DISCRETE EVENT SIMULATION OF WHOLE CARE PATHWAYS TO ...
Transcript of DISCRETE EVENT SIMULATION OF WHOLE CARE PATHWAYS TO ...
Proceedings of the 2015 Winter Simulation Conference
L. Yilmaz, W. K. V. Chan, I. Moon, T. M. K. Roeder, C. Macal, and M. D. Rossetti, eds.
DISCRETE EVENT SIMULATION OF WHOLE CARE PATHWAYS TO ESTIMATE COST-
EFFECTIVENESS IN CLINICAL GUIDELINES
Julie Eatock Joanne Lord
Department of Computer Science Health Economics Research Group
Brunel University London
Kingston Lane
Brunel University London
Kingston Lane
Uxbridge, UB8 3PH, UK Uxbridge, UB8 3PH, UK
Marta Trapero-Bertran Anastasia Anagnostou
Economics and Business Department; CRES Department of Computer Science
(Health Economics Research Centre)
University Pompeu Fabra
Brunel University London
Kingston Lane
c/Ramon Trias Fargas 25-27 Uxbridge, UB8 3PH, UK
Barcelona, 08005, SPAIN
ABSTRACT
For pragmatic reasons, cost-effectiveness analyses performed for NICE Clinical Guidelines use a
piecemeal approach, evaluating only selected aspects of diagnosis, treatment or care. A Whole Pathway
approach, considering diagnosis-to-death, may provide more realistic estimates of costs and health
outcomes, taking account of the healthcare context and individual risk factors, history and choices for
patients with long-term conditions. A patient-level DES model using the characteristics of 12,766 real
patients was created to reflect the NICE guideline for Atrial Fibrillation. Of eight topics suggested for
inclusion in an update of the guideline, the model was capable of fully answering four topics, and
partially answering two topics. The remaining topics were beyond the scope of the model. The model was
used by NICE in their recent update of the Atrial Fibrillation Clinical Guidelines.
1 INTRODUCTION
In England, the National Institute for Health and Care Excellence (NICE) is responsible for assessing the
cost-effectiveness of drugs and other healthcare to be made available through the National Health Service
(NHS). NICE produce clinical guidelines recommending care pathways for certain medical conditions,
and defining criteria for access to services. This guidance is not binding and clinicians are expected to use
their judgment to provide the best care for their patients, but it does influence local decisions about how
much of which services to commission. The guidelines can be very large and complex, and need to be
updated at regular intervals to reflect changes in knowledge and technology.
Initial development and updates are led by a Guideline Development Group (GDG), which includes
clinicians and patient representatives, alongside a methodological team who are responsible for
identifying and analyzing evidence of effectiveness and cost-effectiveness. Due to time constraints, the
process starts with a scoping exercise, in which a limited number of review questions is selected. Cost-
effectiveness modeling is also restricted to selected aspects of the pathway, and therefore runs the risk of
neglecting systemic effects and interactions (Lord et al. 2013).
1447978-1-4673-9743-8/15/$31.00 ©2015 IEEE
Eatock, Lord, Trapero-Bertran, and Anagnostou
The MAPGuide (Modeling Algorithm Pathways in Guidelines) project was set up to test the
feasibility and usefulness of modeling the entire pathway, from diagnosis to death at an individual patient
level, within the context of clinical guidelines. Two exemplars, Prostate Cancer and Atrial Fibrillation,
were chosen to illustrate the approach. These topics were selected as they already had relatively well-
articulated care pathways defined in published NICE guidelines, and were due to be updated within the
timescales of the project. This provided an opportunity to develop models to reflect the existing
pathways, and then to test the models’ ability to address selected cost-effectiveness questions within the
context of guideline updates.
This paper describes the design of the MAPGuide Atrial Fibrillation model, based on the original
NICE guideline (National Collaborating Centre for Chronic Conditions 2006). The ability of this model to
assess the cost-effectiveness of potential changes to the 2006 pathway that were included in the scope for
2014 guideline update is assessed, and subsequent use of the model in the ‘live’ update process is
discussed. Finally some conclusions are drawn about the appropriateness of this type of modeling for
assessing cost-effectiveness in clinical guidelines, and areas for further research are identified.
2 ATRIAL FIBRILLATION
Atrial Fibrillation (AF) is a condition characterized by an irregular or rapid heartbeat (Camm et al. 2010).
Incidence of AF increases with age, and is correlated with comorbidities such as hypertension and
diabetes (National Collaborating Centre for Chronic Conditions 2006).
AF is often asymptomatic and intermittent, but progressive (Kirchhof et al. 2007). In paroxysmal AF
the fibrillation occurs for a short time, usually minutes or hours, but stops spontaneously within 7 days.
The frequency of episodes of paroxysmal AF can vary greatly, from several episodes per day, to months
in between (Shehadeh, Liebovitch and Wood 2002). However, the frequency generally increases as the
condition progresses, due to electrical remodeling in the heart tissue (Nattel 2002). Persistent AF is
defined by the need for medical intervention (pharmacological or electrical cardioversion) to terminate
the fibrillation within 7 days. AF is said to be permanent when it lasts for more than a year or when
cardioversion attempts have failed.
Due to the non-symptomatic nature of many AF cases, it is difficult to diagnose and to classify.
Typically, diagnosis is done through a 12 lead ECG, which is often performed for some other purpose and
for relatively short periods of time. Hence the AF may be missed if the heart is in normal sinus rhythm
during this time and, even once diagnosed, without longer monitoring it is difficult to determine the
duration of fibrillation and whether it will spontaneously terminate. Diagnosis is improved by using
Holter recorders for up to 7 days (Stahrenberg et al. 2010), but still many episodes may be missed.
