This may be the author’s version of a work that was submitted/accepted

for publication in the following source:

Burdett, Robert, Kozan, Erhan, Sinnott, Michael, Cook, David, & Tian,
Glen
(2017)
A mixed integer linear programing approach to perform hospital capacity
assessments.
Expert Systems with Applications, 77, pp. 170-188.

This file was downloaded from: https://eprints.qut.edu.au/98721/

c Consult author(s) regarding copyright matters

This work is covered by copyright. Unless the document is being made available under a
Creative Commons Licence, you must assume that re-use is limited to personal use and
that permission from the copyright owner must be obtained for all other uses. If the docu-
ment is available under a Creative Commons License (or other specified license) then refer
to the Licence for details of permitted re-use. It is a condition of access that users recog-
nise and abide by the legal requirements associated with these rights. If you believe that
this work infringes copyright please provide details by email to qut.copyright@qut.edu.au

License: Creative Commons: Attribution-Noncommercial-No Derivative

Works 4.0

Notice: Please note that this document may not be the Version of Record
(i.e. published version) of the work. Author manuscript versions (as Sub-
mitted for peer review or as Accepted for publication after peer review) can
be identified by an absence of publisher branding and/or typeset appear-
ance. If there is any doubt, please refer to the published source.

https://doi.org/10.1016/j.eswa.2017.01.050
A Mixed Integer Linear Programing Approach to Perform Hospital Capacity
Assessments
Robert L. Burdett1, Erhan Kozan2, Michael Sinnott3, David Cook4,Yu-Chu Tian5
1,2
School of Mathematical Sciences, Queensland University of Technology, GPO Box 2434, 2 George Street Brisbane Qld 4000, Australia
Princess Alexandra Hospital, 2 Ipswich Rd, Woolloongabba, Brisbane, Qld 4102, Australia
3,4
5
School of Electrical Engineering and Computer Science, Science and Engineering Faculty, Queensland University of Technology, GPO Box
2434, 2 George Street Brisbane Qld 4000, Australia

Abstract – An approach to perform a system wide analysis of hospital resources and capacity has
been developed. Embedded within an intelligent system it would provide planners and management
capability to strategically improve the efficiency of their hospitals today and a means to create more
efficient hospitals in the future. In theory this approach can help hospitals with a variety of capacity
planning and resource allocation activities. On a day to day basis it can be used to perform a variety
of important capacity querying activities. In addition it can be used to predict the future
performance of a hospital and the effect of structural and parametric changes within the hospital.
The approach consists of a mixed integer linear programming (MILP) model and a number of
advanced extensions. The MILP models can determine the maximum number of patients of each
type that can be treated within a given period of time or the time required to process a given cohort
of patients. A case study of a large public hospital has been performed to validate our approach.
Extensive numerical investigations successfully demonstrate the applicability of the approach to real
sized health care applications and the great potential for further research and development on this
topic.

Keywords: capacity analysis, theoretical capacity, health care, hospitals, hospital resource planning,
capacity querying

1. Introduction

1.1. Background

Hospitals are critical elements of health care systems. Analysing their capacity and productivity is a
very important topic (Akcali et al, 2006) as many people need to be treated and everyone who
attends a hospital undeniably expects a high level and quality of care, within the shortest period of
time. To provide the best care, hospitals must have sufficient physical capacity and be managed and
operated well. Evidently this is not a simple task (Utley and Worthington, 2012). For instance
hospitals are very large integrated systems. They operate twenty four hours per day, seven days a
week. They consist of many treatment spaces, staff, resources and medical equipment. Health care
activities may also change at short notice and take an uncertain amount of time, due to patient
factors, medical complications and comorbidities. Hospitals must equitably treat many different
patient types. As hospital resources are limited, access to their services is very competitive. The
competition between known elective patients and unknown acute patients is another significant
issue (Burdett and Kozan, 2015).
Capacity planning is a way in which to avoid shortfalls in service and to minimise expensive
overruns, each of which arise from capacity-activity mismatches. Without loss of generality the
challenge in capacity planning lies in matching the assets, resources and services available to the
intended demand in service provision. Another significant difficulty is the implementation of
adequate IT systems to facilitate capacity planning. The importance of information technology (IT)
systems in health care cannot be understated. IT systems have a direct impact on quality which in
turn permits operational efficiencies to be realised (Gholami et al (2015)). To be most useful,
capacity planning approaches must be embedded within an intelligent system (IS) that has the ability
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
to gather and analyse data and communicate with other systems. Decision makers must be able to
extract the correct information from hospital IT systems in order to run hospital capacity
assessments. In the past many hospitals have used Hospital Based Corporation Information System
(HBCIS), Operating Room Management Information System (ORMIS), and Emergency Department
information systems (EDIS). As demonstrated in this article, those systems are still sufficient to
facilitate hospital capacity assessments. Fortunately there are many new information systems being
developed for hospitals. These new integrated IT systems have the capability of recording and
manipulating far more information and will soon replace the many separate systems that have been
used in the past. The position and care requirement of every patient and every activity is now
recordable. That information can facilitate the application of more advanced capacity assessments.
There are many different “capacity queries” that need to be answered when performing
capacity planning. The following is a small sample of some high level queries that hospital managers
and planners might consider:

 What can and cannot be achieved with existing hospital resources?

 How does the integration and interaction of hospital resources with respect to the hospital’s
activity affect the overall efficiency and effectiveness?
 Is the hospital adequate for the purposes and activities that it was created for? If not, does there
need to be more capacity or should the system be utilized better, by improved planning and
scheduling.
 How well was the hospital utilised in a previous year?

To meet increased demands the costs of health care have increased in recent years and many
hospitals have become larger and more sophisticated (Qld Health 2014, 2015). To our knowledge
there are few scientific methods to quantify the capability of this very complex system and to justify
those expansions. A typical hospital today has hundreds of beds and treatment spaces and future
hospitals will inevitably need to be even larger. Upon investigation it is neither simple nor
straightforward for hospital staff and health care managers to comprehensively answer capacity
related questions. These difficulties greatly motivate this article. This technically challenging topic is
very important, contemporary and timely given the pressures and challenges placed upon hospitals
worldwide, i.e. for increased productivity.

1.2. Research Aims and Methodology

This article considers how an assessment of hospital capacity can be achieved. It directly extends the
previous research in Burdett and Kozan (2016) whereby multiple criteria are handled and a model
with fewer technical conditions and complexity is solved. This article also builds upon capacity
assessment and planning research, most notably from Kozan and Burdett (2005) and Burdett (2015).
Those articles however consider domains other than health care.
The main objective of hospital capacity assessment (HCA) is to determine what performance is
theoretically possible from a hospital and to provide some form of activity and capacity statements
relative to this. To achieve this, mixed integer linear programming models are developed. They can
be used for capacity analysis and capacity querying purposes as part of an expert system. The exact
details of those models can be found in Section 4. How they are used for capacity querying is then
discussed in Section 5. A sensitivity analysis of hospital capacity is further discussed in Section 6 and
is warranted given the stochastic nature of treatment times and lengths of stay.
Our approach to hospital capacity analysis includes the recovery wards, operating theatres,
intensive care unit and the emergency department of a hospital; hence it is quite comprehensive
and holistic. The mathematical models of this article have been explicitly formulated to help answer
the following questions:

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
 i. What number of patients of each type can be treated within a specified time?
ii. Is there sufficient capacity to process and treat an intended number of patients of particular
types within a specified time period?
iii. Is there unused capacity available to treat an additional number of patients?
iv. How should the hospital be expanded to meet future demands?

These questions constitute low level capacity queries. The first question reflects the need to identify
a hospitals capacity generally. The second and third determines whether the system has sufficient
capacity for a specific intention. The second question analyses a given demand which could be
widely optimistic or pessimistic, if not anywhere in between. The third question however is different
as it involves the analysis of extra demand placed above the current capacity, as determined by the
first question. It should be noted that in ii) the patients on the current elective patient waiting list
may be used. The elective waiting list can be used to forecast the percentage mix of patients and
treatment types for the coming time period. The fourth question is most difficult and controversial
as the focus is how to increase the system’s capacity from the multitude of avenues available.
This article’s approach provides an upper bound on the productivity of a hospital and its
capacity to provide treatments and care. It can be viewed as an estimate of the operational capacity
of a hospital. Our upper bound denoted by 𝔸 is a measure of theoretical capacity because it
describes the best possible performance. The proposed approach does not presently consider
operational factors (i.e. staffing and rostering) that can degrade the systems performance. In reality
detailed planning and scheduling is necessary to optimize the hospital’s performance and to
maximize utilization. From these plans and schedules the operational capacity can be exactly
determined. As those tasks are computationally difficult (i.e. non-deterministic polynomial-time (NP)
hard), they are not yet advocated for capacity analysis (Burdett and Kozan, 2015).
The aforementioned upper bound is a measure of the systems structural attributes and is a
worthwhile reference point to compare other approaches - for example those that determine
operational capacity. The results of simulation activities can be compared to this reference point.
The theoretical capacity can be used as a bound for detailed planning and scheduling. For instance it
can be used as a proxy to measure the gap. In our opinion determining one metric in isolation is less
beneficial. Both capacity metrics should be identified as jointly they provide evidence of the
possibility of improving the system via changes to operational practices and procedures. If both
operational and theoretical capacity values are relatively close then it can be inferred that without
capacity expansion, an increase in productivity is unachievable.
The logic behind our approach is based upon resource saturation principles. The MILP models
try to utilize each resource as much as possible. This means that the resources should be assigned
sufficient work to be continuously utilised over a specified period of time. Resources that can be
saturated restrict the system. As bottlenecks they restrict further output. The proposed capacity
analysis approach is based upon an optimization model. A by-product of solving this model is a high
level plan that describes how a hospital functions over time. The plan specifies the number of
patients of each type that can be processed (i.e. a patient case mix (PCM)) and describes the
resource assignments and resource utilisations needed to treat that case mix.
In the next section, prior research is discussed. In Section 3 the requirements for capacity
analysis and mathematical modelling are examined. As previously mentioned, the main technical
developments are presented in Section 4-6. In Section 7, a full case study is presented. Concluding
remarks are then made in Section 8.

