Annales F 34 Sarhaddi DISS

CONTINUOUS IOT-BASED
MATERNAL MONITORING:
SYSTEM DESIGN, EVALUATION,
OPPORTUNITIES, AND
CHALLENGES
Fatemeh Sarhaddi
TURUN YLIOPISTON JULKAISUJA – ANNALES UNIVERSITATIS TURKUENSIS

SARJA – SER. F OSA – TOM. 34 | TECHNICA – INFORMATICA | TURKU 2023
CONTINUOUS IOT-BASED
MATERNAL MONITORING:
SYSTEM DESIGN, EVALUATION,
OPPORTUNITIES, AND
CHALLENGES
Fatemeh Sarhaddi
TURUN YLIOPISTON JULKAISUJA – ANNALES UNIVERSITATIS TURKUENSIS

SARJA – SER. F OSA – TOM. 34 | TECHNICA – INFORMATICA | TURKU 2023
University of Turku
Faculty of Technology
Department of Computing
Information and communication Technology
Doctoral programme in Technology (DPT)
Supervised by
Professor, Pasi Liljeberg Professor, Amir M. Rahmani
University of Turku University of California, Irvine
Associate Professor, Anna Axelin Adjunct Professor, Iman Azimi

University of Turku University of Turku
Reviewed by
Professor, Raquel Bailón Associate Professor, Teemu Myllylä
University of Zaragoza University of Oulu
Opponent
Associate Professor, Frida Sandberg
Lund University
The originality of this publication has been checked in accordance with the University
of Turku quality assurance system using the Turnitin OriginalityCheck service.
ISBN 978-951-29-9581-3 (PRINT)

ISBN 978-951-29-9582-0 (PDF)
ISSN 2736-9390 (Painettu/Print)
ISSN 2736-9684 (Sähköinen/Online)
Painosalama, Turku, Finland, 2023
UNIVERSITY OF TURKU
Faculty of Technology
Department of Computing
SARHADDI, FATEMEH: Continuous IoT-based maternal monitoring: system
design, evaluation, opportunities, and challenges
Doctoral dissertation, 186 pp.
Doctoral programme in Technology (DPT)
December 2023
ABSTRACT
Maternal care encompasses health care services for pregnant women during preg-
nancy, childbirth, and the postpartum period. Maternity care providers aim to ensure
a healthy pregnancy, safe delivery, and smooth transition to motherhood. Traditional
maternal care is offered through regular check-ups by health care professionals.
In recent years, the emergence of Internet-of-Things (IoT)-based systems has trans-
formed the way health care services are provided. These systems offer low-cost ubiq-
uitous monitoring in everyday life settings and can be used for maternal monitoring.
However, IoT-based maternal monitoring systems lack a comprehensive approach in
maternal care because they are limited by sensing capabilities, specific health prob-
lems, and short periods of monitoring. Moreover, the use of IoT-based systems for
maternal health monitoring requires addressing critical quality attributes, such as fea-
sibility, energy efficiency, and reliability and validity of the collected physiological
parameters. Quality assessment methods also must be integrated with such systems
to discard the noisy part of collected parameters and improve the data quality. Fur-
thermore, long-term, continuous IoT-based maternal monitoring by collecting data
that was not traditionally available provides new opportunities, including analyzing
the trend of physiological parameters during pregnancy and postpartum, as well as
detecting maternal health issues.
This thesis presents an IoT-based maternal monitoring system and explores its po-
tential in maternal care. We evaluate the system’s feasibility, reliability, and energy
efficiency. We also discuss the practical challenges of implementing the system.
Then, we validate the heart rate (HR) and heart rate variability (HRV) parameters
that the system collects while the user is asleep and awake. In addition, we propose
a deep-learning-based method for quality assessment of HR and HRV parameters to
discard unreliable data and improve health decisions. We use the system to collect
data from 62 pregnant women during pregnancy and three-months postpartum. Then,
the reliable HR and HRV parameters are used to track the trends during pregnancy
and postpartum.
Finally, we investigate maternal loneliness as a major mental health problem. We
develop two predictive models to detect maternal loneliness during late pregnancy
and the postpartum period. The models use the objective health parameters passively
collected by the system and achieve high performance (weighted F1 scores > 0.87).
KEYWORDS: maternal health, heart rate variability, remote health monitoring,

loneliness, photoplethysmography, quality assessment, Internet of things
iii
TURUN YLIOPISTO
Teknillinen tiedekunta
Tietotekniikan laitos
Tieto- ja viestintätekniikka
SARHADDI, FATEMEH: Continuous IoT-based maternal monitoring: system
design, evaluation, opportunities, and challenges
Väitöskirja, 186 s.
Teknologian tohtoriohjelma (DPT)
Joulukuu 2023
TIIVISTELMÄ
Äitiysterveydenhuollon tavoitteena on varmistaa terve raskaus, turvallinen synnytys

ja sujuva siirtyminen äitiyteen. Perinteistä äitiysneuvontaa tarjotaan säännöllisten
tarkastuskäyntien avulla terveydenhuollon ammattilaisten toimesta.
Viime vuosina Internet of Things (IoT) -pohjaisten järjestelmien kehittyminen on
mahdollistanut terveydenhuoltopalvelujen tarjoamisen entistä monipuolisemmin.
Nämä järjestelmät mahdollistavat edullisen ja laaja-alaisen terveydenseurannan ajasta
ja paikasta riippumatta. Kuitenkin tällä hetkellä IoT-pohjaiset järjestelmät ovat hyvin
vähäisessä määrin käytössä äitiysterveydenhuollossa.
Tämä väitöskirja esittelee IoT-pohjaisen äitiyshuollon seurantajärjestelmän ja tutkii
sen tuomia mahdollisuuksia ennenkaikkea äitiysterveyshuollossa. Järjestelmän to-
teutettavuutta, luotettavuutta ja energiatehokkuutta analysoidaan sekä käsittellään
myös järjestelmän käyttöönoton käytännön haasteita. Järjestelmä kerää käyttäjän
sydämen sykkeen (HR) ja sydämen sykkeen vaihtelun (HRV), jotka järjestelmä kerää
käyttäjän ollessa unessa ja hereillä. Lisäksi esitellään syväoppimiseen perustuvaa
menetelmää kerätyn sydämen sykkeen ja sydämen sykkeen vaihtelun laadun arvioin-
tiin epäluotettavan ja mahdollisesti virheellisen datan poistamiseksi ja sitä kautta
järjestelmän luotettavuuden ja käytettävyyden parantamiseksi. Tässä tutkimuksessa
hyödynnettiin dataa joka on kerätty tutkimuksen aikana 62 raskaana olevalta naiselta
raskauden aikana sekä kolme kuukautta synnytyksen jälkeen.
Tässä väitöskirjassa tutkittaan myös äitiyden yksinäisyyttä merkittävänä mielenter-
veysongelmana. Työssä kehitettiin kaksi ennustemallia äitiyden yksinäisyyden havait-
semiseksi myöhäisraskauden ja synnytyksen jälkeisen ajanjakson aikana. Mallit
käyttävät järjestelmän passiivisesti keräämiä objektiivisia terveysparametreja ja saavut-
tavat korkean suorituskyvyn (painotetut F1-pisteet > 0,87).
AVAINSANAT: äitiysterveydenhuolto, sykevälivaihtelu, etäterveyden seuranta,

yksinäisyys, laadunarviointi, esineiden internet
iv
Acknowledgements
I want to express my gratitude to my supervisors, Prof. Pasi Liljeberg, Prof. Amir.

M. Rahmani, Associate Prof. Anna Axelin, and Adjunct Prof. Iman Azimi for guid-
ing my throughout this journey and offering the chance to conduct research in a
multidisciplinary group. Your advises, support, encouragement, and patience were
crucial in completing this thesis.
Additionally, I want to acknowledge the members of Digital health technology
lap and my coauthors especially Dr. Hannakaisa Niela-Vilén, Kianoosh Kazemi, Mo-
hammad Feli, Yuning Wang, Dr. Jennifer Auxier, Dr. Sina Labbaf, Dr. Milad Asgari
Mehrabadi , and Dr. Emad Kasaeyan Naeini for their collaboration and shared en-
thusiasm. Our discussions and joint efforts have contributed immensely to the depth
and breadth of this research. I also want to extend my appreciation to Elisa Lankinen,
Henrika Merenlehto, and Ani Heinonen for their contribution in data collection.
This research was supported by the Academy of Finland (AKA) and University
of Turku Graduate School fellowship. I acknowledge these organizations for their
financial support, which made it possible for me to dedicate time and resources to
this research.
I am also grateful to have Professor Raquel Bailón and Associate Professor
Teemu Myllylä for their valuable feedback on the manuscript. My appreciation also
extend to Associate Professor Frida Sandberg for accepting to act as the opponent
for public examination.
I want to extend my gratitude to all my friends supporting me during this jour-
ney, Mahdieh, Bahare, Maryam, Elham, Farimah, Saeide, Sepide, Mina, Talaye and
others. I am also deeply grateful to my parents, and my brothers Reza and Ali, for
their support, understanding and encouragement. Lastly, I want to express my love
and gratitude to Mohsen and my son Hossein, who have been a constant source of
support for me. Their unwavering belief in my abilities has been a driving force
behind the completion of this thesis.
Date December 11, 2023

Fatemeh Sarhaddi
v
Table of Contents
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Original Publications . . . . . . . . . . . . . . . . . . . . . . xi
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 IoT-Based Maternal Monitoring Opportunities and Challenges 2
1.2 Aims and Objectives . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . 5
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Internet of Things-based Health-Monitoring Systems . . . . 7
2.2 Cardiac Activity Measurement . . . . . . . . . . . . . . . . . 8
2.2.1 Electrocardiogram . . . . . . . . . . . . . . . . . . . 9
2.2.2 Photoplethysmogram . . . . . . . . . . . . . . . . . . 10
2.3 Heart Rate Variability . . . . . . . . . . . . . . . . . . . . . . 11
3 IoT-Based Maternal Health Monitoring . . . . . . . . . . . . . 13

3.1 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Maternal Health Monitoring Services . . . . . . . . . . . . . 14
3.3 Presented Long-term IoT-Based Maternal Health Monitoring
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 Perception Layer . . . . . . . . . . . . . . . . . . . . 15
3.3.2 Gateway Layer . . . . . . . . . . . . . . . . . . . . . 17
3.3.3 Cloud Layer . . . . . . . . . . . . . . . . . . . . . . . 17
3.3.4 Application Layer . . . . . . . . . . . . . . . . . . . . 18
3.4 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 System Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5.1 Feasibility . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5.2 Robustness and Reliability of Measurements . . . . 22
3.5.3 Energy Consumption . . . . . . . . . . . . . . . . . . 23
vi
3.5.4 Practical Challenges . . . . . . . . . . . . . . . . . . 24
4 Validation and Quality Assessment of HR and HRV Param-

eters for Health Monitoring . . . . . . . . . . . . . . . . . . . . . 25
4.1 Necessity of Accuracy and Quality Assessment of HR and
HRV Parameters Collected by Wearable Devices . . . . . . 25
4.2 Accuracy Assessment of the Smartwatch HR and HRV Pa-
rameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 Related Works . . . . . . . . . . . . . . . . . . . . . . 27
4.2.2 Participants and Recruitment . . . . . . . . . . . . . 28
4.2.3 Data Collection . . . . . . . . . . . . . . . . . . . . . 28
4.2.4 Data Analysis . . . . . . . . . . . . . . . . . . . . . . 29
4.2.5 Results and Discussion . . . . . . . . . . . . . . . . 31
4.3 A CNN-Based PPG Quality Assessment Approach for HR
and HRV Parameters . . . . . . . . . . . . . . . . . . . . . . 37
4.3.1 Related Works . . . . . . . . . . . . . . . . . . . . . . 37
4.3.2 Data Collection and Data Analysis . . . . . . . . . . 39
4.3.3 Proposed PPG Quality Assessment Approaches . . 40
4.3.4 Model Implementation and Evaluation . . . . . . . . 41
4.3.5 Comparison with the State-of-the-Art Models . . . . 42
5 IoT-Based Maternal Monitoring During Pregnancy and Post-

partum: Analyzing the Trends of HR and HRV Parameters 43
5.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.1 Participants and Setup . . . . . . . . . . . . . . . . . 44
5.2 HR and HRV Trends During Pregnancy and Postpartum . . 44
5.2.1 HR and HRV Parameters Analysis . . . . . . . . . . 46
5.2.2 Results and Discussion . . . . . . . . . . . . . . . . 48
6 Maternal Loneliness Detection Based on Passive Data Sens-

ing Using Our IoT-Based Maternal Monitoring System . . . 53
6.1 Maternal Loneliness . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Setup and dataset Preparation . . . . . . . . . . . . . . . . . 54
6.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . 54
6.2.2 Data Collection . . . . . . . . . . . . . . . . . . . . . 55
6.2.3 Data Analysis . . . . . . . . . . . . . . . . . . . . . . 55
6.2.4 Feature Extraction . . . . . . . . . . . . . . . . . . . 56
6.2.5 Labeling . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2.6 datasets . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2.7 Missing Data Imputation . . . . . . . . . . . . . . . . 58
6.3 Predictive Machine-Learning Models . . . . . . . . . . . . . 58
vii
6.3.1 Decision Tree . . . . . . . . . . . . . . . . . . . . . . 58
6.3.2 Gradient Boosting . . . . . . . . . . . . . . . . . . . . 59
6.4 Model Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . 59
6.5.1 Performance of Maternal Loneliness Prediction . . . 59
6.5.2 Important Features in Maternal Loneliness . . . . . 62
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8 Overview of Original Publications . . . . . . . . . . . . . . . . 66

8.1 Long-Term IoT-Based Maternal Monitoring: System Design
and Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 66
8.2 A comprehensive accuracy assessment of Samsung smart-
watch heart rate and heart rate variability . . . . . . . . . . . 66
8.3 A Deep Learning-based PPG Quality Assessment Approach
for Heart Rate and Heart Rate Variability . . . . . . . . . . . 67
8.4 Trends in Heart Rate and Heart Rate Variability During Preg-
nancy and the 3-Month Postpartum Period: Continuous Mon-
itoring in a Free-living Context . . . . . . . . . . . . . . . . . 67
8.5 Maternal Social Loneliness Detection Using Passive Sens-
ing Through Continuous Monitoring in Everyday Settings:
Longitudinal Study . . . . . . . . . . . . . . . . . . . . . . . . 68
Appendix A: questions provided by the mobile application . 70
List of References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Original Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
viii
Abbreviations
AI Artificial intelligence
ANS Autonomic nervous system
API Application Programming Interface
AUC Area Under the Curve
AVNN Average of normal to normal interbeat intervals
BMI Body mass index
CI Confidence Interval
CNN Convolutional Neural Networks
ECG Electrocardiography
GAF Gramian Angular Field
GCG Gyrocardiogram
HF Absolute power in high frequency band
HR Heart rate
HRV Heart rate variability
HLM Hierarchical linear model
IBI Interbeat interval
IMU Inertial measurement unit
IoT Internet of Things
KNN K-nearest neighbors
LF Absolute power in low frequency band
LF/HF Absolute power in low frequency band to absolute power in high fre-
quency band ratio
LSTM Long Short Term Memory
NNI Normal interbeat interval
PCG Phonocardiogram
PNN50 Percentage of successive NN intervals that differ by more than 50 ms
PPG Photoplethysmography
PRV Pulse rate variability
RFE Recursive feature elimination
RMSSD Root mean square of successive NN interval differences
SCG Seismocardiogram
SD Standard deviation
SDNN Standard deviation of NN intervals
ix
SSL Secure sockets layer
SVM Support vector machine
TST Total sleep time
UCLA University of California, Los Angeles
WASO Wake after sleep onset
x
List of Original Publications
This dissertation is based on the following original publications, which are referred
to in the text by their Roman numerals:
I Fatemeh Sarhaddi, Iman Azimi, Sina Labbaf, Hannakaisa Niela-Vilén, Nikil

Dutt, Anna Axelin, Pasi Liljeberg, and Amir M. Rahmani. ”Long-term
IoT-based Maternal monitoring: system design and evaluation.” Sensors
21, no. 7 (2021): 2281.
II Fatemeh Sarhaddi* , Kianoosh Kazemi* , Iman Azimi, Rui Cao, Hannakaisa

Niela-Vilén, Anna Axelin, Pasi Liljeberg, and Amir M. Rahmani. ”A com-
prehensive accuracy assessment of Samsung smartwatch heart rate and
heart rate variability.” PloS one 17, no. 12 (2022): e0268361.
III Emad Kasaeyan Naeini, Fatemeh Sarhaddi, Iman Azimi, Pasi Liljeberg,
Nikil Dutt, and Amir M. Rahmani. ”A Deep Learning-based PPG Quality
Assessment Approach for Heart Rate and Heart Rate Variability”. ACM
Transactions on Computing for Healthcare 4, no. 4 (2023): 1-22.
IV Fatemeh Sarhaddi, Iman Azimi, Anna Axelin, Hannakaisa Niela-Vilen,

Pasi Liljeberg, and Amir M. Rahmani. ”Trends in Heart Rate and Heart
Rate Variability During Pregnancy and the 3-Month Postpartum Period:
Continuous Monitoring in a Free-living Context.” JMIR mHealth and uHealth
10, no. 6 (2022): e33458.
V Fatemeh Sarhaddi, Iman Azimi, Anna Axelin, Hannakaisa Niela-Vilen,

Pasi Liljeberg, and Amir M. Rahmani. ”Maternal Social Loneliness Detec-
tion Using Passive Sensing Through Continuous Monitoring in Everyday
Settings: Longitudinal Study”. JMIR Formative Research 7, no. 1 (2023):
e47950.
The original publications have been reproduced with the permission of the copyright
holders.
*
Equal contribution
xi
The following peer-reviewed publications were published during doctoral studies.
These publications are not included in this thesis but are closely related.
VI Hannakaisa Niela-Vilén, Jennifer Auxier, Eeva Ekholm, Fatemeh Sarhaddi,

Milad Asgari Mehrabadi, Aysan Mahmoudzadeh, Iman Azimi, Pasi Lil-
jeberg, Amir M. Rahmani, and Anna Axelin. ”Pregnant women’s daily
patterns of well-being before and during the COVID-19 pandemic in Fin-
land: Longitudinal monitoring through smartwatch technology.” PloS one
16, no. 2 (2021): e0246494.
VII Milad Asgari Mehrabadi, Iman Azimi, Fatemeh Sarhaddi, Anna Axelin,
Hannakaisa Niela-Vilén, Saana Myllyntausta, Sari Stenholm, Nikil Dutt,
Pasi Liljeberg, and Amir M. Rahmani. ”Sleep tracking of a commer-
cially available smart ring and smartwatch against medical-grade actigra-
phy in everyday settings: instrument validation study.” JMIR mHealth and
uHealth 8, no. 11 (2020): e20465.
VIII Rui Cao, Iman Azimi, Fatemeh Sarhaddi, Hannakaisa Niela-Vilen, Anna
Axelin, Pasi Liljeberg, and Amir M. Rahmani. ”Accuracy Assessment of
Oura Ring Nocturnal Heart Rate and Heart Rate Variability in Comparison
With Electrocardiography in Time and Frequency Domains: Comprehen-
sive Analysis.” Journal of Medical Internet Research 24, no. 1 (2022):
e27487.
IX Hannakaisa Niela-Vilen, Iman Azimi, Kristin Suorsa, Fatemeh Sarhaddi,

Sari Stenholm, Pasi Liljeberg, Amir M. Rahmani, and Anna Axelin. ”Com-
parison of Oura Smart Ring Against ActiGraph Accelerometer for Mea-
surement of Physical Activity and Sedentary Time in a Free-Living Con-
text.” CIN: Computers, Informatics, Nursing (2022).
X Jennifer Auxier, Kaisu T. Savolainen, Miriam Bender, Amir M. Rahmani,

Fatemeh Sarhaddi, Iman Azimi, and Anna M. Axelin. ”Exploring access
as a process of adaptation in a self-monitoring perinatal ehealth system:
Mixed methods study from a sociomaterial perspective.” JMIR Formative
Research 7 (2023): e44385.
XI Hannakaisa Niela-Vilen, Eeva Ekholm, Fatemeh Sarhaddi, Iman Azimi,

Amir M. Rahmani, Pasi Liljeberg, Miko Pasanen, and Anna Axelin. ”Com-
paring prenatal and postpartum stress among women with previous adverse
pregnancy outcomes and normal obstetric histories: A longitudinal cohort
study.” Sexual & Reproductive Healthcare 35 (2023): 100820.
xii
1 Introduction
Maternity care refers to the health care services provided to women during preg-
nancy, childbirth, and the postpartum period. This care plays a crucial role in en-
suring the health and well-being of both the mother and her fetus. Maternity care
providers aim to provide comprehensive and compassionate care to women during
this special time to ensure that pregnant women have a healthy pregnancy, a safe
delivery, and a smooth transition into motherhood. In addition, health complica-
tions during pregnancy, such as hypertension or gestational diabetes, may increase
the mother’s risk of corresponding health issues in the future [1; 2]. Therefore, ma-
ternity care is essential to ensure the best possible pregnancy outcome as well as
to promote health at the population level. Conventionally, maternal care is offered
through regular monitoring and check-ups by health care professionals.
In recent years, digital health and systems based on the Internet of things (IoT)
have transformed the way health care services are provided. IoT-based health moni-
toring services enable cost-effective, continuous health monitoring everywhere [3].
The monitoring systems can collect real-time data from a user and her or his envi-
ronment, transmit the data to remote servers, analyze the data, and provide feedback.
The effectiveness of IoT-based maternal monitoring systems in improving health out-
comes for pregnant women and their fetuses has been shown in several studies [4; 5].
IoT-based maternal-monitoring systems can also provide opportunities for a deeper
understanding of physiological changes during pregnancy by continuously collect-
ing data from pregnant women that are not traditionally available. At the same time,
such systems enable personalized monitoring and intervention based on the history of
collected data by using machine learning and artificial intelligence (AI) techniques.
The current works on IoT-based maternal monitoring are constrained by their
narrow focus on specific health problems, limited sensing capabilities, and short-
term monitoring during pregnancy [6; 7; 8; 9; 10]. Furthermore, the implementation
challenges of long-term IoT-based systems for maternal monitoring have not been
explored in research. Moreover, using IoT-based systems for maternal health moni-
toring necessitates addressing some essential quality attributes, including feasibility
and usability, energy efficiency, and reliability and validity of the collected physi-
ological parameters. These attributes are crucial for enhancing the user experience
and engagement of users with the system as well as for ensuring the accuracy of the
results in health care services [11; 12; 13; 14; 15; 16; 17]. In addition, the effec-
1
Fatemeh Sarhaddi
tiveness of using such systems to gain a deeper understanding of normal trends in

physiological parameters and to predict or prevent health issues during pregnancy
and postpartum should be investigated. However, IoT-based maternal monitoring
systems lack continuous data collection during the pregnancy and postpartum peri-
ods that can be used to analyze trends and detect health issues.
1.1 IoT-Based Maternal Monitoring Opportunities and

Challenges
IoT-based maternal monitoring systems offer many opportunities and present several
challenges and requirements. One opportunity these systems offer is investigating
physiological trends during pregnancy and postpartum. A pregnant woman’s body
undergoes physiological changes to prepare for a fetus’s development and delivery
[18; 19]. These changes affect the mother’s health parameters. For example, sev-
eral studies have shown a decrease in the level of physical activity and an increase
in heart rate (HR) as pregnancy proceeds [9; 20; 21]. Sleep quality also decreases
from the second to the third trimester [22]. Although some physiological changes are
normal during pregnancy, they can also be a sign of disease or disorder. Therefore,
it is important to distinguish normal physiological changes during pregnancy from
abnormal changes that could be signs of disease or health issues [18]. The early de-
tection of disease or health complications during pregnancy provides the opportunity
to perform the proper intervention and help to improve the health and well-being
of the pregnant woman and her child. Although several studies have investigated
the trends of maternal physiological parameters, they are limited to short-term and
episodic data collection [21; 23; 24; 25]. An IoT-based maternal monitoring system
that acquires health data continuously during pregnancy and postpartum can offer
adequate trends of maternal physiological parameters.
Another opportunity of IoT-based maternal monitoring is detecting mental health
issues that usually remain undetected during pregnancy and that affect the mother
and baby’s future health. Pregnancy increases the risk of mental health issues such
as anxiety and depression [26]. In the United Kingdom, mental health issues dur-
ing pregnancy are one of the major causes of maternal death [27]. However, these
issues usually remain undetected because the symptoms are assumed to be due to
pregnancy-related changes [28]. Detecting these issues and providing appropriate
intervention will reduce the risk of further complications during the perinatal period.
The rich continuous data collected by a long-term IoT-based maternal monitoring
system can be used to develop predictive models for major health issues such as
maternal loneliness.
Although IoT-based systems offer many opportunities for maternal health care,
several requirements and challenges must be addressed to use these systems. IoT-
based maternal monitoring systems should be feasible for mothers to use despite
2
Introduction
the physiological changes during pregnancy and postpartum. The system requires
reliability and energy efficiency. Moreover, because the system aims to be used for
a long time, it should provide other quality attributes (e.g., a good user experience)
and motivate users for long-tern use. Considering the lack of long-term, IoT-based
maternal monitoring in the literature, it is necessary to design, develop, deploy, and
evaluate such a system and explore the challenges.
IoT-based monitoring systems use wearable devices to collect data continuously
in everyday life settings. Using wearable devices as a part of an IoT-based system
enables collecting and analyzing comprehensive health data and providing personal-
ized health care over an extended duration [11]. Therefore, wearable devices offer
a wide range of health care services that are not available in traditional health care
[29]. For example, wearable devices can detect health risk indicators such as falls or
worsening disease conditions for older people who live alone [30; 29]. However, the
devices are prone to noise, especially when used in everyday life activities, which can
result in low-quality data and subsequently invalid health parameters [31]. Therefore,
wearable devices should be validated before use in health care applications.
In addition, the accuracy of physiological trends and mental health detection
offered by IoT-based maternal monitoring systems highly depends on the accuracy
of the collected data. Low-quality data may result in false predictions or alarms.
Therefore, it is necessary to use quality assessment methods to discard unreliable
data and only use reliable data in the analysis.
1.2 Aims and Objectives

