DEGREE PROJECT IN COMPUTER SCIENCE AND ENGINEERING,

SECOND CYCLE, 30 CREDITS

STOCKHOLM, SWEDEN 2020

A time series forecasting approach

for queue wait-time prediction

ANTON STAGGE

KTH ROYAL INSTITUTE OF TECHNOLOGY

SCHOOL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE
A time series forecasting
approach for queue wait-time
prediction

ANTON STAGGE

Master in Computer Science

Date: June 10, 2020
Supervisor: Mats Nordahl
Examiner: Pawel Herman
School of Electrical Engineering and Computer Science, KTH
Swedish title: Förutsägelser av väntetid med hjälp av
tidsserieprognoser
 iii

Abstract
Waiting in queues is an unavoidable part of life, and not knowing how long
the wait is going to be can be a big source of anxiety. In an attempt to mitigate
this, and to be able to manage their queues, companies often try to estimate
wait-times. This is especially important in healthcare, since the patients are
most likely already under some distress.
In this thesis the performance of two different machine learning (ML) ap-
proaches and a simulation approach were compared on the wait-time predic-
tion problem in a digital healthcare setting. Additionally, a combination ap-
proach was implemented, combining the best ML model with the simulation
approach.
The ML approaches used historical data of the patient queue in order to
produce a model which could predict the wait-time for new patients joining the
queue. The simulation algorithm mimics the queue in a virtual environment
and simulates time moving forward until the new patient joining the queue is
assigned a clinician, thus producing a wait-time estimation. The combination
approach used the wait-time estimations produced by the simulation algorithm
as an additional input feature for the best ML model.
A Temporal Convolutional Network (TCN) model and a Long Short-Term
Memory network (LSTM) model were implemented and represented the se-
quence modeling ML approach. A Random Forest Regressor (RF) model and
a Support Vector Regressor (SVR) model were implemented and represented
the traditional ML approach. In order to introduce the temporal dimension to
the traditional ML approach, the exponential smoothing preprocessing tech-
nique was applied.
The results indicated that there was a statistically significant difference be-
tween all models. The TCN model and the simulation algorithm had the lowest
Mean Square Error (MSE) of all individual models. Both sequence modeling
models had lower MSE compared to both of the traditional ML models. The
combination model had the lowest MSE of all, adopting the best performance
traits from both the ML approach and the simulation approach. However, the
combination model is the most complex, and thus requires the most mainte-
nance.
Due to the limitations in the study, no single approach can be concluded as
optimal. However, the results suggest that the sequence modeling approach is a
viable option in wait-time prediction, and is recommended for future research
or applications.
iv

Sammanfattning
Att vänta i köer är en oundviklig del av livet. Att inte veta hur lång väntan
kommer att bli kan framkalla ångest. I ett försök att lindra denna ångestkälla,
samt för att kunna hantera sina köer, försöker företag ofta uppskatta vänteti-
den. Detta är särskilt viktigt inom hälso- och sjukvården, eftersom patienterna
troligtvis redan upplever någon typ av oro.
Syftet med denna uppsats är att jämföra prestandan hos tre olika metoder
för att förutspå väntetiden hos en digital vårdstjänst. Två olika maskininlär-
ningsmetoder (ML) samt en simuleringsmetod jämfördes. Utöver detta jäm-
fördes även en kombinationsmetod, som kombinerade den bästa ML-modellen
med simuleringsmetoden.
ML-metoderna använde sig av historisk data från patientkön för att skapa
en modell som kunde förutsäga väntetiden för nya patienter som ställer sig i
kön. Simuleringsalgoritmen imiterar kön i en virtuell miljö och simulerar att
tiden går framåt i denna miljö tills den nya patienten som anslöt sig till kön
kan tilldelas en ledig kliniker. På detta sätt kan en prediktion av väntetiden ges
till patienten. Kombinationsmetoden använde simuleringsprediktionerna som
ytterligare indata till den bästa ML-modellen.
En Temporal Convolutional Network (TCN)-modell samt en Long Short-
Term Memory (LSTM)-modell implementerades och representerade sekvens-
modelleringsmetoden (eng: sequence modeling). En Random Forest Regressor
(RF)-modell samt en Support Vector Regressor (SVR)-modell implementera-
des och representerade den traditionella ML-metoden. För att den traditionella
ML-metoden skulle få tillgång till tidsdimensionen applicerades förbehand-
lingstekniken exponentiell utjämning på dess data.
Resultatet visade att det fanns en statistiskt signifikant skillnad i kvadrat-
felet mellan alla modellerna. TCN-modellen samt simulationsalgoritmen hade
lägst medelkvadratfel av de ensamstående modellerna. Båda sekvensmodelle-
ringsmodellerna hade lägre medelkvadratfel än de traditionella ML-modellerna.
Kombinationsmodellen hade absolut lägst medelkvadratfel, då modellen be-
höll fördelarna från både ML- samt simuleringsmetoden. Däremot är kombi-
nationsmetoden den metod som kräver mest underhåll.
På grund av begränsningarna i studien kan ingen enstaka metod hävdas
vara optimal. Resultaten tyder emellertid på att sekvensmodelleringsmetoden
kan användas för väntetidsprediktion i ett kösystem, och rekommenderas där-
för för framtida forskning eller applikationer.
Contents

1 Introduction 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Background 5
2.1 Research Areas . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Sequence Modeling . . . . . . . . . . . . . . . . . . . 5
2.1.2 Queuing Systems . . . . . . . . . . . . . . . . . . . . 6
2.2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Exponential Smoothing . . . . . . . . . . . . . . . . . 8
2.2.2 Artificial Neural Networks . . . . . . . . . . . . . . . 8
2.2.3 Convolutional Neural Networks . . . . . . . . . . . . 9
2.2.4 Temporal Convolutional Networks (TCN) . . . . . . . 10
2.2.5 Recurrent Neural Networks . . . . . . . . . . . . . . . 15
2.2.6 Long Short-Term Memory (LSTM) . . . . . . . . . . 16
2.2.7 Regression Tree . . . . . . . . . . . . . . . . . . . . . 18
2.2.8 Random Forest Regressor (RF) . . . . . . . . . . . . . 19
2.2.9 Support Vector Regression (SVR) . . . . . . . . . . . 20
2.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Wait-time Prediction . . . . . . . . . . . . . . . . . . 23
2.3.2 Wait-time Prediction in Healthcare . . . . . . . . . . . 25

3 Methodology 27
3.1 The Queuing System and its Data . . . . . . . . . . . . . . . . 27
3.1.1 The Queuing System . . . . . . . . . . . . . . . . . . 27
3.1.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

v
vi CONTENTS

3.2.1 Naive . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.2 The Simulation Algorithm (Sim) . . . . . . . . . . . . 31
3.2.3 TCN . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.4 LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.5 RF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.6 SVR . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.7 Combination Model (Comb) . . . . . . . . . . . . . . 35
3.3 Hyperparameter Optimization . . . . . . . . . . . . . . . . . 35
3.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Results 38
4.1 Hyperparameter Optimization . . . . . . . . . . . . . . . . . 38
4.1.1 TCN . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.1.2 LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.3 RF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.4 SVR . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

5 Discussion 45
5.1 Model Comparison . . . . . . . . . . . . . . . . . . . . . . . 45
5.1.1 Forecasts . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2.1 Model Selection . . . . . . . . . . . . . . . . . . . . 47
5.2.2 Temporal Dimension . . . . . . . . . . . . . . . . . . 48
5.2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2.4 Hyperparameter Optimization . . . . . . . . . . . . . 49
5.3 Social Aspects and Usability . . . . . . . . . . . . . . . . . . 50
5.4 Sustainability and Ethics . . . . . . . . . . . . . . . . . . . . 51
5.5 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . 51

6 Conclusion 53

Bibliography 54

A Forecasts 61
 CONTENTS vii

Terminology
The following terms and abbreviations are used in this report:
ML - Machine Learning
TCN - Temporal Convolutional Network
LSTM - Long Short-Term Memory
RF - Random Forest
SVM - Support Vector Machine
SVR - Support Vector Regressor
Feature / predictor / parameter - Input variable
QL - Queue length, i.e. the number of consumers actively waiting in a queue
ED - Emergency Department
DLAMI - Deep Learning Amazon Machine Instance
Consultation / meeting - Used interchangeably to describe a consultation
meeting between a patient and a clinician.
Chapter 1

Introduction

A large number of industries that provide products or services require con-

sumers to wait in queues. Several studies have indicated that the waiting pro-
cess has a strong impact on the perceived quality of the service [1], [2]. In
certain queues it is relatively easy for the consumer to estimate how long the
wait is going to be, for example cash register queues, due to the fact that the
progress of the queue is visible. In other queues, however, the wait-time can
not be easily estimated by the consumer, e.g. online or healthcare queues. You
might know your position in the queue, but not how long each person in front
of you is going to need for their errand. An unknown wait-time is a bigger
source of anxiety compared to a known, finite, wait-time [1]. For this reason,
and to be able to manage the queues, companies (service/product providers)
often try to estimate the wait-time.
The accuracy of wait-time predictions is very important. If the predictions
are too short the consumers tend to become frustrated as they were essentially
lied to. If the predictions are too long the consumers might consider other
alternatives or rival companies instead [3], [4]. Wait-time prediction is espe-
cially important in healthcare, since the patients seeking help are most likely
already under some distress [4], [5].
Digital healthcare has become increasingly popular during recent years.
Digital healthcare providers offer services where patients can have video or
text consultations with clinicians online [6]. One of the main advantages of
digital healthcare compared to traditional healthcare is time efficiency. In tra-
ditional healthcare, one has to physically move to the treatment location and
then wait for an unknown amount of time before getting to see a clinician. In
digital healthcare, the inconvenience of having to move location does not exist.
Furthermore, since faster access to more information is available, predictions

1
2 CHAPTER 1. INTRODUCTION

for the wait-time can be made more easily. Patients seeking help consequently
do not have to waste their time waiting idly in a queue.
A substantial amount of research has been conducted with focus on wait-
time prediction [5], [7]–[13]. Some of this work will be covered in detail in
section 2.3. Several approaches to the problem have previously been inves-
tigated: Queuing Theory, Simulations and Machine Learning [6], [9], [10],
[13]. In the Machine Learning (ML) approach, historical data of the state of
the queue is used in order to obtain a model which can make predictions of the
future state of the queue. The state-of-the-art ML model used for this purpose
is the Random Forest (RF) [14] model but other models have been examined
[7], [9], [13].
Another research area that has had some advances more recently is Se-
quence Modeling, which consists of mapping an input sequence to an output.
This results in a new dimension added to the input, namely the temporal di-
mension. Sequence modeling is traditionally applied on problems involving
texts, speech, DNA or other naturally occurring sequences [15]. The current
state-of-the-art sequence modeling models are the Long Short-Term Memory
(LSTM) [16] network and the (more recently presented) Temporal Convolu-
tional Network (TCN) [15]. The historical data of a queue can be considered
a sequence, since the current state of the queue is highly dependent on pre-
vious states of the queue. While simpler linear techniques have been applied
[9], [17], no studies applying the more advanced techniques from the sequence
modeling research area on the wait-time prediction problem have been identi-
fied.
The aim of this thesis is to compare the performance of three different
approaches to the queue wait-time prediction problem in digital healthcare: a
sequence modeling approach, the traditional ML approach and a simulation
approach implemented by the company KRY [6]. In order to incorporate the
temporal dimension to the traditional non-sequence modeling ML approach,
the exponential smoothing [18] preprocessing technique is applied.

1.1 Problem Statement

The research question of this thesis is the following:

Does a sequence modeling approach, trained on data from the

queue’s history, perform better at predicting patient wait-times
in digital healthcare compared to a non-sequence modeling ML
 CHAPTER 1. INTRODUCTION 3

approach utilizing exponential smoothing, and a simulation ap-

proach?

In order to answer the research question stated above the performance of two
sequence modeling models (TCN and LSTM), two non-sequence modeling
ML models utilizing exponential smoothing (RF and SVR), and the simula-
tion algorithm implemented by the company KRY [6] will be compared. Ad-
ditionally, the performance of a combination model, comprising of the best
performing ML model and the simulation algorithm, will be examined.

