development of a real-time algorithm for the detection of ... · and communication which further...

Post on 01-Nov-2019

0 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Development of a Real-Time Algorithm for the Detection ofPhysiological Arousal in the Presence of Motion Among Children with

Autism

by

Akshay Sainag Reddy Puli

A thesis submitted in conformity with the requirementsfor the degree of Master of Health Science, Clinical Engineering

Institute of Biomaterials & Biomedical EngineeringUniversity of Toronto

c© Copyright 2018 by Akshay Sainag Reddy Puli

Development of a Real-Time Algorithm for the Detection of Physiological Arousal in the Presence of

Motion Among Children with Autism

Akshay Sainag Reddy Puli

Master of Health Science, Clinical Engineering

Institute of Biomaterials & Biomedical Engineering

University of Toronto

2018

Abstract

Anxiety is a clinical concern among some children with autism, for whom such an emotional onset may

exacerbate core symptoms. Current treatment options are challenged as many of these children are

unaware of their emotional state, and due to autism some children also have difficulties with language

and communication which further complicates anxiety diagnoses. In this thesis, we proposed a novel

automated anxiety detection algorithm capable of interpreting heart rate arousal in the presence of

physical movements , and intuitively notifying the user to engage in anxiety management techniques.

The novelty of our proposed algorithm is based on its ability to detect anxiety in presence of motion;

this is a challenge as motion and anxiety influence the heart rate similarly. This algorithm was able to

account for the changes in heart rate due to motion and detection anxiety with an accuracy, specificity

and sensitivity of 92.7% in children with autism.

ii

Acknowledgements

My journey was laced with obstacles and hurdles, that I could not have overcome without the continuous

support of my supervisor Dr. Azadeh Kuski.

I would like to thank my family and friends for their consistent encouragement, especially by S.B.

who braved the long days at the library beside me and my annoying whining.

Lastly, I would like to dedicate this thesis to the participants and their family, who took the time to

take part in this research.

iii

Contents

Acknowledgements iii

Contents iv

List of Tables vi

List of Figures vii

1 Introduction 1

2 Background 4

2.1 Nervous System Physiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1.1 Autonomic Response Pathway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1.2 Cardiovascular Autonomic Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1.3 Autonomic Response to Anxiety . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Measuring Cardiovascular Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2.1 Electrical Activity of the Heart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2.2 Electrode Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.3 Motion Artefacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.4 Denoising ECG Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.5 Heart Rate Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Accelerometery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.4 Arousal Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.4.1 Current Emotion Detection Review . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.4.2 Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.4.3 Maneuver Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.4.4 Kalman Filter Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.5 Multimodal Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Methods 21

3.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2.1 Sensor Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.2.2 Behavioural Questionnaire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.3 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

iv

3.4 Pre-processing of Sensor Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.4.1 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.5 Performance Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4 Filter Design 27

4.1 Multimodal Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.2 Activity Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.3 Arousal Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5 Results 33

5.1 Characterisation of Heart Rate Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5.2 STAI Questionnaire Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.3 Detection Algorithm Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.3.1 Evaluation of Motion Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.3.2 Anxiety Detection Algorithm Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 37

5.3.3 Anxiety Detection Algorithm Optimised Results . . . . . . . . . . . . . . . . . . . 43

6 Discussion 45

7 Conclusion 49

Bibliography 50

Appendix 57

Appendix-A : STAI Questionnaire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

Appendix-B: Participant average heart rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

Appendix-C: Anxiety and motion detection participant result . . . . . . . . . . . . . . . . . . . 73

Appendix-D: Visual Reference for Electrode Placement . . . . . . . . . . . . . . . . . . . . . . . 81

v

List of Tables

5.1 Participant Demographics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5.2 Summary of Optimised Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

vi

List of Figures

2.1 Heart rate variation due to motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Electrocardiogram Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.3 Skin-Electrode interface electrical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.4 ECG signal distorted due to motion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.5 Frequency spectrum of an ECG signal with the absence of Motion Artifacts . . . . . . . . 10

2.6 Frequency spectrum of an ECG signal with the motion artifacts . . . . . . . . . . . . . . . 10

2.7 The frequency response of Butterworth, Chebyshev Type-1, and Chebyshev Type-2 win-

dows. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.8 Adaptive filter algorithm with acceleration noise reference . . . . . . . . . . . . . . . . . . 13

2.9 An overview of the Pan-Tompkins R-peak detection [1]. . . . . . . . . . . . . . . . . . . . 14

2.10 Overview of Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.11 Overview of the proposed anxiety detection algorithm . . . . . . . . . . . . . . . . . . . . 19

3.1 Shimmer 2r chest strap and typical placement of ECG electrodes [2] . . . . . . . . . . . . 22

3.2 Testing Environment Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.3 Experimental Protocol Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.4 Stroop test screen example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.5 An overview of the steps required to extract features from the ECG and acceleration signals 25

4.1 Overview of the proposed anxiety detection algorithm . . . . . . . . . . . . . . . . . . . . 28

4.2 Overview of the proposed anxiety detection algorithm . . . . . . . . . . . . . . . . . . . . 29

4.3 Overview of the arousal detection algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.1 The average heart rate (N=15) during each of the twelve phases of the testing protocol . . 34

5.2 Effect of threshold τm on the classification of motion (N = 15, Qm = 0, ρm = 5, ws =

5, wstd = 5, wIM = 10 and overlap=50%) . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5.3 Effect of innovation moving average window length wIM (N = 15, Qm = 0, ρm = 5, ws =

5, wstd = 5, τm = 0 and overlap = 50%) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5.4 Effect of Process Covariance Qm on the performance of the motion detection (N =

15, τm = 0, ρm = 5, ws = 5, wstd = 5, wIM = 10 and overlap=50%) . . . . . . . . . . . . . . 38

5.5 Effect of Measurement Covariance Multiplier ρm on the performance of the motion detec-

tion (N = 15, Qm = 0, τm = 0, ws = 5, wstd = 5, wIM = 10 and overlap=50%) . . . . . . . 38

vii

5.6 Effect of the moving average window length ws applied to the acceleration resultant mag-

nitude vector, on the performance of the motion detection algorithm (N = 15, Qm =

0, ρm = 5, τm = 0, wstd = 5, wIM = 10 and overlap=50%) . . . . . . . . . . . . . . . . . . . 39

5.7 Effect of moving standard deviation window length Wstd applied to the raw accelerometer

signal, on the performance of the motion detection algorithm (N = 15, Qm = 0, ρm =

5, ws = 5, τm = 0, wIM = 10 and overlap=50%) . . . . . . . . . . . . . . . . . . . . . . . . 39

5.8 Effect of the threshold τanx on the performance of the arousal detection algorithm (N =

15, Qanx = 0.001, ρanx = 1, wRR = 10, wn = 50, and overlap = 50%) . . . . . . . . . . . . 40

5.9 Effect of measurement covariance Qanx on the performance of the arousal detection algo-

rithm (N = 15, τanx = 0.55, ρanx = 1, wRR = 10, wn = 50, and overlap = 50%) . . . . . . . 40

5.10 Effect of the measurement covariance multiplier ρanx on the performance of the arousal

detection algorithm (N = 15, Qanx = 0.001, τanx = 0.55, wRR = 10, wn = 50, and

overlap = 50%) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

5.11 Effect of R-R interval moving average window WRR on the performance of the arousal

detection algorithm (N = 15, Qanx = 0.001, ρanx = 1, τanx = 0.55, wn = 50, and

overlap = 50%) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

5.12 Effect of innovation moving average window wn on the performance of the arousal detec-

tion algorithm (N = 15, Qanx = 0.001, ρanx = 1, wRR = 10, τanx = 0.55, and overlap =

50%) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

5.13 Acceleration signal and detected anxiety level by the motion detection algorithm, during

baseline (BL) and stroop task (SA), while standing, slow walking, and fast walking . . . . 43

5.14 Innovation ζ, and identified arousal states during standing still, slow walking, and fast

walking for one participant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

7.1 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-1 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 58

7.2 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-2 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 59

7.3 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-3 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 60

7.4 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-4 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 61

7.5 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-5 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 62

7.6 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-6 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 63

7.7 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-7 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 64

7.8 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-8 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 65

7.9 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-9 while standing, slow walking, and fast walking . . . . . . . . . . . . . . . 66

7.10 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-10 while standing, slow walking, and fast walking . . . . . . . . . . . . . . 67

viii

7.11 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-11 while standing, slow walking, and fast walking . . . . . . . . . . . . . . 68

7.12 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-12 while standing, slow walking, and fast walking . . . . . . . . . . . . . . 69

7.13 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-13 while standing, slow walking, and fast walking . . . . . . . . . . . . . . 70

7.14 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-14 while standing, slow walking, and fast walking . . . . . . . . . . . . . . 71

7.15 Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded

for participant-15 while standing, slow walking, and fast walking . . . . . . . . . . . . . . 72

7.16 Arousal detection for participant-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

7.17 Arousal detection for participant-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

7.18 Arousal detection for participant-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

7.19 Arousal detection for participant-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

7.20 Arousal detection for participant-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

7.21 Arousal detection for participant-6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

7.22 Arousal detection for participant-7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

7.23 Arousal detection for participant-8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

7.24 Arousal detection for participant-9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

7.25 Arousal detection for participant-10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

7.26 Arousal detection for participant-11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

7.27 Arousal detection for participant-12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

7.28 Arousal detection for participant-13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

7.29 Arousal detection for participant-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

7.30 Arousal detection for participant-15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

ix

Chapter 1

Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental condition defined in the Diagnostic and Statis-

tical Manual of Mental Disorders (DSM-V) as an impairment in social communication, and the presence

of repetitive behaviour or restrictive interests [3]. Social communication difficulties are characterized

by unusual behaviours during social interactions such as an inability to maintain conversations, identify

social cues, abnormal eye contact, and unsocial body postures [4]. Restrictive and repetitive behaviours

(RRB) were first defined as a characteristic symptoms of ASD in 1943 by Kanner, who described such

behaviour as ”the childs resistance to environmental change from anyone else but themselves” [5]. An

insistence for sameness and high frequency, repetitive motor movements are the hallmarks of ASD [6].

Prevalence of ASD has significantly increased in recent years among children; at present in the United

States, 1 in 68 children are classified to have been diagnosed with ASD, according to the Centre for

Disease Control and Prevention (CDC) [4]. The high prevalence of ASD has encouraged research into

understanding the pathology, etiology and symptomatology of ASD.

The heterogeneous nature of ASD, manifests uniquely in each individual with a spectrum of core,

and co-occurring symptoms [7]; 60% of children diagnosed with ASD experience numerous co-occurring

disorders that influence the expression of the core ASD symptoms [8]. Anxiety is one such co-morbidity

that is is highly prevalent in the ASD population; 85% of children, and adolescents are diagnosed with co-

occurring form of anxiety [9]. It is known to cause functional impairment which further exacerbates core

ASD symptoms in particular social-communication, and repetitive and restrictive behaviour [10]. Thus,

anxiety impedes an individual’s daily activities such as securing employment, maintaining relationships

or succeeding in academia. High functioning children with ASD, are also greatly impaired by anxiety due

to the awareness of their symptoms, but as a consequence of language, and communication impairment

they tend to seclude themselves from their peers, avoiding any social interactions[11].

Anxiety poses a challenging diagnosis, as it often overshadowed by clinically more significant condi-

tions such as ASD itself [12]. Current anxiety detection measures depend on parental questionnaires,

self questionnaires, and observations that are subjective, and rely on the patient’s ability to be aware

of their emotions, or internal state [13]. Cognitive behavioural therapy (CBT) is a common treatment

regimen used to develop techniques with a patient to overcome, and manage their anxiety [14, 6, 15].

CBT also heavily depends on one’s ability to be self-aware, and recognise that they are feeling a sense

of anxiety. Children with autism often have difficulties with self-awareness, and are unaware of the emo-

tions that they are experiencing[16]. Thus, the effectiveness of CBT can be limiting as the child does

1

Chapter 1. Introduction 2

not recognise the appropriate time in which to engage with activities learnt through CBT. Additionally,

impairments in language, and communication further challenges traditionally anxiety treatments, as the

child is unable to articulate their emotions. This especially places a huge burden on families, as they

are left guessing to the care their child is wanting, and this creates frustration that negatively impacts

their quality of life.

There is an immediate need for a technology that has the capacity to detect anxiety, so that children

with autism can be taught to correlate an identifiable alert to the onset of anxiety. Then encouraged to

engage with techniques learnt through CBT to manage the symptoms they are currently experiencing.

This technology should provide a means of objective, and language-free detection of anxiety, that a

child could rely on to reinforce their current symptoms. We envision such a product to encompass

two components: 1) a wearable device, and 2) an intelligent anxiety detection algorithm. Wearable

technologies allows for a unique access to our physiology; sensors are able to be camouflaged into clothing

for an unintrusive measure of physiological signals such as heart rate, and respiration. Such devices

are increasingly becoming popular among the consumer market; fitness based smart-watches are now

able to accurately capture heart rate measurements by being placed on the wrist. Additionally, the

close proximity of this technology allows researchers to develop assistive algorithms that are capable of

continuous monitoring a user’s health, and deliver of care in realtime. Thus, with the easy accessibility

of wearable hardware from of-the-shelf products; the next innovations are expected to be brought on

by algorithms designed specifically for the continuum of healthcare, beyond the confines of traditional

institutions, and into the everyday lives of individuals using wearable technologies.

Our motivation was to develop an assistive technology that would be able provide an objective, and

language-free measurement of anxiety; implemented into a wearable device, such an algorithm would

be designed to autonomously tracks the user’s anxiety level, and provide notifications during elevated

levels. In order to achieve this goal, we had defined two research questions:

1. Is there a change in heart rate in response to anxiety while at rest, as well as during motion?

2. Can an algorithm analyse heart rate patterns to detect anxiety at rest and in the presence of

motion?

Previous researchers have developed algorithms that utilise non-invasive physiological measurements

as a proxy to detect ones level of anxiety or stress. Heart rate, respiration, perspiration, pupil dilation,

and blood pressure are common measurements that were recorded to explore a physiological response

to anxiety or stress [17, 18, 19]. Thus, the first research objective was to provide evidence of significant

elevation of heart rate due to anxiety, while the participant is at rest, as well as when engaged in

physical activity. This would inform the development of a realtime anxiety detection algorithm, that

could provide an autonomous, objective, and language free method of communicating a child’s state of

anxiety. This technology could translate one’s internal state, into a recognizable form of notification,

that a child could use to then learn to relate with anxiety; used in conjunction with CBT, a child could

engage in coping techniques when notified by the algorithm. Additionally, this would reduce the burden

on caregivers, and parents guessing at their child’s internal state, as the alerts sent out by the anxiety

detection algorithm would assist them in selecting the right care for their child, in realtime.

