaina.marchesi@tecgraf.puc-rio.br; hamacher@puc-rio.br;

jfleck@puc-rio.br

1
Department of Industrial Engineering, Pontifícia Universidade Católica do Rio de Janeiro, Rua Marquês

T
de São Vicente, 225, Rio de Janeiro, RJ, 22451-900, Brazil

IP
*
corresponding author

CR
Abstract

Among the many challenges involved in efficient healthcare resource planning, the Patient Admission

US
Scheduling Problem is of particular significance, impacting organizational decisions at all planning levels.
The problem of scheduling patient admissions involves assigning patients to beds over a given time
horizon so as to maximize treatment efficiency, patient comfort and hospital utilization, while satisfying
AN
all necessary medical constraints and taking into consideration patient preferences as much as possible. A
number of different variants of the Patient Admission Scheduling Problem exist at the strategic, tactical,
and operational levels. In this paper, we consider a static offline operational level variant for which we
M

propose a comprehensive mixed integer programming formulation and advance an exact solution method.
We generate new best found solutions for 9 out of 13 benchmark instances from a publicly available
ED

repository. Additionally, we prove the optimality of two best known solutions reported in the literature.

Keywords: OR in health services; patient admission scheduling; mixed integer programming; bed
PT

allocation; healthcare

Funding: This work was supported by the National Council for Scientific and Technological
CE

Development (CNPq) [grant number 146684/2016-8 to LSLB, grant numbers 306802/2015-5 and
403863/2016-3 to SH]; Fundação Carlos Chagas Filho de Amparo à Pesquisa do Estado do Rio de Janeiro
AC

(Faperj) [grant number E-26/201.959/2017 to JFM]; Coordination for the Improvement of Higher
Education Personnel (CAPES); and the Pontifícia Universidade Católica do Rio de Janeiro.

Conflicts of Interest: The authors declare no competing conflicts of interest.

 ACCEPTED MANUSCRIPT

1. Introduction

Life expectancy doubled worldwide in the course of the 20th century (Riley, 2005), and new
health delivery models and technology are expected to further increase healthy life expectancy (Frieden,
2015). In recent decades, an ageing population and improvements in preventive care have increased
demand for healthcare services; at the same time, the healthcare industry remains under pressure to cut
costs and improve quality of care (Merrild, 2015). To offset a prospective rise in clinical medicine costs

T
unaccompanied by significant improvements in health outcomes (Bayer & Galea, 2015), the healthcare

IP
sector has largely switched its focus to a value-based strategy (Porter & Lee, 2013). In this context, the
goal is to achieve the best outcomes at the lowest cost, meaning that efficient use of resources and patient

CR
satisfaction are important, albeit competing, objectives that must be met by healthcare administrators.
Addressing this challenge involves attending to practical issues such as admissions control, process
design, aggregate planning, capacity allocation, and appointment scheduling. Among these concerns, the

US
Patient Admission Scheduling Problem (PASP) is of particular significance since its solution impacts
essentially all other healthcare-related problems.
AN
The PASP entails the assignment of patients to beds in a healthcare institution over a given time
horizon so as to maximize treatment efficiency, patient comfort and hospital utilization, while satisfying
all necessary medical constraints and taking into consideration patient preferences as much as possible.
M

The problem of scheduling patient admissions has implications at all planning levels within a healthcare
institution. At the strategic level, the goal is to establish long-term structural decisions that maximize
ED

organizational efficiency. Historical data on patient admittance is frequently used to estimate future
demand and determine the corresponding amount of required resources (e.g., available beds). Numerous
strategic level variants of PASP have been reported in the literature and we briefly review some of them.
PT

Kusters and Groot (1996) used real patient data from Dutch hospitals to develop predictive models of
resource utilization and emergency admissions in order to optimally allocate beds, nursing staff, and
CE

operating theater facilities. A framework for modeling beds, workforce needs, and operating theaters was
proposed by Harper (2002) to quantify planning policies so as to increase operational efficiency and
AC

effectiveness. Focusing on the principles driving admission planning, Vissers et al. (2007) advanced a
methodology for contrasting the impact of different service concepts on hospital admission strategies.
Chen et al. (2010) designed a genetic algorithm for optimizing long-term admission scheduling strategies
seeking to balance the competing objectives of efficiency and fairness. The issue of admission planning
under stochastic demand requirements was addressed by Jittamai and Kangwansura (2011) through a
mixed integer linear programming model that optimized the utilization of operating room resources
conditioned on emergency admissions, unattended patients, and no-shows. Lee and Kwak (2011)
 ACCEPTED MANUSCRIPT

combined a multi criteria decision-making modeling framework with goal programming to design,
evaluate and implement strategic enterprise resource planning in a healthcare organization in South
Korea.

At the tactical level, the focus is on process organization and execution for an intermediate
planning horizon. Seasonal variations in available resources and demand are commonly accounted for in
order to specify admission policies in terms of the amount of necessary resources, mix of admitted

T
patients, and guidelines for operational planning. Tactical level variants of the PASP are typically solved

IP
a few weeks to a few months in advance and have also been vastly reported in the literature. Adan and
Vissers (2002) developed an integer linear programming model to generate specialty-specific patient

CR
admission profiles conditioned on target patient throughput and resource utilization. Focusing on the
impact of variable length of stay on hospital operational effectiveness, Gallivan et al. (2002) modeled the
operation of a hypothetical booking system using data from postoperative intensive care patients and

US
cautioned that booked surgery admissions should be conditioned on a high degree of reserve capacity.
Elective patient admission control was modeled as a Markov decision process by Nunes, de Carvalho, &
AN
Rodrigues (2009) and value iteration was used to optimize admission policies for multiple specialties
across successive planning periods. Addressing the issue of emergency department crowding, Bair et al.
(2010) applied discrete event simulation to investigate the impact of inpatient boarding on quality of care
M

at the emergency department. A discrete event simulation model was also used by Holm, Luras, & Dahl
(2013) to maximize bed utilization across different wards in a hospital. Barz and Rajaram (2015)
ED

formulated the problem of elective patient admission control as a Markov decision process and developed
approximate dynamic programming-based heuristics accounting for multiple resource constraints and
uncertain requirements.
PT

