Data-driven health monitoring and lifestyle interventions

Data-driven health monitoring and lifestyle interventions

Citation for published version (APA):

Radha, M. G. (2020). Data-driven health monitoring and lifestyle interventions: towards management of hypertension and other lifestyle diseases through data-driven modelling of physiology and behaviour. [Phd Thesis 1 (Research TU/e / Graduation TU/e), Electrical Engineering]. Technische Universiteit Eindhoven.

Document status and date:

Published: 18/02/2020

Document Version:

Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.

• The final author version and the galley proof are versions of the publication after peer review.

• The final published version features the final layout of the paper including the volume, issue and page numbers.

Link to publication

General rights

Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.

• You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:

www.tue.nl/taverne Take down policy

If you believe that this document breaches copyright please contact us at:

openaccess@tue.nl

providing details and we will investigate your claim.

Download date: 17. Sep. 2022

Data-driven health monitoring Data-driven health monitoring and lifestyle interventions

and lifestyle interventions

Mustafa Radha

Data-driven health monitoring and lifestyle interventions

Towards management of hypertension and other lifestyle diseases through data-driven modelling of physiology and behavior

interventions

Towards management of hypertension and other lifestyle diseases through data-driven modelling of physiology and

behaviour

PROEFSCHRIFT

ter verkrijging van de graad van doctor aan de Technische Universiteit Eindhoven, op gezag van de rector magnificus prof.dr.ir. F.P.T. Baaijens, voor een commissie aangewezen door het College voor Promoties, in het

openbaar te verdedigen op dinsdag 18 februari 2020 11:00 uur

door

Mustafa Ghassan Radha

geboren te Baaqubah, Irak

Dit proefschrift is goedgekeurd door de promotoren en de samenstelling van de promotiecommissie is als volgt:

voorzitter: prof. dr. A.M.J. Koonen 1e promotor: prof. dr. R.M. Aarts 2e-promotor: prof. dr. W.A. IJsselsteijn compromotor: dr. M.C. Willemsen

leden: prof. dr. L. van Gemert-Pijnen (University of Twente) dr. J. Allen (Newcastle University)

prof. dr. E. Korsten (Catharina Hospital)

adviseur(s): dr. V. Menkovski (Eindhoven University of Technology)

Het onderzoek of ontwerp dat in dit proefschrift wordt beschreven is uitgevoerd in overeen- stemming met de TU/e Gedragscode Wetenschapsbeoefening.

Abstract

Modern lifestyle patterns are associated with chronic diseases that are highly prevalent world-wide, despite the fact that these conditions are preventable. A main reason for this is that the signs and risk factors for chronic lifestyle diseases are difficult to detect in a timely manner. In addition, even when the presence of risk factors is confirmed, individuals still experience difficulty to make the necessary lifestyle changes. The aim of the work presented in this thesis is to explore data-driven modelling and prediction techniques to develop tools that can assist in prevention of chronic lifestyle diseases. The developed techniques have the aim to (1) monitor a person’s health continuously and unobtrusively in order to detect and inform people of health risk factors, (2) model a person’s behaviour and habits in order to recommend suitable lifestyle improvement interventions and (3) track a person during lifestyle interventions to ensure that the intervention is performed effectively and safely.

In the first part of the thesis, the aim was to develop and evaluate a methodology for unobtrusive, wearable sleep hypnogram monitoring at home. A deep learning approach is proposed to learn the temporal characteristics of sleep architecture from heart-rate variability data, which is first trained on electrocardiogram data and later adapted through a transfer learning methodology to photoplethysmography data. The model achieves state-of-the-art performance by combining a rich set of physiological features extracted from heart beat measurements with a versatile long short-term memory neural network algorithm. This method has applications in self-monitoring of sleep and screening of sleep disorders.

In the second part of the thesis, the same unobtrusive photoplethysmography sensor is used to estimate variations in blood pressure throughout the day, with a focus on estimating the nocturnal blood pressure dip, which is highly predictive of cardiovascular disease risk. Again, a long short- term memory neural network is combined with a rich physiological feature set to show promising results in comparison to several baselines.

In the final part of the thesis, the research focus turns to the modelling of behaviour in order to help individuals achieve their desired lifestyle goals, with a focus on hypertensive individuals. First, the Rasch model is evaluated as a way to model the feasibility of lifestyle recommendations. This model is used to generate lifestyle advice and compared against various baselines in a comparative study. Finally, resistance exercise is selected as a particularly challenging intervention and a feedback system is proposed to help individuals perform the intervention safely and effectively at home.

In conclusion, sleep can provide a physiological baseline for health monitoring. Utilising unobtrusive sensing technology, signal processing techniques and novel machine learning ap- proaches, it is possible to measure meaningful health parameters during sleep that are otherwise too cumbersome or expensive to monitor frequently. Combining such monitoring technologies with virtual recommendations and guidance for lifestyle interventions has a strong potential to enable self-management of health and the prevention of many chronic lifestyle diseases. It remains a goal for future research to combine these different data-driven technologies into holistic systems and evaluating the total combined effect on peoples health over longer timeframes.

