Forecasting Ambulance Demand using Machine
Learning: A Case Study from Oslo, Norway
Anna Haugsbø Hermansen
Department of Computer Science
NTNU, Trondheim, Norway
Email: annahhermansen@gmail.com
Abstract—In Emergency Medical Services (EMS), time is of the
essence. It is crucial for one or more ambulances to reach the
scene of a medical incident quickly to ensure timely, sometimes
2021 IEEE Symposium Series on Computational Intelligence (SSCI) | 978-1-7281-9048-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/SSCI50451.2021.9659837
life-saving, assistance to people in need. To do so, one needs
good estimates of when and where incidents are likely to occur.
This work investigates how to best forecast EMS demand in and
around Oslo, Norway’s capital. Demand forecasts are computed
using machine learning models created from historical EMS
and weather data. We use a fine spatio-temporal resolution of
1x1km spatial regions and 1-hr time intervals. Two different
approaches are implemented and compared against each other:
a split approach that forecasts the volume and distribution of
the EMS demand separately, and a complete approach that
directly models the EMS demand over our spatial domain. We
use multi-layer perceptron (MLP) and long short-term memory
(LSTM) models to forecast the EMS demand, and compare them
to simple aggregation methods and baselines. We also study
different feature sets and online learning to improve the models’
performance. Experiments suggest that a split online model
consisting of a simple aggregation distribution model and an
MLP volume model with simple temporal features produces the
best forecasts. In particular, the split model produces competitive
volume, distribution, and complete forecasts relative to a state-
of-the-practice method, MEDIC, and a well-established complete
MLP model proposed by Setzler et al.
I. I NTRODUCTION
Emergency medical services (EMS) are a crucial but com-
plex part of modern health care systems. For EMS there is
substantial uncertainty in the volume, severity, and location
of medical incidents, which impacts ambulance demand and
ambulance travel times. Further, the trade-off between cost,
effect, and equity is a substantial issue. Allocating many
resources, including ambulances, is necessary to handle sudden
high workloads. But resources that are never or seldom used
pose a high and unnecessary cost. It is more cost-effective to
focus resources in areas with high demand, but that means
people living in low-demand areas will have less access to
those critical resources. On the other hand, providing equal
access to resources when the cost of doing so is higher in
some areas may imply that people are valued higher in those
areas [4].
In light of these complexities, it is not surprising that studies
of EMS demand has received and is receiving much attention
from the general public, policymakers, medical professionals,
and researchers. To structure the discussion in this paper, we
group previous studies of EMS demand according to their
978-1-7281-9048-8/21/$31.00 ©2021 IEEE
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
Ole Jakob Mengshoel
Department of Computer Science
NTNU, Trondheim, Norway
Email: ole.j.mengshoel@ntnu.no
spatial and temporal granularity. Daily forecasts are considered
a low temporal resolution, while a time interval of one or a few
hours is considered high resolution. For the spatial resolution,
studies using regions of up to a few square kilometers are
considered high resolution, while coarser resolution (including
those studying entire cities) are low resolution.
Previous works using low spatial and temporal resolution
[15], [16], [13], [7], [14] have considered Hong Kong (Hong
Kong), New Taipei City (Taiwan), Birmingham (UK), and
Ningbo (China). These studies have established (i) the fea-
sibility of ambulance demand forecasting at this spatial and
temporal scale and (ii) the impact of extreme weather on EMS
demand [15], [16].
Research using a low spatial but high temporal resolution
typically forecast the hourly demand for large cities such as
Toronto (Canada) and Calgary (Canada) [9], [1]. This research
demonstrates, for EMS forecasting, the feasibility and relative
performance of different forecasting methods. Further, it shows
how to improve hourly forecasts via online learning, in which
the hourly forecast for the later hours of a day is updated in
light of correct data from the early parts of the day.
Using a high spatial but low temporal resolution, researchers
have successfully studied forecasting of daily demand in
small spatial regions of a few square kilometers or on a
continuous scale in Singapore [8] and Athens (Greece) [5]. An
interesting finding from the Athens study is the reduction of
distance between ambulances and incidents by over 1km when
distributing the units according to a novel model using their
forecast demand as input, compared to the current location of
ambulances in Athens [5].
Using both high spatial and temporal resolution typically
means performing hourly forecasts in either small spatial
regions of a few square kilometers, often inside large cities,
or on a continuous spatial domain [11], [2], [3], [17], [18],
[20], [19]. Such studies have been performed with success
in cities like Charlotte-Mecklenburg (USA), New Taipei City
(Taiwan), Toronto (Canada), and Melbourne (Australia). Using
such fine-grained spatio-temporal scales is one of the main
considerations of this work, thus we provide a more detailed
analysis in Section III. However, we wish to emphasize Zhou
et al.’s study on decomposing or splitting ambulance demand
into components [17], [18], [20], [19] as well as Setzler et
al.’s [10] use of neural networks.
 A. Contributions and Perspectives
In this work, we consider the above complexities, uncer-
tainties, and trade-offs in the context of a real-world study in
and around Oslo, Norway. Building on previous research, we
develop machine learning models for forecasting EMS demand
in Oslo and Akershus, Norway. Doing so involves several
challenges. In addition to the challenges mentioned above,
demand forecasts are only useful for staffing and positioning
resources if the forecasts have high resolutions on a spatio-
temporal scale. However, high resolutions result in very sparse
data which is difficult to model accurately.
The main contributions of our work1 are as follows:
• Using machine learning methods, we study ambulance
demand forecasting using datasets from the Oslo region
in Norway. To our knowledge, such a study has not been
performed before.
• We emphasize the use of a split approach, inspired
by Zhou [17] et al., which splits or decomposes the
forecasting problem into a volume forecasting problem
and a spatial distribution forecasting problem.
• We use online learning to enhance our proposed models.
• Based on experiments with the Oslo datasets, where we
compare several split approaches to baselines and bench-
marks from the literature, we recommend a split online
model consisting of a simple aggregation distribution
model and an MLP volume model using simple temporal Figure 1: Illustration of the total number of incidents per grid
features. cell in the filtered incident dataset from Oslo. The demand
Inspired by the work of Zhou [17], we study how her exhibits extreme variation and locality, with many cells having
split approach performs compared to the more established zero incidents and some cells having between 10,000 and
complete approaches. We also build on Setzler et al.’s [10] 100,000 (10k and 100k) incidents.
use of neural networks to model EMS demand and search
for the best possible network architecture. In addition, we More specifically, this is the overall goal of the health-care
go beyond Setzler et al. by testing LSTM models, which to effort of which this work is a part:
our knowledge have not been tested in this domain before.