2. Literature Review and Discussion

In this section we review previous research on the topic of capacity analysis and capacity planning in
hospitals. Hospital capacity assessment is a topic that has received comparatively little attention in
the literature, particularly from a mathematical perspective. We have found that there are few well
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
documented capacity assessment approaches for hospitals and that the definition of hospital
capacity is often unclear. Capacity planning is a related topic that has been considered more
frequently. It is a different topic to capacity assessment in the sense that it seeks to determine the
number of resources that are required to meet notional levels of demand. Capacity assessment in
contrast considers what can be achieved with specified resource levels. There are a number of
models to determine for instance how many beds are required for notional levels of demand (Utley
and Worthington, 2012). In addition there are models to determine how to manage demands
exceeding agreed levels of capacity.
A good source of information on the issues involved in capacity planning in hospitals is the
article by Green (2004). In that article examples of how Operations Research (OR) models can be
used have been demonstrated and future research opportunities and challenges have been
identified. Rechel et al (2010) has also reviewed hospital capacity planning recently. They found that
there is a trend towards planning based on service volume and activity, rather than upon bed
capacity. Furthermore departments seeking to optimize their own functioning, without considering
how this affects the performance of others are a major cause of bottlenecks in hospitals. In their
opinion Diagnostic Related Groups (DRG) are not an appropriate methodology for capacity planning.
In addition anything that eases throughput by releasing bottlenecks, potentially adds value to the
system. Their article provides evidence of the need to focus on care pathways when designing and
constructing health-care facilities. This viewpoint motivates the approach taken in this article. Care
pathways seem to be a promising way of conceptualizing hospital capacity, but they are a
methodology that needs to be further developed. Challenges include the systematization of care
pathways, their large number and propensity to change, the integration of health-care demand and
supply, and the linking of resources to care pathways. Other challenges include the lack of
consistency and evidence for particular path details, often being locally derived and dependent on
local conditions, and the propensity for care paths to change mid path due to changes in patient
characteristics, and availability of components (i.e. imaging, appointments delays).
The article by Roth and Van Dierdonck (1995) is noteworthy as they developed a detailed
blueprint for performing hospital wide resource planning (HRP). Their HRP approach is essentially an
operation planning and control system and is based upon material requirements planning (MRP)
theory. In their approach bills of material (BOM) are utilised and only critical resources are modelled.
BOM are required for patients within each diagnostic related group (DRG). Additionally the expected
resource utilization times are needed. They demonstrated that this information can in theory be
obtained. As there are 470 categories this is a large job. They advocate that free capacity be
reserved as a buffer for emergencies. It is important to remember that DRG is a means of classifying
patient cases with similar resource and care requirements (i.e. patient care load). They are also used
as a measure of the products a hospital. Consequently they have been used for billing and
reimbursement purposes. For example DRG’s allow standardized diagnostic and treatment charges
to be defined. The weakness of DRG’s is that there may be significant diversity and variation within
individual categories; hence a small number of cases are disproportionately responsible for cost and
bed occupancy within any DRG, work unit or specialty area, and the average resource consumption
and length of stay is not accurate. In that article, a data analysis has demonstrated this to be true.
The validity of using DRG’s is hence questionable, and evidently the use of another grouping
criterion seems to be necessary.
On a similar topic, Van Merode et al (2004) investigated whether enterprise resource planning
(ERP) systems can be used for health care purposes. They found that ERP is best used for
deterministic processes as a shortcoming of ERP is that it needs fixed patient routings. In addition
ERP is not well suited to hospital processes that are characterized by a high variability and reactive
decision-making and whose length of stay vary considerably among patients.
Kunzt et al (2007) considered how to measure the efficiency of a hospital. They proposed and
demonstrated an alternative metric for hospital capacity planning. Their metric explicitly
incorporates economic efficiency and differs from prior metrics that are based upon hospital
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
occupancy. Their approach seeks to avoid the weaknesses of ratio metrics, for instance, that focus
on a single input and output, and neglect other important factors. Ma and Demeulemeester (2013)
developed an integrated multi-level approach for hospital planning. Their approach consists of three
phases. In the first phase an optimal patient mix and volume are selected that brings the maxiumm
profit. Then bed capacity is reallocated and a master surgery schedule is created. In the third phase
simulation is performed to evaluate operational policies. Optimization models are developed to
faciltate the first two phases. The three stages are part of an iterative approach.
Queuing theory has recently been used for capacity planning in hospital departments because
of its ease of calculation, minimal data requirements, and ability to be delivered in spreadsheets
(Cochrane and Roche, 2009). Izady and Worthington (2012) considered staff planning in an
emergency department (ED). As demand varies over the course of a day, their iterative staffing
approach determines the minimal hour-by-hour levels of medical staff needed to meet a four hour
sojourn target. Doctors, emergency nurse, ECG technicians, lab technicians, radiologists, and nurses
are all included in the planning process. The ED was treated as a time dependent queuing network.
Their algorithm uses the square root staffing law, infinite server networks to compute the resources’
time dependent workloads, and simulation to come up with a good solution. Cochrane and Roche
(2009) also considered capacity planning in an ED. They derived and implemented (in a spreadsheet)
an open queuing network model in order to increase the ED’s capacity to treat patients. Their
methodology captures hospital-specific differences in patient acuity mix, arrival patterns and
volumes, and efficiencies of processes. In their approach, routing matrices have been derived to
capture the flow pattern of each patient type. They include routing probabilities only on those node-
pairs that are directly connected. Gorunescu et al (2002) proposed an alternative type of model for
capacity planning. Their approach utilizes queuing theory to optimise the use of hospital resources.
In particular the optimal number of beds required to maintain a specified level of patient delay is
determined. The proposed model determines mean bed occupancy and the probability that demand
is lost because beds are occupied.
Many papers have addressed how to schedule patient’s surgeries and or hospital stays in
order to optimally utilize hospital capacity. Recent example include Burdett and Kozan (2015),
Braubach et al (2014), Cappanera et al (2014), Fugener et al (2014), Gartner and Kolisch (2014),
Meskens et al (2013), Vijayakumar et al (2013), Carnes et al (2011), Chow et al (2011), Cardoen et al
(2010), Cardoen and Demeulemeester (2009), Fei et al (2009). It is important to note that some of
the aforementioned research can be used, at least in theory, for capacity analysis purposes.
However as those problems are NP-hard, they do they provide exact solutions of the underlying
mathematical problem. They are not suitable for repeated application, as part of a sensitivity
analysis of capacity.
Hospital utilization can be vastly improved if patient flows are improved. Past research has
also focussed upon finding deficiencies and bottleneck in patient flows. Recently Khanna et al (2012)
analysed the effect of patient flow parameters and discharge initiatives to the occupancy levels of
hospitals of various size and settings. Twenty three hospitals were analysed in their study. Their
conclusion is that there is a need for a system-wide effort to address bottlenecks in patient flow
through the hospital. This comment motivates the approaches developed in this article.
To perform capacity analysis and modelling, it is necessary for patients and their treatments to
be categorised appropriately. For quantifying hospital output, patient classification is particularly
important. The definition of patient type is often used as a primary indicator of the level of resources
needed and the resource consumption, to provide care for the patient. Managers require
homogenous measures of output in order to effectively perform planning. The absence of
homogenous measures hinders analysis. To make the most accurate statements and conclusions,
past research suggests that within categorisations, resource requirements, treatment times and
recovery times should be homogenous. This task has been found to be difficult as patients and the
specific details of their illness and treatment vary greatly. Hence some categorisations like the
international classification of disease (ICD) have too many categories, while others exhibit large
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
variation or heterogeneity (i.e. DRG). The use of the severity of illness (SVI) index has been
evaluated by Horn and Horn (1986a, 1986b) and Horn and Sharkey (1983). The SVI is used to
facilitate a severity adjusted DRG that reduces variation within DRG categories. Numerical
experimentations have demonstrated that the SVI is a better case mix grouping strategy and
produces groups that are more homogenous. Rosko (1988) critically reviewed patient classification
systems such as DRG, disease staging, SVI, and patient management category (PMC). They concluded
that although hospital patient classification systems are not perfect they have improved greatly.
Their analysis shows that DRGs are not very homogeneous with respect to resource consumption,
and although DRGs can be augmented with severity measures, too many patient categories result,
and these gains may not be worth it. Carpenter et al (1999) examined the effect of severity of illness
on cost, revenue, and profit. They found that severity has a negative effect on profitability and that
concerns about the impact of severity of illness on hospital finances may be well-founded. For
example they found that hospitals were not compensated for severity of illness, and financial losses
were incurred. Forthman et al (2005) reported that severity adjustment alone does not adequately
explain legitimate differences in charges, costs, and length of stays within DRGs. In their opinion
other patient attributes such as age and the number of complications and comorbidities managed
during the patient stay must be accounted for. Consequently they proposed and tested a new
method that includes severity, intensity and complexity (SIC). This new approach was shown to be
statistically sound and was better able to explain variation in medical practice.

3. Requirements for Capacity Modelling

In this section we present important information that is relevant to the formulation of our later
hospital capacity models (HCM). Hospital processes and the activities and treatments that patients
receive is reviewed. Important terminology is also introduced. For fast access and to ease
understanding, a full list of all parameters and decision variables can be found in Appendix A. It is
worth mentioning that numbered superscripts are used to differentiate between variables that
describe very similar quantities. This practice greatly simplifies and reduces the number of variables
that are required. The superscripts could be removed in most circumstances without confusion, but
are required for mathematical correctness. When reading this article’s equations the superscripts
can be overlooked and emphasis should be placed on the subscripts.

3.1. Hospital Processes and Terminology

Hospitals are complicated structures and vary from place to place. Without loss of generality
however all hospitals contain places where treatments and care are provided. In this article two
types of places are defined, namely hospital areas and treatment spaces. A hospital area is simply a
collection of treatment spaces. The set of hospital areas and treatment spaces is denoted
respectively by 𝑊 and 𝛱. It is important to note that most hospital areas are self-sufficient and have
their own administrative staff. The hospital areas and treatment spaces within the hospital are used
to perform particular types of activity and are visited by patients for medical and surgical activities.
Patients occupy beds, chairs or other furniture within treatment spaces. Hospital wards are a special
type of hospital area. They are assigned to each surgical unit for patients to recover in. They consist
of beds and spaces. As beds are relatively cheap, numerous and mobile, bed occupancy actually
refers to the occupancy of a bed and a space at the same time, i.e. a bed-space.
Most hospitals have a combination of medical and surgical units. Each hospital unit 𝑢 ∈ 𝑈
typically provides treatments for a specific medical or surgical specialisation (i.e. a specialty) 𝑠 ∈ 𝑆.
Surgical procedures are performed by surgical teams in operating rooms (a.k.a. theatres). Some
theatres are used for specific types of surgeries and others are configured more generally. The
operating theatres are managed centrally. This is required in order to regulate the competition
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
between different surgical specialties. Due to the nature of surgery and the availability of key staff
such as surgeons, blocks of time are assigned each day of the week and patients are assigned to slots
within those blocks. Each hospital unit has particular resources assigned to it and its own budget.
Hospitals utilize a variety of different resources to perform health care activities. The set of these
resources are defined by 𝑅. Every resource is of a particular type and therefore belongs to a
particular resource group 𝑔 ∈ 𝐺. The predominant resource groups are staff and medical
equipment. A hospital is hence formally described as follows: 𝐻 = (𝑊, 𝛱, 𝑈, 𝑅).
The patients who are treated at the hospital are referred to as patient cases. A patient whose
visit to the hospital is planned is generally referred to as an elective patient. The level of urgency for
those patients waiting for elective surgery is called the clinical urgency category. There are three
categories: urgent, semi urgent, not urgent. The waiting time limit of each is 30, 90 and 365 days
respectively. Patients who present themselves at the emergency department are commonly called
emergency patients. A patient who is admitted for an extended stay (i.e. overnight or longer than
24hr) is called an in-patient. Out-patients are those patients whose stay is short, i.e. 4-6 hours
roughly. They visit the hospital briefly for medical appointments and treatments, or day surgery, and
are not admitted.

3.2. Patient Types and Care Plans

For capacity modelling it is necessary to define detailed information about the activities patients
require and their respective processing times. This information is called a patient care plan (PCP).
PCP are a predetermined set of tasks for patients of a particular type 𝛾 ∈ 𝛤. A variety of PCP can be
defined for each patient type. The set of those PCP is denoted by 𝛹𝛾 . It should be noted that PCP 𝜓
for patient type 𝛾 is not related in any way to PCP 𝜓 for patient type 𝛾 ′ . A PCP 𝜓 ∈ 𝛹𝛾 is formally
defined as follows: 𝑝𝑙𝑎𝑛 = {(𝜙, 𝑢, 𝑡, 𝑟)|𝜙 ∈ 𝛷, 𝑢 ∈ 𝑈, 𝑡 ∈ ℝ}. Each tuple (𝜙, 𝑢, 𝑡, 𝑟) represents a
stage in the PCP and describes the activity type, the unit performing the activity, the time to perform
the activity and the set of resources required. A reference to the kth task in PCP 𝜓 for patient type 𝛾
is denoted by o𝛾,𝜓,𝑘 . The number of tasks in the PCP is denoted by 𝐾𝛾,𝜓 . The following sets are
particularly important in later models and greatly simplify mathematical descriptions: ℘ =
{(𝛾, 𝜓)|∀𝛾 ∈ Γ, ∀𝜓 ∈ 𝛹𝛾 } and ℘1 = {(𝛾, 𝜓, 𝑘)|∀𝛾 ∈ Γ, ∀𝜓 ∈ 𝛹𝛾 , ∀𝑘 ∈ [1, 𝐾𝛾,𝜓 ]}. These sets
describe different patient type, PCP and PCP stage pairings.
In each PCP it is assumed that the set of resources required for each PCP task can be
determined. The unit-activity tuple (𝑢, 𝜙) is used to describe the places where the activity can be
performed. It is assumed that each unit has specific areas to perform specific types of activities. This
4
is denoted by 𝑊𝑢,𝜙 . It is worth mentioning that the superscript is used to differentiate between this
set of hospital areas and other sets that will be utilized later in this article.
In principle it is possible to aggregate PCP activities if they are performed in the same hospital
area. However if different resource requirements are needed, it may not be best to do so.
The same PCP may occur multiple times within a patient category as long as the resource and
treatment times are different in each. This approach would be taken for example, when the
variation in processing times is large. In theory PCP may be used to directly characterise patients. For
example each unique PCP can be regarded as a specific patient category.
Valid PCP can be obtained in a variety of ways. An analysis of hospital records recorded in
proprietary software such as the Operating Room Management System (ORMIS), Emergency
Department Information System (EDIS) and Hospital Based Corporation Information System (HBCIS)
is needed to identify appropriate processing times. From observations and hospital visits, the most
plausible PCP tasks are as follows: administration and registration (ADM), assessment (ASM), pre opt
care (POC), surgery (OP), post op care and post anaesthesia care (PAC), recovery (REC), observation
(OBS), emergency care (EC), miscellaneous care (MISC), intensive care (IC), discharge (DIS),
resuscitation (RES), stabilisation (STB), medical care and medical treatment (MC), preliminary care

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
(PRC), transfer (TRF), imaging (IM), pathology (PAT), triage (TRI), rehabilitation (REH). Figure 1
describes typical precedence’s between those PCP tasks and possible patient flows for inpatients,
outpatients and emergency patients, for both medical and surgical interventions.
Many PCP can be defined from Figure 1. There are many additional sub tasks and sub
activities; these are performed within the aforementioned ones. That level of detail is unnecessary
for high level capacity analysis and planning where sub activities and sub tasks do not interact. To
facilitate task pre-emption in our capacity analysis approach, it is necessary to divide existing
activities and nodes into several tasks and nodes respectively. In the event that more or less detail is
needed, it is a simple matter to define other activities and to construct an alternative network
diagram. The validity of the network and the appropriateness of the activities used forthwith is the
responsibility of hospital planners and staff and is not the immediate concern of this article.

ADM POC REC

Inpatient &
Outpatient OP
(surgical)
IM DIS
Only
Inpatient
PAT IC PAC OBS
Only
Outpatient
RES STB TRF MC OBS

Emergency

OBS
ASM
ASM DIS

DIS PAT

TRI MC

ADM IM

ADM IM PAT
Inpatient & Outpatient (medical)

Fig.1. Treatment tasks and possible precedence’s for patients

3.3. Patient Case Mixes

Determining the mix of patients that use hospital resources is important for planning as this
information is used to quantify what resources are exactly needed and how their time is
apportioned. The mix of patients treated within the hospital is called the patient case mix (PCM).
This is a generic term. It can be used to describe the actual number of different types of patients (i.e.
𝛾 ∈ 𝛤) or the proportional number. It is important to note that each patient type has a specific
attribute that is different to other groups of patients. Here “type” refers to DRG, ICD, specialty or
something else. As there are many different patient types, there are many different PCM that can be
defined and identified.
2
A proportional PCM can be described by the introduction of two parameters, 𝜇𝛾1 and 𝜇𝛾,𝜓 . The
first parameter describes the mix of patient types, whereas the second describes the mix of PCP
2
within each patient category. It should be noted that ∑∀𝛾∈𝛤 𝜇𝛾1 = 1 and ∑𝜓∈𝛹𝛾 𝜇𝛾,𝜓 = 1 ∀𝛾 ∈ 𝛤.
𝑡𝑜𝑡 1 2
The proportion of patients with each PCP is hence: 𝜇𝛾,𝜓 = 𝜇𝛾 × 𝜇𝛾,𝜓 ∀(𝛾, 𝜓) ∈ ℘. These mixes can
be defined and altered as part of a “what-if “analysis. They could also be identified from historical
hospital information.
The elective patient waiting list describes which patients are waiting to access the hospital for
the treatment of diagnosed illnesses. Many patient mixes can be extracted and used for planning
purposes. The mix of future emergency patients is something that should also be determined. As
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
there is no emergency patient waiting list per se, this mix can only be predicted empirically and
statistically. When using this mix, it is assumed that past demands are an indication of future
demands, with the limitation that in practice this may rarely be true, particularly over short intervals
of time.