The present state of research into anxiety, and stress detection is based on supervised machine

learning algorithm, such as the support vector machine (SVM) [20]. These algorithms require a large

dataset from multiple participants, for the learning, and construction of a generalised model of the

Chapter 1. Introduction 3

correlation between heart rate and anxiety. Though such algorithms have been effective in classifying

state of anxiety, the generalised model does not account for interpersonal variation. However, some

researchers have overcome this issue by training models uniquely to each participant [20]; as a practical

implementation, this approach is limiting due to the requirement for the collection of lengthy training

data from each user prior to the algorithm producing accurate measurements. Additionally, the initial

model is solely based on the collected training data, and does not dynamically adapt to new observations.

In the case of tracking heart rate, the child’s baseline might change with time, and a system must adapt

to such variations in real-time in order to continue to produce an accurate analysis of anxiety level.

Thus, a supervised approach seems limited in various aspects required to function in a naturalistic

environments.

As an unsupervised classifier, the Kalman Filter is able to overcome much of the limitations of the

SVM algorithms. It is able to rapidly build a baseline model based on initial measurements of the users

heart rate [21]. Based on this model, the algorithm could tack the baseline heart rate, and adjust for

any long-term trend deviations through new observations [21]. A sudden deviations of the heart rate

measurement from the model predictions, are labelled as an onset of anxiety. A.Kuski [22] had developed

such an algorithm, and tests on children with autism had showed sensitivity and specificity of 99% and

92%, respectively. Although a limitation of this study was that it measured its data while the participant

was sitting, and engaged in minimal physical activity; this is unrealistic to environments in which a child

is typical engaged in.

Detection of anxiety is challenged by the presence of motion, which users commonly experience in a

naturalist environment. Physical activity includes any movement of the limbs, from a change in posture

to walking. When we engage in such motion, we do so at the cost of some energy expenditure. This

creates an imbalance as there is a sudden demand for more nutrient supply. The autonomic nervous

system responds by activating pathways in an effort to restore homoeostasis. Some of the pathways

that are activated overlap those that also respond to the onset of anxiety; this is particularity true for

the sympathetic nervous system pathways that innervate at the heart. As a result, physical activity

manifests similarly, with respect to the acceleration of the heart rate, to the onset of anxiety. Thus,

current anxiety detection algorithms that fail to account for such deviation, would inevitably detect false

positives when applied to heart rate measurements that are recorded for participants who are not at

rest.

The main contributions of this thesis was to develop a robust anxiety detection algorithm that

is capable of functioning in naturalistic conditions. Additionally, the algorithm was designed to utilise

measurements such heart rate, and motion that are typically available on accessible wearable technologies

such as smart-watches. The development, and performance testing of our algorithm is organized into

the following six chapters. Chapter 2 is the background section that presents further details on the

rationale behind the choice of physiology measurement, the basics of the Kalman, and multimodal

kalman filters which form the foundation of the anxiety detection algorithm. The methodology that was

developed to recruit, and collect data from participant in order to test the performance of the proposed

anxiety detection algorithm is presented in chapter three, along with the implementation of the detection

algorithm in chapter four. The results from the sensitivity analysis of the algorithm are presented in

chapter 5, and their significance discussed in chapter six, along with a conclusion of the scope of this

research in the future.

Chapter 2

Background

This chapter provides the foundational material for the methodology presented in the latter sections.

Firstly, we describe the roll of the nervous system in maintaining homoeostasis among visceral organs.

Follow by an overview of the degree to which emotion activates neurological pathways, and the effect on

various organ systems. We then explore non-invasive measures to monitor cardiovascular activity, and

lastly conclude with a review of a few algorithms that analyse physiological signals as a marker for one’s

internal state or emotion.

2.1 Nervous System Physiology

Neural pathways span the entirety of the human body, forming a communication circuit between the

brain, body and the surrounding environment [23]. Neurons are the fundamental unit of the nervous

system, through which encoded messages are transmitted by an electrochemical medium [24]. Based on

their function and location, neurons are classified into the central nervous system (CNS), that comprises

of the brain and spinal cord, or the peripheral nervous system (PNS) [25]. The PNS is further categorised

into the autonomic nervous system (ANS), which controls involuntary functions such as heart rate,

digestion, blinking etc, and the somatic nervous system (SNS) that controls voluntary functions [25].

The autonomic nervous system is further divided into the parasympathetic and sympathetic path-

ways, through which granular control over the functioning of visceral organs is achieved [26] [27] [28]

. This is established through the antagonistic effect produced through the activation of each of the

pathways[29]. The sympathetic nervous system prepares the body for a ‘fight or flight’ response which

evokes changes to several different organs in particular the acceleration of the cardiovascular system

[29]. The parasympathetic nervous system, on the other hand induces a ‘sleep and digest’ response that

reduces physiological function such as heart rate, respiration, and an increased excretion of enzymes to

aid in digestion [29].

The activation of the ANS is a response to an environmental or internal stimuli, in an attempt to

maintain homoeostasis among the various organ systems[30]. The effects of an environmental change,

such as a variation in temperature, activate specific pathways that are observable through the onset of

physical changes such as shivering, raised hair follicles, or sweating. Emotions and changes in mental

state can also induces the activation of ANS pathways resulting in physiological changes, however; there

is much debate on the cohesiveness between emotions and a specific ANS response[31][32].

4

Chapter 2. Background 5

2.1.1 Autonomic Response Pathway

An autonomic response to emotion is often defined as the measurable physiological variations that occur

surrounding the period in time when an emotion is felt [33]. There are two primary pathways in which an

autonomic response could be enlisted: 1) innervation of parasympathetic and sympathetic neurons onto

visceral organs, and 2) release of specific excitatory and inhibitory hormones that transverse through the

endocrine system [33]. Each of these pathways play a role in the temporal activation of an autonomic

response. Activation of the neural pathways, induce an instantaneous but short lasting effect. On the

other hand, a hormonal pathway produces a response that is longer lasting however; requires a relatively

prolonged onset duration [34].

2.1.2 Cardiovascular Autonomic Control

The heart functions to continuously circulate oxygenated, and nutrient rich blood to the cells of the body;

in addition to transporting de-oxygenated blood to the lung for replenishment [35]. This is achieved

through a series of synchronised contractions, and relaxations by each of the four chambers within the

heart; top two chambers are refereed to as atria, and the bottom chambers are called ventricles [36] .

Similar to other muscles, the heart contracts when exposed to electrical impulses from innervating nerve

endings [37]. The heart is able to produce its own set of electrical impulse that ensure synchronous

contractions between all the chambers [36]. These electrical impulses are produced by the sinoatrial

node (SA) located near the right atrium. Electrical impulses from the SA node propagate through out

the heart inducing a wave of contractions[36]. They terminate at the boundaries of the atria, and also

activates the atrioventricular (AV) node that generates another set of electrical impulses ,after a delay,

to contract both ventricles; resulting in the departure of the blood from the heart [36].

Further control of the cardiovascular system is established through the innervation of the sympathetic

and parasympathetic neurons, that release excitatory and inhibitory neurotransmitters [37][38]. The

temporal summation of each such neurotransmitters at the SA node, dictates the rate at which the

heart contracts [37]. Thus, it is often difficult to differentiate specific influences by the sympathetic

and parasympathetic pathways for a given stimuli [39]. However, analysing the net effect of the ANS

provides a substantial understanding of the ratio at which each of the pathways are being expressed[40].

Based on this assumption we can explore the relationship between one’s internal state, and the effects

on cardiovascular system.

It has been speculated that negative emotions such as anxiety, fear, and anger results in a state

of cardiovascular arousal; more specifically, such emotions are correlated with increased heart rate,

elevated breathing, and excess perspiration [41][42]. These physiological changes are typically attributed

to a sympathetic response, as the body prepares for an onset of an energy intensive period [43]. Similar,

such changes are comparable to a response during which an individual is taking part in physical activities

[44].

Skeletal muscles are the functional unit that enable us to engage in physical movements of our limbs

[45]. During such a movement, muscles are contracted to force a change in position of an attached

skeletal structure [46]. Muscles contractions require additional energy, and during physical activities

numerous muscles simultaneously are needed to contract [46]. This creates a sudden demand of energy

above the baseline levels, resulting in the activation of a sympathetic response to meet the bodies energy

needs. An acceleration in heart rate, blood pressure, and breathing can be observed as body ensures

Chapter 2. Background 6

sufficient supply of oxygen and nutrients in maintained [47]. Figure 2.1 shows a positive correlation

between heart rate, and the level of physical movement; there is a clear increase in heart rate during the

fast walking stage, as compared to standing still.

Figure 2.1: Heart rate variation due to motion

2.1.3 Autonomic Response to Anxiety

A physiological response to emotion has been a debated topic in research, with a variety of theories

explaining a correlation or the lack there off [33]. For the purposes of this study, we had based our

assumption on one of the theories which believes that physiological response is an evolutionary trait that

had developed as a measure to protect, and prepare for an imminent change in surrounding conditions

[48].

Researchers [49] [50] [51], had explored the theory of coherence between mental state and a autonomic

response. They had made use of markers such as heart rate, skin conductance, facial response, and

skin temperature; as a convenient and non-invasive proxy to measure the activation of the numerous

pathways within the autonomic nervous system. Jacobsen [49], had measure several physiological metrics

and found that skin temperature and heart rate responded to verbally provoked emotions. Collet [50],

further explored the idea of autonomic response to basic emotions, as he defines them as happiness,

surprise, anger, fear, sadness, and disgust. The research concluded that the complexity of the ANS

pathways, resulted in no-one physiological marker differentiating all the basic emotions. However; skin

conductance, body temperature, and respiration together were able to classify each emotional response.

Chapter 2. Background 7

Though, these results show coherence between emotional reaction and physiological change, there is

much controversy among the specificity of an ANS pathway to an emotion [52].

2.2 Measuring Cardiovascular Function

Electrocardiogram (ECG) is a non-invasive measure of cardiovascular activity. Electrodes on the surface

of the skin record the electrical activity of the heart; providing an insight into the internal state. In

this section we explore the fundamentals of measuring the electrical activity of the heart, processing the

signal, and extracting accurate heart rate reading.

2.2.1 Electrical Activity of the Heart

The sinoatrial (SA) node and atrioventricular (AV) node are part of the hearts conduction system,

that controls the rate of contraction of the atria and ventricles [53]. The SA node functions as a

pacemaker, producing electrical impulses that propagate through the myocardium surrounding the atria,

and terminate at the AV node [53]. Upon a short delay, the AV node fires resulting in the contraction of

the ventricles [53]. The electrical impulses produced by each of the nodes, produces a charge imbalance

in the surroundings that propagates in all directions [54]. Upon reaching the epidermis, the waves of

charge imbalances manifest as a potential difference that can be measures across a pair of electrodes

attached on the surface of the skin [54].

Skin electrodes placed on the surface of the skin are connected to an electrocardiogram, that is able

to measure minute fluctuations in voltages across a pair of electrodes to capture an ECG signal [54].

This signal is representative of the electrical activity of the heart. Figure 2.2 depicts a typical ECG

signal, as recorded through two electrodes placed on the midclavicular line, and few centimetres inferior

to the clavicle.

Figure 2.2: Electrocardiogram Signal

The topology of an ECG signal contains several peaks referred to as P, Q, R, S, and T; detonated

Chapter 2. Background 8

in Figure 2.2. Each of the peaks are representative of a different state of the heart during a beat. The

largest and most prominent peak, is the R-peak. It represents the contraction of the large myocardial

muscle surrounding the left ventricle; this is the time point at which oxygenated blood exists from the

heart to the rest of the body.

2.2.2 Electrode Placement

Electrocardiograms are connected to a variety of electrode orientations depending on the quality of

ECG signal that is to be measured [55]. For the purposes of detecting heart rate, it is often sufficient

to use two or three electrodes placed across the chest [55] [56]. Though for a clinical review of ones

cardiovascular activity, the placement of electrodes is crucial [56]. However, in most applications a ECG

signal is measurable given a pair of electrodes that are located bilateral to the heart.

A drawback to a 2-lead electrode placement, is that the signal is more prone to artifacts. In order

to improve the signal-to-noise ratio, a third electrode is often attached inferior to the diaphragm, and

refereed to as the right-leg-drive electrode [57]. Through this electrode, small fluctuation in voltage

are produced to be proportional to the inverse of the common-mode noise present between the pair

of detection electrodes [57]. These signals then propagate through the surface of the skin to summate

at each of the electrodes [57] [58]. This would result in a destructive interference, and attenuate the

influence of the common-mode noise. Thus, improving the quality of the ECG signal [58].

2.2.3 Motion Artefacts

An electrical signal originating at the heart, traverses through several layers made up of different material

properties. Signals can encounter layers of muscles, adipose tissue, and fluids that attenuate and distort

the signal prior to manifesting on the surface of the skin.

An additional source of noise, is the skin-electrode interface. This interface is often attributed to

much of the noise observed on an ECG signal [59]. In order to develop a rationale for the noise at the

interface of the electrode, we model the physical properties as an electrical circuit. Figure 2.3 depicts

one such model, where the dermis and epidermis behaving as resistors with a non-linear capacitative

effect in parallel [59].

Upon placing an electrode with conductive gel on the surface of the skin. A half-cell potential is

induced in the gel due to a flow of current between the surface of the skin and the electrode [60].

Eventually, the ions in the gel and at the surface of the skin redistribute to reach an equilibrium, and

establishing a potential difference [60]. This half-cell potential modelled in Figure 2.3 as the voltage

source Esweat; manifests as a DC offset on a ECG signal. During motion, the movement of the electrode

causes an imbalance in the ion equilibrium, and and results in fluctuation of the half-cell potential that

is superimposed onto the ECG signal as motion artifacts. Figure 2.4 illustrates an ECG signal that is

affected by noise due to motion.

2.2.4 Denoising ECG Signals

Physical activity has a profound effect on the quality of the ECG signal acquired by an electrocardiogram

[61]. The excessive movements results in the deterioration of the ECG signal. Thus, several post-

processing techniques have been developed to counter much of the distortions present in the ECG signals

that are measured during the presence of motion.

Chapter 2. Background 9

Figure 2.3: Skin-Electrode interface electrical model

Figure 2.4: ECG signal distorted due to motion.