There is evidence to suggest that the presence of AF increases the incidence of ischemic strokes five-
fold (Camm et al. 2010) and that AF may be the cause of up to 5% of all strokes of unknown origin
(Christensen et al. 2014). Ischemic strokes are caused when a thromboembolism (blood clot) moves and
blocks blood from reaching the brain cells. In order to reduce this risk, oral anti-coagulants are
administered. This in turn, however, increases the risk of a hemorrhagic stroke – where blood vessels in
the brain burst and damage cells. Strokes, of either type, can be catastrophic for the patient and very
expensive for the health and social care system. The risks of recurrence, mortality and disability after
stroke are also related to the type of AF (Ntaios et al. 2013). Algorithms to predict the occurrence of
thromboembolic events in AF patients have had limited success (see Lip et al. 2010 for a review).
Treatment for AF falls under the broad categories of rhythm control or rate control. Rhythm control
drugs aim to maintain sinus rhythm and therefore act to slow the progressive nature of the condition. In
the Clinical Guidelines issued in 2006 (CG36), these drugs were recommended for all patients with
paroxysmal AF, and for some with persistent AF (National Collaborating Centre for Chronic Conditions
2006). For persistent AF patients who met certain criteria and those with permanent AF, ventricular rate
control drugs were recommended. Rate control drugs do not attempt to put the heart in sinus rhythm, but
try to maintain a regular heartbeat of 60-80 beats per minute at rest. There are many classes of rhythm
1448
Eatock, Lord, Trapero-Bertran, and Anagnostou
control drugs available that work by blocking one or more different ion channels (Savelieva and Camm
2008). The cost-effectiveness of rhythm control drugs has been called into question (Camm, Savelieva
and Lip 2007).
3 INDIVIDUAL LEVEL VS COHORT LEVEL IN CONTEXT OF ECONOMIC MODELING
3.1 Economic Modeling
The remit set for NICE by the UK government states that cost-effectiveness as well as clinical
effectiveness should be considered when developing guidance. NICE recommends a standardized set of
methods (a ‘reference case’) to be used in cost-effectiveness analysis (NICE 2013). This includes the
core concept of the quality-adjusted-life-year (QALY) as the method to value health outcomes. QALYs
are defined as value-weighted time i.e. life years weighted by their quality (Weinstein, Torrance and
McGuire 2009). A QALY is the product of a patient’s current health state (health utility) and the time
spent in that state. Health utility is measure on a sliding scale, where 0 represents a state of health
equivalent to death and 1 represents perfect health. So, one year in perfect health yields one QALY, as
would 2 years with a health utility of 0.5. It is further assumed that QALYs are additive across people, so
6 months in perfect health for each of two people also yields one QALY. NICE recommends that QALYs,
like costs, should be discounted to adjust for societal preferences over timing; thus a particular health
state is ‘worth’ more in the near future than the distant future.
The health economists who conduct cost-effectiveness analyses for NICE use decision models to
estimate the long-term costs and health outcomes (QALYs) associated with alternative treatment options
for defined patient groups (Briggs, Sculpher and Claxton 2006). Most usually these take the form of
aggregate (cohort) models, such as decision trees or state-transition Markov models. In Markov models,
distinct health states have utilities and costs attached to them. Time advances in discrete time-steps, and at
each time-step patients move to the next appropriate state, based on transition probabilities. As all patients
move in unison, costs and QALYs can be easily calculated for each time step, and accumulated over the
time horizon. This type of model is generally far quicker to run than a discrete-event simulation (DES)
model, but it has limited ability to reflect complex patterns of disease progression related to individual
characteristics and history (Karnon 2003).
3.2 Advantages of Individual Level Modeling
Clinical guidelines are intended to assist clinicians in making appropriate and cost-effective treatment
decisions given the health status, history and other relevant characteristics of individual patients.
Modeling a whole treatment pathway where decisions can be based on individuals’ characteristics
provides greater flexibility to reflect reality than a cohort-based approach, where movement through the
pathway is based only on average probabilities across a group of patients. Cohort models can be
elaborated to tailor decisions for patient characteristics – for example, by increasing the number of
branches in a decision tree, or by increasing the number of health states or introducing tunnel states in a
Markov model. However, these approaches can quickly become unwieldy if there are many
characteristics to consider. In such cases, the ability to record and update an individual’s characteristics
and to make progress through the pathway conditional on those characteristics can give a more efficient
and compact model structure, and DES is an appropriate tool to achieve this.
The ability to remember patients’ characteristics also allows cost and QALY estimates to be
individualized, which enables assessment of whether particular sub-groups of patients would benefit from
different treatment strategies, while still providing population level results from a single batch of model
runs. The conventional approach in economic decision analysis is to run a series of separate subgroup
analyses, by adjusting the model to reflect different transition probabilities or outcomes for people with
different sets of characteristics (Briggs, Sculpher and Claxton 2006). But again this can become unwieldy,
particularly when there are multiple correlated characteristics that are related to event rates or outcomes,
1449
Eatock, Lord, Trapero-Bertran, and Anagnostou
for example key risk factors for stroke, such as high blood pressure, age and BMI (body mass index) are
highly correlated. Using a DES approach, the correlations between initial characteristics can be
automatically captured by using a real patient dataset.
3.3 Disadvantages of Individual Level Modeling
One of the most obvious disadvantages of modeling at an individual level is the data requirements. A key
input to economic modeling is clinical trial data, which is usually reported as aggregated results with
limited subgroup analysis. So for example, the effectiveness of a particular medication for patients with
paroxysmal AF may be reported, but with no indication of whether it is more or less effective for patients
with particular characteristics such as high blood pressure or BMI. It is also difficult to obtain more basic
data to characterize the incidence, progression and risks of AF over time. As outlined above, diagnosis
and monitoring of AF is problematic due to its intermittent and often asymptomatic nature. While
implanted pacemakers can record data on AF episodes over a long period of time, these are generally only
used for patients with recurrent and symptomatic disease, or who have other heart conditions. Follow up
data for patients diagnosed with AF have been used to estimate how the risk of strokes and other adverse
events are associated with patient characteristics in the short to medium term. There are also population
survey data providing a snapshot of patient characteristics at a point in time. However, to predict long
term outcomes we need to consider how these characteristics change over the patient’s lifetime. So for
example, we would need to model how blood pressure, BMI and smoking change with age, and in
response to lifestyle choices or other health issues.