Finally, the operational level is concerned with defining scheduling rules that govern day-to-day
patient admission planning. Short-term decisions are commonly made through analysis of daily demand
CE

fluctuations and enforcement of priority rules. In its classic form, the PASP operational level variant is a
static offline problem that assumes deterministic demand requirements and fixed admission dates and
AC

length-of-stay. The classic PASP variant was introduced by Demeester et al. (2010) and has been shown
to be NP-hard (Vancroonenburg et al., 2011). Demeester et al. (2010) proposed a hybrid token-
ring/variable neighborhood descent-based tabu search algorithm and also generated and made publicly
available a set of instances based on real hospital situations that have become benchmark data sets for
evaluating and comparing PASP solutions. A number of algorithms have since been proposed that aim at
improving on the approximate solution from Demeester et al. (2010). Bilgin et al. (2012) improved upon
a subset of the benchmark instances using a hyper-heuristic framework in which local search
 ACCEPTED MANUSCRIPT

neighborhoods served as low-level heuristics. Further improvements were obtained by Ceschia and
Schaerf (2011) by means of a multi-neighborhood local search procedure that adapted neighborhood
relations to account for different weights of the objective function components. Hammouri and Alrifai
(2014) reported the first evolutionary algorithm-based approach to the classic PASP based on
biogeography optimization, but failed to improve upon the state-of-the-art. An optimization-based
heuristic was proposed by Range, Lusby, & Larsen (2014) where branch-and-bound, column generation,
and dynamic constraint aggregation techniques were used to achieve tighter lower bounds and generate

T
new best known solutions for a subset of the benchmark instances. Using meta-heuristics based on an

IP
adaptive non-linear great deluge algorithm, Kifah and Abdullah (2015) solved a subset of the benchmark
instances in a considerably smaller amount of computational time, but failed to improve upon the

CR
objective function values from Range, Lusby, & Larsen (2014). Focusing on decreasing the run time,
Turhan and Bilgen (2017) applied two mixed integer programming-based heuristics to decomposed sub-

US
problems of the PASP and obtained solutions with optimality gaps of the order of 5-15% in less than
three minutes of CPU time.
AN
A great number of operational level approaches to the PASP have relied on metaheuristics in
combination with a variety of search strategies. Moreover, although Range, Lusby, & Larsen (2014) and
Turhan and Bilgen (2017) incorporated mixed integer programming formulations into heuristic-based
M

approaches, only lower bounds on the optimal solution were obtained. Hence, to the best of our
knowledge, no solution method has yet been reported to provide guaranteed optimality for the static
ED

PASP; this constitutes the main motivation for our work. Of note, a dynamic variant of the classic PASP
was introduced by Ceschia and Schaerf (2012) to account for uncertain length of stay, admission delays,
and non-elective patients. This variant was solved using simulated annealing-based metaheuristics and
PT

subsequently extended to incorporate operating theater restrictions and a local search approach (Ceschia
and Schaerf, 2016).
CE

In this work, we focus on the static offline operational PASP, as defined in Demeester et al.
(2010), and advance a mixed integer programming formulation through which optimal solutions are
AC

obtained. Our contributions are twofold. First, we propose a mathematical formulation incorporating all
restrictions from the original model of Demeester et al. (2010), in contrast to existing work, where a
simplified version of the allocation model is commonly used and instances are typically pre-processed
using the algorithm from Ceschia and Schaerf (2011). Second, we apply an exact method to obtain
optimal solutions and benchmark its performance against existing approaches. Unlike heuristic and
metaheuristic-based methods, in which run time is used as stopping criteria, we generate optimal
solutions. We generate new best found solutions for 9 out of 13 benchmark instances from a publicly
 ACCEPTED MANUSCRIPT

available repository, two of which with proven optimality. Moreover, we obtain solutions for two
instances that are equal to best known solutions reported in literature and also show them to be optimal.
To the best of our knowledge, ours is the first work to provide solutions for the static PASP with
guaranteed optimality.

The remainder of this paper is organized as follows. In Section 2 we revisit the PASP formulation
and present relevant definitions and terminology. Section 3 details the proposed model and solution

T
method. Results are presented in Section 4 and concluding remarks and suggestions for future research

IP
directions are given in Section 5.

2. Problem Statement

CR
We consider a static offline variant of the operational PASP, as defined in Demeester et al.
(2010), where only elective patients are accounted for, admission dates are assumed to be known in

US
advance, and a default average length-of-stay is defined a priori based on each patient’s diagnosed
disease. Considering that physicians are frequently capable of accurately estimating the length-of-stay
AN
needed for a given treatment and that emergency patients can be implicitly accounted for by slack beds,
the aforementioned assumptions may be deemed reasonable for medium or short-term planning. In what
follows, we present the problem definition and terminology in a manner that is consistent with that of
M

Demeester et al. (2010).

2.1. Problem Definition

ED

We begin by introducing the terminology that will be used in the remainder of this paper. A
planning horizon consists of a set of subsequent days of finite and fixed length within which admission
PT

and discharge days are specified. The Length of Stay (LoS) of each patient, defined in terms of the
admission and discharge days, is also assumed to be known in advance. A patient is a person who needs
CE

medical treatment, and only elective patients (those who stay more than one night) are eligible to be
scheduled; elective patients may be classified as planned or admitted patients. The former are patients
waiting to be hospitalized, with defined admission and discharge days, and the latter consist of patients
AC

who already occupy one of the hospital beds. Each bed belongs to a room, which contains a certain
number of beds. The capacity of a room defines its type: single (one bed), double (two beds) or wards
(four beds), and each patient has a preference for a certain type of room, termed room preference. Each
room belongs to a department that is specialized in the treatment of certain pathologies. Therefore, each
department attends a set of specialisms, whose requirements are met by certain rooms. Moreover, each
department and room has its own priority degree for those specialisms. Room properties are defined in
 ACCEPTED MANUSCRIPT

terms of different available equipment, such as oxygen and telemetry. Beds in the same room benefit from
the same equipment. Depending on the treatment, a patient may require or prefer to be allocated to a room
with certain equipment. The room policy refers to the gender policy in effect. Rooms that require patients
to be of the same gender enforce policy M (only Male) or policy F (only Female). In contrast, rooms
where both genders are allowed enforce policy N (mixed gender at any time) or policy D (any gender may
be present; however, on any given night, only patients of the same gender must be present, and the policy
is defined by the first patient to be scheduled in that room). The age policy is defined for departments that

T
impose a maximum or minimum age limit for acceptance. Transfers correspond to the exchange of

IP
patients between rooms throughout their length of stay and should be avoided. Table 1 presents the
domains in which each of the aforementioned attributes are defined.