4. Capacity Formulations

In this section several MILP models are introduced. They identify the theoretical capacity of a
hospital. Different levels of detail and complexity are involved; hence there are several variants. The
models presented here rely upon the definition of patient types and patient care plans (PCP) as
previously discussed in Section 3. The mathematical models assign patients to each PCP within each
patient category as demonstrated in Figure 2.
Type 1. 𝛾 = 1. Capacity/Plan

𝜓=1 PCP 1.1 # patients

𝜓=2 PCP 1.2 # patients

𝜓 = 𝛹1 PCP 1.𝛹1 # patients

Type 2. 𝛾 = 2 Case Mix

𝜓=1 PCP 2.1 # patients

𝜓=2 PCP 2.2 # patients

𝜓 = 𝛹2 # patients
PCP 2.𝛹2

Figure 2. The assignment of patients to each PCP within each patient category

The theoretical capacity is defined as the total number of patients treated within a specified time
period 𝑇. Although this number is important to know, identifying the number of patients that can be
treated, of each type is equally if not more important. This information for instance is the case mix.
It is used to describe the resource consumption of the hospital and is an essential piece of
information that can be used in planning activities, i.e. like scheduling.
As previously discussed, there are different ways to characterise patients and their treatments
or illnesses. Each way, can result in a different mathematical formulation and a different level of
capacity. This is because different patient types have different resource requirements. In order for
our mathematical models to remain generic, the patient type specification is unrestricted. Hence
each of the aforementioned patient types can be facilitated by our approach.
The mathematical notation used including the parameters and decision variables are
described as they are needed. We reiterate that a complete description is listed in the Appendix. The
main hospital entities (i.e. care plans, units, treatment spaces, areas, wards, resources, patients,
patient types and resource types) have been previously described.

4.1. Hospital Capacity Model 1

The first mathematical model (HCM1) assigns patients to hospital areas and determines the number
2
of patients that can be treated of each type. An important decision variable in HCM1 is 𝑛𝛾,𝜓 . It refers
to the number of patients of type 𝛾 with PCP 𝜓. The primary decision variable of this model however
is denoted by 𝛼𝛾,𝜓,𝑘,𝑤 . It describes in which hospital area each stage of the PCP, namely 𝑜𝛾,𝜓,𝑘 , is
performed, and how many patients have that treatment. The exact details of HCM1 are shown
below:

Maximize 𝑍 = 𝔸 − ℧. 𝐶𝑀𝑉 (1)

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
Subject To
2
𝔸 = ∑∀(𝛾,𝜓)∈℘ 𝑛𝛾,𝜓 [Total number of patients treated] (2)
2
𝐶𝑀𝑉 = 𝛾∈𝛤 ∀(𝜓,𝜌)∈𝜇𝛾2|𝑛𝛾,𝜓 − 𝜌. 𝑛𝛾 | + (𝛾,𝜌)∈𝜇1 |𝑛1𝛾 − 𝜌. 𝔸| ≤ ℧ [Case mix violation]
∑ ∑ 1 ∑ (3)
̇
∑∀(𝛾,𝜓,𝑘)∈℘1 |𝑤∈𝑊 2 (𝛼𝛾,𝜓,𝑘,𝑤 . 𝑡𝛾,𝜓,𝑘,𝑤 ) ≤ 𝑇𝑤2 . |𝛱𝑤
2|
∀𝑤 ∈ 𝑊 [Ward utilisation] (4)
𝛾,𝜓,𝑘
1 2
𝛼𝛾,𝜓,𝑘,𝑤 = 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘ , ∀𝑤 ∈ 𝑊|𝑤 ∉ 𝑊𝛾,𝜓,𝑘 [No assignment] (5)
𝛼𝛾,𝜓,𝑘,𝑤 ≥ 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑤 ∈ 𝑊 [Positivity requirement] (6)
2
𝑛1𝛾 = ∑𝜓∈𝛹𝛾 (𝑛𝛾,𝜓 ) ∀𝛾 ∈ 𝛤 [Patients treated by type] (7)
2
𝑛𝛾,𝜓 = ∑𝑤∈𝑊(𝛼𝛾,𝜓,𝑘,𝑤 ) ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 [Comparative relationship] (8)

The objective function seeks to maximize the total number of patients treated. This total is
partitioned amongst each PCP. This model can cope with different time availabilities for each
hospital area. For example parameter 𝑇𝑤2 specifies how long (i.e. in hours) each area may be used
within the specified time period. Hence constraint (4) ensures that there is a total time of 𝑇𝑤2 . |𝛱𝑤 2|

hours to process (i.e. treat) patients in each area, when all spaces within each area can be used for
𝑇𝑤2 time. The second constraint is appropriate when an assignment of tasks to treatment spaces is
not explicitly considered. Constraint (5) removes redundant decisions from the problem. This
constraint specifies a non-assignment if area w is not an option for any task in the PCP. Constraint
(6) is a standard positivity requirement. Constraint (7) provides a link between the two main decision
2 ∗
variables. This equation is true because 𝑛𝛾,𝜓 = 𝑛𝛾,𝜓,𝑘 = ∑𝑤∈𝑊(𝛼𝛾,𝜓,𝑘,𝑤 ) ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 . For
2 ∗
example, the number of patients treated (i.e. 𝑛𝛾,𝜓 ) is distributed to each stage of the PCP (i.e. 𝑛𝛾,𝜓,𝑘 )
and must be distributed across the available areas in some way.

4.1.1. Book keeping

HCM1 has several presumptions. For example the patient’s length of stay (LOS) in each unit is not
2
affected by the hospital area that is assigned, i.e. 𝑡̇𝛾,𝜓,𝑘,𝑤 = 𝑡𝛾,𝜓,𝑘 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑤 ∈ 𝑊𝛾,𝜓,𝑘 . It
is also assumed that all spaces within a hospital area can be used to treat patients, with a given
activity, if that area has been defined as permissible. We remind the reader that each activity tuple
(𝜙, 𝑢) has a set of permissible areas and this is denoted by 𝑊𝑢,𝜙 4
. That set is an input and is then
2
used to define the set of permissible areas 𝑊𝛾,𝜓,𝑘 for 𝑜𝛾,𝜓,𝑘 . The exact details of the relationship
2 4
between 𝑊𝛾,𝜓,𝑘 and 𝑊𝑢,𝜙 can be found in the Appendix.
The solution of HCM1 facilitates the calculation of additional “book keeping” information that
may be useful to hospital staff. Equations (9) – (15) compute the number of patients involved in
various activities, locations and of different types:

𝑛𝑢3 = ∑𝛾∈𝛤(𝑛𝛾,𝑢
4
) ∀𝑢 ∈ 𝑈 [Patients treated by units] (9)
4 2
𝑛𝛾,𝑢 = ∑𝜓∈𝛹𝛾 |𝑢∈𝑈 1 (𝑛𝛾,𝜓 ) ∀𝑢 ∈ 𝑈, ∀𝛾 ∈ 𝛤 [Patients treated by unit by type] (10)
𝛾,𝜓
5
𝑛𝛾,𝜓,𝜙,𝑤 = ∑𝑘∈[1,𝐾𝛾,𝜓 ]|𝑤∈𝑊 2 𝛼𝛾,𝜓,𝑘,𝑤 ∀(𝛾, 𝜓) ∈ ℘, ∀𝜙 ∈ 𝛷, ∀𝑤 ∈ 𝑊 (11)
𝛾,𝜓,𝑘
6 5
𝑛𝛾,𝜙,𝑤 = ∑𝜓∈𝛹𝛾 𝑛𝛾,𝜓,𝜙,𝑤 ∀𝛾 ∈ 𝛤, ∀𝜙 ∈ 𝛷, ∀𝑤 ∈ 𝑊 (12)
7 6
𝑛𝜙,𝑤 = ∑𝛾∈𝛤 𝑛𝛾,𝜙,𝑤 ∀𝜙 ∈ 𝛷, ∀𝑤 ∈ 𝑊 (13)
8
𝑛𝛾,𝜓,𝑤 = ∑𝑘∈[1,𝐾𝛾,𝜓 ] 𝛼𝛾,𝜓,𝑘,𝑤 ∀(𝛾, 𝜓) ∈ ℘, ∀𝑤 ∈ 𝑊 [Number of activities performed] (14)
9 2
𝑛𝑠 = ∑∀(𝛾,𝜓)∈℘|𝜎𝛾,𝜓 =𝑠 𝑛𝛾,𝜓 ∀𝑠 ∈ 𝑆 [Patients treated of each specialty] (15)
𝜆1𝑤 = 100. ∑∀(𝛾,𝜓,𝑘)∈℘1 (𝛼𝛾,𝜓,𝑘,𝑤 . 𝑡̇𝛾,𝜓,𝑘,𝑤 ) /𝑇𝑤2 . |𝛱𝑤2|
∀𝑤 ∈ 𝑊 [Ward % utilisation] (16)
1 2 6 7 2 6 7
𝑛𝛾 , 𝑛𝛾,𝜓 , 𝑛𝑢 , 𝑛𝛾,𝑢 , 𝑛𝛾,𝜙,𝑤 , 𝑛𝜙,𝑤 ≥ 𝑑𝑒𝑚𝛾 , 𝑑𝑒𝑚𝛾,𝜓 , 𝑑𝑒𝑚𝑢3 , 𝑑𝑒𝑚𝛾,𝑢
3 4 1 4
, 𝑑𝑒𝑚𝛾,𝜙,𝑤 , 𝑑𝑒𝑚𝜙,𝑤 (17)
𝑛𝑠9 ≥ 𝜌. 𝑛EL ∀(𝑠, 𝜌) ∈ 𝜇3 |𝑠 ∈ 𝑆 EL [Mix of specialties within electives] (18)
EL EL e EL
𝑛𝛾 ≥ 𝜌. 𝑛 ∀(𝛾, 𝜌) ∈ 𝜇 |𝛾 ∈ 𝛤 [Mix of patient type within electives] (19)
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
 Constraint (16) computes the utilization level for each hospital area. Constraint (17) is a short-
hand for six alternative (i.e. optional) performance targets. Various selections may hence be chosen.
A variety of different proportional patient case mix may be used in HCM1. If patient case mix
requirements are not defined then the model will choose the mix itself, to the detriment of PCP with
higher lengths of stay. Constraints (18) and (19) enforce patient case mixes associated with elective
patients. The two main patient mixes however are computed in constraint (3). The absolute value
terms should be linearized in the usual way. The right hand side is optional. An alternative is to
penalise the objective function and that option is shown in equation (1). It should be noted that
2
patient mix 𝜇𝛾1 may or may not be defined. Similarly PCP mix 𝜇𝛾,𝜓 may not be defined or may be
defined for some patient types 𝛾 and not others. Hence those parameters are redefined using sets.
For example:

1
𝜇1 = ∅ or 𝜇1 = {(1, 𝜇11 ), (2, 𝜇21 ), … , (|𝛤|, 𝜇|𝛤| )} (21)
2 2 2
𝜇𝛾2 = ∅ or 𝜇𝛾2 = {(1, 𝜇𝛾,1 ), (2, 𝜇𝛾,2 ), … , (|𝛹𝛾 |, 𝜇𝛾,|𝛹𝛾|
)} (22)

Consequently the presence or absence of different mixes can be identified and when present
enforced. It should be noted that equation (3) and the objective function penalisation approach are
alternatives to the inclusion of the following constraints:

𝑛1𝛾 ≥ 𝜌. 𝔸 ∀(𝛾, 𝜌) ∈ 𝜇1 [The mix of patient type] (23)

2
𝑛𝛾,𝜓 ≥ 𝜌. 𝑛𝛾 ∀𝛾 ∈ 𝛤, ∀(𝜓, 𝜌) ∈ 𝜇𝛾2 [The PCP mix]
1
(24)

Extreme care should be taken when including mix constraints of the aforementioned type. The
regulation of patient treatments within categorisations is strict and the resulting capacity can be low
due to resource bottlenecking. For example, one group of PCP within a patient type category may
use a resource that is quite limited. Because those PCP are restricted, it means that the other PCP
may be restricted too because of the exact requirements described by the given mix, even though
resources are actually available.

4.2. Hospital Capacity Model 2

In this section the details of HCM2 are given. It is an extension of HCM1. HCM2 however is more
“fine grained” and consequently more accurate as it assigns patients to individual spaces within
hospital areas. Let the number of patients assigned to space 𝜋 for the kth stage of PCP (𝛾, 𝜓) be
2
𝛽𝛾,𝜓,𝑘,𝜋 . The set of spaces that can be used for each 𝑜𝛾,𝜓,𝑘 is defined as 𝛱𝛾,𝜓,𝑘 . This set is determined
4
from the mapping (𝜙, 𝑤) → 𝜋 that is given by set 𝛱𝜙,𝑤 . There is no requirement for each space to
be used for a single type of care activity using this approach. Each space is utilizable for a period of
time denoted by 𝑇𝜋3 . The following constraints are added to HCM1:

𝛽𝛾,𝜓,𝑘,𝜋 ≥ 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝜋 ∈ 𝛱 [Positivity requirement] (25)

2
𝛽𝛾,𝜓,𝑘,𝜋 = 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝜋 ∈ 𝛱|𝜋 ∉ 𝛱𝛾,𝜓,𝑘 [No assignment] (26)
∑∀(𝛾,𝜓,𝑘)∈℘1 |𝜋∈𝛱 2 (𝛽𝛾,𝜓,𝑘,𝜋 . 𝑡̈𝛾,𝜓,𝑘,𝜋 ) ≤ 𝑇𝜋3 ∀𝜋 ∈ 𝛱 [Space utilization limit] (27)
𝛾,𝜓,𝑘
𝛼𝛾,𝜓,𝑘,𝑤 = ∑𝜋∈𝛱𝑤3 | 𝛽𝛾,𝜓,𝑘,𝜋 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑤 ∈ 𝑊 [Comparative relationship] (28)

Constraint (27) is added to restrict the over utilisation of spaces. The addition of this
constraint however requires the removal of constraint (4) that restricts the utilization of hospital
areas. This is necessary for reasons of accuracy.