Another obvious disadvantage of individual level modeling is the time it takes to develop and run the
model. While model run time is becoming less of an issue with the power of computing and the option of
using a distributed system, it takes longer to run than the decision tree and Markov models generally used
by health economists. This problem can be a significant barrier if probabilistic sensitivity analysis is to be
used to characterize uncertainty over population-level parameters, as this requires an outer loop of
iterations, multiplying the time needed to run a model by many times. Another important barrier may be
development time, as when a model becomes more complex the time to de-bug, and validate it increases
exponentially.
Patients may live with AF for many years, so the model needs to run over a long simulated time span
to capture the, possibly lifelong, costs and effects of changes to the pathway. This makes it difficult to
validate that the model is producing realistic results, as outcomes seen in reality in 20 years’ time will
differ from the results in the model in 20 simulated years, not due to model inaccuracies, but due to
technological advance and change in the population and healthcare environment.
4 DESIGN OF MODEL: THE CASE OF AF
The first stage of the project was to build a model based on the 2006 clinical guidelines for AF (CG36)
(National Collaborating Centre for Chronic Conditions 2006). The model development team started by
turning the recommendations in the guideline into a series of flowcharts, to represent the sequence of
treatment decisions and events. The modeling team consulted with clinical experts to resolve ambiguities
and uncertainties over the guideline recommendations. The flowcharts were coded for easy referral to the
specific item in guideline, and formed the basis of the discrete-event simulation model.
The model was designed for evaluating the cost-effectiveness of treatment options rather than as a
planning tool, so while costs were assigned to tests, visits and medications, resource constraints were not
included in the model. For similar reasons, the model was run as a cohort model, with all simulated
patients arriving together at the start of the run and the model terminating when the last patient died.
The sample of patients was drawn from the THIN (The Health Improvement Network) database, and
included 12,766 patients who had been diagnosed with AF. For each patient, 24 characteristics were
recorded that would be used within the model to evaluate risks, and decide treatment options. The
correlations between characteristics was therefore implicit in the data. Some characteristics, such as
1450
Eatock, Lord, Trapero-Bertran, and Anagnostou
hypertension and history of diabetes and stroke were allowed to increase as the patients aged in the
model. However, for simplicity other characteristics such as smoking and BMI remained static.
Figure 1. The current guidelines were converted to a series of flowcharts which combined to provide a
comprehensive view of the management of AF which formed the basis of the DES model.
Table 1. Main competing event types in the model.
Main event types Event sub-types
Thromboembolic (TE) ischemic stroke (IS)
transient ischemic attack (TIA)
other thromboembolic event (OTE)
Hemorrhagic (HE) hemorrhagic stroke (HS)
major bleed
AF progression undocumented events,
self-terminating events
non-terminating events requiring cardioversion
acute arrhythmic events requiring emergency cardioversion
coronary heart disease (CHD)
diabetes
hypertension
age threshold
death
Progression of AF is difficult to characterize, and to some extent the terms paroxysmal, persistent and
permanent are simply ‘labels’ that help determine the appropriate treatment option. A flowchart of
progression and control of AF, incorporating documented and undocumented instances of AF, and
cardioversion events was designed to provide a definitive framework to capture progression. A risk
model was also developed to quantify the risks of key outcomes (as shown in Table 1) conditional on
individual characteristics and treatment effects.
STROKE RISK /ANTICOAGULATION THERAPY (R11, R20,R52, R55, R56, R57, R58) (SR)
Guideline CHAPTER 11
Aspirin (longterm) (R11) SR4
Warfarin (longterm)
(R11, R20, R52)
SR6
Warfarin or aspirin? (R11)
SR5
Moderate riskAge≥65 with no high risk factors
Age <75 with hypertension,
diabetes or vascular disease
WarfarinAspirin
Assess stroke risk
(SR)
contraindicated to warfarin/dabigatran?
SR1yes
DabigatranSR7
Which drug?SR3
dabigatran
assess stroke risk (R55, R56, R57, R58)
SR2
High risk previous ischaemic stroke/
TIA or throboembolic event
age≥75 with hypertension,
diabetes or vascular disease
valve disease or heart failure,
or impaired lef t ventricular
function on echocardiography
Treatment
stra tegy?
SR8
Rhythm control (RY)
Rate control (RA)
AF for the rest of
their lives
no
warfarin
Low risk(age<65
with no moderate
or high risk factors)
TREATMENT OPTIONS FOR PAROXYSMAL (RYpx)(R27, R28, R29, R30, R31, R32, R66)
Guideline CHAPTER 6
Rhythm control for paroxysmal
AF (RYpx)
Is NDT or PiP appropriate? RYpx1
(R33)
No drug treatment (NDT) (R27)
RYpx1b
Pill-in-the-pocket (PiP) (R27, R33)
RYpx1c
Standard beta-blockers (R28)
RYpx2
no
Treatment failure? RYpx3
no
LVD (or CHF)? (R31) RYpx4
yes
CAD? (R29) RYpx4a
1c agent (R29) RYpx5
sotalol (R29, R30)
RYpx6
amiodarone (R29, R30, R31)
RYpx9
Tertiary specialist
(R66) RYpx11
Ongoing management
(R32) (OM)
Patient choice? RYpx1a
yes
Treatment failure? RYpx7
Treatment failure? RYpx8
Treatment failure? RYpx10
no
yesno
nono
yes
yes
no
1st line treatment
2nd line treatment
3rd line treatment
4th line treatment
yes
Pre-treatment
Treatment failure? RYpx1d
yes
no
ACUTE ONSET TREATMENT
(R35, R36, R37, R38, R39, R40. R42)
Guideline CHAPTER 9
ACUTE onset AF
(haemodynamic
instability (HI)
(A)
To check electrolytes and
review CXR. Attempt to
establish aetiology of
acute haemodynamic
instability
no yes
yes
ECG and chest
x-ray A1
Urgent
rate contro l
required?
(R37) A10
!"#$
%&'(#)'**$+,
A21
Intravenous
amiodarone
(R38) A24
Intravenous
heparin
(R39) A7
Intravenous BB
(R38) A17
Is the AF known to
be permanent? A8not known
permanent
life threatening HI?