CR
Table 1: Sets and corresponding domains
Set Index Domain
Patient (P)
Female (F)
Male (M)
p
f
m
US {1, …, |P|}
Subset of Patients (P)
Subset of Patients (P)
AN
Specialism (S) s {1, …, |S|}
Department (D) d {1, …, |D|}
Room (R) r {1, …, |R|}
M

*
Non-dependent Rooms (ND) nd Subset of Rooms (R)
**
Dependent Rooms (DR) dr Subset of Rooms (R)
Room Property (K) k {1, …, |K|}
ED

Gender (G) g {1, …, |G|}

Day/Night (N) n {1, …, |N|}
Room Preference (W) w {1, …, |W|}
PT

*
Subset of Rooms which have a Defined Room Policy (Only Female, Only Male or Mixed)
**
Subset of Rooms whose Room Policy depends on the first patient assigned
CE

In this context, the objective of the PASP is to allocate patients to beds subject to room and
department policies as well as the requirements of each patient’s treatment. Increased patient satisfaction
AC

is naturally associated with the ability to meet room preferences and avoid transfers. Based on the
terminology presented above, a set of constraints may be defined in terms of necessary and desirable
room characteristics. The solution to the PASP aims at minimizing the total sum of penalized violations
of constraints regarding room policies as well as required and preferred patient preferences. Next we
present the constraints defined for the PASP and discuss the penalties associated with violating them.
 ACCEPTED MANUSCRIPT

2.2. Problem Constraints

In a manner similar to Turhan & Bilgen (2017), we define a total of 13 constraints and distinguish
between hard and soft constraints. The set of hard constraints (HC) includes the following:

HC1: A room must be available to receive a patient in the course of the planning horizon.

HC2: Admission and discharge days and the length of stay cannot be changed.

T
HC3: The length of stay is continuous, and a patient is scheduled until his/her discharge date.

IP
HC4: Only one patient must be scheduled to a given bed on any given night.

CR
The set of soft constraints (SC) includes the following:

SC1: Gender policy must be respected for each room.

US
SC2: A patient must be scheduled respecting the department age limit, when applicable.
AN
SC3: A patient must be allocated to a room with required room properties for his/her treatment.

SC4: For clinical reasons (e.g., quarantine), some patients must be assigned to a single room when
M

required.

SC5: The room type preference of a patient (single, double or ward) should be considered whenever
ED

possible. If the patient is assigned to a room in which the number of beds is higher than his/her
preference, a violation is incurred. Otherwise, no violation is considered.
PT

SC6: A patient should be assigned to a department that attends to his/her treatment specialism.

SC7: A patient should be allocated to a room that attends to his/her specialism in first degree of priority.
CE

If the degree of priority is secondary or tertiary, a violation is considered as the diference between
degrees.
AC

SC8: A patient should be scheduled to a room with his/her preferred room properties. This constraint is a
complementary version of SC3.

SC9: Transfers should not be allowed.

Constraints represent both necessary and desirable scheduling conditions, and the distinction
between these is given in terms of the penalty values associated with them. On the most stringent end of
the spectrum are constraints for which no penalty value is defined; these correspond to hard constraints
 ACCEPTED MANUSCRIPT

that must necessarily be respected (HC1-HC4). Slightly more flexibility is afforded to soft constraints for
which a high penalty value is associated (SC1-SC4 and SC9). Lastly, desirable conditions are cast as soft
constraints that may be violated by incurring in a relatively low penalty (SC5-SC8). Penalties are
computed on a per patient, per night basis, and the cost of violating constraints is tallied up over the entire
length of stay. For consistency, we adopt the same penalty values as in Turhan & Bilgen (2017), as shown
in Table 2.

T
IP
CR
Table 2: Penalties for violating hard and soft constraints

Constraint Penalty
SC1
SC2
SC3
5
10
5
US
AN
SC4 10
SC5 0.8
SC6 1
M

SC7 1
SC8 2
ED

SC9 11
PT

3. Solution Method

Based on the terminology defined in Section 2, the PASP can be formulated as a Mixed Integer
CE

Program (MIP) whose solution follows a two-step approach. The first step consists of implementing the
MIP as a patient-room assignment problem, a highly efficient format for the PASP through which best
known solutions for most instances are obtained. The second step includes post-processing the solution in
AC

order to obtain patient-bed assignments, which are subsequently validated by an application made
available online1 by Demeester et al. (2010). Beds within the same room are equal and we assume that a
patient assigned to a certain room can occupy any of the beds available in that room. Hence, with the
information regarding patient-room assignments, we obtain the final patient-bed solution by randomly

1
https://people.cs.kuleuven.be/~wim.vancroonenburg/pas/.
 ACCEPTED MANUSCRIPT

allocating patients to beds. A complete listing of parameter and variable definitions is given in Tables 3
and 4.

T
IP
CR
US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Table 3: Model parameters

Parameter Description Unit
Penalty on Room Gender Policy violation (SC1) -
Penalty on Age violation (SC2) -
Penalty on Mandatory Room Propeties violation (SC3) -
Penalty on Room Type Requirement violation (SC4) -
Penalty on Room Type Preference violation (SC5) -

T
Penalty on Specialism - Department violation (SC6) -
Penalty on Specialism - Room priority violation (SC7) -

IP
Penalty on Preferred Room Propeties violation (SC8) -
Penalty on Transfers (SC9) -

CR
Number of beds in room -
Admission day of patient -
Discharge day of patient
Length of stay of patient p [DD - AD ]
Age of patient
US -
Day
-
AN
Minimum age in department -
Maximum age in department -
Mandatory room property of patient {0,1}
M

Room property present in room {0,1}

Gender of patient {0,1}
ED

Gender of room (defined room policy) {0,1}

Room Type of room {0,1}
Room Type preference of patient {0,1}
PT

Room Type Requirement of patient (indicates need for single room) {0,1}
Indicates if a Room is of type “Single Room” {0,1}
CE

Specialism associated with patient in night {0,1}

Specialism accepted in room (and its respective department ) {0,1}
Specialism priority in room -
AC

Preferred room property of patient {0,1}

Auxiliary binary parameter to indicate the admission plan of each hospitalized {0,1}
patient (computed using AD and DD )
Auxiliary binary parameter to indicate if room belongs to a certain {0,1}
department
 ACCEPTED MANUSCRIPT

Table 4: Model variables

Variables Description Domain
Violation: 1 if patient is assigned to a room whose gender policy does not match {0, 1}
his/her gender, on night ; 0 otherwise (SC1 – Mixed, Female and Male policies)
Violation: 1 if room contains at least two patients of diferent gender for each night ; 0 {0, 1}
otherwise (SC1 – Dependent policy)
Violation: 1 if patient is assigned to a room whose age policy does not match his/her {0, 1}

T
age, on night ; 0 otherwise (SC2)
Violation: 1 if patient is assigned to a room that does not contain his/her mandatory {0, 1}

