See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/274390870

Demand Forecasting and Capacity Management

for Hospitals

Article · September 2011

READS

233

2 authors, including:

Oscar Barros
University of Chile
40 PUBLICATIONS 148 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Oscar Barros
letting you access and read them immediately. Retrieved on: 17 May 2016
 1

DEMAND FORECASTING AND CAPACITY

MANAGEMENT FOR HOSPITALS

1
Oscar Barros , Richard Weber, Carlos Reveco, Eduardo Ferro, and Cristian Julio

Department of Industrial Engineering, University of Chile, Santiago, Chile

Abstract:

Demand forecasting and capacity management are complicated tasks for certain healthcare
services due to the inherent uncertainty, complex relationships, and typically high public exposure
involved. Health service demand in three Chilean hospitals has been studied concluding that it
can be forecast with high accuracy using Neural Networks and Support Vector Regression. This
investigation has allowed us design a process to manage demand by transforming the respective
demand forecasts into the resources needed to proper attention. Comparing required resources
with available resources and simulating various scenarios permits taking corrective actions when
capacity is not aligned with demand.

The proposed forecasting methods and the capacity management process have been accepted
by hospital management and staff and are currently in use in one hospital. To support the efficient
use of the developed forecasting and management methods, advanced IT systems have been
implemented that allow the routine use of the respective processes. We are currently
implementing processes and systems in one of the other participating hospitals. The results have
been so encouraging that National Health Authorities are considering the extension of the
proposed demand forecasting and management practices to close to one hundred public
hospitals in Chile.

Keywords: Health care management, forecasting, process management, capacity management

1
Corresponding author: [email protected]; Republica 701, Santiago, Chile, tel. 562
9784037, fax 562 9784011
 2

1. INTRODUCTION
Public hospitals in Chile have, in general, more demand for health services than available
capacity permits to attend within a reasonable time frame. Hence it is important for a hospital to
forecast demand with great precision, in order to adjust capacity or take alternative courses of
action, e.g. transferring patients to other facilities. For example, it is possible to discharge patients
from a hospital to a local health service center for non-complex pathologies; also private services
can be hired in case of an emergency that cannot be treated at a public hospital. Since demand
forecasts are not sufficient on their own, but serve as input for hospital management, it is required
that service demand be predicted not only on an aggregated level but for different pathology
types separately, which makes them technically more demanding.

The forecasted demand for each pathology type allows determining the required resources,
such as doctors of different specialties, reception areas, emergency room cubicle capacity and
operating room capacity. Comparing the resources needed to satisfy demand with available
capacity permits making decisions to adjust capacity or prevent or transfer demand.

Public hospitals in Chile, which process 75% of the country’s demand for health services
(MIDEPLAN, 2006) are not using any formal way of forecasting demand and managing capacity.
Current procedures are informal and defined based on the experience of the participants in the
process; furthermore such procedures are mainly oriented to solving the problem of excess
demand when it occurs. To be fair, there are some informal attempts to foresee how bad the
winter period, when most excess demand is produced, is going to be and to make some
decisions regarding the number of doctors and hospital beds that will be made available during
this season at a given hospital.

Given the situation outlined, we agreed with the Chilean Health Authority to perform an
applied research program that would use state-of-the-art analytical tools, process design
methodologies, and IT to develop a general solution for demand forecasting and capacity
management that could eventually be used at all Chilean hospitals.

Benefits expected from this work are:

 Significantly improved service to hospital patients, with a shorter waiting time.

 Better use of resources for the entire health system, due to a better distribution of demand
to the level that could best serve it.
 Better use of resources at each hospital, since their planning can be made with advanced
knowledge regarding demand that allows capacity optimization.

We started the research in March, 2009 and selected three hospitals to be studied to
develop the methods, processes, and systems that will eventually be used in all Chilean
hospitals.

Demand forecasting and management is part of a larger design that intends to provide a
systematic solution to global hospital management. Such a solution is based on the design of a
general process structure that we developed for hospitals and that defines the management
processes which are needed to ensure a predefined service level for patients and to optimize the
use of the required resources. The general process structure allowed us to determine the key
processes where implementation of new practices would generate most value (Barros and Julio,
2010, 2011). In agreement with health authorities we selected the process described here and
another one related to operating room scheduling. In each of the selected hospitals we evaluated
 3

the current situation of demand forecasting and capacity management to determine the feasibility
of introducing analytical and formal practices to improve the respective processes.

The results we present in this paper have been developed in collaboration with hospital
staff. They reviewed each of the steps described below, which led to a working process for
forecasting and managing demand. Emphasis is also given to the experiences we documented
during this work and that could be beneficial in similar future projects.

Section 2 of this paper reviews the literature on the use of analytical methods in forecasting
and other experiences in hospital capacity management. Section 3 presents how hospital
demand has been modeled using several methods with the results obtained. The processes that
convert forecasts into the resources needed to satisfy demand and manage capacity are
described in Section 4. Section 5 presents conclusions and provides suggestions for future work.

2. REVIEW OF RELEVANT EXPERIENCE

Demand forecasting is a useful and well-studied subject (Armstrong, 2001) that has
generated important results in different areas, such as the retail industry (Aburto and Weber,
2007) and inventory control at several enterprises such as Dell (Kapuscinki et al, 2004).
Forecasts provide relevant information for making decisions on the resources needed to provide
adequate service to meet the potential demand and to avoid stock breakdowns or overstocking.