1.2 Purpose
The purpose of this research is to investigate whether the more advanced tech-
niques from sequence modeling research area can be applied to the queue wait-
time prediction problem in healthcare, and if so, investigate how it performs
compared to the traditional approaches. Recent studies have applied other ML
approaches with success, but the ML models used do not take the temporal
dimension into account. The TCN model has recently demonstrated its capa-
bility on a wide variety of tasks and datasets [15]; thus, there lies an interest
in exploring its usability on the wait-time prediction problem. If the sequence
modeling approach proves to be viable, it could be applied to other wait-time
prediction problems, not just wait-time prediction in a digital healthcare set-
ting, as long as there exists historical data of the state of the queue.

1.3 Scope
This paper compares four different ML models, where two are sequence mod-
eling models and the other two are not. Other models are not considered. The
scope does not include the application of queuing theory to the problem. Fur-
thermore, the data used in this thesis comes from a single digital healthcare
provider and from Swedish patients only. The consumer behaviour might dif-
fer in other markets.

1.4 Outline
This report is divided into six chapters. In Chapter 2 (Background) the re-
search areas are introduced, the theory behind all models and techniques ap-
plied in this paper is explained, and finally related work is presented. In Chap-
4 CHAPTER 1. INTRODUCTION

ter 3 (Methodology) the experiments conducted in this thesis are presented in

detail. In the following chapter, Chapter 4 (Results), the results of the experi-
ments are reported. Next, in Chapter 5 (Discussion), the results are interpreted
and a discussion is held regarding both the results and the limitations in the
methodology used in finding said results. Finally, the report and its findings
are concluded in Chapter 6 (Conclusion).
Chapter 2

Background

This chapter starts by introducing the research areas within which this thesis
is carried out. Secondly, the theoretical background of all methods used is
explained in detail. Lastly, related work and their similarities and differences
are presented.

2.1 Research Areas

2.1.1 Sequence Modeling
A sequence is a collection of observations in which order matter. The research
area known as sequence modeling consists of mapping an input sequence to
an output sequence, where every output depends only on input observations
before it. More formally, as described in [15], a sequence modeling network is
a function that given an input sequence x0 , ..., xT predicts an output sequence
y0 , ..., yT
f (x0 , ..., xT ) = y0 , ..., yT (2.1)
under the constraint that every yt only depends on x0 , ..., xt and not on any
of the future input xt+1 , ..., xT . This is known as the causal constraint. The
goal of any sequence modeling network is to minimize the prediction error
between the predicted values and the actual observations. This is commonly
done with ML and Artificial Neural Networks [15], [19]. The current state-
of-the-art network models are LSTMs [16] and TCNs [15]. The LSTM model
has been a popular choice for recurrent networks for some time [19]. However,
the TCN has proven to be superior to several network models (including the
LSTM) in multiple cases more recently [15], [19]–[21].

5
6 CHAPTER 2. BACKGROUND

Time Series Forecasting

A special type of sequence known as a time series is a sequence that consists of
observations made at subsequent points in time for one or more attributes. A
single attribute measured/observed is known as univariate time series, whereas
if multiple attributes are considered, it is known as a multivariate time series.
The research area of predicting future values in a time series is known as time
series forecasting [22]. Time series forecasting is a rather hot research topic
in which several papers have been released in the most recent years [23]–[26].
Time series forecasting techniques most often predicts the next value in the
series, which makes it a special case of sequence modeling explained by the
function:
f (x0 , ..., xt ) = yt (2.2)
where x0 , ..., xt are input vectors (in the case of multivariate time series fore-
casting) and yt is the output variable. This is known as one-step forecasting.

2.1.2 Queuing Systems

The most classical type of queuing system is a single-queue single-server,
which follows the First-in First-out (FIFO) discipline [27]. In this system con-
sumers are waiting for a single limited resource, as shown in Figure 2.1. Since
they can not access it all at the same time a queue is required [13]. The next
logical extension of this queuing system is the single-queue multi-server sys-
tem, in which a single line of consumers are waiting for multiple servers that
provide the same resource, as shown in Figure 2.2. [27]

Figure 2.1: A single-queue single-server system.

 CHAPTER 2. BACKGROUND 7

Figure 2.2: A single-queue multi-server system.

Queue Modeling
The research area known as queuing theory focuses on using mathematical
models to explain a queuing system. This is the traditional way of modeling a
queue [13]. The different mathematical formulas provided by queuing theory
each have different constrains which have to be met [12], [27]. Alternative less
restricted methods of modeling a queue include simulations or the use of ML.
Simulations can be used to model a queue. This is the algorithmic ap-
proach. A simulation is designed to mimic a queue as perfectly as possible,
and then time is moved forward in a virtual environment. By using assump-
tions or calculated estimations for the amount of time each consumer is going
to spend at the server, time can be moved forward until every consumer has
visited a server. This queuing model, therefore, heavily relies on the accu-
racy of the estimated service time. A detailed description of the simulation
algorithm investigated in this thesis is given in subsection 3.2.2.
When modeling a queue with ML, historical data of the state of the queue is
used. A set of features is selected to represent the state, and a model is trained
on the data of said features. The goal is for the model to portray the system as
accurately as possible [5], [10], [13], [28]. This approach is a viable alternative
to queuing theory [10], and all it requires is data regarding the history of the
queue. When a model of the queue has been acquired it can be used to make
predictions about future states of the queue, and certain variables in its state,
for instance the wait-time.
8 CHAPTER 2. BACKGROUND

2.2 Theory
2.2.1 Exponential Smoothing
The technique and use of exponential smoothing and exponential windows
have existed for a long time. In the statistical literature it is often credited to
Brown [29], and referred to as Brown’s simple exponential smoothing. The
method was originally introduced as a forecasting method, but have in the
more recent years been proven to be a good preprocessing method [30]. The
forecasting method is described as:

ŷt+1 = αxt + (1 − α)ŷt (2.3)

where xt is the raw observation at time t, ŷt is the smoothed observation at

time t, and α is a constant [31]. When used as a preprocessing method, the
same formula is applied, but the forecasts for time t + 1 are used as inputs for
time t.

2.2.2 Artificial Neural Networks

The first model of an artificial neuron was introduced as early as 1943 in [32],
a mathematical representation of the neurons in a human brain. An artificial
neuron is designed to take multiple inputs and multiply every input by a weight.
The result is summed to produce a single value which is fed through an acti-
vation function f to produce an output [33]. This is illustrated by Figure 2.3.
Ever since the introduction of artificial neurons, more sophisticated models,

Figure 2.3: A generic artificial neuron.

that connect multiple neurons into a large network, known as Artificial Neural
Networks (ANNs), have been invented. Rosenblatt introduced the first percep-
tron model in [34]. An example of a multi-layer perceptron (MLP) network,
 CHAPTER 2. BACKGROUND 9

Figure 2.4: An example of a multi-layer perceptron network. The network has

2 hidden layers, with 5 and 4 neurons respectively, and a single output neuron.

connecting multiple artificial neurons together to form a network with 2 layers

and a single output neuron is displayed in Figure 2.4.
ANNs are supervised universal function approximators. These networks
can learn to approximate any given function by being trained on labeled data.
During training the weights in the network are slightly updated in order to
reduce the error over all data. This is done with an algorithm known as the
backpropagation [35]–[37] training algorithm. [33]

2.2.3 Convolutional Neural Networks

A Convolutional Neural Network (CNN) [38] is a regular ANN that uses the
mathematical operation called convolution in at least one of its layers [39],
making it a convolutional layer. Convolutional layers take the spatial relation
of grid-like inputs into account. A spatial relation in, for example, an image
is that closer pixels are more likely to be correlated. In a convolutional layer,
window formations of weights (known as filters or kernels) stride over the
input in order to produce an output known as a feature map. An example of a
convolution with a single filter over a 2-dimensional input image can be seen
in Figure 2.5a [39]. In each step the filter stops at, it produces one new value
to the feature map by taking the dot product of its weights and the inputs the
filter covers and passing the result through an activation function. The filters
are thus made of neurons. [39]
10 CHAPTER 2. BACKGROUND

(a) An example of a convolution on a 2-dimensional input image of size 3x4, with a

filter of size 2x2, and the resulting feature map of size 2x3.

(b) 10 feature maps stacked together to form a single output with depth 10.

Figure 2.5: A detailed illustration of a convolutional layer.

A single convolutional layer can apply multiple filters to the same input.
The feature maps produced by each filter are stacked on top of each other in
order to produce a single output, as shown in Figure 2.5b. The number of
filters, the size of the filters and the size of the stride are hyperparameters to
the layer. [39]

2.2.4 Temporal Convolutional Networks (TCN)

The technique applied by a TCN has existed for a long time with many varia-
tions, but the name was established in 2018 by Bai, Kolter, and Koltun. The
name was said to be a descriptive term for a family of architectures, and a
generic TCN model was provided which is described below [15].
 CHAPTER 2. BACKGROUND 11

A TCN applies the concepts of a CNN on sequential input data by apply-

ing causal 1-dimensional convolutions on multiple subsequent inputs together
instead of on a single one [20]. The filter size k determines how many inputs
are considered in a single layer. The output is always zero padded to be the
size of the input, i.e. zero padding of length (k − 1). Figure 2.6 illustrates this
further.

Figure 2.6: An illustration of a 1-dimensional convolution applied to a se-

quence of length 10 with 6 features and a kernel size k = 2.

The convolutional layers are stacked to provide depth to the network as

shown in Figure 2.7.
12 CHAPTER 2. BACKGROUND

Figure 2.7: An example of a basic TCN architecture with 3 stacked layers each
with k = 3.

In order to expand the number of inputs used to compute the output (known
as the receptive field) beyond linear with the filter size and network depth, a
technique knows as dilated convolutions [40] are made. A dilation is a fixed
step d between two filter taps. The filter skips the inputs between the taps as
shown in Figure 2.8 [15].

Figure 2.8: An example of a dilation layer with d = 4 and k = 3. Not all lines
have been drawn to simplify the image. In reality, each output ŷt is connected
to k inputs xt .

Stacking dilated convolutional layers and increasing the dilation d expo-

nentially for each layer makes the receptive field exponentially large while still
making sure that there is some filter tapping each input [15]. An example of
 CHAPTER 2. BACKGROUND 13

such a network is shown in Figure 2.9.

Figure 2.9: An example of 3 dilation layer stacked with d = 1, 2, 4, k = 3

yielding a receptive field of 29.

A temporal convolutional layer is made of a residual block, which addition-

ally connects the input directly to the output in order to allow the layer to learn
modifications to the mapping instead of the whole transformation [15]. There
exists substantial evidence that the use of residual blocks make optimization
easier and improves accuracy for deep architectures [15], [41]. As is shown
in Figure 2.10, a single residual block consists of 2 repeated sub-blocks con-
taining a dilated causal convolution followed by a weight normalization [42],
a ReLU activation function [43] and a Dropout [44] layer. Moreover, a single
1-dimensional convolutional layer with a kernel size of 1 connects the input to
the output with element-wise addition [15].
14 CHAPTER 2. BACKGROUND

Figure 2.10: The architecture of a single residual block, as described by [15].

Because of the residual blocks, the calculation of the receptive field changes.
Without residual blocks, the formula for the receptive field R would be:

R(n) = R(n − 1) + (k − 1) × d(n) (2.4)

where n is the layer, d(n) is the dilation for layer n and R(0) = 1. For example,
looking back at Figure 2.7 which has k = 3 and d = 1, the receptive field for
the output layer is 15. However, since each residual block contains 2 causal
convolutions, the formula becomes:

R0 (n) = R0 (n − 1) + 2 × (k − 1) × d(n) (2.5)

If the dilation rate increases exponentially with a base 2, the formula can be
simplified to:

R0 (n) = 1 + 2 × (k − 1) × (20 + 21 + 22 + ... + 2n−1 )

(2.6)
= 1 + 2 × (k − 1) × (2n − 1)

An illustration of how the residual blocks affect the receptive field is shown
in Figure 2.11. As one can see the number of inputs considered for a single
output is 13. Using Equation 2.6:

R(2) = 1 + 2 × (3 − 1) × (22 − 1) = 1 + 2 × 2 × 3 = 13 (2.7)

 CHAPTER 2. BACKGROUND 15

Figure 2.11: An example of the receptive field of two stacked residual blocks
with k = 3 and dilation increasing exponentially with a base of 2.

