The 34th International Conference of the System Dynamics Society, July 2016
The Economic Burden of Hip Fractures among Elderly Patients
in Ireland: A Combined Perspective of System Dynamics and
Machine Learning
Mahmoud E lbattah Owen Molloy
National University of Ireland Galway National University of Ireland Galway
m.elbattah1@ nuigalway.ie owen.molloy@ nuigalway.ie
Abstract
Population ageing is increasing in a rapid pace worldwide, and especially within developed
countries. Extraordinary economic challenges are therefore in prospect with regard to
healthcare delivery. In this respect, healthcare executives increasingly need tools that can
accurately assess the impacts of the foreseen demographic transition. The paper investigates
the economic implications in relation to the incidence of hip fractures among elderly patients
in Ireland. A combined approach is adopted that utilises System Dynamics (SD) and machine
learning. At the macro-scale level, an SD model is used to produce projections of elderly
populations who are susceptible to sustain hip fractures. In addition, the SD model is
disaggregated to properly depict the demographic structure of the healthcare system in
Ireland. At the micro-scale level, machine learning models are used to make careful
predictions on the inpatient length of stay and discharge destinations for simulation-
generated patients. The study is claimed to deliver useful insights regarding the potential
economic burden on the Irish healthcare system implied by elderly hip-fracture patients.
More broadly, we attempt to provide a multi-methodology perspective that combines
simulation modeling and machine learning towards increasing the confidence and credibility
of the simulation model predictions for decision making purposes.
Keywords:
System Dynamics; Machine Learning; Elderly Healthcare; Hip Fracture Care.
1. Introduction
In tandem with climate change and global terrorism, the UN identified population ageing as
one of the three main global challenges [1]. In Europe, the proportion of people aged 65 years
or over has already exceeded that younger than 15 years in 2008, and that proportion is
expected to double by 2060 [2]. More importantly, the proportion of very old people aged 80
years or over is expected to triple between 2008 and 2060 [3]. Likewise in Ireland, the
population has been experiencing a pronounced transition of ageing. The Health Service
Executive (HSE) of Ireland reported in 2014 that the increase in the number of people over
65 is approaching 20,000 per year [4]. Population ageing is therefore expected to have
profound impacts on a broad range of economic and social areas. Figure 1 plots the trend of
ageing worldwide as reported by the UN [5].
In the context of elderly-related care, the study focused its attention on the care scheme of hip
fracture in Ireland. Hip fractures are a major cause of injuries and morbidity among elderly
patients. As acknowledged by numerous studies [6-8], hip fractures were observed to be
exponentially increasing with age, despite the existence of rate variability from country to
another. Furthermore, the burden of hip fractures on the healthcare system may unavoidably
increase owing to the continuous improvement of life expectancy of the population [9-10].
The 34th International Conference of the System Dynamics Society, July 2016
21 billion
in 2050
Figure 1. Global projections of the elderly aged 60 and over compared to children under 15 (1950- 2090) [5].
In light of that, the paper mainly aimed at estimating the potential economic cost in relation
to hip fracture treatment for elderly patients in Ireland. The economic burden is two-fold: i)
Direct costs, and ii) Indirect costs. The direct costs typically include the amount of
expenditure spent on ED (Emergency Department) admission, and inpatient/outpatient care.
While indirect costs can involve various adverse effects on the quality of life. However, the
study endorsed only the direct costs, which can be mostly tangible to quantify and assess. The
direct costs of treatment are highly contingent on the inpatient length of stay and the
discharge destination whether home, or a long-stay care such as nursing home for example.
As a result, it was imperative to endorse other auxiliary questions related to the inpatient
length of stay and discharge destination in order to be able to make a valid estimation of that
economic burden. Table 1 lists the principal and auxiliary questions in detail.
Table 1. Questions of interest.
Questions
Principal Question Auxiliary Questions
Q1) With the growing trend of population ageing,
what is the potential economic burden of elderly hip-
fracture patients on the healthcare system in Ireland
Q1) What is the expected proportion of elderly
patients discharged to home, or long-stay care after
the hip fracture
over the next 10 years? Q2) Given the characteristics of an elderly hip-
fracture patient, how to predict the length of stay in
acute fi
Q3) Given the characteristics of an elderly hip-
fracture patient, how to predict the discharge
dectinaigcd
Specifically, we attempted to make contributions in two aspects. First, useful insights were
delivered in relation to the expected economic burden of hip fracture care owing to
population ageing. The insights were provided based on a well-rounded picture
corresponding to the demographic profiles and structure of the healthcare system in Ireland.
Second, the study presented the prospective application of a multi-methodology approach that
combined simulation modeling and data-driven techniques using machine learning. Machine
learning is used in an attempt to improve the predictive power of the simulation model, in
tum improving its credibility for decision making.