There is another line of demand forecasts focused on services. In it the variable to predict
is the number of clients who will demand the service, in order to manage capacity needed to
provide a given level of service. In a recent work, joint demand and capacity management have
been proposed for services in a restaurant (Hwang et al, 2010) where the main focus lies on
optimizing revenue for a given dynamic demand without considering, however, demand
forecasting explicitly. A similar study has been proposed for scheduling elective surgery under
uncertainty (Min and Yih, 2010) but again without considering uncertain demand which is the
main focus of our paper.

In the case of hospital services the capacity is determined by available physical facilities,
such as medical cubicles, operating rooms and beds, and human resources, such as doctors,
who perform diagnostic procedures and treatments on patients. This capacity should be planned
to guarantee a given service level and optimize use of resources; for this an accurate forecast of
the number and type of patients who will arrive in the future is needed.

Many different methods have been proposed for forecasting (Armstrong, 2001; Box et al,
1994), and several studies compare such methods in terms of accuracy of results. One of these
studies that is relevant to our work compares Neural Networks with other econometric methods
and concludes that the former give, in general, better results (Adya and Collopy, 1998). As will be
shown below, in our experiments the technique of Support Vector Regression can even
outperform Neural Networks.

Few studies of formal demand forecast in the health area have been published. Some of
these have focused mostly on predicting the number of beds required to meet emergency
demand (Jones et al, 2002; Schweigler et al, 2009; Farmer and Emani, 1990). These studies
have focused on forecasting demand in the emergency room where all patients must be attended
to, even with a considerable delay. This is important because there is no possibility of changing
the appointment to another date, or of having patients leave without attention, which is relevant to
 4

the input data, because historical demand is equal to the number of patients attended. This fact
will be important for the present work, since we were only able to find good data for emergency
services. Several studies have shown, however, that in practice a small difference between
patient arrival and care service could exist (Kennedy et al, 2008) a fact that has been taken care
of in our system (see Section 4.4). Another work that uses an approach similar to ours is reported
in Shirxia et al, (2009) but we will show that our approach provides superior results. For capacity
management the usual procedure has been to simulate the flow of patients through emergency
facilities. None of the papers we have reviewed considers an explicit state of the art demand
forecasting technique, except the one by Marmor et al, (2009) that estimates demand based on a
long term moving average over the demand. Other papers that use the common approach of
static arrival distribution are the following: Garcia et al, (1995); Samaha et al, (2003); Rojas and
Garavito, (2008); and Khurma and Bacioiu, (2008).

3. FORECASTING METHODS: APPLICATIONS

AND RESULTS
In this section we describe exploratory analysis and preprocessing the available data,
show how the different forecasting models has been generated and present the respective
results.

3.1. Analysis and Pre-processing of Available Data for

Forecasting
To be able to forecast effectively one of the key factors is the quality of historical
information. In addition, the hospital operating conditions and environment should remain
relatively stable. This work focuses on two public pediatric hospitals: Luis Calvo Mackenna (from
now on referred to as HLCM) and Exequiel González Cortés (HEGC), and a general purpose
hospital, San Borja Arriarán (HSBA). These hospitals have quality data in their emergency room
areas.

However, to turn this quality data into useful information for the forecasting models, further
analyses and a series of transformations were necessary. By analyzing the demand that arrives
at the emergency department outliers were detected. We found that two months had substantially
higher demand than the average of the remainder of the months and decided therefore to replace
them with the respective average; see Figure 1. In both cases of removed outliers one particular
situation occurred: a special kind of flu virus led to the abnormally increased demand which
consequently could not be predicted based on historical demand data. As a solution we propose
to use expert knowledge of the physicians with experience in emergency room data in order to
adjust the forecast proposed by our system in cases where such special events happen.

Visual inspection of aggregated demand shown in Figure 1 reveals a strong seasonal

pattern. We observe a low demand during the summer months (December-January-February)
and a high influx of patients during the months of the winter season (May – June – July; in the
southern hemisphere). This is due to the fact that high air pollution, smog, and low temperatures
lead to respiratory diseases increasing the number of emergencies. In general a downward trend
can be observed over the years.
 5

12000
11000
10000
Number of Patients

9000
8000
7000
6000 Actual Demand

5000 Adjusted Demand

4000
3000
2000
2001 2002 2003 2004 2005 2006 2007 2008 2009
Year

Figure 1: Actual vs. Adjusted Demand per Month in Luis Calvo Mackenna Hospital

When data is disaggregated by pathology type, e.g. medical and surgical, we notice huge
differences: the first one is much more volatile since it depends on factors such as temperature
and influenza like illness rate, as suggested in Jones et al, (2002) while the second one is more
stable, as shown in Figures 2 and 3. From the data it is also possible to conclude that medical
demand comprises 70% of the emergency cases and surgical demand corresponds to 30% of the
cases.

Figure 2: Medical demand for HLCM Figure 3: Surgical demand for HLCM

Demands at HEGC and HLCM have a very similar behavior, since both are children
hospitals with comparable size and target populations. HSBA data also follows a similar pattern.
 6

On arrival at the emergency facilities each patient is registered, including personal data,
time of arrival, diagnosis, and classification according to severity of illness. For the purposes of
this work we obtained all this historical data for the three hospitals as follows:

 HLCM: from January, 2001 to December, 2009

 HEGC: from January, 2001 to July, 2009
 HSBA: from January, 2000 to December, 2009.

For the purpose of capacity management it would also be interesting to have the exact time
when medical attention starts. This could differ significantly from arrival time but it has not been
registered to date.

Unusual demand for treatment of pathologies that appear occasionally, such as allergies
and A H1N1, was also discarded, because there was not enough data to detect a pattern; and, in
general, outliers were discarded replacing them with an average as mentioned earlier. Daily
individual data for patients was aggregated for each month to conform to the time series that we
modeled.