2.2.5 Recurrent Neural Networks

A Recurrent Neural Network (RNN) is another sequence modeling extension
of the ANN in which information from previous inputs x are incorporated in
a hidden state h which is passed forward in time. This allows the network to
retain a memory of previously observed inputs for each produced output [39].
Figure 2.12 displays the computational graph for a basic RNN with a single
input unit.

Figure 2.12: The computational graph of a basic RNN. U W and V are weight
matrices, x is the input, h is the hidden state which is passed forward in time,
o is the output, and ŷ is the output after the activation function. The black box
represents a delay of one time-step.

The process of producing outputs ŷ (t) (known as the forward propagation),

16 CHAPTER 2. BACKGROUND

begins with an initial state h(0) and then for each time-step t in the sequence
the following equations are applied:

a(t) = b + Wh(t−1) + Ux(t)

h(t) = Activation(a(t) )
(2.8)
o(t) = c + Vh(t)
ŷ (t) = Activation(o(t) )

where b, c are bias vectors [39].

2.2.6 Long Short-Term Memory (LSTM)

An LSTM [16] is an extension of the RNN that has an additional memory cell.
The long-term memory comes from the weights, and the short-term memory
comes from said memory cells. Another important feature added to the LSTM
is its gates. A gate is a sigmoidal neuron that takes the current input x(t) and
the hidden state h(t−1) as input [45]. It is called a gate because its result is
multiplied with the result from another node. Therefore, if its value is close
to zero the gate is closed, and vice versa. The LSTM unit has an input gate
connected to the input node, a forget gate connected to the memory node, and
an output gate connected to the output node [45]. Figure 2.13 displays the
computational graph of a basic LSTM unit.
 CHAPTER 2. BACKGROUND 17

Figure 2.13: A computational graph of a basic LSTM unit. U W and V are

weight matrices, x is the input, h is the hidden state which is passed forward
in time, s is the memory node also passed forward in time, and ŷ is the output
after the activation function. i, f and o are the input, forget and output-gates.
The black box represents a delay of one time-step, the flattened S represents
an arbitrary activation function, and σ is the sigmoid activation function.

Similar to the RNN, the process of producing outputs ŷ (t) , i.e. forward
propagation, for an LSTM unit starts with a initial hidden state h(0) , and then
the equations in Equation 2.9 are applied for each time-step t of the sequence.
Here i(t) denotes the input gate, f (t) the forget gate, o(t) the output gate, and
s(t) the memory node.

i(t) = σ(Wi h(t−1) + Ui x(t) )

f (t) = σ(Wf h(t−1) + Uf x(t) )
o(t) = σ(Wo h(t−1) + Uo x(t) )
(2.9)
s(t) = f (t)  s(t−1) + i(t)  Activation(Ws h(t−1) + Us x(t) )
h(t) = o(t)  Activation(s(t) )
ŷ (t) = Activation(Vh(t) )
18 CHAPTER 2. BACKGROUND

where σ(·) is the sigmoid activation function and  is element wise multipli-
cation [39]. All bias vectors have been excluded from the equations for sim-
plicity. The size of all matrices W and U is a hyperparameter to the LSTM.
Much like the TCN, the modern LSTM utilizes dropout [44], both between
layers, and between the recurrent units.

2.2.7 Regression Tree

A Regression Tree [46] is a series of questions about the features in the dataset
built up in a tree-like structure. Each node in the tree contains a question or test
on a specific feature, thus, each branch represents an answer or outcome of the
test. The leafs in the tree hold the prediction values [47]. Figure 2.14 displays
a simple constructed example of a regression tree modeling the wait-time at a
hospital. A regression tree divides the prediction space into non-overlapping
regions, each region being a leaf node in the tree. This is illustrated further by
Figure 2.15.

Figure 2.14: An example of a Regression Tree modeling the wait-time in a

hospital queue. The leaf nodes show the predicted wait-time in minutes. If the
outcome of the test is true, the right branch is selected.

When constructing a tree, an approach called recursive binary splitting is

commonly used. This is a greedy top-down approach where you start with one
region containing all data points. The region is then, successively, split into
two smaller regions, where each split is optimal considering only this step,
optimal meaning that all predictors and cutpoints are considered. The chosen
 CHAPTER 2. BACKGROUND 19

Figure 2.15: The prediction space divided into regions based on the Regres-
sion Tree in Figure 2.14.

predictor and cutpoint is the one that minimizes the Residual Sum of Squares
(RSS) given by Equation 2.10, where J is the number of regions (R), and ŷRj
is the mean prediction value of the points within region Rj . [47]
J X
X
RSS = (yi − ŷRj )2 (2.10)
j=1 i∈Rj

2.2.8 Random Forest Regressor (RF)

As the Regression Tree model described in subsection 2.2.7 has very high
variance, meaning that two models trained on two partitions of the same data
tend to differ, the ensemble learning technique known as Random Forest [14]
greatly improves its accuracy. A RF model is, as suggested by the name, a
collection of trees. The final prediction is simply the average of all the trees.
Before constructing a single tree in the forest, a unique training set is first cre-
ated by repeatedly sampling from the original training set, a process known
as bootstrapping. The tree is then built by considering only a fixed number
of predictors m during each split. Those predictors are chosen at random for
√
each split, and usually m = p where p is the total number of predictors
(size of input dimension). Because of the random selection of predictors and
20 CHAPTER 2. BACKGROUND

bootstrapped training data, each tree in the forest become less correlated to
the others, and the variance of the model is reduced. The number of trees, the
maximum depth of each tree and the number of predictors m are hyperparam-
eters to the model. [47]

2.2.9 Support Vector Regression (SVR)

Support Vector Classifier
A Support Vector Classifier (SVC) is a hyperplane that separates 2 classes
of points with the widest possible margin M to the plane. The formula for
a hyperplane is given by w ~ · ~x − b = 0, where w ~ is a weight vector, b is a
constant and · is the scalar product. All points on one side on the plane will
~ · ~x − b > 0, and all points on the other side will have w
have w ~ · ~x − b < 0.
Thus ∀i, the point xi and its corresponding class yi ∈ {−1, 1} should satisfy
Equation 2.11 [47].

yi × (w
~ · x~i − b) ≥ M (1 − i ) (2.11)
Every point should be on the correct side of the plane, with margin, but allow-
ing for some (i ) slack. The slack relaxes the problem so that a solution can
2
be found for more cases. The size of the margin M can be calculated to ||w|| ~
.
Maximizing the margin is therefore equivalent to minimizing the length of w. ~
The optimization problem for the SVC is

X
~ 2+C
Minimize ||w|| i
i
(2.12)
Under the constraints yi × (w
~ · x~i − b) ≥ 1 − i , ∀i
i ≥ 0, ∀i

where C is a hyperparameter that tunes the amount of slack allowed [48]. Fig-
ure 2.16 shows an example of a SVC in 2D space.
 CHAPTER 2. BACKGROUND 21

Figure 2.16: An example of a SVC in 2D space. The solid line represents

the hyperplane, the dashed lines are the margins. The red and blue point being
inside the margin is an effect of the their respective slack variables being larger
than 0.

An interesting property of the SVC is that it is only the points that lay on or
within the margin that affect the position of the hyperplane. These are known
as the support vectors [47]. For example, if you were to move any of the top
four blue points in Figure 2.16 to any other location on the correct side of the
margin, it would not change the classifier at all.
In 2015 the authors in [49] proved that the linear regression model known
as the Elastic Net [50] could be reduced to a SVC. For every Elastic Net in-
stance, there exists a solution to a constructed binary classification problem in
which the hyperplane is identical to the Elastic Net solution. The practical im-
plications of this is that it enables the use of optimized SVC solvers for Elastic
Net problems [49].

Support Vector Machine

A Support Vector Machine (SVM) is an extension of a SVC that can handle
data with non-linear boundaries between classes by transforming the feature
space to a higher dimension and solving the (linear) problem there. This is
done with kernels [47] (not to be confused with convolutional kernels). The
idea behind kernels is to use the advantage of a high-dimensional space with-
out actually moving to a higher dimension. The only operation done between
two points in the high-dimension space is the scalar product, this operation
can be computed using the low-dimensional representation of the points. An
example of how this is done is shown in Equation 2.14.
22 CHAPTER 2. BACKGROUND

2
 
 
√ x 1
x
~x = 1 φ(~x) =  2x1 x2  (2.13)
x2
x22

φ(~x) · φ(~y ) = x21 y12 + 2x1 x2 y1 y2 + x22 y22 =

(2.14)
= (x1 y1 + x2 y2 )2 = (~x · ~y )2 = K(~x, ~y )

Equation 2.13 shows a transformation φ of a 2D vector into a 3D space, and

Equation 2.14 shows how you can calculate the dot product in the 3D space
by using the 2D space representations of the two vectors ~x and ~y . This is an
example of the common polynomial kernel

K(~x, ~y ) = (~x · ~y + d)p (2.15)

where d = 0 and p = 2. Another common kernel is the radial basis function
(rbf):
2
K(~x, ~y ) = e−γ||~x−~y|| (2.16)
where γ is a hyperparameter. [47]
The optimization problem for the SVM with the transformations is given
by Equation 2.17 [48]. However, in practice, the optimization problem is con-
verted into a corresponding dual problem using the Lagrangian approach [51].
It is in this dual problem, which is out of scope for this paper, the kernel func-
tions are utilized.

1 X
~ 2+C
Minimize ||w|| i
2 i
(2.17)
Under the constraints yi × (w
~ · φ(~
xi ) − b) ≥ 1 − i , ∀i
i ≥ 0, ∀i

Support Vector Regression

The SVR model uses all the same techniques as the SVM, except that the prob-
lem has been converted in to a regression problem which focuses on finding a
function f (x) that has an error |f (xi )−yi | at most ε [52]. The hyperparameter
ε is an upper bound on the error tolerated, and much like the support vectors of
the SVM, the SVR does not care about errors less than ε [53]. The optimiza-
tion problem for the SVR, given training vectors x~i and their corresponding
target values yi , is:
 CHAPTER 2. BACKGROUND 23

1 X
~ 2+C
Minimize ||w|| (ζi + ζi∗ )
2 i
Under the constraints yi − w
~ · φ(xi ) − b ≤ ε + ζi , (2.18)
~ · φ(xi ) + b − yi ≤ ε + ζi∗ ,
w
ζi , ζi∗ ≥ 0, ∀i

where ζi∗ denotes training errors above ε and ζi denotes training errors below
ε [48], [54]. The ζ variables, analogous to the slack variables in the SVM
optimization problem, allow for some pairs (xi , yi ) to be approximated with
worse precision than ε and thus relaxes the constraints of the optimization
problem [53]. Figure 2.17 illustrates a linear SVR model.

Figure 2.17: An example of a linear SVR in 2D space.

2.3 Related Work

In this section some of the identified related work is presented. Most of the
work presented is focused on the ML approach. This is due to the fact that the
choices of ML models and input features in said work were used as a basis for
the choice of models and features in this thesis.