The 34th International Conference of the System Dynamics Society, July 2016
2. Scale of the Problem of Hip Fractures in Ireland
Around 3,000 people sustain hip fractures annually in Ireland [15]. Specifically, the rates of
fracture for the total population aged 50 years and over were reported as 407 and 140 per
100,000 for females and males respectively [16]. It was also reported that about 80% of the
elderly patients are over 75 years of age [17]. Therefore, these figures can inevitably increase
owing to the growing trend of ageing as shown in Figure 2 that plots projections of elderly
population in Ireland from 2016 to 2026.
From an economic perspective, hip fractures can represent a major burden on the Irish
healthcare system. According to the HSE, hip fractures were identified as one of the most
serious injuries resulting in lengthy hospital admissions and high costs [18]. The median LOS
was recorded as 13 days, and less than one-third go directly home after their hospital
treatment [15]. As a result, numerous studies attempted to investigate the costs associated
with hip fracture incidents [19-24]. For instance, the cost of treating a typical hip fracture
was estimated around €12,600 [18], while a different study reported a higher cost of €14,300.
Given these statistics, it can be inferred that hip fractures are, and will be, a major concern to
healthcare in Ireland, and there will be a critical need to develop evidence-based strategies in
order to meet the foreseen challenges.
1,400,000
1,200,000
1,000,000 +
800,000
600,000
seoaao BEEEEEEE
ie oe eeeeeee eo]
2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026
Population
Year
60-65 66-84 m 85+
Figure 2. The projections of elderly population in Ireland (2016- 2026) [25].
3. The Status of the Healthcare System in Ireland
A sound simulation-based study should start with understanding the problem or system of
interest [26]. In this sense, understanding the underlying structure of the Irish healthcare
system represented a major concern of the study. This section delivers a concise picture of the
underpinning components of the healthcare system in Ireland.
The healthcare system in Ireland has been undergoing a radical reform based on a phased
strategy since 2012. The fundamental goal of the reform is to transition the healthcare system
towards the integrated delivery of healthcare services. The integrated care is adopted as a
means to improve the services in relation to accessibility, quality and user satisfaction of care
services. According to the WHO, integrated care is a concept that brings together inputs,
delivery, management and organisation of services related to diagnosis, treatment, care,
rehabilitation and health promotion [27].
The 34th International Conference of the System Dynamics Society, July 2016
The transitional arrangements included structuring the Irish healthcare system into 9
geographic regions, called “Community Health Organizations”, commonly abbreviated as
CHO. The newly established CHOs are aimed to serve as integrated service areas that can
deliver better, more integrated and responsive services to people in the most appropriate
setting. Every CHO is responsible for the delivery of primary and community-based services
within national frameworks responsive to the needs of local communities. Specifically, the
CHOs include 90 Primary Care Networks (PCNs) across the country, where each PCN is
intended to serve an average population of 50,000 inhabitants. The quality of care provided
within a PCN will highly depend on how healthcare staff are organised in a way that
promotes teamwork to responsively address needs of local people. Figure 3 shows the
geographic boundaries of the 9 designated CHOs.
Equally important, the reform strategy endeavours to reorganise public hospitals into a small
number of hospital groups, each with its own govemance and management. The hospital
groups are named as follows [28]: i) Dublin North East, ii) Dublin Midlands, iii) Dublin East,
iv) South/South West, v) West/North West, and vi) Midwest. On one hand, the formation of
hospital groups can harness the benefits of increased independence and a greater control at
local level. On the other hand, grouping hospitals can allow appropriate integration in order
to improve patient flow across the continuum of care.
Figure 3. The geographic boundaries of the Community Health Organisations (CHOs) [43].
4. Approach Overview
The study endeavored to embrace a multi-methodology approach. The needs and benefits of
multi-methodologies were acknowledged within different contexts. For instance, two
arguments were made by study [12]. First, the complexity and multi-dimensionality of real-
world problems require using different methodologies to enable focusing on different aspects
of the situation. Second, a problem can go through different phases, and more than one
methodology might be required to tackle all phases. In addition, study [13] argued that the
triangulation of a situation using different methodologies can generate new insights while
enhancing confidence in the results through a reciprocal validation.
In our case, the adopted approach attempted to combine simulation modeling and machine
learning. On one hand, System Dynamics is a well-established simulation methodology that
can be used to explore the behaviour of systems over time. Further, an SD model can help
understand the long-term implications in situations where the complexity of change is
compounded by secondary effects [11]. On the other hand, machine learning has become an
instrumental artifact for building powerful prediction models. Specifically, we utilised
machine learning to provide robust data-driven predictions of variables that have a significant
influence on the problem of interest. The incorporation of the two methods is claimed to yield
more credible results with respect to decision making scenarios.