Figures 2.5 - 2.6 depict the fast Fourier transform (FFT) of an ECG signal, and it can be observed

that much of the information contained within the ECG signal is between the frequency range of 0.01Hz

to 100Hz [62]. Typically the high frequency R-component of the ECG signal is represented by the

frequencies around 2Hz [63]. Additionally, comparing the FFT plots of an ECG signal with the presence

of motion artifacts (Figure 2.6), and without such artifacts (Figure 2.5); there is an observable additional

frequency components that are superimposed within the frequency range of the ECG signal. Much of the

human motion, such as walking, manifests at around 40Hz, and thus the addition frequency components

are the source of the distortions seen in the time domain of the ECG signal. Another source of noise, is

due to the human body acting as an antenna to pick 60Hz or 50Hz oscillations due to mains powerline

Chapter 2. Background 10

that exists in the surroundings [64]. This type of noise is clearly observable in the frequency domain of

the ECG signal, as a peak at a frequency of 60Hz. There are several post-processing techniques which

are capable of reducing the effects due to motion artifacts, and powerline interference.

Figure 2.5: Frequency spectrum of an ECG signal withthe absence of Motion Artifacts

Figure 2.6: Frequency spectrum of an ECG signal withthe motion artifacts

Digital Filters

Filters are intended to segregate a signal based on specific characteristic properties. Digital filters

are designed to selectively remove specific frequency components, that produce distortions in the time

domain[65].

There are two implementations of digital filters, finite impulse response (FIR) and infinity impulse

response (IIR) filters that gain their characteristics through the design of filter coefficients [65]. FIR

filters are chosen for applications that require a linear-phase characteristics during the passband. IIR

filters have sharper cut-offs with lower lobs in the stop-ban, when in comparison to a FIR filter with the

same number of coefficients [65]. However, IIR induces a phase shift in the pass-band [65].

Filter characteristics are defined by their transfer function, which specifics the co-efficient that in

return defines the frequency and phase response of the filter. Through tuning each of these co-efficients,

filters are able to be customised to selectively attenuate undesired frequencies [66]. A generalised transfer

function of an IIR filter in the z-domain is given by

H(z) =B(z)

A(z)=

∑mi=0 biz

−i∑nk=0 akz

−i (2.1)

where bi and ak are co-efficients that define the filters performance, and i and k are the number of

co-efficients [66]. Through such a transfer function we are able to identify the poles (from the numerator

Chapter 2. Background 11

co-efficients) and zeros (from the denominator coefficients) of the filter in the z-plane, that examin the

stability of any given filter [67].

There are three basic filter characteristics that are often implemented for the processing of digital

signals.

1. Low Pass Filter: Attenuates frequencies that are higher than the cut-off frequency Fs. It is used

to extract low frequency components of signal.

2. High Pass Filter: Frequencies greater than the cut-off frequency Fs are amplified, and lower

frequencies are attenuated. Such a a filter is used to extract high frequency components of a

signal.

3. Bandpass Filter: The passband of this filter exists between a range of frequencies, and as

frequencies outside this range are attenuated.

An idealist implementation of a filter would be a sharp, vertical cut-off at a defined frequency Fs.

However, this is unrealisable in practical implementations, thus suboptimal windows have tried to a

provide solution with varied compromises. Figure 2.7 compares the response of three types of windows:

Butterworth, Chebyshev Type-1, and Chebyshev Type-2.

Figure 2.7: The frequency response of Butterworth, Chebyshev Type-1, and Chebyshev Type-2 windows.

Windows allow for a simple implementation of a specific type of filter. The choice of a type of window,

is dependent on the application, and the tolerance of the window specific compromises. For example,

Butterworth filters have a ripple-free pass, and stop band. However; the cutt-off is not as sharp as the

Chebyshev filters, which achieve sharper cutt-off by tolerating ripples.

A Bandpass filter is typically designed to initially process an ECG signal. The frequency range of

passband is set to be between 5Hz and 15Hz, to extract the R-wave of the ECG signal. This is specific

to heart rate detection, as the rest of the ECG signal is not required for such a calculation. Additionally,

Chapter 2. Background 12

the narrow pass band of the Bandpass filter, removes much of the noise inducing frequency components.

However, there still exists frequency components from motion artifacts that are superimposed in the

passband frequency range. A Bandpass filter, which differentiates based on frequency, is unable to

remove such a type of noise. In order to overcome these limitations, we next explore adaptive filters that

are specifically able to attenuate noise that is present in the frequency range of interest.

Adaptive Filter

ECG signals are made up of a deterministic and stochastic component. Stochastic encompasses the

random variation or noise in the signal and deterministic is the signal information that is representative

of the physiological source [68]. Equation 2.2 is a common model of a physiological signal (y) that is

correlated to the linear combination of the deterministic component (S) and stochastic component (ε)

[69]. The goal of an adaptive filtering algorithm is to reproduce a signal with a reduction in variance

due to noise.

y = S + ε (2.2)

Adaptive filters function on the assumption that motion artifacts are a linear addition to the determin-

istic physiological signal. Thus a removal of these artifacts would entail a subtraction of the captured

signal by a noise reference that is directly correlated to the motion artifacts. An overview of a common

adaptive filter is shown in Figure 2.8.

Lui et al. reviewed several noise references [69] and found that the most effective was the use of 3-

dimensional acceleration data that captures the users chest movements. Using the X,Y and Z acceleration

data in combination with tunable coefficients, the adaptive filter algorithm attempts to minimize the error

function stated in Equation 2.3. Where X is the reference input X(n) = [C,Accx(n), Accy(n), Accx(n)]

and W is the tunable filter coefficients W = [W0,W1,W2,W3] [69]. The adaptive filter algorithm min-

imises the cost function by recursively updating the filter coefficients to produce an output with least

amount of variance.

e = S + ε−XWT (2.3)

2.2.5 Heart Rate Inference

ECG signals contain a variety of information about the state, and functioning of the cardiovascular

system. In particular, ECG reading are able to reveal details such as left ventricular function, atrial

flutter, and onset of tachycardia [70]. Additionally, inference of heart rate from ECG is one of the most

common applications, as it provides a quick glimpse into one internal state. Fluctuations in heart rate

are thus regarded as a precursor to the onset of clinically adverse conditions.

As the topology of an ECG signal is representative of the hearts physical contractions, we can assume

that the time period between adjacent PQRST complex is proportional to the rate at which the heart

beats. Thus, to obtain an accurate measurement of heart rate, subsequent peaks of the complex must be

Chapter 2. Background 13

Figure 2.8: Adaptive filter algorithm with acceleration noise reference

accurately detected. Pan and Tompkins had developed just such an algorithm that specifically detected

the R-peaks of an ECG signal[1]. The reason for this choice was due to the R-peaks manifesting as a

sharp increase to a significantly large amplitude, which is observed in Figure 2.2[1]. Finally, calculating

the time difference between adjacent R-peaks ,referred to as R-R intervals, produced measurements that

are inversely proportional to heart rate [71].

Figure 2.9 shows an overview of the steps that make up the Pan-Tompkins algorithm. The input

signal y(t) is filtered through a bandpass filter with a cut-off at 5Hz and 15Hz, prior to applying the

Pan-Tompkins algorithm. As the first step, the input signal is differentiated to amplify parts of the signal

that change rapidly in a short period of time. Then the resultant signal is squared to non-linearly amplify

peaks with rest to the surrounding signal. The third step, is to apply a moving average integrator that

merges spikes that occur too close together. As it assumed that R-peaks can not physically manifest in

such short time periods, and thus such peaks are to be disregard for the heart rate calculations. Lastly,

a thresholding scheme is applied such that if x(t) is greater a threshold τ , then that instant in time is

recorded as containing a R-peak.

2.3 Accelerometery

We constantly engage in movement constantly during our daily lives; as this is how we interact, and

navigate through our surroundings. Muscles are laced across our skeletal system, enabling us to perform

precise movements through the contractions of such muscles. As dictated my newtons laws of motion;

upon muscle contraction, the force induced on the limb will accelerate while traversing a distance. Thus,

by measuring the acceleration of ones limbs, we are able to correlate to their level of physical activity.

Movements such as running that require gross, and rapid limb motion; would produce larger acceleration

amplitudes as compared to while walking.

Accelerometers are a type transducer that are able to measure the magnitude of acceleration [72].

As an object overcomes inertia, it experiences acceleration in the direction of the external force being

applied. Thus, acceleration is represented as a vector, described through a magnitude, and a direction.

Chapter 2. Background 14

Figure 2.9: An overview of the Pan-Tompkins R-peak detection [1].

However; accelerometers are unaware of their spacial orientation [72]. Thus, in order to measure the

direction of movement of an object in space; three accelerometers must be placed orthogonal to one and

other. This provides the capabilities of measuring acceleration in each of the three dimensional planes

(X, Y, and Z dimensions)[73]. The resultant acceleration magnitude (V (t)) is calculate by

V (t) =√S2x(t) + S2

y(t) + S2z (t) (2.4)

the square root of the sum of the squared magnitude of acceleration (Sx, Sx, andSx) in each of the three

dimensions. The parameter V (t), directly correlates to the pace at which the object could be observed

to be moving at. Thus, it forms a more reliable proxy to measuring ones activity level.

2.4 Arousal Detection

Inference of arousal and anxiety level from physiological measurements, has been explored by many

researchers. This section, begins with a discussion of the innovations and limitations of few of the

existing algorithms. Then we present an overview of the Kalman filter, which forms the basis of our

proposed algorithm for the detection of anxiety.

2.4.1 Current Emotion Detection Review

V. Sandulescu [74], had explored the concept of using electro-dermal activity (EDA) and pulse plethys-

mograph (PPG) as markers to determine neutral and stressful states. The EDA signals measure changes

in perspiration on the participants finger, and the PPG signal provided cardiovascular information such

heart rate and heart rate variability. In order to induce stress in the participant, the researchers decided

to employ the Trier Social Stress Test (TSST); which consisted of a neutral task, followed by a public

speaking and cognitive task. This was meant to replicate a stressful situation, and finally the TSST

ended with a neutral task. The physiological signals collected during the session, were then used to train

Chapter 2. Background 15

a personalised model using a state vector machine (SVM) algorithm. An individualised detection model

takes into consideration variations in human response to stress and avoids producing a generalised model

for the population. Thus, the researchers were able to obtain promising results with detection accuracies

of around 80% and precision of over 80% [74].

Our literature search showed that SVM was a popular choice for use with detection of anxiety

using physiological signals. C. Liu presents a review paper investigating the performance of machine

learning algorithms such as K-Nearest Neighbours (KNN), Regression Tress (RT), Bayesian Network

(BNT), and Support Vector Machine (SVM) in affect recognition using physiological markers [75]. The

researcher concluded that SVM and RT had significantly superior performance at detection of anxiety

with accuracies of over 88% [75]. A limitation to these supervised classifiers is that they must be

individually trained for each participant. This is impractically in naturalistic setting, as mincing the

experimental protocol to collect the training dataset will be cumbersome and hinder the adoption of such

technologies. Especially when developing devices for children with autism; the system must be capable

of adjusting and learning an individuals physiological response as they vary vastly across the spectrum

but still correspond to anxiety.

F. Sun presents a paper in which they incorporate activity-awareness during stress detection [76].

The researcher employed a similar protocol as V. Sandulescu [74], in which they collected the participants

heart rate and galvanic skin response (GSR). Additionally, they had incorporated accelerometer sensors

to detect the participants level of activity or motion. Participants were asked to sit, stand, and walk

while they completed neutral and stress inducing tasks. In this case, researchers choose the Stroop

test and mental arithmetics as the stressful tasks. Using the collected data, the research compared the

performance of machine learning algorithms, decision tree, Bayesian network, and SVM, in classifying

stressful and baseline states during different levels of activity. Unlike previous researchers, F. Sun

constructed a generalised model by training the system on a subset of the participants. Their results

showed that SVM significantly outperformed other algorithms, however; they were only able to obtain

accuracies above 80% when training on 18 out of the 20 participants. This outlines another limitation

of these supervised learning algorithms; the heavy dependency on the training dataset produces models

that are incapable of adjusting and relearning as new data is collected from participant.

The paper presented by O. N. Mozos [77], follows a similar experimental protocol as presented by

F. Sun [76]. However; O. N. Mozos compared the SVM to other algorithms such as AdaBoost and K-

nearest neighbor. Also, personalised models were trained for each participant unlike F. Suns generalised

model. The data set consisted of several features extracted from the four incorporated sensors; photo-

plethysmography (PPG) was used to obtain a measurement on heart rate, electrodermal activity (EDA),

accelerometer to detect activity level, and a microphone to record speech. The results showed that the

AdaBoost algorithm could detect anxiety at an accuracy of 94% while utilising all the sensor and without

the environmental (motion and voice) sensors, the accuracy drops to 79%. This provides the evidence

that incorporation of motion-awareness does improve the detection rate of stress in more realistic settings.

The ability to detect stress while in motion is a critical advancement in this technology as movement

induces physiological changes, such as increase in heart rate, that can be incorrectly classified as anxiety

if the system is unaware of the participants activity level. Incorporating environmental awareness enables

the algorithm to better perform in naturalistic environments in which a plethora of stimuli can induce

similar physiological changes to anxiety or stress.

Presently, to our knowledge there has been limited research into a language-free and objection anxiety

Chapter 2. Background 16

detection tool for children with autism. The papers presented previous have focused on individuals aged

between 18 and 39. In [78], A. Kushki explored the relationship between the autonomic nervous system

and anxiety in children with ASD, and the potential of using physiological markers such as heart rate

and EDA to detect anxiety. Following up from this research, A. Kushki developed a Kalman filter

based algorithm that analyses heart rate trends to automatically detect periods of heightened levels

of anxiety [22]. The use of the Kalman filter enables the algorithm to develop a personalised model

based on the childs physiological response to stress. Additionally, is capable of learning and adjust

as new measurements are collected from the sensors. This overcomes the major limitations of previous

algorithms and proves similar detection rates can be achieved among youth with ASD. In [22], A. Kushki

had obtained a specificity and sensitivity of 92% and 99% respectively. However; this was conducted

while the participant was at rest and does not incorporate any sense of motion-awareness into the

algorithm.

2.4.2 Kalman Filter

The Kalman filter assumes a stochastic linear dynamic system that is modeled by a difference equation

with additive white gaussian noise [79]. This linear-gaussian assumption is true for both the state model

xk and measurement model yk.

xk+1 = F (k)xk + wk (2.5)

yk = H(k)xk + υk (2.6)

In the above equation, the process noise wk and measurement noise vk, are assumed to be mutually

independent and gaussian with a covariance of Qk and Rk respectively [79][80]. The Kalman filter

fundamentally, predicts the state model and corrects the predictions based on new observations zk [80].