The thesis investigates IoT-based maternal monitoring to improve the maternal care
provided during pregnancy and postpartum. For this purpose, the first aim of the
thesis is to design and develop a continuous IoT-based maternal monitoring system
offering low-cost continuous monitoring during the perinatal period. Such a sys-
tem could improve current maternal health care by providing essential monitoring
services that traditionally could not be provided by maternity care. Then, health pa-
rameters, including HR and heart rate variability (HRV), collected by the system will
be validated to ensure accurate monitoring and decisions. The aim is also to assess
the quality of collected HR and HRV parameters to improve the quality of decisions
and feedback the system offers. In the next step, the presented system will be used to
collect data from pregnant women continuously during pregnancy and postpartum.
The data are expected to be used to understand the trends in HR and HRV parameters
during pregnancy and postpartum. This understanding enables us to discriminate be-
tween normal and abnormal changes, which can be a sign of health complications.
Finally, the presented system will be further developed to include predictive models
to detect maternal loneliness based on the data the system collects.
To summarize, the main research objectives of the thesis are as follow.
3
Fatemeh Sarhaddi
• Research Objective I: Present, develop, and evaluate a continuous IoT-based

maternal health-monitoring system
• Research Objective II: Validate and assess the quality of HR and HRV pa-
rameters collected by the developed IoT-based monitoring system
• Research Objective III: Deploy the IoT-based system for collecting data from
pregnant women during pregnancy and 3 months postpartum
• Research Objective IV: Investigate the trends of HR and HRV parameters

during pregnancy and the postpartum period
• Research Objective V: Develop predictive models to detect maternal lone-

liness during pregnancy and the postpartum period based on data collected
passively by the monitoring system
1.3 Contributions
In summary, the contributions of this thesis are as follow:
• Develop, implement, and evaluate a continuous maternal monitoring sys-

tem and analyze the implementation challenges: The dissertation presents a
long-term maternal monitoring system to monitor mothers’ objective and sub-
jective health parameters continuously during pregnancy and postpartum. The
system provides several monitoring services (e.g., sleep, stress, and physical
activity monitoring), resulting in a holistic view of the mothers’ health con-
ditions. The system is evaluated considering user engagement, reliability of
collected data, and energy efficiency. Then, the challenges and practical issues
in implementing such systems are investigated.
• Investigating the validity of HR and HRV parameters collected by the

presented system: The dissertation investigates the accuracy of HR and HRV
parameters collected by the presented IoT-based maternal monitoring system
against the corresponding gold standard of electrocardiogram (ECG) signals.
• Proposing and implementing a deep learning-based PPG quality assess-

ment method based on HR and HRV parameters: We introduce a deep
learning-based photoplethysmogram (PPG) quality assessment method to de-
termine the quality of the PPG signals collected by our presented system based
on the desired HR and HRV parameters. The proposed method uses convolu-
tional neural network (CNN) architecture and outperformed the existing PPG
quality assessment methods.
4
Introduction
• Continuous long-term monitoring of physiological health parameters of

pregnant women in everyday settings: We used the developed IoT-based ma-
ternal monitoring system to collect data continuously from 62 pregnant women
during pregnancy and 3 months postpartum.
• Analyzing the trends of HR and HRV parameters during pregnancy and

the postpartum period: The dissertation noninvasively assessed the trends of
HR and HRV parameters during pregnancy and postpartum noninvasively. For
this purpose, the continuous data collected by the developed maternal mon-
itoring system were used. Then, reliable HR and HRV parameters were ex-
tracted. The hierarchical linear mixed models (HLM) were used to investigate
the trends of HR and HRV parameters during the second and third trimesters
as well as the postpartum period.
• Proposing predictive models for maternal loneliness detection using pas-

sive sensing: Two predictive machine-learning models were developed for
maternal loneliness detection during late pregnancy and postpartum. The mod-
els use the objective data collected passively by our presented IoT-based ma-
ternal monitoring system.The results show the promising performance of the
presented models. In addition, important physiological parameters that are
associated with maternal loneliness are investigated.
1.4 Thesis Organization

This thesis is based on five original publications that were published in international
peer-reviewed journals. All five papers resulted from collaborations with other re-
searchers. The thesis is article-based and organized into two main parts. Part 1 con-
sists of eight chapters and provides an overview of the research, and Part 2 contains
original publications.
Part 1 aims to provide a big picture of the research published in the original pub-
lication. Chapter 1 presents the introduction, research objectives, and contributions.
Chapter 2 outlines the background of the thesis. Chapter 3 describes the state-of-
the-art IoT-based maternal health-monitoring systems and our presented long-term
IoT-based maternal monitoring system. The presented system is used in the follow-
ing chapters as well. Chapter 4 shows the validity and accuracy of HR and HRV
parameters collected by the smartwatch used in our IoT-based maternal monitoring
system. This chapter first provides the validation of HR and HRV parameters col-
lected by the smartwatch against an ECG device. Then it introduces a deep learning-
based PPG-based quality assessment method to assess the quality of HR and HRV
parameters extracted from PPG signals. The PPG signals were collected by smart-
watch that used in our presented system in free-living conditions. Chapter 5 presents
the data collection from pregnant women using the presented IoT-based system in
5
Fatemeh Sarhaddi
Chapter 3. Then, we analyze the changes in nighttime HR and HRV parameters dur-
ing pregnancy and postpartum. Chapter 6 explains two machine-learning methods
developed for detecting maternal loneliness during late pregnancy and postpartum
as a major mental health issue. The developed models use passively collected data
using the presented IoT-based system. Research conclusions and future directions
are presented in Chapter 7. Finally, Chapter 8 presents a summarized overview of
the original publications and the author’s contributions to each paper.
The second part of this thesis consists of five original publications. The attached
original publications support the research aspects presented in the first part of the
thesis. Paper I contributes to the content presented in Chapter 3. This paper presents
an IoT-based system for long-term remote maternal health monitoring and covers
Research Objective I. Paper II and Paper III correspond to the contents of Chapter 4.
These papers contribute to Research Objective II. Paper II validates the HR and HRV
parameters acquired by the smartwatch used in our presented system, and Paper III
introduces a deep learning-based PPG quality assessment method in everyday life
settings. Paper IV is in accordance with Chapter 5 and represents the HR and HRV
trends during pregnancy and postpartum and contributes to Research Objectives III
and IV. Finally, in Paper V, two machine-learning methods were developed to predict
loneliness and analyze physiological parameters that are associated with loneliness,
which is related to Chapter 6 and Research Objective V. An overview of the thesis
organization is illustrated in Figure 1.
Research Objective I
Chapter 3
Paper I
Chapter 4 Chapter 5 Chapter 6

Paper II
Paper IV Paper V
Paper III
Research Objective II Research Objectives III & IV Research Objective V
Figure 1. Overview of original publications, chapters, and research objectives of the thesis
6
2 Preliminaries
This chapter briefly explains the essential concepts that have been used in the fol-
lowing chapters. First, IoT-based health-monitoring systems are described. Then,
the chapter outlines the cardiac activity measurement and HRV parameters that are
fundamental concepts in the following chapters.
2.1 Internet of Things-based Health-Monitoring Systems

The Internet of Things (IoT) is a network of physical and virtual objects or “things”
that interact with each other and a remote cloud server [32]. Each thing or object
in an IoT-based system is uniquely identifiable and has self-configuring capabilities,
and they can be integrated as an information network.
Advancements in wearable devices and activity trackers facilitate and increase
the use of IoT-based systems in health and well-being applications. The paradigm
enables ubiquitous monitoring everywhere continuously 24 hours everyday. In this
thesis, we presented an IoT-based maternal monitoring system which could contin-
uously monitor physical activity and sleep parameters. The system also collects 12
minutes of PPG signals every second hour to monitor heart activity due to battery-life
constraints of the wearable device. The PPG signals could provide an approximation
of changes in heart activity during the day.
IoT-based health-monitoring systems conventionally perform four main func-
tionalities: collecting health data using various data sources, transmitting data to
a cloud server, analyzing the data, and presenting the analyzed data [33; 34]. The
architecture of an IoT-based health-monitoring system based on the mentioned func-
tionalities consists of four layers and is represented in Figure 2:
1. Perception layer: The perception layer collects health data from the user
using various types of data sources such as wearable devices, smartphones, portable
devices, and background and demographic information.
2. Gateway layer: The gateway layer transmits the data collected in the percep-
tion layer to the cloud layer. The gateway can be a router or smartphone providing
the transmission functionality or can be a smart gateway [35] and provide more ad-
vanced functionalities such as local data analysis, data compression, and security.
3. Cloud layer: The cloud layer receives the data collected via data sources
in the perception layer using the gateway layer. The cloud layer enables the secure
7
Fatemeh Sarhaddi
Perception Layer
Wearable Device Gateway Layer Cloud Layer
Smartphone Data Modeling Application Layer

Smartphone
Router Data Storage Web Application

Portable Device
Background and Smart

demographic Gateway Data Analysis Mobile Application
Information
Figure 2. IoT-based health-monitoring architecture [34]
storage of the collected health data. Moreover, the cloud layer provides data anal-
ysis functionalities by incorporating AI, machine learning, and deep learning meth-
ods. As a result, the cloud layer provides personalized modeling, trend and anomaly
detection, and early detection of health issues. Moreover, the cloud layer allows
analyzed data to be visualized on the application layer.
4. Application layer: The application layer acts as an interface and facilitates
interaction with the IoT-based health monitoring system. This layer usually con-
sists of web and mobile applications to visualize and monitor the collected data and
communicate between caregivers and users. The mobile application enables users to
visually monitor their health data and gain awareness of their health. In addition, the
web application visualizes and models the users’ data for caregivers.
2.2 Cardiac Activity Measurement

Cardiac activity can be measured by various bio-signals. Bio-signals refer to sig-
nals that can be measured continuously from living organisms such as the human
body. Bio-signals represent the activities and nature of the corresponding physio-
logical process and can be in different forms, such as biochemical, electrical, and
physical [36]. Accurate measurement of cardiac activity is crucial for assessing car-
diovascular health and diagnosing various cardiac conditions. Several bio-signals,
including electrocardiogram (ECG), phonocardiogram (PCG), photoplethysmogram
(PPG), seismocardiogram (SCG), and gyrocardiogram (GCG), are employed to mon-
itor heart activity and extract related vital sign, contributing to the determination of
the health and well-being of individuals [36]. PCG represents a recording of the
sound signals generated by the heart [36]. SCG and GCG utilize accelerometers and
gyroscopes attached to the chest, typically positioned on top of the heart, to monitor
mechanical activity of the heart [37]. In this section, we will provide more detailed
explanations of ECG and PPG signals as they are used in subsequent chapters.
8
Preliminaries
2.2.1 Electrocardiogram
ECG is a conventional method for monitoring cardiovascular activity and related
health parameters, such as HR and HRV. ECG signals depict the electrical activity of
the heart and can be easily recorded by electrodes connected to the skin. This signal
is the most commonly used bio-signal and is used widely in clinical settings. Clinical
standard ECG measurement uses four electrode attached to the chest and limbs [36].
ECG considered as gold standard for collecting heart-related vital sign, This recog-
nition is primarily attributed to the simplicity and readily identifiable wave-forms
present in the ECG signals. Numerous studies have firmly established clinical ECG
as a standard procedure.
Additionally, ECG signals can be used to detect cardiovascular diseases and ab-
normalities because these diseases affect the shape of ECG signals [36]. However,
this method cannot be used for long-term health monitoring due to complicated stan-
dard ECG setup, which needs four ECG electrodes be attached to the limbs or chest
of the user. This may limits the user ability to engage in their daily activity. Fur-
thermore, loose or misplaced electrode connections can compromise the quality of
collected ECG signals.
Figure 3 depicts a 1-minutes cardiac template of ECG signals along with the
first 5 seconds of the signals. The ECG template maps all the cardiac cycles of the
ECG signal to one chart, therefore the abnormalities in the cardiac cycles can be
easily identified. The ECG template in Figure 3 shows that the cardiac cycles in a
normal ECG signal are aligned and similar. Additionally, in the five-minute signal
aslo shown that cardiac cycles have similar waveform. Figure 4 shows the 1-minute
template of an ECG sample of atrial fibrillation, along with the first 5 seconds of
corresponding ECG signals. As shown in the ECG template in Figure 4, the cardiac
cycles are different, and the alteration in the ECG signal are readily apparent. More-
over, the different shapes of cardiac cycles are noticeable in the first five seconds of
the corresponding signal.
Templates of normal ECG signals Normal ECG signal

4 5
3 4
3
2
Amplitude
Amplitude
2
1
1
0
0
1
1
2
0.2 0.1 0.0 0.1 0.2 0.3 0.4 0 1 2 3 4 5
Time (seconds) Time (seconds)
Figure 3. Template of 1-minute normal ECG signal and the first 5 seconds of the signal
9
Fatemeh Sarhaddi
Templates of ECG signals with AF ECG signal for AF

3 4
2 3
Amplitude
Amplitude
1 2
0 1
0
1
1
0.2 0.1 0.0 0.1 0.2 0.3 0.4 0 1 2 3 4 5
Time (seconds) Time (seconds)
Figure 4. Template of 1-minute ECG signal with Atrial fibrillation and the first 5 seconds of the
signal
2.2.2 Photoplethysmogram
Photoplethysmography is an optical method used to monitor heart activity. This
method measures the volumetric variation of blood flow using a light emitter and
light detector. The method is enabled by emitting light to the skin and measuring
the light absorption using a light detector [38; 39]. The PPG sensors can easily be
placed on the finger or wrist and are used widely in various wearable devices for HR
and HRV monitoring. PPG signals can also be used to extract respiration rate and
oxygen saturation.
The PPG method is an easy-to-implement, inexpensive, energy efficient and con-
venient method that is widely used in both clinical and commercial devices [15].
PPG-based wearable devices can be used in remote health monitoring system and in
every day settings. However, PPG method is highly susceptible to motion artifacts
and environmental noise, which are inevitable in everyday life setting. For instance,
the light sensors might be exposed by environment light sources, or motion artifact
affect the signals when users are involved in various daily activities while using the
PPG-based wearable devices. A 5-second sample of filtered PPG signals is shown in
Figure 5.
30000
20000
Amplitude
10000
10000
0 1 2 3 4 5
Time(seconds)
Figure 5. A 5-second sample of filtered PPG signal
10
Preliminaries
Table 1. HRV parameters with definitions [51]
Variable Units Description

AVNN ms Average of NNIs
The square root of the mean of the sum of the
RMSSD ms
squares of differences between adjacent NNIs
SDNN ms Standard deviation of all NNIs
LF ms2 Power in the low-frequency range (0.04–0.15 Hz)
HF ms2 Power in the high-frequency range (0.15–0.4 Hz)
LF/HF - LF to HF ratio
2.3 Heart Rate Variability

HRV shows the variation in the time interval between successive heartbeats and re-
flects alterations in the regulation of the autonomic nervous system (ANS) [40; 41].
Studies have shown that HRV is associated with stress level, sleep quality, and pain
intensity [42; 43; 44; 45; 46], and changes in HRV can be a sign of health issues such
as anxiety [47], preeclampsia [48], and hypertension [49].
HRV parameters are extracted based on interbeat intervals (IBIs), defined as the
interval between two consecutive heartbeats. IBIs will be processed further to re-
move abnormal IBIs based on the human HR range, which results in normal interbeat
intervals (NNIs). The time-domain HRV parameters show the NNIs variations in the
time domain, and the frequency-domain HRV parameters show the NNIs variations
in different frequency ranges. Time-domain and frequency-domain HRV parameters
used in this dissertation and their definitions are explained in Table 1. Root mean
square of the successive differences (RMSSD) reflects the vagal changes. Standard
deviation of NNIs (SDNN) during rest is associated with the fluctuations in respira-
tory sinus arrhythmia modulated by the parasympathetic nervous system. and is an
indicator of stress resilience. Power in the low-frequency range (LF) is affected by
the activity of both parasympathetic and sympathetic systems, whereas power in the
high-frequency range (HF) is mainly affected by the parasympathetic nervous system
[40]. LF and HF are correlated with mental stress [42]. LF/HF shows the LF to HF
ratio and is associated with psychological stress. Moreover, studies have shown that
stressful conditions affect average NNIs (AVNN), RMSSD, HF, and LF/HF [42; 50].
Wearable devices provide opportunities for low-cost continuous monitoring of
physiological health parameters, including HR and HRV parameters. The devices
mostly implement PPG sensors for HR and HRV monitoring. As described in Sec-
tion 2.2.2, PPG is a noninvasive optical method that measures changes in blood flow
volume. PPG signals are then used to extract pulse rate variability (PRV), which rep-
resents the variation of peak-to-peak intervals. PRV is widely used as a substitute for
HRV and can be used to extract HRV parameters such as AVNN, SDNN, RMSSD,
11
Fatemeh Sarhaddi
PNN50, LF, and HF. Many studies have evaluated the agreement between PRV and
HRV. Few studies have indicated the difference between PRV and HRV in specific
situations such as cold exposure [52; 53]. However, even these studies claim that
HRV is the primary determinant of PRV. Moreover, authors in [51] showed that HRV
parameters could reliably be estimated by PPG signals with sufficient confidence.
Here, we also used HRV for PRV derived from PPG signals.
HRV parameters can be extracted from segments of IBIs with different lengths.
Based on the literature, three standards for HRV analysis exist [40; 54; 55]:
• Long-term HRV analysis: This method is the gold standard for clinical HRV
analysis and uses 24-hour recordings to extract HRV parameters.
• Short-term HRV analysis: This standard uses 5-minute recordings to extract

HRV parameters.
• Ultrashort-term HRV analysis: This method uses less-than-5-minute record-

ings to extract some of the HRV parameters.
In this thesis, the short-term HRV analysis was used to extract HRV parame-
ters. This standard was selected based on the battery life of the wearable device and
collected data in the case studies.
12
3 IoT-Based Maternal Health Monitoring
In this chapter, we present a long-term IoT-based remote maternal monitoring sys-

tem. The system continuously monitors maternal stress, sleep, and physical activity,
and it can be extended to include other parameters such as diet. The system collects
subjective and objective data using various data sources, and it stores and analyzes
the data remotely on a cloud server. The data can then be visualized through web and
mobile applications. We evaluate the performance of the system in terms of feasibil-
ity, reliability, and energy consumption. We also discuss the practical challenges in
developing such systems.
3.1 Related Works

Many studies provide remote maternal-health monitoring that aims to improve ma-
ternal health. Some studies have used periodic measurements of health parameters
such as blood pressure, weight, and blood glucose to investigate specific issues such
as hypertension [6] and diabetes [56]. Other studies have used mobile applications
with periodic health parameter measurements to predict the risk of health issues such
as hypertension [57] and preterm birth [5].
Studies have also used wearable devices to collect health parameters continu-
ously from pregnant women. Lopez et al. [58] used a wristband to monitor HR,
sleep, and activity for 3 months. They used the collected data to monitor hyperten-
sion. Another study [8] presented a personalized sleep quality assessment method
using an IoT-based monitoring system. The system used a wristband to collect HR,
sleep, and physical activity. Kumar et al. [4] proposed a general architecture for
IoT-based maternal-monitoring systems.
These studies are limited as they only consider specific health issues in a short
time and use limited data sources. Thus, a long-term, IoT-based maternal monitoring
system is needed to provide continuous monitoring of health parameters and enable
a better understanding of normal and abnormal changes during pregnancy. Such a
system would improve the well-being of pregnant women by providing a holistic
view of their health and identifying the risk factors.
In the following, we first discuss the monitoring services that are essential for
maternal monitoring. Next, we introduce our presented IoT-based maternal health
system. Then, we evaluate the quality attributes of the system, including feasibility,
13
Fatemeh Sarhaddi
reliability, and battery life. Finally, we discuss the practical challenges of this system.
3.2 Maternal Health Monitoring Services

This section highlights the importance of extended maternity care services, such as
physical activity monitoring, sleep monitoring, and stress monitoring, which could
be provided through IoT-based systems for remote maternal health monitoring.
• Physical activity monitoring: Physical activity is a key factor in well-being.

Physical activity normally decreases during the course of pregnancy [59; 9].
However, a decrease in physical activity during pregnancy increases the risks
of obstetric complications [60]. Continuous objective monitoring of physi-
cal activity provides accurate and detailed information for both health care
providers and pregnant women to improve counseling and to be alert to the
risks at an early stage.
• Sleep monitoring: Sleep disorders are common during pregnancy, especially

during the third trimester [61]. In addition, sleep disorders can increase the
risk of gestational hyperglycemia [62], preterm birth [63], and mood disorders
[64]. Continuous sleep monitoring will enable a deeper understanding of in-
dividual sleep parameters. As a result, health care providers can better help
a pregnant woman to have a healthy circadian rhythm, which also results in
better weight management [65].
• Stress monitoring: Pregnant women can experience more stress compared to

nonpregnant women due to pregnancy. For example, physiological changes,
pregnancy-related symptoms, and the fetus’s health can be sources of stress
during pregnancy. A high maternal stress level is associated with a higher risk
of depression, hypertensive disorders, and preterm birth, to name a few [66;
67]. Therefore, continuous stress monitoring during pregnancy will provide
the possibility to detect high-stress levels early and better enable health care
providers to consult with and support highly stressed pregnant women.
3.3 Presented Long-term IoT-Based Maternal Health

Monitoring System
In this section, we present our long-term, IoT-based maternal monitoring system.
This system continuously collects subjective and objective health data from preg-
nant women and provides a holistic view of their well-being. Figure 6 indicates
an overview of the presented system, which consists of four layers: perception layer,
gateway layer, cloud layer, and application layer. The data are collected using various
data sources in the perception layer. Then, collected data are transferred to the cloud
14
IoT-Based Maternal Health Monitoring
Perception Layer
Healthcare
provider
Wearable Device
Smartphone Gateway Layer Cloud Layer Application Layer
Portable Device Smartphone Data Storage Web Application
Background Router Data Analysis Mobile Application

Information
Figure 6. An overview of the presented IoT-based maternal health monitoring system [34]
server using the gateway layer. Next, the data are stored and analyzed within the
cloud layer. Finally, the application layer presents the data to the users and provides
communication functionalities between users and caregivers and/or researchers. In
the following, we briefly describe these four layers in our implemented system.
3.3.1 Perception Layer

The perception layer collects data from various data sources including wearable de-
vices, smartphones, portable devices for periodic monitoring, and background and
demographic information. In our system, the data sources consist of a smartwatch,
a customized cross-platform mobile application, a blood pressure device, and back-
ground information.
Wearable Devices
We used the Samsung Gear Sport smartwatch [68] as a wearable device in the mon-
itoring system. The smartwatch was selected based on its sensors, configurability,
and providing access to raw signals. It also has an inertial measurement unit (IMU),
acceptable battery life, weight, and internal memory.
The watch has PPG, accelerometer, and gyroscope sensors that can be utilized
to extract HR, HRV, physical activity, and sleep parameters. The watch uses the
sensors and continuously collects HR, physical activity, and sleep parameters. The
smartwatch processes the collected signals and provides step counts, walking steps,
running steps, distance, activity duration, and activity intensity per 10 minutes. The
watch also provides duration, start and end of the sleep, and intensity of hand move-
ment during sleep. Moreover, HRV parameters as an indicator of stress can be ex-
tracted using raw PPG signals.
In addition, the watch runs the Tizen operating system (OS) [69]. Tizen OS is an
open-source OS used in various wearable devices [70] and enables the development
of customized data collection applications. In the presented maternal monitoring
15
Fatemeh Sarhaddi
system, we developed several programs for 12 minutes PPG signals collection every
second hour and acquire daily activity and sleep data supplied by the watch. The
collected data were stored in the smartwatch’s internal memory. We developed an
application for compressing and sending the data to the cloud server using a Wi-Fi
connection. After successfully transferring the data to the server, the application
removes the data from the internal watch memory. The smartwatch has sufficient
memory to store data for more than a month. We asked the participants to wear the
watch continuously and upload the data to the cloud server frequently.
Smartphone
Smartphone can be used to collect self-report and sensor data. In our system, we
developed a customized cross-platform mobile application for smartphones to col-
lect self-report data by providing questionnaires on daily, weekly, and specific time
and random bases. (See Appendix A: questions provided by the mobile application
for more details). The mobile application includes components for users to send
health data (e.g., blood pressure) or report technical problems, or for caregivers or
researchers to send push notifications, as well as for authentication and authorization.
We developed the customized mobile application with Angular 2 [71] and Cordova
[72] open-source frameworks.
Portable Devices
Portable devices can be used to measure physiological parameters periodically. In

our monitoring system, each participant was provided with an OMRON M3 Intel-
lisense blood pressure device [73] and was asked to measure their blood pressure
at least once a week. The device is clinically validated, and even nonexpert users
can measure blood pressure accurately due to the 360-degree accuracy feature of the
cuff [74]. Moreover, a nurse researcher instructed the participants to use the blood
pressure device properly in a face-to-face meeting. The blood pressure values were
manually entered into our customized mobile application’s blood pressure measure-
ment component and sent to the cloud.
Background Information
Background information (age, previous miscarriage, weight before pregnancy, etc.)

can help to assess the risk of pregnancy complications. We collected background in-
formation by sending a questionnaire via our mobile application to the participants at
the beginning of the study. The questionnaire considers the history of previous preg-
nancies, diagnosed diseases, risk factors in the current pregnancy, lifestyle before
and during the pregnancy, and stress level.
16
3.3.2 Gateway Layer

Two types of gateway devices (i.e., smartphone and Wi-Fi router) can be used in
the monitoring system to transmit data to the cloud layer. Our customized mobile
application uses the Internet connectivity of the smartphone to transfer the data to
our cloud server. In addition, the smartwatch connects to the Internet through a Wi-
Fi router to send the data.
3.3.3 Cloud Layer

The cloud server consists of user management, data management, and data analysis
modules. The user management module is responsible for creating and modifying
user accounts with different levels of authorization, creating and modifying question-
naires, creating and modifying various groups of users with specified question sets,
and scheduling questionnaires and notifications.
The data management module is responsible for receiving, basic validating, and
storing the received data. This module uses the functionality of the user management
module for authentication upon receiving the data from the smartwatch or the cross-
platform mobile application. Then, the data management module checks whether the
received data are corrupted. If necessary, the user will be notified to upload the data
again. For privacy concerns, no personal data are sent to or stored in the cloud server.
The data analysis module is responsible for preprocessing, analyzing, and mod-
eling the collected data and providing monitoring services. The monitoring services
provided by the data analysis module are stress monitoring, sleep monitoring, and
physical activity monitoring. This module can also be extended to provide other
monitoring services, such as diet monitoring.
Stress monitoring is performed by HR and HRV monitoring using PPG signals
collected from the smartwatch (see Section 2.3 for more details). Studies have shown
the correlation of several HRV parameters with different types of stress [42; 50].
Based on their results, stress is associated with AVNN, RMSSD, SDNN, LF, HF, and
LF/HF HRV parameters [42; 50]. Stress-monitoring services in this system provide
monitoring of HR and the aforementioned HRV parameters.
The sleep monitoring service uses the sleep events, hand movements, and phys-
ical activity events collected by the watch to extract sleep parameters such as to-
tal sleep time, sleep fragmentation, and wake after sleep onset, as described by
Mehrabadi et al. [75].
The physical activity monitoring service uses the activity data provided by the
watch along with wear-time data to extract physical activity parameters and time
spent being sedentary. To provide these monitoring services, the data analysis pipeline
used in this module comprised several steps as shown in Figure 7.
• Preprocessing: In the preprocessing step, unreliable data are detected and re-
17
Fatemeh Sarhaddi
Pre-processing Parameter extraction
PPG quality Stress (HRV)

assessment parameters
Sleep validation and Missing data

Collected data Sleep parameters Modeling
adjustment imputation
Physical activity Physical activity

validation parameters
Figure 7. Data analysis pipeline in the presented IoT-based maternal monitoring system [34]
moved from processing. The data are collected in everyday life settings and are
prone to noise. Assessing the quality of the collected data improves the accu-
racy of decisions made based on the data. We used simple rule-based methods
for physical activity and sleep data [75] as well as deep-learning-based PPG
quality assessment [76] for PPG signals. The PPG quality assessment method
is described in detail in Chapter 4.
• Parameter extraction: This step uses reliable data from the previous step to
extract HR, HRV, sleep, and physical activity parameters.
• Missing-data imputation: The proposed missing-data imputation methods

by Azimi et al. [77] were used to fill in the missing values in the data.
• Modeling: In this step, machine-learning methods were used to detect anoma-

lies, find trends in the data, and create personalized modeling using methods
previously proposed by our research group [8; 78].
The server was implemented using Apache 2 [79], the Flask framework [80],
and MongoDB [81]. Apache 2 is an efficient, extensible, open-source server that
is widely used. Flask is a scalable and flexible open-source framework in Python.
In addition, MongoDB is a NoSQL flexible database that we used to store various
types of collected data. We also used the Secure Sockets Layer (SSL) Application
Programming Interface (API) to enhance the security of communication.
3.3.4 Application Layer