The 34th International Conference of the System Dynamics Society, July 2016
In a pipelined fashion, the approach comprised four stages. The first stage included the
development of an SD simulation model, which provided the population projections of
elderly patients. The SD model was disaggregated in accordance with the structure of the
Irish healthcare system as described in Section 5.4. Secondly, the produced projections were
used to generate individual elderly patients, whereas each patient was assigned a set of
characteristics that accurately mimicked reality. The simulation-generated patients
represented a fine-grained perspective that can be used to estimate patient outcomes on
individual basis. Thirdly, two machine leaning models were developed in order to predict the
LOS and discharge destination for every elderly patient generated by the simulation model.
The prediction models were developed and tested using Microsoft Azure Machine Learning
[14]. Based on the predicted outcomes, the cost of treatment is calculated for every patient.
Finally, the aggregation of costs can provide an overall view of the economic burden of hip
fractures. Figure 4 sketches an overview of the approach.
1, Projection 2. Generation of Patients 3. Prediction
(System Dynamics) i (Machine Learning)
—_ pPrei ted
> ui | 08 4. Cost
rediction Aggregation
i
nicht qt Predicted
Patient “> Discharge}
et Pest:
Model.
Population-Level Patient. Level H Patient-Level
Figure 4. Overview of the approach stages.
5. System Dynamics Model:
5.1 Sources of Data
The study utilised a dataset acquired from the Irish Hip Fracture Database (IHFD) [29]. The
IHFD repository is the national clinical audit developed to capture care standards and
outcomes for hip-fracture patients in Ireland. The IHFD records contain ample information
about the patient’s journey from admission to discharge. Specifically, a typical patient record
included 38 data fields such as gender, age, type of fracture, date of admission, time to
surgery and LOS. The dataset consisted of 2,024 patient records for the year 2013. Full
descriptions of the dataset fields were also available in the form of a data dictionary [30].
Mainly, the assumptions, limitations and parameters of the simulation model were
constructed based on admission and discharge records from that dataset.
With respect to population statistics, the study used projections prepared by the Central
Statistics Office (CSO) [25]. The population information contained comprehensive
information about the population in terms of age groups and sex. However, the simulation
model focused only on population aged 60 years and over, in line with the study scope.
Furthermore, we acquired additional demographic statistics from the HSE Health
Intelligence. The demographic information was prepared in connection with the 9 CHOs that
structure the healthcare system in Ireland as described in Section 3. Although, the CHOs’
statistics included only the year 2014, they were useful for setting necessary assumptions,
which are described in the next section. Figure 5 plots the reported elderly populations in
2014 with respect to every CHO.
The 34th International Conference of the System Dynamics Society, July 2016
800,000
600,000
400,000
Population
200,000
CHO1 CHO2 CHO3 CHO4 CHOS CHO6 CHO7 CHO8 CHO9
Community Health Organisations
Total Population m Elderly Population (>=60)
Figure 5. The population profiles of the 9 CHOs in 2014. The left-sided column represents the total population,
while the right-sided column represents the elderly population aged 60 and over.
5.2 Model Assumptions and Simplifications
A set of assumptions and simplifications were decided while maintaining the simulation
model as a reasonably approximate representation of the actual system. Table 2 presents what
assumptions were made and why.
Table 2. Model
and simplifications.
Purpose / Reason
A ion / Simplification
The rate of hip fracture in the total population aged 60
and over was set as 407 for females and 140 for males
per 100,000.
The rate was defined by [16].
Elderly patients were assumed as those aged 60 and
over, although usually considered as aged 65 and
older [31].
To conform to the preset hip fracture rate, which
included those aged 60 and over.
The model did not consider the scenario of patient
transfer from an acute hospital to another during
treatment course.
Only for si where the
i course
was bounded within a single acute hospital.
The elderly population for each CHO was computed
by applying a (fixed) percentage of the nation-wide
projected population on a yearly basis from 2016 to
Due to lack of population information about the 9
CHOs. The study obtained the population profiles of
the CHOs for the year 2014 only.
2026. For example, the elderly population of CHO1
was computed as 9.5% of the total elderly projected
population in 2016, whereas 9.5% was the actual
percentage in 2014.
5.3 Initial Model
The initial model provided a bird’s-eye view of the care scheme of hip fracture with respect
to the principal question of interest. The preliminary version of the model aimed at capturing
the relationships within the system components in an SD manner. Specifically, the model
focused on capturing the major dynamic behaviour in relation to the continuous growth of
ageing, and the consequent implications on the incidence of hip fractures among elderly
patients. The model defined the main actors within the system as follows: i) Elderly patients,
ii) Acute hospital, and iii) Discharge destinations including home or long-stay care facilities.