Goal To minimize the EMS response time in Oslo2 through
We use MEDIC and existing neural network models [10]
ambulance demand forecasting and strategic place-
as benchmarks. Channouf et al. [1] show that the use of
ment of ambulances.
special day effect flags for New Year’s Eve and a local festival
As a step towards achieving this goal, one needs to forecast
improves demand volume forecasts. We try to capture such
when and where incidents are likely to occur. Consequently,
special days implicitly by adding day of the month and month
we study data-driven and machine learning methods for fore-
of the year to our feature set. We also use weather features as
casting of ambulance demand. The datasets used, research
suggested by Setzler [10] and Zhou [17]. Our study extends
questions, and metrics are discussed below in this section.
the work by Wong and Lai [15], as we use a higher spatio-
temporal resolution. Unlike much previous work on EMS A. Dataset
forecasting, we use online learning (as suggested by Chen et The EMS incident dataset used in this work was provided
al. [2], [3]) and compare it to traditional batch learning. by the EMCC department of the Oslo University Hospital
II. BACKGROUND (OUH) and the Norwegian National Advisory Unit for Prehos-
pital Emergency Medicine (NAKOS). It includes the location
In a medical emergency, a minimal ambulance response time
and timestamps of missions completed by the ambulance
will benefit the medical outcome for the patient or patients. At
department between January 1st, 2015, and February 11th,
the same time, we want to optimize the utilization of available
2019. Due to privacy concerns, we use anonymized data in
resources so that we can reduce the response time without
which the exact incident locations have been mapped to a
incurring large expenses.
standard 1x1km grid, see Figure 1.3
Ultimately, our goal is to minimize ambulance travel time by
positioning ambulances strategically and according to demand. 2 For
simplicity, we often say “Oslo” instead of “Oslo and Akershus.”
3 More information about the grid is available [12]. The grid can be
1 This paper summarizes results of Hermansen’s master thesis [6] downloaded from Statistics Norway at www.ssb.no/natur-og-miljo/geodata.

Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
 Topography, climate, population, and traffic are factors that
may impact ambulance demand, and these vary from city to
city. Oslo—being “wrapped around” a fjord, with a humid
continental climate and relatively low population density—
clearly differs from many previously studied cities. Zhou [17],
for example, studies ambulance demands in Toronto (Canada)
and Melbourne (Australia). Toronto, Melbourne, and Oslo
have population densities of approximately 5 000 persons/km2 ,
500 persons/km2 , and 200 persons/km2 , respectively. Even
within the same city, for example New Taipei City, it has
been observed that different ML models perform best within
different parts of the city [2], perhaps due to differences in
population density.
B. Research Questions (RQs)
We have defined specific research questions (RQs) that we
study in this work, as discussed below.
1) Research Question 1 (RQ 1): We want to forecast the
demand λ(t) = ut ∈ RM t
+ such that uk is the forecast number
of incidents at time step t in region k for k ∈ {1, 2, ..., M },
where M is the number of spatial regions. The higher the
spatio-temporal resolution of the EMS demand forecasts, i.e.,
the larger M is for a given geographical area and the shorter
each time step is, the more information we have for positioning
ambulances strategically. The disadvantage of using a high
spatio-temporal resolution is that the data becomes more
sparse and stochastic, making forecasting more difficult. This
leads us to our first research question:
RQ 1 How can EMS demand in Oslo be forecast accurately
at a high spatio-temporal resolution?
In this work, we use a high spatial resolution of 1x1km
regions, which is the highest possible granularity in our case
due to the privacy restrictions of our dataset. Our temporal
resolution of 1 hour is the highest temporal granularity used
in the literature, making state-of-the-art comparisons easier.
We test various methods, models, and features to investigate
how to model the EMS demand as accurately as possible at
these spatio-temporal resolutions.
2) Research Question 2 (RQ 2): Existing studies on EMS
demand forecasting have, with a few exceptions, tried to
forecast the complete demand λ(t) directly. However, it is
also possible to use a split approach that forecasts the volume
and the distribution of the demand separately. Let δ(t) ∈ R+
be the aggregated
PM volume
t
of incidents at time step t such
that δ(t) = u
k=1 k . Let f (t) = vt ∈ RM + be the spatial
distribution of the incidents at time step t such that vkt is the
fraction
PMof the events in time step t that occurs in region k
t
and k=1 kv = 1. A complete forecast at time step t can
be obtained by combining volume and distribution forecasts:
λ(t) = δ(t)f (t).
Almost all previous research has been on complete models,
with the notable exception of Zhou and co-authors [18];
see Section III. We are interested in how different methods
influence the forecasting of EMS demand:
RQ 2 Is a split model or a complete model better at
forecasting EMS demand in Oslo?
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
Here, the definition of “better” is not straightforward be-
cause we consider several dimensions. Which is better, a
model with good complete forecasts or a model with good
volume or distribution forecasts? We believe that a model that
makes good distribution and volume forecasts is better because
such separate forecasts have much practical value. First, a good
volume forecast is useful for staffing decisions—it can be used
to determine how many employees should staff a shift. Second,
a good distribution forecast helps determine where resources
should be positioned to ensure low response times. These two
points speak to the practical value of a split approach in which
the focus is on both volume and distribution forecasts. Having
a good complete forecast, on the other hand, is not as easily
interpreted and therefore may have less practical value.
C. Metrics
Error metrics are used to quantitatively evaluate the perfor-
mance of forecasting models.
1) Mean Absolute Error: The mean absolute error (MAE)
avoids the cancellation of negative and positive errors by
taking the absolute value of each error. For forecasts ŷ ∈ Rm
and targets y ∈ Rm , the mean absolute error is defined as:
m
1 X
MAE (ŷ, y) = |ŷi − yi |. (1)
m i=1
2) Mean Squared Error: The mean squared error (MSE) is
a popular metric for regression problems. Similar to MAE, it
avoids cancellation of negative and positive errors, but MSE
emphasizes large errors. For model predictions ŷ ∈ Rm and
targets y ∈ Rm , the MSE is defined as:
m
1 X
MSE (ŷ, y) = (ŷi − yi )2 . (2)
m i=1
MSE is often used as a loss function for neural networks
because of the efficient calculation of its gradient.