IP
room properties , on night ; 0 otherwise (SC3)
Violation: 1 if room type requirement is violated for patient , in room on night ; 0 {0, 1}

CR
otherwise (SC4)
Violation: 1 if room type preference is violated for patient on night ; 0 otherwise {0, 1}
(SC5)

US
Violation: 1 if patient , on night , is assigned to a room
does not attend his/her specialism; 0 otherwise (SC6)
(department) that originally {0, 1}
AN
Violation: If patient , on night , is assigned to a room that does not prioritize his/her
specialism, the diference is computed; 0 otherwise (SC7)
Violation: 1 if patient is assigned to a room that does not contain his/her preferred {0, 1}
M

room properties , on night ; 0 otherwise (SC8)

Violation: 1 if patient is transferred from room in night ; 0 otherwise (SC9) {0, 1}
Alocation variable: 1 if patient is assigned to room on night ; 0 otherwise {0, 1}
ED

Indicator variable: 1 if room contains only female patients on night ; 0 otherwise {0, 1}
PT
CE
AC
 ACCEPTED MANUSCRIPT

The proposed mathematical formulation is as follows:

min (1)

with ∑ ∑ ∑( )

∑∑∑

T
∑∑∑∑

IP
∑∑∑

CR
∑∑

∑∑∑

∑∑∑
US
AN

∑∑∑∑
M

∑∑∑
ED

s.t. ∑ (2)

(3)
PT

∑ (4)
CE

(5)
(6)
AC

( ) (7)
( ) (8)
( ) (9)
(10)
(11)
(12)
 ACCEPTED MANUSCRIPT

(13)
(14)
(15)
(16)
(17)
(18)
(19)

T
(20)

IP
(21)
(22)

CR
(23)
(24)

US (25)
(26)
(27)
AN
(28)
M

The objective function in (1) entails minimizing the total sum of soft constraint violations.
Constraint (2) limits assignments to the capacity of each room for each night. Constraint (3) enforces that
patients are scheduled in compliance with the admission plan, while constraint (4) guarantees that a
ED

patient is assigned to only one room for each night (HC1-HC4). Constraints (5)-(7) enforce room gender
policy requirements (SC1). Constraint (5) applies to rooms with defined gender policy, while (6)-(7) refer
PT

to dependent policies. Using variable to designate the presence of female patients in the room,
constraint (6) enforces female patient restrictions and (7) enforces male patient restrictions. Both
constraints seek to avoid the assignment of two distinct genders to the same room, penalizing allocations
CE

in which different genders share a room. Constraints (8)-(9) enforce maximum and minimum age
requirements of the department in which the room is located (SC2), when applicable. Instead of a Big M
AC

approach, the difference between the patient’s age and the required age is added so as to allow allocation
to occur when a violation occurs. Constraint (10) matches the patient’s mandatory room properties to the
existing ones (SC3); failure to provide the required properties results in a violation. Constraint (11)
allows the assignment of patients to rooms that fully or partially meet their capacity preferences (SC5),
while constraint (12) enforces single-room requirements associated with treatment (e.g., in cases of
quarantine) (SC4). Constraint (13) ensures that a patient is assigned to a room that attends his/her
specialism (SC6). Constraint (14) matches the priority of the required specialism (SC7), and (15) matches
 ACCEPTED MANUSCRIPT

the patient’s preferred room properties (SC8). Finally, constraint (16) tracks patient transfers (SC9), and
constraints (17)-(28) define the domain of the decision variables.

One of the main distinguishing features of our model is the fact that constraint violations are
modeled separately, in contrast with existing formulations that bundle several constraints. A case in point
are the models from Ceschia & Schaerf (2011), Range, Lusby, & Larsen (2014), and Turhan & Bilgen
(2017), which also considered patient-room assignment prior to patient-bed allocation, explicitly

T
accounting for constraints (2)-(4) but combined the total penalty of soft and hard constraints (with the

IP
exception of SC1 and SC9) into a penalty matrix. This matrix represents the violation cost of assigning a
patient to a bed (Demeester et al., 2010; Kifah and Abdullah, 2015) or a room (Ceschia & Schaerf, 2011;

CR
Range, Lusby, & Larsen, 2014; Turhan & Bilgin, 2017). In addition, a version of constraints (5)-(7)
without the violation variable is present in Demeester et al. (2010), Ceschia & Schaerf (2011), Range,
Lusby, & Larsen (2014), and Turhan & Bilgin (2017). Constraint (16) and the transfer variable in (18)

US
were adapted from the approach for patient transfers presented in Demeester et al. (2010), Ceschia &
Schaerf (2011), and Turhan & Bilgin (2017). The remaining constraints, namely (8)-(15) and the violation
AN
variables in (20)-(28), were not considered in previous MIP approaches for PASP and constitute an
original contribution of this work. With regards to the objective function (1), previous models minimize
penalties incurred in all violations, accounting for the sum of the penalty cost matrix as well as violations
M

of SC1 and SC9. In what follows, we address dimensionality concerns by defining conditions for variable
and constraint generation.
ED

3.1 Sparsity Conditions

Vexed by the curse of dimensionality, large MIP models are commonly unsuccessful in
PT

obtaining feasible solutions or good solutions in acceptable time. To address this limitation, we propose a
set of spasity conditions that reduce the solution space of our model and ultimately improve MIP
performance. A straightforward, albeit impactful, way of decreasing sparsity is through the LoS and
CE

admission plan parameters. Since we assume that the hospitalization period of each patient is known in
advance and does not change, we can limit our solution space both in terms of the time periods
AC

(accounting only for those periods where LoS are defined) and in terms of the type of patients
(considering only elective patients).

Given that model variables are mostly limited by the planning horizon and LoS, violation
variables follow the binary parameters regarding the constraints to which they are related. This allows
variables to be created only in situations where a violation may occur, thus decreasing the number of
feasible solution combinations. Sparsity conditions for variables resemble the penalty-matrix approach of
 ACCEPTED MANUSCRIPT

Ceschia & Schaerf (2011). Constraints are also limited by the planning horizon and LoS, when applicable.
The age policy enforced in (8) and (9) is limited to the rooms of certain departments. In the case of
constraints (14) and (15), which match the specialisms between patient-department and patient-specialism
priority, respectively, constraints are generated only when a patient is treated in certain specialisms.
Furthermore, (10) and (15) are generated only when a patient needs and prefers (binary parameters)
certain room property or equipment. Those conditions reduce the total number of restrictions to the cases
when a violation may occur, which improves the overall efficiency of the approach.

T
Sparsity conditions used in our formulation are defined by the following rules:

IP
1. The variables , , ,, , , , , ,

CR
can be omitted from the formulation when = 0 and = 0.