However, the initial model did not accurately describe the structure of the healthcare system
in Ireland as described in Section 3. Figure 6 illustrates the initial SD model. Table 3 lists the
model variables and Table 4 presents the model equations.
The 34th International Conference of the System Dynamics Society, July 2016
The model included a single reinforcing loop implied by the elderly patients of a fragility
history, who are susceptible to re-sustain hip fractures or fall-related injuries. According to
the HSE [18], one in three older people fall every year and two-thirds of them fall again
within six months.
Hip Fracture Rate for
Elderly Males
Potential Male
NN. +
+ )
Patients
Hip Fracture Rate for.
Elderly Females
"Total Discharged|
Patients
Fraction
Figure 6. Initial SD Model. The model depicts two potential discharge destinations as Home or Long-Stay Care.
The discharge destinations will be excluded from the disaggregated model, and will be predicted by machine
learning models on individual patient basis.
Table 3. Model variables.
‘iabl Description
Total Elderly Represents the total number of elderly population, aged 60 and over, nationwide
Population ina particular year.
Potential Male Patients | Total male patients aged 60 and over.
Potential Female Total female patients aged 60 and over.
Patients
Hip Fracture Rate for The rate of hip fracture in the total elderly male population = 140 cases per
Elderly Males 100,000.
Hip Fracture Rate for The rate of hip fracture in the total population aged 60 and over =407 for females
Elderly Females per 100,000.
InHospital Stock variable represents the total number of elderly hip-fracture patients in acute
hospitals nationwide.
Discharge Fraction
Proportion of total elderly patients discharged to home or long-stay care.
Total Discharged
Patients
Represents the total number of elderly patients discharged to home and long-stay
care.
Recurrence Rate
The rate that defines the proportion of discharged patients who are susceptible to
re-sustain a hip fracture and return to an acute hospital.
Table 4. Model equations.
Equation Type
(1) Hip Fracture Rate for Elderly Males = 140 cases per 100,000. Auxiliary
(2) Hip Fracture Rate for Elderly Females = 407 cases per 100,000. Auxiliary
(3) New Male cases = Hip Fracture Rate for Elderly Males * Potential Male Patients Inflow
(4) New Female Cases = Hip Fracture Rate for Elderly Females * Potential Female Patients Inflow
(5) Home-Discharged= InHospital * Discharge Fraction Outflow
(6) Long-Stay Care Discharged= InHospital * (1-Discharge Fraction) Outflow
(7) Recurrent Patients = (Home- Discharged * Recurrence Rate) Inflow
+(Long-Stay Care Discharged * Recurrence Rate)
(8) InHospital =Integ( (New Male cases+ New Female Cases) Stock
- (Home-Discharged + Long-Stay Care Discharged)
+ Recurrent Patients, Initial Value)
The 34th International Conference of the System Dynamics Society, July 2016
5.4 Disaggregated Model
The preliminary model did not properly consider the structure underlying the Irish healthcare
system. In contrast, the disaggregated model aimed to provide an accurate representation of
the healthcare system in terms of the 9 CHOs, and their associated elderly populations. It was
important to disaggregate the preliminary model, so that the results can be interpretable with
reference to the geographic areas that structure the healthcare system. Furthermore, the
discharge destinations (Home, Long-Stay Care) were excluded from the disaggregated model.
The discharge destinations were predicted on individual patient basis by machine learning
models. Figure 7 illustrates the disaggregated model. The model was implemented using the
R package deSolve [32].
‘Tal Elderly
Population
\\
\\I
aN
‘Toa Discharge
Patients
a
Figure 7. Disaggregated SD Model.
6. Generation of Patients
Based on the projections produced the disaggregated SD model, individual patients were
generated within every CHO. Every patient was assigned a set of characteristics including: i)
Age, ii) Sex, iii) Fracture Type, iv) Fragility History, v) ICD-10 Diagnosis, vi) Residence
Area, and vii) Host Hospital. The values of the characteristics were sampled based on the
distributions within the real dataset extracted from the IHFD (Section 5.1). The generation
8
The 34th International Conference of the System Dynamics Society, July 2016
process was implemented using the R language. The total number of generated patients
reached 1,599,000 for 50 simulation experiments. Table 5 presents the counts of elderly
patients generated within every CHO.
Table 5. Counts of patients d per CHO over 50 l experiments.
C ‘ity Health Organisation (CHO) No. of Si ‘ion-G ‘ated Patients
CHO1 151,850
CHO2 169,550
CHO3 142,450
CHO4 247,750
CHO5 187,050
CHO6 140,750
CHO7 191,900
CHO8 187,050
CHO9 180,650
7.0 Prediction: Machine Learning Models
7.1 Source of Training Data
The study used a dataset extracted from the IHFD for training both of a regression and a
classification model. As mentioned in Section 5.1, a typical patient record included 38 data
fields such as gender, age, type of fracture, date of admission, time to surgery and LOS. The
dataset consisted of 2,024 patient records for the year 2013.