A web application was developed to visualize the data collected from participants.
The application provides monitoring functionalities for researchers to follow partici-
pants’ health parameters. The web application provides various reporting tools (e.g.,
daily and weekly trends and participants’ answers to the questionnaires). As with our
mobile application, the web application was implemented in Angular 2 [71]. This
18
approach allows the same component to be used in both applications and decreases
the implementation overhead.
3.4 Data collection

The presented IoT-based maternal monitoring system was investigated and evalu-
ated in a longitudinal study on high-risk pregnant women. Pregnant women were
asked to wear a smartwatch programmed to continuously collect sleep and physi-
cal activity and 12 minutes of PPG signals every other hour. Moreover, participants
were asked to install our cross-platform mobile application on their smartphones
(Android or iOS). The mobile application provides a questionnaire that includes 2-
3 questions daily about the participants’ subjective mental and physical parameters.
(Daily questions can be found in Appendix A). The mobile application also provides
a component for participants to enter their blood pressure measurements and send
them to our server. A blood pressure measurement device was also provided for each
participant, and they were asked to measure their blood pressure and enter it into the
mobile application at least once a week.
Participants and Recruitment: We recruited 32 pregnant women with single-
ton pregnancies and previous preterm births (gestational weeks 22–36) or late mis-
carriages (gestational weeks 12–21) through advertisements in social media and ma-
ternity clinics. The participants were in early pregnancy (12–15 gestational weeks),
were older than 18 years, and were able to understand Finnish. The recruitment
happened in January-December 2019 in Southwest Finland.
The eligible pregnant participants attended a face-to-face meeting with the re-
searcher, where the study’s objectives were explained, and written consent was ob-
tained. Each participant received a smartwatch and a blood pressure device, along
with instructions on proper usage. The researcher also guided participants on in-
stalling our custom app on their phones. The participants agreed to participate in the
study from gestational weeks 12–15 to 3 months postpartum. Four women withdrew
from the study, resulting in a total of 28 participants who successfully completed the
study. This study was later extended by recruiting another group of pregnant women
with low-risk pregnancies. We used the data from the initially recruited group to
evaluate the system.
Research Ethics: This study received ethics approval from the Ethics Com-
mittee of the Hospital District of Southwest Finland (Dnro: 1/1801/2018). Written
informed consent was obtained from all participants.
3.5 System Evaluation

We evaluated the presented IoT-based maternal monitoring system by implementing
the system for recruited high-risk pregnant women. The evaluation considered the
19
Fatemeh Sarhaddi
feasibility, reliability, and energy consumption of the presented system. We also

discuss the practical issues in the development of the system.
3.5.1 Feasibility
We investigated the feasibility of the presented maternal monitoring system in terms
of the wear-time of the smartwatch and the use of our customized mobile application
during pregnancy and postpartum.
Wearable Device Usage

Figure 8 depicts the average daily wear-time of the smartwatch during pregnancy
and postpartum. As shown in Figure 8, the participants used the watches around
18 hours/day in the first weeks after recruitment and the time decreased during the
course of pregnancy to around 15.5 hours/day. After the delivery, the average daily
wear-time was less than the wear-time during pregnancy. The wear-time during the
postpartum period began at 11 hours, then increased after the first weeks, reach-
ing 14.5 hours at the end of the study. In this study, the average wear-time during
pregnancy and the postpartum period was 17.01±4.20 hours/day and 13.72±5.71
hours/day, respectively.
Pregnancy Postpartum Period

22.5
Wearing Time (Mean Hours/Day)
20.0
17.5
15.0
12.5
10.0
7.5
5.0
13 16 17 20 21 24 25 28 29 32 33 36 37 41 12 34 56 78 9 10 11 12
Pregnancy Weeks PostPartum Weeks
Figure 8. Average wear-time of the 28 high-risk pregnant women during pregnancy and
postpartum [34]
It should be noted that several technical and practical issues affected the wear-
time during the study, such as hospitalization during pregnancy, preterm births, hos-
pitalization of the newborn baby after delivery, work-related restrictions regarding
wearing the device, and technical issues in the server for a few days.
The results show the feasibility of using the wearable device in this system during
pregnancy and the postpartum period. Our findings align with [12], showing the
average wear-time of 17.3 hours/day during pregnancy and 14.4 hours/day during
1 month postpartum. However, our results are slightly lower compared with this
20
work, which could have resulted from hospitalization and pregnancy complications
because our participants had high-risk pregnancies.
Cross-Platform Mobile Application Usage
This section investigates the usage of our cross-platform application in terms of an-
swering daily questions and uploading blood pressure measurements using the appli-
cation. Figures 9 and 10 show the average responses to daily questions and average
weekly blood pressure measurements during the study, respectively. As shown in
Figure 9, the question response rate of the questions began at 70% and slightly de-
creased during the pregnancy. During postpartum, the response rate was less than it
was during pregnancy. It began at around 50% after delivery and increased between
the second and fourth week postpartum, and then decreased until week 10 and again
slightly increased. In total, participants replied to 5493 daily questionnaires. Of the
replies, 3879 were during pregnancy and 1614 were after the delivery. The average
mobile application usage in terms of the response rate was 67.5% in pregnancy and
57.0% in the postpartum period.

100
90
Average App Usage (Percentage)
80
70
60
50
40
30
20
13 16 17 20 21 24 25 28 29 32 33 36 37 41 12 34 56 78 9 10 11 12
Figure 9. Average mobile application usage (response rate) of the 28 high-risk pregnant women
during pregnancy and postpartum [34]
Figure 10 displays the weekly average of blood pressure measurements that were
uploaded via the mobile application throughout pregnancy and the postpartum pe-
riod. We removed one participant as an outlier because she measured her blood
pressure daily. Moreover, 10 participants stopped uploading their blood pressure
measurement after delivery. The results show that the average weekly blood pressure
measurements during pregnancy and postpartum was 0.74 and 0.29, respectively.
To the best of our knowledge, this is the first study to explore the feasibility of
mobile application usage in terms of answering daily questions during pregnancy and
postpartum. Our results show the feasibility of the mobile application in our system
for maternal monitoring during pregnancy and postpartum.
21
Fatemeh Sarhaddi

1.4
Average Number of Measurement (Per Week)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.2
13 16 17 20 21 24 25 28 29 32 33 36 37 41 12 34 56 78 9 10 11 12
Figure 10. Weekly average blood pressure measurements of 27 high-risk pregnant women during
pregnancy and postpartum [34]
3.5.2 Robustness and Reliability of Measurements

Robustness and reliability of measurements are crucial features in remote health
monitoring systems. The reliability of the health decisions in these systems is corre-
lated with the reliability of the collected data. Therefore, it is critical to ensure the
reliability of collected data in such systems.
In the presented system, PPG signals are collected in everyday life settings to
extract HR and HRV parameters. The quality of PPG signals is highly affected by
environmental noise and motion artifacts. In this section, we discuss the reliability
of HR and HRV parameters by investigating the frequency and duration of collected
PPG signals, and we show that our system satisfies the requirements for reliable HR
and HRV data collection. The reliability and validity of the collected PPG signals
will be discussed in Chapter 4. We also introduced a PPG signal quality assessment
method for improving the quality of collected HRV data in Chapter 4.2.4.
Duration of PPG Signal Recording
In this study, the PPG signal was used to extract HR and HRV parameters. As ex-
plained in Section 2.3, three standards exist for HR and HRV measurements with
different durations of recording (e.g., 24-hour recording, 5-minute recording, and
less than 5 minutes recording [40; 54; 55]). A 24-hour recording can result in the
most accurate results because it captures changes across an entire day. However, it is
inapplicable in long-term health monitoring because wearable devices have limited
battery capacity. See the Section 3.5.3 for more details.
Considering battery limitation, we decided to use the short-term HRV measure-
ment standard with 5-minute PPG recordings in our long-term maternal monitoring
system. We developed an application for the watch to acquire 12-minute PPG data
every second hour. Therefore, after removing the sensor calibration data, we would
22
have two consecutive 5-minute PPG recordings. In this way, we reduced the effect
of noisy signals and increased the reliability of extracted HRV parameters.
Sampling Frequency of the PPG Signal
The sampling frequency of the PPG signal is another factor that affects the reliability
of HR and HRV parameters. Choi et al. [51] compared the reliability of HR and HRV
parameters extracted from PPG signals with sampling frequencies ranging from 5 to
10000 Hz with an ECG signal with 10000 Hz sampling frequency. They showed that
desired HRV parameters in this study, including AVNN, RMSSD, SDNN, LF, and
HF, could reliably be extracted with a minimum frequency of 20 Hz. Therefore, our
setup met the minimum requirement for extracting reliable HRV parameters.
3.5.3 Energy Consumption
The wearable devices used in remote health-monitoring systems have limited battery
capacity and must frequently be recharged. The energy consumption of wearable
devices is a critical issue in remote health-monitoring systems and can affect the
feasibility and usability of such systems. In this section, we investigate the energy
consumption of the smartwatch used in our system as one of the feasibility aspects
of the system.
PPG sensors in wearable devices and in the smartwatch used in our system have
high energy consumption [14; 82]. In the previous section, we discuss the PPG
collection duration and sampling frequency of PPG signals. Here, we investigate
the time interval between PPG recordings to optimize the energy consumption of the
smartwatch.
We performed an experiment on the battery life of the smartwatch with different

time intervals between PPG signal recordings. In this experiment, we programmed
three smartwatches to collect 12 minutes of PPG signals with 20-Hz frequency at
different time intervals (e.g., 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours).
The watches were on a table and in flight mode to remove the usage bias. The
average battery life of the watches is reported in Figure 11. The battery life for 15-
minute intervals is about 25 hours and increased to 157 hours for 4-hours intervals.
Considering that daily usage of the smartwatch will also decrease the battery life,
we decided to have 2-hour intervals between PPG recordings. This interval results
in 2 to 3 days of battery life in a normal situation and ensures the feasibility of the
smartwatch in long-term monitoring.
23
Fatemeh Sarhaddi
160
140
120
Battery life (hours)
100
80
60
40
20
0
15
30
45
60
75
90
105
120
135
150
165
180
195
210
225
240
PPG collection intervals (minutes)
Figure 11. Smartwatch battery life with different intervals of PPG signal collection [34]
3.5.4 Practical Challenges

Long-term remote health monitoring raises several practical challenges, such as long-
term engagement of users, maintenance of systems and devices despite evolving
technologies, and user experience that we also faced during the course of this study.
To keep users motivated to use an IoT system for a long time, the system should
be usable and comfortable in everyday life settings, and the provided monitoring ser-
vices should be effective [11]. We selected a smartwatch that can be used easily in
everyday life activities. We also tried to minimize the effort to use the system by
minimizing the interaction needed with the system and selecting a reasonable bat-
tery life for the wearable device. We provided technical support through the mobile
application to help solve any technical issues our participants faced when using the
system. In addition, the monitoring of stress, sleep, and physical activity provided
by our system are essential to improve the health and well-being of pregnant women
as described in Section 3.2.
Moreover, the IoT-based monitoring system should be easily integrated with the
evolving technology in the long term. We used open-source frameworks such as
Flask, Angular, and MongoDB and development methodologies such as RESTFul
API that improve the maintenance of the system regardless of changing technology.
Moreover, backward compatibility in Tizen OS, which runs in the smartwatch, en-
ables our program to run on the watch regardless of updates on the OS. In the future,
we will also use other APIs, such as Validic API, to enable us to collect data from
various monitoring devices easily.
24
4 Validation and Quality Assessment of
HR and HRV Parameters for Health
Monitoring
In Chapter 3, we presented a long-term IoT-based maternal monitoring system. We

indicated that the HRV collection in the presented system meets the requirements of
reliable measurements in terms of frequency and duration. In this chapter, we show
the necessity of validation and quality assessment of PPG-based HR and HRV pa-
rameters for health-monitoring applications. Then, we will investigate the accuracy
of the collected HR and HRV parameters in our monitoring system. Finally, we pro-
pose a deep-learning-based PPG quality assessment method to further improve the
quality of collected HRV parameters.
4.1 Necessity of Accuracy and Quality Assessment of

HR and HRV Parameters Collected by Wearable
Devices
HR and HRV parameters can be continuously collected by wearable devices at a
low cost. The devices mostly implement PPG sensors for HR and HRV monitor-
ing. PPG-based wearable devices provide opportunities to collect data in everyday
life settings. However, the devices are prone to noise, such as motion artifacts and
environmental noise, when people engage in various activities. The noise is unavoid-
able and negatively affects the PPG-signal quality and, consequently, the reliability
of extracted HR and HRV parameters [83; 38]. The unreliable physiological param-
eters can lead to false predictions or decisions in health care applications, such as
loneliness prediction, which is discussed in Chapter 6.
Moreover, the noise affects HR and HRV parameters differently, because the
accuracy and reliability of HR and HRV parameters depend on different factors of
the signals. For example, RMSSD, which shows the short-term variation in the NNIs,
can be affected by small distortion in signals [40]. However, SDNN, which indicates
the long-term variation of NNIs, will be unreliable if the noise affects the signal
variations [40]. Therefore, a PPG signal can be reliable for extracting some HRV
parameters and unreliable for extracting other parameters, as illustrated in Figure 12.
Figure 12 shows three different 1-minute segments of PPG signals that are af-
25
Fatemeh Sarhaddi
(a) PPG with reliable HR, AVNN, and (b) PPG with reliable HR, AVNN, and
SDNN, and unreliable RMSSD RMSSD, and unreliable SDNN
(c) PPG with reliable HR and AVNN,

and unreliable RMSSD and SDNN
Figure 12. Three 1-minute samples of PPG signals with varying accuracy for HR and HRV
parameters [84]
fected by different types of noise. The extracted HR and HRV parameters are shown
with their error compared to the ECG baseline. In Figure 12a, only a small part of
the signal was affected by the noise, highlighted in red. In this example, HR, AVNN,
and SDNN have acceptable errors compared with the ECG baseline. However, the
error rate for RMSSD is high because noise affects the short-term variations of the
signals. In Figure 12b, the peaks are not distorted, but the noise affected the variation
of the NNIs compared to ECG peak-to-peak intervals. In this example, HR, AVNN,
and RMSSD can be extracted reliably. However, SDNN, which is correlated with
long-term variations of the signal, is unreliable. In Figure 12c, part of the signal is
corrupted, and the noise affects the short-term and long-term variation of the signals.
Therefore, SDNN and RMSSD extracted from this signal are unreliable, whereas
HR and AVNN can be extracted reliably. The above examples were for time-domain
HRV parameters, but the situation is the same for frequency-domain parameters.
We must assess the quality of the smartwatch used in our IoT-based maternal
monitoring system to ensure the reliability of extracted parameters. Then, quality as-
sessment of the signals is required to distinguish the low- and high-quality segments
26
Validation and Quality Assessment of HR and HRV Parameters for Health Monitoring
of PPG signals and discard the unreliable parts of the signals. Moreover, based on
the above examples, the validation and quality assessment should be performed sep-
arately for HR and each HRV parameter.
4.2 Accuracy Assessment of the Smartwatch HR and

HRV Parameters
In this section, we assessed the accuracy of the Samsung Gear Sport smartwatch in
terms of HR and HRV parameters. We selected this smartwatch for accuracy assess-
ment because it was utilized in our system developed in Chapter 3. The evaluation
was performed against a medical-grade ECG device during 24 hours of data collec-
tion in an everyday life setting. The evaluation was performed separately for HR and
each HRV parameter during sleep and when awake.
4.2.1 Related Works

Several studies have assessed the HR accuracy extracted by wristbands in different
situations in various population groups. Accuracy of HR measurement using multi-
ple wristbands, including Apple Watch [85], Basis Peak [85], Empatica E4 [86], Fit-
bit Surge [85], Microsoft Band [85], PulseOn [85], Samsung Gear S2 [85], Garmin
Forerunner [87], Mio Alpha 2 [85], the Everlast Smartwatch[88], Fitbit Charge HR
[89], and Basis Peak [90] have been examined in the previous studies. The stud-
ies indicate high accuracy of HR measurement for the PPG-based wristband while
participants engaged in low-intensity activities. They also showed that HR accuracy
decreased with an increase in the intensity of activity. The studies only assessed the
accuracy of HR and were limited to predefined activities in laboratory settings.
Other studies have investigated the accuracy of HRV parameters considering var-
ious smartwatches and smart rings. Accuracy of HRV parameters extracted from the
Empatica E4 [91; 92], Microsoft Band 2 [93], the Wavelet wristband [94], Apple
Watch [95], and Oura ring [31; 96; 97] were evaluated against a medical-grade ECG
device. The studies showed that HR and HRV parameters collected by these devices
have high accuracy during rest. Moreover, they showed that the accuracy of the HRV
parameter was highly decreased by noise and motion artifacts. The studies were per-
formed in laboratory settings (mostly in seated positions) and are limited to less than
1-hour data collection. Thus, assessing the accuracy of the PPG-based smartwatch in
everyday settings involving various activities and situations is required. Moreover,
the accuracy of different HRV parameters should be evaluated separately because the
parameters are affected differently by noise.
Based on the limitation of previous works, the validation of the smartwatch HR
and HRV parameters should be performed separately for each parameter and in ev-
eryday life settings when participants can be involved in various activities.
27
Fatemeh Sarhaddi
4.2.2 Participants and Recruitment

We recruited healthy adults aged 18 to 55 in Southern Finland between July and
August 2019. The exclusion criteria included having a previous diagnosis of cardio-
vascular disease or displaying symptoms of illness during the recruitment, as well as
having any restrictions in doing physical activities or wearing wearable devices.
The participants were recruited using snowball sampling. The researcher sched-
uled a face-to-face meeting with the eligible participants. In the meeting, the study’s
objectives were explained to the participants. After obtaining written consent from
the participants, the wearable devices (a Gear Sport smartwatch [68] and a Shim-
mer ECG device [98]) and the instructions /were delivered to the participants. The
Shimmer device was positioned on the chest, and the smartwatch was worn on the
nondominant hand. The researcher helped the participants to use and wear the de-
vices if needed. The participants were asked to wear the devices continuously for 24
hours while carrying out their daily routines. Forty-six participants (23 female and
23 male) were recruited.
The data from 28 participants, including 14 males and 14 females, were used
in this assessment. We excluded the data from other participants due to various
issues during the data collection, such as missing sleep data, ECG electrodes loosely
attached to skin resulting in noisy ECG signals, and insufficient ECG data due to
practical issues.
Research Ethics The study was conducted according to the ethical principles
based on the Declaration of Helsinki and the Finnish Medical Research Act (No.
488/1999). The study protocol received a favorable statement from the ethics com-
mittee (University of Turku, Ethics Committee for Human Sciences, Statement no:
44/2019). The participants were informed about the study, both orally and in writ-
ing, before the written informed consent was obtained. Participation was voluntary,
and each of the participants had the right to withdraw from the study at any time
and without giving any reason. To compensate for the time used for the study, each
participant received a 20-euro gift card to a grocery store at the end of the monitoring
period when returning the devices.
4.2.3 Data Collection

We collected data using a Samsung Gear Sport smartwatch and a Shimmer3 ECG
device as the gold standard. We instructed participants to wear the smartwatch on
the wrist of the non-dominant hand. In addition, participants wore the ECG device
using a chest strap with four limb electrodes placed on the arms and the legs (see
[99] for more details). Moreover, participants filled out self-report questionnaires to
log their non-wear, sleep, and awake times.
Samsung Gear Sport Smartwatch: As described in Chapter 3, the Gear Sport
28
smartwatch [68] is an open-source smartwatch that enables customized data collec-

tion remotely. For this study, a customized smartwatch application was developed to
collect 16 minutes of PPG signals every 30 minutes with a frequency of 20 Hz. The
data collection duration was chosen to have a battery life of more than 24 hours. The
data were stored in the internal storage of the smartwatch and then sent to our cloud
server.
Shimmer ECG Device: We used the Shimmer3 ECG device [98] as a gold
standard for the collection of ECG data. The Shimmer is a lightweight device that
can be placed on the chest with a strap and utilizes four electrodes connected to the
skin for ECG signal collection. We configured the shimmer device to collect ECG,
accelerometer, gyroscope, and skin temperature data continuously. The collected
data were saved on the device. The Shimmer device has a battery life of more than
24 hours and did not need to be charged to collect our data. We sent the data to
our cloud server after data collection. We used Lead II ECG (right arm - left leg) to
extract HR and HRV parameters.
4.2.4 Data Analysis

In this case study, as with other case studies in this thesis, we used short-term HRV
analysis based on the duration of recordings (see Section 2.3). Therefore, we used
5-minute segments of ECG and PPG signals to extract HR and HRV parameters. The
segments were used as an input to our data analysis pipeline. The pipeline consisted
of preprocessing, peak detection, abnormal peak removal, feature extraction, and sta-
tistical analysis, which are explained in the following. Figure 13 shows the overview
of our data analysis pipeline.
Abnormal peak Feature Statistical

Preprocessing Peak detection
removal extraction analysis
Figure 13. Data analysis pipeline [100]
Preprocessing
Preprocessing consists of synchronization and filtering.
• Synchronization: The ECG and PPG signals were collected by different de-
vices. Therefore, it was necessary to synchronized the signals. To find the time
difference between ECG and PPG, we used cross-correlation between acceler-
ation signals collected by the Shimmer ECG device and the smartwatch. Then
we shifted the ECG signals to synchronize them with the PPG signals.
29
Fatemeh Sarhaddi
• Filtering: For HR and HRV parameter extraction, only the frequencies in hu-
man HR ranges are used. Consequently, other frequencies were eliminated
using the Butterworth filter. A 5th-order high-pass Butterworth filter with a
cutoff frequency of 0.5 Hz was applied to the PPG signals, while a Butter-
worth bandpass filter with cutoff frequencies of 0.5–100 Hz was employed
for the ECG signals. The choice of cutoff frequencies was determined by the
frequency of the PPG and ECG signals.
Peak Detection
The peak detection step involves identifying peaks in both ECG and PPG signals.
The following will describe the peak detection methods for PPG and ECG signals
separately.
PPG Peak Detection
The PPG peak detection method proposed by Kazemi et al. [101] was used for
PPG peak detection. The method uses dilated CNNs and returns a probability of
being a peak for each signal point. Then, the peaks are detected as local maximums
within the points with a higher probability than a predefined threshold. Figure 14
shows the detected peak in a 30-second window of PPF signals.
ECG Peak Detection
A two-round moving threshold-based ECG peak detection method was used to
detect peaks in ECG signals. First, the average of all signal points in an ECG signal
window was used to detect peaks. Then, the average of the detected peaks served
as a new threshold to update the detected peaks. Moreover, the normal human HR
range was used to add undetected peaks. The method was used in [31; 100] and
performs well compared with the state-of-the-art methods [31]. Figure 14 shows
detected peaks in a 30-second ECG signal window.
Abnormal Peak Removal

The normal HR for humans ranges from 20 to 200 per minute. As a result, NNIs
range from 300 ms to 3000 ms. The ranges were used to remove abnormal peaks.
Moreover, the average NNIs in a single segment of the signal were used as a thresh-
old, and the NNIs with a variation of more than 20% of the threshold were discarded.
If more than half of a signal segment were removed due to abnormal peaks and NNIs,
the segment would be removed from the analysis.
Feature Extraction
We used the peak detected in the previous step to extract HR and HRV parameters.
The NNIs extracted from detected peaks were used to extract the following time-
30
30000
20000
PPG 10000
0
10000
0.2
0.1
ECG
0.0
0.1
0 5 10 15 20 25 30
Time (seconds)
Figure 14. A peak detection sample for a 30-second segment of PPG and ECG signals
domain and frequency-domain parameters: AVNN, RMSSD, SDNN, PNN50, LF,

HF, and LF/HF.
Statistical Analysis
We used the Pearson correlation coefficient, linear regression analysis with R-squared
value (𝑟2 ), and Bland-Altman analysis to validate the parameters extracted from the
watch against the ECG device. The statistical analysis was implemented in Python
using Scipy [102], sklearn [103], and Statsmodels [104] libraries.
4.2.5 Results and Discussion