(R42)? A2
Emergency ECV
(R35) A13
delay in organising
electrical card io-
version? A9
ECV
(R36) A14
Known Wolff-
Parkinson-White
syndrome?
A11
intravenous
amiodarone
(R36) A15
intravenous
flecainide
(R36) A16
Taking
anticoagulation?
A5
known
permanent
Contraindicated to
BB or RLCA?
(R38) A12
Intravenous RLCA
(R38) A18
!"#$%
&'(#)'**$+,
A20
yes
yes
no
heparin
contraindicated?
A6
Onset <48h?
A25
4 weeks of
warfarin
(R40) A27
Factors indicative of a
high risk recurrence?
(R40) A26
Fatal?A3
Death A4
yes
no
no
no
Cardioversion success? A19
in sinus rhythm A23
Not in sinus rhythm A22
no
yes
no
yes
yes
no
no yes
yes
yes
Classification (CL)
nono/unknown
label as persistentnot in sinus rhythm
label as persistent in sinus rhythm
yes yes
!"#$%
&'(#)'**$+,
A28
no
yes / no
no
label as permanent (onset >48h)
!"#$%&'(#)%&*+,-.+/0.1213043+"5+6!7
6#89+#:9+#;<9+#;89+#;=9+#;:9+#>;7
?@2A0B240+!C"DE(#+F
!"#$%&
'&()*&+,-.&
(R7) C1
heparin (R19) C2
warfarin (3 wks) and/or TOE
(R10, R17, R19) C3
Electrical or pharmacological
cardio? (R7) C4
Electrical cardioversion
(ECV) (R7) C7
High risk of failure C5
sotalol or amiodarone (4 wks) (R9)
C8
/$#
Onset confirmed < 48h ago? (R21) C16
warfarin (4 wks) (R18, R19) C15
Cardioversion
inmediate success
C13
Cardioversion
(C)
noyes
+ maintenance of sinus
rythm at 1 and 6 months
(R60) At the 1 month
follow-up the frequency
of subsequent reviews
should be tailored to the
individual pat ient taking
into account
comorbidities and
concomitant drug
therapies (R61) Clinician
to re-assess the need for
continued
anticoagulation (R62) ***
Cardioversion
inmediate success?
C11
assess stroke risk (SR)
CAD or LVD? (R8) C6
Intravenous class 1c
C9
Intravenous amiodarone
C10
Cardioversion
inmediate success? C12
no yes
yes yes
nono
yes
yes
no
> 2 attempts at ECV? C14
no
no
yes (not in sinus rhythm)
permanent
no
no
in sinus rhythm(persistent)
label as persistent
!"#$!%#&!'()"'"$!#'*)&!")+',"$-
,"./0'".10'"22-
345678597'*:$;!#"'<
Rate control (persistent or permanent)
(RA)
The specialist referral
should be considered
especially in those with
lone AF or ECG evidence
of an underlying
electrophysiological
disorder (eg WPW) or
where pharmacological
therapy has failed
beta-blocker (BB) (R23) RA2
Tertiary specialist
(R66) RA12
amiodarone RA10
Ongoing
management
(R32)
(OM)
rate limiting calcium antagonist (RLCA) (R23) RA3
Treatment failure?
RA4
Treatment failure?
RA5
BB+ D (R24) RA6
RLCA + D (R24) RA7
Treatment failure?
RA8
Treatment failure?
RA9
Treatment failure?
RA11
yes yes
yes yes
yes
1st line treatment
2nd line treatment
3rd line treatment
4th line treatment
no
no
no
no
Control during exercise
required? RA1
no yes
no
ONGOING MANAGEMENTGuideline CHAPTER 12
Ongoing MAnagement
(OM)
Event? OM2
Major ‘event’ TE (IS/TIA/other)
Bleed (HS/major)
Other (CHD (MI/UA)/CHF)
OM4
routine follow up
OM7
Classification of AF (CL)
Drug intolerance
OM6
Acute onset AF (A)
Wait for next event
OM1
Calculation of time
to next event
AF control
OM3
Death
New risk factorsMinor events: HT/T2D/
age
OM5
Acute onset?OM12
noGP/OP
silent(undocumented
recurrence)?OM8
no
yes
yes(emergency)
Fatal?OM9
yes
no
Drop in treatment?
OM10
yes
no
undocumentedrecurrence?
OM11
yes
no
CLASSIFICATION OF AF (CL)(R11, R12, R13)
DIAGNOSIS (D)(R2, R3)
Guideline CHAPTER 4
NON ACUTE AF presentation
Asymptomatic and symptomatic patients (NA)
Perform ECG (R2) D1
Discharge D13
Symptomatic
patients
(breathlessness/
dyspnoea;
palp ita tions;
syncope/
dizziness; chest
discomfort;
stroke/TIA); patients
with associated
medical problems
(Heart failure; stroke
or thromboembol ism)
no
Rhythm or rate? (R11)
CL1
assess stroke rIsk
(SR)
cardioversion and assess stroke risk (R13)* (C)
assess stroke risk
(SR)
Rhythm (R13)
Rate: (a) over 65; or
(b) with coronary heart disease (MI or UA); or
(c) with contraindication to ant iarrhythmic drugs; or
(d) with contraindications to ant icoagulants
(bleed); or
(e) with SHD (CAD or LVD); or
(f) history of failed attempts of cardioversion; or
(g) without CHF (R12 and R13)
paroxysmal
Symptomatic
episodes more
than 24h apart?
D7
Event recorder ECG (R3) D8
24h ambulatory ECG monitor
(R3) D9
* patients unsuitable for cardioversion
include those with:
• Contraindications to anticoagulation
• Structural heart disease that
precludes long-term maintenance of
sinus rhythm
• A long duration of AF
• A history of multiple failed attempts
at cardioversion and7or relapses,
even with concomitant use of
antiarrhytmic drugs or non-
pharmacological approaches
• An ongoing but reversible cause of
atrial fibrillat ion
permanent
yes
Symptoms
such as
having
arrhythmia
last night, or
fast heart
beat, etc.
