A v a i l a b l e o n l i n e a t
w w w . s c i e n c e d i r e c t . c o m
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / j v a l
Constrained Optimization Methods in Health Services
Research
—An Introduction: Report 1 of the ISPOR Optimization
Methods Emerging Good Practices Task Force
William Crown, PhD
1
,
*, Nasuh Buyukkaramikli, PhD
2
, Praveen Thokala, PhD
3
, Alec Morton, PhD
4
,
Mustafa Y. Sir, PhD
5
, Deborah A. Marshall, PhD
6
,
7
, Jon Tosh, PhD
8
, William V. Padula, PhD, MS
9
,
Maarten J. Ijzerman, PhD
10
, Peter K. Wong, PhD, MS, MBA, RPh
11
, Kalyan S. Pasupathy, PhD
12
,
*
1
OptumLabs, Boston, MA, USA;
2
Scienti
fic Researcher, Institute of Medical Technology Assessment, Erasmus University Rotterdam,
Rotterdam, The Netherlands;
3
Research Fellow, University of Shef
field, Sheffield, UK;
4
Professor of Management Science, Department
of Management Science, Strathclyde Business School, University of Strathclyde, Glasgow, Scotland, UK;
5
Assistant Professor, Health
Care Policy & Research, Information and Decision Engineering, Mayo Clinic Kern Center for the Science of Health Care Delivery,
Rochester, MN, USA;
6
Canada Research Chair, Health Services & Systems Research; Arthur J.E. Child Chair in Rheumatology
Research; Director, HTA, Alberta Bone & Joint Health Institute;
7
Associate Professor, Department Community Health Sciences, Faculty
of Sciences, Faculty of Medicine, University of Calgary, Calgary, Alberta, Canada;
8
Senior Health Economist, DRG Abacus,
Manchester, UK;
9
Assistant Professor, Department of Health Policy & Management, Johns Hopkins Bloomberg School of Public Health,
Baltimore, MD, USA;
10
Professor of Clinical Epidemiology & Health Technology Assessment (HTA); Head, Department of Health
Technology & Services Research, University of Twente, Enschede, The Netherlands;
11
Vice President and Chief Performance
Improvement Of
ficer, Illinois Divisions and HSHS Medical Group, Hospital Sisters Health System (HSHS), Belleville, IL. USA;
12
Associate Professor - Healthcare Policy & Research, Lead, Information and Decision Engineering, Mayo Clinic Kern Center for the
Science of Health Care Delivery, Rochester, MN, USA
A B S T R A C T
Providing health services with the greatest possible value to patients
and society given the constraints imposed by patient characteristics,
health care system characteristics, budgets, and so forth relies heavily
on the design of structures and processes. Such problems are complex
and require a rigorous and systematic approach to identify the best
solution. Constrained optimization is a set of methods designed to
identify ef
ficiently and systematically the best solution (the optimal
solution) to a problem characterized by a number of potential
solutions in the presence of identi
fied constraints. This report identifies
1) key concepts and the main steps in building an optimization model; 2)
the types of problems for which optimal solutions can be determined in
real-world health applications; and 3) the appropriate optimization
methods for these problems. We
first present a simple graphical model
based on the treatment of
“regular” and “severe” patients, which max-
imizes the overall health bene
fit subject to time and budget constraints.
We then relate it back to how optimization is relevant in health services
research for addressing present day challenges. We also explain how
these mathematical optimization methods relate to simulation meth-
ods, to standard health economic analysis techniques, and to the
emergent
fields of analytics and machine learning.
Keywords: decision making, care delivery, modeling, policy.
Copyright
& 2017, International Society for Pharmacoeconomics and
Outcomes Research (ISPOR). Published by Elsevier Inc.
Introduction
In common vernacular, the term
“optimal” is often used loosely
in health care applications to refer to any demonstrated superi-
ority among a set of alternatives in speci
fic settings. Seldom is
this term based on evidence that demonstrates that such sol-
utions are, indeed, optimal
—in a mathematical sense. By “opti-
mal
” solution we mean the best possible solution for a given
problem given the complexity of the system inputs, outputs/
outcomes, and constraints (budget limits, staf
fing capacity, etc.).
Failing to identify an
“optimal” solution represents a missed
opportunity to improve clinical outcomes for patients and eco-
nomic ef
ficiency in the delivery of care.
Identifying optimal health system and patient care interven-
tions is within the purview of mathematical optimization models.
There is a growing recognition of the applicability of constrained
optimization methods from operations research to health care
problems. In a review of the literature
[1]
, note more than 200
constrained optimization and simulation studies in health care. For
example, constrained optimization methods have been applied in
problems of capacity management and location selection for both
health care services and medical supplies
[2
–
5]
.
1098-3015$36.00
– see front matter Copyright & 2017, International Society for Pharmacoeconomics and Outcomes Research (ISPOR).
Published by Elsevier Inc.
http://dx.doi.org/10.1016/j.jval.2017.01.013
* Address correspondence to: Kalyan S. Pasupathy, Mayo Clinic, Health Sciences Research, 200 First Street SW, Harwick 2-43, Rochester,
MN 55905.
E-mail:
William.Crown@optum.com; Pasupathy.Kalyan@mayo.edu.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
Constrained optimization is an interdisciplinary subject, cutting
across the boundaries of mathematics, computer science, econom-
ics, and engineering. Analytical foundations for the techniques to
solve the constrained optimization problems involving continuous,
differentiable functions and equality constraints were already laid in
the 18th century
[6]
. However, with advances in computing tech-
nology, constrained optimization methods designed to handle a
broader range of problems trace their origin to the development of
the simplex algorithm
—the most commonly used algorithm to solve
linear constrained optimization problems
—in 1947
[7
–
11]
. Since that
time, various constrained optimization methods have been devel-
oped in the
field of operations research and applied across a wide
range of industries. This creates signi
ficant opportunities for the
optimization of health care delivery systems and for providing value
by transferring knowledge from
fields outside the health care sector.
In addition to capacity management, facility location, and
ef
ficient delivery of supplies, patient scheduling, provider
resource scheduling, and logistics are other substantial areas of
research in the application of constrained optimization methods
to health care
[12
–
16]
. Constrained optimization methods may
also be very useful in guiding clinical decision making in actual
clinical practice where physicians and patients face constraints
such as proximity to treatment centers, health insurance bene
fit
designs, and the limited availability of health resources.
Constrained optimization methods can also be used by health
care systems to identify the optimal allocation of resources across
interventions subject to various types of constraints
[17
–
23]
. These
methods have also been applied to disease diagnosis
[24
,
25]
, the
development of optimal treatment algorithms
[26
,
27]
, and the
optimal design of clinical trials
[28]
. Health technology assessment
using tools from constrained optimization methods is also gaining
popularity in health economics and outcomes research
[29]
.