7.2. Data Anomalies
A data anomaly was defined as an observation that appears to be inconsistent with the
remainder of the dataset [33], or more generally as any data that is unsuitable for the intended
use [34]. This section describes data anomalies exposed within the IHFD dataset, and the
procedures conducted to deal with them.
7.3 Outlier Removal
In order to prevent the odd influence of outliers, we considered only the samples whose LOS
were no longer than 40 days. The excluded outliers represented approximately 8% of the
overall dataset. Figure 8 plots a histogram of the LOS used to identify the outliers.
105 Oy
Figure 8. Histogram and probability density of the LOS variable. The outliers can be observed when the LOS
becomes longer than 40 days.
7.4 Dealing with Data Imbalances
The training data was originally obtained in a form of an imbalanced dataset, which was
accounted for having an adverse impact on prediction quality [35]. The problem of
imbalanced data was acknowledged as one of the profound challenges in machine learning
research [36]. In our case, imbalanced training samples were outstanding for inpatient LOS
longer than 20 days, and discharge destinations where a patient was transferred to another
acute hospital after surgery. In addition, training samples for male patients, and particular
age groups were obviously underrepresented. Figure 9 shows the imbalanced histograms of
9
The 34th International Conference of the System Dynamics Society, July 2016
the LOS and discharge destination. In order to cope with the imbalance constraint, over-
sampling technique [37] was adopted. The underrepresented samples were resampled at
random until they approximately contained as many examples as the other well-represented
samples.
il. I
aa eee si
Figure 9. The imbalanced training samples, where figures (a) and (b) plot histograms of inpatient LOS and
discharge destination respectively.
7.5 Prediction of LOS: A Regression Model
The inpatient LOS has a pivotal importance within healthcare schemes. In addition, the LOS
was recognised as the main component of the overall cost of hip fracture care [38]. A
regression forest [39] model was developed to predict the inpatient LOS.
7.6 Prediction of Discharge Destination: A Multi-Class Classifier
The intuition was that a discharge destination can be predicted based on patient’s
characteristics and the LOS, which was predicted separately by the regression model. A
random forest [39] model was developed for predicting the discharge destination.
7.7 Feature Selection
Initially, the dataset contained 38 features, however they were not all relevant. Intuitively
irrelevant feature were simply excluded. In addition, the most influential features were
decided based on the technique of permutation feature importance [40]. Table 6 presents the
set of features used by the predictors, and their associated importance scores.
Table 6. Selected features in descending order with respect to importance score.
Predictor Model Selected Features
Feature Importance
Score =
Host Hospital 0.71
Patient Age 0.50
Los ICD Diagnosis 0.46
Reoressian Motel Patient Residence Area 0.40
Fracture Type 0.39
Patient Sex 0.29
Fragility History 0.22
Host Hospital 0.44
iiinn _ Patient Age 0.35
ae Cecfier LOS 0.21
Patient Residence Area 0.20
Patient Sex 0.13
10
The 34th International Conference of the System Dynamics Society, July 2016
7.8 Predictors Evaluation
The predictive models were tested using a subset from the dataset described in Section 5.1.
The randomly sampled test data represented approximately 40% of the overall dataset. The
prediction error of each model was estimated by applying 10-fold cross-validation. Table 7
presents evaluation metrics of the LOS regression model, while Figure 10 shows the
confusion matrix of the discharge destination classifier.
Table 7. Average 10-fold cross-validation accuracy of the LOS predictor.
| Relative Absolute Error | Relative Squared Error | Coefficient of Determination |
i 0.26 | =0.17 | 50.83
Predicted Class
Transferred-to- 4 Nursing
Hospital ™® Home
ctual Class
classifier.
Figure 10. Average 10-fold cro:
8.0 Calculation of C ost
This section explains how the cost of treatment was calculated for every simulation- generated
patient. We utilised the study [23] that provided comprehensive information on the costs
within hip fracture treatment in Ireland. Generally, the cost of treatment was calculated as the
equation below. Table 8 and Table 9 provide detailed information on the cost of every item.
[ Cost of Treatment= (ED Cost) + (Hospital Inpatient Cost) + (Outpatient Visits Cost) + (Long-Stay Care Cost) ]
Table 8. The description of cost assumptions.
Item Description Approximate C ost
ED Cost The cost of admission to the Emergency €602
Department
Hospital Inpatient Cost: The inpatient stay at an acute hospital According to Table 9
Outpatient Visits Cost The cost of outpatient visits after discharge. €154 * 9 Visits= €1386
Long-Stay Care Cost The cost implied by staying in long-stay care, €700 * 32 Weeks = € 22400
(Optional) such as nursing homes.