AF for the rest of
their lives
Acute onset AF
(A)
ACUTE ONSET AF presentation (A)
yes no
TTED4
Further assessment needed? D5
TOE D6
yes
no
positive
More investigations?
D12
yes
no
Further
investigation? D3
Suspected paroxysmal?
D2negative
yes
no
* Before assess
stroke risk
Wait for next event
Confirmed paroxysmal
AF? D10
Confirmed paroxysmal
AF? D11
no no
yes
Classification of AF (CL)
persistent
Classification of AF (CL)
CLASSIFICATION OF AF
RHYTHM CONTROL
STRATEGY Parox & Persis AF (RY)
Guid. CHAPTERS 6 & 8
ONGOING MANAGEMENT (OM)
Guid. CHAPTER 12
RATE
CONTROL
STRATEGY (RA)
Guid. CHAPTER 7
CARDIOVERSION (C)
Guid. CHAPTER 5
STROKE RISK STRATIFICATION (SR)
Guid. CHAPTER 11
DIAGNOSIS (D)
Guid. CHAPTER 4
ECG D1
TTE D4
24hECG D9ERECG D8
Assess risk
factor? SR2
Aspirin SR4Dabigatran
SR7
Onset <48hrs
ago? C1
Heparin C23 wks
warfarin C3
4 wks sotalo l or
amiodarone C8
Electrical
cardioversion C7
Pharmacological
cardioversion
C9/C10
4 wks
warfarin C15
Schedule
next
appointment
M1
Is PiP or
antiarrhythmic
appropriate? RY1
y
No drug
treatment
RY1b
low/med med/high
n
ACUTE onset AF (A)
Guid. CHAPTER 9
ECG and
chest x-ray A1
AF control OM3
Drug introlerance
OM6
Routine
follow-up OM7
Referral from
GP (NA)
Discharge
D13
TOE D6
ECV A14
Intra flecainide
A16
Intra RLCA A18
Emergency ECV
A13
Intra amiodarone
A24
Warfarin SR6
med/high
PiP
RY1c
Amiodarone
RA10
Death M5
Wait
for
next
event
M2
Classification?
paroxysmal permanentpersistent
Contraindicated
warf/dabig? SR1
ny
Beta-blocker
(BB) RA2
Rate limiting
calcium antagonist
(RLCA) RA3
BB+D
RA6
RLCA+D
RA7
Tertiary
specialist
RA12
Standard
BB RY2
1c agent
RY5
Sotalol
RY6
Amiodarone
RY9
Tertiary
specialist
RY11
4w of warfarin
Intra amiodarone
A15
Intra BB A17
Major event
OM4
New risk factors
OM5
!"#$%&'($)#$&&%"$*+,'-./'0%+%1/0/+"'*('%"#$%&'
($)#$&&%"$*+'!"#$%&'()*+(,*-%.#/012'23+/'4556
1451
Eatock, Lord, Trapero-Bertran, and Anagnostou
The model used a time to next event design, as illustrated in Figure 2. After initial diagnosis, the
model calculates the time to all the competing events (described in more detail in section 4.1), and
identifies the event that is due to occur first. The patient then waits, accumulating costs and QALYs
associated with their current health condition and treatment, until the event occurs. If this is a non-AF
related event, such as reaching an age threshold or developing diabetes, the patient’s characteristics are
updated and the risks and time to competing events re-calculated. However following an AF-related
event, such as disease progression or stroke, the patient receives short-term treatment if appropriate, their
costs and QALYs are updated, and their subsequent AF treatment is reviewed and possibly changed. This
cycle is repeated until the next event scheduled is death, when the patient exits the model.
4.1 Calculating Time to Event
Published epidemiological sources were used to estimate individuals’ risks of the included outcomes in
the absence of treatment. Rates of AF progression were based on longitudinal studies of patients with AF
in Canada and Europe (Kerr et al. 2005; Nieuwlaat et al. 2005; Nieuwlaat et al. 2008; Camm et al. 2011).
Rates of hemorrhagic and thromboembolic events were estimated from a large AF cohort in Sweden
(Friberg, Rosenqvist and Lip 2012), with rates varying according to individual risk factors as quantified
by the HAS-BLED and CHA2DS2-VASc algorithms (Lip et al. 2010; Pisters et al. 2010). Rates of
coronary heart disease, diabetes and hypertension were all based on equations from the Framingham
Heart Study (Anderson et al. 1991; Wilson et al. 2007; Parikh et al. 2008). A proportion of patients who
experienced AF-related hemorrhagic and thromboembolic events died within 30 days. Non-AF related
mortality rates came from national life table data (www.ons.gov.uk). These background risks were
modified to account for the effectiveness of the treatments where appropriate. So, for instance, an
anticoagulant decreased the risk of thromboembolic events but increased the risk of hemorrhagic events.
The relative effects of treatments were estimated from the clinical trial literature.
The chance of not having a particular event until a time t, given an adjusted hazard rate of λ is
modelled by an exponential survival function (i.e. 1-e-λt
). This assumes that the hazard remains constant
Figure 2. Overview of model structure (left) and final DES model built using Simul8 (right).
Rate Control
Diagnosis
Classification
Emergencies
Cardioversion
Stroke Risk
Rhythm Control
On-going
Management
1452
Eatock, Lord, Trapero-Bertran, and Anagnostou
over the period of time modelled. However, as patients move through the model, their risk of a particular
type of event may change in response to other associated risks, or as they age. We modelled this using a
piecemeal exponential distribution, in which the hazard is constant except for changes at defined points in
time (when an event has occurred or an age-threshold has been crossed).
At model entry, each patient is assigned a random number for each of the main types of event (see
Table 1). This could be considered as a proxy for unknown factors that influence a patient’s propensity
for that particular type of event. For each of the three composite events in Table 1 (i.e. thromboembolic,
hemorrhagic, or progression), patients are assigned second random number, which determines which sub-
type of event will occur next (i.e. IS, TIA or OTE in the case of a thromboembolic event). This approach
ensures that related groups of events are not treated as independent.