Recently, the ISPOR Emerging Good Practices Task Force on
Dynamic Simulation Modeling Applications in Health Care Deliv-
ery Research published two reports in Value in Health
[30
,
31]
and
one in Pharmacoeconomics
[32]
on the application of dynamic
simulation modeling (DSM) to evaluate problems in health care
systems. Although simulation can provide a mechanism to
evaluate various scenarios, by design, they do not provide
optimal solutions. The overall objective of the ISPOR Emerging
Good Practices Task Force on Constrained Optimization Methods
is to develop guidance for health services researchers, knowledge
users, and decision makers in the use of operations research
methods to optimize health care delivery and value in the
presence of constraints. Speci
fically, this task force will 1)
introduce constrained optimization methods for conducting
research on health care systems and individual-level outcomes
(both clinical and economic); 2) describe problems for which
constrained optimization methods are appropriate; and 3) iden-
tify good practices for designing, populating, analyzing, testing,
and reporting results from constrained optimization models.
The ISPOR Emerging Good Practices Task Force on Constrained
Optimization Methods will produce two reports. In this
first report,
we introduce readers to constrained optimization methods. We
present de
finitions of important concepts and terminology, and
provide examples of health care decisions in which constrained
optimization methods are already being applied. We also describe
the relationship of constrained optimization methods to health
economic modeling and simulation methods. The second report
will present a series of case studies illustrating the application of
these methods including model building, validation, and use.
De
finition of Constrained Optimization
Constrained optimization is a set of methods designed to ef
ficiently
and systematically
find the best solution to a problem
characterized by a number of potential solutions in the presence
of identi
fied constraints. It entails maximizing or minimizing an
objective function that represents a quanti
fiable measure of
interest to the decision maker, subject to constraints that restrict
the decision maker
’s freedom of action. Maximizing/minimizing
the objective function is carried out by systematically selecting
input values for the decision from an allowed set and computing
the objective function, in an iterative manner, until the decision
yields the best value for the objective function, a.k.a optimum.
The decision that gives the optimum is called the
“optimal
solution.
” In some optimization problems, two or more different
decisions may yield the same optimum. Note that, programming
and optimization are often used as interchangeable terms in the
literature, for example, linear programming and linear optimiza-
tion. Historically, programming referred to the mathematical
description of a plan/schedule, and optimization referred to the
process used to achieve the optimal solution described in the
program.
The components of a constrained optimization problem are its
objective function(s), its decision variable(s), and its constraint(s).
The objective function is a function of the decision variables that
represents the quantitative measure that the decision maker
aims to minimize/maximize. Decision variables are mathematical
representation of the constituents of the system for which
decisions are being taken to improve the value of the objective
function. The constraints are the restrictions on decision variables,
often pertaining to resources. These restrictions are de
fined by
equalities/inequalities involving functions of decision variables.
They determine the allowable/feasible values for the decision
variables. In addition, parameters are constant values used in
objective function and constraints, like the multipliers for the
decision variables or bounds in constraints. Each parameter
represents an aspect of the decision-making context: for exam-
ple, a multiplier may refer to the cost of a treatment.
A Simple Illustration of a Constrained Optimization
Problem
Imagine you are the manager of a health care center, and your
aim is to bene
fit as many patients as possible. Let us say, for the
sake of simplicity, you have two types of patients
—regular and
severe patients, and the demand for the health service is
unlimited for both these types. Regular patients can achieve
two units of health bene
fits and severe ones can achieve three
units. Each patient, irrespective of severity, takes 15 minutes for
consultation; only one patient can be seen at any given point in
time. You have 1 hour of total time at your disposal. Regular
patients require $25 of medications, and severe patients require
$50 of medications. You have a total budget of $150. What is the
greatest health bene
fit this center can achieve given these inputs
and constraints?
At the outset, this problem seems straightforward. One might
decide on four regular patients to use up all the time that is
available. This will achieve eight units of health bene
fit while
leaving $50 as excess budget. An alternate approach might be to
see as many severe patients as possible because treating each
severe patient generates more per capita health bene
fits. Three
patients (totaling $150) would generate nine health units leaving
15 minutes extra time unused. There are other combinations of
regular and severe patients that would generate different levels
of health bene
fits and use resources differently.
This is graphically represented in
Figure 1
, with regular
patients on the x-axis and severe patients on the y-axis. Line
CF is the time constraint limiting total time to 1 hour. Line BG is
the budget constraint limiting to $150. Any point to the south-
west of these constraints (lines), respectively, will ensure that
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
311
time and budget do not exceed the respective limits. The
combination of these, together with non-negativity of the deci-
sion variables, gives the feasible region.
The lines AB-BD-DF-FA form the boundary of the feasibility
space, shown shaded in the
figure. In problems that are three- or
more dimensional, these lines would be hyperplanes. To obtain the
optimal solution, the dashed line is established, the slope depends
on the relative health units of the two decision variables (i.e., the
number of regular and severe patients seen). This dashed line moves
from the origin in the northeast direction as shown by the arrow.
The optimal solution is two regular patients and two severe patients.
This approach uses the entire 1-hour time as well as the $150 budget.
Because regular and severe patients achieve two- and three-unit
health bene
fits, respectively, we are able to achieve 10 units of health
bene
fit and still meet the time and budget constraints.
No other combination of patients is capable of achieving more
bene
fits while still meeting the time and budget constraints. Note
that not all resource constraints have to be completely used to
attain the optimal solution. This hypothetical example is a small-
scale problem with only two decision variables; the number of
regular and severe patients seen. Hence, they can be represented
graphically with one variable on each axis.
With the dif
ficulty in representing larger problems graphi-
cally, we turn to mathematical approaches, such as the simplex
algorithm, to
find the solutions. The simplex algorithm is a
structured approach of navigating the boundary (represented as
lines in two dimensions and hyperplanes in three or more
dimensions) of the feasibility space to arrive at the optimal
solution.
Table 1
summarizes the main components of the
example and notes several other dimensions of complexity
(linear vs nonlinear, deterministic vs stochastic, static vs
dynamic, discrete/integer vs continuous) that can be incorpo-
rated into constrained optimization models.
The mathematical formulation of the model is as follows:
Max
f
R
x
R
þ f
L
x
L
(objective function)
subject to c
R
x
R
þ c
L
x
L
r B (budget constraint)
t
R
x
R
þ t
L
x
L
r T (time constraint)
x
R
,x
L
Z 0 and integer (decision variables)
where c
R
,c
L
is the cost of regular and severe patients, respec-
tively; B is the total budget available; t
R
,t
L
is time to see regular
and severe patients, respectively; T is the total time available; f
R
,f
L
is health bene
fits of regular and severe patients, respectively; and
x
R
,x
L
is the number of regular and severe patients, respectively.