An overview of the Kalman filter is show graphically in Figure 2.10.

We begin at the first iteration of the Kalman filter; the initial conditions are unknown and are mod-

elled as a random variable with known Gaussian mean and variance [80]. Thus, the initial prior estimate

x−0 and covariance P−0 used to calculate the Kalman filter gain, through the following equation,

Gk = P−(k+1|k)(P−(k+1|k) +Rk)

−1(2.7)

where Rk is the measurement covariance [80]. The Kalman gain modulates the update of the state

estimation; noisy measurements associated with a large covariance, are attenuated to reduce outliers

influencing the state update [21]. On the other hand, prior state predictions with large associated co-

variances are corrected with observations zk amplified by the Kalman gain [21]. Thus, Kalman gain is

integral in calculating the a priori estimate xk and covariance Pk as part of the update step,

xk+1|k = x−k|k +Gkζk (2.8)

where the difference between prior state estimate x−k and the system measurement zk , is referred to as

Chapter 2. Background 17

Figure 2.10: Overview of Kalman Filter

innovation ζk. The innovation sequence follows the gaussian assumption, and is considered to be zero

mean and white [21].

The prediction step or descriptively named the time update step,

x−k|k = xk−1|k−1 (2.9)

translates the past a prior estimation into the present to form the new prior state prediction and covari-

ance at time k [21]. These steps are repeated at the frequency of new measurements to calculate the

state of the system in real-time [21].

2.4.3 Maneuver Detection

A maneuver is the deviation of the observation from a baseline, that is induced by changes in environ-

mental factors [21]. The Kalman filter tracks such systems by including a control input parameter in

the system model [21]. The control input augments the state estimate based on environmental changes

to accurately track the system during a maneuver [21].

Dynamic systems on the other hand, are typically not fully defined as the control inputs are usually

unknown [21]. Especially while tracking anxiety, the environmental factors that influence onset of anxiety

are neurological and vary vastly across the population. Thus, physiological measurements are used as a

proxy for anxiety detection [21].

The onset of arousal or increased activity level are assumed to follow a binomial distribution, such

Chapter 2. Background 18

that the state of the system takes the form of one of two possible outcome {baseline = 0; elevated = 1}.A threshold τ is chosen to handle the decision of switching states. τ is experimentally selected such that

P (εk < τ) = 1− α (2.10)

the probability of a measurement less than the threshold, is equal to a predefined confidence intervalα.

The maneuver onset is based on the innovation ζk exceeding the threshold τ . The innovation tracks

deviations in observation as compared to the state estimate, and is assumed to be white gaussian with

zero mean. The maneuver is normalised based on a slide-standard score

εk =µζ − ζkσζ

(2.11)

ζk =1

N

k−Wm∑i=k

ζi (2.12)

σζ =

√√√√ 1

N − 1

k∑i=0

|ζi − µζ |2 (2.13)

µζ =1

N

k∑i=0

ζi (2.14)

where the innovation mean ζk is calculated over a sliding window of length Wm, and the long-term

innovation mean σζ , and standard deviation µζ considers all innovation samples till time k.

2.4.4 Kalman Filter Modifications

The Kalman filter models a state change by an adjustment of the measurement noise Rk. Upon the

normalized innovation k exceeding the threshold , the system changes from a baseline state to an elevated

state with increased measurement noise by a factor of N . This results in noisy or outlier observations

from influencing the state estimate. Equation 6-2.16 incorporate the measurement noise switching into

the Kalman gain calculations

Σk =

Rk if εk = 0

ρRk if εk = 1(2.15)

Gk = P−(k+1|k)(P−(k+1|k) + Σk)

−1(2.16)

2.5 Multimodal Kalman Filter

The interactive multimodal Kalman (IMM) filter is a practical implementation of a multimodal filter,

which assumes that a system exists in one of r modes of operation. A Bayesian framework provides the

capability of switching between each operating mode based on a likelihood function. The IMM algorithm

is structured to be computational less intensive due to the fact that the filter starts each iteration with

Chapter 2. Background 19

a single mixed estimate,and covariance from r Kalman filters; where r is the number of modes assigned

to model the system. An overview of the IMM algorithm is shown in Figure 2.11.

Figure 2.11: Overview of the proposed anxiety detection algorithm

The first stage of the IMM algorithm, calculates the probability for transition from the current mode

i to mode j; this is referred to as the mixing probability µi|jk−1 with dimension rxr. The mixing proba-

bility is written as

µi|jk−1 =

1

cjµik−1pij (2.17)

where the transition probability pij , is chosen as design parameter to dictate the behaviour of the algo-

rithm during mode switching. The mixing probability is normalised by cj that is defined as

cj =

r∑i=1

µik−1pj (2.18)

The IMM algorithm achieves computational efficiency through mixing the state estimates and co-

Chapter 2. Background 20

variances of each filter. Thus, for a cycle of the IMM algorithm a single mixed state estimate x0jk−1, and

mixed covariance P 0jk−1 is required; these are calculated using the following equations,

x0jk−1 =

r∑i=0

xik−1µi|jk−1 (2.19)

P 70jk−1 =

r∑i=0

µ(k − 1)i|j{P ik−1[xik−1 − x0jk ]2} (2.20)

The previous state estimation x(k− 1)i and covariance P ik−1 from each Kalman filter i, are mixed based

on the mixing probability µi|jk−1. The mixed covariance P 0j

k−1 and state estimate x0jk−1 form the a priori

inputs to the Kalman filter. The Kalman filter that implementation are modified with the steps detailed

in section 2.4.4, that ensured adaptability against noisy measurements affecting the prediction model.

The final step is to update the mode mixing probability based on the likelihood the system is in mode

j. Following the gaussian assumption, the probabilities for each observation zk is derived from a normal

distribution with mean of a posterior state estimate xjk and variance Sjk. This is mathematical written as,

Λjk = N [zk; xjk, Sjk] (2.21)

µjk =1

cΛjk cj (2.22)

where Sjk is the variance of innovation ζk , and c is a normalization parameter.

c =

r∑j=0

Λjk cj (2.23)

Chapter 3

Methods

A research methodology was developed to collect physiological and movement data of participants to

analyses their heart rate response during a baseline task, as compared to a stressor task. The effect of

motion during these tasks was also measured. In this section we present our test protocol, from which

data was collected to tune the performance of the anxiety detection algorithm.

3.1 Participants

Participants for the study were recruited from the Province of Ontario Neurodevelopmental Disorders

(POND) database, and were chosen based on these criteria:

The inclusion criteria were as follows:

• A primary diagnosis of ASD using DSM-V criteria, and supported supported by the Autism Di-

agnostic Observation Schedule (ADOS), and the Autism Diagnostic Interview - revised (ADI-R)

• Full scale IQ of 50 or higher

• Normal, or corrected to normal, hearing, and vision

• Able to communicate in English

• Between the ages of 8 and 16

The exclusion criteria were as follows:

• Unable to tolerate ECG electrodes on the chest

• Unable to walk or stand on a treadmill

• Use of Beta-blockers

3.2 Data Collection

Fifteen participants were invited to attend a test session, during which physiological, and behavioural

measurements were captured. Physiological signals were captured through sensors that were mounted

on the participant’s chest.

21

Chapter 3. Methods 22

3.2.1 Sensor Measurements

The Shimmer 2r from Shimmer technologies, was the primary device used to collect physiological, and

motion data. It consisted of an ECG acquisition system, accelerometer, and wireless bluetooth capabil-

ities that allowed for untethered, wireless communication to a data collection computer. Gel-electrodes

where attached to four loci on the chest ( appendix-D contains the reference guide used for the placement

of these electrodes), leads were then connected the electrodes to the associated channels on the Shimmer

2r. This was then mounted on the chest via a elastic chest band; Figure 3.1 shows a typical placement

of electrodes, and the Shimmer device on a participant.

Figure 3.1: Shimmer 2r chest strap and typical placement of ECG electrodes [2]

Additionally, on-board the Shimmer 2r is a tri-axial accelerometer that is capable of measuring changes

in acceleration along the X, Y, and Z directions. The Shimmer device was positioned to capture during

the torso movements of the participants. The acceleration data and ECG data is samples at 250Hz and

transferred to a data collection computer through a wireless Bluetooth protocol.

The ECG signal was collected using a four-electrode system, the need for such precision was due

to the degrading effect motion artefacts can have on the signal-to-noise ratio. A four-electrode system

provides two leads of ECG signals from associate pairs of electrodes, and a fourth lead is used as a

right-leg drive to remove common mode noise present in the ECG signal.

3.2.2 Behavioural Questionnaire

Questionnaires were used to evaluate the participants’ experience of the anxiety in response to the

stressor task. The state-trait anxiety inventory questionnaire is a tool widely used in physiological

studies to quantify one feeling either at present (state) or generally (trait). For this study a short-form

of the STAI was employed [81], consisting of six questions to which the participant had to respond

with a number between 1 and 4. One representing a disagreement of how they feel as compared to the

question asked, and 4 being in total agreement. Appendix-A contains the form used to collect responses

for the STAI questionnaire. The questions were verbally administrated as we wanted to ensure the

baseline heart for a given motion level was maintained, prior to start of the stroop test. The parents or

guardians of each participants were also asked to complete a child & adolescent symptom inventory-5

Chapter 3. Methods 23

(CASI-5) questionnaire [82], this was used to evaluate their child’s anxiety symptoms. In addition, we

queried the POND database for the participant’s full-scale IQ. Three of the participants had IQ scores

based on the Wechsler Abbreviated Scale of Intelligence (WSI-I), StanfordBinet intelligence scales, and

Wechsler Preschool and Primary Scale of Intelligence (WPPSI-III). Additional, the Autism Diagnostic

Observation Schedule (ADOS), Social Communication Questionnaire (SCQ), and Revised Children’s

Anxiety and Depression Scale (RCAD), that were previously administered, and recorded.

3.3 Experimental Protocol

Participants and their parents or guardians, are provided an informed consent. In the cases where the

child was unable consent, an assent was obtained. They were then introduced into a room containing

the data collection computers, and a treadmill as shown in Figure 3.2.

Figure 3.2: Testing Environment Layout

The data collection computers recorded signals that were being wirelessly transmitted from the

Shimmer device, mounted on the chest of the participant. During breaks between each stage of the

protocol, participants were asked to sit in-front of a monitor to view a calming animated video clip.

After which they are asked to mount the treadmill and follow the tasks being displayed on the monitor

in front of them. The treadmill was then set to speeds appropriate to the activity level being tested;

during the standing stage, the treadmill was not turned on, slow walking was set to the first speed setting

supported by the treadmill, and during the fast walking, the treadmill was set to a speed that replicated

the participant’s gait.An overview of the protocol is detailed in Figure 3.3.

The testing session consisted of three stages during which the participants were asked to either stand,

slow walk, or walk at a faster pace on a treadmill. Prior to start of each stage, a resting baseline was

captured while the participant is seated, and engaged in a 5-minute calming childrens movie clip. This

is referred to as the Rest Phase in Figure 3.3, and is present prior to the start of each stage. The testing

protocol starts off with a 15 minute Rest Phase that collects the baseline data required for the initial

training of the anxiety detection algorithm.

During each of the stages, participants were asked to carry out a baseline and stressor activities;

before and after each stressor activity, participants where asked a series of six questions that made up

the State-Trait Anxiety Inventory (STAI). This was used to gauge the effect of the stressor activity in

Chapter 3. Methods 24

Figure 3.3: Experimental Protocol Overview

inducing anxiety. The baseline phase consisted of watching a 5-minute clip from BBCs Planet Earth 2

movie. Clips were chosen specifically to not include scenes that can induce anxiety; rapid changes in

music, violence, or frightening scenes were excluded from any of the clips shown to the participants.

The Stroop Colour-Word Interference test has been previously shown to be effective in inducing

anxiety among children with autism, and has also has been a popular choice among researchers as a

stressor activity [83, 22, 84, 74, 85]. During the test, participants are asked to name the font colour of

the word that is being displayed in front of them. The displayed words are chosen at random, as are the

font colour, from a list of colours: blue, red, green, purple, and yellow. Figure 3.4 is an example of the

stroop test that was presented on the screen during the test session.

Figure 3.4: Stroop test screen example

The five-minute stroop test is divided into five, one-minute blocks. Blocks 1, 3 and 5 were set to present

words at 2-second intervals, and blocks 2 and 3 at 1.25-second intervals. The congruent section (matching

colour name and print colour) were made up of blocks 1, and 5. The rest of the blocks made up the

incongruent section (conflicting colour name and print colour). This protocol was presented in [22], and

showed positive results in inducing anxiety among children with ASD.

Chapter 3. Methods 25

3.4 Pre-processing of Sensor Data

The accelerometry, and electrocardiogram (ECG) signals from each of the participants had to be pro-

cessed to extract signal features that would be used by the anxiety detection algorithm. In this section,

we discuss the steps that were implemented for feature extraction, and conclude with a brief descrip-

tion the techniques used to quantify the classification performance of the motion detection, and anxiety

detection algorithms.

3.4.1 Feature Extraction

The raw acceleration and ECG signals captured during the test sessions, were processed to extract

features required by the anxiety detection algorithm. Figure 3.5 depicts an overview of the feature

extraction step.

Figure 3.5: An overview of the steps required to extract features from the ECG and acceleration signals

We started by processing the acceleration data to analyse the physical movements of the participants.

As mentioned previously, acceleration is measured in three orthogonal axis (X,Y, andZ). The contribu-

tion of each axis was combined to calculate a resultant magnitude vector −→ν , by using the equation in

3.1

−→ν =√

(A2x +A2

y +A2z), (3.1)

where A is a vector of acceleration in each orthogonal plane. The amount of physical activity is encoded

by the peak values of the resultant magnitude vector (−→ν ). This was converted to a time-varying signal

of physical activity σk by applying a moving standard deviation over −→ν .

σk =

√√√√ 1

N − 1

k+w∑i=k−w

|−→ν i − µ|2 (3.2)

µk =1

N

k+w∑i=k−w

−→ν (3.3)

Chapter 3. Methods 26

In the equation 3.2, the window length of 2w was selected experimentally, and µ was the mean of the N

samples within the window. The time-varying envelope sequence σk, formed the input into the motion

detection algorithm.