2. Variable can be omitted from the formulation when = 0.

3. Variable
4. Variable US
can be omitted from the formulation when
can be omitted from the formulation when = 0 and
= 0.

= 0.
AN
5. In general, contraints (4)-(16) can be omitted from the formulation when =0

and = 0.

6. Contraints (8)-(9) can be omitted from the formulation when = 0. In addition, constraint
M

(8) can be omitted when = 0 and constraint (9) when = 0.

7. Contraint (10) can be omitted from the formulation when = 0.

ED

8. Contraints (13)-(14) can be omitted from the formulation when = 0.

9. Contraint (15) can be omitted from the formulation when = 0.

PT

3.2 Special Case: Model without Transfers

Patient transfers (SC9) are associated with the highest penalty value among all violations (Table
CE

2) and may ultimately lead to other violations in the static offline operational PASP formulation. This
reflects the many challenging issues involved in transferring patients during their hospitalization, such as
ensuring room availability and reevaluating requirements to guarantee quality of care. Transfers should
AC

therefore be cautiously considered and it is reasonable to expected that optimal solutions present as few of
these violations as possible.

In the heuristics methods proposed by Ceschia & Schaerf (2011) and Range, Lusby, & Larsen
(2014), different approaches to dealing with transfer penalties were tested. In this work, we model SC9
using (16) so as to reduce the number of transfers by evaluating pairs of consecutive nights for each
patient. However, since our solution space contains transfers, and recalling that they are associated with
 ACCEPTED MANUSCRIPT

the highest penalty cost, it is reasonable to expect that solutions without transfers would not only be
feasible, but also close to the real optimal solution. In fact, a model without transfers will always provide
a feasible solution to the complete model, since the solution space of the former is contained in that of the
latter. More importantly, using the solution to the model without transfers as an initial solution to the
complete model yields better results for the tested benchmark instances than directly solving the complete
formulation (see Section 4 for a complete discussion of this matter).

T
The model we present next is a special case of the general model (1)-(28). In what follows, we
detail its underlying simplifications. First, transfers are modeled as a hard constraint, so that a patient

IP
must stay in the same room during his/her entire LoS. This is modeled by exchanging variable by

CR
, a binary variable taking value of 1 if a patient is allocated to a certain room, and 0 otherwise. In
addition, contraints (3) and (16) are removed and the following constraints are added:

(29)
∑

∑
US (30)
AN
(31)

Constraint (29), which substitutes (3), ensures that if a patient is allocated to a certain room, they
M

must remain in that room during the entire length-of-stay defined for that patient’s hospitalization.
Constraint (30) enforces that a patient must be assigned to at most one room, and, for generality, it also
ED

allows patients not be scheduled, if necessary. Constraint (31) defines the domain of the new allocation
variable.

We end by noting that although new constraints are added, the new decision variable is associated
PT

with a smaller number of sets ( , ) when compared to transfer variable . Since the solution space for
the simplified model is a subset of the one for the complete model, it is reasonable to expect that optimal
CE

solutions will be found in less computation time, provided that a feasible integer solution exists for the
model without transfers. Finally, regarding sparsity conditions, contraints (29)-(31) can be omitted from
AC

the formulation when = 0.

4. Computational Results

In the discussion that follows, we will refer to the model without transfers (defined in Section
3.2) as simplified model, while the general model with transfers, given by (1)-(28), will be referred to as
complete model. It is clear that the simplified model accounts for a subset of the complete model’s
solution space, meaning that a solution with zero gap for the former is not necessarily an optimal solution
 ACCEPTED MANUSCRIPT

to the latter. Hence, optimality can only be guaranteed by solving the complete model, whose formulation
accounts for all possible transfers. Nevertheless, an integer solution without transfers is a feasible solution
to the PASP and it is, therefore, possible to use it as an initial solution to the complete model. In what
follows, we verify that the use of such warm start approach can be advantageous for solving the 13
benchmark instances generated by Demeester et al. (2010).

We begin by characterizing the 13 instances considered in this study in terms of the number of

T
rooms (R), beds (B), patients (total – P and elective – E), the length of the planning horizon (PH), number

IP
of departments (D) and specialisms (S), as well as the average percentage bed occupancy (BO), as shown
in Table 5.

CR
Table 5: Characteristics of the problem instances

Instance R B P E PH (days) D S Properties BO (%)

1
2
3
98
151
131
286
465
395
693
778
757
652
755
708
14
14
14
US
4
6
5
4
6
5
2
2
2
59.69
59.98
57.07
AN
4 155 471 782 746 14 6 6 2 54.23
5 102 325 631 587 14 4 4 2 59.69
6 104 313 726 685 14 4 4 2 64.38
7 162 472 770 519 14 6 6 4 33.52
M

8 148 441 895 895 21 6 6 4 43.90

9 105 310 1400 1400 28 4 4 4 79.08
10 104 308 1575 1575 56 4 4 4 47.76
ED

11 107 318 2514 2514 91 4 4 4 45.86

12 105 310 2750 2750 84 4 4 4 54.86
13 125 368 907 907 28 5 5 4 51.90
PT

As discussed in Range, Lusby, & Larsen (2014), instances 1-6 benefit from greater patient-room
CE

compatibility when compared with the remaining instances. Finding feasible solutions for instances 7-12
is, therefore, a more complex task. Instance 13 reflects a multi-specialism approach, in which, during a
certain LoS, a patient will need to visit, and hence transfer between two specialisms. Additionally,
AC

instance 5 is characterized by the lowest bed occupancy rate among the first six instances, and hence
greatest scheduling flexibility. On the other hand, the last six instances are more heterogeneous, and the
highest bed occupancy rate is associated with instance 9. We note that, although higher occupancy rates
correlate with difficulty in finding a solution, the number of patients and rooms (or beds) that must be
allocated directly impacts solution time, given that the PASP is NP-hard. Moreover, it is worth
mentioning that the total number of patients includes elective patients as well as patients whose LoS is
 ACCEPTED MANUSCRIPT

zero. The former require hospitalization, while the latter need not be scheduled. Patients whose discharge
date lies beyond the planning horizon are scheduled until the last planning day.