We assess the validity of HR and HRV parameters during sleep and while awake.
The sleep time was extracted from self-report questionnaires that participants filled
in manually.
Table 2 illustrates the Pearson correlation coefficient, 95% confidence interval
(CI), mean difference, and 𝑟2 values of the HR and HRV parameters of the smart-
watch compared to the ECG gold standard. All p values for Pearson correlation are
less than 0.001; therefore, Pearson correlation coefficient values are statistically sig-
nificant. As shown in Table 2, all values are positively correlated. During sleep,
the correlation values for HR, AVNN, SDNN, and pNN50 are high. In addition,
RMSSD, LF, and HF have high but slightly lower correlations than previously men-
tioned parameters. LF/HF values are also moderately correlated. When awake,
AVNN values are highly correlated, whereas HR values are moderately correlated.
31
Fatemeh Sarhaddi
All other HRV parameters have low positive correlations.
Table 2. Pearson correlation coefficient, 95% CI, mean difference, and 𝑟2 values between the
smartwatch and Shimmer3 HR and HRV parameters in 5-minute segments of PPG signals [100]
Pearson
95% Confidence Mean
Parameters Correlation 𝑟2
Interval Difference
Coefficient
HR 0.941 [-7.53, 6.77] -0.38 0.882
AVNN 0.960 [-83.87, 108.59] 12.36 ms 0.909
RMSSD 0.778 [-68.49, 32.01] -18.24 ms 0.405
Sleep SDNN 0.802 [-72.66, 28.29] -22.19 ms 0.246
time PNN50 0.964 [-13.21, 11.58] -0.81 0.926
LF 0.784 [-1763.66, 834.77] -464.45 ms2 0.206
HF 0.782 [-1188.67, 693.23] -247.72 ms2 0.462
LF/HF 0.622 [-2.24, 1.72] -0.26 0.216
HR 0.675 [-43.58 , 24.65] -9.47 0.293
AVNN 0.833 [-135.12 , 254.44] 59.66 ms 0.582
RMSSD 0.251 [-79.59 , 91.89] 6.15 ms -0.191
Awake SDNN 0.404 [-76.99 , 61.36] -7.81 ms 0.099
time PNN50 0.277 [-26.0 , 60.99] 17.5 -1.62
LF 0.350 [-1727.72 , 1402.18] -162.77 ms2 0.075
HF 0.130 [-1215.82 , 1599.31] 191.75 ms2 -0.493
LF/HF 0.211 [-3.38 , 2.12] -0.63 -0.453
Figures 15 and 16 depict the regression lines for parameters extracted from PPG
collected by the smartwatch compared with ECG during sleep and while awake,
respectively. The ideal line is also illustrated in black. As shown in these figures,
during sleep, the regression line of HR, AVNN, and PNN50 are close to the ideal
line, and when awake, only the fitted line of HR and AVNN follow the ideal line.
The fitted lines of other HRV parameters diverge both during sleep and when awake.
In addition, the 𝑟2 values, which indicate how well the data fit the regression line,
are shown in Table 2. The 𝑟2 values of HR, AVNN, and pNN50 during sleep and 𝑟2
of AVNN when awake are high. HR also has moderate 𝑟2 when participants were
awake. However, other parameters have low 𝑟2 values during sleep and while awake.
The Bland-Altman plots during sleep and when awake are shown in Figures 17
and 18, respectively. The mean biases and 95% CIs are given in Table 2. As shown
in Table 2, during sleep, the smartwatch’s AVNN values have very low mean bias
with AVNN extracted from the ECG device. RMSSD, SDNN, pNN50, and LF/HF
also show low mean biases, whereas LF and HF have moderate mean biases. During
waking hours, RMSSD and SDNN have relatively low mean biases, whereas other
parameters show moderate mean biases. In addition, during sleep, HR, RMSSD,
32
Scatter plots and regretion analysis for HR and HRV parameters during sleep
HR AVNN
100 1400
80 1200
1000
ECG
ECG
60
800
40
600
20
20 40 60 80 100 600 800 1000 1200 1400
PPG PPG
RMSSD SDNN
250 200
200 150
150
ECG
ECG
100
100
50 50
0 0
0 50 100 150 200 250 0 50 100 150 200
PPG PPG
100 PNN50 5000 LF
80 4000
60 3000
ECG
ECG
40 2000
20 1000
0 0
0 20 40 60 80 100 0 1000 2000 3000 4000 5000
PPG PPG
5000 HF 8 LF-HF
4000 6
3000
4
ECG
ECG
2000
1000 2
0 0
0 1000 2000 3000 4000 5000 0 2 4 6 8
PPG PPG
Figure 15. The scatter plots and regression analysis of the HR and HRV parameters collected
from the Samsung smartwatch and Shimmer ECG in 5-minute segments during sleep. The
regression lines and ideal lines are indicated in red and black, respectively. [100]
SDNN, and pNN50 have narrow CIs, whereas AVNN, LF, HF, and LF/HF have wide
CIs. During the awake time, HR and all HRV parameters have wide CIs.
As shown in Figure 17 and Figure 18, the smartwatch underestimates AVNN
while overestimating HR and other HRV values during sleep. However, during wak-
ing hours, the watch overestimates AVNN, RMSSD, and pNN50 and overestimates
33
Fatemeh Sarhaddi
Scatter plots and regretion analysis for HR and HRV parameters while awake
160 HR AVNN
1400
140
120 1200
100 1000
ECG
ECG
80 800
60 600
40
400
20
20 40 60 80 100 120 140 160 400 600 800 1000 1200 1400
PPG PPG
RMSSD SDNN
250 200
200 150
150
ECG
ECG
100
100
50
50
0 0
0 50 100 150 200 250 0 50 100 150 200
PPG PPG
80 PNN50 5000 LF
60 4000
3000
40
ECG
ECG
2000
20 1000
0 0
0 10 20 30 40 50 60 70 80 0 1000 2000 3000 4000 5000
PPG PPG
5000 HF 8
LF-HF
4000
6
3000
ECG
ECG
4
2000
1000 2
0 0
0 1000 2000 3000 4000 5000 0 2 4 6 8
PPG PPG
Figure 16. The scatter plots and regression analysis of the HR and HRV parameters collected
from the Samsung smartwatch and Shimmer ECG in 5-minute segments when awake. The
regression lines and ideal lines are indicated in red and black, respectively. [100]
other parameters (i.e., HR, SDNN, LF, and LF/HF).

The results show that during sleep, the smartwatch has high accuracy regard-
ing HR, AVNN, and pNN50 and acceptable accuracy for RMSSD, SDNN, LF, and
HF. However, during waking hours, only AVNN and HR values extracted from the
smartwatch have satisfactory accuracy. Therefore, we need to improve the signal
34
Bland-Altman analysis for HR and HRV parameters during sleep

10
HR AVNN
+1.96 SD: 6.8 150 +1.96 SD: 1.1e+02
100
5
50
Difference
Difference
0 mean diff: mean diff:
-0.38 0 12.36
5 50
-1.96 SD: -7.5 100 -1.96 SD: -84
10
40 50 60 70 80 90 700 800 900 1000 1100 1200 1300 1400
Means Means
RMSSD SDNN
+1.96 SD: 32 40 +1.96 SD: 28
40
20 20
0 0
Difference
Difference
mean diff: 20 mean diff:
20 -18.24 -22.19
40 40
60 60
80 -1.96 SD: -68 80 -1.96 SD: -73
25 50 75 100 125 150 175 200 20 40 60 80 100 120 140 160
Means Means
PNN50 LF
15 +1.96 SD: 12 +1.96 SD: 8.3e+02
1000
10 500
5 0
Difference
Difference
0 mean diff: 500

mean diff:
-0.81 -464.45
5 1000
10 1500
15 -1.96 SD: -13 2000 -1.96 SD: -1.8e+03
0 20 40 60 80 0 500 1000 1500 2000 2500 3000
Means Means
HF LF-HF
1000 +1.96 SD: 6.9e+02 +1.96 SD: 1.7
2
500
1
Difference
Difference
0
-247.72 -0.26
500 1
1000 2
1500
-1.96 SD: -1.2e+03 -1.96 SD: -2.2
3
0 500 1000 1500 2000 2500 3000 3500 0 1 2 3 4 5 6 7
Means Means
Figure 17. Bland-Altman plots of the HR and HRV parameters in 5-minute segments obtained by
the smartwatch and Shimmer3 during sleep [100]
quality by using noise cancellation techniques [105] or quality assessment methods

[106; 107] to discard the noisy part of the signals and only use the reliable part of
signals to extract HRV parameters accurately. In the following section, we introduce
a deep learning-based quality assessment method to distinguish the reliable and un-
reliable segments of the PPG signals compared with the ECG gold standard, which
can be used to improve the accuracy of extracted HR and HRV parameters.
35
Fatemeh Sarhaddi
Bland-Altman analysis for HR and HRV parameters while awake

40
HR AVNN
+1.96 SD: 25 300 +1.96 SD: 2.5e+02
20
200
Difference
Difference
0 100
mean diff: mean diff:
-9.47 59.67
20 0
40 100
-1.96 SD: -44 200
-1.96 SD: -1.4e+02
60
40 60 80 100 120 140 160 180 600 800 1000 1200 1400
Means Means
RMSSD SDNN
+1.96 SD: 92 +1.96 SD: 61
100
50
50
Difference
Difference
0 6.15 -7.81
50 50
100 -1.96 SD: -80 100

-1.96 SD: -77
40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140
Means Means
80
PNN50 LF
+1.96 SD: 61 2000 +1.96 SD: 1.4e+03
60
1000
40
Difference
Difference
20 mean diff: 0 mean diff:

17.49 -162.98
0 1000
20
40 -1.96 SD: -26 2000 -1.96 SD: -1.7e+03
0 10 20 30 40 50 60 70 0 500 1000 1500 2000 2500 3000
Means Means
HF LF-HF
2000 +1.96 SD: 1.6e+03 +1.96 SD: 2.1
2
1000
Difference
Difference

0 191.75 -0.63
2
1000
-1.96 SD: -1.2e+03 4 -1.96 SD: -3.4
0 500 1000 1500 2000 2500 3000 1 2 3 4 5
Means Means
Figure 18. Bland-Altman plots of the HR and HRV parameters in 5-minute segments obtained by
smartwatch and Shimmer3 during awake [100]
36
4.3 A CNN-Based PPG Quality Assessment Approach

for HR and HRV Parameters
In this section, we propose CNN-based PPG quality assessment methods based on
extracted HR and HRV parameters. One 1-dimensional (1D) CNN and three 2-
dimensional (2D) CNN models are proposed for each HRV parameter. The proposed
models were evaluated using the data described in the previous section (see Section
4.2.2 and 4.2.3).
In the following, we first briefly explore the previous works in PPG quality as-
sessment methods. Then, we describe the data collection and data analysis. After
that, we introduce our proposed deep learning-based PPG quality assessment meth-
ods for HR and HRV parameters. The best method was selected for HR and each
HRV parameter based on the performance. Finally, we compare the performance of
the best proposed methods with the state-of-the-art PPG quality assessment methods.
The results show that the proposed methods outperform the other methods for HR
and HRV parameters.
4.3.1 Related Works

Various studies have proposed PPG quality assessment methods or PPG quality in-
dicators to distinguish reliable and unreliable segments of PPG signals. Several PPG
quality indicators based on the morphology of the signal, including baseline wan-
dering, skewness, and kurtosis, were used to estimate the quality of the PPG signals
[106; 108; 109].
Rule-based and template-matching methods have also been proposed to identify
low-quality segments of PPG signals [110; 111; 112; 113; 114]. These methods use
hierarchical rules or the similarity of the current segment of the signal with template
signals based on morphological features and predefined thresholds.
Moreover, several conventional machine-learning methods, including both su-
pervised and unsupervised methods, were used for PPG quality assessment. Several
traditional supervised methods such as decision tree [115; 116], random forest [117],
support vector machine (SVM) [118; 119; 116], K-nearest neighbors (KNN) [116],
and neural network [120] were developed for PPG quality assessment considering
the shape, and time and frequency domain parameters of the signals.
In addition, several deep learning-based PPG quality assessment methods have
recently been proposed, including real-time PPG quality assessment based on HR
value [76], 1D and 2D deep learning-based quality assessment for participants with
atrial fibrillation [121], a CNN-based method called DeepBeat [122], 1D CNN-based
quality assessment [123], and 2D CNN-based model in [124]. The deep learning-
based methods extract features automatically. However, even in these methods, the
labeling was performed based on HR and morphological features. In addition, most
37
Fatemeh Sarhaddi
of the PPG quality assessment methods were evaluated based on the collected data
in predefined lab settings with limited noise or with a small number of participants.
Table 3 summarizes the PPG quality assessment methods and their features.
Source code
Reference Features Method Annotation
Availability
Automatic Automatic labeling
Proposed
feature 1D CNN/ 2D CNN based on HR and ✓
method
extraction HRV parameters
Extracted HR Manual
Orphanidou et al [110] and morphological Threshold based rules based on the ×
Rule based features signal shape
Methods
Hierarchical
Reddy et al [111] Predictor Coefficient ” ×
decision rules
Statistical parameters Predefined

Tyapochkin et al [112] ” ×
of IBIs rules
Amplitude and Hierarchical

Vadrevu et al [125] ” ×
time-series features decision rules
Manual
Morphological
based on the
Alam et al [117] features and Random forest ×
signal shape
Supervised HR value
and HR values
methods
Frequency domain Manual
Zhang et al [119] and time series SVM based on the ×
characteristics signal shape
Frequency-domain,
Preira et al [118] time-domain, and SVM ” ×
non-linear features
Frequency-domain,
Preira et al [116] and time-domain SVM ” ×
features
Unsupervised Mahmoudzadeh et al [126] Statistical features Elliptical envelope ” ✓
methods
Entropy and
Roy et al [127] signal complexity Self-organizing map ” ×
features
Automatic
Preira et al [121] ResNet18 ” ×
feature extraction
Deep learning
methods Multi-task CNN
Soto et al [122] ” (pre-training with ” ×
CDAE)
Goh et al [123] ” 1D CNN ” ×
Roh et al [124] ” 2D CNN ” ×
Automatic labeling
Naeini et al [44] ” 1D CNN ×
based on HR
Table 3. Comparing PPG quality assessment methods [84]
To address previous works’ limitations, we propose PPG quality assessment

methods for HR and each HRV parameter. Moreover, the methods were evaluated
using data collected in everyday life settings. Thus, the model can be better general-
38
ized for data in different conditions. In addition, we proposed automatic labeling of

the data based on HR and HRV parameters using the gold standard baseline.
4.3.2 Data Collection and Data Analysis

The data described in Sections 4.2.2 and 4.2.3 were used to evaluate the proposed
methods. The 210 hours of simultaneous ECG and PPG signals collected from 36
subjects were used to evaluate the proposed method.
We used the same steps as the Section 4.2 for data analysis and extracting HRV
parameters (see Section 4.2.4). However, here, different peak detection methods
were used, and HR, AVNN, RMSSD, SDNN, and LF/HF HRV parameters were
extracted. Then, the extracted parameters from ECG and PPG signals were used for
labeling. Here, we first explain the peak detection methods. Then, the automatic
labeling will be described.
Peak Detection
We used the deep learning-based method proposed in [128] for ECG peak detection.
This method uses Long Short-Term Memory (LSTM) architecture to find the location
of the peaks. Then it removes the false peaks by a peak-to-peak distance-based
anomaly detection method. This method was selected because it performs better
than traditional methods such as Hamilton [129] and Pan-Tompkins [130] (see [128]
for more details).
For PPG peak detection, the method Van Gent et al. [131] implemented in the
HeartPy package in Python was used. This method uses moving average, adaptive
threshold, and outlier detection to detect real peaks in PPG signals. This method also
has acceptable accuracy compared with traditional methods [131].
Automatic Labeling
We used the HRV features extracted from each 5-minute segment of PPG signals for
labeling. We compared each feature (e.g., HR, AVNN, RMSSD, SDNN, and LF/HF)
from one PPG segment with the same feature extracted from the corresponding ECG
segment. If the difference is less than a predefined threshold, then the signal is re-
liable for that HRV parameter; otherwise, it is unreliable with regards to that HRV
parameter. The result of this step was five labels. The threshold for each HRV param-
eter was selected based on the normal range of the parameter as described in [132].
The threshold can be changed according to the acceptable accuracy for the desired
application.
39
Fatemeh Sarhaddi
4.3.3 Proposed PPG Quality Assessment Approaches

We present 1D and 2D CNN-based deep learning approaches for PPG quality assess-
ments. CNN-based architectures automatically identify the significant features with
low-to-zero preprocessing of the input data [133]. CNN also reduces the number of
parameters by sharing the parameters between layers, which speeds up the network
and training [134].
The architecture of our proposed CNN methods is illustrated in Figure 19. The
proposed 1D CNN method comprised two convolution layers, as shown in Figure 19.
A batch normalization layer follows each convolutional layer to prevent overfitting of
the model. Then, we used a max pooling layer to reduce the dimensionality. Finally,
we added a flattened layer, and a sigmoid function provided the classification results.
The kernel size used in this model was 1*3, which is common. We used 32 filters in
the first and 64 filters in the second convolutional layer. The optimization function
was stochastic gradient descent with a learning rate of 0.00001, and the loss function
was binary cross entropy.
Moreover, we used pretrained 2D CNN models VGG16 [135], ResNet50 [136],
and MobileNetV2 [137], which have shown good performance in similar applica-
tions, and we tuned them for our purpose. To make the models work with 2D input
(images), we converted PPG signals to images using the Garmin angular field (GAF)
method [138]. The GAF preserves the temporal dependencies in PPG signals while
transforming them into images. We used the transformed PPG image along with the
corresponding label as an input for the models.
Fully Connected
1*2 Max
Flattening &
Convolution
32 Filters 64 Filters
Convolution
5-min PPG Pooling

time-series
Sigmoid
Pooling
Normalization
Normalization
Batch
Batch
Reliable
Unreliable
Automatic PPG
Annotation
HR
AVNN
RMSSD 224*3*3 Deep Models
SDNN Reliable
VGG16
Unreliable
LF/HF
Reliable
ResNet50
Unreliable
5-min PPG Reliable

MobileNetV2
image Unreliable
Figure 19. Overview of our CNN-based PPG quality assessment models [84]
40
4.3.4 Model Implementation and Evaluation

The Keras Sequential API, which implements Tensorflow in Python, was used to im-
plement and train the presented models [139]. We used Adam optimizer and binary
cross-entropy as optimizer and loss functions for our models. Moreover, the accu-
racy, F1 score, and area under the curve (AUC) were used to evaluate the models.
The proposed models were trained separately for HR, AVNN, RMSSD, SDNN,
and LF/HF. Each model used the corresponding label generated from our automatic
annotation. Therefore, five models were trained for each method.
We first trained and validated the proposed models. Then, the model with the
best performance was chosen for comparison with state-of-the-art methods. We used
80%, 20%, and 20% of the collected data for training, validation, and test sets, re-
spectively. The distribution of class label in dataset is shown in table 4.
Table 4. Distribution of class labels for each parameter [84]
Train+ validation Test

parameters
Reliable Unreliable Reliable Unreliable
HR 46175 12413 11544 3104
AVNN 50925 7663 12732 1916
RMSSD 16327 42261 4082 10566
SDNN 25823 32766 6456 8192
LF/HF 42388 16200 10597 4050
Table 5 shows the performance of the proposed models on the validation set. As
shown in the table 5, the 1D CNN models outperform 2D CNN models for all HR and
HRV parameters. Therefore, we used the 1D models for HR and HRV parameters
and compared the 1D models with current methods. LF/HF achieved the highest
performance, whereas RMSSD had the lowest performance for all models.
Table 5. Validation Set performance results of different CNN models for various HR-HRV features
[84]
Model 1D CNN MobileNet ResNet50 VGG16

Metric ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC
HR 95.63 96.11 96.21 94.58 93.93 94.63 92.49 92.44 93.52 92.38 94.42 94.87
AVNN 96.71 97.68 97.71 95.68 95.95 96.93 90.43 90.22 91.32 93.36 94.29 94.85
RMSSD 91.42 91.48 91.69 89.68 89.44 89.46 86.47 85.81 87.17 85.72 85.09 85.79
SDNN 96.01 96.97 97.09 92.66 93.91 93.66 92.52 92.12 93.66 93.42 93.82 93.99
LF/HF 97.71 97.71 97.71 91.92 91.95 92.93 90.22 90.22 91.31 94.36 94.29 93.99
41
Fatemeh Sarhaddi
Table 6. Performance of the proposed and state-of-the-art models on Test sets [84]
Label HR AVNN RMSSD SDNN LF/HF

Metric ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC
Rule
61.19 77.28 67.80 63.08 72.44 68.35 52.13 59.95 61.81 63.08 75.44 71.35 67.65 76.15 65.70
Based[125]
KNN[118] 87.56 92.20 79.66 92.47 95.71 80.69 93.70 88.62 92.01 85.78 83.60 85.40 79.56 86.23 72.38
SVM[118] 78.70 88.08 50.00 86.84 92.95 50.00 89.54 79.52 84.46 78.39 75.84 78.24 72.42 84.00 50.00
DT[118] 78.70 88.08 50.00 86.84 92.95 50.00 89.54 79.52 84.46 78.39 75.84 78.24 72.42 84.00 50.00
Elliptical
71.54 81.94 57.39 78.89 87.85 53.64 67.35 41.28 59.33 61.76 56.54 61.19 62.54 74.14 53.10
Envelope[126]
Xception[121] 89.30 90.28 85.80 89.54 90.44 85.35 81.03 70.95 77.81 88.54 90.44 85.35 85.09 88.15 84.70
Proposed
95.64 96.12 97.71 96.71 97.68 97.71 91.43 91.48 93.59 94.02 94.98 95.02 94.82 95.22 95.31
Method
4.3.5 Comparison with the State-of-the-Art Models

The evaluation results showed that the proposed 1D CNN model outperformed other
deep learning-based models for HR and all HRV features. Thus, we compared the
1D CNN model with various types of PPG quality assessment methods, includ-
ing a rule-based method, several supervised machine-learning models, an unsuper-
vised machine-learning model, and a deep learning-based model. For the rule-based
method, we selected the hierarchical rule-based model introduced in [125]. For su-
pervised machine-learning methods, we selected SVM, KNN, and decision tree as
described in [116]. The elliptical envelope was chosen as an unsupervised learning
model [126]. We also selected Xception [121] as a deep learning-based model.
The performance of our 1D method in comparison with the selected state-of-
the-art PPG quality assessment method is shown in Table 6. For each HRV param-
eter, the corresponding label was used in the classification of the PPG signals and
evaluation of the performance of the methods. As shown in Table 6, the proposed
1D CNN method, in general, had better performance than the other methods had.
KNN had slightly better accuracy for RMSSD than the proposed 1D CNN method
had. However, for other performance metrics and for other HRV parameters, the
proposed 1D CNN method performed better. Moreover, Xception had better over-
all performance than the other selected methods had. After Xception, supervised
machine-learning methods had good performance. KNN has better performance for
AVNN and RMSSD than Xception had. Unsupervised elliptic envelope and rule-
based methods had the lowest performance.
42
5 IoT-Based Maternal Monitoring During
Pregnancy and Postpartum: Analyzing
the Trends of HR and HRV Parameters
Chapter 3 presented a long-term IoT-based maternal monitoring for remote health

monitoring during pregnancy and postpartum. Chapter 4 investigated the accuracy
of collected HR and HRV parameters against the ECG gold standard in everyday life
settings. Moreover, Chapter 4 provided a quality assessment method to improve the
quality of collected HR and HRV parameters by discarding noisy parts of the data.
This chapter uses the presented IoT-based system to monitor physiological pa-
rameters during pregnancy and postpartum. Therefore, the system was used to col-
lect data from pregnant women. The data were then used to investigate HR and HRV
changes during pregnancy and postpartum. Hence, this chapter aims to provide a
better understanding of autonomous nervous system alteration during pregnancy and
postpartum. Here, we first describe our data collection using our IoT-based maternal
monitoring system, where physiological data of 62 pregnant women were continu-
ously collected during pregnancy and postpartum. Then, using the collected data,
we investigate the trends in nighttime HR and HRV during the second and third
trimesters and the postpartum period.
5.1 Data Collection

The data were collected by the IoT system introduced in Chapter 3. The partici-
pants used a smartwatch and our customized cross-platform mobile application dur-
ing the study. They also used a blood pressure device to measure their blood pressure
once a week. The smartwatch collected PPG signals and sleep and activity parame-
ters. Moreover, the mobile application collected responses to our questionnaires and
background information.
In this chapter, we used the nighttime PPG signal collected by the smartwatch
from the participants to monitor HRV during pregnancy and postpartum. The PPG
signals collected from gestational week 16 until 3 months postpartum were used in
this chapter. Moreover, we used background information such as BMI, age, edu-
cation level, and delivery- and infant-related information collected by the mobile
application in our analysis.
43
Fatemeh Sarhaddi
5.1.1 Participants and Setup

Two groups of pregnant women (high-risk and low-risk) were recruited in this study.
The recruiting was performed in Southwest Finland through advertisements in ma-
ternity clinics and on social media. The inclusion criteria were as follow: (a) a
singleton pregnancy, (b) ability to understand the Finnish language, (c) possession of
an android or iOS smartphone, (d) 12–15 gestational weeks of pregnancy, (e) age 18
years or older, and (f) willingness to use the provided smartwatch and our customized
mobile application from recruitment to 3 months postpartum.
The high-risk pregnancy group had histories of preterm births (gestational weeks
22–36) or late miscarriages (gestational weeks 12–21) as described in Section 3.4.
The low-risk pregnancy group had no previous miscarriages or preterm births and
had histories of full-term births (gestational weeks 37–42).
Interested participants contacted the researcher via email and arranged a face-
to-face meeting. In the meeting, the researcher explained the procedures and de-
tails of the study. After obtaining written informed consent from the participant, a
smartwatch and a blood pressure device was delivered to the participant. Moreover,
our mobile application was installed on her smartphone. In the high-risk pregnancy
group, 32 pregnant women were recruited from January to December 2019, but four
participants withdrew from the study, resulting in 28 participants in this group. In
the low-risk pregnancy group, 30 participants were recruited from October 2019 to
March 2020.
To investigate the trend of HR and HRV, we used the PPG signals collected from
our participants. We combined the data from the low-risk pregnancy group with
the high-risk pregnancy group because no significant difference existed between the
HRV trends of the two groups. Therefore, the data of 58 pregnant women were
included in our analysis.
5.2 HR and HRV Trends During Pregnancy and Post-

partum
Studies have shown that HR increases and HRV parameters normally decrease dur-
ing pregnancy [21; 23; 24; 25]. At the same time, variations in HRV parameters
can be a sign of physical or mental health complications during pregnancy. For ex-
ample, higher LF/HF in early pregnancy compared with a normal pregnancy may
indicate hypertension [140]. Pregnant women with pre-eclampsia have lower LF and
higher LF/HF than normal pregnant women have [141; 142]. Moreover, women with
normal pregnancies have higher time-domain parameters compared with depressed
pregnant women [143]. Anxiety or induced stress during pregnancy could also cause
a decrease in HF [144]. Moreover, mindful pregnant women who can manage their
stress experience less reduction in RMSSD and HF during pregnancy [145]. There-
44
IoT-Based Maternal Monitoring During Pregnancy and Postpartum: Analyzing the Trends of HR
and HRV Parameters
fore, it is necessary to investigate normal HRV variation during pregnancy to distin-

guish normal changes from complication-related changes in HRV. This may help in
the early detection of health issues during pregnancy and, subsequently, the improve-
ment of the mother and fetus’s health. It also should be noted that HRV interpretation
is challenging because HRV behavior differs between people [40].
Many studies have investigated HRV during pregnancy. Table 7 summarizes
works on HRV trends during pregnancy and postpartum. As shown in Table 7, previ-
ous works used ECG signals (see Section 2.2.1) to measure HRV. Most of the studies
measured HRV once from pregnant women at several pregnancy weeks or during
labor, and they have one measurement from nonpregnant women [146; 147; 148;
149; 150]. These studies evaluated HRV changes by comparing HRV from differ-
ent women. Although these studies provided normal ranges of HRV parameters and
their changes during pregnancy, they could not consider individual variation in HRV
parameters [40].
Table 7. Summary of research on HR and HRv trends during pregnancy and postpartum
Duration of study and