This data cleaning procedure was also performed to HEGC and HSBA data.

3.2. Forecasting Methods and their Testing

Four forecasting methods were tested: Linear Regression, Weighted Moving Averages,
Neural Networks, as suggested in McLaughlin and Hays, (2008), and Support Vector Regression
(SVR). The first two are well known techniques used for forecasting and described in the
literature (see Armstrong, 2001). Neural Networks and SVR are recently used techniques for
forecasting and will be summarized below.

The particular type of network we used is the Multi-layer Perceptron (MLP). Its basic units
are neurons that are grouped in layers and are connected by means of weighted links between
two layers. Each neuron receives inputs from other neurons and generates a result that depends
only on the information locally available and which serves as input to other neurons. The
architecture of the network is shown in Figure 4.

Figure 4: Architecture of a Neural Network (MLP).

 7

Each neuron operates according to the structure in Figure 5, where the output y is
determined as a function of the weighted inputs.

Output
Neuron
Weight
Input

Figure 5: Neuron Details

Function f in Figure 5 is called the activation function and may take different forms; the
most commonly used one for continuous outputs is the logistic function, as shown in Eq. (1).

The network was trained with the above mentioned historical data. The basic idea is that
previous data predicts a given future month. In particular we assumed that the pattern was
seasonal and therefore used previous values of the month we want to predict as some of the
inputs to the model. The structure of the network consists of an output layer with one neuron that
generates the desired forecast. The input layer contains the variables we will use to explain the
demand. In the hidden layer we used a number of neurons between input and output neurons,
since a high number will tend to copy the data (over fitting) and a small number will not produce
good forecasts.

As mentioned already, previous months were used as input data. However, there are
months that are more relevant than others. We tried to select relevant attributes using a genetic
algorithm, as suggested in Shirxia et al, (2009), but results were not encouraging. In Shirxia et al,
(2009) a common pitfall in neural network design was made, which is to separate the data set into
just two groups: one for training and one for testing (Zhang, 2007). This results in trying to
minimize the error over the testing data and leads to an over fitting of the resulting model. In our
case, we divided the data into three sets: 70% for training; 20% for testing, where the network is
trained to minimize the test error. The third set with 10% of the data is independently used to
validate results. This use of an independent set provides a better evaluation of future results.

Regarding the network architecture we tested several parameters, such as the number of
epochs to use, the learning rate, and the number of hidden neurons. Best results were obtained
for 10,000 training epochs, maintaining the model with minimum error in the training set; a
learning rate of 0.2 with a momentum of 0.3. Also, decaying was introduced, but this only helps to
get to the solution faster with no significant changes in results.

Based on results which will be shown later, we selected a Neural Network with 18 input
neurons. If N is the index of the month to be forecast , three neurons corresponding to the values
of the same month in previous years, N-12, N-24, and N-36, were included; 3 neurons
representing the tendency between months given by the differences between N-12 and N-13, N-
 8

24 and N-25, N-36 and N-37 and a set of 12 binary variables to represent the months of the year
are also part of the network`s input layer. This provided a solution that can forecast up to a year
in advance and takes account of tendencies. Thus the network has 18 input neurons plus an
additional bias neuron that helps to separate cases and allows having smaller Neural Networks
than without this bias.

The output layer contains simply one neuron that generates the forecasted demand in
month N. The hidden layer contains 10 neurons providing the model an adequate degree of
freedom, usually calculated by (Number of input neurons + Number of output neurons)/2. The
resulting network is shown in Figure 6.

Figure 6: Resulting Neural Network Architecture

Another method we tested is Support Vector Regression (Chen and Schölkopf, 2005;
Hofmann et al, 2008; Smola and Schölkopf, 2004), which is a variation of Vector Support
Machines (SVM) based on the following idea. SV regression (SVR) performs linear regression in
 9

a high-dimensional feature space generated by a kernel function as described below, using the -
insensitive loss function proposed by Vapnik, (1995). This function allows a tolerance degree to
errors not greater than  as shown in Figure 7. The description is based on the structure and
terminology used in Smola and Schölkopf, (2004).

We start with a set of training data where each denotes the

input space of the sample and has a corresponding target value being the
number of available data points to build the regression model. The SVR algorithm applies a
function transforming the original data points from the initial Input Space ( ) to a generally
higher dimensional feature space ( ). In this new space, a linear model is constructed,
which represents a non-linear model in the original space:

With and

In Eq. (3), denotes the dot product in . When the identity function is used, i.e.
, no transformation is carried out and linear SVR models are obtained.

The goal when using the -insensitive loss function is to find a function that fits given
training data with a deviation less or equal to  and, at the same time, is as flat as possible in
order to reduce model complexity. This means that one seeks a small weight vector . One way
to ensure this is by minimizing the norm (Smola and Schölkopf, 2004), leading to the
following optimization problem:

This problem could be infeasible. Therefore, slack variables , are

introduced to allow error levels greater than  (see Figure 7), arriving at the formulation in Eqs. (6)
to (9).

s.t.
 10

Figure 7: Support Vector Regression to Fit a Tube with Radius ε to the Data and
Positive Slack Variables ξi

This is known as the primal problem of the SVR algorithm. The objective function takes into
account generalization ability and accuracy in the training set, and embodies the structural risk
minimization principle (Vapnik, 1998). Parameter C > 0 determines the trade-off between
generalization ability and accuracy in the training data, and the value up to which deviations
larger than are tolerated. The -intensive loss function has been defined as in Eq. 10.