3) Categorical Cross-Entropy: Categorical cross-entropy
(CCE) is a measure of the distance between two probability
distributions. It is based on cross entropy Pfrom information
m m
theory. For a model prediction ŷ ∈ R , i=1 ŷi = 1 and a
Pm
target distribution y ∈ Rm , i=1 yi = 1, the CCE is:
m
X
CCE(ŷ, y) = − yi log ŷi . (3)
i=1
For many predictions and targets, we take the average of the
cross-entropies between each prediction-target pair.
III. R ELATED R ESEARCH
Existing research has used datasets of highly varying spatial
and temporal resolutions. Given our goal and research ques-
tions, as well as limited space, we focus mainly on previous
research using hourly forecasts in either small spatial regions
of a few square kilometers or on a continuous spatial domain.
Using such a high spatio-temporal resolution is one of the
main challenges of our work.
 Steins et al. [11] use Zero-Inflated Poisson (ZIP) regression
to model the hourly EMS call volume in 6x6km, 4x4km, and
3x3km spatial regions of three different Swedish counties. The
ZIP regression combines two models: one that generates zeroes
and a Poisson process that generates counts. This makes it
suitable for modeling point processes with excess zero-count
data, characteristic of EMS demand. Steins et al. [11] create
one model for each county, using socioeconomic, geographic,
and temporal features. The ZIP model performs slightly better
than the existing model, which calculates the total average
demand in the county per hour of the week using data from
the previous year and then assigning a fraction of this demand
to each grid based on their day and night population.
Setzler et al. [10] propose a multi-layer perceptron (MLP)
to forecast EMS call volume in Charlotte-Mecklenburg, North
Carolina (USA) for a variety of grid cell sizes and time in-
tervals. Their 2002–2004 EMS data were mapped to different
combinations of grid cell sizes (2x2mile and 4x4mile) and time
intervals (1hr and 3hrs); we focus on the 1hr case below. The
MLPs use four temporal features (hour, day of week, month,
and season), a hidden layer with four nodes, and one output
value for each grid cell. Setzler et al. [10] compare their MLP
model to the MEDIC forecasting method used by the EMS
agency in Charlotte-Mecklenburg.
MEDIC makes, in the 1hr case, forecasts by averaging the
call volumes at the same hour for the previous four weeks over
the past five years. Let Ah,d,w,y be the actual call number for
a given hour h, day d, week w, and year y. Let Y = 5 be
the number of years and W = 4 be the number of weeks to
average over. Then the MEDIC forecast Fh,d,w,y is:
PY −1 PW
i=0 j=1 Ah,d,w−j,y−i
Fh,d,w,y = . (4)
YW
In practice, one MEDIC model is created for each spatial re-
gion. Setzler et al. find that their new MLP model outperforms,
according to MSE, MEDIC for 4x4mile grids (both 1hr and
3hr time intervals). For the 2x2mile grid with 3hr intervals,
there is no statistical difference between the two methods.
Interestingly, MEDIC was better for the 2x2mile grid with 1hr
time intervals. Upon further inspection, the authors found that
forecasting only zeros outperformed the two other approaches
for this configuration.
Chen et al. [2] study daily EMS demand in 3x3km spatial
regions in New Taipei City, Taiwan, using 2010–2012 data.
Their forecasts use moving average, artificial neural network
(ANN), linear regression, and support vector machine (SVM)
models [2]. They train one model per spatial region for each
model type. The SVM and ANN models are trained with
month, day of the week, season, hour, and year features.
Models are retrained with more data after each forecast is
made on the test set. Chen et al. find that the ANN model is
best overall but that the SVM is better for regions with high
EMS demand. Thus, they propose a model selection phase in
which the best of the four proposed models is selected for
each region.
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
In a follow-up study, Chen et al. [3] adopt a two-phase
model-selection approach to make 3hr forecasts in 3x3km
spatial regions of three districts in New Taipei City. First, they
use cross-validation to select the best features and potential
hyper-parameters for the four different models. Second, the
best of the four proposed models is selected for each region.
The authors test seven different feature sets for the SVM and
ANN models. Several different models with different feature
sets are chosen for the regions. For regions with the highest
EMS demand, the ANN model is chosen in 4 out of 7 regions.
In low-demand regions, linear regression is chosen in 16 out
of 27 regions. The different feature sets are chosen one to
three times each. Overall, Chen et al. [3] conclude in 2016
that this two-fold model selection approach produces better
results than their 2014 single-model approach [2].
Zhou, in her PhD thesis [17] and related publications
[20], [18], [19], [17], proposes several novel approaches
for forecasting ambulance demand on a discrete time and
continuous space domain. She and her co-authors assume
that the ambulance demand in each 1- or 2-hr time period
independently follows a non-homogeneous Poisson process
whose expected value can be represented as a positive intensity
function λt (s), where s ∈ S ⊆ R2 is a spatial location in the
relevant geographic area S. The intensity functionRis decom-
posed into an aggregate demand intensity R δt = s λt (s) ds
and spatial density ft (s) such that S ft (s) ds = 1. Then,
λt (s) = δt ft (s). The aggregate demand intensity δt is simply
the total demand of the entire geographic area of interest in
time interval t, often studied in the literature. In contrast, Zhou
[17] focuses on forecasting the spatial density ft (s). We call
this approach of forecasting the aggregated demand and its
spatial density separately a “split” approach.
Zhou et al. [17] discuss case studies from Toronto, Canada
(see [18] and [20]), and Melbourne, Australia (see [19]), with
data from 2007-2008 and 2011-2012, respectively. Three main
models for spatial density ft (s) are studied:
1) A Gaussian Mixture Model (GMM) with weekly season-
ality constraints and conditional autoregressive (CAR)
priors to capture daily seasonality.
2) A spatio-temporal kernel density estimation (stKDE)
approach.
3) An extended stKDE with kernel warping.