Study Results Number of participants
frequency of recording
One recording during the 27 women with normal
No significant changes
Voss 2000 [149] second half of pregnancy pregnancy
in HRV parameters
(Gestational week 21-40) 14 nonpregnant women
One time before pregnancy,
HRV declined during 8 pregnant
Stein 1999 [151] and 4 time during pregnancy
pregnancy 12 nonpregnant women
(week 6, 10, 18, and 34)
AVNN: decreased significantly Once per month
Baumert 2012 [152] 32 pregnant women
SDNN: no significant changes during weeks 20 - 40
SDNN and HF: decreased
Once per trimester
significantly
Garg 2020 [153] (gestational weeks 66 pregnant women
LF and LF/HF: increased
11-13, 20-22, and 30-32)
significantly
16 pregnant women
HR and LF/HF: increased Once every four weeks during
Walther 2005 [154] with chronic hypertension
HF: decreased the second half of pregnancy
35 healthy pregnant women
HR: slightly increased Once per gestational weeks
Moertl 2009 [155] Frequency-domain HRV: 10-13, 15-18, 20-22, and >30 20 healthy pregnant woman
no significant changes (During pregnancy)
During pregnancy
LF, HF, SDNN and RMSSD:
Once at 1st trimester
decreased significantly 30 pregnant women
Gandhi 2014 [156] (weeks 6-12),
HR and LF/HF:increased 30 non-pregnant women
and one at 3rd trimester
significantly
(Weeks 25-36)
HR: No significant changes
LF and HF: significantly lower Once at 3rd trimester
Heiskanen 2008 [25] 28 pregnant women
during the 3rd trimester and once at postpartum
compared to postpartum
Other studies have few ECG recordings from the same pregnant women ranging
from per month to per trimester [25; 154; 155; 152; 153; 156]. These studies used
short-term ECG recordings ranging from 10 to 30 minutes while resting in laboratory
settings. The studies are limited because they have few short-term ECG recordings
45
Fatemeh Sarhaddi
Normalized
Normalized
HRV
IBIs
parameters
Reliable Statistical
Collected PPG quality Peak Peaks & analysis
PPG
PPG signals assessment detection IBIs
signals
HR and HRV
parameters
Figure 20. HR and HRV parameters analysis pipeline [157]
from pregnant women that could not accurately reflect the changes during pregnancy.
Moreover, the recordings are restricted to predefined positions in laboratory settings.
Only Stein et al. [151] performed four 24-hour ECG recordings during pregnancy in
everyday life settings. However, this work still lacks continuous HRV measurement.
Continuous HRV measurement during pregnancy and the postpartum period provides
accurate and reliable information about HRV trends and patterns of HRV changes.
Therefore, in this case study, we performed a continuous HRV measurement starting
from the second trimester to 3 months postpartum in everyday life settings. To the
best of our knowledge, this is the first study that has continuously measured PPG
signals from pregnant women to investigate HRV trends. In the following sections,
we will explain the method and the results of collecting and analyzing the trend of
HRV parameters collected continuously during pregnancy and postpartum.
5.2.1 HR and HRV Parameters Analysis

An overview of the HR and HRV parameters analysis pipeline, which was imple-
mented in our cloud server, is illustrated in Figure 20. As shown in the figure, we
first used a PPG quality assessment method to extract reliable signals from the col-
lected data. Then, we used the peak detection method to detect peaks and extract
IBIs. We then normalized the extracted IBIs. In the following step, we used IBIs and
normalized IBI to extract HR and HRV and normalized HRV parameters. Finally, we
used statistical analysis (i.e., HLM models) to investigate the trend of HR and HRV
and normalized HRV parameters during pregnancy and the postpartum period.
Reliable Signal Extraction
As mentioned earlier, PPG is a convenient, noninvasive optical method to measure

HR and HRV parameters. However, PPG is prone to noise such as ambient light or
motion artifacts when participants engage in their physical activity. The inevitable
noise affects the quality of the signals negatively [106]. The low-quality PPG sig-
46
and HRV Parameters
nals could result in inaccurate extracted parameters (e.g., HRV parameters [106]).
Therefore, we employed a PPG quality assessment method to identify and discard
unreliable segments of PPG signals. For this purpose, an SVM classifier was utilized
to differentiate between reliable and unreliable signals. The classifier was trained
using morphological features extracted from PPG signals, including skewness, kur-
tosis, approximate entropy, Shannon entropy, and spectral entropy, as described in
[126]. Subsequently, we applied this SVM classifier to detect and discard unreliable
5-minute segments from the collected PPG signals.
Peak Detection and IBI Extraction
In this step, we first enhanced the reliable PPG signals by filtering the noise that was
outside the human HR range. Therefore, we used a bandpass filter with 0.7 Hz and
3.5 Hz cut-off frequencies. Then, we used a moving average-based peak detection
method as described in [158]. The method included an adaptive threshold based on
the morphology and amplitude of PPG signals. After detecting the peaks, the IBIs
were calculated as the interval between two consecutive peaks. In the error detection
phase, too-large or too-small IBIs compared to the average IBIs of the segments were
discarded. The method was implemented in the HeartPy library in Python [159].
Parameter Normalization
Changes in HRV parameters are caused by variation in HR or alteration in average

HR [160]. Therefore, we must remove the effect of normal HR increase during
pregnancy [151; 21] on HRV parameters to investigate the trend of HR variation
during pregnancy. For this purpose, we computed normalized IBIs by dividing the
IBIs by the average IBIs in each 5-minute window of reliable signals [161].
HR and HRV Extraction
We used short-term HRV analysis to extract HR, AVNN, RMSSD, SDNN, LF, HF,
and LF/HF from IBIs as described in Section 2.3. The HRV parameters were cho-
sen because a previous study [51] showed that they can be reliably extracted at the
frequency at which we collected PPG signals (e.g., 20 Hz). Moreover, we extracted
nRMSSD, nSDNN, nLF, nHF, and nLF/nHF, which represent normalized RMSSD,
normalized SDNN, normalized LF, normalized HF, and normalized LF to normal-
ized HF ratio from normalized IBIs. Table 8 shows the normalized HRV parameters
and their definitions.
47
Fatemeh Sarhaddi
Table 8. Normalized HRV parameters definitions [157]
Parameters Unit Description

nSDNN ms Standard deviation of normalized IBIs
Time-domain
Square root of the mean of the sum of the squares of
nRMSSD ms
differences between adjacent normalized IBIs
Power in low-frequency range from normalized IBIs
nLF ms2
Frequency- (0.04- 0.15 Hz)
domain Power in high-frequency range from normalized IBIs
nHF ms2
(0.15- 0.4 Hz)
nLF/nHF - Ratio of nLF to nHF
Statistical Analysis
We used the HLM model to investigate the HR and HRV trends during pregnancy and
the postpartum period. The HLM model is a multilevel statistical analysis method
that is used widely in longitudinal analysis. The model can analyze the trends in
between- and within-person variances while considering the correlation of data in
measurements from each individual [162].
We investigated the HR and HRV trends in three time periods; that is, the second
trimester (16–28 weeks of gestation), the third trimester (29–40 weeks of gestation),
and the postpartum period (12 weeks after delivery). Then, we compared the trends
in each period with other periods. In our analysis, we used days as the independent
variable and added age, education level, and BMI before pregnancy as controlling
factors.
We included the second trimester and the postpartum data from all participants
in our analysis. However, we removed seven participants who had preterm births
from the third trimester data. Therefore, the data of 51 participants were included in
the third trimester analysis. The data were analyzed using the statsmodels library in
Python [104].
5.2.2 Results and Discussion

Figures 21 and 22 represent the weekly mean of HR, time-domain, and frequency-
domain HRV parameters averaged from all participants during pregnancy and the
3-month postpartum period. In addition, trends (slope), and average initial value
(intercept) of HR and HRV parameters are shown in Table 9 and Table 10.
As indicated in Table 9 and Table 10, during the second trimester, HR increased
significantly, whereas the AVNN, SDNN, nSDNN, RMSSD, nRMSSD LF, HF, and
nHF decreased significantly. In addition, nLF/nHF increased significantly during the
second trimester. During the third trimester, HR decreased significantly, whereas
AVNN, SDNN, nSDNN, RMSSD, and HF increased significantly.
48
and HRV Parameters
Pregnancy Postpartum period

75
70
HR
65
60
55
1100
1050
1000
AVNN
950
900
850
800
75
70
65
SDNN
60
55
50
8.5
8
7.5
nSDNN
7
6.5
6
5.5
80
70
RMSSD
60
50
8.5
8
nRMSSD
7.5
7
6.5
6
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 1 2 3 4 5 6 7 8 9 10 11 12
Pregnancy week Postpartum week
Figure 21. Weekly mean and 95% CI of HR and time-domain HRV parameters during pregnancy
and the postpartum period. (The vertical line separates the second and third trimesters) [157]
During pregnancy, beginning from gestational week 16, HLM showed that HR
increased significantly, whereas HRV parameters AVNN, SDNN, RMSSD, LF, and
HF decreased significantly. However, analysis shows that starting from gestational
week 35, HR began to decrease, and the mentioned HRV parameters started to in-
49
Fatemeh Sarhaddi
Pregnancy Postpartum period

1750
1500
1250
LF
1000
750
500
15
10
nLF
2500
2000
HF
1500
1000
8
7
6
nHF
5
4
3
1.2
1
LF/HF
0.8
0.6
6
5
4
nLF/nHF
3
2
1
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 1 2 3 4 5 6 7 8 9 10 11 12
Pregnancy week Postpartum week
Figure 22. Weekly mean and 95% CI of frequency-domain HRV parameters during pregnancy
and the postpartum period. (The vertical line separates the second and third trimesters) [157]
crease. However, HR and HRV parameters did not reach their initial values (i.e., the
value at pregnancy week 16). In the postpartum period, most parameters were sta-
ble except nRMSSD, which decreased slightly, and LF/HF, which increased slightly.
Previous works have also shown an increase in HR during pregnancy [21; 154] and
50
and HRV Parameters
Table 9. Trends of HR and time-domain HRV parameters during the second trimester, the third
trimester, and the postpartum period [157]
HR AVNN SDNN nSDNN RMSSD nRMSSD

Intercept 62.736 916.443 84.023 9.079 89.293 9.986
Second
(P-value) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001)
trimester
Slope 0.045 -0.585 -0.082 -0.006 -0.103 -0.007
(P-value) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001)
Intercept 81.324 708.181 69.787 9.237 70.275 8.366
Third
(P-value) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001)
trimester
Slope -0.025 0.345 0.084 0.007 0.071 0.006
(P-value) (0.006) (0.002) (P<.001) (P<.001) (P<.001) (0.971)
Intercept 47.237 1216.255 115.321 9.481 135.233 11.620
Postpartum
(P-value) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001)
period
Slope -0.009 0.130 0.001 -0.001 -0.037 -0.004
(P-value) (0.366) (0.464) (0.954) (0.689) (0.115) (0.012)
a decrease in HRV parameters during the course of pregnancy [23; 24; 25].
Table 10. Trends of frequency-domain HRV parameters during the second trimester, the third
trimester, and the postpartum period [157]
LF nLF HF nHF LF/HF nLF/nHF

Intercept 1045.893 2.422 2990.343 4.727 -0.248 0.565
Second
(P-value) (P<.001) (P<.001) (P<.001) (P<.001) (P<.001) (0.002)
trimester
Slope -3.109 0.001 -4.224 -0.007 -0.061 0.001
(P-value) (P<.001) (0.718) (P<.001) (0.003) (0.103) (0.007)
Intercept 1008.979 4.248 1743.362 6.166 0.67 0.995
Third
(P-value) (0.026) (0.001) (0.010) (0.001) (0.839) (0.003)
trimester
Slope 0.424 0.006 2.767 0.004 -0.097 0.001
(P-value) (0.443) (0.093) (0.007) (0.209) (0.113) (0.276)
Intercept 2124.653 2.486 4969.652 6.670 -0.397 0.416
Postpartum
(P-value) (0.002) (0.087) (P<.001) (0.001) (0.269) (0.022)
period
Slope 2.415 -0.005 -0.682 -0.001 0.234 -0.002
(P-value) (0.053) (0.402) (0.626) (0.769) (0.004) (0.050)
Association Between BMI, Age, and Education with HRV During Pregnancy
and Postpartum
We also investigated the association between BMI before pregnancy, age, and edu-
cation level with the HRV trends during pregnancy and the postpartum periods. We
added these factors as independent variables in our HLM models. Our models indi-
cated that BMI before pregnancy is not significantly associated with changes in HRV
during pregnancy and the postpartum period. Moreover, the results showed that age
is negatively associated with nSDNN, nRMSSD, and HF in the second trimester.
However, in the third trimester and in the postpartum period, age is positively asso-
51
Fatemeh Sarhaddi
ciated with SDNN, nSDNN, RMSSD, nRMSSD, HF, and LF/HF. Studies have also
illustrated that HRV normally decreases when people age [40].
Our results showed a positive correlation between education level and SDNN,
nSDNN, RMSSD, nRMSSD, and HF during the third trimester. However, education
level is not significantly associated with HRV parameters during the second trimester
or the postpartum period. Lower HRV levels are associated with higher stress lev-
els [42], and these results may indicate lower stress levels in pregnant women with
higher education levels. Cardwell [163] showed that low education level is a predic-
tor of higher stress during pregnancy. Other studies have also shown that people with
higher education levels experience lower stress levels compared with less educated
people in the same stressful situation [164].
52
6 Maternal Loneliness Detection Based
on Passive Data Sensing Using Our
IoT-Based Maternal Monitoring System
In this chapter, we focus on loneliness as an important mental health issue during

pregnancy. We develop two predictive models for detecting maternal loneliness us-
ing data collected passively by the IoT-based system introduced in Chapter 3. We
used the data collected in Chapter 5 from pregnant women during pregnancy and
postpartum to train and test the developed model.
6.1 Maternal Loneliness

Loneliness is a negative subjective feeling related to the perception of being alone
or isolated or having deficient meaningful relationships [165]. Social loneliness and
emotional loneliness are two kinds of loneliness. Social loneliness refers to the lack
of desired network or social relationships, and emotional loneliness refers to the lack
of intimate and close relationships [166; 167]. Loneliness is a global health issue
that considerable parts of the populations in many countries face [168], and it is
increasing, especially in the era of social isolation due to the COVID-19 pandemic
[169; 170; 171; 172].
Several studies in the literature have investigated the health-related effects of
loneliness in different population-based groups. Studies have indicated that loneli-
ness is associated with psychiatric and physiological disorders such as depression
[173], anxiety [173], alcohol abuse [174], Alzheimer’s disease and cognitive impair-
ment [175], sleep problems [176], hypertension [177], autoimmune disorders [175],
obesity [178], and an increased rate of physiological deterioration. Loneliness dur-
ing adolescence is associated with a worse health condition and is worse in females
compared with males [179]. Loneliness can also increase the risk of mortality [180].
Loneliness during pregnancy and the postpartum period is also correlated with
various health problems for both the mother and her child. Many studies in different
countries showed that maternal loneliness during pregnancy is associated with de-
pression [181; 182; 183; 184; 185; 172]. Other studies have shown the correlation
between loneliness and postpartum depression [186; 187]. Moreover, several studies
during the COVID-19 pandemic showed the correlation of maternal loneliness with
53
Fatemeh Sarhaddi
anxiety [188], cognitive distortion [183], psychological distress [189], and increase
of perceived stress [181]. Maternal loneliness also has adverse effects on newborn
babies, such as increasing the risk of respiratory tract infection [190]. Hence, the
early detection of maternal loneliness helps to improve the health and well-being of
the mother and her child by preventing the associated health issues through interven-
tion.
Conventionally, loneliness was investigated by self-reported questionnaires [172;
182; 181; 183] and interviews [191]. The subjective standard questionnaires or in-
terviews examined the association between loneliness and health problems such as
depression, anxiety, lack of social support, cognitive distortion, and stress. However,
these methods are not able to predict loneliness in individuals.
In addition, few studies have taken advantage of the rapid growth of IoT-based
systems and wearable devices and used passive sensing for loneliness prediction
[192; 193]. Wu et al. [193], investigated the association of momentary loneliness
with location and Bluetooth data collected from college students. In another study,
Doryab et al. [192] used sleep, activity, location, screen time, call logs, and Blue-
tooth information collected by activity trackers and smartphones of college students
for loneliness prediction. These studies were also restricted because they only con-
sidered college students on a campus.
To the best of our knowledge, no study in the literature has predicted mater-
nal loneliness based on objective data collected passively. This case study presents
two machine-learning models to predict maternal loneliness using passively collected
data. Using such models can help to improve the well-being of the mother and her
child with low cost and minimal engagement from mothers. We used a smartwatch
to collect objective physiological parameters (HR, HRV, physical activity, and sleep)
passively because previous studies have shown the association of these parameters
with loneliness [194; 195; 196; 197]. We developed decision tree and gradient boost-
ing models to predict maternal loneliness using the passively collected data. Finally,
we investigated the importance of physiological health parameters in each model.
6.2 Setup and dataset Preparation

In this section, we first briefly explain our participants and data collection. Then,
we describe our data analysis pipeline, which was used to extract features and create
datasets for this study.
6.2.1 Participants
For this study, we used the data collected from pregnant participants described in
Section 5.1.1. As mentioned, 62 pregnant women were recruited in total. However,
we only included participants in our loneliness prediction case study if they had an
54
Maternal Loneliness Detection Based on Passive Data Sensing Using Our IoT-Based Maternal
Monitoring System
acceptable amount of missing data, resulting in 31 participants in this study.
6.2.2 Data Collection

We collected data using the IoT-based maternal monitoring system introduced in
Chapter 3. As mentioned in Chapter 3 and Chapter 5, we used a smartwatch and a
cross-platform mobile application for data collection. The smartwatch continuously
collected HR, HRV, sleep, and physical activity data during pregnancy and 3 months
postpartum. In addition, we used the University of California, Los Angeles (UCLA)
standard questionnaire to evaluate loneliness. The 12-item Revised UCLA ques-
tionnaire was provided to our participants using our customized mobile application.
The revised UCLA questionnaire includes six questions regarding social loneliness
and six questions regarding emotional loneliness [198]. Participants responded to
this questionnaire at week 36 during pregnancy and at 12 weeks after delivery. The
background information was also collected by our mobile application.
6.2.3 Data Analysis

We used the data collected by the smartwatch to generate seven datasets. Each data
record in our datasets contained the data from 7 days before and the day of respond-
ing to the UCLA questionnaire. Seven days was chosen because the UCLA instru-
ment examines a person’s feelings during the prior week. We only considered valid
days in our analysis, defined as having sleep data and having worn the smartwatch
for at least 10 hours while awake. Moreover, data records with less than 4 valid days
were discarded due to missing data. Our analysis of the data records comprised the
following steps:
1. Feature extraction
2. Labeling
3. dataset generation
4. Missing data imputation
5. Machine-learning models development
6. Important features investigation
Figure 23 summarizes these steps. We describe these steps in the following subsec-
tions.
55
Fatemeh Sarhaddi
Sleep Dataset
HRV Dataset
HRV features Physical activity

Dataset
Sleep features Sleep & HRV
8 days data Dataset
(with at least 4
Physical activity
valid days) Sleep & physical
features
activity Dataset
HRV & physical

activity Dataset
Labels
Sleep & HRV &
physical activity
Dataset
Machine learning
models
Decision tree Important
Missing data
classifer features
imputation
Gradient boosting
investigation
classifier
Figure 23. Data analysis pipeline [199]
6.2.4 Feature Extraction
We used the data collected by the smartwatch to extract resting HR and HRV, sleep,
and physical activity features. Here, we briefly explain the extracted features:
HR and HRV Features
We used the same pipeline described in Section 5.2.1 to extract HR and HRV param-
eters. We used resting HR at night and corresponding HRV parameters (i.e., AVNN,
SDNN, RMSSD, LF, HF, and LF/HF) for our analysis.
56
Monitoring System
Sleep Features
We extracted total sleep time (TST), wake after sleep onset (WASO), sleep fragmen-
tation, the average intensity of hand movement during sleep, a sleep quality indicator,
and a sufficient sleep parameter using the collected data from the watch. We defined
the sleep quality indicator as being awake for less than 20 minutes during the night
sleep and the sufficient sleep parameter as having TST between 7 and 8.5 hours dur-
ing the night [200]. Other sleep parameters were defined and calculated as described
in [75].
Physical Activity Features
The smartwatch continuously collected physical activity parameters, including step

counts, walking steps, running steps, distance, activity duration, and activity intensity
per 10 minutes. We accumulated the data to extract daily values. We also extracted
statistical parameters (i.e., mean, minimum, median, maximum, SD, interquartile
range, range, skewness, kurtosis, and root mean square) from the distribution of
hourly step counts and hourly activity duration. We defined a sufficient activity in-
dicator that indicates more than 7000 step counts during the day. Moreover, we
calculated sedentary time by subtracting active time from time spent awake.
The summary of extracted features in this study is illustrated in Table 11.
Table 11. Summery of extracted features [199]
HR and HRV features Sleep features Physical activity features

HR TST Step counts
AVNN Sleep fragmentation Walking steps
RMSSD WASO Running steps
SDNN Average hand movement Distance
LF Sleep quality indicator Activity duration
HF Sufficient sleep parameter Activity intensity
LF/HF Sedentary time
Sufficient activity indicator
Statistical features
from distribution of
daily steps
Statistical features
from distribution
of daily activity duration
57
Fatemeh Sarhaddi
6.2.5 Labeling
UCLA questionnaire scores were used as labels. The questionnaire has social and
emotional scores ranging from 0 to 24. The emotional scores were ignored because
we had a few participants with emotional loneliness, and only social scores were
used. UCLA social score ≥ 12 is considered loneliness, and UCLA social score
< 12 non-loneliness [198].
6.2.6 datasets
The extracted data from each data record, along with the corresponding label, were
used to generate seven datasets. The datasets include different combinations of sleep,
HRV, and physical activity features, which result in the sleep dataset, HRV dataset,
physical activity dataset, sleep and HRV dataset, sleep and physical activity dataset,
HRV and physical activity dataset, and sleep, HRV, and physical activity dataset.
The datasets were used for training and testing the developed machine-learning clas-
sifiers.
6.2.7 Missing Data Imputation

The data record in our datasets has at least four valid days. Therefore, the data record
may have missing data. We used the average value of the missing features within the
data record to fill in the missing values.
6.3 Predictive Machine-Learning Models

Several machine-learning models have been developed for maternal loneliness pre-
diction. The decision tree and gradient boosting models were chosen for prediction
because they have better performance than other models. We briefly explain the
decision tree and gradient boosting models in the following.
6.3.1 Decision Tree

The decision tree classifier is a supervised model that stands out because of its sim-
plicity and intelligibility. This model is a robust model well suited for complex data
[201; 202]. The model has a tree-like structure, including internal nodes and leaves.
Internal nodes represent a feature, and a split rule based on that feature and leaves
show the class label. The features in internal nodes are selected using the Gini index,
which shows the purity of classification.
58
Monitoring System
6.3.2 Gradient Boosting

The gradient boosting method is a machine-learning method that performs well on
complex data with high cardinality. The model provides speedy and accurate pre-
diction and usually outperforms other conventional machine-learning models [203].
Gradient boosting is an ensemble of weak prediction models that tries to build sev-
eral models sequentially, and the added models in each step eliminate the error of
previous models [203]. This model can find nonlinear relationships. Moreover, it
performs well in loneliness prediction studies [192].
6.4 Model Evaluation

We first employed the recursive feature elimination (RFE) method on physical activ-
ity features to remove the least important feature and prevent overfitting. Then, we
used the leave-one-participant-out cross-validation method to evaluate the developed
methods and report the average performance. The predictive models, RFE method,
training, testing, and evaluation were implemented in Python using the scikit-learn
library [204]. The predictive models were evaluated based on the following mea-
sures:
• Precision: percentage of predicted samples that actually belonged to a class
• Recall: percentage of correctly predicted samples per class
• F1 score: harmonic mean of precision and recall per class
• Weighted F1 score: weighted average of F1 scores
• Sensitivity: percentage of lonely participants correctly detected
• Specificity: percentage of non-lonely participants correctly detected
6.5 Results and Discussion

This section evaluates the performance of machine-learning models for maternal
loneliness prediction. Then, the important features of each model are discussed.
6.5.1 Performance of Maternal Loneliness Prediction

The performance of decision tree and gradient boosting classifiers based on preci-
sion, recall, F1 score, weighted F1 score, sensitivity, and specificity are summarized
in Table 12. Moreover, Figure 24 represents the performance of models in terms of
F1 scores.
For the decision tree model, the best performance was reached in the physical ac-
tivity dataset, HRV and physical activity dataset, physical activity and sleep features
dataset, and all features dataset. The results showed that physical activity features
are the most important features for loneliness prediction in decision tree models. In
59
Fatemeh Sarhaddi
Table 12. Per class precision, recall and F1 score, weighted F1 score, sensitivity, and specificity
performance measures for the decision tree and gradient boosting models [199]
Precision Recall F1 score Weighted