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
4.2.1 Book Keeping

The solution of HCM2 also facilitates the calculation of additional “book keeping” information. It is
important to calculate the following patient number calculations concerning spaces:

7
𝑛𝜙,𝑤 = ∑𝜋∈𝛱𝑤3 𝑛12
𝜙,𝜋 ∀𝜙 ∈ 𝛷, ∀𝑤 ∈ 𝑊 (29)
10
𝑛𝜋 = ∑∀(𝛾,𝜓,𝑘)∈℘1 |𝜋∈𝛱 2 (𝛽𝛾,𝜓,𝑘,𝜋 ) ∀𝜋 ∈ 𝛱 (30)
𝛾,𝜓,𝑘
𝑛11
𝛾,𝜋 = ∑∀(𝛾 ′ ,𝜓,𝑘)∈℘1 |𝜋∈𝛱 2 ′ 𝛽𝛾,𝜓,𝑘,𝜋 ∀𝜋 ∈ 𝛱, ∀𝛾 ∈ 𝛤 (31)
𝛾,𝜓,𝑘 ,𝛾=𝛾
𝑛12
𝜙′ ,𝜋 = ∑∀(𝛾,𝜓,𝑘)∈℘1 |𝜋∈𝛱 2 ′
′
𝛽𝛾,𝜓,𝑘,𝜋 ∀𝜋 ∈ 𝛱, ∀𝜙 ∈ 𝛷 (32)
𝛾,𝜓,𝑘 ,𝜙𝛾,𝜓,𝑘 =𝜙
𝜆2𝜋 = 100 ∑∀(𝛾,𝜓,𝑘)∈℘1 |𝜋∈𝛱 2 (𝛽𝛾,𝜓,𝑘,𝜋 . 𝑡̈𝛾,𝜓,𝑘,𝜋 ) /𝑇𝜋3 ∀𝜋 ∈ 𝛱 [Space % utilization] (33)
𝛾,𝜓,𝑘

12
The following relationships also hold: 𝑛10 11
𝜋 = ∑𝛾∈𝛤 𝑛𝛾,𝜋 = ∑𝜙∈𝛷 𝑛𝜙,𝜋 ∀𝜋 ∈ 𝛱. As previously
mentioned, the processing time of PCP tasks is the same regardless of the space the activity is
2
performed in, i.e. 𝑡̈𝛾,𝜓,𝑘,𝜋 = 𝑡𝛾,𝜓,𝑘 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝜋 ∈ 𝛱𝛾,𝜓,𝑘 . The space availability is also
3 2 3
assumed to be the availability of the hospital area, i.e. 𝑇𝜋 = 𝑇𝑤 ∀𝑤 ∈ 𝑊, ∀𝜋 ∈ 𝛱𝑤 .

4.3. Extensions

HCM1 and HCM2 can be further strengthened to reflect real life conditions. This section details a
number of improvements but at the expense of added complexity.

4.3.1 Unequal Task Durations

Different activity durations and lengths of stay may eventuate in real life depending on which
hospital area or treatment space is assigned. An assumption of equal durations was previously
made, but this assumption could be removed. For this situation, the model does not need to be
changed; only the input data. For example the values 𝑡̇𝛾,𝜓,𝑘,𝑤 and 𝑡̈𝛾,𝜓,𝑘,𝜋 must be defined
∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑤 ∈ 𝑊, ∀𝜋 ∈ 𝛱. These values could be input within the PCP tuple. For example
𝑡𝛾,𝜓,𝑘 could be re-defined as a vector of |𝑊| or |𝛱| values. As many areas and spaces are not
applicable this parameter is sparse. Consequently these values should to be defined separately and
𝑡𝛾,𝜓,𝑘 should be directly removed from the PCP tuple. Using a set based approach let ℘2 =
2
{(𝛾, 𝜓, 𝑘, 𝑤)|(𝛾, 𝜓, 𝑘) ∈ ℘1 , 𝑤 ∈ 𝑊𝛾,𝜓,𝑘 } and redefine parameter 𝑡𝛿̇ ∀𝛿 ∈ ℘2 . Similarly let
2
℘3 = {(𝛾, 𝜓, 𝑘, 𝜋)|(𝛾, 𝜓, 𝑘) ∈ ℘1 , 𝜋 ∈ 𝛱𝛾,𝜓,𝑘 } and redefine parameter 𝑡̈𝛿 ∀𝛿 ∈ ℘3 . It is necessary
to rewrite constraint (4) and (27) in the following way:

∑(𝛾,𝜓,𝑘,𝑤 ′ )∈℘2 |𝑤 ′ =𝑤(𝛼𝛾,𝜓,𝑘,𝑤 . 𝑡̇(𝛾,𝜓,𝑘,𝑤) ) ≤ 𝑇𝑤2 . |𝛱𝑤

2|
∀𝑤 ∈ 𝑊 (34)
3
∑(𝛾,𝜓,𝑘,𝜋′ )∈℘3 |𝜋′ =𝜋(𝛽𝛾,𝜓,𝑘,𝜋 . 𝑡̈(𝛾,𝜓,𝑘,𝜋) ) ≤ 𝑇𝜋 ∀𝜋 ∈ 𝛱 (35)

When an activity can be carried out equally well in a number of places, an argument can be made
that there is no difference between those places. Differences only occur because of other decision
making activities and operating conditions in the hospital and if so, these should not be present.

4.3.2. Multiple Time Periods

The aforementioned models determine the hospital’s capacity over a period of time T where each
hospital area and treatment space is utilisable for 𝑇𝑤2 and 𝑇𝜋3 time respectively. Another approach is
to perform an analysis of capacity over many different time periods. There are a few reasons for
doing this. For example there may be different demands in each time period or the hospital may be
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
configured differently, in each time period. Conceptually it is quite easy to facilitate an approach like
this by adding a time period subscript to all sets, parameters and decision variables. For example,
2
the main decision variables become: 𝑛𝛾,𝜓,𝑡 , 𝛼𝛾,𝜓,𝑘,𝑤,𝑡 , 𝛽𝛾,𝜓,𝑘,𝜋,𝑡 . The objective is hence the total
2
number of treatments over all time periods. i.e. 𝔸 = ∑𝑡 ∑∀(𝛾,𝜓)∈℘ 𝑛𝛾,𝜓,𝑡 .

4.3.3. Including Other Resources

Different resources are needed to provide health care. For example, there are many different types
of physicians and nurses and medical equipment. This section includes those resources and builds
upon the core models defined in earlier sections to provide a third model namely HCM3. The
resources modelled here are different from the physical spaces previously modelled. Their number
and availability can significantly restrict the hospitals capacity. For example they affect the previous
2
decisions 𝛼𝛾,𝜓,𝑘,𝑤 , 𝛽𝛾,𝜓,𝑘,𝜋 and 𝑛𝛾,𝜓 .
In regard to the availability of resources within the hospital, it should be noted that some
resources are fixed in place while others are mobile. The resources that are specifically assigned to
units, areas and spaces are denoted by the following sets: 𝐹𝑢1 ,𝐹𝑤2 ,𝐹𝜋3 . Hence 𝑅 = 𝐹𝑢1 ∪ 𝐹𝑤2 ∪ 𝐹𝜋3 . It is
assumed that resource assignments are hierarchical. This means that resources assigned to units are
automatically available to hospital areas assigned to that unit and to all spaces within those areas.
Similarly, resources assigned to hospital areas are available to all spaces within that area, but not to
other areas, even those assigned to the unit. Consequently the actual set of resources available to
units, areas and spaces is as follows: 𝑅𝑢1 = 𝐹𝑢1 ∀𝑢 ∈ 𝑈, 𝑅𝑤 2
= 𝐹𝑤2 ∪ (⋃𝑢∈𝑈𝑤 𝑅𝑢1 ) ∀𝑤 ∈ 𝑊,
𝑅𝜋3 = 𝐹𝜋3 ∪ 𝑅𝑤 2
∀𝜋 ∈ 𝛱, 𝑤 = 𝐿𝜋 . The location of treatment spaces is given by Lπ .
In order to identify the hospitals capacity, the resource requirements for each PCP are
required. For each PCP they are given by 𝑟̇𝛾,𝜓,𝑘,𝑔 . This is the number of resources of type g required
for 𝑜𝛾,𝜓,𝑘 . The time that resources of type g are required during activity 𝑜𝛾,𝜓,𝑘 is given by 𝑡𝛾,𝜓,𝑘,𝑔
where 𝑡𝛾,𝜓,𝑘,𝑔 ≤ 𝑡𝛾,𝜓,𝑘 . Each resource belongs to a specific resource group 𝑔 ∈ 𝐺. The set of
resources of type g is defined as 𝑅𝑔4 . Further sets may be defined to describe which resources of
5
each type are available to each unit, area and space. They are as follows: 𝑅𝑔,𝑢 = {𝑟 ∈ 𝑅𝑢1 |𝑟 ∈
𝑅𝑔4 }, 𝑅𝑔,𝑤
6
= {𝑟 ∈ 𝑅𝑤2
|𝑟 ∈ 𝑅𝑔4 }, 𝑅𝑔,𝜋
7
= {𝑟 ∈ 𝑅𝜋3 |𝑟 ∈ 𝑅𝑔4 }. For all 𝑟 ∈ 𝑅𝑔4 the time required is initially
assumed to be the same. Several additional sets are helpful and are now introduced. The set of
activities that require resources of type g is given by ℘5𝑔 = {(𝛾, 𝜓, 𝑘) ∈ ℘1 |𝑟̇𝛾,𝜓,𝑘,𝑔 > 0 ∨ 𝑡𝛾,𝜓,𝑘,𝑔 >
0}. The set of activities that may be performed by resource r, based purely on the resources type, is
also specified by ℘6𝑟 = (⋃∀𝑔∈𝐺|𝑟∈𝑅𝑔 ℘5𝑔 ). The set of resources available for 𝑜𝛾,𝜓,𝑘 is 𝑅𝛾,𝜓,𝑘 = {𝑟 ∈
𝑅|(𝛾, 𝜓, 𝑘) ∈ ℘6𝑟 }.
For capacity modelling, further detail can be provided concerning how each resource is to be
8
used. For instance the set of resources that perform activities of type 𝜙 can be denoted by 𝑅𝜙 .
Within units, areas and spaces, the following mappings can be provided: (𝑢, 𝜙) → 𝑟, (𝑤, 𝜙) → 𝑟,
4 5 6
(𝜋, 𝜙) → 𝑟. These mappings are given by the following sets: 𝐹𝑢,𝜙 ⊂ 𝑅𝑢1 , 𝐹𝑤,𝜙 2
⊂ 𝑅𝑤 , 𝐹𝜋,𝜙 ⊂ 𝑅𝜋3 .
Because of the aforementioned hierarchy, the actual set of resources available for each activity type
9 4 10 5 9 11 6 10
are as follows: 𝑅𝑢,𝜙 = 𝐹𝑢,𝜙 , 𝑅𝑤,𝜙 = 𝐹𝑤,𝜙 ∪ (⋃𝑢∈𝑈𝑤 𝑅𝑢,𝜙 ), 𝑅𝜋,𝜙 = 𝐹𝜋,𝜙 ∪ 𝑅𝑤,𝜙 ∀𝜋 ∈ 𝛱, 𝑤 = 𝐿𝜋 .
To include resources in the capacity models, a variety of different modelling approaches were
considered. The approach discussed and advocated here considers how to assign individual
resources to tasks. For example we assign resources to specific PCP tasks in the same way that
hospital areas and treatment spaces were previously. This is a detailed low level approach. We
believe this approach to be more accurate and robust than other strategy investigated during our
preliminary modelling. One of those involves the formulation of a model that decides how long
resources are specifically assigned to spaces, areas and activities. That approach is quite problematic
upon reflection. It would result in resources performing fractions of activities. It is better to decide
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
how many activities a resource will perform as this makes the time requirement transparent. Our
decision variable is 𝛿𝛾,𝜓,𝑘,𝑟 , the number of activities of type 𝑜𝛾,𝜓,𝑘 who are treated by resource r. The
following constraints are hence necessary:
∑(𝛾,𝜓,𝑘)∈℘6𝑟 (𝛿𝛾,𝜓,𝑘,𝑟 . 𝑡𝛾,𝜓,𝑘,𝑟 ) ≤ 𝑇𝑟4 ∀𝑟 ∈ 𝑅 [Resource time availability limit] (36)
2 5
∑𝑟∈𝑅𝑔4 𝛿𝛾,𝜓,𝑘,𝑟 = 𝑟̇𝛾,𝜓,𝑘,𝑔 . 𝑛𝛾,𝜓 ∀𝑔 ∈ 𝐺, ∀(𝛾, 𝜓, 𝑘) ∈ ℘𝑔 [Comparative relation] (37)
2
𝛿𝛾,𝜓,𝑘,𝑟 ≤ 𝑛𝛾,𝜓 ∀𝑟 ∈ 𝑅, ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 [Bound] (38)
1
𝛿𝛾,𝜓,𝑘,𝑟 = 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘ , ∀𝑟 ∈ 𝑅|𝑟 ∉ 𝑅𝛾,𝜓,𝑘 [Invalid assignment] (39)
9 9
𝛿𝛾,𝜓,𝑘,𝑟 = 0 ∀𝑜 = (𝛾, 𝜓, 𝑘) ∈ ℘1 ||𝑅𝜙 𝑜 ,𝑢𝑜
| > 0, ∀𝑟 ∈ 𝑅|𝑟 ∉ 𝑅𝜙 𝑜 ,𝑢𝑜
(40)
𝛿𝛾,𝜓,𝑘,𝑟 ≥ 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑟 ∈ 𝑅 (41)