These methods are compared against each other and against
MEDIC [10] using the Average Logarithmic Score (ALS).
The stKDEs models produce equally accurate forecasts as
the GMM models while having considerably lower time
complexity. The warped stKDE model outperforms the other
models for Melbourne, as the warping adapts the stKDE to
Melbourne’s complex geographical boundaries.
IV. D EMAND F ORECASTING M ETHODS AND M ODELS
A. Categories of Forecasting Models
Our demand forecasting models for Oslo are split into three
main categories: distribution, volume, and complete models.
The volume models forecast aggregate hourly demand, while

the distribution models forecast how this hourly demand is
distributed spatially across regions. Volume and distribution
forecasts can be combined into a complete model that forecasts
the number of incidents per region. The complete models make
such complete forecasts directly. We evaluate the volume and
complete models using MSE and MAE, while the distribution
models are evaluated using CCE.
The complete models have M = 5569 non-negative output
values—one per grid cell. The target values for the complete
models are simply the number of incidents in each of the M
regions at each time step t; yt ∈ RM .
The volume models have a single non-negative output. The
target values for the volume models are the sum of the number
of events across all grid cells in a time step t; yt vol = δt =
P M vol
i=1 (yt )i , with yt ∈ R.
The distribution models have M = 5569 output values—
one per region. The output values must be positive and sum to
one. The distribution target values are calculated by normaliz-
ing the number of incidents across a time step; yt dist = δ1t yt .
If there are no incidents (δt = 0) in a time step, an even
1 1 1 For each of these 3 · 2 · 4 = 24 models we cross-validate
distribution is used as the target; yt dist = [ M ,M , ..., M ].
B. A Small Numerical Example
As a simple example, consider M = 5 regions with
the following number of incidents at time step t: yt =
[1, 2, 0, 0, 1]. Then, the target distribution will be: yt dist =
[0.25, 0.5, 0, 0, 0.25], the target volume will be yt vol = 4, and
the complete target will be yt comp = [1, 2, 0, 0, 1].
Next, suppose that for a different time step t + k there are
no incidents: yt+k = [0, 0, 0, 0, 0]. Then the target distribution
is yt+k dist = [0.2, 0.2, 0.2, 0.2, 0.2], the target volume is
yt+k vol = 0, and the complete target will be yt+1 comp =
[0, 0, 0, 0, 0]. Note that ∀t (yt vol · yt dist = yt comp = yt ).
C. Offline Forecasting Methods
We study several offline models for forecasting the hourly
EMS demand in Oslo and Akershus. These include several
neural network models and some simple baseline models based
on averages of the historical EMS demand; see Section IV-C1
and Section IV-C2. We also study benchmark models from the
literature [10]; see Section V-D.
1) Neural Networks: We consider neural networks because
of their ability to model complex and non-linear patterns in
data. We propose both MLP and LSTM neural networks to
forecast EMS demand. The LSTM model can leverage the
sequential nature of the problem and may be able to pick up
patterns over time.
The neural networks are trained and tested with four differ-
ent sets of features (inputs), see Table I. The basic feature set
includes the hour of the day, day of the week, and month.
These features are all one-hot encoded. The other feature
sets are extensions of this basic set. The weather feature
set includes rainfall and temperature measurements. For each
hourly forecast, the entire day’s worth of weather data is
included. The weather data is collected at 3-hour intervals,
meaning each day has eight measurements of each weather
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
Feature Set Id Features Nodes
Basic b Hour, day of week, month 43
Hour, day of week, month,
Weather w 59
precipitation, temperature
Date d Hour, day of week, month, day 74
Hour, day of week, month, day,
Weather and date w&d 90
precipitation, temperature
Table I: Overview of the four different feature sets. The
identifier of the feature set is used in the result tables. The
nodes column states the number of feature nodes each feature
set has.
type. Section V-E2 details how the weather data is collected.
The date feature set includes the day of the month. The day of
the month is one-hot encoded with 31 classes. We hope that
the inclusion of the day of the month will allow the neural
network to leverage special day effects. Finally, we consider
a combined weather and date feature set.
For each of our three categories, we propose two different
model types (MLP and LSTM) and four different feature sets.
2 · 7 = 14 configurations in the architecture validation phase:
2 configurations for the number of hidden layers (two or three
layers) and 7 configurations for the number of nodes in each
hidden layer (2, 4, 8, 16, 32, 64, or 128 nodes). This makes
a total of 24 · 14 = 336 different architectures.
The distribution models are optimized using CCE loss and
use the softmax activation function in the output layer to
ensure that the sum of the outputs is one. Meanwhile, the
volume and complete models are optimized using the MSE
loss and have no activation function in the output layer.
2) Baselines: Our baselines are simple but specialized
volume and distribution models that capture basic patterns in
the data. They compute average values for disjunct groups
induced by time or space. The baselines make forecasts by
computing the average value of the relevant temporal (for
volume forecasts) or spatial (for distribution forecasts) group,
as detailed in Section V-C
D. Online Forecasting Methods
We create online versions of our proposed offline forecast-
ing methods to make the models dynamic and get the most
out of the available data. We use a hybrid approach for the
online models by first training them offline on the training set
and then continue with online learning on the validation/test4
set. Our online models extend their offline counterparts: we
perform online learning after initial offline learning.
1) Neural Networks: The pseudo-code in Figure 2 details
how we perform online learning for our neural network
models. We start off with a trained offline version of a model.
Then, we lower the learning rate of the optimizer (from 0.001
to 0.0001) to mitigate the catastrophic forgetting problem by
avoiding putting too much weight on new samples. We make
4 We use the terminology “validation/test” since the validation dataset is
used for online learning in the validation phase while the test dataset is used
in the test phase, as detailed in Section V-B
 1: procedure F ORECAST O NLINE(model, inputs, targets)
2: all forecasts ← () . Empty sequence, initially
3: ` ← length(inputs)
4: for i = 0 to ` step 24 do . Use 24 inputs at a time
5: x ← inputs[i : i + 24]
6: y ← targets[i : i + 24]
7: z ← model.predict(x)
8: all forecasts.append all(z)
9: model.train(x, y) . Train on new samples
10: end for
11: return all forecasts
12: end procedure
Figure 2: Online forecasting algorithm.
forecasts for each hour of the first day of the validation or test
set and store those forecasts in a vector as seen in lines 7 and
8 in Figure 2. Next, in line 9, we train the model one epoch
on those 24 samples. We continue forecasting and training
one day at a time until we have made forecasts for the entire
validation set. These forecasts are then returned and are used
to compute the error of the model.