(10)

It is more convenient to represent this optimization problem in its dual form. For this
purpose, a Lagrange function is constructed and, once applying saddle point conditions, the dual
problem in Eqs. (11) to (14) is obtained (Vapnik, 1995).

s.t.
 11

This is the quadratic optimization problem that has to be solved to obtain the solution of the
SVR model, which is a function of the dual variables and . Using saddle point conditions it
can be shown that Eq. (15) holds (Vapnik, 1998). Replacing this expression in Eq. (3), the final
solution of the SVR algorithm is obtained as Eq. (16).

Here, the expression is equal to which is known as the Kernel

Function (Vapnik, 1995). The existence of such a function allows us to obtain a solution for the
original regression problem, without knowing explicitly the transformation applied to the
data. Some examples of common Kernel Functions are shown in Table 1.

Name Kernel Functions

Linear
Polynomial
Radio Basis Function

Table 1: Common Kernel Functions

It is well-known that the generalization performance of SVR (estimation accuracy) depends on a

good setting of parameters C, and the kernel parameters (Vladimir, Ma, 2004). We use a Grid-
Search to find good parameters for SVR with Radial Basis Function as kernel ( ) within
-1 0 10
previously specified ranges for parameters C, ε, and . In our case we use, C=10 ,10 …10 ;
-10 -9 -1 -8 -7 0
ε=2 , 2 ,…, 2 ; =2 , 2 ,…, 2 .

3.3. Results
We used Mean Average Percentage Error (MAPE) and Mean Square Error (MSE) as
performance measures to determine model accuracy, as defined in Eqs. (17) and (18).
 12

For the Linear Regression, Weighted Moving Average, and SVR the same inputs as the
ones described for the Neural Network were used (except for the bias). Results obtained using
these four methods for the validation sets of all hospitals are displayed in Table 2.

Weighted Moving
Linear Regression Neural Network SVR
Average

MAPE MSE MAPE MSE MAPE MSE MAPE MSE

HLCM Medical
12.67% 150,686 7.53% 144,729 7.45% 161,689 5.61% 154,861
Demand
HLCM Surgery
6.54% 27,097 7.36% 20,137 8.99% 22,947 5.09% 25,199
Demand
HEGC Medical
15.91% 3,114,376 16.5% 1,978,332 7.7% 1,043,753 6.86% 606,324
Demand
HEGC Surgery
8.55% 14,302 8.96% 11,730 8.3% 12,155 5.88% 8,120
Demand
HEGC Orthopedic
8.41% 35,940 8.60% 28,247 5.12% 29,851 4.44% 25,460
Surgery Demand
HSBA Medical
8.27% 3,125,071 11.83 % 850,342 7.9% 1,226,165 6.97% 643,984
Demand
HSBA Maternity
10.54% 23,738 6.98% 12,408 10.6% 38,629 3.24% 7,867
Demand

Table 2: Forecast Result Errors (best results in bold)

As shown in Table 2, in four out of seven cases best results are obtained with SVR, when
using MSE as criterion to compare the performance of the different models. When using MAPE
as criterion for comparison, SVR appears as the best option for demand forecasting in all cases.
These results were obtained using Rapid Miner 4.6.000, the Neural Network library from WEKA
and SVR from LIBSVM (Chang and Lin, 2001) Library but using Rapid Miner as graphic user
interface (GUI).

Figures 8 and 9 display the results graphically showing forecast with SVR and actual
demand for HLCM, with a 90% confidence interval for the forecast. This interval has been
calculated based on the hypothesis that the forecast error has a normal distribution with a mean
zero. This hypothesis has been confirmed using a Kolmogorov-Smirnov test.
 13

7000
6500
6000
5500
Number of Patients

5000
4500 Actual Demand
4000 Forecast
3500
Upper Bound
3000
2500 Lower Bound
2000

Figure 8: Medical Demand in Luis Calvo Mackenna Hospital

1400
1300
1200
1100
Number of Patients

1000
900 Actual Demand
800 Forecast
700
Upper Bound
600
500 Lower Bound
400

Figure 9: Surgical Demand in Luis Calvo Mackenna Hospital

Similar results are shown in Figures 10, 11 and 12 for HEGC, and in Figures 13 and 14 for
HSBA.
 14

10000
9000
Number of Patients 8000
7000
6000 Actual Demand
5000 Forecast
4000 Upper Bound
3000 Lower Bound
2000

Figure 10: Medical Demand in Exequiel Gonzales Cortes Hospital

1400
1300
1200
1100
Number of Patients

1000
900 Actual Demand
800 Forecast
700
Upper Bound
600
500 Lower Bound
400

Figure 11: Surgical Demand in Exequiel Gonzales Cortes Hospital

 15

3000

2500
Number Of Patients
2000
Actual Demand
1500
Forecast
1000 Upper Bound
Lower Bound
500

Figure 12: Orthopaedic Surgical Demand in Exequiel Gonzales Cortes Hospital

9000

8000
Number of Patients

7000

6000 Actual Demand

Forecast
5000
Upper Bound
4000
Lower Bound
3000

Figure 13: Medical Demand in San Borja Arriarán Hospital

 16

1700

1600

Number of Patients 1500

1400
Actual Demand
1300
Forecast
1200 Upper Bound
1100 Lower Bound
1000

Figure 14: Maternity Demand in San Borja Arriarán Hospital

Based on the results presented above, we conclude that Support Vector Regression is an
appropriate method for predicting demand in hospitals, but without discarding the possibility of
using simpler methods that may provide acceptable results under certain conditions. Since all the
aforementioned models generate forecasts in less than one minute on a standard PC, the run
time is irrelevant for the proposed monthly use f the algorithm.