In the current version of the problem, the parameters are as
follows:
f
R
¼ two health benefit units, f
L
¼ three health benefit units
c
R
¼ $25, c
L
¼ $50, B ¼ $150
t
R
¼ 0.25 hours, t
L
¼ 0.25 hours, T ¼ 1 hour
So the model is as follows:
Max
2 x
R
þ 3x
L
(objective function)
subject to 25x
R
þ 50x
L
r 150 (budget constraint)
0.25x
R
þ 0.25x
L
r 1 (time constraint)
x
R
,x
L
Z 0 and integer
As described above,
Figure 1
illustrates the graphical solution to
this model. However, problems with higher dimensionality must use
mathematical algorithms to identify the optimal solution. The
problem described above falls into the category of linear optimization,
because although the constraints and the objective function are
linear from an algebraic standpoint, the decision variables must be in
the form of integers. As will be discussed further in the section
“Steps
in a Constrained Optimization Process,
” there are other optimization
modeling frameworks, such as combinatorial, nonlinear, stochastic,
and dynamic optimization.
Because the algorithms for integer optimization problems can
take much longer to solve computationally than those for linear
optimization problems, one alternative is to set the integer
optimization problem up and solve it as a linear one. If fractional
values are obtained, the nearest feasible integers can be used as
the
final solution. This should be done with caution, however.
First, rounding the solution to the nearest integers can result in
an infeasible solution or, and second, even if the rounded
solution is feasible, it may not be the optimal solution to the
original integer optimization problem. Nonlinear optimization is
suitable when the constraints or the objective function is non-
linear. In problems in which there is uncertainty, such as the
estimated health bene
fit that each patient might receive in the
above example, stochastic optimization techniques can be used.
Dynamic optimization (known commonly as dynamic programming)
formulation might be useful when the optimization problem is
not static, the problem context and parameters change in time,
and there is an interdependency among the decisions at different
time periods (for instance, when decisions made at a given time
interval, say number of patients to be seen now, affects the
decisions for other time periods, such as the number of patients
to be seen tomorrow).
Table 1
summarizes the model compo-
nents in the hypothetical problem, relates it to health services
with examples, and identi
fies the specific terminology.
Problems That Can Be Tackled with Constrained
Optimization Approaches
In this section, we list several areas within health care in which
constrained optimization methods have been used in health
services. The selected examples do not represent a comprehen-
sive picture of this
field, but provide the reader a sense of what is
possible. In
Table 2
, we compare problems using the terminology
of the previous section, with respect to decision makers, deci-
sions, objectives, and constraints.
Steps in a Constrained Optimization Process
An overview of the main steps involved in a constrained optimiza-
tion process
[33]
is described here and presented in
Table 3
. Some of
the steps are common to other types of modeling methods. It is
important to emphasize that the process of optimization is iterative,
rather than comprising a strictly sequential set of steps.
Problem Structuring
This involves specifying the objective, that is, goal, and identifying
the decision variables, parameters, and the constraints involved.
Fig. 1
– Graphical representation of solving a simple integer
programming problem.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
312
These can be speci
fied using words, ideally in nontechnical
language, so that the optimization problem is easily understood.
This step needs to be performed in collaboration with all the
relevant stakeholders, including decision makers, to ensure all
aspects of the optimization problem are captured. As with any
modeling technique, it is also crucial to surface key modeling
assumptions and appraise them for plausibility and materiality.
Mathematical Formulation
After the optimization problem is speci
fied in words, it needs to
be converted into mathematical notation. The standard mathe-
matical notation for any optimization problem involves specify-
ing the objective function and constraint(s) using decision
variables and parameters. This also involves specifying whether
the goal is to maximize or minimize the objective function. The
standard notation for any optimization problem, assuming the
goal is to maximize the objective, is as shown below:
Maximize z
¼ f(x
1
, x
2
,
…, x
n,
p
1
, p
2
,
…, p
k
)
subject to
c
j
(x
1
, x
2
,
…, x
n,
p
1
, p
2
,
…, p
k
)
r C
j
for j
¼ 1, 2,…, m
where, x
1
, x
2
,
…, x
n
are the decision variables, f(x
1
, x
2
,
…, x
n
) is
the objective function, and c
j
(x
1
, x
2
,
…, x
n,
p
1
, p
2
,
…, p
k
)
r C
j
represent the constraints. Note that the constraints can include
both inequality and equality constraints and that the objective
function and the constraints also include parameters p
1
, p
2
,
…, p
k
,
which are not varied in the optimization problem. Speci
fication of
the optimization problem in this mathematical notation allows
clear identi
fication of the type (and number) of decision variables,
parameters, and the constraints. Describing the model in math-
ematical form will be useful to support model development.
Model Development
The next step after mathematical formulation is model develop-
ment. Model development involves solving the mathematical
problem described in the previous step, and often performed
iteratively. The model should estimate the objective function and
the left-hand side values of the constraints, using the decision
variables and parameters as inputs. The complexity of the model
can vary widely. Similar to other types of modeling, the complex-
ity of the model will depend on the outputs required, and the
level of detail included in the model, whether it is linear or
nonlinear, stochastic or deterministic, static or dynamic.
Perform Model Validation
As with any modeling, it is important to ensure that the model
developed represents reality with an acceptable degree of
fidelity
[33]
. The requirements of model validation for optimization are
more stringent than for, for example, simulation models, due to
the need for the model to be valid for all possible combinations of
the decision variables. Thus, appropriate caution needs to be
taken to ensure that the model assumptions are valid and that
the model produces sensible results for the different scenarios. At
the very least, the validation should involve checking of the face
validity (i.e., experts evaluate model structure, data sources,
Table 1
– Model summary and extensions.
Hypothetical problem
Real-life health services
Terminology
Aim
Maximize health/health care bene
fits
Maximize health/health care bene
fits
Objective function
Options
available
Regular or severe patients
Service lines, case mix, service mix, etc.
Decision variables
Constraints
Total cost
r$15
Budget constraint
Constraints
Total time
r1 h
Time constraint
Resource constraint (e.g., staff and beds)
Evidence
base
Cost of each patient, health bene
fits of each
patient, and the time taken for consultation
Costs, health bene
fits, and other relevant data
associated with each intervention to be
selected
Model (to
determine the
objective
function and
constraints)
Complexity
Static
Dynamic
Optimization
method
The problem does not have a time component;
decision made in one time period does not
affect decisions made in another
The optimization problem and parameters may
change in different time points, and the
decision made at any point in time can affect
decisions at later time points (e.g., there can
be a capacity constraint de
fined on 2 mo,
whereas the planning cycle is 1 mo)
Deterministic
Stochastic
All the information is assumed to be certain (e.
g., cost of each patient, health bene
fits of
each patient, and the time taken for
consultation)
Know that the information is uncertain (i.e.,
uncertainty in the costs and bene
fits of the
interventions)
Linear (i.e., each additional patient costs the
same and achieves same health bene
fits)
Nonlinear (objective function or constraints may
have a nonlinear relationship with the model
parameters, e.g., total costs and QALYs
typically have a nonlinear relationship with
the model parameters)
Integer/discrete
Continuous
The decision variables (number of patients) can
take only discrete and integer values
The decision variables can take fractional
values (e.g., number of hours)
QALYs, quality-adjusted life-years.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
313
assumptions, and results), and veri
fication or internal validity
(i.e., checking accuracy of coding).