Our model was implemented and solved using AIMMS 4.4 and GUROBI 7.5 with default
settings. Computational tests were performed on an Intel i7 3.3 GHz 64GB RAM computer with a run
time limit of 24h. In order to assess the advantages of warm starting the complete model with an integer
solution from the simplified model, two sets of results were generated. The first was obtained by solving

T
the complete model with a run time limit of 24 h given no initial solution. The second was obtained by

IP
solving the simplified model with a run time limit of 12 h (or until convergence, provided that an optimal
solution could be found in less than 12h) and then using its integer solution as a warm start to the

CR
complete model. The total run time limit of the warm start procedure was also set to 24h, meaning that in
this case the complete model was executed for 12h + any remaining amount of time not required for
convergence by the simplified model (results from the simplified model are presented in Table A of the

US
Appendix). The choice of a 12h run limit for the simplified model was based on the analysis of this
model’s convergence rate. When executing the simplified model for 24h, we verified that instances 1, 5, 6
AN
and 7 converged reasonably fast ( 4h), while more computational time was required to decrease the
optimality gap of all other instances. Moreover, the rate of convergence of these remaining instances
declined considerably after 12h of execution, when compared to the first hours of run time. In light of
M

this, we chose to limit the execution time of the simplified model to 12h or until zero gap was achieved,
whichever occurred first.
ED

For the purposes of comparing our results with those obtained in previous works, we adjusted the
reported processing times taking into account the characteristics of the CPU, following the approach from
Da Silva et al. (2012). Estimates of computer clocks for different CPU makes and models were obtained
PT

online2 and used to compute average, minimum and maximum estimated clock speeds (Table 6). These
estimates were subsequently used to adjust run times reported in Demeester et al. (2010), Ceschia and
CE

Schaerf (2011), Bilgin et al. (2012), Range, Lusby, & Larsen (2014), Hammouri &Alfrai (2014), and
Turhan & Bilgin (2017).
AC

2
https://asteroidsathome.net/boinc/cpu_list.php.
 ACCEPTED MANUSCRIPT

Table 6: Estimated clock speeds

Estimated Clock Speed (GFLOPs)

Reference Processor
Average Min Max
Demeester et al. (2010) Intel Quad Core PC/4GB RAM 4.98 2 8.13
Ceschia & Schaerf (2011) Intel Quad Core PC 4.98 2 8.13
Bilgin et al. (2012) Intel Core 2 Duo 3GHz 6.52 6.7 6.34
Range et al. (2014) Intel i7 CPU/8GB RAM 27.33 9.75 97.78

T
Hammouri & Alfrai (2014) Intel i3-4130 3.4GHz/4GB RAM 13.11 12.05 14.18
Turhan & Bilgin (2017) Intel i3 2.27GHz/3GB RAM 8.22 8.18 8.26

IP
Proposed model Intel i7-3960X 3.3Ghz/64GB RAM 44.14 – –

CR
Table 7 contrasts existing best known solutions with our MIP results. Under header “Literature
Results”, we present average solution times (adjusted following the procedure detailed previously)

generated by our complete model with

US
associated with existing best known solutions for each of the 13 benchmark instances. We report on the
results obtained using our complete model without warm start under header “MIP”, and show results
warm start under header “Warm Start MIP”.
AN
A number of points are worthy of discussion. First, using an initial solution (generated by the
simplified model) allowed for faster convergence of instances 1 and 5, when compared to the complete
M

model without warm start. In the case of instance 6, an optimal solution for the MIP was found, within the
defined run time limit, only when the warm start procedure was used. Moreover, the optimality gap was
ED

significantly reduced when instances 2, 3, 4 and 13 were solved using the warm start procedure and better
solutions were obtained for instances 8 and 10 when an initial solution was given vs. solving the MIP
without warm start. Additionally, warm starting the MIP improved the solution to instance 9 and allowed
PT

for obtaining integer solutions for instances 11 and 12. Although convergence was not achieved through
the warm start procedure for instance 7, the integer solution obtained was the same as the optimal solution
CE

generated without warm start. Taken together, these results indicate that warm starting the complete
model with the solution generated by the simplified model is an advantageous approach to solving the
MIP formulation.
AC

Second, we generated new best found solutions for 9 out of the 13 tested benchmark instances
(note that solutions obtained for instances 5 and 6 were the same as best known solutions reported in
literature and, for this reason, we do not consider them here as new best found solutions; nevertheless,
they were shown to be optimal by our MIP formulation). Furthermore, optimality of solution was also
proven for instances 1 and 7.
 ACCEPTED MANUSCRIPT

Third, although we report considerably higher run times than previous best known solutions, the
method we propose is parameter-free and does not require pre-processing of violation penalties. This
contrasts with the approach from Range, Lusby, & Larsen (2014), which required that a suite of
parameters be calibrated on a case-by-case basis and for which no dominant parameter setting has been
defined. Moreover, the simplified formulation from Ceschia and Schaerf (2011) relied on a pre-
processing step in which a patient-room penalty matrix was computed using a brute force approach.
However, there is no mention of the time spent on such pre-processing, nor is this time accounted for in

T
the simulation times reported. Additionally, the problem we address is a static offline variant of the PASP

IP
for which relatively large computational time does not constitute an impediment. Of note, our model
generates patient-room assignments that must be post-processed to yield patient-bed solutions. Our post-

CR
processing times were no greater than 2s, a negligible value when compared with the run times we report.
Therefore, similarly to previous works in which this conversion was necessary, we do not account for the

US
time required for postprocessing our solution. Finally, we note that we were not able to improve upon the
best known solutions for instances 11 and 12 within a run time limit of 24h. We attribute this to the fact
that these instances consider a significantly higher number of patients (of the order of 2-5 times higher
AN
than other instances) as well as more planning periods than the remaining test instances, which, as
discussed previously, are the main factors that influence our model’s performance.
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

T
IP
Table 7: Comparison between best known solutions and MIP results (new best found solutions in bold)

Literature Results MIP Warm Start MIP

Time* Time IP Lower GAP Time IP Lower GAP

CR
Instance Reference BKS**
(s) (s) Solution Bound (LP) (%) (s) Solution Bound (LP) (%)
1 Range et al. (2014) 368.65 654.40 77349 651.20 651.20 0.00 41437 651.20 651.20 0.00
2 Range et al. (2014) 515.74 1130.40 86400 1272.80 1115.80 14.07 86400 1128.00 1111.60 1.48
3 Range et al. (2014) 439.40 768.20 86400 768.00 753.00 1.99 86400 761.60 758.60 0.40
4
5
6
7
Ceschia & Schaerf (2011)
Range et al. (2014)
Range et al. (2014)
Range et al. (2014)
8043.33
156.36
276.56
153.79
1172.20
624.00
792.60
1193.00
86400
10277
86400
12121
1168.40
624.00
796.40
1176.40
1142.05