Once an event has occurred, care is needed in adjusting the times for competing events. For example,
it would be counterintuitive if the predicted time to a TE event were to increase following a bleed. To
avoid this, the random numbers used to determine the time to the next events are only resampled when
that specific type of event has occurred. This approach produced a piecemeal increasing function for each
of the competing risks (as illustrated in Figure 3).
5 MODEL OUTPUTS
The model was designed to produce results at two levels, individual diaries of patients’ experiences and
interactions with the healthcare system, and aggregated cohort-level outputs for use in cost-effectiveness
analysis. The individual level diaries (see Figure 4) provided a simple but effective way for clinicians to
verify and assess the model. The diary recorded the individual’s medical history and their associated risks
when they entered the process, and their subsequent treatment and events as the patient progressed
through the model. Costs and QALYs were also recorded at each stage of the process (see Figure 4) to
enable reviewing of calculations.
The cohort-level outputs were designed for use in cost-effectiveness analyses. These recorded the
numbers of events, costs and QALYs across the patient sample. For each version of the pathway model,
results were summarized using a net benefit statistic. This is calculated by multiplying the total QALYs
by a monetary value – the cost-effectiveness threshold λ – and then subtracting total costs. We used a
threshold of £20,000 per QALY in our analyses, the lower limit of the range recommended by NICE.
Probabilistic Sensitivity Analysis (PSA) was used to account for uncertainty over the input parameters. In
PSA, the model is repeatedly run, each time drawing a set of input parameters from defined probability
distributions (using random, Monte Carlo simulation). There were therefore two nested sets of iterations:
an outer layer of PSA sampling and an inner layer of individual simulation. The output from the model
was presented as the mean and variance of the net benefits across the PSA iterations. Different versions of
the care pathway can be compared in terms of net benefit: the most cost-effective pathway (at a defined
level of λ) being that which produces the maximum net benefit.
Figure 3. Cumulative probability of three competing events over 30 years for an individual patient.
1453
Eatock, Lord, Trapero-Bertran, and Anagnostou
Figure 4. Model Outputs: Patient Diary (top) and Cost and QALY reports (bottom).
1454
Eatock, Lord, Trapero-Bertran, and Anagnostou
6 USE OF MODEL DURING GUIDELINE UPDATE PROCESS
The model was designed to reflect the current guideline and answer questions on cost effectiveness on
potential guideline updates. During the building of the model, the model developers were unaware of the
specific questions that would be asked as part of the update process. This ensured that the model was not
designed specifically to answer the update questions, but to determine whether a whole pathway DES
model could be utilized and adapted as part of the update process. There were 8 topics identified by the
GDG that would be considered in the update process (see Table 2).
Table 2. Update topics.
Topic Description
A Prophylaxis for the prevention of post-operative AF
B Anti-arrhythmic drugs as pharmacological cardioversion (PCV) for people with atrial fibrillation
C Rhythm versus rate control for persistent AF; subgroups including those with hypertension,
previous MI and congestive heart failure
D Treatment for maintaining Sinus Rhythm in people with AF after cardioversion.
E Alternative risk factor based scoring systems to estimate stroke and embolism risk
F Stratification tools to assess bleeding risk before prescription of antithrombotic medication
G Apixaban, rivaroxaban or Dabigatran etexilate versus warfarin as for patients at moderate or high
risk of stroke or systemic embolism
H Cather ablation for paroxysmal and persistent AF patient
The model was therefore able to answer some of the update questions but as the nature of the
questions were unknown to the modeling team during development, other update questions were
unanswerable in the time given. The GDG were unfamiliar with the simulation software, so Topics C, E
and F were included by the model development team as menu driven options from the start screen of the
model, while Topics B and G were addressed by making changes in the spreadsheet containing input data
to the model, which the GDG were able to do themselves.
An additional benefit of the process of building the model was that during development we were able
to highlight to the GDG any existing inconsistencies and ambiguities within the current guideline which
could be corrected in the update.
7 DISCUSSION
The model accurately captured the AF care pathway as described by the clinical guideline, and was able
to produce estimates of costs and health outcomes from diagnosis-to-death for the simulated patient
cohort. Using a patient level DES model allowed realistic modeling of the treatment pathway, with
decisions based on individual characteristics and history. The model used the characteristics of real
patients, which provided implicit correlations between the characteristics that would be difficult to
replicate with distribution-generated characteristics. Arguably, this individual-level simulation approach
resulted in a much more complex model with far greater data requirements than a more conventional cost-
effectiveness model.
Once the model was complete it was provided to the GDG for their clinical update of the AF
guideline. Despite the modeling team being unaware of the questions that would be raised during the
update process during the development of the model, it allowed some of the update questions to be
answered, but other questions would have required re-modeling of certain aspects of the model. As the
GDG were unfamiliar with the software, the model development team designed a user interface to enable
testing of some of the topics, while others could be tested by altering the data within the input parameter
spreadsheet. This highlighted the need to be familiar with both the software and the model itself, in order
1455
Eatock, Lord, Trapero-Bertran, and Anagnostou
to use the model to address specific cost-effectiveness questions. This suggests that access to specialist
DES expertise or training for economic modelers would be necessary to implement this approach in
routine guideline development. The time taken to build the model was greater than that usually available
to the GDG, and therefore would be prohibitive. Now built, however, the model has the potential to be re-
used, with little additional development for subsequent guidelines. The model could also be extended to
include prevalent (existing) cases as well as incident (new) cases, which would help assess the budget
impact of the changes across a population.
Some members of the GDG were unfamiliar with, or did not have access to, the software and
therefore couldn’t fully review the model. The flowcharts that were created as part of the model
development process were highly accessible and understandable and linked to the existing guidelines.
These together with the individual patient diaries were used by the members to review the model design
and output, providing the development group with confidence in the model’s results. An additional
benefit of creating the flowcharts was that inconsistencies and ambiguities within the original guideline
could be addressed in the update.