Select Optimization Method
This step involves choosing the appropriate optimization
method, which is dependent on the type of optimization problem
that is addressed. Optimization problems can be broadly classi-
fied, depending on the nature of the objective functions and the
constraints, for example, into linear versus nonlinear, determin-
istic versus stochastic, continuous versus discrete, or single
versus multiobjective optimization. For instance, if the objective
function and constraints consist of linear functions only, the
corresponding problem is a linear optimization problem. Simi-
larly, in deterministic optimization, the parameters used in the
optimization problem are
fixed whereas in stochastic optimiza-
tion, uncertainty is incorporated. Optimization problems can be
continuous (i.e., decision variables are allowed to have fractional
values) or discrete (e.g., a hospital ward may be either open or
closed; the number of computed tomography scanners that a
hospital buys must be a whole number).
Most optimization problems have a single objective function;
however, when optimization problems have multiple con
flicting
objective functions, they are referred to as multiobjective opti-
mization problems. The optimization method chosen needs to be
in line with the type of optimization problem under consider-
ation. Once the optimization problem type is clear (e.g., discrete
or nonlinear), a number of texts may be consulted for details on
solution methods appropriate for that problem type
[33
–
36]
.
Broadly speaking, optimization methods can be categorized
into exact approaches and heuristic approaches. Exact approaches
iteratively converge to an optimal solution. Examples of these
include simplex methods for linear programming and the New-
ton method or interior point method for nonlinear programming
[34
,
37]
. Heuristic approaches provide approximate solutions to
optimization problems when an exact approach is unavailable or
is computationally expensive. Examples of these techniques
include relaxation approaches, evolutionary algorithms (such as
genetic algorithms), simulated annealing, swarm optimization,
ant colony optimization, and tabu-search. Besides these two
approaches (i.e., exact or heuristic), other methods are available
to tackle large-scale problems as well (e.g., decomposition of the
large problems to smaller subproblems).
There are software programs that help with optimization;
interested readers are referred to the Web site of the Institute
for Operations Research and the Management Sciences
Table 2
– Examples of health care decisions for which constrained optimization is applicable.
Type of health
care problem
Typical decision
makers
Typical decisions
Typical objectives
Typical constraints
Resource
allocation
within and
across disease
programs
Health authorities,
insurance funds
List of interventions to be funded
Maximize population
health
Overall health budget,
other legal constraints
for equity
Resource
allocation for
infectious
disease
management
Public health
agencies, health
protection
agencies
Optimal vaccination coverage
level
Minimize disease
outbreaks and total
costs
Availability of medicines,
disease dynamics of the
epidemic
Allocation of
donated
organs
Organ banks,
transplant service
centers
Matching of organs and
recipients
Maximize matching of
organ donors with
potential recipients
Every organ can be
received by at most one
person
Radiation
treatment
planning
Radiation therapy
providers
Positioning and intensity of
radiation beams
Minimizing the
radiation on healthy
anatomy
Tumor coverage and
restriction on total
average dosage
Disease
management
models
Leads for a given
disease
management plan
Best interventions to be funded,
best timing for the initiation of
a medication, best screening
policies
Identify the best plan
using a whole
disease model,
maximizing QALYs
Budget for a given disease
or capacity constraints
for health care providers
Workforce
planning/
staf
fing/shift
template
optimization
Hospital managers,
all medical
departments (e.g.,
ED and nursing)
Number of staff at different
hours of the day, shift times
Increase ef
ficiency and
maximize
utilization of health
care staff
Availability of staff,
human factors, state
laws (e.g., nurse-to-
patient ratios), budget
Inpatient
scheduling
Operation room/ICU
planners
Detailed schedules
Minimize waiting time
Availability of beds, staff
Outpatient
scheduling
Clinical department
managers
Detailed schedules
Minimize
overutilization and
underutilization of
health care staff
Availability of
appointment slots
Hospital facility
location
Strategic health
planners
Set of physical sites for hospitals
Ensure equitable
access to hospitals
Maximum acceptable
travel time to reach a
hospital
ED, emergency department; ICU, intensive care unit; QALYs, quality-adjusted life-years.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
314
(
www.informs.org
) for a list of optimization software. The users
need to specify, and more importantly understand, the parame-
ters used as an input for these optimization algorithms (e.g., the
termination criteria such as the level of convergence required or
the number of iterations).
Perform Optimization/Sensitivity Analysis
Optimization involves systematically searching the feasible
region for values of decision variables and evaluating the objec-
tive function, consecutively, to
find a combination of decision
variables that achieve the maximum or minimum value of the
objective function, using speci
fic algorithms. Once the optimiza-
tion algorithm has
finished running, in some cases, the identified
solution can be checked to verify that it satis
fies the “optimality
conditions
” (i.e., Karush-Kuhn-Tucker conditions)
[38]
, which are
the mathematical conditions that de
fine the optimality. Once the
optimality is con
firmed, the results need to be interpreted.
First, the results should be checked to see whether there is
actually a feasible solution to the optimization problem, that is,
whether there is a solution that satis
fies all the constraints. If
not, then the optimization problem needs to be adjusted (e.g.,
relaxing some constraints or adding other decision variables) to
broaden the feasible solution space. If a feasible optimal solution
has been found, the results need to be understood
—this involves
interpretation of the results to check whether the optimal
solution, that is, values of decision variables, constraints, and
objective function, makes sense.
It is also good practice to repeat the optimization with differ-
ent sets of starting decision variables to ensure the optimal
solution is the global optimum rather than local optimum.
Sometimes, there may be multiple optimal solutions for the
same problem (i.e., multiple combinations of decision variables
that provide the same optimal value of objective function). For
multiobjective optimization problems (i.e., problems with two or
more con
flicting objectives), Pareto optimal solutions are con-
structed from which optimal solution can be identi
fied on the
basis of subjective preferences of the decision maker
[39
,
40]
.
It is good practice to run the optimization problem using
different values of parameters, to verify the robustness of the
optimization results. Sensitivity analysis is an important part of
building con
fidence in an optimization model, addressing the
structural and parametric uncertainties in the model by analyz-
ing how the decision variables and optimum value react to
changes in the parameters in the constraints and objective
function, which ensures that the optimization model and its
solution are good representations of the problem at hand.