2.3.1 Wait-time Prediction

The research done in [11] is focused on producing new valuable predictors for
a multi-server queue with time-varying arrival rate, service rate and customer
abandonment, much like the queue in place at the company KRY. This study
suggests that the predictors Queue-Length (QL) and Head of Line (HOL) are
significant. HOL is the elapsed waiting time for the consumer in the front of
24 CHAPTER 2. BACKGROUND

the line. Although [11] focuses on the queuing theory research area, the result
of their study can be applied in this thesis.
In a study from 2017 by Mourao, Carvalho, Carvalho, and Ramos [12]
several different models to predict (classify) wait-time overflows in banking
queues were compared. Banks in Brazil can be served fines if their customers
have to stand in line over a certain amount of time. Such a case was consid-
ered an overflow. The models compared were 3 ML models (Deep Learning,
Gradient Boost Machine and RF) as well as a model based on queuing theory.
The ML models were trained on historical data about the queue state for each
costumer. The features of this data included, but were not limited to: Arrival
Rate, Service Rate, Arrival Time, Average Service Time, Queue Length, Head
of Line (HOL). All ML models outperformed the queuing theory model, both
in accuracy and F1-score. The GBM achived the highest accuracy of 97%, and
the RF model had an accuracy of 96%. The work done in [12] differs from this
paper because of the single-server queuing system used and the classification
approach.
The ML approach was proven to be viable for wait-time prediction by
Kyritsis and Deriaz in their study from 2019 [10]. Kyritsis and Deriaz im-
plemented a ANN model with 2 hidden layers and trained it with an Adam
Optimizer on historical data from banking queues. The dataset used is pub-
licly available and consisted of, for each person that joined the queue, the 5
features: queue length, day of week, hour of day, minute of hour and the wait-
time. The data had a mean wait-time of 13.18 minutes, a median of 12 minutes
and a standard deviation of 5.95 minutes. The final model produced and Mean
Absolute Error (MAE) of 3.35 minutes, compared to a naive model which al-
ways predicted the mean with a MAE of 4.71 minutes. The work in [10] differs
from this paper by the fact that the number of servers in the multi-server queue
system was an unknown variable, and by the choice in ML model.
In a study by Zhang, Nguyen, and Zhang in 2013, several algorithms, in-
cluding some ML models, were implemented to predict the wait-time on a
time series dataset from the Department of Motor Vehicles (DMV) in San
Jose, California [17]. The data consisted of the wait-time for every 10 minutes
of the day. The simple linear regression model achieved the lowest average
error, followed closely by the SVM. The ML models were trained on the fea-
tures: latest wait-time, time of the day and day of the week. Much like the
work in this thesis, [17] compares the performance of multiple ML models on
wait-time prediction. However, the work differs by the fact that the data used
as well as the algorithms chosen were much simpler.
 CHAPTER 2. BACKGROUND 25

2.3.2 Wait-time Prediction in Healthcare

As early as 1987 attempts at modeling queues in healthcare were made in [55].
Accurate wait-time predictions are especially important in healthcare, because
the patients are more likely to already be in a worried state of mind when they
are seeking help [5], [56].
In the study from 2017, performed by Joseph, Hijal, Kildea, Hendren, and
Herrera the treatment durations in a Radiation Oncology clinic were predicted
and compared by four different ML regression models [5]. The predicted treat-
ment durations were used to produce waiting time estimations. The ML mod-
els were trained on historical data of treated patients. The features of the data
included general information about the patient, which physician the patient
had, at which time the appointment was scheduled as well as actually held,
general information about the treatment received and more. The 4 different
ML regression models compared were a linear regression model, a SVM, a
Decision tree and a RF. The RF model was found to be the best with a MAE of
4.6 minutes and an R2 -score of 0.47. Even though the study done in [5] does
not directly predict the wait-time as is done in this thesis, a different approach
is proposed, and the authors raise the importance of doing research regarding
wait-time prediction within a healthcare setting.
In 2018, Curtis, Liu, Bollerman, and Pianykh examined the possibility
to predict patient wait-time using ML in the Massachusetts General Hospital
Department of Radiology [9]. Four different examination types were consid-
ered: CT, MRI, ultrasound and radiography. Only the radiography examina-
tion modality had a drop-in queue system with 9 devices that could be used in
parallel. From the existing information system at the hospital, the authors were
able to extract 40 different predictors for each patient that visited the hospital
during the span of half a year. Multiple ML models (such as GBM, Elastic Net,
RF, SVM, Bagging, Neural Network and more) were trained and evaluated on
the collected historical data. For the radiography examination modality, the
RF model achieved the lowest RMSE of 8.816, and had the 3:rd highest R2
score of 0.4582. The Elastic Net followed by the Linear Regression model
had the 2 highest R2 scores, both slightly above 0.46. The most important
predictors (applicable to the drop-in radiography) were identified as the queue
length (current and most recent) and the median examination time of the five
most recent examinations. The work done in [9] provides further indication
that RF is an good choice in model for wait-time prediction in patient queues.
Moreover, it provides a new potential predictor which can be adopted in the
work done in this paper.
26 CHAPTER 2. BACKGROUND

Sanit-in and Saikaew made a comparative study of three wait-time predic-

tion approaches in 2019 [13]. The models considered were a queuing theory
model, a naive model predicting the mean wait-time and a RF model. The
study used two datasets collected from two different queues. The first dataset
had 3480 records from a post office queue, and the second had 1348 records
from an ear, nose and throat clinic from a hospital. Both datasets had the same
features. These features included, but were not limited to: the number on the
queue ticket, hour of the day, day of the week, period of day (morning, after-
noon), queue length, arrival rate, service rate and wait-time. The wait-time
was divided in to 5 classes ranging from very short to very long in order to
convert the problem in to a classification problem. The RF model achieved
the highest accuracy on both datasets. For the post office dataset it had an ac-
curacy of 81.70%, while the queuing theory model had an accuracy of only
65.23%. The naive baseline model achieved an accuracy of 64.94%. The RF
model had an accuracy of 85.76% on the ear nose and throat clinic dataset,
where the naive model had an accuracy of 68.89%. The queuing theory model
could not be applied on the ear, nose and throat clinic, since it did not satisfy
all constraints that the model requires. In addition to the comparative study,
[13] also explored the feature importance for the RF model. Three features
that proved to affect the outcome of the predictions the most were the queue
length, the arrival rate and the service rate. The arrival rate was measured as
the number of consumers arriving per minute at the time, and the service rate
was measured as the number of consumers receiving the service per minute
at the time. Compared to this thesis, the work in [13] had a lot less data.
Moreover, they approached the wait-time prediction problem as a classifica-
tion problem rather than as a regression problem.
Chapter 3

Methodology

This chapter describes the methods used to carry out the experiments. A de-
scription of the queuing system in place at KRY (hereby referred to as the
company), and the historical data used for training is given. A detailed de-
scription of the implementation of all the prediction models is given (including
the simulation algorithm implemented by the company), as well as the hyper-
parameter optimization executed for the ML models. Finally, the evaluation
metrics are defined.
All hyperparameter optimization, training and evaluation were executed
on a Deep Learning Amazon Machine Instance (DLAMI) of type g4dn.xlarge
(4x vCPU, 16GB RAM, 1x NVIDIA T4 GPU), running Ubuntu 16.04.

3.1 The Queuing System and its Data

3.1.1 The Queuing System
The queue system in place at the company is a single-queue multi-server sys-
tem, as described in subsection 2.1.2. The consumers are the patients and the
servers are clinicians offering the consultation resource. However, there are
some differences, making this system slightly more complex.
One difference in this system, compared to the system described earlier,
is that patients can not be received by just any clinician. Patients seek help
for a certain symptom and belong to certain age groups, and some symptoms
or age groups require specialization to treat. One example is child care. A
child can only receive help from a clinician specialized in child care. To solve
this, patients are assigned labels when joining the queue, which are based on
variables such as age and symptom. The patients can only be matched with

27
28 CHAPTER 3. METHODOLOGY

clinicians carrying the same labels, indicating that they can treat those patients.
Another difference is the fact that there exist some special cases of patients,
which are given different priorities in the queue. Such cases include: infants,
prescription renewals, and internal referrals. In the case of prescriptions re-
newals, patients do not have to have a video consultation but can do this via
a text meeting. These meetings require much less time compared to video
consultations. Internal referrals is the case where a clinician at the company
has referred a patient to another clinician at the company. The patients in the
queue are sorted firstly by priority, and secondly by arrival time.
The last difference in the queue system is that patients are unable to join
the queue if the wait-time estimation by the simulation model exceeds 4 hours.
This is a feature implemented to avoid overcrowding.

3.1.2 Data
Data from 767246 patients between the period of 2018-04-22 to 2020-02-17
was collected. In the data cleaning process patients with missing values and
patients with a recorded wait-time exceeding 6 hours were considered anoma-
lies and were therefore dropped. Next, patients with a difference in wait-time
compared to the previous patient of more than 3 times the standard deviation
(σ = 9.3214 minutes) were considered outliers and were also dropped. This
was done due to the fact that admins can access the queue system and manu-
ally assign patients to clinicians, resulting in a very small wait-time. After the
cleaning process, data from a total of 757086 patients remained.
The data collected from the queue is not uniformly sampled. Sequence
modeling applications implicitly require that the data is uniformly sampled
to be effective. The solution applied to this problem was to append the time
intervals ∆t (difference in time between each patients) to the input vector as
suggested in [57]. A description of each feature included in an observation is
displayed in Table 3.1.
For the sequence modeling models, the input features were normalized by
removing the mean and scaling to unit variance. This was done using Scikit-
learns StandardScaler [48]. The sequence length was set to 110, thus, each
multivariate input sequence consisted of 110 inputs, x0 , ..., x110 , where each
vector xt contained information about all input features described in Table 3.1.
The output target y110 was the actual wait-time feature for patient x110 , as de-
scribed in Equation 2.2.
The technique of Exponential Smoothing, as described in subsection 2.2.1,
was applied in order to incorporate the temporal dimension in the data for the
 CHAPTER 3. METHODOLOGY 29

non-sequence modeling models. A smoothened representation for each in-

put feature (except for year, month, day and minute) was calculated by Equa-
tion 2.3. The smoothened representations of the input features thus contained
information about all previous input features, decreasing exponentially with a
factor α. The value of α was left as a hyperparameter. Whether to use only
the smoothened representations or to append them to the input vector was also
left as a hyperparameter.
The data was divided into 3 different sets: a training, validation and test-
set. This division was done in chronological order, in order to preserve the
temporal dimension of the time series. 80% of the data was used for the train-
set, 10% was used for the validation-set, and the remaining 10% for the test-set.
30 CHAPTER 3. METHODOLOGY

Feature Description
Year Year of the visit.
Month Month of the year.
Day Day of the month.
Hour Hour of the day.
Minute Minute of the hour.
∆t Seconds elapsed since the latest patient
joined the queue.
HoL How long the patient currently in the front
of the queue has waited.
Queue length Number of patients currently in the queue.
Arrival rate Number of patients arriving per second,
measured over the latest 15 patients.
Gender Gender of the patient. 0, 1 or 2, where 0
means not specified.
Age Age of the patient.
Symptom The symptom the patient is seeking help
for. This is selected by the patient from a
list of 61 symptoms, and was represented
by the index in said list.
Latest meeting length The length of the latest meeting that
ended.
Latest wait-time The wait-time of the latest patient that
received help.
Doctors working Number of doctors currently working.
Receiving service Number of patients in meetings.
Mean average meeting length The mean of the currently working
doctors average meeting length.
Service rate Mean number of patients in meetings
measured over the latest 15 patients.
Simulation wait-time The wait-time estimation given by the
simulation model.
Actual wait-time The actual wait-time in minutes for the
patient.