4. CAPACITY MANAGEMENT BASED ON THE

FORECAST
As stated in (Jones et al, 2002), having a forecast is not, by itself, a useful contribution to
hospital management. Managers also need to know whether their capacity will suffice to attend
such demand with defined quality standards, and how the hospital resources could be re-
arranged or modified to achieve that goal. The linkage between forecasting and resources
provides managers a quantitative basis for hospital capacity planning and management.

As mentioned earlier, medical hours is one of the most scarce and expensive resources
in Chilean public hospitals. To estimate the number of medical hours required to attend the
expected demand in the emergency service, the following distributions are considered:

 Demand categorization per month, based on illness severity.

 Demand arrival per hour.

For the purpose of this paper, the capacity management analysis, results, and
recommendations presented in this section will focus on the emergency service of the Luis Calvo
Mackenna Hospital.
 17

4.1. Demand Categorization

Following health care conventions used in Chilean hospitals, patients are classified into 4
categories according to the severity of their illness, as shown in Table 3. Each category is
associated to different uses of the medical resource, as will be explained below.

Category Description
C1 Dying Patient
C2 High Risk Patient
C3 Low Risk Patient
C4 No risk Patient

Table 3: Category Types and Description

As shown in Figure 15, the illness severity distribution varies across the different months of
the year. Nevertheless, the severity distribution per month remains relatively stable over the
years. Therefore, in the following calculations each month will be considered to have a
deterministic distribution of patients for each category.
 18

60%

50%
Patients per Category (%)

40%

30%
C1
C2

20% C3
C4

10%

0%

Month

Figure 15: Monthly Categorization Distribution

Given the emergency patients forecast and the illness severity distribution, the expected
number of patients per category can be calculated. In order to determine the number of doctors
required to attend such demand, the next step is to characterize the behavior of the attention time
for each category. For this purpose, a representative sample of C1, C2, C3 and C4 individuals
was used.

Each C1 patient is referred to the reanimation room for resuscitation, upon its arrival to the
emergency service. When this occurs, and depending on the complexity of the surgery or
diagnosis, between one and three the doctors currently working in the attention cubicles leave
immediately their activities to focus on the dying patient. After the medical attention, the time
required to stabilize and treat the C1 patient is registered in a logbook, along with the names of
the doctors that performed the medical procedure. Using this information, we run a Kolmogorov-
Smirnov test to determine the distribution of the C1 patients´ attention time. We concluded that it
follows a log normal distribution with a mean of 108 minutes and a standard deviation of 121
minutes. We also noticed that the number of doctors required to attend these patients has a
 19

distribution highly concentrated in two doctors, therefore using this value for the following
calculations.

When trying to characterize the consults of C2 patients, the data used did not provide
enough information to determine a distribution of the attention time. However, in a discussion with
doctors a consensus was reached in that the average time to attend C2 patients is 60 minutes,
with a standard deviation of 20 minutes. A normal distribution was chosen to represent the
behavior of this attention time. Finally, and with a high level of confidence, we determined the
distribution of the attention time for C3 and C4 patients as lognormal with means of 10 and 7
minutes, and standard deviations of 7 and 3 minutes, respectively. The non-dying patients are
attended by only one doctor.

The time distributions presented above provide a stochastic basis to estimate the time
that doctors will spend to attend the patients from each category, expected to arrive to the
emergency room. A summary of the attention time distributions found for the different severity
categories is presented in Table 4. The number of doctors required to attend each patient per
category is presented in Table 5.

Category Mean (min) Standard

Distribution
deviation (min)
C1 Log Normal 108 121
C2 Normal 60 20
C3 Log Normal 10 7
C4 Log Normal 7 3

Table 4: Distribution of Attention Time per Category

Category Doctors Req.

C1 2
C2 1
C3 1
C4 1

Table 5: Number of Doctors per Category

4.2. Demand Arrival Distribution per Hour

Patients’ arrivals were analyzed at different times of the day, as shown in Figure 16. This
study concluded that 59% of the patients arrived at the emergency service from 12:00 until 20:00.
Using a representative sample we found that this distribution does not vary significantly among
different days of the week nor among the same days of different weeks. Therefore, this
distribution is considered as fixed for every day of the year.
 20

8%

Patients Arrival Frequency (%)

7%
6%
5%
4%
3%
2%
1%
0%
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time (Hours)

Figure 16: Patients Arrival Distribution per Hour

4.3. Balance of Resources

Once the demand and its behavior have been characterized, the next step was to
determine whether the existing resources would suffice to meet the forecasted demand. The unit
selected to perform this comparison is medical hours per month.

On the supply side, the total available medical hours per month are determined by the
simple multiplication of the number of doctors available by the number and length of shifts per
doctor in that period of time. In the case of HLCM, the number of doctors does not vary inter- or
intra- shifts. On the demand side, a simple method to calculate the medical hours required to
meet the forecasted demand is by considering a deterministic monthly behavior obtained from the
model, divided uniformly within each month and distributed within each day as presented in
Figure 19. As explained above, the severity distribution of these patients can be considered as
deterministic for each month. With these considerations regarding the expected demand, the
forecasted number of patients per category can be calculated multiplying the total number of
forecasted patients by the proportion of patients per category.

The final step is to convert the demand per category into medical hours, distributed along
each hour of the day. To get a quick idea of the medical hours required to meet such demand, we
multiplied the forecasted number of patients per category by the mean of the corresponding
attention time distributions presented in Table 4. Table 6 illustrates the expected behavior of
demand during the day and the availability and rate of use of the medical resources.