Model Dataset Sensitivity Specificity
Non- Non- Non- F1 score
Loneliness Loneliness Loneliness
lonely lonely lonely
Sleep 0.778 0.619 0.636 0.765 0.7 0.684 0.693 0.765 0.636
HRV 0.6 0.528 0.614 0.514 0.607 0.521 0.567 0.513 0.614
PA 0 .905 0.833 0.864 0.882 0.884 0.857 0.872 0.882 0.864
Decision
Sleep & HRV 0.7 0.579 0.636 0.647 0.667 0.611 0.642 0.647 0.636
Tree
Sleep & PA 0 .905 0.833 0.864 0.882 0.884 0.857 0.872 0.882 0.864
HRV & PA 0.864 0.824 0.864 0.824 0.864 0.824 0.846 0.824 0.864
All 0.864 0.824 0.864 0.824 0.864 0.824 0.846 0.824 0.864
Sleep 0.611 0.476 0.5 0.588 0.55 0.526 0.540 0.588 0.5
HRV 0.553 0.471 0.591 0.432 0.571 0.451 0.516 0.432 0.591
PA 0.87 0.875 0.909 0.824 0.889 0.848 0.871 0.823 0.909
Gradient
Sleep & HRV 0.667 0.524 0.545 0.647 0.6 0.579 0.591 0.647 0.545
Boosting
Sleep & PA 0.833 0.867 0.909 0.765 0.87 0.812 0.845 0.765 0.909
HRV & PA 0.909 0.882 0.909 0.882 0.909 0.882 0.897 0.882 0.909
All 0.87 0.875 0.909 0.824 0.889 0.848 0.871 0.823 0.909
addition, the high specificity of the decision tree on the datasets with high perfor-
mance indicate that the model can identify non-lonely individual with high accuracy.
Moreover, the high sensitivity of the decision tree for physical activity or all features
datasets demonstrates that the decision tree performs well in accurately detecting
individuals experiencing loneliness.
Similar to the decision tree, the gradient boosting model performed best on the
datasets containing physical activity features. The highest performance of the model
was achieved on the physical activity and HRV dataset. For gradient boosting, adding
physical activity features to the dataset improved the results. Adding HRV features
also resulted in improvement, though less than with physical activity. However,
adding sleep features to physical activity features resulted in a decrease of perfor-
mance. Consequently, this model highlights the importance of physical activity and
HRV features for detecting maternal loneliness. In addition, the gradient boosting
model can identify lonely individuals with an accuracy higher than 88% for HRV
and physical activity features, while the accuracy of the model for identifying non-
lonely individuals is more than 90% for all the datasets containing physical activity
features. As you can see in Table 12, sleep features can have a negative impact on
the accuracy of detecting individuals experiencing loneliness, while having no effect
on detecting non-lonly individual.
The developed predictive models performed well in maternal loneliness predic-
tion. The gradient boosting and decision tree models achieved weighted F1 scores of
0.897 and 0.872, respectively. Additionally, both gradient boosting and decision tree
achieved the same sensitivity; therefore, both model had the same performance in de-
tecting participants experiencing loneliness. However, the gradient boosting model
achieved higher specificity, indicating its better performance in correctly detecting
non-lonely individuals. The results showed the feasibility and possibility of passive
sensing for maternal loneliness prediction.
60
Monitoring System
Decision Tree
non Lonely
Lonliness
0.8
0.6
F1 score
0.4
0.2
0.0
HRV & PA
PA
sleep & PA
HRV
sleep & HRV
All features
sleep
Datasets
Gradient boosting
non Lonely
Lonliness
0.8
0.6
F1 score
0.4
0.2
0.0
HRV & PA
PA
sleep & PA
HRV
sleep & HRV
All features
sleep
Datasets
Figure 24. Performance evaluation (F1 score) of decision tree and gradient boosting for different
datasets [199]
As mentioned, a few have works used passive sensing with wearable devices to
predict loneliness in college students. Wu et al. [193] collected self-report ques-
tions regarding loneliness along with location and Bluetooth data. However, our
models achieved better performance and used standard loneliness measures as la-
bels. Doryab et al. [192] also used activity tracker and smartphone data and pre-
sented machine-learning models for loneliness prediction in college students. They
achieved an accuracy of 80.2%. Our models for maternal loneliness prediction
achieved higher performance. In addition, their model needed sensitive information
such as phone numbers of close friends, whereas our models only used physiological
parameters collected by a smartwatch.
61
Fatemeh Sarhaddi
6.5.2 Important Features in Maternal Loneliness

In this section, the important features of loneliness detection in each predictive model
are discussed. For this purpose, the most important features in the datasets that
have weighted F1 scores ≥ 0.8 were investigated. Therefore, the datasets containing
physical activity features (four datasets) were considered for both models.
For the decision tree model, the most important features were the intensity of
activity and kurtosis of hourly steps per day. Other effective but less important fea-
tures include SDNN, LF, LF/HF, statistical features of hourly activity duration. For
the gradient boosting model, same as the decision tree, the most important features
were also the intensity of activity and kurtosis of the step count. Moreover, statistical
parameters of hourly step count and activity duration also contribute in loneliness
detection.
To summarize, the results show that the intensity and pattern of daily physical
activity during the day are the most important features of loneliness. Moreover, rest-
ing HR and HRV parameters are also correlated with loneliness. The results indicate
that more intensive physical activity is correlated with a lower level of loneliness. In
contrast, less intensive activity, when most of the activity happens in the morning or
early evening, along with low resting HRV, can be a sign of loneliness.
The finding of the negative association between physical activity and maternal
loneliness is very important for maternal health. It is known that physical activity
decrease while pregnancy proceeds due to physiological changes in the body [9].
Therefore, maternal care providers should be aware that lower physical activity could
also be a sign of maternal loneliness and be alert to other signs of loneliness.
In addition, Studies on college students have also shown the same results regard-
ing the negative correlation of loneliness with the duration of the activity, total step
counts, and activity duration in the evening [138; 192; 205]. Moreover, Ben-Zeev
et al. [206] reported that loneliness was not related to sleep duration, which is in
accordance with our results.
It is important to note that, considering the typical decrease in physical activity
during pregnancy, in order to generalize our models for the entire pregnancy or other
demographic groups, the models should also incorporate data from different popu-
lation groups. This is crucial, as the threshold for activity levels may vary among
different population groups.
62
7 Conclusion
IoT-based health monitoring systems provide new opportunities in health care ser-
vices by providing low-cost continuous health monitoring all times in everyday set-
tings. However, using such systems in practice raises many challenges. In this re-
search, we presented and evaluated an IoT-based system for maternal health mon-
itoring during pregnancy and postpartum. We also investigated and discussed the
requirements and challenges of deploying such maternal monitoring systems. Then,
we validated the HR and HRV parameters collected by the presented system. In ad-
dition, we proposed a deep learning-based quality assessment method to assess the
quality of HR and HRV parameters collected with our system.
We used our system to conduct a study on pregnant women and collected contin-
uous data from 62 women during pregnancy and 3 months postpartum. The collected
data were used to analyze the changes in HR and HRV parameters comprehensively
during pregnancy and postpartum. Finally, we used the collected data to develop
predictive machine-learning models for maternal loneliness as a major mental health
problem during pregnancy and postpartum.
To achieve Research Objective I (i.e., present, develop, and evaluate a continuous
IoT-based maternal health monitoring system), Chapter 3 presents a long-term IoT-
based maternal monitoring system. The system was evaluated against both system-
level and user-level requirements, including reliability, feasibility, usability, and en-
ergy efficiency. The practical challenges of developing the system were also dis-
cussed in detail.
Research Objective II (i.e., validate and assess the quality of HR and HRV pa-
rameters collected by the developed IoT-based monitoring system) was achieved in
Chapter 4. In Chapter 4, we first validated the accuracy of the smartwatch, which was
used in our monitoring system, against a clinical-level ECG gold standard. The vali-
dation was in terms of HR and HRV parameters and was performed during sleep and
waking hours. The results revealed the need for a PPG quality assessment method,
especially during waking hours, because the accuracy of most parameters was not
acceptable during waking hours. We then introduced a deep-learning-based quality
assessment method for HR and HRV parameters. Four CNN-based models for HR,
RMSSD, SDNN, and LF/HF were proposed. The models demonstrated superior per-
formance compared to existing state-of-the-art models on the data set collected from
46 participants in everyday settings.
63
Fatemeh Sarhaddi
Research Objective III (i.e., deploying the presented IoT-based system to collect
data from pregnant women during pregnancy and 3 months postpartum) was realized
in Chapter 5. In this chapter, we described our data collection from 62 pregnant
women. We collected data from two groups of pregnant women (i.e., high-risk and
low-risk groups). The high-risk group had a history of miscarriage or preterm birth,
whereas the low-risk group had a history of full-term birth. Our system continuously
collected the objective and subjective data from the participants for 9 months.
Research Objective IV (i.e., investigate the trends of HR and HRV parameters
during pregnancy and the postpartum period) was addressed in Chapter 5. In this
chapter, we employed the data collected during pregnancy and postpartum, which
was described in Research Objective II. We analyzed the trends of minimum HR and
corresponding HRV and normalized HRV parameters at nighttime using the hierar-
chical linear model. With this model, we analyzed the trend of parameters during
the second and third trimesters and the postpartum period. We then compared the
trend of changes in different trimesters and with the postpartum period. The effect
of age, BMI, and education level on the trend of HR and HRV parameters was also
investigated.
Research Objective V (i.e., develop predictive models to detect maternal loneli-
ness during pregnancy and the postpartum period based on data collected passively
by the developed monitoring system) was accomplished in Chapter 6. In this chap-
ter, we proposed two predictive models to detect maternal loneliness. The physio-
logical data collected passively through the IoT-based maternal monitoring system
developed in the thesis, along with loneliness scores, were used to train and test the
models. The results showed the high performance of the developed models in ma-
ternal loneliness detection. Moreover, the results indicated that intensity of activity,
activity in the evening, and resting HR and HRV are important features in predicting
loneliness.
In this thesis, we have contributed toward the above Research Objectives. How-
ever, there are still open directions to be explored. The IoT-based maternal moni-
toring system can collect a large amount of previously unavailable data. The data
can be used to predict mental and physiological health issues, including depression
and hypertension. Moreover, our current system lacks intervention functionalities.
The system should be extended to provide appropriate notifications, interventions,
and alarms in case of risk conditions. Furthermore, the system should be further de-
veloped to include other monitoring services, for example, diet, weight, and mobile
usage, and connect with other wearable and mobile applications that are currently
available.
Moreover, the accuracy validation of HR and HRV parameters was performed
using data from healthy nonpregnant adults in a 24-hour data collection. In the fu-
ture, we should consider validating the data with pregnant women experiencing var-
ious health conditions. During pregnancy, changes in body weight, hormonal levels,
64
Conclusion
and the growing fetus can influence sleep patterns and body movements, potentially
resulting in different motion artifacts. These variations may impact HR and HRV
parameters and need further investigation. Additionally, the proposed CNN-based
PPG quality assessment method was based on the same data as the accuracy vali-
dation. This method can also be extended, providing data from different population
groups with various health conditions. Moreover, other HRV parameters should be
considered, including frequency domain and nonlinear HRV parameters.
We analyzed the nighttime trends of HR and HRV parameters during the second
and third trimesters and the postpartum period. Analyzing daytime HR and HRV
trends is an open direction in this research. Moreover, as suggested in [151; 147],
HRV changes occur mostly during the first trimester. In the future, we will analyze
the trends of HR and HRV parameters during the first trimester and compare them to
other trimesters and the postpartum period.
We presented machine learning models for maternal loneliness detection. How-
ever, our data set was small, and we only considered data at two-time points during
late pregnancy and postpartum. In the future, we should use data during the whole
pregnancy with a greater number of participants to generalize our models. Moreover,
participants with different health conditions should be considered.
65
8 Overview of Original Publications
This chapter presents a brief overview of original publications and the author’s role
in each publication.
8.1 Long-Term IoT-Based Maternal Monitoring: Sys-

tem Design and Evaluation
In this paper, we present a long-term IoT-based maternal health monitoring system.
This system enables remote and continuous health monitoring of pregnant women
during pregnancy and the postpartum period. The system provides several monitor-
ing services, including stress, physical activity, and sleep, by collecting objective and
subjective physiological health data. The data are collected using wearable devices
and smartphones and transferred to our cloud server to be stored and analyzed. We
leveraged several machine learning-based methods to enhance the data analysis, such
as anomaly detection, personalized modeling, and quality assessment methods. The
presented system was implemented and used by pregnant women from the second
trimester to 3 months postpartum. The paper extensively investigates and discusses
the challenges of such a long-term maternal monitoring system.
Author’s contribution
The author is the first author of this publication. She had a significant role in de-
signing, developing, implementing, and deploying the maternal monitoring system.
She had a major role in collecting and analyzing the system’s usage data. She also
contributed to drafting the manuscript.
8.2 A comprehensive accuracy assessment of Sam-

sung smartwatch heart rate and heart rate variabil-
ity
In this paper, we investigated the accuracy of the HR and HRV parameters extracted
from PPG signals collected by the Samsung smartwatch in free-live settings. We uti-
lized short-term HRV analysis to assess the validity of HR, AVNN, RMSSD, SDNN,
pNN50, LF, HF, and LF/HF extracted from the collected PPG-signal in comparison
66
Overview of Original Publications
with a medical-grade ECG monitor during sleep time and awake time. We used the
24-hour continuous data from 28 participants in home-based monitoring for evalua-
tion. The accuracy of HR and HRV parameters obtained from the Samsung smart-
watch was investigated using the linear regression method, the Pearson correlation
coefficient, and the Bland–Altman plot.
The author is the joint first author of this publication. She has contributed to the
design of the study. She had a major role in the setup and data collection. She also
contributed to the data analysis. She had a significant role in drafting the manuscript.
8.3 A Deep Learning-based PPG Quality Assessment

Approach for Heart Rate and Heart Rate Variability
In this paper, we propose a deep learning-based quality assessment for PPG signals
based on the HR and HRV parameters. We first indicated that conventional PPG
quality assessment methods could be improved by considering the desired HRV pa-
rameters. The quality of each extracted HRV parameter is affected by different signal
features. Therefore, we proposed different CNN-based models for each HRV param-
eter. We evaluated the proposed method for HR, RMSSD, SDNN, and LF/HF pa-
rameters compared with the state-of-the-art PPG-quality assessment methods. The
evaluation was performed using the simultaneous PPG and ECG data collected in
everyday life settings. Moreover, we presented an automated annotation method for
labeling PPG segments as reliable/unreliable based on the ECG baseline. The pro-
posed quality assessment models are implemented as an open-source portable model
in Python and are available for others to use in their studies.
The author is the second author of this publication. She had a major role in data
collection and data analysis. She contributed to the study setup, model development,
literature review, and manuscript drafting.
8.4 Trends in Heart Rate and Heart Rate Variability

During Pregnancy and the 3-Month Postpartum Pe-
riod: Continuous Monitoring in a Free-living Con-
text
In this paper, we investigate the trends of HR and HRV during pregnancy and the
postpartum period. We collected PPG signals from 58 pregnant women using the
67
Fatemeh Sarhaddi
IoT-based maternal monitoring system presented in the first paper. The PPG signals
were acquired continuously during the second trimester, the third trimester, and three
months in the postpartum period in everyday life settings. We analyzed the minimum
HR and HRV parameters corresponding to min HR at night. We also investigate
the trends of normalized HRV parameters based on average HR. This normalization
enabled us to remove the impact of normal HR changes during pregnancy on HRV
trends. We utilized hierarchical linear modeling to investigate the trend of desired
parameters considering in-person and between-person differences. Moreover, we
investigate the effect of age, BMI before pregnancy, and education level on HRV
trends.
The author is the first author of this publication. She played a significant role in the
design of the study. She was the main contributor to the design, development, and
deployment of the long-term maternal monitoring system utilized for data collection
in this paper. Moreover, she had a significant role in data analysis and investigating
the parameters’ trends. She contributed to drafting the manuscript.
8.5 Maternal Social Loneliness Detection Using Pas-

sive Sensing Through Continuous Monitoring in Ev-
eryday Settings: Longitudinal Study
In this paper, we developed two machine learning models to predict maternal so-
cial loneliness using passive sensing. We used physiological data, i.e., HR, HRV,
physical activity, and sleep, collected passively by a smartwatch during pregnancy
and postpartum. We also collected the UCLA loneliness questionnaire in gestational
weeks 36 and 12 weeks after delivery, which was used to classify participants as
lonely or non-lonely. We utilized eight days of collected physiological data from
the smartwatch and the corresponding response to the UCLA questionnaire from 31
pregnant women. We leveraged these data for training and testing the decision tree
and gradient boosting models for loneliness prediction. These two models achieved
high F1 scores. Moreover, the results show that activity pattern, intensity of activity,
and resting nighttime HR and HRV parameters are important features of the models
to predict loneliness. This paper demonstrated the feasibility of maternal loneliness
prediction based on objective data and passive sensing. Therefore, it provides oppor-
tunities for maternal well-being improvement through early detection of loneliness.
68
The author is the first author of this publication. She was the major contributor
to the study design. She also has a significant role in developing machine learning
models and data analysis. Moreover, she contributed to the design, development, and
deployment of the long-term maternal monitoring system used for data collection in
this paper. She contributed to drafting the manuscript.
69
Appendix A: questions provided by the
mobile application
This appendix includes daily questions that were asked through our cross-platform
mobile application throughout pregnancy and postpartum. The questions were based
on the pregnancy week and day. The questions for day 7 for all pregnancy and
postpartum weeks are the same and are shown in Table 13 along with the answers.
Questions for other days of the week are yes/no questions and are determined based
on the specific pregnancy day and week, as shown in Table 14. Participants received
two or three questions available for that day.
Table 13. Day 7 questions
Questions Answers
Finnish English translation Finnish English Translation
Millaiseksi arvioisit
How would you rate 0 (Ei lainkaan stressiä) – 0 (No stress at all) -
stressitasosi
your stress level 100 (pahin mahdollinen 100 (Worst possible
viimeisen viikon
during the last week? stressi) stress)
aikana?
Millaiseksi arvioisit How would you rate
0 (Ei lainkaan) – 0 (Not at all)-
liikunnan määräsi your exercises level
100 (Erittäin paljon) 100 (Very much)
viimeisen viikon aikana? during the last week?
How would you assess
Millaiseksi arvioisit
the quality of 0 (Erittäin huono) – 0 (Very poor) -
unen laatusi
your sleep over 100 (Erittäin hyvä) 100 (Very good)
viimeisen viikon aikana?
the past week?
70
Table 14
Question in Finnish Question in English Weeks