Table 3.1: A description of all the features in the data collected from the queue.
The first group consists of all input features used for the ML models. The
second group consists of the simulation estimation and the target feature.
 CHAPTER 3. METHODOLOGY 31

3.2 Models
3.2.1 Naive
Two naive models were implemented. The first model, Naive (previous), was
defined as:
ŷ (t+1) = y (t) (3.1)
This model is not possible to apply in the real world application, because the
wait-time for the previous patient is most often not known when the next pa-
tient arrives. The previous patient is most likely still in the queue, thus, this
prediction model uses information from the future. However, the model was
used merely as a benchmark for the other models.
The second naive model, Naive (latest known), predicts the latest known
wait-time. This is one of the input features as described in Table 3.1. More
formally, the model was defined as:
(t)
ŷ (t+1) = xlatest_wait_time (3.2)

3.2.2 The Simulation Algorithm (Sim)

The simulation algorithm (Sim) implemented by the company executes peri-
odically for continuous predictions, as well as on certain triggers, for instance,
when a new patient joins the queue. The initial value that the simulation algo-
rithm predicts for each new patient is saved. The algorithm was therefore not
re-implemented, instead, the existing saved values were used. The simulation
algorithm has access to all schedules for all clinicians, and thus knows when
each clinician is working and when each clinician is on a break. Moreover, it
uses the information in the labels for both patients and clinicians during the
matching, to be able to mimic the queue as perfectly as possible. In order to
calculate an expected end time for meetings which have started, the algorithm
uses hard coded meeting length estimations. There is one estimation for each
type of meeting. The estimation for text meetings is much shorter than the
one for video meetings. These estimations are based on the average meeting
length for meetings of the same type, but are used as a management tool in
the company. For example, if the estimation is set to x minutes, all clinicians
will be informed that they are expected to spend on average x minutes per
consultation.
When the simulation runs, it moves time forward by iterating over all future
events and adds to said event list continuously as more events are scheduled.
32 CHAPTER 3. METHODOLOGY

Initially, the events consist of all the start and end times of the clinician shifts,
including their breaks, up to 7 days in to the future. Additionally, if any shift
has an ongoing meeting, the expected end time of said meeting is added as
an event. This end time is calculated based on the meeting length estimations
explained above. At the time for each event in the ordered event list, a check
is done if any ongoing meeting has ended, signifying that a clinician has be-
come available for matching. Furthermore, all scheduled shifts are searched
for shifts that have either ended or started at this point in time. Clinicians
whose shifts just ended are no longer considered for matching, and vice versa
for clinicians whose shifts just started. Next, the patients waiting in the queue
are matched with the currently available clinicians based on their labeling in-
formation. An event is created for each matched patient, signifying the end
time of their meeting. The matched patients are assigned wait-time estima-
tions calculated by taking the difference of the simulated matching time and
the actual current time. The algorithm is presented in pseudo code below in
Algorithm 1.
 CHAPTER 3. METHODOLOGY 33

Algorithm 1 Pseudo code for the simulation algorithm.

Input: queue - the queue of patients
shif ts - all scheduled clinician shifts
now - the current time
Output: waitT imes - a list of all patients from queue with assigned wait-time
estimations
1: function Simulation(queue, shif ts, now)
2: – initialization –
3: events ← empty ordered collection . Timestamps of events
4: ongoing ← empty collection . Shifts with ongoing meetings
5: available ← empty collection . Active shifts with no meeting
6: waitT imes ← empty collection . Patients with assigned wait-times
7: for all shifts s in shif ts do
8: Add start and end time of s to events
9: Add all start and end times of breaks in s to events
10: if s has ongoing meeting then
11: Add s to ongoing
12: Add expected end time of meeting to events
13: else if s is active now then
14: Add s to available
15: end if
16: end for
17: – main loop –
18: while events is not empty do
19: tick ← events.getAndRemoveF irst()
20: for all shifts s in ongoing do
21: m ← s.getOngoingM eeting()
22: if m expected end time ≤ tick then . m has ended
23: Move s to available
24: end if
25: end for
26: for all shifts s in shif ts do
27: if s’s end time ≤ tick then . s has ended
28: Remove from available
29: else if s started at tick then
30: Add s to available
31: end if
32: end for
33: Match patients from queue to applicable clinicians in available
34: Assign each matched patient a wait-time estimation (tick − now)
35: Add their expected meeting end times to events
36: Move matched patients from queue to waitT imes
37: Move matched clinician shifts from available to ongoing
38: end while
39: return waitT imes
40: end function
34 CHAPTER 3. METHODOLOGY

3.2.3 TCN
The TCN model was implemented using TensorFlow2 [58] and the Keras li-
brary [59]. The implementation of the TCN layer in [60] was used, together
with a final dense layer with linear activation. The TCN layer in [60] was
made following the specification from [15]. The sole difference is that the
weight normalization in the residual block was swapped for an optional batch
normalization [61]. Whether to use batch normalization or not was left as a hy-
perparameter. The other hyperparameters were kernel size, number of filters,
number of residual blocks (with exponentially increasing dilation), dropout
rate, and kernel initializer. The Adam Optimizer [62] from Keras was used
during training, with the default learning rate of 0.001. The model was trained
for a maximum of 100 epochs with a batch size of 64 and a sequence length
of 110. Early stopping was used with a patience of 20, i.e. the training was
terminated if the validation MSE did not decrease for 20 consecutive epochs.

3.2.4 LSTM
The LSTM model was implemented using TensorFlow2 [58] and the Keras
library [59]. The LSTM layer from Keras was used together with a dense
layer with linear activation. All hyperparameters to the layer were set to the
default values except for the number of units, kernel initializer, dropout rate,
and recurrent dropout rate. Identically to the TCN, the LSTM was trained for
a maximum of 100 epochs using: the Adam Optimizer [62] with the default
learning rate of 0.001, early stopping with a patience of 20, a batch size of 64,
and a sequence length of 110.

3.2.5 RF
The Scikit-learn implementation RandomForestRegressor [48] was used for
the RF model. All hyperparameters were left to their default values except
n_estimators (for the number of trees), max depth, and max features (maxi-
mum features considered at each split).

3.2.6 SVR
The GPU accelerated implementation of the SVR in the cuML library from
the open source project Rapids [63] was used. The hyperparameters were left
to default except for the type of kernel, which degree for polynomial kernels,
amount of slack C allowed, and epsilon.
 CHAPTER 3. METHODOLOGY 35

3.2.7 Combination Model (Comb)

A final combination model (Comb) was created which was a combination of
the simulation model and the best performing ML model. This was done by
simply appending the simulation predictions to the input vector of the ML
model as an additional feature.

3.3 Hyperparameter Optimization

For each ML model a hyperparameter optimization was made. This was done
using the library Hyperopt [64]. This library applies Bayesian Optimization,
a probabilistic model, in order to find the hyperparameters that minimizes the
validation loss by using a Tree-structured Parzen Estimator algorithm [65].
This has proved to be better than a random search, and it reduces the number
of search iterations required by taking past iterations into account when start-
ing new ones [64]. The details of this algorithm are described in [65]. The
search spaces for all hyperparameters for all models are defined in Table 3.2,
Table 3.3, Table 3.4, and in Table 3.5. For each model 100 trials (search iter-
ations) were executed.
In order to speed up the hyperparameter optimization process, all models
where trained on 40% percent of the training data (the most recent, i.e. the last
in chronological order). The TCN and LSTM were trained for 4 epochs each
trial, and the SVR had a limit on the maximum iterations made in the fitting
algorithm of 20000.

Parameter Search Space

kernel size [2, 6]
filters [32, 192]
blocks [2, 6]
dropout rate [0.0, 0.5]
kernel initializer [he_normal, he_uniform]
use batch normalization [False, True]

Table 3.2: TCN hyperparameter search space. The maximum amount of train-
able parameters is 2845633, and the maximum receptive field is 631.
36 CHAPTER 3. METHODOLOGY

Parameter Search Space

units [32, 832]
kernel initializer [glorot_normal, glorot_uniform]
dropout rate [0.0, 0.5]
recurrent dropout rate [0.0, 0.5]

Table 3.3: LSTM hyperparameter search space. The maximum amount of

trainable parameters is 2832961.

Parameter Search Space

n_estimators [15, 512]
max depth [0, 128]
max features [1, n]
alpha [0.2, 1]
only smooth [False, True]

Table 3.4: RF hyperparameter search space. If only smooth was set to false,
the smoothened representations of all features would be appended to the input
vector instead of replacing it. Therefore, n = 18 or n = 31, depending on the
value of only smooth.

Parameter Search Space

kernel [linear, poly, rbf]
degree [2, 4]
C eN (0,1)
epsilon [0.0, 1.0]
alpha [0.2, 1]
only smooth [False, True]

Table 3.5: SVR hyperparameter search space, where N (0, 1) denotes the nor-
mal distribution with µ = 0 and σ = 1. The degree parameter was only used
for the polynomial kernel. If only smooth was set to false, the smoothened
representations of all features would be appended to the input vector instead
of replacing it.
 CHAPTER 3. METHODOLOGY 37

3.4 Evaluation
In order to reduce the variance and effect of randomness for each ML model,
5 separate models were trained and combined into a final model by averaging
their separate predictions. This does not apply to the SVR model, since its
training process is deterministic, i.e. does not include randomness.
To determine if there was a statistically significant difference in the er-
rors of any model, a Friedman test [66] was done.. The Friedman test is
the non-parametric version of a analysis of variance test (ANOVA). The non-
parametric version was used because the errors can not be assumed to be nor-
mally distributed. The null hypothesis was that all errors had the same distri-
bution. The hypothesis was tested by picking a random sample of N = 25000
patients from the test data, which total size was 75600. The squared errors for
the predictions from each final model were calculated for each patient in the
sample. The Friedman test implementation from the SciPy library [67] was
then used for testing said error samples. The significance level was set to 0.01.
The Friedman test was followed by a post-hoc pairwise comparison us-
ing the non-parametric version on the dependent (paired) t-test which is the
Wilcoxon signed-rank test [68]. This was done to determine between which
models there was a statistical significant difference in the errors. The null
hypothesis tested was that the two paired samples come from the same distri-
bution. Similar to the Friedman test, this was done by sampling N = 25000
patients from the test data, calculating the square errors in the predictions for
each models, and then using the Wilcoxon singed-rank test implementation
from the SciPy library [67] with a significance level of 0.01.
To measure the performance of each final model their MSE and R2 was
calculated on the test data. The MSE is the most commonly used loss function
for regression, and it is calculated as:
n
1X
MSE = (ŷi − yi )2 (3.3)
n i

The R2 score ranges between 0 and 1, and is calculated as:

Pn
(ŷi − ȳ)2
R2 = Pin 2
(3.4)
i (yi − ȳ)

where ȳ is the mean of the observed data.

Chapter 4

Results

This chapter starts by presenting the results of the hyperparameter optimiza-

tion for each ML model. Secondly, it presents the result of the evaluation of
all models. This includes the MSE, R2 , training times and the results of the
statistical tests. Lastly, a selection of forecast values are displayed for each
model.

4.1 Hyperparameter Optimization

The following sections present the hyperparameters used for the final ML mod-
els referenced in the rest of the paper. The hyperparameters chosen were the
ones from the model that achieved the lowest validation loss out of the 100
different trial models.

4.1.1 TCN
The hyperparameters with the lowest validation loss are displayed in Table 4.1.
The lowest validation loss was 52.502 and the average s/trial were 295.363.
This yields a receptive field of 33 according to Equation 2.6.

kernel size filters blocks dropout rate kernel initializer batch norm
3 64 3 0.0637 he_normal False

Table 4.1: The hyperparameters used for the TCN model.

38
 CHAPTER 4. RESULTS 39

4.1.2 LSTM
Table 4.2 presents the hyperparameters for the LSTM model which resulted in
the lowest validation loss of 50.470. The average s/trial were 2391.482.

dropout rate kernel initializer recurrent dropout rate units

0.0457 glorot_normal 0.235 384

Table 4.2: The hyperparameters used for the LSTM model.

4.1.3 RF
The hyperparameters that received the lowest validation loss of 51.227 are
displayed in Table 4.3. The RF model took on average 140.498 s/trial.

n_estimators max depth max features alpha only smooth

448 16 12 0.364 False

Table 4.3: The hyperparameters used for the RF model.

4.1.4 SVR
The SVR model with the lowest validation loss of 59.021 had the parameters
displayed in Table 4.4. The SVR model took on average 216.340 s/trial.

kernel C epsilon alpha only smooth

rbf 339.705 0.968 0.201 False

Table 4.4: The hyperparameters used for the SVR model.