Table 9. Inpatient hospital costs for patients aged 65 and over.
Age Group Average Cost per Average Cost per
Case (Male) € Case (Female) €
65-69 7,020 5,909
70-74 8,365.64 6,353
75-79 9,249 7,879
80-84 10,418 9,376
85+ 11,094 9,902
11
The 34th International Conference of the System Dynamics Society, July 2016
9.0 Visualisation of Results
This section aims at interpreting the results in a visual manner. All the results were obtained
by averaging the outputs over 50 simulation experiments. Figure 11 shows the overall
predicted cost of hip fracture treatment per year from 2016 to 2026. The cost is expected to
continuously increase, and reach around 84 M by the year 2026.
Average Cost per Year over 50 Simulation Experiments
Cost (Million Euros)
3
2016 2018 2020, 2022 2024 2026
Year
Figure 11.The average cost per year over the simulated period from 2016 to 20126. The cost was averaged over
50 simulation experiments.
Figure 12 plots the average accumulative cost for home-discharged patients compared to
those who are expected to be discharged to long-stay care such as nursing homes. The figure
reveals that there is a clear discrepancy between the two proportions. Though, this
discrepancy agrees with that less than one-third of elderly hip-fracture patients go directly
home after their hospital treatment, as reported by the HSE [15].
700
600
500
400
300
* [|
i)
Home-Discharged Long-Stay Discharged
Accumulative Cost (€ Million)
Figure 12.The average ive cost for h i patients comp: to those di to long-
stay care.
12
The 34th International Conference of the System Dynamics Society, July 2016
Figure 13 refines the model results with respect to every CHO individually. Specifically, the
figure plots the average accumulative cost for patients discharged to home and long-stay care
within every CHO. It can be clearly observed that particular CHOs are expected to have
significant higher levels of patients who are discharged to long-stay care. This prediction can
be reasonable as CHO4, for example, has the highest proportion of elderly people nationwide.
200,000,000
175,000,000
150,000,000
125,000,000
100,000,000
75,000,000
50,000,000 +
25,000,000 +
o 4
CHO1 CHO2 CHO3 CHO4 CHOS CHO6 CHO7 CHO8 CHO9
mHome-Discharged =m Long-Stay Care Discharged
Figure 13.The average accumulative cost for home-di patients comp: to those di to long-
stay care in every CHO.
Figure 14 visualises a heat map of the 9 CHOs with regard to the overall predicted cost over
the simulated period. The CHO4, CHO7 and CHO3 were expected to have the highest levels
of costs in relation to elderly hip-fracture patients.
wid
Figure 14. Heatmap: Overall predicted cost within every CHO. The predicted cost with respect to a given CHO
is visually indicated by red (high) and green (low) in Figure (a).
13
The 34th International Conference of the System Dynamics Society, July 2016
10. Verification and Validation
10.1 Model Verification
In order to examine the logic and suitability of the model, it was verified qualitatively and
quantitatively. Throughout the simulation model’s development, a set of verification tests [41] were
conducted as follows:
e Structure-V erification Test: The model structure was checked compared to the actual system.
Specifically, it was verified that the model structure reflected reality in terms of the
underlying CHOs, and associated elderly populations.
e Extreme Conditions Test: The equations of the simulation model were tested in extreme
conditions. For example, flows of patients were set at extreme conditions (e.g., there is no
elderly population aged 60 or over).
e Parameter-Verification Test: The model parameters and their numerical values were
inspected to correspond conceptually and numerically to reality. Specifically, probability
distributions of patient attributes output from the model were compared against those derived
from the real system, such as age, sex and fracture types for example.
10.2 Model Validation
According to [42], the most definitive test of simulation model validity is comparing outputs of the
simulation model to those of the actual system. Similarly, we used the variables of discharge
destination and LOS as a measure of the approximation between the simulation model and the actual
system.
On one hand, Figure 15 provides a histogram-based comparison between the actual system and the
simulation model regarding the discharge destination. The comparison showed that the distributions
of the actual and simulated data were relatively close. However, the comparison revealed that the
model slightly underestimated and over-estimated the proportion of patients discharged to long-stay
care.
(@) (b)
Figure 15. Hi of the destination for the actual system and simulation model, where (a) and (b)
represent the actual system and simulation model respectively.
On the other hand, Figure 16 compares the actual system’s average LOS to that of the simulation
model with respect to the 9 CHOs separately. The figure clearly shows that the simulated CHOs’
average LOS matched the actual system very well, without any significant over- or under
estimations. Overall, validation and verification tests proved that the simulation model can be
suitable for answering questions from the perspective of the study’s intended objectives.