Constraint (36) ensures that the total time that resource r is working is not greater than the time it is
available. Constraint (37) ensures that all activities are assigned their required resources. The third
constraint (38) limits the number of times resource r can be assigned to activity 𝑜𝛾,𝜓,𝑘 . The fourth
constraint (39) removes invalid assignments. Constraint (40) is activated only if a mapping exists. In
that constraint no assignment is possible if the resource works for the unit but the unit does not do
the activity, i.e. the resource is not working in unit 𝑢𝛾,𝜓,𝑘 . That constraint addresses the following
issue, “physician A from unit B is assigned a surgical operation that should be done by physicians
from unit C”. It should be noted that constraint (40) is not always needed. It is possible in the
aforementioned example to sub-divide physician into those of different specialties. Consequently
the aforementioned issue would vanish.
Constraints (36)-(39) do not take into account the geographical position of resource r. In
practice it is impossible to assign resource r to an activity if it is not assigned to the location of the
activity. It is also important to take into account the possibility that some resources are not the
same even within the same type. For example, some activities can only be performed by staff of a
given specialty. In light of these additional complexities a new decision variable is introduced. Let the
number of activities of type 𝑜𝛾,𝜓,𝑘 that resource r treats in space 𝜋 be 𝛿𝛾,𝜓,𝑘,𝑟,𝜋 . Given the
relationship 𝛿𝛾,𝜓,𝑘,𝑟 = ∑𝜋∈𝛱 𝛿𝛾,𝜓,𝑘,𝑟,𝜋 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑟 ∈ 𝑅, constraint (36)-(41) are rewritten in
the following way:

∑(𝛾,𝜓,𝑘)∈℘6𝑟 ∑∀𝜋∈𝛱(𝛿𝛾,𝜓,𝑘,𝑟,𝜋 . 𝑡𝛾,𝜓,𝑘,𝑟,𝜋 ) ≤ 𝑇𝑟4 ∀𝑟 ∈ 𝑅 [Resource time availability limit] (42)

∑𝑟∈𝑅𝑔4 𝛿𝛾,𝜓,𝑘,𝑟,𝜋 = 𝑟̇𝛾,𝜓,𝑘,𝑔 . 𝛽𝛾,𝜓,𝑘,𝜋 ∀𝑔 ∈ 𝐺, ∀𝜋 ∈ 𝛱, ∀(𝛾, 𝜓, 𝑘) ∈ ℘5𝑔 (43)
2
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 ≤ 𝑛𝛾,𝜓 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑟 ∈ 𝑅𝛾,𝜓,𝑘 , ∀𝜋 ∈ 𝛱 [Bound] (44)
2
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 = 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝑟 ∈ 𝑅, ∀𝜋 ∈ 𝛱|𝜋 ∉ 𝛱𝛾,𝜓,𝑘 (45)
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 = 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝜋 ∈ 𝛱, ∀𝑟 ∈ 𝑅|𝑟 ∉ 𝑅𝜋3 [For all fixed resources] (46)
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 = 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1 , ∀𝜋 ∈ 𝛱, ∀𝑟 ∈ 𝑅|(𝛾, 𝜓, 𝑘) ∉ ℘6𝑟 (47)
9 9
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 = 0 ∀𝑜 = (𝛾, 𝜓, 𝑘) ∈ ℘1 ||𝑅𝜙 𝑜 ,𝑢𝑜
| > 0, ∀𝜋 ∈ 𝛱, ∀𝑟 ∈ 𝑅|𝑟 ∉ 𝑅𝜙 𝑜 ,𝑢𝑜
(48)
1
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 ≥ 0 ∀(𝛾, 𝜓, 𝑘) ∈ ℘ , ∀𝜋 ∈ 𝛱, ∀𝑟 ∈ 𝑅 (49)

Limitations on the time available to use different resources, is ensured by constraint (42). Each
activity requires a specific number of resources of type g. That number must be satisfied by the
resource assignment. This is enforced by constraint (43). Constraint (44) ensures that each resource
can only be assigned to an activity so many times. In particular the number of assignments must be
less than the total number of patients. Constraints (45)-(48) remove invalid assignments. For
instance constraint (45) ensures that no assignment is possible if the activity is not performed in the
space. Furthermore, constraint (46) ensures that no assignment is possible if the resource is not
working in the space. Constraint (47) ensures that no assignment is possible if the resource does not
perform activities of the specified type.

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
4.3.4. Transfers Between Care Plans

Patients who present themselves at the emergency department are treated in bed-spaces and by
staff and equipment present in that area of the hospital. This implies that the ED is perhaps a
separate entity and can be analysed separately. However this is not true as a significant number of
those patients change status and become inpatients. Many of these require surgery and are re-
categorized as elective patients. Those patients require more hospital resources than other patient
types. The capacity model does not facilitate this per se. A naïve first approach may be to specify
additional PCP’s, but the number of those would be excessive. Hence another mechanism is
proposed. Let us assume that the percentage number of patients that convert from one (𝛾, 𝜓)
pairing to another is given. Let us also assume that these conversions, where present, are stored, for
example in the following set: 𝑇𝑅𝐴𝑁𝑆𝐹𝐸𝑅 = {(𝛾 ′ , 𝜓 ′ , 𝛾, 𝜓, 𝜌)|(𝛾, 𝜓) ∈ ℘, (𝛾 ′ , 𝜓 ′ ) ∈ ℘, 𝜌 ∈ [0,1]}.
Each tuple specifies that 𝜌 % of patients of type (𝛾′, 𝜓′) are converted to patients of type (𝛾, 𝜓). The
actual number of patients treated of each (𝛾, 𝜓) pairing is therefore given by equation (50).
Consequently in equation (51) the comparative relation given by equation (7) must be revised, to
ensure that the correct assignments are made to each ward and to each space:

2
𝑛̅𝛾,𝜓 2
= 𝑛𝛾,𝜓 + ∑∀(𝛾′ ,𝜓′ ,𝛾,𝜓,𝜌)∈𝑇𝑅𝐴𝑁𝑆𝐹𝐸𝑅 (𝜌. 𝑛𝛾2′ ,𝜓′ ) ∀(𝛾, 𝜓) ∈ ℘ (50)
2 1
𝑛̅𝛾,𝜓 = ∑𝑤∈𝑊(𝛼𝛾,𝜓,𝑘,𝑤 ) ∀(𝛾, 𝜓, 𝑘) ∈ ℘ (51)

5. Capacity Querying

Capacity querying is an important activity. This section discusses how the MILP models may be
altered to perform capacity querying. Four methods are proposed. Each is used for different
purposes. The input for analysis is an actual mix of patients. It is forthwith described by parameter
2
𝑛̃𝛾,𝜓 ̃ = ∑∀(𝛾,𝜓)∈℘ 𝑛̃𝛾,𝜓
. For this patient case mix the total number of patients is as follows: 𝔸 2
.

Query 1: The first query takes into account a set of treatment targets and as such there are no
2 2
patient mix constraints. The following constraints are added: 𝑛𝛾,𝜓 ≥ 𝑑𝑒𝑚𝛾,𝜓 ∀(𝛾, 𝜓) ∈ ℘ where
2 2
𝑑𝑒𝑚𝛾,𝜓 = 𝑛̃𝛾,𝜓 . If the demand targets cannot be met, then the model has no solution. In that event
no information is provided as to how feasible the actual patient case mix is and whether the
hospital’s capacity is close to sufficient, or completely unrealistic. The advantage of this approach is
that the demand targets are first met, and any free capacity is then assigned opportunistically, in an
unrestricted way to maximise the overall number of treatments. This may be useful where the
emphasis is on value, with an expectation that minimum work targets be met.

Query 2: The second query utilises an implied proportional patient case mix defined from the actual
specified mix of patients. The first step is to compute 𝜇𝛾1 = ∑𝜓∈𝛹𝛾 (𝑛̃𝛾,𝜓 2 ̃ ∀𝛾 ∈ 𝛤 and
) /𝔸
2 2 2
𝜇𝛾,𝜓 = 𝑛̃𝛾,𝜓 / ∑𝜓′ ∈𝛹𝛾 (𝑛̃𝛾,𝜓 ′) ∀(𝛾, 𝜓) ∈ ℘. The next step is to solve the capacity model to obtain
2 2 2
𝑛𝛾,𝜓 and 𝔸. An analysis of the results is then necessary. This involves a comparison of 𝑛𝛾,𝜓 with 𝑛̃𝛾,𝜓
and a comparison of 𝔸 with 𝔸 ̃ . These values allow free capacity to be computed in the following
F
way: 𝔸 = max(𝔸 − 𝔸 ̃ , 0). Hence capacity does not exist if 𝔸 ̃ > 𝔸 and then free capacity does
̃
exist if 𝔸 < 𝔸. If capacity does exist then the specified mix is maintained and 𝑛𝛾,𝜓 2
is scaled
2 2
appropriately. It should be noted that in that scenario 𝑛𝛾,𝜓 > 𝑛̃𝛾,𝜓 ∀(𝛾, 𝜓) ∈ ℘. Where capacity is
2 2
lacking, the opposite occurs, 𝑛𝛾,𝜓 < 𝑛̃𝛾,𝜓 ∀(𝛾, 𝜓) ∈ ℘. This strategy may be useful when there is
uncertainty in minimum work targets being met, or if Query 1 has identified an infeasible case mix.

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
Query 3: The third query is applied when insufficient capacity is identified. This query is a variation of
the first query but in contrast involves the relaxation of the demand constraints and the penalization
of unmet demands. In other words no patient mix constraints are required here as in query 1, and
2 2
the penalised target violation is calculated as 𝑃𝑇𝑉 = ∑∀(𝜓,𝛾)∈℘ ℧𝛾,𝜓 . |𝑛𝛾,𝜓 − 𝑑𝑒𝑚𝛾,𝜓 |. Penalty value
℧𝛾,𝜓 is introduced as a mechanism to preference different PCP and to allow PCP prioritisation. It
may be removed and replaced with a single penalty ℧. In essence this query searches for an
alternative case mix that is feasible, that is not too different. The case mix is changed only where
necessary by this approach.

Query 4: This query is a variation of the first, and requires demand constraints to be satisfied. The
purpose however is not to maximize the number of treatments but to minimize the time required to
process the specified patient case mix, namely 𝑇 ′ . If 𝑇 ′ > 𝑇 then capacity does not exist. If 𝑇 ′ < 𝑇,
then free capacity exists. For this approach 𝑇𝑤2 should be replaced with 𝑇 ′ in constraint (4). In
constraint (27), 𝑇𝜋3 should also be replaced with 𝑇 ′ . As 𝑇 ′ is a decision variable, constraint (16) and
(33) must be removed. Constraint (3), which concerns case mix violations, is also unrequired.
To demonstrate these four queries, a simple hospital scenario with 16 PCP and two different
case mix has been chosen. The results are shown in Table 1. The first PCM is feasible. Query 1 in
particular demonstrates that there is a great opportunity to treat an additional number of patients,
after the specified target demands have been met. If the PCM is strictly enforced then a further
142.5 patients can be treated and this indicates the presence of free capacity. The second PCM is not
feasible and this is indicated by Query 2. For example if the PCM is strictly enforced then 169 fewer
patients can be treated. Query 3 however shows that the specified PCM is close to feasible, if it is
altered slightly. For example 20730 patients can be treated if the second PCP of patient type one is
reduced by six treatments. Using Query 4 the first PCM can be processed in 8700 hours (i.e. 60 hours
less). This is another indicator of free capacity. The second case mix however requires 8832 hours
(i.e. an extra 72 hours).

Table 1. A demonstration of capacity checking for feasible and infeasible case mixes
𝜸 𝝍 Proposed Strategy 1 Strategy 2 Proposed Strategy 2 Strategy 3
Mix [1] [Strict Targets] [Enforced Mix] Mix [2] [Enforced Mix]
1 1 680 685 [+5] 684.6897 [+4.6897] 683 677.4321 [-5.5679] 683
2 680 685 [+5] 684.6897 [+4.6897] 686 680.4076 [-5.5924] 680 [-6]
3 680 682.5 [+2.5] 684.6897 [+4.6897] 680 674.4565 [-5.5435] 680
4 680 685 [+5] 684.6897 [+4.6897] 690 684.375 [-5.625] 690
2 1 45 45 45.31034 [+0.3103] 45 44.63315 [-0.3668] 45
2 45 45 45.31034 [+0.3103] 50 49.59239 [-0.4076] 50
3 45 47.5 [+2.5] 45.31034 [+0.3103] 35 34.71468 [-0.2853] 35
4 45 45 45.31034 [+0.3103] 2 1.983696 [-0.0163] 2
3 1 3000 43563.33 [+40563.33] 3020.69 [+20.6897] 3100 3074.728 [-25.2717] 3100
2 3000 3000 3020.69 [+20.6897] 3000 2975.543 [-24.4565] 3000
3 1500 1500 1510.345 [+10.3448] 1500 1487.772 [-12.2283] 1500
4 3000 3000 3020.69 [+20.6896] 3000 2975.543 [-24.4565] 3000
5 2500 2500 2517.241 [+17.2414] 2500 2479.62 [-20.3804] 2500
6 3000 3000 3020.69 [+20.6896] 3000 2975.543 [-24.4565] 3000
7 1500 1500 1510.345 [+10.3448] 1500 1487.772 [-12.2283] 1500
8 261 270 [+9] 262.8 [+1.8000] 265 262.8397 [-2.1603] 265
Totals 20661 61253.33[+40592.33] 20803.49 [+142.49] 20736 20566.96 [-169.043] 20730 [-6]

6. Sensitivity Analysis and Capacity Buffering

In hospitals the activity processing times may vary significantly and this can greatly affect overall
hospital utilization. Evidently it is necessary to analyse the sensitivity of the hospitals theoretical
capacity under these circumstances. To perform such an analysis, two algorithms are suggested.
These algorithms take as input a PCM, i.e. a particular set of demands. These approaches can be
used to identify the necessary level of capacity buffering. The first approach determines and

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
analyses the range of total treatments that may be obtained, i.e. the different capacity values. The
specific details are as follows:

Algorithm 1. Sensitivity analysis of capacity for different processing times

Solve model with maximize 𝔸 ̃ as objective. Record capacity value 𝔸
̃.
For trial 𝑗 = 1, … , 𝐽:
Arbitrarily select 𝜁𝛾,𝜓,𝑘 ∈ [−1,1] ∀(𝛾, 𝜓, 𝑘) ∈ ℘1
Compute ∆𝑡𝛾,𝜓,𝑘 = 𝜁𝛾,𝜓,𝑘 × 𝑡𝛾,𝜓,𝑘 and set 𝑡𝛾,𝜓,𝑘 = 𝑡𝛾,𝜓,𝑘 + ∆𝑡𝛾,𝜓,𝑘 ∀(𝛾, 𝜓, 𝑘) ∈ ℘1
Solve model with Query 2 and record capacity as 𝔸𝑗
Compute ∆𝔸𝑗 = 𝔸 ̃ − 𝔸𝑗
avg
Compute: 𝔸 = ∑𝑗 𝔸𝑗 /𝐽

This sensitivity analysis involves the repeated solution of the capacity model with different randomly
generated activity times. The minimum number of trials needed can be determined by the approach
of Burdett et al (2012). Parameter 𝜁𝛾,𝜓,𝑘 is a % increase factor when in [0,1], and a reduction factor
when in [−1,0]. In the event that activity time reductions are unimportant, the above algorithm can
easily be modified to remove this feature. If ∆𝔸𝑖 < 0 then the hospital has insufficient capacity. If
∆𝔸𝑖 = 0 then the hospital has sufficient capacity. If ∆𝔸𝑖 > 0 then the hospital has excess capacity.
The second proposed approach determines and analyses the range of total time required to process
the given PCM. This approach (i.e. Algorithm 2) is similar to the preceding algorithm but the
objective is to minimize 𝑇̃ and the model is solved via Query 4. It should also be noted that 𝑇𝑗 is
recorded. Hence: ∆𝑇𝑗 = 𝑇̃ − 𝑇𝑗 , and 𝑇 avg = ∑𝑗 𝑇𝑗 /𝐽.
To demonstrate these two algorithms the first patient case mix in Table 1 was analysed. The
histogram in Figure 3 describes the sensitivity of the hospitals capacity over the course of one year,
for a 5% level of positive and negative deviation, and over 1000 different trials. The average capacity
is 20304.6, i.e. a reduction of 498.89 treatments or in other words a 2.46% decrease. The standard
deviation is 182.63 treatments. At worst 800 fewer patients were treatable, and at best an
additional 111. Of the 1000 trials, 994 resulted in a capacity decrease while 6 an increase. In
summary this demonstrates a considerable variation in capacity and the need for capacity buffering.
The results of the application of the second algorithm are summarised in Figure 4. In summary an
average of 8940.89 hours was required, i.e. an increase of 240.89 hours over the original 8700 hours
required. This is a 2.76% increase. The standard deviation was 95.88 hours. At worst 349 hours are
needed, and at best 157.8 hours less.
70

60

50
Frequency

40

30

20

10

0
-800
-775
-750
-725
-700
-675
-650
-625
-600
-575
-550
-525
-500
-475
-450
-425
-400
-375
-350
-325
-300
-275
-250
-225
-200
-175
-150
-125
-100
-75
-50
-25
0
25
50
75
100
125

Deviation in treatment numbers

Fig.3. Sensitivity of capacity (i.e. total treatment numbers) for a 5% deviation in activity times

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
 9100 120

9000 100
Time required (hrs)
8900 80

Frequency
8800 60

8700 40

8600 20

8500 0
0 100 200 300 400 500 -35 -10 15 40 65 90 115140165190215240265290315340365
Trial # Time deviation (hrs)

Fig.4. Sensitivity of the PCM; a) scatter plot, b) histogram

Both algorithms suggest that additional time will be required in practice to process the
specified PCM. Some form of buffering is therefore required to ensure that capacity is more readily
available. In other words the probability of not having sufficient capacity should be made small by
selecting a reduced PCM. To determine this reduced PCM, Query 2 can be applied with one small
change and then algorithm 2 can be applied for verification purposes. The exact steps will now be
demonstrated on the aforementioned example. For this scenario it should first be noted that the
specified PCM comprises 20661 patients which in theory can be treated within 8700 hours. The
average time required however is 8940.89. This is a 2.769% increase (i.e. 8940.89/8700 – 1). Hence
the original time of 8700 should be decreased, perhaps by the same percentage. In this case it is
8549.11 hours (i.e. 8700 * (1-0.02769)). With 𝑇 = 8459.11 the capacity model is resolved via Query
2 to obtain a new PCM with a total of 20088.93 patients (i.e. 572 fewer patients). The percentage
reduction in total treatments is also 2.769% (i.e. 572 / 20661). The obtained PCM is as follows:
[661.17, 661.17, 661.17, 661.17, 43.75, 43.75, 43.75, 43.75, 2916.93, 2916.93, 1458.47, 2916.93,
2430.78, 2916.93, 1458.47, 253.77]. The application of Algorithm 2 to analyse the sensitivity of this
PCM, resulted in an average time of 8700.83 with a standard deviation of 90.96. The average
increase was 241.72 hours, the min duration was 8421.77 and the max was 8797.47. Hence the min
deviation was -37.34 hours and the max was 338.36 hours. The difference between the obtained
time and the required time 8700 was also analysed. Those results are summarised in Figure 5.
In summary, the analysis has shown that a revised PCM can be created with specified average
time to complete of 8700. Hence under a wide range of activity time variations, there is sufficient
capacity to process the entire PCM “on time”. The results shown in Figure 5b describe a probability
density function (pdf) for the completion of the PCM on time. In this case, the PCM will be
completed on time 17% of the time. It will be completed early 25% of the time and late 58% of the
time. To reduce the probability that the PCM will be completed late requires an iterative process of
analysis. For example the percentage reduction can be reduced in increments, until a desired level of
robustness is obtained.

150 25

100
20
Relative Frequency (%)

50

0 15
Deviation (hrs)

0 20 40 60 80 100
-50
10
-100

-150 5
-200
0
-250
-80
-60
-40
-20

100
-300
-280
-260
-240
-220
-200
-180
-160
-140
-120
-100

0
20
40
60
80

-300
Trial # Time deviation (hrs)

Fig.5. Sensitivity of the revised PCM; a) scatter plot, b) histogram

7. Case Study
 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
In this section IBM’s CPLEX Studio (Version 12.6) was used to solve the capacity model and its
variants. A quad core Dell personal computer (PC) with a 2.5 gigahertz processor and 4 gigabytes
memory under Windows 7 was used. To demonstrate the proposed methods our case study is based
upon de-identified data from the Princess Alexandra Hospital, Brisbane, Qld, Australia. It is a
metropolitan tertiary referral and university and college affiliated teaching hospital. The hospital is
spread over a number of different floors. It provides acute and elective adult medical and surgical
care and emergency department services. The intensive care unit, the operating theatres and the
cardiac care unit are positioned adjacently for easy transfer. Our case study includes 59 specialty
units, 74 wards, 840 bed-spaces and 20 operating theatres or thereabouts. To demonstrate our
models, the activities and work performed in 2012 were selected. The surgical cases treated in that
year were extracted from the operating room management and information system (ORMIS)
records. Post-operative care, intensive care and recovery times were separately identified from ward
occupancy information. ED data was extracted from the Emergency Department Information system
(EDIS). There were 19174 patient episodes, of which 13084 were elective and 6090 emergency.
There are many ways to define patient type. Here the medical and surgical specialties have
been used. Within those specialties, patient care pathways have been created. This choice was made
because the patient cases in each specialty are known and the proportional patient case mix can be
computed from these numbers. For each specialty four PCP were created. Two PCP are for elective
patients, with and without an ICU stay. The other two are for emergency patients, with and without
an ICU stay. Each PCP consists of four stages: pre-operative care, surgery, post- operative care or
intensive care, and recovery.

7.1. Analysis of Surgical Activities

The hospitals capacity to perform surgical operations was first analysed. The activities performed
within the emergency department were not included. This analysis includes both elective and
emergency surgical patients. The PCP transfer mechanism of section 4.7 was not utilised but it is
possible for patients with a PCP that does not include an ICU stay to be converted to one that does.
It is also possible for elective patient to be converted to emergency patient. The proportional patient
case mix previously described was enforced using the penalty approach and then via strict
constraints. The time to solve the model was around 12 seconds. The results of the first approach
are shown in Table 2 and Figure 6. These results show that only three values (i.e. of the total number
of treatments) are obtainable. This occurs because the aforementioned penalty method is
synonymous with the standard weighted sum method that cannot identify non convex regions of the
Pareto frontier (Kim and DeWeck, 2005). If the product of the CMV and the penalty is less than the
number of treatments then the model is solvable, otherwise it is not.

Table 2. Model solution for different penalty

Penalty CMV #Patients CMZ*Penalty
334 na na na
333 82.053 27351 27323.65
40 82.053 27351 3282.12
[1,39] 349.2 37245 [349.2, 13618.8]
0 139395 72702 0

Fig.6. Visual summary of Table 2 results

The second approach involving the satisfaction of case mix constraint (3) is now discussed. The
results are shown in Figure 7.

 Corresponding author: Tel: +61 7 3138 1029; Fax: +61 7 3138 2310.
E.mail Address: e.kozan@qut.edu.au.
 Fig.7. Pareto frontier of CMV versus number of treatments; full plus enlarged portion

This chart describes a Pareto frontier. For small increases to the CMV initially, the number of
treatments can be increased greatly. As the CMV increases further however, the number of
treatments does not increase as much. At a CMV of 6000, the number of treatments is no longer
increasing by very much and is quite flat. The penalty method in comparison is shown in red (i.e.
dashed). Using a Euclidean distance metric this chart suggests that the best solution is 53767.336
patients with a CMV of 30000. If no CMV is allowed then the solution is only 333.33 patients.
Such a low value of capacity exists because it is not possible to satisfy the proportional case mix in
any other way, without bottleneck issues coming into effect. Upon reflection it is undesirable to
have a solution with a small value of capacity and low CMV or else a solution with a large value of
capacity and large CMV. Hence the ratio between the CMV level and the total number of treatments
is very important. It is proposed that a solution with the lowest ratio is selected. In other words it is
necessary to choose a solution with a relatively high number of treatments and a relatively low CMV.
Amongst those with the same ratio, the solution with the higher number of treatments is superior as
the other solution points are dominated solutions. This chart indicates that the tenth solution point
with a CMV of 82 and number of treatments 27351 is best. This is a much better solution as the
number of treatments is high and the CMV is quite low. For the aforementioned solution the
number of patients of each type is as follows: 𝑛1𝛾 =[984.64, 902.59, 738.48, 1531.66, 0, 984.64,
765.83, 328.21, 0, 738.48, 133.81, 1176.10, 4458.22, 5907.83, 2981.27, 0, 1121.39, 82.05, 1042.29,
2352.19, 1121.39]. The ratios are shown in Figure 8.

Fig.8. Ratio of CMV to number of treatments; full chart plus enlarged portion

It is interesting to note that the application of constraint (23) and (24) resulted in a similar solution
with the same CMV and same number of treatments. The ward and space utilizations associated
with this solution were also obtained, see Figure 9a. They demonstrate that three wards are highly
utilized, i.e. at 70-83%. These include one operating theatre and the post anaesthetic care ward
within the surgical care unit. The other wards have utilizations around 20%. The ward that cares for
orthopaedic patients is most heavily utilized as the proportional case mix for that specialty is
highest.

20
 6000

5000

4000

#Treatments
Historical
3000 Capacity Model

2000

1000

Surgical Specialty

Fig.9. Surgical care activities. a) Ward utilizations; b) a comparison with historical result

In comparison to what actually occurred (see Figure 9b), the theoretical capacity is approximately
30% greater. This is reasonable as there are inefficiencies and stochastic variations in real life. The
OPHT, ORTH and PLAS units (i.e. specialty 13, 14, 15) are noteworthy because the difference in their
treatment numbers is quite large. These results indicate that the model with the specified PCP is
able to successfully describe the utilisation of the hospital’s physical spaces in regard to the
treatment of surgical elective patients.
To demonstrate capacity querying, Query 1 was applied. The historical patient case numbers
were added as target constraints, and the proportional case mix constraints were removed. The
number of treatments obtained from the solution of the model was 68750.9. The exact case mix and
levels of free capacity are shown in Table 3 below. Once demands are met, free capacity is assigned
opportunistically. These results indicate that wards belonging to nine units have ample capacity to
process more patients.

Table 3. Capacity querying results via Query 1

Unit Historical #Cases Free Unit Historical # Cases Free Unit Historical #Cases Free
ASU 692 692.00 0.00 GAS 229 328.21 99.21 PLAS 2099 2099.00 0.00
B&E 631 10449.04 9818.04 GYN 4 17552.57 17548.57 RESP 4 6942.54 6938.54
COLO 516 516.00 0.00 HPB 513 513.00 0.00 RTPT 780 4629.15 3849.15
CSU 1082 1082.00 0.00 LTPT 82 82.00 0.00 TRMA 59 59.00 0.00
DENT 3 3.00 0.00 NSUR 823 823.00 0.00 UGI 691 691.00 0.00
ENT 698 788.32 90.32 OPHT 3136 4518.77 1382.77 UROL 1657 7921.94 6264.94
FMAX 533 533.00 0.00 ORTH 4151 4151.00 0.00 VASC 791 4376.35 3585.35

7.2. Inclusion of Medical Patients

There are many medical patients that visit the hospital for non-surgical interventions. The hospitals
capacity to perform medical treatments is analysed here. Many of these medical treatments are
performed in the spaces and wards used by elective surgical patients. The historical PCM was used
and the solution of the capacity model for different CMV values is shown in Figure 10. It is worth
noting that the penalty approach was also applied and is also shown. The ratio of the CMV to the
number of treatments is also shown in Figure 10.

21
Fig.10. Trade-off curves for the number of treatments versus the CMV and the CMV ratio

For small increases to the CMV initially, the number of treatments can be increased greatly. As the
CMV increases further however, the number of treatments does not increase as much, but it does
continue to increase. With respect to the minimum ratio criteria, the best solution has 35417.14
treatments and a CMV of 80. This number of treatments is less than the historical value of 55946.
Upon closer scrutiny, six of the 32 medical units differ greatly and contribute to the overall
difference of about 20000 patients per year. Otherwise, the capacity model aligns very well with
what actually occurred. This is demonstrated in Figure 11.