Because the LSTM volume and distribution models are
trained on normalized outputs, we have to normalize the
validation/test targets before feeding them into the online
forecasting function. We also have to inverse the normalization
on the returned forecasts to transform the raw forecasts back
to the original scale.
V. E XPERIMENTAL R ESULTS
A. Computing Platforms
The models were implemented in Python 3.6.7 with the
Keras 2.4.3, Pandas 1.1.2, GeoPandas 0.8.1, and Numpy 1.19.2
libraries. Default values have been used unless otherwise
specified. We use Keras to implement ANN architectures to
forecast the EMS demand. We use the RMSProp optimizer for
all of the ANNs and the ReLU activation function in the hidden
layers of the MLPs. For the LSTM volume and complete
models, we normalize the output values of the training set
using Keras’ MinMaxScaler to ensure efficient training.
B. Experimental Protocol: Data, Models, and Phases
The forecasting models are trained, validated, and tested in
different phases with different datasets. Figure 3 shows how
the EMS incident time series is split into disjunct training,
validation, and test dataset. The training dataset is the largest,
with just over 2.5 years of data (January 1st, 2015 – July 1st,
2017). The validation dataset is the smallest, with just over
0.5 years of data (July 2nd, 2017 – February 10th, 2018). The
test dataset consists of 1 year’s worth of data (February 11th,
2018 – February 11th, 2019).
Common to all three phases is the use of early stopping
with patience 5 in the training of the neural network models.
The maximum number of iterations is set to 100, but this limit
is never reached.
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
Figure 3: Illustration of how the full dataset is split into
training, validation and test datasets.
First, in the architecture selection phase, we use cross-
validation on the training set to select the best architecture for
each of our proposed neural networks. We use 5-fold cross-
validation for the MLP models and hold-out cross-validation
for the LSTM models, as they are sensitive to the sequential
order of the time series. We test 14 different architectures
for each of 24 models (see Section IV-C) and pick the best
architecture for each model.
Next, in the validation phase, we train all of our proposed
models (8 neural networks in each of 3 categories + 3
distribution baseline + 1 volume baseline, making a total of 28
models) on the training set and evaluate them on the validation
set, with and without online training. Because neural networks
have random initialization, we create 5 instances of each neural
network model and train them using hold-out cross-validation.
The instance with the lowest cross-validation error is selected,
and only this instance is evaluated on the validation set. We
select the two best models within each of the 3 categories to
proceed to the testing phase based on their performance on
the validation set, giving 6 models to test (the bottom 6 rows
in Table II).
Finally, in the testing phase, the 6 models selected in the
validation phase and the 3 benchmark models (MEDIC, MLP,
and All 0s from [10]) are trained anew on the combined
training and validation set and tested on the test set, producing
our final results. Also here, we create 5 instances of each
neural network model and select the best according to hold-
out cross-validation, which is then evaluated on the test set.
C. Baselines
1) Volume Baseline: The volume baseline, α1hr , makes
forecasts based on the average number of incidents in each
hour of the week. This captures the weekly seasonality of the
EMS demand. By grouping the incidents in the training data
by the day of week and hour of the day, α1hr creates a total of
7 · 24 = 168 groups and computes the average for each, which
is later used to make forecasts. For example, α1hr forecasts
8.52 for Mondays (1am, 2am] because there was an average
of 8.52 incidents occurring for that Monday time interval in
the training data.
2) Distribution Baseline 1: The first distribution baseline,
βtotal , makes forecasts based on the spatial distribution of the
incidents in the training set across all time steps. It calculates
this distribution by summing the number of events per grid
cell and then dividing by the total number of events:
tc
1 X
βtotal (tc , th ) = Ptc PM yt . (5)
(yt )i t=1
t=1 i=1

Here, tc is the current time step, th > 0 is the forecast time
horizon, and yt ∈ RM is a vector of the number of incidents
occurring in the M regions at time t. Note that the forecast
βtotal (tc , th ) does not depend on the time horizon th ; βtotal
forecasts the same distribution vector for all time steps.
3) Distribution Baseline 2: The second distribution base-
line, β8hr , creates 21 different distribution predictions: one for
each 8hr bucket of the week. These are created much in the
same way as βtotal , but we group the training set by day of
the week and 8hr bucket before averaging within each group.
Similar to α1hr , it makes forecasts by outputting the average
distribution of the group of the time step to forecast.
4) Distribution Baseline 3: The third distribution baseline,
β1hr , creates a distribution forecast for each hour of the week,
similarly to β8hr but with 168 different values.
D. Benchmarks
We use the MEDIC, All 0s, and MLP models [10] as
forecasting benchmarks. The models are in the complete cate-
gory, but we can derive their implicit volume and distribution
forecasts as shown in the example in Section IV-B.
1) MEDIC: We implement the MEDIC method; see (4).
We have around four years’ worth of data, while the original
MEDIC method uses five years of data. Thus, we adapt the
method by reducing the number of points included in the
model to what is available. For the test set, forecasts before
07.01.2019 are made with 16 data points, while the forecasts
for the next three weeks are made with 17, 18, and 19 points,
respectively. The remaining forecasts use the original MEDIC
method with 20 data points. For simplicity, we subtract 52
weeks to go one year back.
2) All Zeroes: The All 0s model forecasts zeroes for all grid
cells at all times. While being an extremely simple method, it
is often accurate for grid cells with low population density.
3) Setzler: We implement a complete MLP model [10],
hereby denoted as the Setzler model. It has a single hidden
layer with four neurons that use the sigmoid activation func-
tion. The features consist of the one-hot encoded hour (0–
23), day of the week (1–7), month (1–12), and season (0–3),
making a total of 47 binary feature nodes. The output layer has
one node for each spatial region and no activation function.