4.2 Evaluation
The p-value produced by the Friedman test on a sample of the test data with
a size of 25000 was 0.0, signifying that at least one model does not have the
same distribution in errors compared to the rest on a significance level of 0.01.
The resulting p-values of the post-hoc pairwise comparison using the Wilcoxon
singed-rank test is displayed in Table 4.5. The table shows that there is a sig-
nificant difference in the error distribution between all pairs of models with a
significance level of 0.01.
40 CHAPTER 4. RESULTS

Model LSTM TCN RF SVR Comb

TCN 1.3×10−5 - - - -
RF 2.2×10−80 7.9×10−69 - - -
SVR 1.9×10−272 1.5×10−255 1.8×10−84 - -
Comb 4.1×10−17 1.2×10−23 6.7×10−95 2.0×10−260 -
Sim 1.3×10−3 6.7×10−4 1.0×10−17 7.4×10−55 5.2×10−63

Table 4.5: The p-value results from the post-hoc pairwise Wilcoxon signed-
rank test. Sample size N = 25000.

The MSE and R2 for each of the final models are displayed in Table 4.6,
and the training times are displayed in Table 4.7. The TCN model achieved
the lowest MSE out of all standard models. Therefore, this model was used
for the combination model (Comb).

Model MSE R2
Naive (previous) 28.8 0.935
Comb 38.3 0.914
TCN 56.2 0.873
Simulation 56.6 0.872
LSTM 61.1 0.862
RF 70.3 0.841
SVR 78.6 0.823
Naive (latest known) 102.7 0.768

Table 4.6: MSE and R2 for each model sorted on MSE. The Naive (previous)
model is not possible to implement in a real scenario, because the previous
patient has most likely not been assigned a clinician when the next patient
arrives.

The Naive models were implemented as baselines to get a better under-

standing of the performance of the other models. The Naive (previous) model
is not possible to implement in a real scenario due to the fact that it uses infor-
mation from the future. However, it provides a benchmark MSE to the other
models. Table 4.6 therefore shows that a model achieving an MSE close to
28.8 can be considered a high performing model. The Naive (latest known)
model provides another benchmark MSE of 102.7. Models achieving MSEs
close to this value have practically not learned anything, since the latest known
patient wait-time is one of the input features.
The Comb model achieved the lowest MSE of all models, not including
 CHAPTER 4. RESULTS 41

the Naive (previous) model. All ML models and the Simulation model out-
performed the Naive (latest known) baseline model.
The models in the non-sequence modeling ML approach (RF and SVR)
achieved substantially higher MSEs and lower R2 scores compared to the mod-
els in the sequence modeling approach (LSTM and TCN), as can be seen in
Table 4.6. Since the statistical tests showed that there was a statistically sig-
nificant difference between all said models, this provides evidence that the se-
quence modeling approach is superior to the non-sequence modeling approach
at predicting wait-time in this specific digital healthcare setting.
The TCN is significantly faster at training compared to the LSTM. The
non-sequence modeling ML models were significantly faster at training com-
pared to the Neural Network models.

Model training time (s) s/epoch

SVR 225 -
RF 801 -
TCN 1920 52
Comb 1947 52
LSTM 48900 1407

Table 4.7: The average training time and seconds per epoch form all ML mod-
els. The RF and SVR models are not trained in epochs.

4.3 Forecasting
Figure 4.1 and Figure 4.2 display the forecasts from each model from 2 ran-
domly selected days in the test data (not including the Naive (previous) model).
It should be noted that these days only consist of a small part of the test set,
which contained a total of 55 days, and might not represent the whole data set.
More forecast graphs can be found in Appendix A.
As one can see in the figures, the ML models tend to react slower to changes
in the wait-time. The simulation algorithm however seem to overestimate the
changes and predict a higher wait-time than necessary during the peak hours.
In Figure 4.1, the combination model do not seem to suffer from any of the
problems mentioned. Whilst, in Figure 4.2, it seems to have both problems,
but in a much smaller scale.
The simulation seems to be able to handle special cases better than the ML
models. An example of this can be seen at approximately 17:45 in Figure 4.1,
where the wait-time has 4 distinct dips. This happens to be 3 child patients
42 CHAPTER 4. RESULTS

and one patient renewing his/her prescription. These are all considered special
cases as explained in subsection 3.1.1. The simulation is able to assign a much
lower wait-time to these patients, but the ML models does not seem to have
learned this feature in the queue system, even though the input features include
both age and symptom (prescription renewal is considered a symptom).
The combination model seems to have the benefit of fast reactions and spe-
cial case detection from the simulation outputs and the benefit of not overes-
timating spikes in the wait-time from the ML model. This can be seen clearly
in both Figure 4.1 and Figure 4.2.
 CHAPTER 4. RESULTS 43

Figure 4.1: The actual wait-time and the predicted wait-time for each model
for single day in the test data (2019-12-30).
44 CHAPTER 4. RESULTS

Figure 4.2: The actual wait-time and the predicted wait-time for each model
for single day in the test data (2020-01-20).
Chapter 5

Discussion

In this chapter a discussion regarding the results and their implications is held.
The performances of the models are compared according to the measured met-
rics, and other aspects of the different models are brought to light. The forecast
graphs are analyzed and evaluated in detail. The limitations in the data used
and the methodology are underlined, the impact of the work is discussed and
finally future research is suggested.

5.1 Model Comparison

The sequence modeling approach was superior to the non-sequence model-
ing approach at predicting wait-time in this specific digital healthcare setting.
The reason for this is believed to be that introducing the temporal dimension
with sequences provides a better structure to the data which is easier to learn,
compared to the exponential smoothing technique. While utilizing exponen-
tial smoothing did improve the performance of the models compared to not
introducing the temporal dimension at all, it did not improve the performance
as much as the sequencing. This raises the question regarding why no other
research applying the more complex models from sequence modeling to the
wait-time problem can be identified. One theory is that the implicit require-
ment of uniformed sampled data has been considered a bigger hindrance than
it seems to be. Another reason could be the complexity which is introduced
with neural networks.
The TCN achieved a lower MSE compared to the simulation algorithm,
but the LSTM model did not. Therefore, no claim of superiority can be made
for either approach. In order to outperform the simulation algorithm, the ML
models have to learn everything the simulation algorithm is given, and ad-

45
46 CHAPTER 5. DISCUSSION

ditionally learn or pick up on some other pattern unknown to the simulation

algorithm. This is believed to be the reason as to why not all ML models did
outperformed the simulation algorithm. The simulation algorithm has access
to the schedules of all clinicians and the labels of said clinicians and all pa-
tients. This labeling information is very specific to the queuing system of the
company. Therefore, the ML models were not given this information in an
attempt to make the research conducted in this thesis more widely applicable.
Yet, the TCN managed to achieve a MSE lower than the simulation algorithm,
which is considered to be a notable result. The TCN having better performance
compared to the LSTM is aligned with the research in [15], [20], [21].
The combination model achieved the lowest MSE and highest R2 out of all
models. This is believed to be due to the fact that this model does not have to
learn to predict wait-time, rather it only have to learn to predict and prevent
the mistakes the simulation algorithm makes. This model has access to more
information compared to the other ML models, and it is therefore no surprise
it has the best performance. There is only a difference of 9.4 in the MSE be-
tween the combination model and the Naive (previous) baseline model, which
indicates that the performance of the combination model is impressive. A
downside to consider for the combination model is that it will require the most
maintenance. Both the simulation and the ML model (TCN) will have to be
updated in the event of any major changes to the queue system.
The RF model outperformed the SVR model. This was expected, as it has
proven to be the best model for wait-time prediction in numerous other studies
[5], [9], [13].
The non-sequence modeling ML models required considerably less train-
ing time compared to the sequence modeling models. Lower training time
means easier and faster maintenance of the models, which can be of impor-
tance if they were to be commercially used. The TCN was also significantly
faster to train compared to the LSTM. This was expected because of the par-
allelizable structure of the TCN, and aligned with the results in other research
[15], [20]. However, computational power has become cheaper and cheaper
with the years, which might mitigate or even eliminate the problem of long
training times.
Interpretability can be an important factor when using any prediction model
commercially. This is a problem for all ML models in this paper, except the
statistical RF model, and it is known as the Black Box problem. The models
have virtually no interpretability and can therefore be regarded as black boxes,
which receive an input and produces an output with no explanation as to why
said output was produced. If a customer receives an unexpected wait-time
 CHAPTER 5. DISCUSSION 47

prediction, it might be of interest to the company to figure out the reason as to

why, both to be able to provide an explanation to a potentially angry customer,
and to be able to prevent such a scenario from re-occurring. Therefore, there
is an advantage to using a simulation algorithm or the statistical RF model.
The simulation algorithm can be re-run with the same inputs, and be analyzed
step by step in order to exactly explain the reason behind the given wait-time
estimation. The RF model is less interpretable compared to the simulation
since it consists of several trees. However, the trees can still be unraveled and
the questions they contain reveled in order to gain some understanding as to
why a certain wait-time prediction was given.

5.1.1 Forecasts
A pattern can be noticed in the forecast behaviors of all ML models compared
to the simulation. The ML models tend to have slower reactions to changes
in the wait-time, which causes them to lag behind the curve a bit. This is
believed to be due to the latest known wait-time feature, which essentially is
the wait-time curve with a time delay. The ML models have to learn how the
wait-time has changed since the latest known wait-time, and if they are unable
to predict any change, the resulting forecasts will follow this time delayed curve
to the point, which is exactly what the Naive (latest known) model does. The
simulation, on the other hand, has faster reactions but tend to overestimate the
spikes in the wait-time. The wait-time is affected by several different factors,
for instance, at peak hours the working clinicians can potentially take the long
queue into account and start working more effectively as a countermeasure.
This is something the simulation is not able to consider, and might be the
reason for the overestimations. The ML models can potentially learn the effect
of such factors, if they are represented in the training data.
The combination model adopted the best traits from both of the approaches
from which it is combined. This is believed to be one of the reasons behind
superior performance of the combination model.

5.2 Limitations
5.2.1 Model Selection
It’s important to note that the research conducted in this paper has its limita-
tions. First and foremost only two different models were compared for each
48 CHAPTER 5. DISCUSSION

ML approach, two sequence modeling models and two non-sequence model-

ing ML models. These were chosen based on their performance according to
the literature study. There exist several other models which might have bet-
ter performance on this specific problem. The linear regression / Elastic Net
models also showed some potential in the literature study, but was left out of
this thesis. As the hyperparameter optimization of the SVR model deemed
the rbf kernel the best one, not the linear kernel, and due to the fact that any
Elastic Net can be reduced to a linear SVC, an argument can be made that a
linear model would not be better than the two non-linear models implemented
in this thesis.
Due to time and computational limitations the experiment conducted had
to be limited to four ML models, but to be able to conclude with more certainty
that sequence modeling is the superior approach of the two more investigation
might be required.

5.2.2 Temporal Dimension

Due to limitation in computing power and time, the sequence lengths of the
sequence modeling models was set to a fixed number of 110. This thesis did
not investigate the impact sequence length has on model performance. Since
the receptive field of the final TCN model was 33, it could be interesting to
examine what amount of previous patients actually determine the wait-time
of a new patient. With a lower sequence length less information is provided,
however, the training time will also be reduced.
Moreover, this thesis implemented a single method of introducing the tem-
poral dimension to the non-sequence modeling models, namely the exponen-
tial smoothing method. Other related research identified have not attempted
to introduce the temporal dimension at all. There exist other, more sophis-
ticated, methods of doing this, for example Kalman Filters [69]. However,
due to time limitations other methods were not explored. This is considered a
limitation of the work conducted in this thesis, and the several ways in which
the non-sequence modeling models can be improved further are left as future
research.