The input to the multimodal Kalman filter stems from the measured ECG signal. The first step, is to

use a standard Pan-Thompkins peak detection algorithm to identify periods in time for the occurrence

of an R-peak [1]. The time difference between adjacent R-peaks is referred to as R-R intervals (RRk),

and is inversely proportional to the heart rate. The low frequency trend of the R-R interval is correlated

to the activities of the sympathetic nervous system [86]; a moving average filter of the form

zk =1

w

k∑i=k−wrr

RRk, (3.4)

was applied to the RR-intervals to extract the underlying slow-varying trend zk . The window size wrr,

was chosen experimentally. zk was the measurement input to the multimodal Kalman filter, along with

the physical activity level from the motion detection algorithm σk.

3.5 Performance Assessment

Sensitivity, specificity, and accuracy are parameters used to evaluate the performance of the filter in

classifying baseline and condition states. Each of the metrics were defined as

Sensitivity =TP

TP + FN

Specificity =TN

TN + FP

Accuracy =TP + TN

TP + TN + FP + FN

where true positive TP , true negative TN , false positive FP , and false negative FN are calculated

against a truth signal. This is a sequence containing the correct classification of state at every point

in time. The motion detection truth signal was set to high during time periods that the participant

was asked to walk. On the other hand, anxiety periods were identified as an increase of more than two

beats/minute during the stoop test, from a baseline state measurement. As for the motion detection

algorithm, markers were placed in the sampling data to indicate the type of motion the participant was

engaging in. This was used as the validation signal to evaluate the performance of the motion detection

algorithm. The results of the sensitivity analysis were used to inform design decisions in optimizing the

performance of both the anxiety detection and motion detection algorithm.

Chapter 4

Filter Design

Heart rate generally changes to the autonomic nervous system (ANS) in response to environmental or

internal stimuli; such as anxiety that is one such stimulus that induces an acceleration in heart rate.

However; the activation of the ANS is not unique to anxiety, and can also be observed to occur in

response to physical movements. This especially creates a challenge to tease out the effects of physical

activity and anxiety to changes in heart rate. This poses a significant challenge to design a system

capable of anxiety detection in naturalistic conditions.

In this chapter, I propose a novel algorithm that is able to detect anxiety in the presence of motion,

and at rest. This algorithm builds upon the concepts of the implementation of a multimodal kalman

filter, which assumes that the system always operates in one of r modes. The algorithm is able to jump

between each mode based on the likelihood probability which dictates the mode that best describes the

current system observation. The proposed algorithm assumes that the heart rate measurement exists

in one of two modes; a baseline mode that assumes a lack of motion, and a motion mode that assumes

the presence of physical activity. Each of the two modes are associated with a modified Kalman filter,

as mentioned in section 2.4.4, for which the parameters are customised to represent the environmental

state in which the filter is operating.

Figure 4.1 summaries the algorithm for which the inputs to the system are an ECG, and acceleration

signal. These inputs are processed by the feature extraction block (detailed in section 3.3.1) which

produces the heart rate zk, and acceleration resultant vector σk signals; this forms the input to the

multimodal Kalman filter and motion detection algorithms respectively. The final block of the system,

combines the output from each of the two algorithm to produce a binary signal of anxiety, indicating

whether or not anxiety has been detected. In this section, we dwell into the details of the proposed

anxiety detection algorithm.

4.1 Multimodal Kalman Filter

The interative multimodal Kalman filter (IMM) is a practical implementation of a multimodal filter,

that is optimised to be computationally less intensive. This is achieved through the mixing of the

initial estimates, and covariance from each of the two kalman filters, in a ratio dictated by the mixing

probability. The estimate that is representative of the measurement observation would carry a greater

27

Chapter 4. Filter Design 28

Figure 4.1: Overview of the proposed anxiety detection algorithm

influence during the calculation of the mixed estimate, this is possible as the mixing probability is based

on the likelihood function that is proportional to innovation or difference between each of the filter’s

estimate, and the measurement observation.

During the testing protocol, the initial baseline during which the participant is seated and watching a

video clip, is used to train each of the Kalman filters, in producing a prediction model that would be used

to track the baseline to which the filters will compare, to detect the onset of anxiety. The training signal

for the motion-Kalman-filter is increased by a fixed value, determined through the heart rate response

analysis conducted as part of the first research question. The two models were trained with differing

measurement data in the assumption that the baseline-Kalman-filter would model the heart rate during

the lack of physical motion, and the motion-Kalman-filter would model the behaviour similar to when

there is a presence of physical activity.

The motion and baseline Kalman filters are assumed to follow a linear model, with additive white

Gaussian noise. This is mathematically modelled through the following system and measurement equa-

tions

xk = xk−1 + wk (4.1)

yk = xk + vk (4.2)

where the state estimate xk is the slow varying R-R trend, and yk is the measurement model. The

process noise wk and measurement noise vk are assumed to be independent, and defined by a Gaussian

distribution with a variance of Qanx and Ranx. The motion-kalman filter parameters Qanx and Ranx

were defined to be a multiple of the value of these parameters assigned in the baseline Kalman filter. This

was implemented to account for the elevated baseline levels during motions, through the adjustments

of these parameters the motion-kalman filter is expected avoid false positives due to physical activities.

The optimal multiplier was chosen experimentally.

4.2 Activity Detection

A tri-axial accelerometers record the movements in the X,Y, and Z directions, and are combined to

calculate the resultant magnitude vector; this occurs in the feature extraction block, that was previously

mentioned in detail. Figure 4.2 is an overview of the activity detection algorithm, that is subdivided

Chapter 4. Filter Design 29

into the Kalman filter, and thresholding blocks.

Figure 4.2: Overview of the proposed anxiety detection algorithm

The input measurement for the Kalman filter is the acceleration resultant amplitude σk that was

extracted by applying a standard deviation over the acceleration resultant magnitude vector −→ν , as de-

tailed in the feature extraction block. The following equations were used to process the raw tri-axial

acceleration signal A

−→ν =√

(A2x +A2

y +A2z) (4.3)

σk =

√√√√ 1

N − 1

k+wstd∑i=k−wstd

|−→ν i − µk|2 (4.4)

µk =1

N

k+wstd∑i=k−wstd

−→ν (4.5)

The window length 2wstd of the moving standard deviation is a parameter that is selected based on the

sensitivity analysis of the detection algorithm. Depending on the nature of σk, a moving average filter

can be applied to account for outliers in the signal, especially high frequency noise that is evidently

unrealisable through physical movement. A window length of ws was selected experimentally.

The Kalman filter is implemented with modifications that ensures that outlier observations, usually

due to the onset of motion, are not incorporated into the state estimate. This is achieved through

changing the measurement noise Rk by a factor of ρ, as shown in the following equations

Σk =

Rk, σk = 0

ρmRk, σk = 1(4.6)

Gk = P−(k+1|k)(P−(k+1|k) + Σk)

−1(4.7)

where σk is altered based on the motion condition; this affects the Kalman gain Gk that determines the

amount by which the state estimate is updated. During the onset of motion the measurement values are

elevated past the baseline, thus an appropriate ρm value, ensures these observations do not affect the

Chapter 4. Filter Design 30

state estimate. The value of the multiplier ρm is determined experimentally.

The innovation ζ is the difference between the baseline estimate, and the measurement observation; it

tracks the deviation of the observation from a baseline, as dictated by the state estimate. The innovation

value was normalised using a set of standard score equations

ζk =µζ − εkσk

(4.8)

εk =1

N

k−Wsm∑i=k

ζi (4.9)

σk =

√√√√ 1

N − 1

k∑i=0

|ζi − µζ |2 (4.10)

µk =1

N

k∑i=0

ζi (4.11)

where µk and σk are the long-term mean and standard division of all the observations. ζk is a short

term mean of window length Wsm, that is optimised experimentally. Sudden changes in the innovation

values are indicative of the onset of motion activity. A thresholding scheme

α(ζ) =

0, ζ ≤ τm1, ζ > τm

(4.12)

is applied such that values of ζ above the threshold τm, are labelled as motion, and the Kalman filter

ignores these measurements from updating the prediction model. The motion levels α(ζ) is either in a

state of motion (α(ζ) = 1), or in a baseline state with little to no physical activity (α(ζ) = 0). The

threshold was set experimentally through a sensitivity analysis to find an optimal value, the results are

presented in chapter five.

4.3 Arousal Detection

As the onset of physical activity requires the detection algorithm to alter parameters to account for such

an environmental change, the arousal detection algorithm provides the ability for the detection algorithm

to alter thresholds based on the onset of motion. This requires an input of the activity level (αk) from

the motion detection algorithm, and innovations (ζik) from each of two Kalman filter implemented as

part of the IMM algorithm; the innovation from the rest-Kalman filter is represented by ζ1k , and the

motion-kalman filer innovation is shown as ζ2k . The output of the arousal detection is a binary signal of

the state of anxiety. Figure 4.3 shows an overview of the arousal detection algorithm that is subdivided

into the activity level switching, normalization, and thresholding modules.

A state of rest, or low levels of physical activity, is indicated by a low value of αk; during this state the

rest-Kalman filter innovation (ζ1k), acts as the input to the normalization module. With the onset of

motion, the valued of αk is set to high by the motion detection algorithm; this results in the activity

Chapter 4. Filter Design 31

Figure 4.3: Overview of the arousal detection algorithm

level switching module to choice the motion-Kalman filter innovation ζ2k as the input for the preceding

modules. This is mathematically represented in equation 4.14:

ζ′

k = (1− αk)ζ1k + αkζ2k , (4.13)

where ζ′

k is the input into the normalisation module. In addition to innovation switching, a selection of

threshold (τα) is also based on the level of activity αk. A switching scheme was implement to switch the

threshold level as follows

τα =

τanx, αk = 0

2τanx, αk = 1(4.14)

At baseline levels of motion, the threshold τα is set to a lower value τanx; during high activity levels, the

threshold is switched to 2τanx, in order to account for the increase in baseline due to physical activity.

An increase in the threshold will account for the heart rate increase experience due to motion, and avoid

detection such an acceleration of heart rate as anxiety. The value of τanx was set experimentally after

conducting a sensitivity analysis on the anxiety detection algorithm. The threshold switching scheme

was implemented as in equation 4.15 :

τα = (1− αk)τ + αk2τ (4.15)

Threshold τα is used by the thresholding module to determine the onset of anxiety based on the condi-

tion that the normalised innovation ζ′′

k is larger than τα. The selected innovation ζ′′

k is normalised with

the standard score equations

Chapter 4. Filter Design 32

ζ′′

k =µζ − ζ

k

σζ(4.16)

ζ′

k =1

N

k−Wn∑i=k

ζ′

i (4.17)

σζ =

√√√√ 1

N − 1

k∑i=0

|ζ ′i − µζ |2 (4.18)

µζ =1

N

k∑i=0

ζ′

i (4.19)

The window size Wn of the moving average filter determines the nature of the normalized signal; a

large window would remove much of the white gaussian noise, by compromising high frequency details.

A smaller window length, would retain such details, but would be susceptible to noise. The optimal

value was chosen after performing a series of sensitivity analyses on performance of the anxiety detection

algorithm at different window sizes. The results of the analysis are presented in chapter 5.

Chapter 5

Results

The proposed filter was evaluated on the dataset described in Chapter 4. Fifteen participants had com-

pleted our experimental protocol and were between the ages of 9 and 16 years old; Table 5.1 summaries

the demographics of the participants and characterises their anxiety and autism symptoms.

Table 5.1: Participant Demographics

Domain Measure Mean± Standard Deviation

Demographics Age 14± 1.77Sex(Male:Female) 9 : 6

Full-scale IQ 89.9± 15.40

ASD Characteristics ADOS (Module-3) 1.9± 0.32ADOS (Module-4) 2± 0

SCQ 20± 7.37

Anxiety Characteristics RCAD - General Anxiety 55.25± 13.02RCAD - Social Anxiety 51.42± 15.72RCAD - Total Anxiety 56.33± 16.10

Social Communication Questionnaire (SCQ) reported for n=15; Autism Diagnostic Observation Schedule (ADOS),module-4 & module-5 reported for n=15; Revised Childrens Anxiety and Depression Scale (RCAD) reported forn=12

The rest of this section details the characterisation of the heart rate response to anxiety and at a baseline,

while standing, slow walking, as well as fast walking. We then present the results on the performance of

the motion, and anxiety detection algorithms.

5.1 Characterisation of Heart Rate Response

The effect of the anxiety stimulus on heart rate was analysed using ANOVA, we assumed a separate

model for each level of motion (standing, slow walking, and fast walking). Figure 5.1 shows the estimated

33

Chapter 5. Results 34

heart rate response, and corresponding standard error of the mean calculated from all fifteen participant

measurements. Additionally, significance values are labelled between the stroop test and baseline tasks,

and for this analysis the α < 0.017 was assumed to indicate significant difference.

Figure 5.1: The average heart rate (N=15) during each of the twelve phases of the testing protocol

Results show a significant effect of the stroop task on the heart rate (F (2, 28) = 14.68, p < 0.0001)

while standing. For this condition, post-hoc analysis revealed significantly higher heart rate during the

stroop test, as compared to the prior-baseline (estimated different −3.79±1.02, t(28) = −3.7, p = 0.0009).

However, when compared to the post-baseline (estimated difference 1.92 ± 0.8, t(28) = 2.38, p = 0.02),

there was not a significant difference. The heart rate values during the two baselines were not significantly

different (estimated difference −1.87 ± 1.5, t(28) = −1.23, p = 0.2275). For the slow-walking condition,

there was also a significant effect of the stroop task on the heart rate (F (2, 28) = 9.51, p < 0.0007). The

post-hoc analysis similarly revealed significantly higher heart rate during the stroop test as compared

to the prior-baseline (estimated different −4.62± 1.07, t(28) = −4.32, p = 0.0002) and post-baseline (es-

timated difference 2.85± 1.07, t(28) = 2.67, p = 0.0125). The heart rate values during the two baselines

were again not significantly different (estimated difference −1.76± 1.07, t(28) = −1.65, p = 0.1096). For

the fast-walking condition, ANOVA was used to show a significant effect of the stroop task on the heart

rate (F (2, 28) = 3.78, p < 0.0035), and post-hoc analysis revealed significantly higher heart rate dur-

ing the stroop test as compared to the prior-baseline (estimated different −6.0481 ± 2.2053, t(28) =

−2.74, p = 0.0105) and post-baseline (estimated difference 4.87 ± 2.00, t(28) = 2.43, p = 0.0215).

The heart rate values during the two baselines were not significantly different (estimated difference

−1.17± 0.88, t(28) = −1.33, p = 0.1093).

Chapter 5. Results 35

5.2 STAI Questionnaire Results

The State Trait Anxiety Inventory (STAI) questionnaire verbally administrated before and after each

Stroop test, to gauge the individual’s emotional response to the Stroop test. The score for 11 participants,

were average for each of the STAI questionnaire; one being administer prior and after the Stroop test.