Sometimes a solution may be the mathematically optimal
solution to the speci
fied mathematical problem, but may not be
practically implementable. For example, the
“optimal” set of
nurse rosters may be unacceptable to staff as it involves breaking
up existing teams, deploying staff with family responsibilities on
night shifts, or reducing overtime pay to level where the employ-
ment is no longer attractive. Analysts should resist the tempta-
tion
to
spring
their
optimal
solution
on
unsuspecting
stakeholders,
expecting
grateful
acceptance:
rather,
those
affected by the model should be kept in the loop through the
modeling process. The optimal solution may come as a surprise:
it is important to allow space in the modeling process to explore
fully and openly concerns about whether the
“optimal” solution
is indeed the one the organization should implement.
Report Results
The
final optimal solution, and if applicable, the results of the
sensitivity analyses, should be reported. This will include the
results of the optimum
“objective function” achieved and the set
of
“decision variables” at which the optimal solution is found.
Both the numerical values (i.e., the mathematical solution) and
the physical interpretation, that is, the nontechnical text describ-
ing the meaning of numerical values, should be presented. The
optimal solution identi
fied can be contextualized in terms of how
much
“better” it is compared with the current state. For example,
the results can be presented as improvement in bene
fits such as
quality-adjusted life-years or reduction in costs.
It is often necessary to report the optimization method used
and the results of the
“performance” of the optimization algo-
rithm, for example, number of iterations to the solution, compu-
tational time, and convergence level. This is important because it
helps users understand whether a particular algorithm can be
used
“online” in a responsive fashion, or only when there is
signi
ficant time available, for example, in a planning context.
Table 3
– Steps in an optimization process.
Stage
Step
Description
Modeling
Problem
structuring
Specify the objective and
constraints, identify
decision variables and
parameters, and list and
appraise model
assumptions
Mathematical
formulation
Present the objective
function and constraints
in mathematical
notation using decision
variables and
parameters
Model
development
Develop the model;
representing the
objective function and
constraints in
mathematical notation
using decision variables
and parameters
Model
validation
Ensure the model is
appropriate for
evaluating all possible
scenarios (i.e., different
combinations of
decision variables and
parameters)
Optimization
Select
optimization
method
Choose an appropriate
optimization method
and algorithm on the
basis of characteristics
of the problem
Perform
optimization/
sensitivity
analysis
Use the optimization
algorithm to search for
the optimal solution and
examine the
performance of the
optimal solution for
reasonable values of
parameters
Report results
Report the results of
optimal solution and
sensitivity analyses
Decision
making
Interpret the optimal
solution and use it for
decision making
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
315
Dashboards can be useful to visualize these bene
fits and com-
municate the insights gained from the optimal solution and
sensitivity analyses.
Decision Making
The
final optimal solution and its implications for policy/service
recon
figuration should be presented to all the relevant stakeholders.
This typically involves a plan for amending the
“decision variables”
(e.g., shift patterns, screening frequency
—see
Table 2
for examples of
decision variables
—to those identified in the optimal solution).
Before an optimal solution can be implemented, it will require
getting the
“buy-in” from the decision makers and all the stake-
holders, for example, frontline staff such as nurses and hospital
managers, to ensure that the numerical
“optimal” solution found can
be operationalized in a
“real” clinical setting. It is important to have
the involvement of decision makers throughout the whole optimi-
zation process to ensure that it does not become a purely numerical
exercise, but rather something that is implemented in real life. After
the decision is made, data should still be collected to assess the
ef
ficiency and demonstrate the benefits of the implementation of the
optimal solution.
If decision makers are not directly involved in model develop-
ment, they may choose not to implement the
“optimal” solution as it
comes from the model. This is because the model may fail to capture
key aspects of the problem (e.g., the model may maximize aggregate
health bene
fits but the decision maker may have a specific concern
for health bene
fits for some disadvantaged subgroup). This does not
(necessarily) mean that the optimization modeling has not been
useful
—enabling a decision maker to see how much health benefit
must be sacri
ficed to satisfy her equity objective may prove to be
bene
ficial toward the overall objective. After the decision is made,
the story does not come to an end: data should continue to be
collected to demonstrate the bene
fits of whatever solution is
implemented, as well as guiding future decision making.
Table 3
presents the two different stages in optimization, that
is, the modeling stage and the optimization stage, highlighting
that model development is necessary before optimization can be
performed. The goal of constrained optimization is to identify an
optimal solution that maximizes or minimizes a particular
objective subject to existing constraints.
Relationship of Constrained Optimization to Related
Fields
The use of constrained optimization can be classi
fied into two
categories. The
first category is the use of constrained optimiza-
tion as a decision-making tool. The simple illustration in the
section
“A Simple Illustration of a Constrained Optimization
Problem
” and all the examples in the section “Problems That
Can Be Tackled with Constrained Optimization Approaches
” are
considered to fall under this category. The second category is the
use of constrained optimization as an auxiliary analysis tool. In this
category, optimization is an embedded tool and the results of which
are often not the end results of a decision problem, but rather they
are used as inputs for other analysis/modeling methods (e.g.,
optimization used in the multiple criteria decision making; in
calibrating the inputs for health economic or dynamic simulation
models; in machine learning and other statistical analysis methods
such as solving regression models or propensity score matching).
As a decision-making tool, optimization is complementary to
other modeling methods such as health economic modeling,
simulation modeling, and descriptive, predictive (e.g., machine
learning), and prescriptive analytics. Most modeling methods
typically evaluate only a few different scenarios and determine
a good scenario within the available options. In contrast, the aim
of optimization methods is to ef
ficiently identify the best solution
overall, given the constraints. In the absence of using optimiza-
tion methods, a brute force approach, in which all possible
options are sequentially evaluated and the best solution is
identi
fied among them, might be possible for some problems.
However, for most problems, it is too complex and too time
consuming to identify and evaluate all possible options. Optimi-
zation methods and heuristic approaches might use ef
ficient
algorithms to identify the optimal solution quickly, which would
otherwise be very dif
ficult and time consuming.
Also, model development using these other methods might be
necessary before optimization, especially in situations in which
the objective function or constraints cannot be represented in a
simple functional form. Thus, all models currently used in health
care such as health economic models, dynamic simulation
models, and predictive analytics (including machine learning)
can be used in conjunction with optimization methods.
Constrained Optimization Methods Compared with
Traditional Health Economic Modeling in Health Technology
Assessments
Constrained optimization methods differ substantially from
health economic modeling methods traditionally used in health
technology assessment processes
[41]
. The main difference
between the two approaches is that traditional health economic
modeling approaches, such as Markov models, are built to
estimate the costs and effects of different diagnostic and treat-
ment options. If decision makers are basing their judgments on
modeling results, they may not formally consider the constraints
and resource implications in the system. Constrained optimiza-
tion methods provide a structured approach to optimize the
decision problem and to present the best alternatives given an
optimization criterion, such as constrained budget or availability
of resources.