Arrival Time Available Resources Required Resources Medical Hours Available Occupation
[Medical Hours/ [Medical Hours/ [Avail. Resources – Req. Rate (%)
Month] Month] Resources]
0:00 - 0:59 90 41 49 46%
1:00 - 1:59 90 24 66 27%
2:00 - 2:59 90 15 75 17%
 21

3:00 - 3:59 90 11 79 12%

4:00 - 4:59 90 6 84 7%
5:00 - 5:59 90 5 85 6%
6:00 - 6:59 90 10 80 11%
7:00 - 7:59 90 15 75 17%
8:00 - 8:59 90 42 48 47%
9:00 - 9:59 90 87 3 97%
10:00 - 10:59 90 114 -24 127%
11:00 - 11:59 90 136 -46 151%
12:00 - 12:59 90 138 -48 153%
13:00 - 13:59 90 129 -39 143%
14:00 - 14:59 90 117 -27 130%
15:00 - 15:59 90 123 -33 137%
16:00 - 16:59 90 125 -35 139%
17:00 - 17:59 90 122 -32 136%
18:00 - 18:59 90 110 -20 122%
19:00 - 19:59 90 102 -12 113%
20:00 - 20:59 90 101 -11 112%
21:00 - 21:59 90 95 -5 106%
22:00 - 22:59 90 81 9 90%
23:00 - 23:59 90 58 32 64%

Table 6: Forecasted Medical Resources Occupation Rate

With these simple calculations, several interesting observations arise regarding the medical
resources in the emergency service. For example, the period from 0:00 until 8:00 shows a high
rate of idle resources, while between 10:00 and 21:00 the medical resources are overly utilized.
From the management point of view, for instance, some doctors from the night shift could be
reassigned to work during peak hours, but always remaining prepared to meet a potential
emergency by having one doctor on duty at home during the night.

4.4. Simulation for Capacity Management in Hospital

Facilities
With the calculations from Section 4.3, we were able to determine roughly the
performance of the current distribution of resources when attending to the forecasted demand.
But the previously mentioned analysis is static and deterministic. In order to provide a more
dynamic and accurate perspective of the emergency service performance throughout the
attention provided to the patients, we developed a simulation model that incorporates the
stochastic behavior of the demand. Since the waiting time and length of lines have shown to be
significantly higher for medical attention, the simulation will be performed for these patients only.

The forecast detailed in Section 3 has an error with a normal distribution. To simulate the
different demand scenarios for each month, the forecast was adjusted several times by different
 22

values sampled from the normal distribution of the error. Due to the stability of its daily behavior,
the demand of each scenario was distributed uniformly across every day of the month. The daily
demand was further disaggregated into hourly demand using the distribution shown in Figure 16.
As a consequence, we were able to generate several scenarios of monthly demand
disaggregated per hour. Using the hourly forecasted demand from each of the scenarios
generated as described above, the average forecasted demand was calculated for each hour of
the day. We assumed that the hourly demand arrives according to a Poisson process; then this
average corresponds to the mean of the Poisson distribution per hour.

Upon their arrival at the emergency service, patients are categorized and served
according to the time distributions presented in Section 4.1, which are also stochastic. Now that
the stochastic behavior of the demand and the medical attention has been incorporated into the
problem, we will discuss the construction of the simulation model and its role in the management
of hospital capacity.

In capacity management two types of problems are considered: configuration

management, for which we want to determine how different designs of the hospital facilities may
affect the quality of service, measured in length of wait before the first medical attention (LOW);
and resource management, which decides how current available resources should be assigned to
increase the service level, and which and where new resources are required to further improve
the quality of service.

The simulation model allows us to observe how the expected flow of patients will use the
different services offered in the facilities of the hospital, and how the available capacity performs
when attending such demand. As a consequence, capacity can be redistributed or adjusted with
the objective of eliminating bottlenecks and reducing idle resources. This provides a powerful
decision tool for managing capacity in such a way that a given service level can be guaranteed at
minimum cost.

a) Configuration Management

In this section, two main designs of the system will be contrasted with the expected
demand: the current configuration and a Fast Track configuration with Triage.

In the current configuration, only dying patients are given priority when arriving at the
emergency service. They are immediately taken to the resuscitation service for stabilization, and
then referred to the operating room or the intensive care unit service. Patients who are not
critically ill must provide their personal data upon their arrival and subsequently wait for medical
attention. Therefore, the relative importance in terms of severity of illness is not considered for
these patients. The admission time is distributed uniformly between 5 and 10 minutes. After
medical attention, the patients can be referred for hospitalization, for diagnostic testing, or
immediate discharge. The hospitalization service does not belong to the emergency service and
hence does not use its medical resources. The time in diagnostic test services is distributed
uniformly between 2 and 4 hours; patients are usually requested to bring the test results back to
the doctor within the same day of the evaluation.

The Fast Track configuration includes a Triage to categorize the patients upon their
arrival at the emergency service, which is performed by a nurse. The reason behind the creation
of the Triage lies in the importance of providing medical attention promptly to patients with the
most urgent needs (C2 and C3). Half of the patients categorized as C4 in the Triage are referred
to the Fast Track attention, which is performed by a doctor who does educational activities with
 23

medical students from 10:00 to 15:00. This allows reducing the average waiting time within these
hours with resources that otherwise would not be used. The rest of the configuration remains as
in the current system.

In both configurations outlined above, the doctors shift structure consists of two 12-hour
shifts (day and night) with 3 doctors attending in each of them. For each configuration, we
developed a simulation model that includes the demand and emergency service characterizations
presented earlier. An example of the models is shown in Figure 17.