The STAI questionnaire contained questions such as ’I feel calm’, ’I am relaxed’, and ’I feel content’;

that typically received a score of 4; suggested that the participants were in agreeance with the question.

On the other hand, questions like ’I am tense’, ’I feel upset’, and ’I am worried’ received answers of 1;

indicating that the participants strongly disagreed with the question.

Due to the nature of experimental protocol, a shorted version of the STAI was chosen and had to be

verbally administered. The intent of collecting this data was to provide supportive evidence that indeed

children with autism felt anxious due to the stroop test. However, the results show an insignificant

difference between the pre and post Stroop test scores.

5.3 Detection Algorithm Performance

In this section we present the results for the motion detection algorithm, multi-modal Kalman filter,

and anxiety detection algorithm. A sensitivity analysis was conducted to examine the effects of different

parameters on the performance of these filters. The accuracy, specificity, and sensitivity presented are

averages of the performance of the filter on data from all fifteen participants.

Chapter 5. Results 36

5.3.1 Evaluation of Motion Detection

We explore the effect of the threshold τm, innovation window length wIM , process covariance Qm, and

measurement covariance multiplier ρm on the specificity, sensitivity, and accuracy of he motion detection

algorithm. These results informed the optimal value to assign each of the parameters to obtain the peak

performance of the Kalman filter.

Effect of threshold

Figure 5.2 show the effect threshold τm on the specificity, sensitivity, and accuracy of the motion detection

Kalman filter. As can be seen, lower values of threshold (τ) result in low specificity and accuracy

values, and high sensitivity. At higher values of threshold, we observer low sensitivity values, and high

specificity, and accuracy. The optimal threshold is evident at τm = 0, producing a sensitivity, specificity,

and accuracy of 89.4 %, based on Figure 5.2.

Effect of Innovation Window Length

Figure 5.3 depicts the effect of the window length wim; values of 10, 100, and 500 were examined through

comparing corresponding receiver-operator curves. As seen, the performance improves as the window

length wim decrease; the optimal value was found to be at wim = 10.

Effect of Process Covariance

Figure 5.4 shows three character-operator curves for Qm values of 0, 0.01R, and 0.001R. Based on Figure

5.4, Q = 0 results in the best performance of the motion detection algorithm.

Effect of Measurement Covariance Multiplier

The measurement covariance multiplier (ρm), is a parameter that is chosen to enlarge the variance

assumed by the measurement model. Three values of R = 5, 10, 15 were tested, and it was found that

there was little influence on the performance of the filter. A value of R = 5, was chosen for the remaining

analysis.

Effect of Measurement Moving Average Window

Figure 5.6 compares the effect on filter performance due to window lengths 5,100,500, and 1000. Based

on the figure, a lower window length improves the detection performance of the filter; an optimal value

of Ws = 5 was found to produce the best performance.

Effect of Measurement Moving Standard Deviation Window

Figure 5.7 shows receiver operator curves for window lengths wstd of 5, 10, 15. As seen in the figure,

there was little change to the performance due to the window length. For the remaining of this analysis

a value of wstd = 5 was chosen.

Chapter 5. Results 37

Figure 5.2: Effect of threshold τm on the classificationof motion (N = 15, Qm = 0, ρm = 5, ws = 5, wstd =5, wIM = 10 and overlap=50%)

Figure 5.3: Effect of innovation moving average win-dow length wIM (N = 15, Qm = 0, ρm = 5, ws =5, wstd = 5, τm = 0 and overlap = 50%)

5.3.2 Anxiety Detection Algorithm Evaluation

In the following section, the performance of the complete detection algorithm is examined; this included

the multimodal Kalman filter, and the arousal detection algorithm which combines the motion level,

and innovation outputs from the IMM filter to detect anxiety level. The specificity, sensitivity, and

accuracy were calculated for various values of threshold τanx, process covariance Qanx, measurement

covariance multiplier ρanx, measurement moving average window length Wrr, and innovation moving

average window Wn.

Effect of threshold

Figure 5.8 depicts the effect of threshold τanx on the sensitivity, specificity and accuracy of the arousal

detection algorithm. A low threshold value produces low accuracy and specificity results, and a high

sensitivity value. The opposite is observed for high values of threshold; the optimal threshold τanx = 0.55

produced the best performance. As seen in Figure 5.8, the sensitivity, specificity, and accuracy of 91.5%.

Chapter 5. Results 38

Figure 5.4: Effect of Process Covariance Qm on theperformance of the motion detection (N = 15, τm =0, ρm = 5, ws = 5, wstd = 5, wIM = 10 and over-lap=50%)

Figure 5.5: Effect of Measurement Covariance Mul-tiplier ρm on the performance of the motion detection(N = 15, Qm = 0, τm = 0, ws = 5, wstd = 5, wIM = 10and overlap=50%)

The arousal detection module (section 4.3) implements an two-stage thresholding scheme, that as-

signed a threshold of 2τanx in the presence of motion, and τanx with the lack of movement. The analysis

was repeated with a single threshold, and found that the sensitivity, specificity, and accuracy dropped

to 87%. Thus, the threshold multiplier of 2 was selected for the remaining analysis.

Effect of Process Covariance

Figure 6 shows the effect on the performance of the IMM filter due to Qanx of 0.001, 0.01, 0.1, and

1. As can be seen on the figure, a smaller value of Q, tends to perform better; an optimal value of

Qanx = 0.001 was observed. As mentioned previously, the motion-Kalman filter was designed with a

multiple of the Qanx selected for the baseline-Kalman filter. In this analysis a multiple of 2 was chosen,

repeating this analysis with a multiple of 1, resulted in a decline of the filter performance.

Chapter 5. Results 39

Figure 5.6: Effect of the moving average window lengthws applied to the acceleration resultant magnitude vec-tor, on the performance of the motion detection algo-rithm (N = 15, Qm = 0, ρm = 5, τm = 0, wstd =5, wIM = 10 and overlap=50%)

Figure 5.7: Effect of moving standard deviation win-dow length Wstd applied to the raw accelerometer signal,on the performance of the motion detection algorithm(N = 15, Qm = 0, ρm = 5, ws = 5, τm = 0, wIM = 10and overlap=50%)

Effect of Measurement Covariance Multiplier

Figure 6 examines the effect of measurement covariance multiplier ρanx at 0.01, 1, 2, and 4 on the

performance of the IMM algorithm. As can be seen on the figure, the optimal value of ρanx = 1

produced the best results.

Effect of Measurement Moving Average Window Length

The slow-varying trend of the R-R intervals formed the input into the IMM algorithm. A moving

average window of length Wm, is used to extract such a trend. Figure 5.11 depicts the effect on filter

performance due to window lengths of Wm = 5, 10, 20, and an optimal value of Wm = 5 was found to

produce significantly superior results.

Effect of Innovation Moving Average Window

Figure 5.12 depicts multiple receiver-operator curves, that compare the effect of different values of Wn

on the performance of the filter. Visually we noted that the performance peaked when a window length

of Wn = 25 was applied.

Chapter 5. Results 40

Figure 5.8: Effect of the threshold τanx on the per-formance of the arousal detection algorithm (N =15, Qanx = 0.001, ρanx = 1, wRR = 10, wn = 50, andoverlap = 50%)

Figure 5.9: Effect of measurement covariance Qanx

on the performance of the arousal detection algorithm(N = 15, τanx = 0.55, ρanx = 1, wRR = 10, wn = 50, andoverlap = 50%)

Chapter 5. Results 41

Figure 5.10: Effect of the measurement covariance mul-tiplier ρanx on the performance of the arousal detectionalgorithm (N = 15, Qanx = 0.001, τanx = 0.55, wRR =10, wn = 50, and overlap = 50%)

Figure 5.11: Effect of R-R interval moving averagewindow WRR on the performance of the arousal detec-tion algorithm (N = 15, Qanx = 0.001, ρanx = 1, τanx =0.55, wn = 50, and overlap = 50%)

Chapter 5. Results 42

Figure 5.12: Effect of innovation moving average window wn on the performance of the arousal detectionalgorithm (N = 15, Qanx = 0.001, ρanx = 1, wRR = 10, τanx = 0.55, and overlap = 50%)

Chapter 5. Results 43

5.3.3 Anxiety Detection Algorithm Optimised Results

Using the optimal parameters discussed earlier, the performance of the anxiety detection, and motion

detection algorithms are summarised in Table 5.2. The output from each of the filters is a binary signal

that is indicative of either motion, anxiety, or their respective baseline state. As an example, Figure

5.13 depicts a resultant accelerometry signal (black) and the output of the motion detection algorithm

(red), for one participant. Figure 5.14 depicts the zeta signal ζ′′

k (blue), and the detected anxiety level

(red).

Table 5.2: Summary of Optimised Performance

Algorithm Threshold Accuracy Specificity Sensitivity

Motion Detection 0 89.40% 91.30% 86.72%

Anxiety Detection 0.55 92.70% 92.70% 92.70%

Figure 5.13: Acceleration signal and detected anxiety level by the motion detection algorithm, during baseline(BL) and stroop task (SA), while standing, slow walking, and fast walking

Chapter 5. Results 44

Figure 5.14: Innovation ζ, and identified arousal states during standing still, slow walking, and fast walkingfor one participant

Chapter 6

Discussion

The detection of internal or emotional state based on one’s physiology has been an active topic of research

in the biomedical space [?]. Through machine learning algorithms such as support vector machine, K-

nearest neighbour, and decision tree algorithms, models have been trained to detect heart rate arousal

from heart rate, respiration, perspiration, and other physiological measurements [22, 87, 88, ?, 76].

However, most of these algorithms have been evaluated based on data collected while the participant

is at rest. Physical movement activates the autonomic nervous system, through pathways similar to

that of an anxiety response. The cardiovascular system has been shown to accelerate with the presence

of anxiety, and physical movements. This creates a challenge to detect anxiety states through a proxy

measurement of heart rate, which is not specific to anxiety. In realistic conditions, this results in the

detection of false positives in the presence of physical activity; rendering many existing algorithms

ineffective.

The above mentioned challenge motivated use to develop an automatic, and objective detection of

anxiety in the presence of motion. We developed a protocol to measure heart rate and motion signals

during baseline and stressor tasks, while standing, slow walking,and walking quickly. The rationale for

the standing and fast walking conditions was to evaluate the ability to detect anxiety while at rest, as

well as during the presence of motion. The slow walking stage was introduced to provide a boundary

case during which participants engage in slightly increased level of movement as compared to standing,

but on the threshold to warrant a switch in filter properties. Prior to designing the algorithm, the data

collected from all the participants were analysed using ANOVA to determine the significance of the

change in heart rate between baseline and stroop test tasks. The results showed that indeed there was

a significant increase of heart rate during the stroop test across the three activity levels. These results

suggest that anxiety continued to have an influence on the heart rate, even with the presence of motion.

This finding was crucial as it showed that the heart rate did not saturate due to the presence of physical

activity, and that the onset of anxiety is able to further accelerate the heart rate. This then provided the

scope for detection algorithms to identify heart rate arousal relating to anxiety even with the presence

of motion.

Based on the results from the ANOVA analysis, we were able to proceed with the construction

of an anxiety detection algorithm that is aware of motion. A.Kushi [22] had developed an anxiety

detection algorithm based on the Kalman filter; an unsupervised machine learning algorithm that is able

to estimate the anxiety state based on relatively few samples of measurements. It is personalised to the

45

Chapter 6. Discussion 46

user, as initial observations are used to develop the state model, and subsequent measurement updates,

tune its characteristics in real-time. We built off of these results, by developing an anxiety detection

algorithm through a multimodal approach that assumes two distinctive models; a baseline state model

that represents heart rate patterns in the absence of motion, and a motion state model that is used in

the presence of physical activities. For each of the models, a Kalman filter is trained with the 15-minute

baseline measurements captured at the start of each session. However; the measurements that initially

trained the motion state model, were elevated by 15 beats per minute; this value was chosen based on

the average increase in heart rate experienced between baseline and stroop test across the three activity

levels.

The interactive multimodal Kalman filter is at the core of the anxiety detection algorithm, and was

successfully tuned to detect anxiety in the presence of motion activity, as well as while at rest. The

first optimisation was achieved by selecting an appropriate value for the measurement noise covariance

Qanx, and process noise multiplier constant ρanx for each of the two kalman filters. It was assumed that

the Qanx value for the motion-Kalman filter, would be double that set for the baseline model Kalman

filter; this was to account for the increase in baseline heart rate during motion. A sensitivity analysis

was performed (Figures -) on the IMM algorithm from which Qanx and ρanx were assigned values of

0.001, and 1.0 respectively. The results for ρanx were surprising, as they indicated that the Kalman filter

modification mentioned in Equation is irrelevant for the filters implemented among the IMM algorithm.

Additional optimisations were explored through the effect of different window lengths for the moving

average filters used upon the R-R interval measurement wm and innovation signal wn. Wm determined

the extraction of a slow-varying trend from the R-R interval, that corresponds to the sympathetic nervous

system response of the heart. The optimal trend was achieved at a window length of ten samples, from

which we concluded that the high frequency component of the trend was approximately 12.5Hz. A

window size of 20, reduced the high frequency component to 6.25Hz, and as a result produced very poor

classification rates. This lead us to believe that there in lies information at frequency components at

around 12Hz, that are essential for the detection of anxiety.

Motion awareness was incorporated into the above anxiety detection algorithm, through the proposed

motion detection algorithm, which is based on a modified on the Kalman filter. Several algorithms are

currently available for the detection of motion from accelerometry signals. Our choice to use the Kalman

filter allowed us to incorporate acceleration as a control signal into the IMM algorithm, that supports

the integration of different signals with varied sampling rate. However; due to the complexities of

constructing a state-space to incorporate heart rate, and acceleration signals, we choose to implement

them in parallel. Thus, the Kalman filter had to be optimised independently for the detection of motion

level. Sensitivity analysis provided a means to select the most optimal value for the threshold τm,

innovation window length wIM , process covariance Qm, and measurement covariance multiplier ρm.

The window length wIM determines the high frequency detail that is preserved, often the high frequency

components are assumed to be caused by measurement noise. A moving average filter’s window length

can be selected to remove such types of noise, however; with an increased window length, details essential

for detection could be lost. Thus, the sensitivity analysis provided the most optimal value. Again the

results for ρm showed no effect on the performance of the filter; which indicates that the modification

to the Kalman filter might not be necessary. Lastly, a threshold τm = 0 produced excellent accuracy,

specificity, and sensitivity of 89.4%, 91.3%, and 86.72% respectively.