E. Data Collection and Preprocessing
1) Preprocessing: Before model training, we preprocess the
raw EMS data. This includes removing incorrect, redundant,
and irrelevant data and structuring it as a time series, which
we denote as the filtered incident dataset. The filtered incident
dataset used for learning is sorted by grid cell id and hourly
intervals, starting at 00:00 on January 1st, 2015 up to and
including 23:00 on February 11th, 2019. The result is a
multivariate time series of hourly resolution with M = 5569
variables—one for each spatial region. We denote this time
series as y, such that yt ∈ RM is a vector with the incidents
in the M = 5569 regions at time step t, and (yt )k ∈ R is the
number of incidents in region k at time step t.
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
Method MSE MAE
All 0s 0.0043334 0.0027283
MEDIC 0.0040064 0.0046825
Setzler 0.0038105 0.0049510
dist βtotal + vol MLPb 0.0038172 0.0048571
dist βtotal + vol LSTMw 0.0038179 0.0048555
dist LSTMw + vol LSTMw 0.0037996 0.0048874
dist LSTMw + vol MLPb 0.0037995 0.0048889
complete MLPb 0.0038011 0.0044522
complete MLPw 0.0037983 0.0044777
Table II: Test errors of the complete forecasts of the proposed
and benchmark methods. Note that all models except All 0s,
MEDIC and Setzler use online learning.
2) Weather Data Collection: We collect weather data to
be used as features in some of our models, to investigate
the impact of weather [15], [16], [10], [17]. Weather data
has been collected through the Grid Time Series Data API
provided by The Norwegian Water Resources and Energy
Directorate (NVE).5 The API provides access to time series of
weather data for grid cells. Several time series are available,
such as daily temperature, precipitation, snow depth, and wind
speed. We collected temperature and precipitation data at 3-
hour intervals from 01.01.2015 00:00 up to and including
11.02.2019 21:00. For simplicity, and since weather does not
vary much across Oslo, we collect weather data from a single
grid cell. This is the cell in the city centre of Oslo, with the
most incidents (SSB id 22620006649000). We normalize the
temperature and precipitation data using Kera’s MinMaxScaler
fitted on the data belonging to the training set.
F. Experimental Results
We emphasize two key results from the validation phase.6
First, every single model among the 28 models validated
improved with online learning. Second, varying the feature
sets did not influence the performance of the distribution and
complete models much.
We select the two best distribution, volume, and complete
models from the validation phase to retrain and run on the
test set, using the protocol presented in Section V-B. Next,
the two distribution models are combined with each of the
two volume models, making a total of four split models being
tested. As only online models were selected, we omit making
this explicit in the rest of this section for brevity.
1) Complete Model Results: The final complete test results
can be found in Table II. We see that the complete MLPw
model achieves the lowest MSE, closely followed by the two
split models with the internal LSTMw distribution model.
The All 0s model has the lowest MAE. The Setzler method
achieves a better MSE than the MEDIC method in our case,
in contrast to the original study [10]. In the original study, the
MEDIC method outperformed their proposed ANN method at
5 The API is available at http://api.nve.no/doc/gridtimeseries-data-gts There
is also an interactive application built on top of the API, which can be used
to explore the data more intuitively at http://www.xgeo.no/.
6 Hermansen’s thesis [6] has futher details and results from the validation
and architecture selection phases.
 Method MSE MAE
All 0s 280.6049 15.1939
1) Model Type: The βtotal model made the best distribution
MEDIC 25.2394 3.8183 forecasts, as evident in Table IV. The table also suggests that
Setzler 33.8975 4.3933 the neural network models make fairly good spatial forecasts.
complete MLPb 34.5490 4.4373
complete MLPw 35.2900 4.5135
MEDIC, on the other hand, makes poor distribution forecasts.
volume LSTMw 22.9890 3.7027 This is caused by the fact that MEDIC makes forecasts as if
volume MLPb 22.9108 3.6988 an incident occurring increases the likelihood that an incident
Table III: Test errors of the volume forecasts of the proposed will occur in the same area again at the same time point in later
and benchmark methods. Note that all models except All 0s, weeks. However, the EMS incidents are mostly independent
MEDIC and Setzler use online learning. and stochastic, which means MEDIC often forecasts incidents
that are unlikely to occur.
Method CCE Surprisingly, βtotal , being the simplest distribution model
All 0s (even dist) 8.6249930 tested, makes the most accurate distribution forecasts. Intu-
MEDIC 10.487533
Setzler 6.2471037
itively, we expected the distribution of demand to show some
complete MLPb 6.0354133 pattern in time. For example, there might have been a more
complete MLPw 6.0745206 centralized distribution during working hours, while people
dist βtotal 5.8713098
dist LSTMw 5.8967190
might be drawn toward the fjord on a warm and sunny Sunday.
Seeing how our best distribution model has captured no such
Table IV: Categorical cross-entropy errors of the distribution variations, there must either be too few variations in our data
forecasts on the test set. A lower error is favorable. All models or our models are unable to learn them.
except All 0s, MEDIC and Setzler use online learning. Zhou [17] finds, in contrast to us, significant variations in
the density of ambulance demand with patterns in time and
space for both Toronto and Melbourne, despite having just two
their smallest spatial and temporal levels of granularity (grid- years of data. This might be due to Melbourne and Toronto
size of 2x2 miles ≈ 3.2x3.2 km and 1-hr time buckets). having higher population densities than Oslo and Akershus
2) Volume Model Results: The errors of the volume fore- (see Section II-A). Indeed, Zhou [17] notes that patterns are
casts7 can be found in Table III. The dedicated volume models most prominent in the most densely populated areas of the
produce the best volume forecasts, followed by MEDIC, the cities. Another possible explanation is that Zhou [17] uses a
complete neural networks, and finally, All 0s. Volume forecasts continuous spatial domain, while we use grid-mapped incident
made on the first week of the test set are illustrated in Figure locations. This might make it harder for us to detect subtle
4. We see that the dedicated volume models and the MEDIC differences in the underlying demand density function. Either
models produce better volume forecasts than the complete way, it might be interesting to focus on the more densely
neural network models, which tend to be too conservative. populated areas of Oslo to see if we can detect some patterns.