Onko sinulla raskauteen liittyviä
DAY1 Do you have concerns related to pregnancy? gwk 12 to the delivery
huolia?
Onko raskautesi sujunut odotustesi Has your pregnancy progressed
DAY4 gwk 12 to the delivery
mukaisesti? as you expected?
DAY1 Onko sinulla ollut tänään pahoinvointia? Have you experienced nausea today? gwks 12-15
DAY2 Oletko oksentanut tänään? Have you vomited today? gwk 12–15
DAY3 Onko sinulla ollut tänään turvotusta? Have you experienced swelling today? gwks 12 and 14
Onko sinulla ollut tänään kipuja
DAY3 Have you had pain in the pelvic area today? gwks 13 and 15
lantion alueella?
DAY4 Onko sinulla ollut tänään närästystä? Have you had heartburn today? gwks 12 and 14
DAY4 Onko sinulla vauvaan liittyviä huolia? Do you have concerns related to the baby? gwks 13 and 15
DAY5 Onko sinulla ollut tänään suonenvetoja? Have you had cramps today? gwks 12 and 14
DAY5 Onko sinulla ollut tänään selkäkipuja? Have you had back pain today? gwks 13 and 15
DAY6 Oletko tänään onnellinen? Are you happy today? gwks 12 and 14
Oletko tänään tuntenut itsesi Have you felt depressed or
DAY6 gwks 13 and 15
masentuneeksi tai alakuloiseksi? downhearted today?
DAY1 Onko sinulla ollut tänään turvotusta? Have you experienced swelling today? gwk 16 to the delivery
DAY1 Oletko tänään onnellinen? Are you happy today? gwk 16 to the delivery
DAY2 Onko sinulla suonikohjuja? Do you have varicose veins? gwk 16 to the delivery
DAY2 gwk 16 to the delivery
DAY3 Onko sinulla ollut tänään närästystä? Have you had heartburn today? gwk 16 to the delivery
DAY3 Pelottaako synnytys sinua? Are you afraid of childbirth? gwk 16 to the delivery
DAY4 Onko sinulla ollut tänään suonenvetoja? Have you had cramps today? gwk 16 to the delivery
DAY4 Onko sinulla vauvaan liittyviä huolia? Do you have concerns related to the baby? gwk 16 to the delivery
DAY5 Onko sinulla ollut tänään selkäkipuja? Have you had back pain today? gwk 16 to the delivery
DAY5 Onko sinulla ollut tänään supistuksia? Have you had contractions today? gwk 16 to the delivery
Onko sinulla ollut tänään
DAY6 Have you had pain in the pelvic area today? gwk 16 to the delivery
kipuja lantion alueella?
DAY6 Oletko tuntenut tänään vauvan liikkeitä? Have you felt the baby’s movements today? gwk 16 to the delivery
DAY1 Onko sinulla vauvaan liittyviä huolia? Do you have concerns related to the baby? Postpartum weeks 1-12
DAY1 Onko sinulla ollut tänään kipuja? Have you had pain today? Postpartum weeks 1-12
DAY2 Oletko tänään onnellinen? Are you happy today? Postpartum weeks 1-12
DAY2 Oletko tuntenut itsesi virkeäksi tänään? Have you felt energetic today? Postpartum weeks 1-12
DAY2 Onko sinulla ollut tänään supistuksia? Have you had contractions today? Postpartum weeks 1-4
DAY3 Oletko liikkunut tänään ulkona? Have you been outdoors today? Postpartum weeks 1-12
DAY3 Oletko imettänyt tänään? Have you breastfed today? Postpartum weeks 1-12
DAY4 Onko sinulla itseesi liittyviä huolia? Do you have concerns related to yourself? Postpartum weeks 1-12
DAY4 Onko sinulla ollut tänään jälkivuotoa? Have you had postpartum bleeding today? Postpartum weeks 1-4
Do you feel that you have recovered
DAY4 Koetko palautuneesi synnytyksestä? Postpartum weeks 5-12
from childbirth?
Oletko tavannut tänään muita Have you met other adults today
DAY5 Postpartum weeks 1-12
aikuisia puolisosi lisäksi? besides your spouse?
Onko sinulla ollut tänään
DAY5 Have you experienced mood swings today? Postpartum weeks 1-12
mielialan vaihteluja?
DAY6 Postpartum weeks 1-12
DAY6 Onko vauvasi ollut tänään kovin itkuinen? Has your baby been very fussy today? Postpartum weeks 1-12
71
List of References
[1] L Anne Gilmore, Monica Klempel-Donchenko, and Leanne M Redman. Pregnancy as a window
to future health: excessive gestational weight gain and obesity. In Seminars in perinatology,
volume 39, pages 296–303. Elsevier, 2015.
[2] Jatta Puhkala, Jani Raitanen, Päivi Kolu, Pipsa Tuominen, Pauliina Husu, and Riitta Luoto.
Metabolic syndrome in finnish women 7 years after a gestational diabetes prevention trial. BMJ
open, 7(3):e014565, 2017.
[3] Dorien Lanssens, Thijs Vandenberk, Christophe JP Smeets, Hélène De Cannière, Sharona Vonck,
Jade Claessens, Yenthel Heyrman, Dominique Vandijck, Valerie Storms, Inge M Thijs, et al. Pre-
natal remote monitoring of women with gestational hypertensive diseases: cost analysis. Journal
of medical Internet research, 20(3):e102, 2018.
[4] Suman Kumar, Yashi Gupta, and Vijay Mago. Health-monitoring of pregnant women: Design
requirements, and proposed reference architecture. In 2019 16th IEEE Annual Consumer Com-
munications & Networking Conference (CCNC), pages 1–6. IEEE, 2019.
[5] Hisham Allahem and Srinivas Sampalli. Framework to monitor pregnant women with a high
risk of premature labour using sensor networks. In 2017 IFIP/IEEE Symposium on Integrated
Network and Service Management (IM), pages 1178–1181. IEEE, 2017.
[6] Erkan Kalafat, Karin Leslie, Amar Bhide, Basky Thilaganathan, and Asma Khalil. Pregnancy
outcomes following home blood pressure monitoring in gestational hypertension. Pregnancy
Hypertension, 2019.
[7] Kathryn I Marko, Jill M Krapf, Andrew C Meltzer, Julia Oh, Nihar Ganju, Anjali G Martinez,
Sheetal G Sheth, and Nancy D Gaba. Testing the feasibility of remote patient monitoring in
prenatal care using a mobile app and connected devices: a prospective observational trial. JMIR
research protocols, 5(4):e200, 2016.
[8] Iman Azimi, Olugbenga Oti, Sina Labbaf, Hannakaisa Niela-Vilén, Anna Axelin, Nikil Dutt,
Pasi Liljeberg, and Amir M Rahmani. Personalized maternal sleep quality assessment: An
objective iot-based longitudinal study. IEEE Access, 7:93433–93447, 2019.
[9] Johanna Saarikko, Hannakaisa Niela-Vilen, Eeva Ekholm, Lotta Hamari, Iman Azimi, Pasi Lilje-
berg, Amir M Rahmani, Eliisa Löyttyniemi, Anna Axelin, et al. Continuous 7-month internet of
things–based monitoring of health parameters of pregnant and postpartum women: Prospective
observational feasibility study. JMIR formative research, 4(7):e12417, 2020.
[10] Mário WL Moreira, Joel JPC Rodrigues, Antonio MB Oliveira, and Kashif Saleem. Smart
mobile system for pregnancy care using body sensors. In 2016 International Conference on
Selected Topics in Mobile & Wireless Networking (MoWNeT), pages 1–4. IEEE, 2016.
[11] Jochen Meyer, Elke Beck, Merlin Wasmann, and Susanne Boll. Making sense in the long run:
long-term health monitoring in real lives. In 2017 IEEE International Conference on Healthcare
Informatics (ICHI), pages 285–294. IEEE, 2017.
[12] Kirsi Grym, Hannakaisa Niela-Vilén, Eeva Ekholm, Lotta Hamari, Iman Azimi, Amir Rahmani,
Pasi Liljeberg, Eliisa Löyttyniemi, and Anna Axelin. Feasibility of smart wristbands for contin-
uous monitoring during pregnancy and one month after birth. BMC pregnancy and childbirth,
19(1):34, 2019.
[13] Lotta Hamari, Tiina Kullberg, Jukka Ruohonen, Olli J Heinonen, Natalia Dı́az-Rodrı́guez, Johan
Lilius, Anni Pakarinen, Annukka Myllymäki, Ville Leppänen, and Sanna Salanterä. Physical
72
LIST OF REFERENCES
activity among children: Objective measurements using fitbit one® and actigraph. BMC research
notes, 10(1):1–6, 2017.
[14] Arman Anzanpour, Delaram Amiri, Iman Azimi, Marco Levorato, Nikil Dutt, Pasi Liljeberg,
and Amir M Rahmani. Edge-assisted control for healthcare internet of things: A case study on
ppg-based early warning score. ACM Transactions on Internet of Things, 2(1):1–21, 2020.
[15] Sumit Majumder, Tapas Mondal, and M Jamal Deen. Wearable sensors for remote health moni-
toring. Sensors, 17(1):130, 2017.
[16] Mukhtiar Memon, Stefan Rahr Wagner, Christian Fischer Pedersen, Femina Hassan Aysha
Beevi, and Finn Overgaard Hansen. Ambient assisted living healthcare frameworks, platforms,
standards, and quality attributes. Sensors, 14(3):4312–4341, 2014.
[17] Iman Azimi, Arman Anzanpour, Amir M Rahmani, Tapio Pahikkala, Marco Levorato, Pasi Lil-
jeberg, and Nikil Dutt. Hich: Hierarchical fog-assisted computing architecture for healthcare iot.
ACM Transactions on Embedded Computing Systems (TECS), 16(5s):1–20, 2017.
[18] Priya Soma-Pillay, Catherine Nelson-Piercy, Heli Tolppanen, and Alexandre Mebazaa. Physio-
logical changes in pregnancy: review articles. Cardiovascular journal of Africa, 27(2):89–94,
2016.
[19] Andrew Carlin and Zarko Alfirevic. Physiological changes of pregnancy and monitoring. Best
practice & research Clinical obstetrics & gynaecology, 22(5):801–823, 2008.
[20] Deshayne B Fell, KS Joseph, B Anthony Armson, and Linda Dodds. The impact of pregnancy
on physical activity level. Maternal and child health journal, 13(5):597–603, 2009.
[21] Lise Loerup, Rebecca M Pullon, Jacqueline Birks, Susannah Fleming, Lucy H Mackillop,
Stephen Gerry, and Peter J Watkinson. Trends of blood pressure and heart rate in normal preg-
nancies: a systematic review and meta-analysis. BMC medicine, 17(1):1–12, 2019.
[22] Ivan D Sedov, Emily E Cameron, Sheri Madigan, and Lianne M Tomfohr-Madsen. Sleep quality
during pregnancy: a meta-analysis. Sleep medicine reviews, 38:168–176, 2018.
[23] Eeva MK Ekholm, Sampo J Piha, Risto U Erkkola, and Kari J Antila. Autonomic cardiovascular
reflexes in pregnancy. a longitudinal study. Clinical Autonomic Research, 4(4):161–165, 1994.
[24] Ryan L Brown, Christopher P Fagundes, Julian F Thayer, and Lisa M Christian. Longitudinal
changes in hrv across pregnancy and postpartum: Effect of negative partner relationship quali-
ties. Psychoneuroendocrinology, 129:105216, 2021.
[25] Nonna Heiskanen, Heli Saarelainen, Pirjo Valtonen, Tiina Lyyra-Laitinen, Tomi Laitinen, Esko
Vanninen, and Seppo Heinonen. Blood pressure and heart rate variability analysis of orthostatic
challenge in normal human pregnancies. Clinical physiology and functional imaging, 28(6):
384–390, 2008.
[26] Norma I Gavin, Bradley N Gaynes, Kathleen N Lohr, Samantha Meltzer-Brody, Gerald
Gartlehner, and Tammeka Swinson. Perinatal depression: a systematic review of prevalence
and incidence. Obstetrics & Gynecology, 106(5 Part 1):1071–1083, 2005.
[27] R Cantwell. Mental disorder in pregnancy and the early postpartum. Anaesthesia, 76:76–83,
2021.
[28] Diana Carter and Xanthoula Kostaras. Psychiatric disorders in pregnancy. British Columbia
Medical Journal, 47(2):96, 2005.
[29] Sang M Lee and DonHee Lee. Healthcare wearable devices: an analysis of key factors for
continuous use intention. Service Business, 14(4):503–531, 2020.
[30] Tony O’Donovan, John O’Donoghue, Cormac Sreenan, David Sammon, Philip O’Reilly, and
Kieran A O’Connor. A context aware wireless body area network (ban). In 2009 3rd Inter-
national Conference on Pervasive Computing Technologies for Healthcare, pages 1–8. IEEE,
2009.
[31] Rui Cao, Iman Azimi, Fatemeh Sarhaddi, Hannakaisa Niela-Vilen, Anna Axelin, Pasi Liljeberg,
Amir M Rahmani, et al. Accuracy assessment of oura ring nocturnal heart rate and heart rate
variability in comparison with electrocardiography in time and frequency domains: Comprehen-
sive analysis. Journal of Medical Internet Research, 24(1):e27487, 2022.
73
Fatemeh Sarhaddi
[32] Shancang Li, Li Da Xu, and Shanshan Zhao. The internet of things: a survey. Information
systems frontiers, 17(2):243–259, 2015.
[33] Juanli Li, Yang Liu, Jiacheng Xie, Menghui Li, Mengzhen Sun, Zhaoyang Liu, and Shuo Jiang.
A remote monitoring and diagnosis method based on four-layer iot frame perception. IEEE
Access, 7:144324–144338, 2019.
[34] Fatemeh Sarhaddi, Iman Azimi, Sina Labbaf, Hannakaisa Niela-Vilén, Nikil Dutt, Anna Axelin,
Pasi Liljeberg, and Amir M Rahmani. Long-term iot-based maternal monitoring: system design
and evaluation. Sensors, 21(7):2281, 2021.
[35] Amir-Mohammad Rahmani, Nanda Kumar Thanigaivelan, Tuan Nguyen Gia, Jose Granados,
Behailu Negash, Pasi Liljeberg, and Hannu Tenhunen. Smart e-health gateway: Bringing intel-
ligence to internet-of-things based ubiquitous healthcare systems. In 2015 12th Annual IEEE
Consumer Communications and Networking Conference (CCNC), pages 826–834. IEEE, 2015.
[36] Rangaraj M Rangayyan. Biomedical signal analysis. John Wiley & Sons, 2015.
[37] Saeed Mehrang, Mojtaba Jafari Tadi, Tero Hurnanen, Timo Knuutila, Olli Lahdenoja, Jussi
Jaakkola, Samuli Jaakkola, Tuija Vasankari, Tuomas Kiviniemi, Juhani Airaksinen, et al. Reli-
ability of self-applied smartphone mechanocardiography for atrial fibrillation detection. IEEE
Access, 7:146801–146812, 2019.
[38] Peck YS Cheang and Peter R Smith. An overview of non-contact photoplethysmography. Dept.
of Electronics & Electrical Engineering, Loughborough University, LE, 1(1), 2003.
[39] John Allen. Photoplethysmography and its application in clinical physiological measurement.
Physiological measurement, 28(3):R1, 2007.
[40] Fred Shaffer and JP Ginsberg. An overview of heart rate variability metrics and norms. Frontiers
in public health, 5:258, 2017.
[41] Rollin McCraty and Fred Shaffer. Heart rate variability: new perspectives on physiological
mechanisms, assessment of self-regulatory capacity, and health risk. Global advances in health
and medicine, 4(1):46–61, 2015.
[42] Hye-Geum Kim, Eun-Jin Cheon, Dai-Seg Bai, Young Hwan Lee, and Bon-Hoon Koo. Stress
and heart rate variability: A meta-analysis and review of the literature. Psychiatry investigation,
15(3):235, 2018.
[43] Mingzhe Jiang, Riitta Mieronkoski, Amir M Rahmani, Nora Hagelberg, Sanna Salanterä, and
Pasi Liljeberg. Ultra-short-term analysis of heart rate variability for real-time acute pain moni-
toring with wearable electronics. In 2017 IEEE International Conference on Bioinformatics and
Biomedicine (BIBM), pages 1025–1032. IEEE, 2017.
[44] Emad Kasaeyan Naeini, Sina Shahhosseini, Ajan Subramanian, Tingjue Yin, Amir M Rahmani,
and Nikil Dutt. An edge-assisted and smart system for real-time pain monitoring. In 2019
IEEE/ACM International Conference on Connected Health: Applications, Systems and Engi-
neering Technologies (CHASE), pages 47–52. IEEE, 2019.
[45] Lizawati Salahuddin, Jaegeol Cho, Myeong Gi Jeong, and Desok Kim. Ultra short term analysis
of heart rate variability for monitoring mental stress in mobile settings. In 2007 29th annual
international conference of the ieee engineering in medicine and biology society, pages 4656–
4659. IEEE, 2007.
[46] Sandrine Devot, Anna M Bianchi, Elke Naujoka, Martin O Mendez, Andreas Braurs, and Sergio
Cerutti. Sleep monitoring through a textile recording system. In 2007 29th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society, pages 2560–2563. IEEE,
2007.
[47] John A Chalmers, Daniel S Quintana, Maree J-Anne Abbott, and Andrew H Kemp. Anxiety dis-
orders are associated with reduced heart rate variability: a meta-analysis. Frontiers in psychiatry,
5:80, 2014.
[48] Shaza M Musa, Ishag Adam, and Mohamed F Lutfi. Heart rate variability and autonomic mod-
ulations in preeclampsia. PLoS One, 11(4):e0152704, 2016.
74
LIST OF REFERENCES
[49] Paulo Evaristo de Andrade, Joice Anaize Tonon do Amaral, Laercio da Silva Paiva, Fernando
Adami, Juliana Zangirolami Raimudo, Vitor Engrácia Valenti, Luiz Carlos de Abreu, and Ro-
drigo Daminello Raimundo. Reduction of heart rate variability in hypertensive elderly. Blood
pressure, 26(6):350–358, 2017.
[50] Rossana Castaldo, Paolo Melillo, Umberto Bracale, M Caserta, Maria Triassi, and Leandro Pec-
chia. Acute mental stress assessment via short term hrv analysis in healthy adults: A systematic
review with meta-analysis. Biomedical Signal Processing and Control, 18:370–377, 2015.
[51] Ahyoung Choi and Hangsik Shin. Photoplethysmography sampling frequency: pilot assessment
of how low can we go to analyze pulse rate variability with reliability? Physiological measure-
ment, 38(3):586, 2017.
[52] Emi Yuda, Muneichi Shibata, Yuki Ogata, Norihiro Ueda, Tomoyuki Yambe, Makoto Yoshizawa,
and Junichiro Hayano. Pulse rate variability: a new biomarker, not a surrogate for heart rate
variability. Journal of physiological anthropology, 39(1):1–4, 2020.
[53] Elisa Mejı́a-Mejı́a, Karthik Budidha, Tomas Ysehak Abay, James M May, and Panayiotis A Kyr-
iacou. Heart rate variability (hrv) and pulse rate variability (prv) for the assessment of autonomic
responses. Frontiers in physiology, 11:779, 2020.
[54] Marek Malik, J Thomas Bigger, A John Camm, Robert E Kleiger, Alberto Malliani, Arthur J
Moss, and Peter J Schwartz. Heart rate variability: Standards of measurement, physiological
interpretation, and clinical use. European heart journal, 17(3):354–381, 1996.
[55] Leandro Pecchia, Rossana Castaldo, Luis Montesinos, and Paolo Melillo. Are ultra-short heart
rate variability features good surrogates of short-term ones? state-of-the-art review and recom-
mendations. Healthcare technology letters, 5(3):94–100, 2018.
[56] Sarit Polsky and Rachel Garcetti. Cgm, pregnancy, and remote monitoring. Diabetes technology
& therapeutics, 19(S3):S–49, 2017.
[57] Jill M Krapf, Nancy D Gaba, Nihar Ganju, Kathryn Isabel Marko, and Anjali Gokhale Martinez.
Remote capture and monitoring of clinical data during pregnancy. Obstetrics & Gynecology,
125:26S, 2015.
[58] Betsy Diamar Balbin Lopez, Jimmy Alexander Armas Aguirre, Diego Antonio Reyes Coronado,
and Paola A Gonzalez. Wearable technology model to control and monitor hypertension during
pregnancy. In 2018 13th Iberian Conference on Information Systems and Technologies (CISTI),
pages 1–6. IEEE, 2018.
[59] Ruben Barakat, Mireia Pelaez, Carmina Lopez, Rocı́o Montejo, and Javier Coteron. Exercise
during pregnancy reduces the rate of cesarean and instrumental deliveries: results of a random-
ized controlled trial. The Journal of Maternal-Fetal & Neonatal Medicine, 25(11):2372–2376,
2012.
[60] Simony Lira Nascimento, Fernanda Garanhani Surita, Ana Carolina Godoy, Karina Tamy Ka-
sawara, and Sirlei Siani Morais. Physical activity patterns and factors related to exercise during
pregnancy: a cross sectional study. PloS one, 10(6):e0128953, 2015.
[61] Ying Yang, Jing Mao, Zhiying Ye, Xiaoli Zeng, Huimin Zhao, Yueting Liu, and Jie Li. Determi-
nants of sleep quality among pregnant women in china: a cross-sectional survey. The Journal of
Maternal-Fetal & Neonatal Medicine, 31(22):2980–2985, 2018.
[62] Sharon J Herring, Deborah B Nelson, Grace W Pien, Carol Homko, Laura M Goetzl, Adam
Davey, and Gary D Foster. Objectively measured sleep duration and hyperglycemia in pregnancy.
Sleep medicine, 15(1):51–55, 2014.
[63] Jennifer N Felder, Rebecca J Baer, Larry Rand, Laura L Jelliffe-Pawlowski, and Aric A Prather.
Sleep disorder diagnosis during pregnancy and risk of preterm birth. Obstetrics & Gynecology,
130(3):573–581, 2017.
[64] Bei Bei, Soledad Coo, and John Trinder. Sleep and mood during pregnancy and the postpartum
period. Sleep Medicine Clinics, 10(1):25–33, 2015.
[65] Margriet S Westerterp-Plantenga. Sleep, circadian rhythm and body weight: parallel develop-
ments. Proceedings of the Nutrition Society, 75(4):431–439, 2016.
75
Fatemeh Sarhaddi
[66] J Goletzke, R-D Kocalevent, G Hansen, M Rose, H Becher, K Hecher, PC Arck, and A Diemert.
Prenatal stress perception and coping strategies: Insights from a longitudinal prospective preg-
nancy cohort. Journal of psychosomatic research, 102:8–14, 2017.
[67] Fiona Alderdice, Fiona Lynn, and Marci Lobel. A review and psychometric evaluation of
pregnancy-specific stress measures. Journal of Psychosomatic Obstetrics & Gynecology, 33
(2):62–77, 2012.
[68] Gear sport. https://www.samsung.com/us/mobile/wearables/
smartwatches/gear-sport-blue-sm-r600nzbaxar/, 2023.
[69] Tizen. https://www.tizen.org, 2023.
[70] Samsung wearables. https://www.samsung.com/us/mobile/wearables/, 2022.
[71] Angular. https://angular.io/, 2023.
[72] Cordova. https://cordova.apache.org/, 2023.
[73] Omron m3. https://www.omron-healthcare.fi/en/
blood-pressure-monitors/m3_2.html#start=1, 2022.
[74] Intelli wrap cuff. https://www.omron-healthcare.fi/en/
blood-pressure-monitor-accessories/Intelli_Wrap_Cuff.html, 2022.
[75] Milad Asgari Mehrabadi, Iman Azimi, Fatemeh Sarhaddi, Anna Axelin, Hannakaisa Niela-Vilén,
Saana Myllyntausta, Sari Stenholm, Nikil Dutt, Pasi Liljeberg, Amir M Rahmani, et al. Sleep
tracking of a commercially available smart ring and smartwatch against medical-grade actig-
raphy in everyday settings: instrument validation study. JMIR mHealth and uHealth, 8(11):
e20465, 2020.
[76] Emad Kasaeyan Naeini, Iman Azimi, Amir M Rahmani, Pasi Liljeberg, and Nikil Dutt. A
real-time ppg quality assessment approach for healthcare internet-of-things. Procedia Computer
Science, 151:551–558, 2019.
[77] Iman Azimi, Tapio Pahikkala, Amir M Rahmani, Hannakaisa Niela-Vilén, Anna Axelin, and Pasi
Liljeberg. Missing data resilient decision-making for healthcare iot through personalization: A
case study on maternal health. Future Generation Computer Systems, 96:297–308, 2019.
[78] Hannakaisa Niela-Vilén, Jennifer Auxier, Eeva Ekholm, Fatemeh Sarhaddi, Milad As-
gari Mehrabadi, Aysan Mahmoudzadeh, Iman Azimi, Pasi Liljeberg, Amir M Rahmani, and
Anna Axelin. Pregnant women’s daily patterns of well-being before and during the covid-19
pandemic in finland: Longitudinal monitoring through smartwatch technology. PloS one, 16(2):
e0246494, 2021.
[79] Apache web server. https://httpd.apache.org/, 2023.
[80] Flask. https://palletsprojects.com/p/flask/, 2023.
[81] Mongodb. https://www.mongodb.com/, 2023.
[82] Kana Sasai, Shintaro Izumi, Kento Watanabe, Yuji Yano, Hiroshi Kawaguchi, and Masahiko
Yoshimoto. A low-power photoplethysmography sensor using correlated double sampling and
reference readout circuit. In 2019 IEEE SENSORS, pages 1–4. IEEE, 2019.
[83] Hyonyoung Han, Min-Joon Kim, and Jung Kim. Development of real-time motion artifact re-
duction algorithm for a wearable photoplethysmography. In 2007 29th Annual international
conference of the IEEE engineering in medicine and biology society, pages 1538–1541. IEEE,
2007.
[84] Emad Kasaeyan Naeini, Fatemeh Sarhaddi, Iman Azimi, Pasi Liljeberg, Nikil Dutt, and Amir M
Rahmani. A deep learning-based ppg quality assessment approach for heart rate and heart rate
variability. ACM Transactions on Computing for Healthcare, 2023.
[85] Anna Shcherbina, C Mikael Mattsson, Daryl Waggott, Heidi Salisbury, Jeffrey W Christle,
Trevor Hastie, Matthew T Wheeler, and Euan A Ashley. Accuracy in wrist-worn, sensor-based
measurements of heart rate and energy expenditure in a diverse cohort. Journal of personalized
medicine, 7(2):3, 2017.
[86] Liliana Barrios, Pietro Oldrati, Silvia Santini, and Andreas Lutterotti. Evaluating the accuracy
of heart rate sensors based on photoplethysmography for in-the-wild analysis. In Proceedings
76
LIST OF REFERENCES
of the 13th EAI international conference on pervasive computing technologies for healthcare,
pages 251–261, 2019.
[87] Morten Pallisgaard Støve, Emil Haucke, Maria Lysbeck Nymann, Tordur Sigurdsson, and Bir-
git Tine Larsen. Accuracy of the wearable activity tracker garmin forerunner 235 for the assess-
ment of heart rate during rest and activity. Journal of Sports Sciences, 37(8):895–901, 2019.
[88] Christina Hahnen, Cecilia G Freeman, Nilanjan Haldar, Jacquelyn N Hamati, Dylan M Bard,
Vignesh Murali, Geno J Merli, Jeffrey I Joseph, and Noud van Helmond. Accuracy of vital signs
measurements by a smartwatch and a portable health device: Validation study. JMIR mHealth
and uHealth, 8(2):e16811, 2020.
[89] Dustin T Weiler, Stefanie O Villajuan, Laura Edkins, Sean Cleary, and Jason J Saleem. Wearable
heart rate monitor technology accuracy in research: a comparative study between ppg and ecg
technology. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting,
volume 61, pages 1292–1296. SAGE Publications Sage CA: Los Angeles, CA, 2017.
[90] Sungjoo Hwang, JoonOh Seo, Houtan Jebelli, and SangHyun Lee. Feasibility analysis of heart
rate monitoring of construction workers using a photoplethysmography (ppg) sensor embedded
in a wristband-type activity tracker. Automation in Construction, 71:372–381, 2016.
[91] Angela AT Schuurmans, Peter de Looff, Karin S Nijhof, Catarina Rosada, Ron HJ Scholte, Arne
Popma, and Roy Otten. Validity of the empatica e4 wristband to measure heart rate variability
(hrv) parameters: A comparison to electrocardiography (ecg). Journal of medical systems, 44
(11):1–11, 2020.
[92] Luca Menghini, Evelyn Gianfranchi, Nicola Cellini, Elisabetta Patron, Mariaelena Tagliabue,
and Michela Sarlo. Stressing the accuracy: Wrist-worn wearable sensor validation over different
conditions. Psychophysiology, 56(11):e13441, 2019.
[93] Davide Morelli, Leonardo Bartoloni, Michele Colombo, David Plans, and David A Clifton. Pro-
filing the propagation of error from ppg to hrv features in a wearable physiological-monitoring
device. Healthcare technology letters, 5(2):59–64, 2018.
[94] Onur Dur, Colleen Rhoades, Man Suen Ng, Ragwa Elsayed, Reinier van Mourik, Maulik D
Majmudar, et al. Design rationale and performance evaluation of the wavelet health wristband:
benchtop validation of a wrist-worn physiological signal recorder. JMIR mHealth and uHealth,
6(10):e11040, 2018.
[95] David Hernando, Surya Roca, Jorge Sancho, Álvaro Alesanco, and Raquel Bailón. Validation of
the apple watch for heart rate variability measurements during relax and mental stress in healthy
subjects. Sensors, 18(8):2619, 2018.
[96] HO Kinnunen and H Koskimäki. 0312 the hrv of the ring-comparison of nocturnal hr and hrv
between a commercially available wearable ring and ecg. Sleep, 41:A120, 2018.
[97] Hannu Kinnunen, Aleksi Rantanen, Tuomas Kenttä, and Heli Koskimäki. Feasible assessment
of recovery and cardiovascular health: accuracy of nocturnal hr and hrv assessed via ring ppg in
comparison to medical grade ecg. Physiological measurement, 41(4):04NT01, 2020.
[98] Shimmer device specification. https://www.shimmersensing.com/products/
shimmer3-ecg-sensor#applications-tab, 2023.
[99] Adrian Burns, Barry R Greene, Michael J McGrath, Terrance J O’Shea, Benjamin Kuris,
Steven M Ayer, Florin Stroiescu, and Victor Cionca. Shimmer™–a wireless sensor platform
for noninvasive biomedical research. IEEE Sensors Journal, 10(9):1527–1534, 2010.
[100] Fatemeh Sarhaddi, Kianoosh Kazemi, Iman Azimi, Rui Cao, Hannakaisa Niela-Vilén, Anna
Axelin, Pasi Liljeberg, and Amir M Rahmani. A comprehensive accuracy assessment of samsung
smartwatch heart rate and heart rate variability. Plos one, 17(12):e0268361, 2022.
[101] Kianoosh Kazemi, Juho Laitala, Iman Azimi, Pasi Liljeberg, and Amir M Rahmani. Robust ppg
peak detection using dilated convolutional neural networks. Sensors, 22(16):6054, 2022.
[102] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cour-
napeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stéfan J. van der
Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nel-
son, Eric Jones, Robert Kern, Eric Larson, C J Carey, İlhan Polat, Yu Feng, Eric W. Moore, Jake
77
Fatemeh Sarhaddi
VanderPlas, Denis Laxalde, Josef Perktold, Robert Cimrman, Ian Henriksen, E. A. Quintero,
Charles R. Harris, Anne M. Archibald, Antônio H. Ribeiro, Fabian Pedregosa, Paul van Mul-
bregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algorithms for Scientific Computing
in Python. Nature Methods, 17:261–272, 2020. doi: 10.1038/s41592-019-0686-2.
[103] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Pret-
tenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot,
and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Re-
search, 12:2825–2830, 2011.
[104] Skipper Seabold and Josef Perktold. statsmodels: Econometric and statistical modeling with
python. In 9th Python in Science Conference, pages 1–10, 2010.
[105] Sayeed Shafayet Chowdhury, Rakib Hyder, Md Samzid Bin Hafiz, and Mohammad Ariful
Haque. Real-time robust heart rate estimation from wrist-type ppg signals using multiple ref-
erence adaptive noise cancellation. IEEE journal of biomedical and health informatics, 22(2):
450–459, 2016.
[106] Mohamed Elgendi. Optimal signal quality index for photoplethysmogram signals. Bioengineer-
ing, 3(4):21, 2016.
[107] Aysan Mahmoudzadeh, Iman Azimi, Amir M Rahmani, and Pasi Liljeberg. Lightweight pho-
toplethysmography quality assessment for real-time iot-based health monitoring using unsuper-
vised anomaly detection. Procedia Computer Science, 184:140–147, 2021.
[108] Hyeon Seok Seok, Sangjin Han, Junyung Park, Donggeun Roh, and Hangsik Shin. Photoplethys-
mographic pulse quality assessment methods based on similarity analysis. In 2018 Joint 10th
International Conference on Soft Computing and Intelligent Systems (SCIS) and 19th Interna-
tional Symposium on Advanced Intelligent Systems (ISIS), pages 350–353. IEEE, 2018.
[109] Jiajia Song, Dan Li, Xiaoyuan Ma, Guowei Teng, and Jianming Wei. Pqr signal quality indexes:
A method for real-time photoplethysmogram signal quality estimation based on noise interfer-
ences. Biomedical Signal Processing and Control, 47:88–95, 2019.
[110] Christina Orphanidou, Timothy Bonnici, Peter Charlton, David Clifton, David Vallance, and
Lionel Tarassenko. Signal-quality indices for the electrocardiogram and photoplethysmogram:
Derivation and applications to wireless monitoring. IEEE journal of biomedical and health
informatics, 19(3):832–838, 2014.
[111] Gangireddy Narendra Kumar Reddy, M Sabarimalai Manikandan, and NVL Narasimha Murty.
On-device integrated ppg quality assessment and sensor disconnection/saturation detection sys-
tem for iot health monitoring. IEEE Transactions on Instrumentation and Measurement, 2020.
[112] Konstantin Tyapochkin, Evgeniya Smorodnikova, and Pavel Pravdin. Smartphone ppg: signal
processing, quality assessment, and impact on hrv parameters. In 2019 41st Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pages 4237–
4240. IEEE, 2019.
[113] Xuxue Sun, Ping Yang, and Yuan-Ting Zhang. Assessment of photoplethysmogram signal qual-
ity using morphology integrated with temporal information approach. In 2012 Annual Interna-
tional Conference of the IEEE Engineering in Medicine and Biology Society, pages 3456–3459.
IEEE, 2012.
[114] Gabriele B Papini, Pedro Fonseca, Xavier L Aubert, Sebastiaan Overeem, Jan WM Bergmans,
and Rik Vullings. Photoplethysmography beat detection and pulse morphology quality assess-
ment for signal reliability estimation. In 2017 39th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC), pages 117–120. IEEE, 2017.
[115] Kejia Li, Steve Warren, and Balasubramaniam Natarajan. Onboard tagging for real-time quality
assessment of photoplethysmograms acquired by a wireless reflectance pulse oximeter. IEEE
Transactions on Biomedical Circuits and Systems, 6(1):54–63, 2011.
[116] Tania Pereira, Kais Gadhoumi, Mitchell Ma, Liu Xiuyun, Ran Xiao, Rene A Colorado, Kevin J
Keenan, Karl Meisel, and Xiao Hu. A supervised approach to robust photoplethysmography
quality assessment. IEEE Journal of Biomedical and Health Informatics, 2019.
78
LIST OF REFERENCES
[117] Shahnawaz Alam, Shreyasi Datta, Anirban Dutta Choudhury, and Arpan Pal. Sensor agnostic
photoplethysmogram signal quality assessment using morphological analysis. In Proceedings
of the 14th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Net-
working and Services, pages 176–185, 2017.
[118] Tania Pereira, Kais Gadhoumi, Mitchell Ma, Rene Colorado, Kevin J Keenan, Karl Meisel, and
Xiao Hu. Robust assessment of photoplethysmogram signal quality in the presence of atrial
fibrillation. In 2018 Computing in Cardiology Conference (CinC), volume 45, pages 1–4. IEEE,
2018.
[119] Yue Zhang and Junjun Pan. Assessment of photoplethysmogram signal quality based on fre-
quency domain and time series parameters. In 2017 10th International Congress on Image and
Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), pages 1–5. IEEE,
2017.
[120] Qiao Li and Gari D Clifford. Dynamic time warping and machine learning for signal quality
assessment of pulsatile signals. Physiological measurement, 33(9):1491, 2012.
[121] Tania Pereira, Cheng Ding, Kais Gadhoumi, Nate Tran, Rene A Colorado, Karl Meisel, and Xiao
Hu. Deep learning approaches for plethysmography signal quality assessment in the presence of
atrial fibrillation. Physiological measurement, 40(12):125002, 2019.
[122] Jessica Torres Soto and Euan Ashley. Deepbeat: A multi-task deep learning approach to assess
signal quality and arrhythmia detection in wearable devices. arXiv preprint arXiv:2001.00155,
2020.
[123] Choon-Hian Goh, Li Kuo Tan, Nigel H Lovell, Siew-Cheok Ng, Maw Pin Tan, and Einly Lim.
Robust ppg motion artifact detection using a 1-d convolution neural network. Computer methods
and programs in biomedicine, 196:105596, 2020.
[124] Donggeun Roh and Hangsik Shin. Recurrence plot and machine learning for signal quality
assessment of photoplethysmogram in mobile environment. Sensors, 21(6):2188, 2021.
[125] Simhadri Vadrevu and M Sabarimalai Manikandan. A new quality-aware quality-control data
compression framework for power reduction in iot and smartphone ppg monitoring devices.
IEEE Sensors Letters, 3(7):1–4, 2019.
[126] Aysan Mahmoudzade, Iman Azimi, Amir M Rahmani, and Pasi Liljeberg. Lightweight pho-
toplethysmography quality assessment for real-time iot-based health monitoring using unsuper-
vised anomaly detection. Procedia Computer Science, 2021.
[127] Monalisa Singha Roy, Rajarshi Gupta, and Kaushik Das Sharma. Photoplethysmogram signal
quality evaluation by unsupervised learning approach. In 2020 IEEE Applied Signal Processing
Conference (ASPCON), pages 6–10. IEEE, 2020.
[128] Juho Laitala, Mingzhe Jiang, Elise Syrjälä, Emad Kasaeyan Naeini, Antti Airola, Amir M Rah-
mani, Nikil D Dutt, and Pasi Liljeberg. Robust ecg r-peak detection using lstm. In Proceedings
of the 35th Annual ACM Symposium on Applied Computing, pages 1104–1111, 2020.
[129] Pat Hamilton. Open source ecg analysis. In Computers in cardiology, pages 101–104. IEEE,
2002.
[130] Jiapu Pan and Willis J Tompkins. A real-time qrs detection algorithm. IEEE transactions on
biomedical engineering, 32(3):230–236, 1985.
[131] Paul van Gent, Haneen Farah, Nicole van Nes, and Bart van Arem. Heartpy: A novel heart rate
algorithm for the analysis of noisy signals. Transportation research part F: traffic psychology
and behaviour, 66:368–378, 2019.
[132] Tania Pereira, Pedro R Almeida, Joao PS Cunha, and Ana Aguiar. Heart rate variability metrics
for fine-grained stress level assessment. Computer methods and programs in biomedicine, 148:
71–80, 2017.
[133] Jiuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting
Liu, Xingxing Wang, Gang Wang, Jianfei Cai, et al. Recent advances in convolutional neural
networks. Pattern recognition, 77:354–377, 2018.
79
Fatemeh Sarhaddi
[134] Laith Alzubaidi, Jinglan Zhang, Amjad J Humaidi, Ayad Al-Dujaili, Ye Duan, Omran Al-
Shamma, José Santamarı́a, Mohammed A Fadhel, Muthana Al-Amidie, and Laith Farhan. Re-
view of deep learning: Concepts, cnn architectures, challenges, applications, future directions.
Journal of big Data, 8(1):1–74, 2021.
[135] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556, 2014.
[136] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition,
pages 770–778, 2016.
[137] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen.
Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference
on computer vision and pattern recognition, pages 4510–4520, 2018.
[138] Zhiguang Wang and Tim Oates. Imaging time-series to improve classification and imputation.
In Twenty-Fourth International Joint Conference on Artificial Intelligence, pages 3939–3945,
2015.
[139] M. Abadi, , et al. Tensorflow: A system for large-scale machine learning. In OSDI’16, pages
265–283, 2016.
[140] S Moors, KJJ Staaks, MEMH Westerhuis, LRC Dekker, KMJ Verdurmen, SG Oei, and JOEH
van Laar. Heart rate variability in hypertensive pregnancy disorders: a systematic review. Preg-
nancy Hypertension, 20:56–68, 2020.
[141] Abdulnasir Hossen. Investigation of the high frequency band of heart rate variability: identifica-
tion of preeclamptic pregnancy from normal pregnancy in oman. Asian Biomed, 7(3), 2013.
[142] Cheryl CH Yang, Te-Chang Chao, Terry BJ Kuo, Chang-Sheng Yin, and Hsing I Chen.
Preeclamptic pregnancy is associated with increased sympathetic and decreased parasympa-
thetic control of hr. American Journal of Physiology-Heart and Circulatory Physiology, 278
(4):H1269–H1273, 2000.
[143] Alison K Shea, Markad V Kamath, Alison Fleming, David L Streiner, Keith Redmond, and Meir
Steiner. The effect of depression on heart rate variability during pregnancy. Clinical Autonomic
Research, 18(4):203–212, 2008.
[144] Taeko Mizuno, Koji Tamakoshi, and Keiko Tanabe. Anxiety during pregnancy and autonomic
nervous system activity: A longitudinal observational and cross-sectional study. Journal of
psychosomatic research, 99:105–111, 2017.
[145] Marijke AKA Braeken, Alexander Jones, Renee A Otte, Ivan Nyklı́ček, and Bea RH Van den
Bergh. Potential benefits of mindfulness during pregnancy on maternal autonomic nervous sys-
tem function and infant development. Psychophysiology, 54(2):279–288, 2017.
[146] Jayesh D Solanki, Freya Hiren Desai, and Kokila Hiren Desai. Heart rate variability is reduced in
normal pregnancy irrespective of trimester: A cross-sectional study from gujarat, india. Journal
of Family Medicine and Primary Care, 9(2):626, 2020.
[147] Giorgio Speranza, Giuseppe Verlato, and Alessandro Albiero. Autonomic changes during preg-
nancy: assessment by spectral heart rate variability analysis. Journal of Electrocardiology, 31
(2):101–109, 1998.
[148] D Chamchad, JC Horrow, L Nakhamchik, and VA Arkoosh. Heart rate variability changes
during pregnancy: an observational study. International Journal of Obstetric Anesthesia, 16(2):
106–109, 2007.
[149] Andreas Voss, Hagen Malberg, Agnes Schumann, Niels Wessel, Thomas Walther, Holger
Stepan, and Renaldo Faber. Baroreflex sensitivity, heart rate, and blood pressure variability
in normal pregnancy. American journal of hypertension, 13(11):1218–1225, 2000.
[150] José Javier Reyes-Lagos, Juan Carlos Echeverrı́a-Arjonilla, Miguel Ángel Peña-Castillo,
Marı́a Teresa Garcı́a-González, Maria del Rocio Ortiz-Pedroza, Gustavo Pacheco-Lopez, Car-
los Vargas-Garcia, Sergio Camal-Ugarte, and Ramon Gonzalez-Camarena. A comparison of
heart rate variability in women at the third trimester of pregnancy and during low-risk labour.
Physiology & behavior, 149:255–261, 2015.
80
LIST OF REFERENCES
[151] Phyllis K Stein, Michael T Hagley, Patricia L Cole, Peter P Domitrovich, Robert E Kleiger, and
Jeffrey N Rottman. Changes in 24-hour heart rate variability during normal pregnancy. American
journal of obstetrics and gynecology, 180(4):978–985, 1999.
[152] Mathias Baumert, Michal Javorka, Andrea Seeck, Renaldo Faber, Prashanthan Sanders, and
Andreas Voss. Multiscale entropy and detrended fluctuation analysis of qt interval and heart rate
variability during normal pregnancy. Computers in biology and medicine, 42(3):347–352, 2012.
[153] Priyanka Garg, Kavita Yadav, Ashok Kumar Jaryal, Garima Kachhawa, Alka Kriplani, and
Kishore Kumar Deepak. Sequential analysis of heart rate variability, blood pressure variability
and baroreflex sensitivity in healthy pregnancy. Clinical Autonomic Research, 30(5):433–439,
2020.
[154] Thomas Walther, Niels Wessel, Mathias Baumert, Holger Stepan, Andreas Voss, and Renaldo
Faber. Longitudinal analysis of heart rate variability in chronic hypertensive pregnancy. Hyper-
tension Research, 28(2):113–118, 2005.
[155] Manfred G Moertl, Daniela Ulrich, Karoline I Pickel, Philipp Klaritsch, Monika Schaffer, Doris
Flotzinger, Isa Alkan, Uwe Lang, and Dietmar Schlembach. Changes in haemodynamic and au-
tonomous nervous system parameters measured non-invasively throughout normal pregnancy.
European Journal of Obstetrics & Gynecology and Reproductive Biology, 144:S179–S183,
2009.
[156] Pritesh Hariprasad Gandhi, Hemant B Mehta, Ashish V Gokhale, Chetan B Desai, Pradnya A
Gokhale, and Chinmay J Shah. A study on cardiac autonomic modulation during pregnancy by
non-invasive heart rate variability measurement. International Journal of Medicine and Public
Health, 4(4), 2014.
[157] Fatemeh Sarhaddi, Iman Azimi, Anna Axelin, Hannakaisa Niela-Vilen, Pasi Liljeberg, Amir M
Rahmani, et al. Trends in heart rate and heart rate variability during pregnancy and the 3-month
postpartum period: Continuous monitoring in a free-living context. JMIR mHealth and uHealth,
10(6):e33458, 2022.
[158] Paul Van Gent, Haneen Farah, Nicole Nes, and Bart van Arem. Heart rate analysis for human
factors: Development and validation of an open source toolkit for noisy naturalistic heart rate
data. In Proceedings of the 6th HUMANIST Conference, pages 173–178, 2018.
[159] Paul van Gent, Haneen Farah, Nicole van Nes, and Bart van Arem. Analysing noisy driver
physiology real-time using off-the-shelf sensors: Heart rate analysis software from the taking
the fast lane project. Journal of Open Research Software, 7(1), 2019.
[160] Jerzy Sacha, Szymon Barabach, Gabriela Statkiewicz-Barabach, Krzysztof Sacha, Alexander
Müller, Jaroslaw Piskorski, Petra Barthel, and Georg Schmidt. How to strengthen or weaken
the hrv dependence on heart rate—description of the method and its perspectives. International
journal of cardiology, 168(2):1660–1663, 2013.
[161] Jakub S Gasior, Jerzy Sacha, Piotr J Jeleń, Mariusz Pawłowski, Bożena Werner, and Marek J
Dabrowski. Interaction between heart rate variability and heart rate in pediatric population.
Frontiers in physiology, 6:385, 2015.
[162] Stephen W Raudenbush and Anthony S Bryk. Hierarchical linear models: Applications and
data analysis methods, volume 1. sage, 2002.
[163] Michael S Cardwell. Stress: pregnancy considerations. Obstetrical & gynecological survey, 68
(2):119–129, 2013.
[164] Alexandra J Fiocco, Ridha Joober, and Sonia J Lupien. Education modulates cortisol reactivity
to the trier social stress test in middle-aged adults. Psychoneuroendocrinology, 32(8-10):1158–
1163, 2007.
[165] Sarvada Chandra Tiwari. Loneliness: A disease? Indian journal of psychiatry, 55(4):320, 2013.
[166] Katharina Diehl, Charlotte Jansen, Kamila Ishchanova, and Jennifer Hilger-Kolb. Loneliness
at universities: determinants of emotional and social loneliness among students. International
journal of environmental research and public health, 15(9):1865, 2018.
[167] Theo G van Tilburg. Social, emotional, and existential loneliness: a test of the multidimensional
concept. The Gerontologist, 61(7):e335–e344, 2021.
81
Fatemeh Sarhaddi
[168] Daniel L Surkalim, Mengyun Luo, Robert Eres, Klaus Gebel, Joseph van Buskirk, Adrian Bau-
man, and Ding Ding. The prevalence of loneliness across 113 countries: systematic review and
meta-analysis. bmj, 376, 2022.
[169] Feifei Bu, Andrew Steptoe, and Daisy Fancourt. Who is lonely in lockdown? cross-cohort
analyses of predictors of loneliness before and during the covid-19 pandemic. Public health,
186:31–34, 2020.
[170] Berta Ausı́n, Clara González-Sanguino, Miguel Ángel Castellanos, and Manuel Muñoz. Gender-
related differences in the psychological impact of confinement as a consequence of covid-19 in
spain. Journal of Gender Studies, 30(1):29–38, 2021.
[171] Balazs Kovacs, Nicholas Caplan, Samuel Grob, and Marissa King. Social networks and loneli-
ness during the covid-19 pandemic. Socius, 7:2378023120985254, 2021.
[172] Sarah ED Perzow, Ella-Marie P Hennessey, M Camille Hoffman, Nancy K Grote, Elysia Poggi
Davis, and Benjamin L Hankin. Mental health of pregnant and postpartum women in response
to the covid-19 pandemic. Journal of affective disorders reports, 4:100123, 2021.
[173] John T Cacioppo, Mary Elizabeth Hughes, Linda J Waite, Louise C Hawkley, and Ronald A
Thisted. Loneliness as a specific risk factor for depressive symptoms: cross-sectional and longi-
tudinal analyses. Psychology and aging, 21(1):140, 2006.
[174] Louise C Hawkley and John T Cacioppo. Loneliness matters: A theoretical and empirical review
of consequences and mechanisms. Annals of behavioral medicine, 40(2):218–227, 2010.
[175] Raheel Mushtaq, Sheikh Shoib, Tabindah Shah, and Sahil Mushtaq. Relationship between lone-
liness, psychiatric disorders and physical health? a review on the psychological aspects of lone-
liness. Journal of clinical and diagnostic research: JCDR, 8(9):WE01, 2014.
[176] Sunil Kumar Raina, Sujeet Raina, Vishav Chander, Ashoo Grover, Sukhjit Singh, and Ashok
Bhardwaj. Identifying risk for dementia across populations: A study on the prevalence of de-
mentia in tribal elderly population of himalayan region in northern india. Annals of indian
academy of neurology, 16(4):640, 2013.
[177] Friedrich V Wenz. Seasonal suicide attempts and forms of loneliness. Psychological Reports,
40(3):807–810, 1977.
[178] William Lauder, Kerry Mummery, Martyn Jones, and Cristina Caperchione. A comparison of
health behaviours in lonely and non-lonely populations. Psychology, Health & Medicine, 11(2):
233–245, 2006.
[179] Andrew Stickley, Ai Koyanagi, Roman Koposov, Marek Blatnỳ, Michal Hrdlička, Mary Schwab-
Stone, and Vladislav Ruchkin. Loneliness and its association with psychological and somatic
health problems among czech, russian and us adolescents. BMC psychiatry, 16(1):1–11, 2016.
[180] Julianne Holt-Lunstad, Timothy B Smith, Mark Baker, Tyler Harris, and David Stephenson.
Loneliness and social isolation as risk factors for mortality: a meta-analytic review. Perspectives
on psychological science, 10(2):227–237, 2015.
[181] Carmen Giurgescu, Ana Carolina Wong, Brooke Rengers, Sarah Vaughan, Alexandra L Nowak,
Mercedes Price, Rhonda K Dailey, Cindy M Anderson, Deborah S Walker, and Dawn P Misra.
Loneliness and depressive symptoms among pregnant black women during the covid-19 pan-
demic. Western Journal of Nursing Research, 44(1):23–30, 2022.
[182] Ellen Moseholm, Inka Aho, Åsa Mellgren, Gitte Pedersen, Terese L Katzenstein, Isik S Jo-
hansen, Diana Bach, Merete Storgaard, and Nina Weis. Psychosocial health in pregnancy and
postpartum among women living with-and without hiv and non-pregnant women living with hiv
living in the nordic countries–results from a longitudinal survey study. BMC pregnancy and
childbirth, 22(1):1–14, 2022.
[183] Rohany Nasir, Z Ahmad Zamani, Rozainee Khairudin, WS Wan Sulaiman, MN Mohd Sani, and
AS Amin. Depression, loneliness and cognitive distortion among young unwed pregnant women
in malaysia: counseling implications. Asian Social Science, 12(8):104–109, 2016.
[184] Prabhaskar Rai. Depression, loneliness and social support among adolescent mothers. Deliber-
ative Research, 30(1):5, 2016.
82
LIST OF REFERENCES
[185] Ilona Luoma, Marie Korhonen, Kaija Puura, and Raili K Salmelin. Maternal loneliness: concur-
rent and longitudinal associations with depressive symptoms and child adjustment. Psychology,
health & medicine, 24(6):667–679, 2019.
[186] Farheen Zaidi, Aruna Nigam, Ruby Anjum, and Rashmi Agarwalla. Postpartum depression in
women: a risk factor analysis. Journal of Clinical and Diagnostic Research: JCDR, 11(8):QC13,
2017.
[187] Violeta Lopez Nahas, Sharon Hillege, and Nawal Amasheh. Postpartum depression: The lived
experiences of middle eastern migrant women in australia. Journal of nurse-midwifery, 44(1):
65–74, 1999.
[188] Virginia Harrison, Michelle L Moulds, and Katie Jones. Perceived social support and prenatal
wellbeing; the mediating effects of loneliness and repetitive negative thinking on anxiety and
depression during the covid-19 pandemic. Women and Birth, 2021.
[189] Takashi Takeda, Kana Yoshimi, Sayaka Kai, and Fumi Inoue. Association between serious
psychological distress and loneliness during the covid-19 pandemic: A cross-sectional study
with pregnant japanese women. International journal of women’s health, 13:1087, 2021.
[190] Linnea Schuez-Havupalo, Elina Lahti, Niina Junttila, Laura Toivonen, Minna Aromaa, Päivi
Rautava, Ville Peltola, and Hannele Räihä. Parents’ depression and loneliness during pregnancy
and respiratory infections in the offspring: a prospective birth cohort study. PLoS One, 13(9):
e0203650, 2018.
[191] Louis Achterbergh, Alexandra Pitman, Mary Birken, Eiluned Pearce, Herman Sno, and Sonia
Johnson. The experience of loneliness among young people with depression: a qualitative meta-
synthesis of the literature. BMC psychiatry, 20(1):1–23, 2020.
[192] Afsaneh Doryab, Daniella K Villalba, Prerna Chikersal, Janine M Dutcher, Michael Tumminia,
Xinwen Liu, Sheldon Cohen, Kasey Creswell, Jennifer Mankoff, John D Creswell, et al. Iden-
tifying behavioral phenotypes of loneliness and social isolation with passive sensing: statistical
analysis, data mining and machine learning of smartphone and fitbit data. JMIR mHealth and
uHealth, 7(7):e13209, 2019.
[193] Congyu Wu, Amanda N Barczyk, R Cameron Craddock, Gabriella M Harari, Edison Thomaz,
Jason D Shumake, Christopher G Beevers, Samuel D Gosling, and David M Schnyer. Improving
prediction of real-time loneliness and companionship type using geosocial features of personal
smartphone data. Smart Health, 20:100180, 2021.
[194] Charlotte M Roddick and Frances S Chen. Effects of chronic and state loneliness on heart rate
variability in women. Annals of Behavioral Medicine, 55(5):460–475, 2021.
[195] Louise C Hawkley, Christopher M Masi, Jarett D Berry, and John T Cacioppo. Loneliness is a
unique predictor of age-related differences in systolic blood pressure. Psychology and aging, 21
(1):152, 2006.
[196] Louise C Hawkley, Kristopher J Preacher, and John T Cacioppo. Loneliness impairs daytime
functioning but not sleep duration. Health psychology, 29(2):124, 2010.
[197] John T Cacioppo, Louise C Hawkley, L Elizabeth Crawford, John M Ernst, Mary H Burleson,
Ray B Kowalewski, William B Malarkey, Eve Van Cauter, and Gary G Berntson. Loneliness and
health: Potential mechanisms. Psychosomatic medicine, 64(3):407–417, 2002.
[198] Dan Russell, Letitia A Peplau, and Carolyn E Cutrona. The revised ucla loneliness scale: con-
current and discriminant validity evidence. Journal of personality and social psychology, 39(3):
472, 1980.
[199] Fatemeh Sarhaddi, Iman Azimi, Hannakaisa Niela-Vilen, Anna Axelin, Pasi Liljeberg, Amir M
Rahmani, et al. Maternal social loneliness detection using passive sensing through continuous
monitoring in everyday settings: Longitudinal study. JMIR Formative Research, 7(1):e47950,
2023.
[200] Max Hirshkowitz, Kaitlyn Whiton, Steven M Albert, Cathy Alessi, Oliviero Bruni, Lydia Don-
Carlos, Nancy Hazen, John Herman, Eliot S Katz, Leila Kheirandish-Gozal, et al. National
sleep foundation’s sleep time duration recommendations: methodology and results summary.
Sleep health, 1(1):40–43, 2015.
83
[201] Glenn De’ath and Katharina E Fabricius. Classification and regression trees: a powerful yet
simple technique for ecological data analysis. Ecology, 81(11):3178–3192, 2000.
[202] Wei-Yin Loh. Classification and regression trees. Wiley interdisciplinary reviews: data mining
and knowledge discovery, 1(1):14–23, 2011.
[203] Peter Prettenhofer and Gilles Louppe. Gradient boosted regression trees in scikit-learn. Unpub-
lished conference/Abstract, 2014.
[204] Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion,
Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-
learn: Machine learning in python. the Journal of machine Learning research, 12:2825–2830,
2011.
[205] Daniel Fulford, Jasmine Mote, Rachel Gonzalez, Samuel Abplanalp, Yuting Zhang, Jarrod Luck-
enbaugh, Jukka-Pekka Onnela, Carlos Busso, and David E Gard. Smartphone sensing of social
interactions in people with and without schizophrenia. Journal of Psychiatric Research, 137:
613–620, 2021.
[206] Dror Ben-Zeev, Emily A Scherer, Rui Wang, Haiyi Xie, and Andrew T Campbell. Next-
generation psychiatric assessment: Using smartphone sensors to monitor behavior and mental
health. Psychiatric rehabilitation journal, 38(3):218, 2015.
Painosalama, Turku, Finland 2023
ISBN 978-951-29-9581-3 (PRINT)