These differences have major implications. There is an oppor-
tunity to learn from optimization methods to improve health
technology assessment processes
[42
–
46]
. Optimization is a
valuable means of capturing the dynamics and complexity of
the health system to inform decision making for several reasons.
Constrained optimization methods can do the following:
1. Explicitly take budget constraints into account: Informed decision
making about resource allocation requires an external esti-
mate of the decision maker
’s willingness to pay for a unit of
health outcome, the threshold. Decision making based on
traditional health economic models then relies on the princi-
ple that by repeatedly applying the threshold to individual
health technology assessment decisions, optimization of the
allocation of health resources will be achieved.
However, the focus of health economics is usually about relative
ef
ficiency without explicit consideration of budget because
many jurisdictions do not explicitly implement a constrained
budget nor do they use mechanisms to evaluate retrospectively
cost-effectiveness of medical technologies currently in use.
2. Address multiple resource constraints in the health system, such as
resource capacity: Constrained optimization methods also allow
consideration of the effect of other constraints in the health
system, such as capacity or short-term inef
ficiencies. Capacity
constraints are usually neglected in health economic models.
In health economics models, the outcomes are central to
decision makers while the process to arrive at these outcomes
is most of the time ignored.
For health policy makers and health care planners, such
capacity considerations are critical and cannot be neglected.
Likewise, some technologies are known for short-term inef
fi-
ciencies; for example, large equipment such as positron
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
316
emission tomography/magnetic resonance imaging devices
are usually not taken into consideration. It takes a certain
amount of time before a new device operates ef
ficiently, and
such short-term inef
ficiencies do influence implementation
[47]
.
3. Account for system behavior and decisions over time: Traditional
health economic models are often limited to informing a decision
of a single technology at a single point in time. Health economic
models with a clinical perspective, such as a whole disease model
[48
,
49]
, or a treatment sequencing model, may allow the full
clinical pathway to be framed as a constrained optimization
problem that accounts for both intended and unintended con-
sequences of health system interventions over time with feed-
back mechanisms in the system.
Each combination of decisions within the pathway can be a
potential solution, constrained by the feasibility of each decision,
for example, the licensed indication for various treatments within
a clinical pathway. These whole disease and treatment sequenc-
ing models can evaluate alternative guidance con
figurations and
report the performance in terms of an objective function (cost per
quality-adjusted life-year, net monetary bene
fit)
[50
,
51]
.
4. Inform decision makers about implementability of solutions that are
recommended: Health economic models are not typically con-
strained
—it is assumed that resources are available as
required and are thus affordable; similarly, the evidence used
in the models come from controlled clinical settings, which
are idealized settings compared with real clinical setting. An
advantage of constrained optimization is the ability to obtain
optimal solutions to decision problems and have sensitivity
analyses performed. Such analyses inform decision makers
about alternate realistic solutions that are feasible and close
to the optimal solution.
Thus, in some sense, classic health economics models are
“hypothetical” to illustrate the potential value as measured by a
speci
fic outcome with respect to cost, whereas optimization is
focused on what can be achieved in an operational context. This
suggests that constrained optimization methods have great value
for informing decisions about the ability to implement a clinical
intervention, program, or policy as they actually consider these
constraints in the modeling approach.
Constrained Optimization Methods Compared with Dynamic
Simulation Models
DSM methods, such as system dynamics, discrete event simu-
lation, and agent-based modeling, are used to design and develop
mathematical representations, that is, formal models, of the
operation of processes and systems. They are used to experiment
with and test interventions and scenarios and their consequen-
ces over time to advance the understanding of the system or
process, communicate
findings, and inform management and
policy design
[30
–
32
,
52
–
54]
. These methods have been broadly
used in health applications
[55
–
57]
.
Unlike constrained optimization methods, DSMs do not pro-
duce a speci
fic solution. Rather they allow for the evaluation of a
range of possible or feasible scenarios or intervention options
that may or may not improve the system
’s performance. Con-
strained optimization methods, in general, seek to provide the
answer to which of those options is the
“best.” Hence, the types
of problems and questions that can be addressed with DSMs
[30
–
32]
are different from those that are addressed with optimization
methods. However, both types of methods can be complementary
to each other in helping us to better understand systems.
Traditionally, constrained optimization methods have served
two distinct purposes in DSM development. 1) model calibration
—fitting suitable model variables to past time series is discussed
elsewhere
[30
–
32]
; 2) evaluating a policy
’s performance/effect
relative to a criterion or set of criteria. However, the complexity
of DSMs compared with simple analytic models may render exact
constrained optimization approaches cumbersome, inappropri-
ate, and potentially infeasible because of the large search space,
for example, using methods of optimal control.
Because of this complexity, alternatives to exact approaches
such as heuristic search strategies are available. Historically,
these types of methods have been used in system dynamics
and other DSMs. Because of their heuristic nature, there is no
certainty of
finding the “best” or optimal parameter set rather
“good enough” solutions. Hence, the ranges assigned need careful
consideration to get
“good” solutions, that is, previous knowledge
of sensible ranges both from knowledge about the system and
from knowledge gained from model building.
Optimization is used as part of system dynamics to gain
insight about policy design and strategy design, particularly
when the traditional analysis of feedback mechanisms becomes
risky due to the large number of loops in a model
[58]
. Similar
procedures to evaluate policies and strategies can be used in
discrete event simulation and agent-based modeling, for exam-
ple, simulated annealing algorithms and genetic algorithms.
Constrained Optimization Methods as Part of Analytics
Constrained optimization methods fall within the area of ana-
lytics as de
fined by the Institute for Operations Research and the
Wilson, ISPOR 2014
Future states
Optimal
decision
Optimization
Fig. 2
– Descriptive, predictive, and prescriptive analytics.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
317
Management Sciences (
https://www.informs.org/Sites/Getting-
Started-With-Analytics
). Analytics can be classi
fied into descrip-
tive, predictive, and prescriptive analytics (
Fig. 2
), and discussed
below. Constrained optimization methods are a special form of
prescriptive analytics.
1. Descriptive analytics concern the use of historical data to
describe a phenomenon of interest, with a particular focus
on visual displays of patterns in the data. Descriptive ana-
lytics is differentiated from descriptive analysis, which uses
statistical methods to test hypotheses about relationships
among variables in the data. Health services research typically
uses theory and concepts to identify hypotheses, and histor-
ical data are used to test these hypotheses using statistical
methods. Examples may include natural history of aging,
disease progression, evaluation of clinical interventions, pol-
icy interventions, and many others. Traditional health serv-
ices for the most part falls within the area of descriptive
analytics.
2. Predictive analytics and machine learning focus on forecasting
the future states of disease or states of systems. With the
increased volume and dimensions of health care data, espe-
cially medical claims and electronic medical record data, and
the ability to link to other information such as feeds from
personal devices and sociodemographic data, big data meth-
ods such as machine learning are garnering increased atten-
tion
[59]
.