The complete neural networks do not capture the daily and It might be worthwhile to use different models for different
weekly seasonality well; their volume predictions look very grid cells as Chen et al. [3] do. An LSTM model should be
similar every day and do not capture the increase in incidents tested in such a case, as it seemed like the LSTM models
during working hours or Friday and Saturday nights. were slightly better than the MLP models at forecasting EMS
3) Distribution Model Results: The errors of the distribu- distributions.
tion forecasts are shown in Table IV. We see that the dedicated The online volume MLPb model makes the best volume
distribution models produce the best forecasts, followed by forecasts, see Table III. The volume MLP and LSTM models
the complete neural network models, the All 0s method, and have similar architecture complexity and validation errors.
finally, the MEDIC method. However, since the LSTM model is significantly more complex
than the MLP models without producing better results, we
G. RQ1: How can EMS Demand in Oslo be Forecast Accu-
prefer the MLP model. All of the volume neural network
rately at a High Spatio-Temporal Resolution?
models outperformed the simple α1hr model, which indicates
We have tested a variety of feature sets, model types, and that there are patterns in time other than the weekly seasonality
training regimes for predicting the EMS demand in Oslo at a that the neural networks manage to leverage.
fine spatial and temporal scale of 1x1km spatial regions and 2) Input Features: Figure 4 suggests that the volume MLPb
1-hr time intervals. In this section, we discuss what we have model is able to capture the weekly seasonality. This shows
found to be the best approaches for making accurate forecasts that the basic features are enough to capture the large-scale
of the EMS demand. patterns of the EMS demand. The neural network models also
7 The forecasting approach used here is as follows. Each 24-hour period,
seem to have leveraged some of the annual seasonality of the
at midnight, a new forecast is created. The forecast is for each hour of the EMS demand since they outperformed the α1hr , which naively
next 24-hour period. Further, we train the model on new data after every models the weekly seasonality.
day. In other words, at the end of the day we train on today’s 24 data points The models with the weather feature set perform reasonably
and predict the demand for the next 24 hours. For a time period that spans
multiple days where forecasts are needed, as is the case here, the above 24- well as shown in Tables II, III, and IV. The complete MLPw
hour forecasting method is repeated for each day. model is the best of the complete models while LSTMw is the

Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
 Figure 4: Volume forecasts on the first week of the test set, starting on Sunday the 11th of February 2018. The complete neural
network methods tend to underestimate demand volume, especially during working hours and Friday and Saturday nights.
second-best volume and distribution model. This indicates that
weather features have the potential of improving forecasts.
The Setzler model adds to our basic feature set the season
of the year as a feature. We did not include season as a feature
in any of our feature sets because it is implicitly covered by
the month feature. The Setzler model makes slightly better
volume forecasts but worse distribution and complete forecasts
compared to our proposed complete MLP models. The fact
that the Setzler model makes better volume forecasts than our
complete models indicates that the inclusion of the season
variable might be helpful.
3) Metrics: Clearly, the choice of performance metric in-
fluences the ranking of different methods in Table II, where
the All 0s method achieves the lowest MAE although it has
limited practical value. This happens because our data is very
skewed towards zero, so most of the zero “forecasts” by All
0s are correct. When All 0s makes mistakes, they can be
quite large, however these are not punished harshly by MAE.
The MSE punishes deviations more than MAE, which makes
MSE a better metric in our case as we are most interested
in the deviations, with zero incidents being dominant in most
regions. However, the complete and (to a lesser degree) volume
models tend to underestimate the EMS demand volume in
Figure 4 even though they are optimized for MSE. A different
approach may work better for our purposes. We could, for
example, consider metrics that punish demand underestimation
errors more than demand overestimation errors or probabilistic
forecasts. Such approaches could be argued to be appropriate
for EMS since it is essential to have enough resources to
respond to the demand, if at all possible.
H. RQ2: Is a Split Model or a Complete Model Better at
Forecasting EMS Demand in Oslo?
The complete models achieve lower MSE on their complete
forecasts than the split models, as seen in Table II. However,
the differences are extremely small. On the other hand, the
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
specialized volume and distribution models produce better
volume and distribution forecasts as seen in Tables IV and
III. Overall, the results are perhaps not very surprising, as
the models have been optimized for making either volume,
distribution, or complete forecasts.
However, it raises the question of which approach has the
most practical value, as discussed in Section II-B2.
The complete models are essentially trying to solve a more
difficult problem than the split models by forecasting the
number and location of incidents simultaneously. Intuitively, it
is easier to forecast the occurrence of ten incidents than to fore-
cast the occurrence of ten incidents at certain exact locations.
Because of the stochastic nature of the EMS demand, it is very
hard to make accurate forecasts in both volume and location,
causing the complete models to only forecast incidents that are
very likely to happen. This makes them often underestimate
the number of incidents, as suggested in Figure 4. This could
perhaps partly be remedied by using a different approach as
discussed in Section V-G3, but the nature of the complete
problem may still cause the complete models to produce worse
volume forecasts than split methods.
The distribution forecasts of the specialized distribution
models are also significantly better than those of the other
models, as demonstrated in Table IV.
The MEDIC model, although a complete model, makes
pretty good volume forecasts. It does not suffer from the
underestimation problem of the neural networks because it
simply averages previous incidents instead of trying to mini-
mize its MSE. While this method works well for the demand
volume, it makes poor distribution forecasts.
We conclude that a split model may in our case produce
more useful information for EMS providers than complete
models, because they make better volume and distribution
forecasts.
 VI. C ONCLUSION AND F UTURE W ORK
A. Conclusion
In this work, we propose and test a variety of models
for forecasting the hourly EMS demand in 1x1km spatial
regions in Oslo and Akershus of Norway. This problem is
challenging because of the sparsity of EMS data at such a fine
spatio-temporal resolution. We study two different approaches
for EMS demand forecasting: (i) a complete approach that
directly forecasts the incidents in each of the M spatial region
M [4] E. Erkut, A. Ingolfsson, and G. Erdoğan. Ambulance location for
y∈R PMand (ii) a split approach that1first forecasts the volume
δ = i=1 yi and distribution f = δ y of the demand sepa-
rately. Then, the two forecasts are integrated to get a complete
forecast. In this setting, we consider two different types of
neural networks (MLP and LSTM) with four different feature
sets for forecasting the complete, volume and distribution of
EMS demand. In addition, we study several simple aggregation
methods for forecasting EMS demand volume and distribution.