US
624.00
789.40
1176.40
2.31
0.00
0.89
0.00
86400
8251
19683
86400
1151.60
624.00
792.60
1176.40
1143.20
624.00
792.60
1175.20
0.73
0.00
0.00
0.10
AN
8 Range et al. (2014) 5222.28 4149.80 86400 4065.40 4024.41 1.02 86400 4063.00 4023.09 0.99
9 Ceschia & Schaerf (2011) > 2h30 21501.80 86400 196756.20 19632.00 > 100 86400 20904.60 19621.73 6.54
10 Ceschia & Schaerf (2011) 6476.29 8036.20 86400 7915.40 7687.33 2.97 86400 7830.40 7676.04 2.01
11 Ceschia & Schaerf (2011) > 2h30 11811.80 86400 - 10987.72 - 86400 11932.00 10914.57 9.32
12 Ceschia & Schaerf (2011) > 2h30 23344.20 86400 - 21886.60 - 86400 24198.40 21600.30 12.03
M

13 Ceschia & Schaerf (2011) 1849.94 9340.80 86400 9724.40 8842.80 9.97 86400 9114.40 8842.11 3.08
*
Solution time reported by the corresponding reference, adjusted following the procedure from Da Silva et al. (2012)
**
BKS – Best Known Solution in literature, as reported by the corresponding reference
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Given that the objective function in (1) consists of multiple cost criteria, we provide a breakdown
of the cost of the Warm Start MIP solutions into different objective components. Table 8 presents the cost
associated with each instance, along with the penalty values associated with different constraint
violations. The penalty for failing to assign a patient to a room with his/her preferred capacity is shown in
column “Room Preference”. The next three columns display the penalty incurred by not assigning a
patient to the appropriate department, not accounting for the prioritized specialism, and not attending to
the preferred treatment properties, respectively. Columns “Age” and “Gender” present the penalty

T
associated with age policy violations and gender policy violations, respectively, while the last column

IP
contains the penalty due to transfers.

CR
For instances 1-6, most penalties result from not being able to satisfy room capacity preferences
and, in some cases, specialism and room properties preferences also contribute to the overall cost; a
similar trend was also reported by Range, Lusby & Larsen (2014). This tendency is also verified for

US
instance 7, but department violations appear as well. Instances 8-13 display a different pattern, where
preferred room properties violations make up for most of the cost, and department, specialism, and
AN
preferred room capacity violations are consistently verified. Moreover, it is interesting to note that age
policy violations account for a considerable portion of the cost for instance 9, and gender violations
appear for instances 9, 12 and 13. Finally, we note that transfer violations were only reported in instances
M

8 and 13.

Table 8: Breakdown of the cost components for the Warm Start MIP solutions
ED

Instance Cost Room Pref. Department Specialism Room Prop. Age Gender Transfer
1 651.2 651.2 0.0 0.0 0.0 0.0 0.0 0.0
PT

2 1128.0 1116.0 0.0 12.0 0.0 0.0 0.0 0.0

3 761.6 753.6 0.0 0.0 8.0 0.0 0.0 0.0
4 1151.6 1041.6 0.0 70.0 40.0 0.0 0.0 0.0
CE

5 624.0 624.0 0.0 0.0 0.0 0.0 0.0 0.0

6 792.6 789.6 0.0 3.0 0.0 0.0 0.0 0.0
7 1176.4 726.4 20.0 158.0 272.0 0.0 0.0 0.0
AC

8 4063.0 1440.0 215.0 891.0 1495.0 0.0 0.0 22.0

9 20904.6 2697.6 300.0 1170.0 11912.0 4770.0 55.0 0.0
10 7830.4 2970.4 2.0 483.0 4375.0 0.0 0.0 0.0
11 11932.0 4352.0 27.0 1001.0 6552.0 0.0 0.0 0.0
12 24198.4 4778.4 400.0 2046.0 16909.0 0.0 65.0 0.0
13 9114.4 2090.4 663.0 1761.0 4475.0 10.0 5.0 110.0
 ACCEPTED MANUSCRIPT

5. Conclusion

Among the many challenges involved in efficient healthcare resource planning, the Patient
Admission Scheduling Problem (PASP) is of particular significance, impacting organizational decisions
at all planning levels. In this paper we focus on the static offline operational PASP for which we propose
a comprehensive mixed integer programming formulation and advance an exact solution method. Unlike
existing approaches, which typically use a simplified version of the allocation model and commonly

T
apply the algorithm from Ceschia & Schaerf (2011) to pre-process the instances, our formulation accounts

IP
for all restrictions from the original model of Demeester et al. (2010). Moreover, unlike heuristic and
metaheuristic-based methods, in which run time is used as stopping criteria, our exact method is capable

CR
of generating optimal solutions. We generated new best found solutions for 9 out of the 13 tested
benchmark instances. Additionally, we obtained solutions for two instances that were equal to best known
solutions reported in literature and also showed them to be optimal. To the best of our knowledge, ours is

US
the first work to provide solutions for the static offline operational PASP with guaranteed optimality.

Of note, the method we propose is parameter-free and does not require pre-processing of
AN
violation penalties. Moreover, we note that ours is the third work, following Demeester et al. (2010) and
Ceschia & Schaerf (2011), to report solving instances 9-12, which are characterized by a significantly
higher number of patients and considerably longer planning periods, as well as instance 13, which
M

includes multi-specialisms; all other existing approaches focused exclusively on instances 1-6 or 1-8. Our
ongoing work incudes refining our model to deal with larger instances. In addition, developing and
ED

incorporating techniques that improve MIP performance while guaranteeing optimality are open
challenges that should be addressed by future research in the field.
PT
CE
AC
 ACCEPTED MANUSCRIPT

References

Adan. I. J. B. F.. & Vissers. J. M. H. (2002). Patient mix optimisation in hospital admission planning: a
case study. International Journal of Operations & Production Management. 22(4). 445–461.

Bair. A. E.. Song. W. T.. Chen. Y.-C.. & Morris. B. A. (2010). The impact of inpatient boarding on ED
efficiency: A discrete-event simulation study. Journal of Medical Systems. 34(5). 919–929.

T
Barz. C.. & Rajaram. K. (2015). Elective patient admission and scheduling under multiple resource

IP
constraints. Production and Operations Management. 24(12). 1907–1930.

Bayer. R.. & Galea. S. (2015). Public health in the precision-medicine era. New England Journal of

CR
Medicine. 373. 499–501.

Bilgin. B.. Demeester. P.. Misir. M.. Vancroonenburg. W.. & Berghe. G. V. (2012). One hyper-heuristic

US
approach to two timetabling problems in health care. Journal of Heuristics. 18(3). 401–434.

Ceschia. S.. & Schaerf. A. (2011). Local search and lower bounds for the patient admission scheduling
AN
problem. Computers & Operations Research. 38(10). 1452–1463.