Fig.11. Comparison of capacity model solution with historical result (Medical Treatments)

We hypothesize that a large part of the observed discrepancy occurs due to the inaccuracy of
medical outpatient data. The flow of surgical patients within the hospital has been more transparent
and data collection systems like ORMIS and HBSCIS are better at recording surgical rather than
medical cases.
To demonstrate capacity querying, Query 1 was applied. The historical patient case numbers
were not added as target constraints because the results in Figure 11 demonstrate that insufficient
capacity exists. Instead the aforementioned results are used as targets. The number of treatments
obtained with Query 1 was 265213. The exact case mix and levels of free capacity are shown in Table
4 below. Once demands are met, free capacity is assigned opportunistically. These results indicate
that wards belonging to many units have ample capacity to process more patients.

22
Table 4. Capacity querying results for Query 1
m Unit Target # Patients Free m Unit Target # Patients Free
1 BIRD 1015 1015 0 17 MED5 385 385 0
2 BIRU 280 462.64 182.64 18 MED6 315 315 0
3 CARD 5145 11256.24 6111.24 19 MED7 245 245 0
4 DEND 175 1783.29 1608.29 20 MSTU 385 4709.31 4324.3
5 DERM 280 16000 15720 21 NEUR 455 455 0
6 GERI 1610 2699.32 1089.32 22 OHU 1925 3753.21 1828.21
7 GMAP 840 9868.57 9028.57 23 PD 700 15215.88 14515.88
8 GRD 700 58174.12 57474.12 24 PPMS 105 47868.85 47763.85
9 HD 10570 58400 47830 25 PSY 3749.63 3749.63 0
10 HYPT 175 175 0 26 PSYC 0 0 0
11 IMMU 0 11076.12 11076.12 27 PSYG 210 346.38 136.38
12 INFD 245 2154.98 1909.98 28 REHD 175 347.58 172.58
13 MED1 770 770 0 29 REN 1715 4756.99 3041.99
14 MED2 700 700 0 30 RHEU 70 5141.06 5071.06
15 MED3 805 805 0 31 ROU 210 1173.32 963.32
16 MED4 450.37 450.37 0 32 SIU 595 960.18 365.18

7.3. Surgical and Medical Patients

In this section a combined analysis of surgical and medical activities has been performed. This is
required as some medical specialities also use the same wards and spaces as surgical specialties thus
resulting in a reduced level of capacity. The solution of the capacity model for different CMV values
is shown in Figure 12 and the ratio of the CMV to the number of treatments is shown in Figure 13.

Fig.12. Trade-off between the treatments and the CMV; full chart plus enlarged portion

Fig.13. Trade-off between treatments and the CMV ratio; full chart plus enlarged portion

In Figure 12 the greatest growth in the total number of treatments occurs when the CMV is less than
100. The increase tapers off for greatly when the CMV exceeds 100000. The ratio of the CMV to the
number of treatments shown in Figure 13 is quite interesting as the curve is both concave and
convex. According to the minimum ratio criteria the best solution occurs when the CMV is 75. The
best solution consists of 47187 treatments, with 12047 surgical and 35140 medical.
A comparison of the results obtained in Section 5.1 and 5.2 are shown below in Figure 14.

23
 Fig.14. Comparison of capacity results; full chart plus enlarged portion

These charts indicate that comparatively speaking, the surgical case mix has less effect on the overall
capacity. The medical patient case mix however has a large effect on the total number of treatments
that the hospital can achieve.

7.4. Analysis of Emergency Department Activities

Due to a lack of detailed information (i.e. in EDIS), about where patients are treated in the
emergency department, each PCP was first given a single activity. That activity is performed in the
acute treatment or resuscitation area of the ED. Emergency care activities were divided into
specialty specific variants. The solution of the model with the historical PCM resulted in the
following case mix: 𝑛1𝛾 =[3.70, 351.10, 36.96, 0, 77.61, 103.48, 3.70, 59.13, 121.96, 140.44, 59.13,
3.70, 291.97, 29.57, 1552.25, 125.66, 0, 51.74, 70.22, 55.44, 144.14, 218.05, 33458.44] and total
treatment of 36958 cases. This number is much reduced from the 60135 that actually occurred. In
conclusion, it can be assumed that a significant number of patients (i.e. about 23177) are being
treated elsewhere in the ED, or else transferred to other parts of the hospital. Hence more PCP need
to be identified for emergency patient.

7.5. Including Other Hospital Resources

In this section hospital resources are added to the case study for added realism. All major staff and
equipment could in theory be added, however obtaining that information is problematic. First it is a
large quantity of information. Second it is not something that the hospital recorded diligently and
accurately in the year that our case study occurred. What we did have was information about the
consultants that performed surgeries, and they have been added. There were 182 consultants in
total. The skill set of those consultants, for example the type of surgical specialties they performed,
was obtained. Of the 182 consultants, 141 performed surgeries for a single specialty, thirty two
consultants were associated with two specialties, eight with three, and one with four.
Each surgical consultant was given a time availability of 1920 hours per year (i.e. 8 hours per
day over 48 weeks). In each surgical PCP the consultants required were assigned for the entire

24
duration of the surgical activity and one consultant was assigned to a proportion of the patients
recovery phase. The full resource model whereby resources are assigned to individual activities and
spaces was attempted, but was unsuccessful due to memory limitations of the aforementioned
personal computer. Consequently the reduced resource model whereby resources are assigned only
to individual tasks was attempted. That model was successfully solved. Based upon the results of
that model, we have found that the number and skill set of the consultants does not restrict the
total number of treatments by much (i.e. 27351 to 27333) if one consultant performs each surgery.
However, if two or three are required, then the total number of treatments (i.e. the capacity) drops
from 27333 to 20574.16 and then to 14134.31.
The utilization level of the consultants is summarised in Figure 15. Those results are associated
with a CMV of 82 as identified in section 5.1. Only a small number (i.e. about 15%) are heavily
utilized. A large group (60%) are not utilized at all. The specialties that are most restricted can also
be determined. They vary for each of the three scenarios. A summary of the most utilized
consultants (i.e. 90 – 100%) and their specialty is shown in Table 5. In conclusion the specialities with
the largest presence in the proportional case mix have the most heavily utilized consultants.

100
90
80
# of Consultants

70
60 One Consultant
50 Two Consultants
40
30 Three Consultants
20
10
0

Utilization Range (%)

Fig.15. Histogram of utilization level

Table 5. Summary of the number of consultants above 90% utilization

Specialty (# of Consultants)
1 CSU (3), ENT (1), OPHT (1), ORTH (7), PLAS (1), TRMA (3), UGI (1), UROL (1), VASC (1)
2 COLO (1), CSU (5), ENT (1), NSUR (1), OPHT (2), ORTH (7), PLAS (2), TRMA (2), UGI (1), UROL (2), VASC (2)
3 CSU (5), OPHT (1), ORTH (9), PLAS (2), TRMA (1), UROL (3)

8. Concluding Remarks

This article has introduced a new approach to analyse the capacity of an entire hospital. Our
approach, based upon resource saturation and bottlenecking principles, is generic and highly
extendable. A key feature of the approach is the solution of a suite of optimization models. The first
mathematical model (HCM1) involves the assignment of patient care activities to hospital areas and
other rooms or zones within the hospital. The second model (HCM2) is more fine grained and
involves the assignment of activities to individual treatment spaces. The assignment of individual
resources to patient care activities is also included in our third model (HCM3). The solution of the
aforementioned models provides an upper bound on the capacity of a hospital, and second, a
patient case mix that describes how many patients of each type can be processed by the hospital
within a specified time period. An analysis of stochastic factors has also been facilitated by a
sensitivity analysis in order to identify operational capacity and to facilitate capacity buffering.
Four different capacity querying strategies were also proposed. In summary, the first query is
appropriate when a yes or no answer is required. When the amount of free capacity should be

25
exactly identified or where the exact capacity deficiency is to be identified, then the second query is
needed. The third query is good at producing corrected case mixes in the event that capacity is
insufficient. The fourth query is also particularly useful in that event, as it determines the time
required to process the given patient case mix. If free capacity is available then the fourth query is
also useful.
The capacity models were applied to a large public adult general hospital in Brisbane Australia.
In our analysis, we have identified the capacity of that hospital to treat surgical and medical patients.
Our results are consistent with what occurred historically, and this result provides a validation of our
capacity analysis approach. Our numerical testing has highlighted that the strict enforcement of case
mix constraints results in a low value of capacity. It is hence better to perform a trade-off (i.e.
Pareto) analysis to see what effect the relaxation of these constraints has. Hence the hospital
capacity is in some sense “selectable”. Our numerical investigations have also shown that the
number of resources available and their time availability restricts overall capacity.
The MILP models incorporate many real life technical conditions and are bespoke to health
care applications. No other approach comes close to our knowledge. It is believed that the
application of our capacity assessment approach can benefit hospitals in many ways and that it is
viable to put this framework in place for instance within a hospital capacity management system.
The MILP models can be integrated into existing information technology platforms and anecdotal
evidence suggests their data requirements are no longer unrealistic.
Although our approach is extremely flexible and comprehensive, this flexibility could be
problematic for end users who will need to carefully select how best to use it. Experts in decision
making and Operational Research may be needed in the early stages to get the most out of this
approach and to supervise how it is used. To assess hospital capacity different kinds of patients and
activities need to be classified and categorized – but this is not necessarily easy. There are many
alternatives and which one should be used is not necessarily transparent. A statistician or other data
scientist may be needed to provide clarity and expertise here.
The model is extremely flexible and can be applied with more or less detail– but that may be
too idiosyncratic to the people who will ultimately use the model. The results provided by the model
are only as good as the data that is used. If detailed accurate information is not available then the
model will provide an incorrect forecast. The model apportions the time available to resources. The
model however is not able to take into account more advanced operating procedures and other
behavioural phenomenon. It can provide a reliable and accurate upper bound and a reasonably
accurate approximation of operational capacity. To evaluate the aforementioned aspects, simulation
based approaches could be used.
This article has identified a number of new research directions. A natural extension of this
work would be to develop a region or organisation based analysis of capacity. The model appears to
be sufficiently robust and could facilitate multiple hospitals using an additional subscript in the MILP.
A capacity expansion approach can also be developed. The choice of which resources to buy and
which hospital areas to expand, can be added as additional decisions in the MILP. The
reconfiguration of an existing hospitals areas (i.e. wards) to increase capacity can be in theory be
facilitated by the MILP. The incorporation of financial aspects of health care would also greatly
improve hospital capacity planning activities and make the proposed hospital capacity assessment
approach more robust and valuable. The integration of all of these extensions and features into one
unified framework is also a task that should be pursued.

Acknowledgements: This research was funded by the Australian Research Council (ARC) Linkage
Grant LP 140100394 and supported by the Princess Alexandra Hospital, Brisbane, Australia. We
would like to thank Dr Andy Wong for his assistance in obtaining data for our case study. We would
also like to thank staff at the PAH for their considerable feedback and their time.

References

26
Akcali, E., Cote, M.J., Lin, C. (2006). A network flow approach to optimizing hospital bed capacity decisions,
Health Care Manage Science, 9, 391-404.
Braubach, L., Pokahr, A., & Lamersdorf, W. (2014). Negotiation-based patient scheduling in hospitals.
Technologies and Decision Support Systems. Studies in Computational Intelligence, 486, 107-121
Burdett, R.L., Kozan, E., Strickland, C. (2012). A practical approach for identifying expected solution
performance and robustness in Operation Research applications. ASOR Bulletin, 31(2), 1-28.
Burdett, R.L., Kozan E. (2015). Techniques to effectively buffer schedules in the face of uncertainties. Computers
and Industrial Engineering. DOI: 10.1016/j.cie.2015.04.024
Burdett, R.L. (2015). Multi-objective models and techniques for railway capacity analysis. European Journal of
Operational Research. DOI: 10.1016/j.ejor.2015.03.020.
Burdett, R.L., Kozan, E. (2016). A multi-criteria approach for hospital capacity analysis. European Journal of
Operational Research, 255(2), 505-521.
Cappanera, P., Visintin, F., Banditori, C. (2014). Comparing resource balancing criteria in master surgical
scheduling: A combined optimisation-simulation approach. Intl J Production Economics, 158, 179-196.
Cardoen, B., Demeulemeester, E., Belien, J. (2010). Operating room planning and scheduling: a literature
review. European Journal of Operational Research, 201, 921-932.
Cardoen, B., Demeulemeester, E., Belien, J. (2009). Optimizing a multiple objective surgical case sequencing
problem. International Journal of Production Economics, 119, 354–366.
Carpenter, C.E., Rosko, M.D., Louis, D.Z., & Yuen, E.J. (1999). Severity of illness and profitability: a patient level
analysis. Health Services Management Research, 12, 217-226.
Carnes, T., Price, D., Levi, R., Dunn, P.F., Daily, B.J., Moss, S. (2011). An optimisation framework for smoothing
surgical bed census via strategic block scheduling.
https://www.gsb.stanford.edu/sites/default/files/documents/oit_01_12_carnes.pdf
Chow, V.S., Puterman, M.L., Salehirad, N., Huang, W., Atkins, D. (2011). Reducing surgical ward congestion
through surgical scheduling and uncapacitated simulation. Production and Operations Management,
20(3), 418-430.
Cochrane, J.K., Roche, K.T. (2009). A multi-class queuing network analysis methodology for improving hospital
emergency department performance. Computers and Operations Research, 36, 1497-1512.
Fei, H., Chu, C., Meskens, N. (2009). Solving a tactical operating room planning problem by a column-
generation-based heuristic procedure with four criteria. Annals of Operations Research, 166 (1), 91–
108.
Forthman, M.T., Dove, H.G., Forthman, C.L., Henderson, R.D. (2005). Beyond severity of illness: evaluating
differences in patient intensity and complexity for valid assessment of medical practice pattern
variation. Managed Care Quarterly, 13(4), 9-17.
Fugener, A., Hans E.W., Kolisch, R., Kortbeek, N., Vanberkel, P.T. (2014). Master surgery scheduling with
consideration of multiple downstream units. European Journal of Operational Research, 239, 227-236.
Gartner, D., Kolisch, R. (2014). Scheduling the hospital-wide flow of elective patients. European Journal of
Operational Research, 233, 689-699.
Gholami, R., Higon D.A., Emrouznejad, A. (2015). Hospital performance: efficiency or quality? Can we have
both with IT?. Expert Systems with Applications, 42, 5390-5400.
Gorunescu, F., McClean, S.I., Millard, P.H. (2002). A queueing model for bed-occupancy management and
planning of hospitals. Journal of the Operational Research Society, 53(1), 19-24.
Green, LV. (2004). Capacity planning and management in hospitals. Operations Research and Health Care.
International Series in Operations Research & Management Science, 70, 15-41.
Horn, S.D., Horn R.A. (1986). The computerized severity index. A new tool for case mix management. Journal of
Medical Systems, 10(1), 73-78.
Horn, S.D., Horn RA. (1986). Reliability and validity of the severity of illness index. Medical Care, 24(2), 159-
178.
Horn, S.D., Sharkey, P.D., Bertram. D.A. (1983). Measuring severity of illness: homogeneous case mix groups.
Medical Care, 21(1), 14-30.
Izady, N., Worthington, D. (2012). Setting staffing requirements for time dependent queueing networks: The
case of accident and emergency departments. European Journal of Operational Research, 219, 531-540.
Khanna, S., Boyle, J., Good, N., Lind, J. (2012). Unravelling relationships: hospital occupancy levels, discharge
timing and emergency department access block. Emergency Medicine Australasia, 24, 510-517.
Kim, I.Y., De Weck, O.L. (2005). Adaptive weighted sum method for bi-objective optimization: Pareto front
generation. Structural Multidisciplinary Optimisation, 29, 149-158.