As mentioned above, several parameters were optimised to obtain current detection performance;

Chapter 6. Discussion 47

however, some of these parameters were interdependent. In order to reduce the complexity of the per-

formance analysis, some of the parameters were selected back on one iteration of adequate performance.

We suspect that some further analysis of these parameters might provide a more optimal value. Another

limitation of the proposed algorithm, is the fixed assumption made on the selection of the process and

measurement covariance within the two Kalman filters of the multimodal algorithm. In order for such

an algorithm to be more adaptable and robust to different environmental conditions, these assumption

potentially need to be dynamically assigned based on the individual.

The final step of optimising the anxiety detection algorithm, was to select a threshold τanx above

which it is assumed to be labelled as the onset of anxiety. The sensitivity analysis showed that a

threshold of (τanx = 0.5) produced a sensitivity, specificity, and accuracy of about 90%. Analysing the

anxiety measurement for each participant (appendix-c), it is evident that the algorithm prematurely

detects anxiety in some occasions. We suspect this is reflective of how the participant felt during the

stressor task. We verified these cases, by correlating the participant’s average heart during each task

(appendix-b) to the anxiety level detected by the proposed algorithm. Results for participants 10 and

11 (Figures 7.25 and 7.26) are observed to be suboptimal. The detection of anxiety by the algorithm

was inconsistent with when the stroop-test was administered. A closer analysis of each participant’s

average heart rate (as shown in figure 7.10 and 7.11) explained this discrepancy. These participant did

not respond as expected to the stroop test, their average heart does not follow our hypothesis which

assumes an elevation of heart rate during the stroop test as compared to the individual’s baseline. Upon

close examination of the detected anxiety level of each individual participant, we noticed that case of

heart rate acceleration prior to the start of the stroop. We suspect that this is due to the anticipation

by the participant of the stroop test. As prior to the start of the stroop test activity, the participants

are asked a series of questions based on the STAI and at times go through a small exercise familiarising

them with the stroop test. Thus, these tasks between the baseline phase and stroop test might explain

for the premature detection of anxiety.

These results suggest a novel automatic anxiety detection algorithm that is capable of identifying

one’s internal state, through a heart rate measurement. This allows for a objective, language and

communication free tool that is able to express one’s level of anxiety. Such a tool is in immediate

need for children who are diagnosed with autism. As anxiety can be debilitating, and exacerbates

core ASD symptoms; cause these children to seclude themselves, and engage in self-harming behaviour.

Current treatment methods for anxiety rely heavily on the patient’s language, and communication ability.

However; autism negatively influences these skills, and thus limiting the efficacy of current treatment

options. This technology could provide has the potential to improve the quality of life for the children,

and their families.

Physiological measurement of heart rate was chosen as a proxy to detect anxiety, as heart monitors

have become increasingly popular among consumer technology. Wearable fitness focused devices are

commonly fitted with sensors capable of detection heart rate from the wrist. This played a roll in

deciding an appropriate physiological measurement, as the focus of this algorithm was to ensure it

could be easily implemented into a wearable device. However, heart rate is influenced by many stimuli

thus limiting the effectiveness in naturalistic environments. Future development of this algorithms,

would focus on the incorporation of other physiological parameters to support the detection of anxiety.

Incorporating such sensors can be done with relative ease through the multimodal kalman filter, which

was originally developed for the inference of state estimates based on data from numerous sensors with

Chapter 6. Discussion 48

variable sampling rate.

Chapter 7

Conclusion

In conclusion, in the this thesis we have presented the heart rate response of children diagnosed with

autism, during baseline and stressor tasks while in motion. The results showed that heart rate does

respond to anxiety conditions, even in the presence of physical activity. With this information, we then

constructed a novel anxiety detection algorithm with activity-awareness. This was a stepping stone in

the development of a technology that could supplement current medical treatments for anxiety. The

autonomous, language-free, and objectivity of the proposed algorithm, could provide an addition form

of communication through which children with autism spectrum disorder are able to navigate their

daily tasks with a bit more ease. Thus, the foundation of this algorithm was based on the processing,

and inferring of anxiety from measurements in real-time. It has the potential to function as a tool in

conjunction with cognitive behaviour therapy (CBT). During an the onset of anxiety, the system would

alert the child that what they are feeling, is actually a response to anxiety, and encourage them to

partake in techniques they had learnt through CBT.

The next steps are to explore the efficacy of this algorithm in an everyday environment; though

the algorithm accounts for motion, there are other stimuli that could manifest similarly to anxiety.

Exploring such stimuli, would improve the robustness, and specificity of detecting anxiety. Additionally,

the incorporation of the anxiety detection algorithm into in a smart-watch, would provide the product

that could significantly help children deal with anxiety during their everyday life. Wearable technologies

have the unique ability to blend into clothing, and unobtrusively monitor the physiology health and

specific physical movement of the user. This could inform the development of future algorithms, that

are more robust against physical movements.

In conclusion, a wearable device with the ability to detect heart rate arousal and inform children of

the onset of anxiety, can drastically improve their quality of life. Children with autism can rely on this

device for guidance during strenuous environments to guide them through activities that could potential

help them overcome their anxiety.

49

Bibliography

[1] J. Pan and W. J. Tompkins, “A Real-Time QRS Detection Algorithm,” IEEE Transactions on

Biomedical Engineering, vol. BME-32, no. 3, pp. 230–236, 1985.

[2] T. K. Borsook, “Shimmer wireless sensor system,” 2011.

[3] American Psychiatric Association, Diagnostic and Statistical Manual of Mental Disorders. American

Psychiatric Association, 5 2013.

[4] D. L. Christensen, J. Baio, K. V. N. Braun, D. Bilder, J. Charles, J. N. Constantino, J. Daniels, M. S.

Durkin, R. T. Fitzgerald, M. Kurzius-Spencer, L.-C. Lee, S. Pettygrove, C. Robinson, E. Schulz,

C. Wells, M. S. Wingate, W. Zahorodny, and M. Yeargin-Allsopp, “Prevalence and Characteristics

of Autism Spectrum Disorder Among Children Aged 8 Years - Autism and Developmental Disabil-

ities Monitoring Network, 11 Sites, United States, 2012.,” Morbidity and mortality weekly report.

Surveillance summaries (Washington, D.C. : 2002), vol. 65, pp. 1–23, 4 2016.

[5] L. Kanner, “Autistic disturbances of affective contact,” 1943.

[6] S. Leekam, M. Prior, and M. Uljarevic, “Restricted and Repetitive Behaviors in Autism Spectrum

Disorders: A Review of Research in the Last Decade.,” Psychol Bull, vol. 137, no. 4, pp. 562–593,

2011.

[7] R. Deth, C. Muratore, J. Benzecry, V. A. Power-Charnitsky, and M. Waly, “How environmental

and genetic factors combine to cause autism: A redox/methylation hypothesis,” NeuroToxicology,

vol. 29, no. 1, pp. 190–201, 2008.

[8] E. Simonoff, A. Pickles, T. Charman, S. Chandler, T. Loucas, and G. Baird, “Psychiatric disorders

in children with autism spectrum disorders: Prevalence, comorbidity, and associated factors in a

population-derived sample,” Journal of the American Academy of Child & Adolescent Psychiatry,

vol. 47, no. 8, pp. 921–929, 2008.

[9] S. White, D. Oswald, T. Ollendick, and L. Scahill, “Anxiety in children and adolescents with autism

spectrum disorders,” Clinical Psychology Review, vol. 29, no. 3, pp. 216–229, 2009.

[10] R. Lang, A. Regester, S. Lauderdale, K. Ashbaugh, and A. Haring, “Treatment of anxiety in

autism spectrum disorders using cognitive behaviour therapy: A systematic review,” Developmental

Neurorehabilitation, vol. 13, no. 1, pp. 53–63, 2010.

[11] A. Gillott, F. Furniss, and W. Ann, “Anxiety in high - functioning children with autism,” Autism,

vol. 5, no. 3, pp. 277–286, 2001.

50

BIBLIOGRAPHY 51

[12] S. B. Helverschou and H. Martinsen, “Anxiety in people diagnosed with autism and intellectual

disability: Recognition and phenomenology,” Research in Autism Spectrum Disorders, vol. 5, no. 1,

pp. 377–387, 2011.

[13] S. H. Spence, “A measure of anxiety symptoms among children,” Behaviour Research and Therapy,

vol. 36, pp. 545–566, 5 1998.

[14] M. Sung, Y. P. Ooi, T. J. Goh, P. Pathy, D. S. S. Fung, R. P. Ang, A. Chua, and C. M. Lam,

“Effects of cognitive-behavioral therapy on anxiety in children with autism spectrum disorders: A

randomized controlled trial,” Child Psychiatry and Human Development, vol. 42, no. 6, pp. 634–649,

2011.

[15] E. J. Costello and A. Angold, Anxiety disorders in children and adolescents. 1995.

[16] R. P. Hobson, J. Ouston, and a. Lee, “Emotion recognition in autism: coordinating faces and

voices.,” Psychological medicine, vol. 18, no. 4, pp. 911–923, 1988.

[17] P. Karthikeyan, M. Murugappan, and S. Yaacob, “A review on stress inducement stimuli for assess-

ing human stress using physiological signals,” Proceedings - 2011 IEEE 7th International Colloquium

on Signal Processing and Its Applications, CSPA 2011, pp. 420–425, 2011.

[18] J. Healey and R. Picard, “Detecting Stress During Real-World Driving Tasks Using Physiological

Sensors,” IEEE Transactions on Intelligent Transportation Systems, vol. 6, pp. 156–166, 6 2005.

[19] A. De Santos Sierra, C. Sanchez Avila, J. Guerra Casanova, and G. Bailador Del Pozo, “A stress-

detection system based on physiological signals and fuzzy logic,” IEEE Transactions on Industrial

Electronics, vol. 58, no. 10, pp. 4857–4865, 2011.

[20] Y. Shi, M. Nguyen, P. Blitz, and B. French, “Personalized stress detection from physiological mea-

surements,” International Symposium on Quality of Life Technology, no. September 2015, pp. 28–29,

2010.

[21] Y. Bar-Shalom, X.-R. Li, and T. Kirubarajan, Estimation with Applications to Tracking and Navi-

gation, vol. 9. 2001.

[22] A. Kushki, A. Khan, J. Brian, and E. Anagnostou, “A Kalman Filtering Framework for Physio-

logical Detection of Anxiety-Related Arousal in Children With Autism Spectrum Disorder,” IEEE

TRANSACTIONS ON BIOMEDICAL ENGINEERING, vol. 62, no. 3, pp. 990–1000, 2015.

[23] R. Llinas, “The intrinsic electrophysiological properties of mammalian neurons: insights into central

nervous system function,” Science, vol. 242, no. 4886, pp. 1654–1664, 1988.

[24] G. Burnstock, “Review lecture. Neurotransmitters and trophic factors in the autonomic nervous

system.,” The Journal of Physiology, vol. 313, pp. 1–35, 4 1981.

[25] R. W. Shields, “Functional anatomy of the autonomic nervous system.,” 1993.

[26] G. Glick, E. Braunwald, and R. Lewis, “Relative roles of the sympathetic and parasympathetic

nervous systems in the reflex control of heart rate,” Circulation research, vol. 16, no. 4, pp. 363–

375, 1965.

BIBLIOGRAPHY 52

[27] U. M. Nater and N. Rohleder, “Salivary alpha-amylase as a non-invasive biomarker for the sym-

pathetic nervous system: Current state of research,” Psychoneuroendocrinology, vol. 34, no. 4,

pp. 486–496, 2009.

[28] V. Rauschel, A. Straube, F. Suß, and R. Ruscheweyh, “Responsiveness of the autonomic nervous

system during paced breathing and mental stress in migraine patients,” Journal of Headache and

Pain, vol. 16, no. 1, 2015.

[29] L. K. Mccorry, “TEACHERS TOPICS Physiology of the Autonomic Nervous System,” vol. 71,

no. 4, 2007.

[30] S. W. Porges, “Emotion: An evolutionary by-product of the neural regulation of the autonomic

nervous system,” Annals of the New York Academy of Sciences, vol. 807, pp. 62–77, 1997.

[31] R. W. Levenson, “The Autonomic Nervous System and emotion,” Emotion Review, vol. 6, no. 2,

pp. 100–112, 2014.

[32] D. Hagemann, S. R. Waldstein, and J. F. Thayer, “Central and autonomic nervous system integra-

tion in emotion,” Brain and Cognition, vol. 52, no. 1, pp. 79–87, 2003.

[33] S. D. Kreibig, “Autonomic nervous system activity in emotion: A review,” Biological Psychology,

vol. 84, no. 3, pp. 394–421, 2010.

[34] B. Folkow, “Perspectives on the integrative functions of the ’sympatho- adrenomedullary system’,”

2000.

[35] A. Farley, E. McLafferty, and C. Hendry, “The cardiovascular system.,” Nursing standard (Royal

College of Nursing (Great Britain) : 1987), vol. 27, no. 9, pp. 35–9, 2012.

[36] V. Mahadevan, “Anatomy of the heart,” Surgery (United Kingdom), vol. 33, no. 2, pp. 47–51, 2015.

[37] E. W. Van Stee, Autonomic Innervation of the Heart, vol. VOL. 26. Berlin, Heidelberg: Springer

Berlin Heidelberg, 2015.

[38] G. G. BERNTSON, J. THOMAS BIGGER, D. L. ECKBERG, P. GROSSMAN, P. G. KAUF-

MANN, M. MALIK, H. N. NAGARAJA, S. W. PORGES, J. P. SAUL, P. H. STONE, and M. W.

VAN DER MOLEN, “Heart rate variability: Origins, methods, and interpretive caveats,” Psy-

chophysiology, vol. 34, pp. 623–648, 11 1997.

[39] T. Vybiral, R. J. Bryg, M. E. Maddens, and W. E. Boden, “Effect of passive tilt on sympathetic and

parasympathetic components of heart rate variability in normal subjects,” The American Journal

of Cardiology, vol. 63, no. 15, pp. 1117–1120, 1989.

[40] B. SAYKRS, “Analysis of Heart Rate Variability,” Ergonomics, vol. 16, no. 1, pp. 17–32, 1973.

[41] A. J. Fisher and M. G. Newman, “Heart rate and autonomic response to stress after experimen-

tal induction of worry versus relaxation in healthy, high-worry, and generalized anxiety disorder

individuals,” Biological Psychology, vol. 93, pp. 65–74, 4 2013.