Figure 17: Fast Track with Triage Simulation Model

As stated above, the average LOW will be used as the main criterion to compare the
performance of the system designs. This metric was calculated weighing the demand per
category by its respective average LOW. The results for this metric for the Base and Fast Track
simulated configurations are presented in Table 7.

Configuration (3 doctors day and night) Avg. (min) Std. Dev. (min)
Base Case 64.2 1.2
Fast-Track with Triage 57.3 0.9

Table 7: Simulated LOW of Different Emergency Service Configurations

 24

Based on the scenarios run in the simulation, a 95% confidence interval was generated
for the LOW of each configuration. The intervals obtained were (55.5, 59.1) and (61.8, 66.6) for
the Base Case and Fast Track with Triage, respectively. To test if the LOW differs significantly
between these two configurations we applied the procedure proposed by Law and Kelton (1982).
This comparison is established based on the difference of their respective statistical distributions,
as displayed in Table 8. Since the confidence interval does not contain the value 0, we confirm
that the difference shown in Table 7 is statistically significant.

Lower Bound Upper Bound

Comparing Configurations Avg. (min) Std. Dev. (min)
95% (min) 95% (min)
Base Case / Fast Track with
6.9 1.5 3.9 9.9
Triage

Table 8: Base Case / Fast Track Configurations Comparison

Using the results presented in Table 7 and 8, we can observe that:

 The simulation model resembles the actual behavior of the system, since current
average LOW is within the confidence interval of the simulated Base case.
 The main bottleneck occurs in the medical consults and during the day-shift, as
addressed in section 4.3.
 The Fast Track Box with Triage reduces the average LOW in 6.9 minutes, which
corresponds to a 10.8% reduction of the current average waiting time.

b) Resource Management

After deciding which configuration performs better on the emergency service, the next
question to be addressed is related to the impact that a redistribution, reduction, or addition of
medical resources would generate on the performance of the system. The resource management
analysis, then, will be performed for the Fast Track configuration only.

The first issue we noticed was that the current shift structure, including the number of
doctors per shift, was not constructed based on the daily behavior of the demand, presented in
Table 5 and previously in this section. The simulation model was then run for several
assignments and number of doctors per shift under the Fast Track configuration with Triage and
assuming a stochastic demand.

If the current structure of two 12-hours shifts is maintained, an initial scenario would
consider only redistributing the six doctors available in a different manner. Given the greater
arrival of patients during the day, as observed in Table 6, a possible redistribution could include
the reassignment of a doctor from the night to the day shift. As a consequence, four doctors
would attend during the day shift and only two at night. The average LOW of this scenario would
be 45.1 minutes.
 25

Further resource management considerations may determine the addition or reduction of

medical hours for attending the patients. Since these resources are known to be quite expensive,
the different scenarios were simulated by changing the existing capacity in 0.5 doctor intervals.
The extra half doctor was included through the creation of a new shift of 6 hours, from 12:00 to
18:00, which is precisely the period in which most patients arrive at the service. Thus, the number
of 6.5 doctors available means that 4 doctors attend on the day shift (8:00 - 20:00), 1 doctor on
the half shift (12:00 – 18:00), and 2 doctors at night (20:00 – 8:00). The average LOW obtained
with this configuration is 40.5 minutes.

The simulation was run using from 5 to 7 doctors within 24 hours, and distributed as
explained above. The idea behind analyzing the reduction of the current number of doctors is to
assess whether the performance of the system is affected in a significant manner when these
resources are lacking, either by management decision or by absenteeism. The average LOW for
different numbers and assignments of resources are summarized in Figure 18, including the 95%
confidence interval for each of the points.

90
82,8
80

70
LOW (min)

60
46,3 Average LOW
50 45,1
40,5 40,2 Upper Bound
40
Lower Bound
30

20
5 5,5 6 6,5 7
Medical Resources (Number of Doctors)

Figure 18: Average LOW and Confidence Intervals for Different Numbers of Doctors

As expected, the addition of medical resources improves the service quality, measured in
average LOW. The interesting result is that the average LOW decreases dramatically when
reducing the number of medical resources from 5 to 5.5 doctors, while it decreases more
gradually when new resources are added. To test whether the LOW difference between all the
scenarios included in Table 18 is statistically significant we applied again the procedure proposed
by Law and Kelton (1982). Table 9 shows the confidence intervals when comparing the LOW of
these scenarios. As it can be observed, increasing from 5 to 5.5 doctors provides a significant
improvement to the performance of the system, while the change between 5.5 and 6 doctors
does not. Nevertheless, increasing from 5.5 to 6.5 doctors does show statistical significance.
 26

Comparing Scenarios 5.5 6 6.5 7

5 [30.4 ; 42.6] [31.6 ; 43.9] [36.3 ; 48.3] [36.5 ; 48.7]
5.5 - [-0.4 ; 3] [4.5 ; 7.2] [4.4 ; 7.8]
6 - [3.3 ; 5.9] [3.1 ; 6.5]
6.5 - [-1.1 ; 1.6]

Table 9: Confidence Intervals for Compared Scenarios

The previous analysis of the system performance provides hospital managers a decision
tool for determining the number and distribution of medical resources on the emergency service,
based on a cost/benefit analysis of resources and service improvement.

5. IMPACT OF THE WORK AND CONCLUSIONS

We have shown that the demand for hospital emergency services can be forecasted with
great confidence. After testing several forecasting methods we found that SVR provides better
results in terms of variance and accuracy. Using this model, we designed a process for managing
hospital capacity based on the comparison between the forecasted demand and the available
medical resources and a simulation model to assess the performance of different configurations
of facilities and resources. Both of these tools assist hospital managers in decision-making
regarding the configuration and distribution of medical resources in the emergency service.