Ceschia. S.. & Schaerf. A. (2012). Modeling and solving the dynamic patient admission scheduling
M

problem under uncertainty. Artificial Intelligence in Medicine. 56(3). 199–205.

Ceschia. S.. & Schaerf. A. (2016). Dynamic patient admission scheduling with operating room
ED

constraints. flexible horizons. and patient delays. Journal of Scheduling. 19(4). 377–389.

Chen. N.. Zhan. Z.-H.. Zhang. J.. Liu. O.. & Liu. H.-L. (2010). A genetic algorithm for the optimization
PT

of admission scheduling strategy in hospitals. In Proceedings of the 2010 IEEE Congress on Evolutionary
Computation. CEC. Barcelona. Spain. July 18-23. 2010 (pp. 1–5).
CE

Da Silva. G.C.. Bahiense. L.. Ochi. L.S.. & Boaventura-Netto. P.O. (2012). The dynamic space allocation
problem: applying hybrid GRASP and tabu search metaheuristics. Computers & Operations Research.
AC

39(3). 671-677.

Demeester. P.. Souffriau. W.. De Causmaecker. P.. & Berghe. G. V. (2010). A hybrid tabu search
algorithm for automatically assigning patients to beds. Artificial Intelligence in Medicine. 48(1). 61–70.

Frieden. T. R. (2015). The future of public health. New England Journal of Medicine. 373. 1748–1754.
 ACCEPTED MANUSCRIPT

Gallivan. S.. Utley. M.. Treasure. T.. & Valencia. O. (2002). Booked inpatient admissions and hospital
capacity: mathematical modelling study. BMJ. 324(7332). 280–282.

Hammouri. A. I.. & Alrifai. B. (2014). Investigating biogeography-based optimisation for patient
admission scheduling problems. Journal of Theoretical & Applied Information Technology. 70(3). 413–
421.

Harper. P. R. (2002). A framework for operational modelling of hospital resources. Health Care

T
Management Science. 5(3). 165–173.

IP
Holm. L. B.. Luras. H.. & Dahl. F. A. (2013). Improving hospital bed utilisation through simulation and

CR
optimisation: With application to a 40% increase in patient volume in a Norwegian general hospital.
International Journal of Medical Informatics. 82(2). 80–89.

US
Jittamai. P.. & Kangwansura. T. (2011). A hospital admission planning model for emergency and elective
patients under stochastic resource requirements and no-shows. In Proceedings of the 2011 IEEE
International Conference on Industrial Engineering and Engineering Management. IEEM. Singapore.
AN
December 6-9. 2011 (pp. 166–170).

Kifah. S.. & Abdullah. S. (2015). An adaptive non-linear great deluge algorithm for the patient-admission
M

problem. Information Sciences. 295. 573–585.

Kusters. R. J.. & Groot. P. M. A. (1996). Modelling resource availability in general hospitals: Design and
ED

implementation of a decision support model. European Journal of Operational Research. 88(3). 428–445.

Lee. C. W.. & Kwak. N. K. (2011). Strategic enterprise resource planning in a health-care system using a
PT

multicriteria decision-making model. Journal of Medical Systems. 35(2). 265–275.

Merrild. P. (2015). The biggest U.S. health care challenges are management. Harvard Business Review.
CE

https://hbr.org/2015/02/the-biggest-us-health-care-challengesare-management-challenges. Accessed
08.23.17.
AC

Nunes. L. G. N.. de Carvalho. S. V.. & Rodrigues. R. C. M. (2009). Markov decision process applied to
the control of hospital elective admissions. Artificial Intelligence in Medicine. 47(2). 159–171.

Porter. M. E.. & Lee. T. H. (2013). The strategy that will fix health care. Harvard Business Review.
https://hbr.org/2013/10/the-strategy-that-will-fix-health-care. Accessed 08.23.2017.
 ACCEPTED MANUSCRIPT

Range. T. M.. Lusby. R. M.. & Larsen. J. (2014). A column generation approach for solving the patient
admission scheduling problem. European Journal of Operational Research. 235(1). 252–264.

Riley. J. C. (2005). Estimates of regional and global life expectancy. 1800-2001. Population and
Development Review. 31(3). 537–543.

Turhan. A. M.. & Bilgen. B. (2017). Mixed integer programming based heuristics for the Patient
Admission Scheduling problem. Computers & Operations Research. 80. 38–49.

T
IP
Vancroonenburg. W.. Goossens. D.. & Spieksma. F. C. R. (2011). On the complexity of the patient
assignment problem. Technical Report. KAHO Sint-Lieven. Gent. Belgium.

CR
Vissers. J. M. H.. Adan. I. J. B. F.. & Dellaert. N. P. (2007). Developing a platform for comparison of
hospital admission systems: An illustration. European Journal of Operational Research. 180(3). 1290–
1301.
US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Appendix

Table A – Results from the simplified model (MIP without Transfer) within a run time limit of 12h

Literature Results MIP without Transfer

Instance Reference Time (s) BKS Time (s) IP Solution Lower Bound (LP) GAP (%)
1 Range et al. (2014) 368.65 654.40 27615 651.20 651.00 0.03
2 Range et al. (2014) 515.74 1130.40 43200 1128.00 1123.00 0.45

T
3 Range et al. (2014) 439.40 768.20 43200 761.60 759.40 0.29
4 Ceschia & Schaerf (2011) 8043.33 1172.20 86400 1151.60 1145.80 0.51

IP
5 Range et al. (2014) 156.36 624.00 3398 624.00 624.00 0.00
6 Range et al. (2014) 276.56 792.60 8766 792.60 792.60 0.00

CR
7 Range et al. (2014) 153.79 1193.00 264 1176.40 1176.40 0.00
8 Range et al. (2014) 5222.28 4149.80 43200 4066.60 4047.76 0.47
9 Ceschia & Schaerf (2011) > 2h30 21501.80 43200 20904.60 20222.85 3.37
10
11
12
Ceschia & Schaerf (2011) 6476.29 8036.20
Ceschia & Schaerf (2011) > 2h30 11811.80
Ceschia & Schaerf (2011) > 2h30 23344.20
43200
43200
43200
US
7831.40
11932.00
24198.40
7714.40
10987.72
21886.60
1.52
8.59
10.56
AN
13 Ceschia & Schaerf (2011) 1849.94 9340.80 43200 9146.40 9012.14 1.49
*
Solution time reported by the corresponding reference, adjusted following the procedure from Da Silva et al.
(2012)
**
BKS – Best Known Solution in literature, as reported by the corresponding reference
M
ED
PT
CE
AC