[42] Y. M. Ulrich-Lai and J. P. Herman, “Neural regulation of endocrine and autonomic stress responses,”

Nature Reviews Neuroscience, vol. 10, pp. 397–409, 6 2009.

BIBLIOGRAPHY 53

[43] B. Oberg, “Overall Cardiovascular Regulation,” Annual Review of Physiology, vol. 38, pp. 537–570,

3 1976.

[44] B. Pomeranz, R. J. B. Macaulay, M. A. Caudill, I. Kutz, D. Adam, D. Gordon, K. M. Kilborn,

A. C. Barger, D. C. Shanon, R. J. Cohen, and H. Benson, “Assessment of autonomic function by

heart rate spectral analysis in humans,” American Journal of Physiology, vol. 248, pp. H151–H153,

1985.

[45] B. Chen and H. Gao, “Motor force homeostasis in skeletal muscle contraction,” Biophysical Journal,

vol. 101, no. 2, pp. 396–403, 2011.

[46] M. a. Geeves and K. C. Holmes, “Structural Mechanism of Muscle Contraction,” Annual Review of

Biochemistry, vol. 68, pp. 687–728, 6 1999.

[47] A. Y. M. Jones, C. Kam, K. W. Lai, H. Y. Lee, H. T. Chow, S. F. Lau, L. M. Wong, and J. He,

“Changes in heart rate and R-wave amplitude with posture,” Chinese Journal of Physiology, vol. 46,

no. 2, pp. 63–69, 2003.

[48] P. J. Lang, “Emotion’s response patterns: The brain and the autonomic nervous system,” Emotion

Review, vol. 6, no. 2, pp. 93–99, 2014.

[49] E. Jacobsen, H. Kehlet, V. Larsen, I. Munkvad, and K. Skinhøj, “INVESTIGATIONS INTO AUTO-

NOMIC RESPONSES DURING EMOTION,” Acta Psychiatrica Scandinavica, vol. 30, pp. 607–625,

12 1955.

[50] C. Collet, E. Vernet-Maury, G. Delhomme, and A. Dittmar, “Autonomic nervous system response

patterns specificity to basic emotions,” Journal of the Autonomic Nervous System, vol. 62, no. 1-2,

pp. 45–57, 1997.

[51] R. W. Levenson, “Autonomic Nervous Systems Differences Among Emotions,” Psychological Science

(Wiley-Blackwell), vol. 3, no. 1, pp. 23–27, 1992.

[52] C. L. Stephens, I. C. Christie, and B. H. Friedman, “Autonomic specificity of basic emotions:

Evidence from pattern classification and cluster analysis,” Biological Psychology, vol. 84, no. 3,

pp. 463–473, 2010.

[53] A. L. Wit, M. B. Weiss, W. D. Berkowitz, K. M. Rosen, C. Steiner, and A. N. Damato, “Patterns

of atrioventricular conduction in the human heart.,” Circulation research, vol. 27, pp. 345–359, 9

1970.

[54] J. G. Webster, Medical Instrumentation: Application and Design. Wiley, 4 ed., 2009.

[55] S. Maheshwari, A. Acharyya, P. Rajalakshmi, P. E. Puddu, and M. Schiariti, “Accurate and reliable

3-lead to 12-lead ECG reconstruction methodology for remote health monitoring applications,”

Irbm, vol. 35, no. 6, pp. 341–350, 2014.

[56] J. Francis, “ECG monitoring leads and special leads,” Indian Pacing and Electrophysiology Journal,

vol. 16, no. 3, pp. 92–95, 2016.

[57] E. M. Spinelli, N. H. Martmez, and M. A. Mayosky, “A transconductance driven-right-leg circuit,”

IEEE Transactions on Biomedical Engineering, vol. 46, no. 12, pp. 1466–1470, 1999.

BIBLIOGRAPHY 54

[58] D. K. Freeman, R. D. Gatzke, G. Mallas, Y. Chen, and C. J. Brouse, “Saturation of the right-

leg drive amplifier in low-voltage ECG monitors,” IEEE Transactions on Biomedical Engineering,

vol. 62, no. 1, pp. 323–330, 2015.

[59] C. Assambo, A. Baba, R. Dozio, and M. J. Burke, “Determination of the parameters of the skin-

electrode impedance model for ECG measurement,” Proceedings of the 6th WSEAS International

Conference on Electronics Hardware Wireless and Optical Communications, p. 9095, 2007.

[60] H. W. Tam and J. G. Webster, “Minimizing Electrode Motion Artifact by Skin Abrasion,” IEEE

Transactions on Biomedical Engineering, vol. BME-24, no. 2, pp. 134–139, 1977.

[61] D. Buxi, S. Kim, N. Van Helleputte, M. Altini, J. Wijsman, R. F. Yazicioglu, J. Penders, and

C. Van Hoof, “Correlation between electrode-tissue impedance and motion artifact in biopotential

recordings,” IEEE Sensors Journal, 2012.

[62] Y. C. Yeh and W. J. Wang, “QRS complexes detection for ECG signal: The Difference Operation

Method,” Computer Methods and Programs in Biomedicine, vol. 91, no. 3, pp. 245–254, 2008.

[63] Cuiwei Li, Chongxun Zheng, and Changfeng Tai, “Detection of ECG characteristic points using

wavelet transforms,” IEEE Transactions on Biomedical Engineering, vol. 42, no. 1, pp. 21–28, 1995.

[64] N. Cho, J. Yoo, S. J. Song, J. Lee, S. Jeon, and H. J. Yoo, “The human body characteristics as

a signal transmission medium for intrabody communication,” IEEE Transactions on Microwave

Theory and Techniques, vol. 55, no. 5, pp. 1080–1085, 2007.

[65] J. G. Proakis and D. K. Manolakis, Digital Signal Processing. 4th editio ed., 2006.

[66] B. Dumitrescu, Positive Trigonometric Polynomials and Signal Processing Applications. No. 4 in

Signals and Communication Technology, Cham: Springer International Publishing, 2017.

[67] M. Haag, “Understanding Pole / Zero Plots on,” Design, pp. 1–5, 2005.

[68] K. J. Blinowska and J. Zygierewicz, Practical Biomedical Signal Analysis Using MATLAB{ R©}.CRC Press, 2011.

[69] S.-H. Liu, “Motion Artifact Reduction in Electrocardiogram Using Adaptive Filter,” Journal of

Medical and Biological Engineering, vol. 31, no. 1, pp. 67–72, 2011.

[70] D. P. Zipes, “Clinical application of the electrocardiogram,” Journal of the American College of

Cardiology, vol. 36, no. 6, pp. 1746–1748, 2000.

[71] D. E. Becker, “Fundamentals of electrocardiography interpretation.,” Anesthesia progress, vol. 53,

no. 2, pp. 53–63, 2006.

[72] J. P. Lynch, A. Partridge, K. H. Law, T. W. Kenny, A. S. Kiremidjian, and E. Carryer, “Design of

Piezoresistive MEMS-Based Accelerometer for Integration with Wireless Sensing Unit for Structural

Monitoring,” Journal of Aerospace Engineering, vol. 16, pp. 108–114, 7 2003.

[73] B. R. O. Reilly, A. Khenkin, and K. Harney, “Sonic Nirvana : Using MEMS Accelerometers as

Acoustic Pickups in Musical Instruments,” Analog Dialogue, vol. 1, no. Figure 2, pp. 1–4, 2009.

BIBLIOGRAPHY 55

[74] O. M. Mozos, V. Sandulescu, S. Andrews, D. Ellis, N. Bellotto, R. Dobrescu, and J. M. Ferrandez,

“Stress Detection Using Wearable Physiological and Sociometric Sensors,” International Journal of

Neural Systems, vol. 1954, no. 09999, pp. 1–17, 2016.

[75] L. Changchun, P. Rani, and N. Sarkar, “An empirical study of machine learning techniques for

affect recognition in human-robot interaction,” in 2005 IEEE/RSJ International Conference on

Intelligent Robots and Systems, no. September, pp. 2662–2667, IEEE, 2005.

[76] F. Sun, C. Kuo, H. Cheng, S. Buthpitiya, P. Collins, and M. Griss, “Activity-aware Mental Stress

Detection Using Physiological Sensors,” Mobile Computing, Applications, and Services, vol. 76,

pp. 1–20, 2012.

[77] O. M. Mozos, V. Sandulescu, S. Andrews, D. Ellis, N. Bellotto, R. Dobrescu, and J. M. Ferrandez,

“Stress Detection Using Wearable Physiological and Sociometric Sensors,” International Journal of

Neural Systems, vol. 0, no. 09999, pp. 1–17, 2016.

[78] A. Kushki, E. Drumm, M. Pla Mobarak, N. Tanel, A. Dupuis, T. Chau, and E. Anagnostou, “Inves-

tigating the Autonomic Nervous System Response to Anxiety in Children with Autism Spectrum

Disorders,” PLoS ONE, vol. 8, no. 4, 2013.

[79] M. S. Grewal and A. P. Andrews, Kalman Filtering. New York, USA: John Wiley & Sons, Inc., 1

2002.

[80] G. Welch and G. Bishop, “An Introduction to the Kalman Filter,” In Practice, vol. 7, no. 1, pp. 1–16,

2006.

[81] T. M. Marteau and H. Bekker, “The development of a six-item short-form of the state scale of the

Spielberger State-Trait Anxiety Inventory (STAI),” British Journal of Clinical Psychology, vol. 31,

pp. 301–306, 9 1992.

[82] W. K. Silverman, W. Fleisig, B. Rabian, and R. A. Peterson, “Childhood Anxiety Sensitivity Index,”

Journal of Clinical Child Psychology, vol. 20, pp. 162–168, 6 1991.

[83] R. Hoehn-Saric and D. R. McLeod, “Anxiety and arousal: Physiological changes and their percep-

tion,” Journal of Affective Disorders, vol. 61, no. 3, pp. 217–224, 2000.

[84] B. Grundlehner, L. Brown, J. Penders, and B. Gyselinckx, “The Design and Analysis of a Real-Time,

Continuous Arousal Monitor,” 2009 Sixth International Workshop on Wearable and Implantable

Body Sensor Networks, 2009.

[85] A. Kushki, E. Drumm, M. Pla Mobarak, N. Tanel, A. Dupuis, T. Chau, and E. Anagnostou, “Inves-

tigating the Autonomic Nervous System Response to Anxiety in Children with Autism Spectrum

Disorders,” PLoS ONE, vol. 8, no. 4, pp. 2–9, 2013.

[86] G. A. Reyes del Paso, W. Langewitz, L. J. M. Mulder, A. van Roon, and S. Duschek, “The utility

of low frequency heart rate variability as an index of sympathetic cardiac tone: A review with

emphasis on a reanalysis of previous studies,” Psychophysiology, vol. 50, pp. 477–487, 5 2013.

[87] A. Sano and R. W. Picard, “Stress Recognition using Wearable Sensors and Mobile Phones,” Hu-

maine Association Conference on Affective Computing and Intelligent Interaction Stress, pp. 671–

676, 2013.

BIBLIOGRAPHY 56

[88] T. Westeyn, K. Vadas, X. Bian, T. Starner, and G. D. Abowd, “Recognizing mimicked autis-

tic self-stimulatory behaviors using HMMs,” Proceedings - International Symposium on Wearable

Computers, ISWC, vol. 2005, pp. 164–167, 2005.

Appendix

Appendix-A : STAI Questionnaire

57

BIBLIOGRAPHY 58

Appendix-B: Participant average heart rate

Figure 7.1: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-1 while standing, slow walking, and fast walking

BIBLIOGRAPHY 59

Figure 7.2: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-2 while standing, slow walking, and fast walking

BIBLIOGRAPHY 60

Figure 7.3: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-3 while standing, slow walking, and fast walking

BIBLIOGRAPHY 61

Figure 7.4: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-4 while standing, slow walking, and fast walking

BIBLIOGRAPHY 62

Figure 7.5: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-5 while standing, slow walking, and fast walking

BIBLIOGRAPHY 63

Figure 7.6: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-6 while standing, slow walking, and fast walking

BIBLIOGRAPHY 64

Figure 7.7: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-7 while standing, slow walking, and fast walking

BIBLIOGRAPHY 65

Figure 7.8: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-8 while standing, slow walking, and fast walking

BIBLIOGRAPHY 66

Figure 7.9: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-9 while standing, slow walking, and fast walking

BIBLIOGRAPHY 67

Figure 7.10: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-10 while standing, slow walking, and fast walking

BIBLIOGRAPHY 68

Figure 7.11: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-11 while standing, slow walking, and fast walking

BIBLIOGRAPHY 69

Figure 7.12: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-12 while standing, slow walking, and fast walking

BIBLIOGRAPHY 70

Figure 7.13: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-13 while standing, slow walking, and fast walking

BIBLIOGRAPHY 71

Figure 7.14: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-14 while standing, slow walking, and fast walking

BIBLIOGRAPHY 72

Figure 7.15: Mean heart rate during the sitting baseline, baseline, and stroop test phases; as recorded forparticipant-15 while standing, slow walking, and fast walking

BIBLIOGRAPHY 73

Appendix-C: Anxiety and motion detection participant results

Figure 7.16: Arousal detection for participant-1

BIBLIOGRAPHY 74

Figure 7.17: Arousal detection for participant-2

Figure 7.18: Arousal detection for participant-3

BIBLIOGRAPHY 75

Figure 7.19: Arousal detection for participant-4

Figure 7.20: Arousal detection for participant-5

BIBLIOGRAPHY 76

Figure 7.21: Arousal detection for participant-6

Figure 7.22: Arousal detection for participant-7

BIBLIOGRAPHY 77

Figure 7.23: Arousal detection for participant-8

Figure 7.24: Arousal detection for participant-9

BIBLIOGRAPHY 78

Figure 7.25: Arousal detection for participant-10

Figure 7.26: Arousal detection for participant-11

BIBLIOGRAPHY 79

Figure 7.27: Arousal detection for participant-12

Figure 7.28: Arousal detection for participant-13

BIBLIOGRAPHY 80

Figure 7.29: Arousal detection for participant-14

Figure 7.30: Arousal detection for participant-15

BIBLIOGRAPHY 81

Appendix-D: Visual Reference for Electrode Placement

Electrodes Visual Reference Sheet – Version2, 01/06/17 Page 1 of 1

LOCATION OF ECG STICKERS

ECG Stickers are shown with a blue circle:

Where to Place Stickers:

• 1 on Left side of chest (below collar bone)

• 1 on Right side of chest (below collar bone)

• 1 on lower right rib cage (just above belly button to the right)

• 1 on lower left rib cage (just above belly button to the left)

top related