The forecasting method and the capacity management process proposed in this paper have
been validated and accepted by hospital managers and staff, and are currently in use in one of
the hospitals. For this to be possible, we embedded the forecasting model and resources
management logic in support computing systems we developed, which may be used in daily work
practices. We are currently implementing these processes and systems in one of the hospitals
included in this study. The results have been so encouraging that National Health Authorities are
considering to extend the demand forecasting and management practices we have designed to
close to one hundred public hospitals in Chile.

Acknowledgements: The second author is grateful for the support provided by the Complex
Engineering Systems Institute (ICM: P-05-004-F, CONICYT: FBO16) (www.isci.cl).
 27

References

Aburto, L., Weber, R. Improved supply chain management based on hybrid demand forecasts.
Applied Soft Computing 2007; 7; 1; 136-144

Adya M, Collopy F. How effective are neural nets at forecasting and prediction? A review and
Evaluation. Journal of Forecasting 1998; 17; 451-461

Armstrong J S. (Ed.). Principles of forecasting. Kluwer Academic Publishers: Norwell, MA; 2001

Barros O, Julio C. Application of enterprise and process architecture patterns in hospitals.

BPTrends April 2010; www.bptrends.com

Barros O., Julio C., "Enterprise and process architecture patterns", Business Process
Management Journal 2011; 17; 4; 598 – 618

Box G E H Jenkins G M, Reinsel G C. Time Series Analysis, Forecasting and Control, 3rd Ed.
Prentice-Hall: Englewood Cliffs, NJ; 1994

Chang C C, Lin C J. LIBSVM: A Library for Support Vector Machines[EB/OL], 2001. Software
available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.

Chen P H, Lin C. J., and Schölkopf B., A tutorial on ν-support vector machines, Appl. Stoch.
Models. Bus. Ind. 2005, 21, 111-136.

Farmer, R D T, Emami, J. Models for forecasting hospital bed requirements in the acute sector.
Journal of Epidemiology and Community Health 1990; 44; 307-312

García M L, Centeno M A, Rivera C, and DeCario N, Reducing Time in an Emergency Room Via
a Fast-Track, Winter Simulation Conference, 1995, 1048-1053

Hofmann, T, Schölkopf, B, Smola, A. J. Kernel Methods in Machine Learning, Annals of Statistics

2008; 36;1171-1220.

Hwang J, Gao L, Jang W. Joint Demand and Capacity Management in a Restaurant System.
European Journal of Operational Research, in press

Jones A J, Joy M P, Pearson J. Forecasting demand of emergency care. Health Care

Management Science 2002; 5; 297-305

Kapuscinski R, Zhang R Q, Carbonneau P, Moore R, Reeves B. Inventory decisions in Dell’s

supply chain. Interfaces 2004; 34; 191–205

Kennedy M, Macbean C E, Sundararajan V, Taylor D M. Review Article: Leaving the Emergency

Department without being seen. Emergency Medicine Australasia 2008; 20;4; 306-313

Khurma N, Bacioiu G M. Simulation-Based Verification of Lean Improvement for Emergency

Room Process, Winter Simulation Conference 2008, 1490-1499
 28

Law A.M., Kelton W.D. Simulation modeling and analysis. Mc Graw Hill, 2001

Marmor Y N, Wasserkrug S, Zletyn S, Mesika Y, Greenshpan, Carmeli B, Shtub A, Mandelbaum

A. Toward Simulation-Based Real-Time Decision-Suppor Systems for Emergency Departments,
Winter Simulation Conference, 2009, 2042-2053.

McLaughlin D, Hays J M. Healthcare operations management. AUPHA Press: Washington, DC;

2008; 378-381

MIDEPLAN. Encuesta Casen 2006. www.mideplan.cl/casen

Min D, Yih Y. Scheduling Elective Surgery under Uncertainty and Downstream Capacity
Constraints. European Journal of Operational Research, to appear,

Rojas L M, Garavito L A. Analysing the Diana Turbay CAMI emergency and hospitalization
processes using an Arena 10.0 simulation model for optimal human resource distribution, Revista
Ingeniería e Investigación 2008; 28;1; 146-153.

Samaha S, Armel W S, Stark D W. The Use of Simulartion to Reduce the Length of Stay in an
Emergency Department, Winter Simulation Conference, 2003, 1907-1911

Schweigler L M, Desmond J S, McCarthy M L, Bukowski K J, Ionides E L, Younger J G.

Forecasting models of emergency department crowding. Academic Emergency Medicine 2009;
16; 301-308

Shirxia Y, Xiang L, Li N, Shang-dong Y. Optimizing neural network forecast by immune algorithm.

School of Business Administration, North China Electric Power University, Beijing, China; 2007;
http://www.springerlink.com/content/t23961t0x38nvu3h/

Smola A J, Schölkopf B. A tutorial on support vector regression. Statistics and Computing 2004;
14; 3

Vapnik V. The Nature of Statistical Learning Theory, Springer-Verlag, 1995

Vapnik V. Statistical Learning Theory, Wiley, 1998

Vladimir C, Ma Y Q. Practical selection of SVM parameters and noise estimation of SVM

regression. Neural Networks 2004; 17;1; 113 126

Zhang G P. Avoiding pitfalls in neural network research. IEEE Transactions on Systems, Man and
Cybernetics—Part C: Applications and Reviews 2007; 37; 3-13