27
Kozan, E., Burdett, R.L. (2005). A railway capacity determination model and rail access charging methodologies.
Transportation Planning and Technology, 28(1), 27-45.
Kuntz, L., Scholtes, S., Vera, A. (2007). Incorporating efficiency in hospital-capacity planning in Germany. The
European Journal of Health Economics, 8(3), 213-223.
Ma, G., Demeulemeester, E. (2013). A multilevel integrative approach to hospital case mix and capacity
planning. Computers and Operations Research, 40, 2198–2207.
Meskens, N., Duvivier, D., Hanset, A. (2013). Multi-objective operating room scheduling considering desiderata
of the surgical team. Decision Support Systems, 55, 650-659.
Rechel, B., Wright, S., Barlow, J., McKee, M. (2010). Hospital capacity planning: from measuring stocks to
modelling flows. Bulletin of the World Health Organization, 88(8), 632-636.
Rosko, M. (1988). DRGs and severity of illness measures: an analysis of patient classification system. Journal of
Medical Systems, 12(4), 37-55.
Roth, A.V., Van Dierdonck, R. (1995). Hospital resource planning: concepts, feasibility and framework.
Production and Operations Management, 4(1), 2-29.
Queensland Health. (2014). Health funding principles and guidelines 2014-2015.
Queensland Health. (2013). Health priorities 2014-2015. Policy and Planning Branch.
Utley, M., Worthington, D. (2012). Capacity planning, Handbook of healthcare system scheduling.
International Series in Operations Research and Management Science, 168, 11-30.
Vijayakumar, B., Parikh, P.J., Scott, R., Barnes, A., Gallimore, J. (2013). A dual bin-packing approach to
scheduling surgical cases at a publicly funded hospital. European Journal of Operational Research, 224,
583-591.
Van Merode, G.G., Groothuis, S., Hasman, A. (2004). Enterprise resource planning for hospitals. International
Journal of Medical Informatics, 73, 493-501.

28
Appendix: Nomenclature

Description Notation Supplementary Information

Capacity 𝔸 ̅ , 𝔸, 𝔸𝐹 : Bounds on capacity and free capacity
𝔸
1
Hospital Units 𝑢∈𝑈 𝑈𝛾,𝜓 , 𝑈𝑤2 : These sub sets describe the hospital units that are involved with
particular PCP and hospital areas.
1
Note: 𝑈𝛾,𝜓 = {𝑢𝛾,𝜓,𝑘 |∀𝑘 ∈ [1, 𝐾𝛾,𝜓 ]}; 𝑈𝑤2 = {𝑢 ∈ 𝑈|𝑤 ∈ 𝑊𝑢3 };
Abbreviation: 𝑈𝛾,𝜓 , 𝑈𝑤
1 2 4
Hospital Areas 𝑤∈𝑊 𝑊𝛾,𝜓 , 𝑊𝛾,𝜓,𝑘 , 𝑊𝑢3 , 𝑊𝑢,𝜙 : These sub sets describe the treatment areas
4
associated with PCP, PCP tasks, units and activities. 𝑊𝑢,𝜙 is an input.
1 2 2 4 4
Note: 𝑊𝛾,𝜓 = ⋃𝑘∈[1,𝐾𝛾,𝜓] 𝑊𝛾,𝜓,𝑘 ; 𝑊𝛾,𝜓,𝑘 = 𝑊𝑢𝛾,𝜓,𝑘,𝜙𝛾,𝜓,𝑘 ; 𝑊𝑢3 = ⋃𝜙∈𝛷 𝑊𝑢,𝜙 ;
Abbreviations: 𝑊𝛾,𝜓 , 𝑊𝛾,𝜓,𝑘 , 𝑊𝑢 , 𝑊𝑢,𝜙
1 2 4
Treatment Spaces 𝜋∈𝛱 𝛱𝛾,𝜓 , 𝛱𝛾,𝜓,𝑘 , 𝛱𝑤3 , 𝛱𝑤,𝜙 : These sub sets describe the treatment spaces
4
associated with PCP, PCP tasks, areas and activities. 𝛱𝑤,𝜙 is an input.
1 2 2 4
Note: 𝛱𝛾,𝜓 = ⋃𝑘∈[1,𝐾𝛾,𝜓] 𝛱𝛾,𝜓,𝑘 ; 𝛱𝛾,𝜓,𝑘 = ⋃𝑤∈𝑊𝛾,𝜓,𝑘
2 𝛱𝑤,𝜙 𝛾,𝜓,𝑘
;
4
𝛱𝑤3 = ⋃𝜙∈𝛷 𝛱𝑤,𝜙 ;
Abbreviations: 𝛱𝛾,𝜓 , 𝛱𝛾,𝜓,𝑘 , 𝛱𝑤 , 𝛱𝑤,𝜙
𝐿𝜋 : The location of the space in a hospital area.
Hospital Activity 𝜙∈𝛷 The medical and surgical activities
Specialty 𝑠∈𝑆 𝑆 𝐸𝑙 : Sub set of specialties associated with elective patients
Patient Type 𝛾∈𝛤 Γ 𝐸𝑙 : Sub set of patient types associated with elective patient
Patient Care Plan 𝜓 ∈ 𝛹𝛾 Set of patient care plans (PCP) for patients of type 𝛾:
𝑃𝐶𝑃𝛾,𝜓 = {(𝜙, 𝑢, 𝑡, 𝑟)|𝜙 ∈ Φ, 𝑢 ∈ 𝑈, 𝑡 ∈ ℝ}; 𝐾𝛾,𝜓 = |𝑃𝐶𝑃𝛾,𝜓 |;
𝑜𝛾,𝜓,𝑘 : Reference to the kth task in PCP 𝜓 for patient type 𝛾
𝜎𝛾,𝜓 : Specialty of the PCP
𝑢𝛾,𝜓,𝑘 , 𝑡𝛾,𝜓,𝑘 , 𝜙𝛾,𝜓,𝑘 , 𝑟𝛾,𝜓,𝑘 : The unit, processing time, activity and resources
associated with 𝑜𝛾,𝜓,𝑘 .
̇
𝑡𝛾,𝜓,𝑘,𝑤 ̈
, 𝑡𝛾,𝜓,𝑘,𝜋 : Tine required in area w and space 𝜋.
𝑟̇𝛾,𝜓,𝑘,𝑔 : The number of resources of type g required.
𝑡𝛾,𝜓,𝑘,𝑔 : Time resource type g is required during activity 𝑜𝛾,𝜓,𝑘
Patient Case Mix 𝜇 𝜇𝛾1 : Mix of patient types
2
𝜇𝛾,𝜓 : Mix of PCP within each patient category (type)
2
Note: ∑∀𝛾∈𝛤 𝜇𝛾1 = 1; ∑𝜓∈𝛹𝛾 𝜇𝛾,𝜓 = 1 ∀𝛾 ∈ 𝛤;
Resource Group 𝑔∈𝐺 Examples: staff, medical equipment, diagnostic equipment
Resources 𝑟∈𝑅 𝑅𝑢1 , 𝑅𝑤 2 5
, 𝑅𝜋3 ,𝑅𝑔4 , 𝑅𝑔,𝑢 6
, 𝑅𝑔,𝑤 7
, 𝑅𝑔,𝜋 , 𝑅𝜙8 , 𝑅𝑢,𝜙
9 10
, 𝑅𝑤,𝜙 11
, 𝑅𝜋,𝜙 : These sub sets describe
resources that are available to units, areas, spaces, groups and for different
activities. Abbreviations: 𝑅𝑢 , 𝑅𝑤 , 𝑅𝜋 ,𝑅𝑔 , 𝑅𝑔,𝑢 , 𝑅𝑔,𝑤 , 𝑅𝑔,𝜋 , 𝑅𝜙 , 𝑅𝑢,𝜙 , 𝑅𝑤,𝜙 , 𝑅𝜋,𝜙
5
Note: 𝑅𝑔,𝑢 = {𝑟 ∈ 𝑅𝑢1 , 𝑟 ∈ 𝑅𝑔 }; 𝑅𝑔,𝑤 6
= {𝑟 ∈ 𝑅𝑤 2
, 𝑟 ∈ 𝑅𝑔 };
7 3
𝑅𝑔,𝜋 = {𝑟 ∈ 𝑅𝜋 , 𝑟 ∈ 𝑅𝑔 };
4 5 6
𝐹𝑢1 ,𝐹𝑤2 ,𝐹𝜋3 , 𝐹𝑢,𝜙 , 𝐹𝑤,𝜙 , 𝐹𝜋,𝜙 : These sub sets describe resources that are fixed in
place. Abbreviations: 𝐹𝑢 ,𝐹𝑤 ,𝐹𝜋 , 𝐹𝑢,𝜙 , 𝐹𝑤,𝜙 , 𝐹𝜋,𝜙 .
𝑅𝛾,𝜓,𝑘 : Set of resources that may be used to perform activity 𝑜𝛾,𝜓,𝑘 .
Note: 𝑅𝛾,𝜓,𝑘 = {𝑟 ∈ 𝑅|(𝛾, 𝜓, 𝑘) ∈ ℘6𝑟 };
Special Tuples ℘, ℘1 , ℘2 , ℘3 , ℘ = {(𝛾, 𝜓)|∀𝛾 ∈ Γ, ∀𝜓 ∈ 𝛹𝛾 };
℘4𝜙 , ℘5𝑔 , ℘6𝑟 ℘1 = {(𝛾, 𝜓, 𝑘)|∀𝛾 ∈ Γ, ∀𝜓 ∈ 𝛹𝛾 , ∀𝑘 ∈ [1, 𝐾𝛾,𝜓 ]} ;
2
℘2 = {(𝛾, 𝜓, 𝑘, 𝑤)|(𝛾, 𝜓, 𝑘) ∈ ℘1 , 𝑤 ∈ 𝑊𝛾,𝜓,𝑘 };
3 1 2
℘ = {(𝛾, 𝜓, 𝑘, 𝜋)|(𝛾, 𝜓, 𝑘) ∈ ℘ , 𝜋 ∈ 𝛱𝛾,𝜓,𝑘 } ;
℘4𝜙′ = {(𝛾, 𝜓, 𝑘) ∈ ℘1 |𝜙𝛾,𝜓,𝑘 = 𝜙 ′ } ⊂ ℘1 ;
℘5𝑔 = {(𝛾, 𝜓, 𝑘) ∈ ℘1 |𝑟̇𝛾,𝜓,𝑘,𝑔 > 0; 𝑡𝛾,𝜓,𝑘,𝑔 > 0} ⊂ ℘1 ;
℘6𝑟 = (⋃∀𝑔∈𝐺|𝑟∈𝑅𝑔 ℘5𝑔 ) ⊂ ℘1 ;
Assignments 𝛼, 𝛽, 𝛿 𝛼𝛾,𝜓,𝑘,𝑤 : Number of activities that are assigned to area w of type 𝑜𝛾,𝜓,𝑘
𝛽𝛾,𝜓,𝑘,𝜋 : Number of activities that are assigned to space 𝜋 of type 𝑜𝛾,𝜓,𝑘
𝛿𝛾,𝜓,𝑘,𝑟,𝜋 : Number of activities 𝑜𝛾,𝜓,𝑘 that resource r performs in space 𝜋.
2 5 6 7 8 12
Counts 𝑛 𝑛1𝛾 , 𝑛𝛾,𝜓 4
, 𝑛𝑢3 , 𝑛𝛾,𝑢 , 𝑛𝛾,𝜓,𝜙,𝑤 , 𝑛𝛾,𝜙,𝑤 , 𝑛𝜙,𝑤 , 𝑛𝛾,𝜓,𝑤 , 𝑛𝑠9 , 𝑛10 11
𝜋 , 𝑛𝛾,𝜋 , 𝑛𝜙,𝜋 : Number of
patients of different types.
Abbreviations: 𝑛𝛾 , 𝑛𝛾,𝜓 , 𝑛𝑢 , 𝑛𝛾,𝑢 , 𝑛𝛾,𝜓,𝜙,𝑤 , 𝑛𝛾,𝜙,𝑤 , 𝑛𝜙,𝑤 , 𝑛𝛾,𝜓,𝑤 , 𝑛𝑠 , 𝑛𝜋 , 𝑛𝛾,𝜋 , 𝑛𝜙,𝜋
Utilisations 𝜆 𝜆1𝑤 , 𝜆2𝜋 : Area and space utilisation.

29