Machine learning methods, such as predictive modeling
and clustering, have an important intersection with con-
strained optimization methods. Machine learning methods
are valuable for addressing problems involving classi
fica-
tion, as well as data dimension reduction issues. And
maybe most importantly, optimization often needs fore-
casts and estimates as inputs, which can be obtained from
the results of machine learning algorithms. A discussion of
machine learning methods is beyond the scope of this
article.
However, the interested reader will
find a detailed intro-
duction elsewhere
[60
,
61]
. Machine learning has the ability
to
“mine” data sets and discover trends or patterns. These
are often valuable to establish thresholds or parameter
values in optimization models, where it is otherwise
dif
ficult to determine the values. Constrained optimization
can also leverage the ability of machine learning to reduce
high dimensionality of data, say with thousands or mil-
lions of variables to key variables.
3. Prescriptive analytics uses the understanding of systems, both
the historical and future based on historical (descriptive) and
predictive analytics, respectively, to determine future course
of action/decisions. Traditional (without optimization) clinical
trials and interventions fall under the category of prescriptive
analytics (
“Change what will happen” in figure). Constrained
optimization is a specialized form of prescriptive analytics
because it helps with determining the optimal decision or
course of action in the presence of constraints (
https://www.
informs.org/Sites/Getting-Started-With-Analytics/
Analytics-Success-Stories
).
Summary and Conclusions
This is the
first report of the ISPOR Constrained Optimization
Methods Emerging Good Practices Task Force. It introduces read-
ers to the application of constrained optimization methods to
health care systems and patient outcomes research problems.
Such methods provide a means of identifying the best policy
choice or clinical intervention given a speci
fic goal and given a
speci
fied set of constraints. Constrained optimization methods
are already widely used in health care in areas such as choosing
the optimal location for new facilities and making the most
ef
ficient use of operating room capacity.
However, they have been less widely used for decision making
about clinical interventions for patients. Constrained optimization
methods are highly complementary to traditional health economic
modeling methods and DSM, providing a systematic and ef
ficient
method for selecting the best policy or clinical alternative in the face
of large numbers of decision variables, constraints, and potential
solutions. As health care data continues to rapidly evolve in terms of
volume, velocity, and complexity, we expect that machine learning
techniques will also be increasingly used for the development of
models that can subsequently be optimized.
In this report, we introduce readers to the vocabulary of
constrained optimization models and outline a broad set of
models available to analysts for a range of health care problems.
We illustrate the formulation of a linear program to maximize the
health bene
fit generated in treating a mix of “regular” and
“severe” patients subject to time and budget constraints and
solve the problem graphically. Although simple, this example
illustrates many of the key features of constrained optimization
problems that would commonly be encountered in health care.
In the second task force report, we describe several case
studies that illustrate the formulation, estimation, evaluation,
and use of constrained optimization models. The purpose is to
illustrate actual applications of constrained optimization prob-
lems in health care that are more complex than the simple
example described in the current article and make recommen-
dations on emerging good practices for the use of optimization
methods in health care research.
R E F E R E N C E S
[1] Rais A, Viana A. Operations research in healthcare: a survey. Int
Transact Oper Res 2011;18:1
–31.
[2] Araz C, Selim H, Ozkarahan I. A fuzzy multi-objective covering-based
vehicle location model for emergency services. Comput Oper Res
2007;34:705
–26.
[3] Bruni ME, Conforti D, Sicilia N, et al. A new organ transplantation
location
–allocation policy: a case study of Italy. Health Care Manag Sci
2006;9:125
–42.
[4] Ndiaye M, Alfares H. Modeling health care facility location for moving
population groups. Comput Oper Res 2008;35:2154
–61.
[5] Verter V, Lapierre SD. Location of preventive health care facilities. Ann
Oper Res 2002;110:123
–32.
[6] Lagrange JL. Théorie des fonctions analytiques: contenant les principes
du calcul différentiel, dégagés de toute considération d'in
finiment
petits, d'évanouissans, de limites et de
fluxions, et réduits à l'analyse
algébrique des quantités
finies. Ve. Courcier, 1813.
[7] Dantzig GB. Programming of interdependent activities, II: mathematical
model. Econometrica 1949;17:200
–11.
[8] Dantzig GB. Linear Programming and Extensions. Princeton, NJ:
Princeton University Press, 1963.
[9] Gass SI, Assad AA. An Annotated Timeline of Operations Research: An
Informal History. Boston, MA: Springer Science & Business Media, 2005.
[10] Kirby MW. Operational Research in War and Peace: The British Experience
from the 1930s to 1970. London, UK: Imperial College Press, 2003.
[11] Wood MK, Dantzig GB. Programming of interdependent activities, I:
general discussion. Econometrica 1949;17:193
–9.
[12] Burke EK, Causmaecker PD, Petrovic S, et al. Metaheuristics for
handling time interval coverage constraints in nurse scheduling. Appl
Arti
ficial Intelligence 2006;20:743–66.
[13] Cheang B, Li H, Lim A, et al. Nurse rostering problems
––a bibliographic
survey. Eur J Oper Res 2003;151:447
–60.
[14] Lin R-C, Sir MY, Pasupathy KS. Multi-objective simulation optimization
using data envelopment analysis and genetic algorithm: speci
fic
application to determining optimal resource levels in surgical services.
Omega 2013;41:881
–92.
[15] Lin R-C, Sir MY, Sisikoglu E, et al. Optimal nurse scheduling based on
quantitative models of work-related fatigue. IIE Transact Healthcare
Systems Eng 2013;3:23
–38.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
318
[16] Sir MY, Dundar B, Steege LMB, et al. Nurse
–patient assignment models
considering patient acuity metrics and nurses
’ perceived workload.
J Biomedical Inform 2015;55:237
–48.
[17] Alistar SS, Long EF, Brandeau ML, et al. HIV epidemic control
—a model
for optimal allocation of prevention and treatment resources. Health
Care Manag Sci 2014;17:162
–81.
[18] Earnshaw SR, Dennett SL. Integer/linear mathematical programming
models. Pharmacoeconomics 2003;21:839
–51.
[19] Earnshaw SR, Richter A, Sorensen SW, et al. Optimal allocation of
resources across four interventions for type 2 diabetes. Med Decis
Making 2002;22:s80
–91.
[20] Lasry A, Sansom SL, Hicks KA, et al. A model for allocating CDC
’s HIV
prevention resources in the United States. Health Care Manag Sci
2011;14:115
–24.
[21] Stinnett AA, Paltiel AD. Mathematical programming for the ef
ficient
allocation of health care resources. J Health Econ 1996;15:641
–53.
[22] Thomas BG, Bollapragada S, Akbay K, et al. Automated bed
assignments in a complex and dynamic hospital environment.