8 CONCLUSION
In summary, the individual level modeling adds complexity, but allows a much more realistic
representation of the clinical pathway based on a patient’s experiences. It must be mentioned that there is
more to guideline development than cost-effectiveness analysis. The GDG are responsible for identifying
areas of clinical interest, and reviewing the available evidence for relevance and quality, and consulting
with stakeholders. Therefore the time available for cost-effectiveness analysis is extremely limited, and
this needs to be taken into account when assessing the usefulness of different modeling techniques.
ACKNOWLEDGMENTS
The MAPGuide project was funded by the Medical Research Council and the National Institute for
Health Research through the Methodology Research Programme (G0901504). Julie Eatock is currently
funded by an EPSRC grant (EP/K037145/1) Forecasting personal health in an uncertain environment.
REFERENCES
Anderson, K. M., P. M. Odell, P. W. F. Wilson and W. B. Kannel. 1991. "Cardiovascular Disease Risk
Profiles." American Heart Journal 121:293-298.
Briggs, A., M. Sculpher and K. Claxton. 2006. Decision Modelling for Health Economic Evaluation
Oxford: Oxford University Press.
Camm, A. J., G. Breithardt, H. Crijns, P. Dorian, P. Kowey, J.-Y. Le Heuzey, I. Merioua, L. Pedrazzini,
E. N. Prystowsky, P. J. Schwartz, C. Torp-Pedersen and W. Weintraub. 2011. "Real-Life
Observations of Clinical Outcomes with Rhythm- and Rate-Control Therapies for Atrial Fibrillation:
Recordaf (Registry on Cardiac Rhythm Disorders Assessing the Control of Atrial Fibrillation)."
Journal of the American College of Cardiology 58:493-501.
Camm, A. J., P. Kirchhof, G. Y. H. Lip, U. Schotten, I. Savelieva, S. Ernst, I. C. Van Gelder, N. Al-Attar,
G. Hindricks, B. Prendergast, H. Heidbuchel, O. Alfieri, A. Angelini, D. Atar, P. Colonna, R. De
Caterina, J. De Sutter, A. Goette, B. Gorenek, M. Heldal, S. H. Hohloser, P. Kolh, J.-Y. Le Heuzey,
P. Ponikowski, F. H. Rutten, E. C. f. P. Guidelines, A. Vahanian, A. Auricchio, J. Bax, C. Ceconi, V.
Dean, G. Filippatos, C. Funck-Brentano, R. Hobbs, P. Kearney, T. McDonagh, B. A. Popescu, Z.
Reiner, U. Sechtem, P. A. Sirnes, M. Tendera, P. E. Vardas, P. Widimsky, D. Reviewers, V. Agladze,
E. Aliot, T. Balabanski, C. Blomstrom-Lundqvist, A. Capucci, H. Crijns, B. Dahlöf, T. Folliguet, M.
Glikson, M. Goethals, D. C. Gulba, S. Y. Ho, R. J. M. Klautz, S. Kose, J. McMurray, P. Perrone
1456
Eatock, Lord, Trapero-Bertran, and Anagnostou
Filardi, P. Raatikainen, M. J. Salvador, M. J. Schalij, A. Shpektor, J. Sousa, J. Stepinska, H. Uuetoa,
J. L. Zamorano and I. Zupan. 2010. "Guidelines for the Management of Atrial Fibrillation." European
Heart Journal 31:2369-2429.
Camm, A. J., I. Savelieva and G. Y. H. Lip. 2007. "Rate Control in the Medical Management of Atrial
Fibrillation." Heart 93:35-38.
Christensen, L. M., D. W. Krieger, S. Højberg, O. D. Pedersen, F. M. Karlsen, M. D. Jacobsen, R. Worck,
H. Nielsen, Aelig, K. gidius, L. L. Jeppesen, S. Rosenbaum, J. Marstrand and H. Christensen. 2014.
"Paroxysmal Atrial Fibrillation Occurs Often in Cryptogenic Ischaemic Stroke. Final Results from the
Surprise Study." European Journal of Neurology 21:884-889.
Friberg, L., M. Rosenqvist and G. Y. H. Lip. 2012. "Evaluation of Risk Stratification Schemes for
Ischaemic Stroke and Bleeding in 182 678 Patients with Atrial Fibrillation: The Swedish Atrial
Fibrillation Cohort Study." 33:1500-1510.
Karnon, J. 2003. "Alternative Decision Modelling Techniques for the Evaluation of Health Care
Technologies: Markov Processes Versus Discrete Event Simulation." Health Economics 12:837-848.
Kerr, C. R., K. H. Humphries, M. Talajic, G. J. Klein, S. J. Connolly, M. Green, J. Boone, R. Sheldon, P.
Dorian and D. Newman. 2005. "Progression to Chronic Atrial Fibrillation after the Initial Diagnosis
of Paroxysmal Atrial Fibrillation: Results from the Canadian Registry of Atrial Fibrillation."
American Heart Journal 149:489-496.
Kirchhof, P., A. Auricchio, J. Bax, H. Crijns, J. Camm, H.-C. Diener, A. Goette, G. Hindricks, S.
Hohnloser, L. Kappenberger, K.-H. Kuck, G. Y. H. Lip, B. Olsson, T. Meinertz, S. Priori, U. Ravens,
G. Steinbeck, E. Svernhage, J. Tijssen, A. Vincent and G. Breithardt. 2007. "Outcome Parameters for
Trials in Atrial Fibrillation: Executive Summary: Recommendations from a Consensus Conference
Organized by the German Atrial Fibrillation Competence Network (Afnet) and the European Heart
Rhythm Association (Ehra)." European Heart Journal 28:2803-2817.
Lip, G. Y. H., R. Nieuwlaat, R. Pisters, D. A. Lane and H. J. G. M. Crijns. 2010. "Refining Clinical Risk
Stratification for Predicting Stroke and Thromboembolism in Atrial Fibrillation Using a Novel Risk
Factor-Based Approach: The Euro Heart Survey on Atrial Fibrillation." Chest 137:263-272.