14
The 34th International Conference of the System Dynamics Society, July 2016
Average LOS (Days)
CHO1 CHO2 CHO3 CHO4 CHO5 CHO6 CHO7 CHO8 CHO9
Community Health Organisations (CHOs)
m Actual System = Simulation Model
Figure 16. CHO-based comparison between the actual system and simulation model in terms of average LOS.
11. Study Limitations
e Only public acute hospitals were considered, from which the IHFD records were obtained.
e The records of the IHFD dataset did not evenly represent the 9 CHOs.
e The real data obtained by the study covered only a single year, which was 2013.
e The rate of hip fractures was assumed as a constant over the simulated interval, however it
might increase or decrease in reality.
e The study considered the in-hospital cost of the patients aged 60-64 the same as 65-69, due to
lack of information on this issue.
e The study did not consider other potential costs such as the ambulance costs.
e The study did not consider the indirect costs such as the quality of life.
e The study did not distinguish between the patients who are discharged to long-stay nursing
homes and rehabilitation institutions, due to lack of information on rehabilitation institutions.
12. Conclusions
The significance of evidence-based decision making for healthcare has increased owing to the
phenomenal challenge of population ageing. The study presented a multi-methodology approach that
integrated System Dynamics with machine leamming techniques. On a population basis, the SD model
realised a population-based perspective of the demand for hip fracture care, regarding elderly
patients in particular. On an individual patient basis, machine leaning models were used to make
accurate predictions on the factors that have a significant impact on the cost of treatment.
Specifically, the inpatient length of stay and discharge destination were predicted for every
simulation- generated patient using regression and classification models respectively. The results are
articulated using a combination of domain knowledge within simulation modeling and robust data-
driven prediction with machine learning.
The predicted costs are provided with reference to the geographic structure of the healthcare system
in Ireland in terms of the Community Health Organisations (CHOs), whereas particular CHOs, such
as CHO4 and CHO7, are predicted to experience considerable costs compared to other CHOs. The
results also emphasise that the significant proportion of costs can mostly be attributed to elderly
patients discharged to long-stay care facilities such as nursing homes. Generally, the study can carry
useful insights for predicting the potential economic burden of elderly hip-fracture patients in
Treland.
15
The 34th International Conference of the System Dynamics Society, July 2016
References
[1] http://www.un.org/en/devel d ll ission/2007/key! hatter} pdf
[2] Rechel, Bemd, Emily Grundy, Jean-Marie Robine, Jonathan Cylus, Johan P. Mackenbach, Cecile Knai, and
Martin McKee. Ageing in the European Union. The Lancet 381, no. 9874 (2013): 1312-1322.
[3] European Commission. 2009 A geing report: ic and budgetary projections for the EU-27 Member States
(2008-2060).Luxembourg: Office for Official Publications of the European Communities, 2009.
[4] Health Service Executive. Anmual Report and Financial Statements 2014, 2014.
[5] http://www.un.org/en/d lation/publi ing/W orldPopulationA geing2015_Info
Chart.pdf
[6] Cooper, Campion, G. Campion, and LJ1992 Melton III. "Hip fractures in the elderly: a world-wide projection."
Osteoporosis international 2, no. 6 (1992): 285-289.
[7] Melton, L. J. "Epidemiology of hip fractures: implications of the exponential increase with age." Bone 18, no. 3
(1996): $121-S125.
Gullberg, B., O. Johnell, and J. A. Kanis. "World-wide projections for hip fracture." Osteoporosis international
7, no. 5 (1997): 407-413.
[9] Melton LJ 3rd (1993) Hip fractures: a worldwide problem today and tomorrow. Bone 14(Suppl 1):S1-S8 2.
[10] Kannus P et al (1996) Epidemiology of hip fractures. Bone 18 (Suppl 1):57S-63S.
[11] Reagan-Cirincione, P., Schuman, S., Richardson, G.P. and Dorf, S.A., 1991. Decision modeling: Tools for
strategic thinking. Interfaces, 21(6), Pp. 52-65.
[12] Brocklesby, J., 1997. B ing hodology literate: An of the cognitive difficulties of
working across paradigms. Multimethodology: The theory and practice of combining management science
methodologies, pp.189-216.
[13] Mingers, J., 2001. Multimethodology: Mixing and matching methods.
[14] Barga, R., Fontama, V. and Tok, W.H., 2015. Predictive Analytics with Microsoft Azure Machine Learning.
[15] Ellanti, Prasad, Breda Cushen, Adam Galbraith, Louise Brent, Conor Hurson, and Emer Ahern. "Improving Hip
Fracture Care in Ireland: A Preliminary Report of the Irish Hip Fracture Database." Journal of osteoporosis 2014
(2014).