We conclude that split models, despite being discussed
little in the literature, are better suited for modeling EMS
demand than complete models. Among the different split
models tested, we find that a simple aggregation model is the
best at modeling the spatial distribution of the EMS demand
in Oslo and Akershus. The demand volume is best captured
by an MLP with two hidden layers of 32 nodes each, using
hour, day of the week, and month as features.
We find that all models studied improved with online
learning. We also find indications that weather may influence
the spatial distribution of EMS demand somewhat. In contrast
to previous research [10], [17], [15], we find that the impact
seems relatively small.
B. Future Work
We now outline three areas of future research. First, in the
EMS field, the goal is to always have enough resources to
respond to the demands of the public. While this is hard to
achieve in practice, it is clear that the volume models often
underestimate demand too much, as evident in Figure 4. One
should perhaps consider metrics that penalize underestimates
of demand more than overestimates, see Section V-G3. Al-
ternatively, probabilistic forecasts or forecasts with prediction
intervals instead of point forecasts could be more useful
to EMS providers. Second, the spatial forecasts might be
improved by looking at the most populous regions separately.
There may also be patterns in demand distribution linked to
population movement, for which data could be obtained from
cell phone providers or road sensors. However, the practical
value of leveraging such patterns is uncertain as most of the
regions with high population density are close to each other.
Third, we believe that the forecasting of the EMS demand
volume might be improved by including flags for special days
such as New Year’s Eve. There might be similar spikes in
incidents related to concerts or other special events, which
could impact the spatial distribution of incidents.
Authorized licensed use limited to: ESCUELA POLITECNICA DEL LITORAL (ESPOL). Downloaded on February 09,2023 at 05:48:03 UTC from IEEE Xplore. Restrictions apply.
R EFERENCES
[1] N. Channouf, P. L’Ecuyer, A. Ingolfsson, and A. Avramidis. The
application of forecasting techniques to modeling emergency medical
system calls in Calgary, Alberta. Health Care Management Science,
10:25–45, February 2007. pages 1, 2
[2] A. Chen and T.-Y. Lu. A GIS-based demand forecast using machine
learning for emergency medical services. In Proc. International Con-
ference on Computing in Civil and Building Engineering, pages 1634–
1641, June 2014. pages 1, 2, 3, 4
[3] A. Y. Chen, T. Lu, M. H. Ma, and W. Sun. Demand forecast using
data analytics for the preallocation of ambulances. IEEE Journal of
Biomedical and Health Informatics, 20(4):1178–1187, 2016. pages 1,
2, 4, 8
maximum survival. Naval Research Logistics (NRL), 55(1):42–58, 2008.
pages 1
[5] G. Grekousis and Y. Liu. Where will the next emergency event occur?
Predicting ambulance demand in emergency medical services using
artificial intelligence. Computers, Environment and Urban Systems,
76:110 – 122, 2019. pages 1
[6] A. Haugsbø Hermansen. Machine learning for spatio-temporal fore-
casting of ambulance demand. Master’s thesis, Norwegian University
of Science and Technology, 2021. pages 2, 7
[7] H. Huang, M. Jiang, Z. Ding, and M. Zhou. Forecasting emergency calls
with a Poisson neural network-based assemble model. IEEE Access,
7:18061–18069, 2019. pages 1
[8] A. X. Lin, A. F. W. Ho, K. H. Cheong, Z. Li, W. Cai, M. L. Chee, Y. Y.
Ng, X. Xiao, and M. E. H. Ong. Leveraging machine learning techniques
and engineering of multi-nature features for national daily regional
ambulance demand prediction. International Journal of Environmental
Research and Public Health, 17(11), 2020. pages 1
[9] D. S. Matteson, M. W. McLean, D. B. Woodard, and S. G. Henderson.
Forecasting emergency medical service call arrival rates. The Annuals
of Applied Statistics, 5, 2011. pages 1
[10] H. Setzler, C. Saydam, and S. Park. EMS call volume predictions: A
comparative study. Computers & Operations Research, 36(6):1843 –
1851, 2009. pages 1, 2, 4, 5, 6, 7, 10
[11] K. Steins, N. Matinrad, and T. Granberg. Forecasting the demand for
emergency medical services. In Proc. of the 52nd Hawaii International
Conference on System Science, pages 1855–1864, January 2019. pages
1, 4
[12] G.-H. Strand and V. V. Holst Bloch. Statistical grids of Norway.
Technical report, Statistics Norway, Department of Economic Statistics,
2009. pages 2
[13] J. E. Thornes, P. A. Fisher, T. Rayment-Bishop, and C. Smith. Am-
bulance call-outs and response times in Birmingham and the impact
of extreme weather and climate change. Emergency Medicine Journal,
31(3):220–228, 2014. pages 1
[14] H. Wong and J.-J. Lin. The effects of weather on daily emergency
ambulance service demand in Taipei: a comparison with Hong Kong.
Theoretical and Applied Climatology, 141:321––330, July 2020. pages
1
[15] H.-T. Wong and P.-C. Lai. Weather inference and daily demand for
emergency ambulance services. Emergency Medicine Journal (EMJ),
29:60–64, October 2010. pages 1, 2, 7, 10
[16] H.-T. Wong and P.-C. Lai. Weather factors in the short-term forecasting
of daily ambulance calls. International journal of biometeorology, 58,
03 2013. pages 1, 7
[17] Z. Zhou. Predicting Ambulance Demand. PhD thesis, Cornell University,
2015. pages 1, 2, 3, 4, 7, 8, 10
[18] Z. Zhou and D. S. Matteson. Predicting ambulance demand: A spatio-
temporal kernel approach. In Proc. of the 21th International Conference
on Knowledge Discovery and Data Mining, page 2297–2303, August
2015. pages 1, 3, 4
[19] Z. Zhou and D. S. Matteson. Predicting Melbourne ambulance demand
using kernel warping. Annals of Applied Statistics, 10(4):1977–1996,
12 2016. pages 1, 4
[20] Z. Zhou, D. S. Matteson, D. B. Woodard, S. G. Henderson, and A. C.
Micheas. A spatio-temporal point process model for ambulance demand.
Journal of the American Statistical Association, 110(509):6–15, 2015.
pages 1, 4