Lord, J., S. Willis, J. Eatock, P. Tappenden, M. Trapero-Bertran, A. Miners, C. Crossan, M. Westby, A.
Anagnostou, S. J. E. Taylor, I. Mavranezouli, D. Wonderling, P. Alderson and F. Ruiz. 2013.
"Economic Modelling of Diagnostic and Treatment Pathways in National Institute for Health and
Care Excellence Clinical Guidelines: The Modelling Algorithm Pathways in Guidelines (MAPGuide)
Project." Health Technology Assessment 17.
National Collaborating Centre for Chronic Conditions. 2006. "Atrial Fibrillation: National Clinical
Guideline for Management in Primary and Secondary Care." CG36, Royal College of Physicians,
London, UK.
National Institute for Health and Care Excellence. 2013. "Guide to the Methods of Technology Appraisal
2013."
Nattel, S. 2002. "New Ideas About Atrial Fibrillation 50 Years On." Nature 415:219-226.
Nieuwlaat, R., A. Capucci, A. J. Camm, S. B. Olsson, D. Andresen, D. W. Davies, S. Cobbe, G.
Breithardt, J.-Y. Le Heuzey, M. H. Prins, S. Lévy and H. J. G. M. Crijns. 2005. "Atrial Fibrillation
Management: A Prospective Survey in Esc Member Countries - the Euro Heart Survey on Atrial
Fibrillation." European Heart Journal 26:2422-2434.
Nieuwlaat, R., M. H. Prins, J.-Y. Le Heuzey, P. E. Vardas, E. Aliot, M. Santini, S. M. Cobbe, J. W. M. G.
Widdershoven, L. H. Baur, S. Lévy and H. J. G. M. Crijns. 2008. "Prognosis, Disease Progression,
and Treatment of Atrial Fibrillation Patients During 1 Year: Follow-up of the Euro Heart Survey on
Atrial Fibrillation." European Heart Journal 29:1181-1189.
Ntaios, G., A. Vemmou, E. Koroboki, P. Savvari, K. Makaritsis, M. Saliaris, G. Andrikopoulos and K.
Vemmos. 2013. "The Type of Atrial Fibrillation Is Associated with Long-Term Outcome in Patients
with Acute Ischemic Stroke." International Journal of Cardiology 167:1519-1523.
1457
Eatock, Lord, Trapero-Bertran, and Anagnostou
Parikh, N. I., M. J. Pencina, T. J. Wang, E. J. Benjamin, K. J. Lanier, D. Levy, S. R. B. D'Agostino, W. B.
Kannel and R. S. Vasan. 2008. "A Risk Score for Predicting near-Term Incidence of Hypertension:
The Framingham Heart Study." Annals of Internal Medicine 148:102-110.
Pisters, R., D. A. Lane, R. Nieuwlaat, C. B. de Vos, H. J. G. M. Crijns and G. Y. H. Lip. 2010. "A Novel
User-Friendly Score (Has-Bled) to Assess One-Year Risk of Major Bleeding in Atrial Fibrillation
Patients: The Euro Heart Survey." Chest 138:1093-1100.
Savelieva, I. and J. Camm. 2008. "Anti-Arrhythmic Drug Therapy for Atrial Fibrillation: Current Anti-
Arrhythmic Drugs, Investigational Agents, and Innovative Approaches." Europace 10:647-665.
Shehadeh, L. A., L. S. Liebovitch and M. A. Wood. 2002. "Temporal Patterns of Atrial Arrhythmia
Recurrences in Patients with Implantable Defibrillators: Implications for Assessing Antiarrhythmic
Therapies." Journal of Cardiovascular Electrophysiology 13:303-309.
Stahrenberg, R., M. Weber-Krüger, J. Seegers, F. Edelmann, R. Lahno, B. Haase, M. Mende, J.
Wohlfahrt, P. Kermer, D. Vollmann, G. Hasenfuß, K. Gröschel and R. Wachter. 2010. "Enhanced
Detection of Paroxysmal Atrial Fibrillation by Early and Prolonged Continuous Holter Monitoring in
Patients with Cerebral Ischemia Presenting in Sinus Rhythm." Stroke 41:2884-2888.
Weinstein, M. C., G. Torrance and A. McGuire. 2009. "Qalys: The Basics." Value in Health 12:S5-S9.
Wilson, P. F., J. B. Meigs, L. Sullivan, C. S. Fox, D. M. Nathan, R. B. D’Agostino and Sr. 2007.
"Prediction of Incident Diabetes Mellitus in Middle-Aged Adults: The Framingham Offspring Study."
Archives of Internal Medicine 167:1068-1074.
AUTHOR BIOGRAPHIES
JULIE EATOCK is a Research Fellow at Brunel University London. Her main area of research is
discrete event simulation modeling, and she has applied this in many different areas including information
systems and policing. In more recent times her interests have tended towards healthcare and she has
applied her skills in areas such as healthcare service provision, the impact of disruptive technology, and
cost-effectiveness. Her email address is [email protected].
JOANNE LORD is a Reader in Health Economics at Brunel University London conducting applied and
methodological research into the cost-effectiveness of health care technologies. She previously worked at
NICE, Imperial College Management School and St George’s Hospital Medical School. Her first job was
in the Operational Research Service at the Department of Health. Joanne has a BSc in Economics and
Mathematics, MSc in Operational Research and PhD in health economics. Her email address is
MARTA TRAPERO-BERTRAN is a Research Fellow at Universitat Pompeu Fabra (UPF) and
Universidad Castilla-La Mancha (UCLM). Her main research interests include modeling techniques for
economic evaluation, transmission of lifestyle behavior, and incorporation of external effects into
economic evaluation. She is the former President of the Spanish Association of Health Economics (AES).
She defended her PhD in Health Economics at HERG-Brunel University. Her email address is
ANASTASIA ANAGNOSTOU is a Research Fellow at the Department of Computer Science, Brunel
University London. She holds a PhD in Hybrid Distributed Simulation. Her research interests are related
to the application of modeling and simulation techniques in the Healthcare sector and Industry. Her email
address is [email protected].
1458