ISBN 978-951-29-9582-0 (PDF)
ISSN 2736-9390 (Painettu/Print)
ISSN 2736-9684 (Sähköinen/Online)

Annales F 34 Sarhaddi DISS

Uploaded by

Copyright:

Available Formats

Annales F 34 Sarhaddi DISS

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Annales F 34 Sarhaddi DISS

Uploaded by

Copyright:

Available Formats

CONTINUOUS IOT-BASED

TURUN YLIOPISTON JULKAISUJA – ANNALES UNIVERSITATIS TURKUENSIS

TURUN YLIOPISTON JULKAISUJA – ANNALES UNIVERSITATIS TURKUENSIS

Associate Professor, Anna Axelin Adjunct Professor, Iman Azimi

ISBN 978-951-29-9581-3 (PRINT)

KEYWORDS: maternal health, heart rate variability, remote health monitoring,

Äitiysterveydenhuollon tavoitteena on varmistaa terve raskaus, turvallinen synnytys

AVAINSANAT: äitiysterveydenhuolto, sykevälivaihtelu, etäterveyden seuranta,

I want to express my gratitude to my supervisors, Prof. Pasi Liljeberg, Prof. Amir.

Date December 11, 2023

List of Original Publications . . . . . . . . . . . . . . . . . . . . . . xi

3 IoT-Based Maternal Health Monitoring . . . . . . . . . . . . . 13

4 Validation and Quality Assessment of HR and HRV Param-

5 IoT-Based Maternal Monitoring During Pregnancy and Post-

6 Maternal Loneliness Detection Based on Passive Data Sens-

8 Overview of Original Publications . . . . . . . . . . . . . . . . 66

Appendix A: questions provided by the mobile application . 70

I Fatemeh Sarhaddi, Iman Azimi, Sina Labbaf, Hannakaisa Niela-Vilén, Nikil

II Fatemeh Sarhaddi* , Kianoosh Kazemi* , Iman Azimi, Rui Cao, Hannakaisa

IV Fatemeh Sarhaddi, Iman Azimi, Anna Axelin, Hannakaisa Niela-Vilen,

V Fatemeh Sarhaddi, Iman Azimi, Anna Axelin, Hannakaisa Niela-Vilen,

VI Hannakaisa Niela-Vilén, Jennifer Auxier, Eeva Ekholm, Fatemeh Sarhaddi,

IX Hannakaisa Niela-Vilen, Iman Azimi, Kristin Suorsa, Fatemeh Sarhaddi,

X Jennifer Auxier, Kaisu T. Savolainen, Miriam Bender, Amir M. Rahmani,

XI Hannakaisa Niela-Vilen, Eeva Ekholm, Fatemeh Sarhaddi, Iman Azimi,

tiveness of using such systems to gain a deeper understanding of normal trends in

1.1 IoT-Based Maternal Monitoring Opportunities and

1.2 Aims and Objectives

• Research Objective I: Present, develop, and evaluate a continuous IoT-based

• Research Objective IV: Investigate the trends of HR and HRV parameters

• Research Objective V: Develop predictive models to detect maternal lone-

• Develop, implement, and evaluate a continuous maternal monitoring sys-

• Investigating the validity of HR and HRV parameters collected by the

• Proposing and implementing a deep learning-based PPG quality assess-

• Continuous long-term monitoring of physiological health parameters of

• Analyzing the trends of HR and HRV parameters during pregnancy and

• Proposing predictive models for maternal loneliness detection using pas-

1.4 Thesis Organization

Chapter 4 Chapter 5 Chapter 6

Research Objective II Research Objectives III & IV Research Objective V

2.1 Internet of Things-based Health-Monitoring Systems

Wearable Device Gateway Layer Cloud Layer

Smartphone Data Modeling Application Layer

Router Data Storage Web Application

Background and Smart

Figure 2. IoT-based health-monitoring architecture [34]

2.2 Cardiac Activity Measurement

Templates of normal ECG signals Normal ECG signal

Templates of ECG signals with AF ECG signal for AF

Figure 5. A 5-second sample of filtered PPG signal

Table 1. HRV parameters with definitions [51]

Variable Units Description

2.3 Heart Rate Variability

• Short-term HRV analysis: This standard uses 5-minute recordings to extract

• Ultrashort-term HRV analysis: This method uses less-than-5-minute record-

In this chapter, we present a long-term IoT-based remote maternal monitoring sys-

3.1 Related Works

3.2 Maternal Health Monitoring Services

• Physical activity monitoring: Physical activity is a key factor in well-being.

• Sleep monitoring: Sleep disorders are common during pregnancy, especially

RMSSD 22433 Deep Models