Interfaces 2013;43:435
–48.
[23] Zaric GS, Brandeau ML. A little planning goes a long way: multilevel
allocation of HIV prevention resources. Med Decis Making 2007;27:71
–81.
[24] Lee EK, Wu TL. Disease diagnosis: Optimization-based methods. In:
Floudas CA, Pardalos PM, eds. Encyclopedia of Optimization. (2nd) The
Netherlands: Springer, 2009. p. 753
–84.
[25] Liberatore MJ, Nydick RL. The analytic hierarchy process in medical and
health care decision making: a literature review. Eur J Oper Res
2008;189:194
–207.
[26] Ehrgott M, Güler Ç, Hamacher HW, et al. Mathematical optimization in
intensity modulated radiation therapy. 4OR 2008;6:199
–262.
[27] Lee H, Granata KP, Madigan ML. Effects of trunk exertion force and
direction on postural control of the trunk during unstable sitting. Clin
Biomech 2008;23:505
–9.
[28] Bertsimas D, Farias VF, Trichakis N. Fairness, ef
ficiency, and flexibility
in organ allocation for kidney transplantation. Oper Res 2013;61:73
–87.
[29] Thokala P, Dixon S, Jahn B. Resource modelling: the missing piece of
the HTA jigsaw? PharmacoEcon 2015;33:193
–203.
[30] Marshall DA, Burgos-Liz L, IJzerman MJ, et al. Selecting a dynamic
simulation modeling method for health care delivery research
—part 2:
report of the ISPOR Dynamic Simulation Modeling Emerging Good
Practices Task Force. Value Health 2015;18:147
–60.
[31] Marshall DA, Burgos-Liz L, IJzerman MJ, et al. Applying dynamic
simulation modeling methods in health care delivery research
—the
simulate checklist: report of the ISPOR Simulation Modeling Emerging
Good Practices Task Force. Value Health 2015;18:5
–16.
[32] Marshall DA, Burgos-Liz L, Pasupathy KS, et al. Transforming
healthcare delivery: integrating dynamic simulation modelling and big
data in health economics and outcomes research. PharmacoEconomics
2016;34:115
–26.
[33] Hillier FS, Lieberman GJ. Introduction to Operations Research. US:
McGraw-Hill Education, 2012.
[34] Minoux M. Mathematical Programming: Theory and Algorithms. Paris:
John Wiley & Sons, 1986.
[35] Puterman ML. Markov Decision Processes: Discrete Stochastic Dynamic
Programming. Hoboken, NJ: John Wiley & Sons, 2014.
[36] Winston WL, Goldberg JB. Operations Research: Applications and
Algorithms. Belmont, CA: Duxbury Press, 2004.
[37] Kelley CT. Iterative Methods for Optimization. Philadelphia, PA: Siam,
1999.
[38] Kuhn H, Tucker A. Proceedings of 2nd Berkeley Symposium. Berkeley:
University of California Press, 1951.
[39] Branke J, Deb K, Miettinen K, et al. Multiobjective Optimization:
Interactive and Evolutionary Approaches. Germany: Springer-VErlag
BErlin Heidelberg, 2008.
[40] Miettinen K. Nonlinear Multiobjective Optimization. New York, NY:
Springer Science & Business Media, 2012.
[41] Drummond MF, Sculpher MJ, Claxton K, et al. Methods for the
Economic Evaluation of Health Care Programmes. Oxford, UK: Oxford
University Press, 2015.
[42] Chalabi Z, Epstein D, McKenna C, et al. Uncertainty and value of
information when allocating resources within and between healthcare
programmes. Eur J Oper Res 2008;191:530
–9.
[43] Epstein DM, Chalabi Z, Claxton K, et al. Ef
ficiency, equity, and
budgetary policies informing decisions using mathematical
programming. Med Decis Making 2007;27:128
–37.
[44] McKenna C, Chalabi Z, Epstein D, et al. Budgetary policies and available
actions: a generalisation of decision rules for allocation and research
decisions. J Health Econ 2010;29:170
–81.
[45] Morton A. Aversion to health inequalities in healthcare prioritisation:
a multicriteria optimisation perspective. J Health Econ 2014;36:
164
–73.
[46] Morton A, Thomas R, Smith PC. Decision rules for allocation of
finances
to health systems strengthening. J Health Econ 2016;49:97
–108.
[47] Van de Wetering G, Woertman WH, Adang EM. A model to correct for
short
‐run inefficiencies in economic evaluations in healthcare. Health
Econ 2012;21:270
–81.
[48] Tappenden P, Chilcott J, Brennan A, et al. Whole disease modeling to
inform resource allocation decisions in cancer: a methodological
framework. Value Health 2012;15:1127
–36.
[49] Vanderby SA, Carter MW, Noseworthy T, et al. Modelling the complete
continuum of care using system dynamics: the case of osteoarthritis in
Alberta. J Simulation 2015;9:156
–69.
[50] Kim E. Sequential drug decision problems in long-term medical
conditions: a case study of primary hypertension. University of
Shef
field, 2015.
[51] Tosh J. Simulation optimisation to inform economic evaluations of
sequential therapies for chronic conditions: a case study in rheumatoid
arthritis. University of Shef
field, 2015.
[52] Banks J. Handbook of Simulation. Wiley Online Library, 1998.
[53] Harrison JR, Lin Z, Carroll GR, et al. Simulation modeling in
organizational and management research. Acad Manag Rev
2007;32:1229
–45.
[54] Sokolowski JA, Banks CM. Principles of Modeling and Simulation: A
Multidisciplinary Approach. Hoboken, NJ: John Wiley & Sons, 2011.
[55] Macal CM, North MJ, Collier N, et al. Modeling the transmission of
community-associated methicillin-resistant Staphylococcus aureus: a
dynamic agent-based simulation. J Translational Med 2014;12:1.
[56] Milstein B, Homer J, Briss P, et al. Why behavioral and environmental
interventions are needed to improve health at lower cost. Health Aff
2011;30:823
–32.
[57] Troy PM, Rosenberg L. Using simulation to determine the need for ICU
beds for surgery patients. Surgery 2009;146:608
–20.
[58] Sterman JD. Business dynamics: systems thinking and modeling for a
complex world. New York, NY: McGraw Hill, 2000.
[59] Crown WH. Potential application of machine learning in health
outcomes research and some statistical cautions. Value Health
2015;18:137
–40.
[60] Kubat M. An Introduction to Machine Learning. Switzerland: Springer,
2015.
[61] Witten, IH, Frank E, Hall MA, Pal CJ. Data MIning: Practical Machine
Learning Tools and Techniques. Cambridge, MA. 2017.
V A L U E I N H E A L T H 2 0 ( 2 0 1 7 ) 3 1 0 – 3 1 9
319
Dostları ilə paylaş: |