5.2.3 Data
In order to make the data collected useful for the ML models, some cleaning
had to be done. This cleaning removed certain patients due to missing or un-
realistic data values. Said patients might have been crucial in explaining the
 CHAPTER 5. DISCUSSION 49

behavior of the wait-time, and their removal can thus have affected the perfor-
mance of all ML models in a negative way.
The ML models require a lot of data, the more the better. However, due
to the pandemic caused by the virus Covid-19 in the beginning of 2020, some
data was deemed non usable. The company experienced a disproportionate
increase of patients per day compared to earlier years, which meant that no
data from patients seeking help past the date of 2020-02-17 could be used.
Moreover, data from a long time ago might not reflect the behavior of the
queue and wait-time more recently. For that reason and because of a change in
the structure of the data, no data from earlier than 2018-04-22 could be used.
The chronological split of the data into training, validation and test sets
could be a source of interference for the ML models as well. This splitting
caused both Christmas (2019-12-24) and new years (2020-01-01), two signif-
icant holidays, to be within the test set. The consumer behavior on said days
differs a lot from other regular days. This could have had an impact on the ML
models performance on the test set.
Another significant limitation in the data is that it is collected from a sin-
gle digital healthcare source. The trends in the data, and therefore the results
of this thesis, might differ for other digital healthcare providers, or in other
markets.
The features used to represent the patients in the data were chosen based on
previous research and availability, but also on reproducibility. That means that
features that were too specific to the queuing system in place at the company
were not used, for instance, clinician and patient labels. However, the work in
this thesis does not contain any analysis of said features. It could be interesting
to investigate how much each feature contributes to the predictions. Some of
the features used may only be unwanted noise. This is something that was not
considered due to the time limitation of the study.

5.2.4 Hyperparameter Optimization

The hyperparameter optimization process of the methodology in this thesis
was affected by the time and computational limitations quite heavily. Because
of the time required to train the LSTM model, only a small part of the training
data (40%) could be used during training for each trial. This might have caused
the final hyperparameters to have a bias towards this small part of the data
which might not represent the rest of the data accurately. Moreover, the models
were only trained for 4 epochs for each trial. This was due to the fact that, as
presented in subsection 4.1.2, the average number of seconds for a single trial
50 CHAPTER 5. DISCUSSION

for the LSTM was 2391. As 100 trials were done, the total time required for
the hyperparameter optimization for the LSTM was around 66 hours.
Not only was the amount of data limited but also the actual hyperparam-
eters and their search space as well. The batch size and, as mentioned previ-
ously, the sequence length was set to a fixed number even though they might
have a considerable impact on the training and performance of the models.
The range for all hyperparameters in the search space were also chosen with
the time limitations in consideration. This might have affected the result of the
hyperparameter optimization in a way that if a larger search space was cho-
sen, another minima might have been found. The current search space might
have caused the search to end up in a local minima, when a better global min-
ima could exist. In an attempt to make the hyperparameter optimization as
fair as possible between the TCN and LSTM models, the search space was
constructed in a way where the maximum possible size of the two models,
measured in number of trainable weights, was roughly the same.
The hyperparameters deemed to be the best for the TCN were also used
for the TCN in the combination model. Since the combination model has a
new input (the simulation output), the best hyperparameters for the regular
TCN might not be best in this case. But due to time limitations, the assump-
tion that the hyperparameters would also be a good choice for the TCN in the
combination model had to be made.

5.3 Social Aspects and Usability

As mentioned previously, the accuracy of wait-time prediction is especially
important in healthcare, since the patients are likely to already be in a wor-
ried state of mind when seeking help. An accurate wait-time could calm the
patients down. Therefore, the research conducted in this paper will hopefully
have a positive societal impact, since it takes steps towards increasing said
accuracy. Furthermore, with an accurate wait-time prediction model, steps
can potentially be taken to reduce the wait-time when the predictions indicate
high values. For instance, to schedule more clinicians. This is left as future
research, but could lead to a better and more accessible healthcare.
More precise predictions can be crucial for patients in a critical condition.
The patient might weigh the option of calling for an ambulance against the
option of digital healthcare, and might not be in a state of mind to make the
best decision. If presented with an accurate wait-time prediction the patient
can make a better educated decision. On the other hand, the predictions are
just predictions. Meaning, the wait-time can be affected by several unforeseen
 CHAPTER 5. DISCUSSION 51

variables. If a ML model is used and marketed as a model with high accuracy,

the patient in a critical condition might place too much trust in the model,
and suffer more damages than if he/she would have called the emergency ser-
vices immediately. This, as well as the Black Box problem mentioned last
in section 5.1, should be taken into consideration when using a ML model
commercially in healthcare.
Based on the findings in this thesis, the sequence modeling approach using
time series forecasting should be considered a viable approach to the wait-time
prediction problem. Since the features used to describe the queue were rela-
tively general, the time series forecasting technique is assumed to be viable not
only for healthcare queues, but for any type of queue. Therefore, a sequence
modeling approach could potentially replace a traditional ML approach ap-
plied to any queue, for an improved performance. Additionally, a sequence
modeling approach could potentially be combined with any existing simula-
tion algorithm to boost its performance.

5.4 Sustainability and Ethics

The data used in this study could be considered highly sensitive, as it includes
personal information about patients. However, the company has anonymized
the data before it was accessed in this study. This was done in such a way that
any data point could not be tied back to a specific patient. Even though this
was the case, the data is still considered sensitive. This has been taken into
account, and an effort has been made to not disclose any potentially damaging
information, while still striving to have as much transparency as possible in
regards to the research conducted.
Worth mentioning is the recent increase in research regarding the energy
consumption of deep learning models. The training process for deep learning
models can be costly, both in regards to hardware and computational power,
but also in regards to the environment [70]. The primary focus in ML is usu-
ally the accuracy and performance of the model, but recent research raises the
importance of environmental sustainability, and having energy consumption
as an evaluation metric as well [71].

5.5 Future Research

As mentioned in subsection 3.1.2, the sequence modeling models implicitly re-
quire the data to be uniformly sampled. The simplest solution to this problem
52 CHAPTER 5. DISCUSSION

was applied in this thesis, appending the time differences between each patient
∆t to the input vector. However, research has shown that there are other, better,
solutions to this problem. More specifically [72] introduced the Phased-LSTM
which extends the regular LSTM by adding a new time gate, which results in
a greatly improved performance in standard RNN applications. Very recently,
similar research has been done for the TCN in [57], which introduces a Con-
tinuous Convolutional Neural Network for non-uniform time series showing
promising results. Future research is needed in order to investigate whether
the models mentioned above could further improve the performance of the
sequence modeling approach on the wait-time prediction problem.
The behavior and trends in the wait-time changes with time. As mentioned
in subsection 5.2.3, data from a long time ago might not contribute as much to
the performance of the ML models as more recent. It might just add unwanted
noise. Future research is required to evaluate if any attempt to mitigate this
problem will provide an increase in performance for the ML models. For in-
stance, one could use sample weights in order to weight the data from more
recent patients more heavily so that they contribute more during training. Fur-
thermore, investigating other, more representative, evaluation techniques for
the ML models could be of interest. Splitting the data in chronological order
will cause the ML models to be evaluated on the latest trend which might not
be included in the training data. In order to alleviate this affect one could, for
instance, evaluate the ML models on one week at a time, allowing them to
train on the data from the previous week before evaluating the next.
Chapter 6

Conclusion

To summarize, the result of the comparison between a sequence modeling ap-

proach, a non-sequence modeling ML approach utilizing exponential smooth-
ing, and a simulation algorithm, was that the sequence modeling TCN model
and the simulation algorithm had almost equal performances on wait-time pre-
diction in digital healthcare. The two models had an MSE around 56, and they
had their respective advantages and disadvantages. When combined together,
the best performing model was attained. The combination model adopted the
advantages from both approaches. The non-sequence modeling approach uti-
lizing exponential smoothing had a considerably worse performance compared
to the other approaches, but excelled in other aspects such as interpretability.
No single approach could be concluded as optimal, due to the limitations in the
conducted research. However, it can be concluded that the sequence modeling
approach is a viable option, and it should therefore be considered in future
wait-time prediction research.
Several other suggestions for future research are proposed. For instance,
investigating whether alternatives to exponential smoothing are better at intro-
ducing the temporal dimension to the non-sequence modeling ML approach.
Alternatively, investigating whether extensions of the sequence modeling mod-
els, designed specifically for non-uniformly sampled data, will have increased
performance.

53
Bibliography

[1] D. H. Maister et al., The psychology of waiting lines. Harvard Business

School Boston, MA, 1984.
[2] M. K. Brady and J. J. Cronin Jr, “Some new thoughts on conceptual-
izing perceived service quality: A hierarchical approach”, Journal of
marketing, vol. 65, no. 3, pp. 34–49, 2001.
[3] R. Davis, T. Rogers, and Y. Huang, “A survey of recent developments
in queue wait time forecasting methods”, in Proceedings of the Inter-
national Conference on Foundations of Computer Science (FCS), The
Steering Committee of The World Congress in Computer Science, Com-
puter . . ., 2016, p. 84.
[4] A. Komashie, A. Mousavi, P. J. Clarkson, and T. Young, “An integrated
model of patient and staff satisfaction using queuing theory”, IEEE
Journal of Translational Engineering in Health and Medicine, vol. 3,
pp. 1–10, 2015, issn: 2168-2372. doi: 10 . 1109 / JTEHM . 2015 .
2400436.
[5] A. Joseph, T. Hijal, J. Kildea, L. Hendren, and D. Herrera, “Predicting
waiting times in radiation oncology using machine learning”, eng, in
2017 16th IEEE International Conference on Machine Learning and
Applications (ICMLA), vol. 2018-, IEEE, 2017, pp. 1024–1029, isbn:
9781538614174.
[6] Kry, https://www.kry.se/, Accessed: 2020-03-30.
[7] G. Arha, “Reducing wait time prediction in hospital emergency room:
Lean analysis using a random forest model”, Master’s thesis, University
of Tennessee, Knoxville, 2017.
[8] M. Kembe, E. Onah, and S. Iorkegh, “A study of waiting and service
costs of a multi-server queuing model in a specialist hospital”, Interna-
tional Journal of Scientific & Technology Research, vol. 1, no. 8, pp. 19–
23, 2012.

54
 BIBLIOGRAPHY 55

[9] C. Curtis, C. Liu, T. J. Bollerman, and O. S. Pianykh, “Machine learning

for predicting patient wait times and appointment delays”, Journal of
the American College of Radiology, vol. 15, no. 9, pp. 1310–1316, 2018.
[10] A. Kyritsis and M. Deriaz, “A machine learning approach to waiting
time prediction in queueing scenarios”, Sep. 2019, pp. 17–21. doi: 10.
1109/AI4I46381.2019.00013.
[11] “Wait-time predictors for customer service systems with time-varying
demand and capacity”, eng, Operations Research, vol. 59, no. 5, pp. 1106–
1118, 2011, issn: 0030-364X.
[12] R. Mourao, R. Carvalho, R. Carvalho, and G. Ramos, “Predicting wait-
ing time overflow on bank teller queues”, eng, in 2017 16th IEEE Inter-
national Conference on Machine Learning and Applications (ICMLA),
vol. 2018-, IEEE, 2017, pp. 842–847, isbn: 9781538614174.
[13] Y. Sanit-in and K. R. Saikaew, “Prediction of waiting time in one-stop
service”, International Journal of Machine Learning and Computing,
vol. 9, no. 3, 2019.
[14] L. Breiman, “Random forests”, eng, Machine Learning, vol. 45, no. 1,
pp. 5–32, 2001, issn: 0885-6125.
[15] S. Bai, J. Z. Kolter, and V. Koltun, “An empirical evaluation of generic
convolutional and recurrent networks for sequence modeling”, CoRR,
vol. abs/1803.01271, 2018. arXiv: 1803.01271. [Online]. Available:
http://arxiv.org/abs/1803.01271.
[16] S. Hochreiter and J. Schmidhuber, “Long short-term memory”, eng,
Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997, issn: 0899-
7667.
[17] Y. Zhang, L. T. Nguyen, and J. Zhang, “Wait time prediction: How
to avoid waiting in lines?”, in Proceedings of the 2013 ACM confer-
ence on Pervasive and ubiquitous computing adjunct publication, 2013,
pp. 481–490.
[18] R. G. Brown, Statistical forecasting for inventory control. McGraw-
Hill, 1959. [Online]. Available: http : / / hdl . handle . net /
2027/mdp.39015001311771.
[19] J. Gehring, M. Auli, D. Grangier, D. Yarats, and Y. N. Dauphin, “Con-
volutional sequence to sequence learning”, CoRR, vol. abs/1705.03122,
2017. arXiv: 1705.03122. [Online]. Available: http://arxiv.
org/abs/1705.03122.
56 BIBLIOGRAPHY