[16] M. K. Dodds, M. B. Codd, A. Looney, and K. J. Mulhall, “Incidence of hip fracture in the Republic of Ireland
and future projections: a population-based study,” Osteoporosis International,vol. 20, no. 12, pp. 2105-2110,
2009.
[17] Laffoy, M., 2008. Strategy to prevent falls and fractures in Ireland's ageing population summary, conclusions
and recommendations.
[18] http://www.hse.ie/eng/services/publications/olderpeople/Executive Summary _-
Strategy _to_ Prevent Falls and Fractures in Ireland%sE2%80%99s Ageing Population.pdf
[19] McGowan, B., Casey, M.C., Silke, C., Whelan, B. and Bennett, K., 2013. Hospitalisations for fracture and
associated costs between 2000 and 2009 in Ireland: a trend analysis. Osteoporosis International, 24(3), pp.849-
857.
[20] Azhar, A., Lim, C., Kelly, E., O'Rourke, K., Dudeney, S., Hurson, B. and Quinlan, W., 2007. Cost induced by
hip fractures. Irish medical journal, 101(7), pp.213-215.
[21] Haentjens, P., Lamraski, G. and Boonen, S., 2005. Costs and consequences of hip fracture occurrence in old
age: an economic perspective. Disability and rehabilitation, 27(18-19), pp.1129-1141.
[22] Cotter, P.E., Timmons, S., O’Connor, M., Twomey, C. and O’Mahony, D., 2006. The financial implications of
falls in older people for an acute hospital. Irish joumal of medical science, 175(2), pp.11-13.
[23] Gannon, B., O’Shea, E. and Hudson, E., 2007. The economic costs of falls and fractures in people aged 65 and
over in Ireland. Irish Centre for Social Gerontology, Galway.
[24] Carey, D. and Laffoy, M., 2005. Hospitalisations due to falls in older persons. Irish medical journal, 98(6),
pp.179-181.
[25] http://www.cso.ie/en/statistics/population/
[26] Kelton, W.D. and Law, A.M., 2000. Simulation modeling and analysis. Boston: McGraw Hill.
[27] Grone, O. and Garcia-Barbero, M., 2001. Integrated care. Intemational journal of integrated care, 1(2).
[28] Higgins, J.R., 2013. The establishment of hospital groups as a transition to independent hospital trusts.
Department of Health and Children, Dublin, Ireland.
[29] https://www.noca.ie/irish-hip-fracture-database.
[30] http://www.hpo.ie/hipe/hipe_data_dictionary/HIPE_Data_Dictionary_2015_V7.0.pdf
[8
16
The 34th International Conference of the System Dynamics Society, July 2016
[31] Rosenberg, M. and Everitt, J., 2001. Planning for aging populations: inside or outside the walls. Progress in
Planning, 56(3), pp.119-168.
[32] Soetaert, K.E.R., Petzoldt, T. and Setzer, R.W., 2010. Solving differential equations in R: package deSolve.
Journal of Statistical Software, 33.
[33] Hodge, V J. and Austin, J., 2004. A survey of outlier detection methodologies. Artificial Intelli Review,
22(2), pp.85-126.
[34] Sarsfield, S., 2009. The data governance imperative. IT Governance Publishing.
[35] Japkowicz, N. and Stephen, S., 2002. The class imbalance problem: A systematic study. Intelligent data
analysis, 6(5), pp.429-449.
[36] Y ang, Q. and Wu, X., 2006. 10 challenging problems in data mining research. International Journal of
Information Technology & Decision Making, 5(04), pp.597-604.
[37] Japkowicz, N., 2000, July. Learning from imbalanced data sets: a comparison of various strategies. In AAAI
workshop on learning from imbalanced data sets (Vol. 68, pp. 10-15).
[38]Johansen, A., Wakeman, R., Boulton, C., Plant, F., Roberts, J. and Williams, A., 2013. National Hip Fracture
Database: National Report 2013. Clinical Effectiveness and Evaluation Unit at the Royal College of Physicians.
[39] Breiman, L., 2001. Random forests. Machine leaming, 45(1), pp.5-32.
[40] Altmann, A., Tolosi, L., Sander, O. and Lengauer, T., 2010. Permutation importance: a corrected feature
importance measure. Bioinformatics, 26(10), pp.1340-1347.
[41] Martis, M.S., 2006. Validation of simulation based models: a theoretical outlook. The Electronic Journal of
Business Research Methods, 4(1), pp.39-46.
[42] Law, A.M., 2009, December. How to build valid and credible simulation models. In Simulation C:
(WSC), Proceedings of the 2009 Winter (pp. 24-33). IEEE.
[43] http://www.hse.ie/eng/services/publications/corporate/CHO_Chapter_1.pdf
17