[20] J. Berglind, Temporal convolutional networks for forecasting patient

volumes in digital healthcare, eng, KTH, Skolan för elektroteknik och
datavetenskap (EECS), 2019.
[21] R. Wan, S. Mei, J. Wang, M. Liu, and F. Yang, “Multivariate tempo-
ral convolutional network: A deep neural networks approach for multi-
variate time series forecasting”, eng, Electronics, vol. 8, no. 8, p. 876,
2019, issn: 2079-9292. [Online]. Available: https://doaj.org/
article/5d25208427d54aa1ba99c368c8f6d6e1.
[22] M. Pickup, Introduction to time series analysis, eng, ser. Quantitative
applications Introduction to time series analysis. 2016, isbn: 1-4833-
9085-3.
[23] S. Siami-Namini and A. S. Namin, “Forecasting economics and finan-
cial time series: ARIMA vs. LSTM”, CoRR, vol. abs/1803.06386, 2018.
arXiv: 1803.06386. [Online]. Available: http://arxiv.org/
abs/1803.06386.
[24] R. Sen, H.-F. Yu, and I. Dhillon, “Think globally, act locally: A deep
neural network approach to high-dimensional time series forecasting”,
arXiv preprint arXiv:1905.03806, 2019.
[25] S. d. O. Domingos, J. F. de Oliveira, and P. S. de Mattos Neto, “An intel-
ligent hybridization of arima with machine learning models for time se-
ries forecasting”, Knowledge-Based Systems, vol. 175, pp. 72–86, 2019.
[26] A. Rubio, J. Bermúdez, and E. Vercher, “Improving stock index fore-
casts by using a new weighted fuzzy-trend time series method”, eng,
Expert Systems with Applications, vol. 76, p. 12, 2017, issn: 0957-
4174. [Online]. Available: http : / / search . proquest . com /
docview/1932176199/.
[27] C. Ng and B. Soong, “Single-queue markovian systems”, eng, in Queue-
ing Modelling Fundamentals, Chichester, UK: John Wiley Sons, Ltd,
2008, pp. 103–140, isbn: 9780470519578.
[28] G.-S. Hu and F.-Q. Deng, “The analysis of queuing system based on
support vector machine”, eng, in ICARCV 2004 8th Control, Automa-
tion, Robotics and Vision Conference, 2004, vol. 3, IEEE, 2004, 2320–
2325 Vol. 3, isbn: 0780386531.
[29] R. G. Brown, Exponential smoothing for predicting demand, INST OP-
ERATIONS RESEARCH MANAGEMENT SCIENCES, 1957. [On-
line]. Available: https://www.industrydocuments.ucsf.
edu/docs/jzlc0130.
 BIBLIOGRAPHY 57

[30] A. Azadeh, M. Sheikhalishahi, M. Tabesh, and A. Negahban, “The ef-

fects of pre-processing methods on forecasting improvement of artificial
neural networks”, Australian Journal of Basic and Applied Sciences,
vol. 5, no. 6, pp. 570–580, 2011.
[31] R. Hyndman, Forecasting with Exponential Smoothing The State Space
Approach, eng, 1st ed. 2008.., ser. Springer Series in Statistics. 2008,
isbn: 1-281-49162-4.
[32] W. McCulloch and W. Pitts, “A logical calculus of the ideas imma-
nent in nervous activity”, eng, The bulletin of mathematical biophysics,
vol. 5, no. 4, pp. 115–133, 1943, issn: 0007-4985.
[33] R. Rojas, Neural networks: a systematic introduction. Springer Science
& Business Media, 2013.
[34] F. Rosenblatt, “The perceptron: A probabilistic model for information
storage and organization in the brain.”, Psychological review, vol. 65,
no. 6, p. 386, 1958.
[35] H. J. Kelley, “Gradient theory of optimal flight paths”, Ars Journal,
vol. 30, no. 10, pp. 947–954, 1960.
[36] P. Werbos, “Beyond regression: New tools for prediction and analysis in
the behavioral science. thesis (ph. d.). appl. math. harvard university”,
PhD thesis, Jan. 1974.
[37] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning represen-
tations by back-propagating errors”, Nature, vol. 323, no. 6088, p. 533,
1986, issn: 0028-0836.
[38] Y. LeCun et al., “Generalization and network design strategies”, Con-
nectionism in perspective, vol. 19, pp. 143–155, 1989.
[39] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press,
2016, http://www.deeplearningbook.org.
[40] A. v. d. Oord, S. Dieleman, H. Zen, et al., “Wavenet: A generative model
for raw audio”, arXiv preprint arXiv:1609.03499, 2016.
[41] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition”, 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), Jun. 2016. doi: 10.1109/cvpr.2016.90.
[Online]. Available: http://dx.doi.org/10.1109/CVPR.
2016.90.
58 BIBLIOGRAPHY

[42] T. Salimans and D. P. Kingma, Weight normalization: A simple repa-

rameterization to accelerate training of deep neural networks, 2016.
arXiv: 1602.07868 [cs.LG].
[43] V. Nair and G. E. Hinton, “Rectified linear units improve restricted
boltzmann machines”, in Proceedings of the 27th international confer-
ence on machine learning (ICML-10), 2010, pp. 807–814.
[44] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut-
dinov, “Dropout: A simple way to prevent neural networks from overfit-
ting”, The journal of machine learning research, vol. 15, no. 1, pp. 1929–
1958, 2014.
[45] Z. C. Lipton, “A critical review of recurrent neural networks for se-
quence learning”, CoRR, vol. abs/1506.00019, 2015. arXiv: 1506 .
00019. [Online]. Available: http://arxiv.org/abs/1506.
00019.
[46] L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen, Classification
and regression trees. CRC press, 1984.
[47] G. James, D. Witten, T. Hastie, and R. Tibshirani, An introduction to
statistical learning. Springer, 2013, vol. 112.
[48] F. Pedregosa, G. Varoquaux, A. Gramfort, et al., “Scikit-learn: Machine
learning in Python”, Journal of Machine Learning Research, vol. 12,
pp. 2825–2830, 2011.
[49] Q. Zhou, W. Chen, S. Song, J. R. Gardner, K. Q. Weinberger, and Y.
Chen, “A reduction of the elastic net to support vector machines with
an application to gpu computing”, ArXiv, vol. abs/1409.1976, 2015.
[50] H. Zou and T. Hastie, “Regularization and variable selection via the
elastic net”, Journal of the royal statistical society: series B (statistical
methodology), vol. 67, no. 2, pp. 301–320, 2005.
[51] I. Steinwart, Support Vector Machines, eng, 1st ed. 2008.., ser. Informa-
tion science and statistics. 2008, isbn: 1-281-92704-X.
[52] C. Cortes and V. Vapnik, “Support-vector networks”, eng, Machine Learn-
ing, vol. 20, no. 3, pp. 273–297, 1995, issn: 0885-6125.
[53] A. Smola and B. Schölkopf, “A tutorial on support vector regression”,
eng, Statistics and Computing, vol. 14, no. 3, pp. 199–222, 2004, issn:
0960-3174.
 BIBLIOGRAPHY 59

[54] F.-K. Wang and T. Mamo, “A hybrid model based on support vector re-
gression and differential evolution for remaining useful lifetime predic-
tion of lithium-ion batteries”, eng, Journal of Power Sources, vol. 401,
pp. 49–54, 2018, issn: 0378-7753.
[55] D. Worthington, “Queueing models for hospital waiting lists”, Journal
of the Operational Research Society, vol. 38, no. 5, pp. 413–422, 1987.
[56] R. A. Nosek and J. P. Wilson, “Queuing theory and customer satisfac-
tion: A review of terminology, trends, and applications to pharmacy
practice”, eng, Hospital Pharmacy, vol. 36, no. 3, pp. 275–279, 2001,
issn: 0018-5787.
[57] H. Shi, Y. Zhang, H. Wu, S. Chang, K. Qian, M. Hasegawa-Johnson,
and J. Zhao, Continuous convolutional neural network fornonuniform
time series, 2020. [Online]. Available: https : / / openreview .
net/forum?id=r1e4MkSFDr.
[58] Martın Abadi, Ashish Agarwal, Paul Barham, et al., TensorFlow: Large-
scale machine learning on heterogeneous systems, Software available
from tensorflow.org, 2015. [Online]. Available: https://www.tensorflow.
org/.
[59] F. Chollet et al., Keras, https://keras.io, 2015.
[60] P. Rémy, Keras tcn, https://github.com/philipperemy/
keras-tcn, 2020.
[61] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep net-
work training by reducing internal covariate shift”, arXiv preprint arXiv:1502.03167,
2015.
[62] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization”,
arXiv preprint arXiv:1412.6980, 2014.
[63] Rapids, https://rapids.ai, 2020.
[64] J. Bergstra, D. Yamins, and D. D. Cox, “Making a science of model
search: Hyperparameter optimization in hundreds of dimensions for vi-
sion architectures”, eng, JMLR, 2013, isbn: 1938-7228.
[65] J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, “Algorithms for
hyper-parameter optimization”, in Advances in Neural Information Pro-
cessing Systems 24, J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett, F.
Pereira, and K. Q. Weinberger, Eds., Curran Associates, Inc., 2011,
pp. 2546–2554. [Online]. Available: http : / / papers . nips .
60 BIBLIOGRAPHY

cc/paper/4443- algorithms- for- hyper- parameter-

optimization.pdf.
[66] M. Friedman, “The use of ranks to avoid the assumption of normality
implicit in the analysis of variance”, Journal of the american statistical
association, vol. 32, no. 200, pp. 675–701, 1937.
[67] P. Virtanen, R. Gommers, T. E. Oliphant, et al., “SciPy 1.0: Fundamen-
tal Algorithms for Scientific Computing in Python”, Nature Methods,
vol. 17, pp. 261–272, 2020. doi: https://doi.org/10.1038/
s41592-019-0686-2.
[68] F. Wilcoxon, “Individual comparisons by ranking methods”, eng, Bio-
metrics Bulletin, vol. 1, no. 6, pp. 80–83, 1945, issn: 00994987.
[69] R. E. Kalman, “A New Approach to Linear Filtering and Prediction
Problems”, Journal of Basic Engineering, vol. 82, no. 1, pp. 35–45,
Mar. 1960, issn: 0021-9223. doi: 10.1115/1.3662552. [Online].
Available: https://doi.org/10.1115/1.3662552.
[70] E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy consid-
erations for deep learning in NLP”, CoRR, vol. abs/1906.02243, 2019.
arXiv: 1906.02243. [Online]. Available: http://arxiv.org/
abs/1906.02243.
[71] E. Garcıa-Martın, C. F. Rodrigues, G. Riley, and H. Grahn, “Estimation
of energy consumption in machine learning”, Journal of Parallel and
Distributed Computing, vol. 134, pp. 75–88, 2019.
[72] D. Neil, M. Pfeiffer, and S.-C. Liu, “Phased lstm: Accelerating recurrent
network training for long or event-based sequences”, in Advances in
neural information processing systems, 2016, pp. 3882–3890.
Appendix A

Forecasts

This section presents the predicted and actual wait-time for all models for 4
randomly selected days in the test-data.

61
62 APPENDIX A. FORECASTS

Figure A.1: The actual wait-time and the predicted wait-time for each model
for single day in the test data (2019-12-28).
 APPENDIX A. FORECASTS 63

Figure A.2: The actual wait-time and the predicted wait-time for each model
for single day in the test data (2020-01-09).
64 APPENDIX A. FORECASTS

Figure A.3: The actual wait-time and the predicted wait-time for each model
for single day in the test data (2020-02-08).
 APPENDIX A. FORECASTS 65

Figure A.4: The actual wait-time and the predicted wait-time for each model
for single day in the test data (2020-02-10).
TRITA EECS-EX-2020:241

www.kth.se