Acknowledgments

First of all I would like to thank Ronald. Thank you for supporting me throughout these years.

As a professor, scientist, colleague and inventor, you set an inspiring example on how to tackle problems creatively and rigorously and how to combine and synergise multiple goals, all the while remaining open-minded and supportive of new ideas. Martijn, I would also like to acknowledge the tremendous guidance you gave, helping me to stay focussed on important things when it got chaotic, getting involved in all aspects of my project and providing profound advice in the conception, execution and reporting of my research. Also thanks to Wijnand as my second promotor. I can say that without exception every conversation with you has been inspiring and widened my perspective.

Finally, I would like to acknowledge and thank Reinder for providing the mentoring and support to get my project rolling and co-shaping the direction of my research.

Over the years I have had the pleasure of meeting and collaborating with inspiring colleagues at Philips Research and beyond and I am very grateful for having them as part of my journey.

Thanks to all the great colleagues in the wearable monitoring projects: Pedro, Koen, Jos, Stefan, Alberto, Jenny, Guido, Lieke, Xi and the rest of the wearable monitoring gang! I’d also like to extend my thanks to the colleagues from the blood pressure management project: Nathalie, Francesco, Boris, Helma, Gerd, Raymond, thanks for the great times and for working with me on these fascinating topics. Finally I’d like to thank the entire personal health group for providing a creative work environment where new ideas are always embraced, we’ve had a great run! Special thanks to Joerg for believing in me and offering me the opportunity to do this work in his department.

Thanks go to the consortium partners from Imperial College London for the collaboration we had on blood pressure monitoring and management: Petra, Cybele, Nikita, Nadja, Valentina, Nikolaos - thank you! My thanks also go to my fellow colleagues in continuous personal health:

Linda, Marta, Saskia, Chao, and the rest of the "bloody" flagship: Niek, Felipe, Heleen, Alberto, Mark, Humberto, Bart, Alok, Laura, Silvia, Marie, Szilard. We were in this together and every meet-up was a beautiful mix of stories with all the recognisable ups and downs of this roller- coaster ride. Acknowledgements go to the master students I had the pleasure of guiding: Niels, Mark, Alexandros, Pradeep and Eefke. Thank you for spending your internships with me, you had a positive impact on my work and I’m sure you’ll achieve great things in your future endeavours.

On a personal note, I’d like to thank my friends who I could always count on to get my mind off of work and have some good old hedonistic fun together...the crazy snack-bar favourites, the basement warriors, the mid-or-feeders... thanks for the memorable experiences over the years. Then I’d like to thank my mother Akbal and my father Ghassan. For always having the best intentions for me and always supporting, nurturing and celebrating the lives of your sons. Thanks for the hard life you went through and the sacrifices you’ve made in order to make our lives better. Also thanks to my brother Safuat, for always being up for a good chat and for hearing out my boring rants about the PhD life! Finally, I want to thank Roos, for being at my side these years. For the warmth and affection you have given me as we shared our happy and sad moments together. You have supported me in so many ways: motivating me when things get tough, reviewing my papers, correcting me every time I misuse medical jargon :) and even designing the cover of this book.

Thank you my love.

-Mustafa

Contents

1

1.1 Wearable Monitoring 16

1.1.1 Activity tracking . . . 16 1.1.2 Physiological tracking . . . 17 1.1.3 Wearable monitoring during sleep . . . 18

1.2 Sleep stage monitoring 20

1.2.1 Automated sleep stage classification using EEG . . . 21 1.2.2 Sleep stage classification from PPG . . . 21

1.3 Blood pressure monitoring 22

1.3.1 Blood pressure as a continuous variable . . . 22 1.3.2 The blood pressure dip . . . 23 1.3.3 The measurement of BP dipping with PPG . . . 24

1.4 Machine learning 24

1.4.1 Supervised machine learning . . . 24 1.4.2 Sequence-to-sequence modelling . . . 26 1.4.3 Transfer learning . . . 28

1.5 Lifestyle change 29

1.5.1 Feasibility . . . 29 1.5.2 Safety . . . 30

1.6 Overall project aims 30

2

2.1 Introduction 35 2.1.1 Non-temporal models . . . 36

2.1.2 Temporal models . . . 37

2.1.3 Long short-term memory model . . . 38

2.2 Materials and Methods 39 2.2.1 Materials . . . 39

2.2.2 Feature extraction . . . 40

2.2.3 Machine learning model . . . 40

2.3 Results 43 2.3.1 Model training and evaluation . . . 43

2.3.2 The effect of demographic factors on performance . . . 44

2.3.3 Performance for different patient groups . . . 44

2.4 Discussion 46 2.5 Conclusion 47

3

3.1 Introduction 49 3.1.1 Manual sleep stage scoring . . . 49

3.1.2 Heart-rate variability . . . 49

3.1.3 Photoplethysmography . . . 50

3.1.4 Recurrent deep neural networks . . . 52

3.1.5 Transfer learning . . . 52

3.1.6 Objective . . . 53

3.2 Materials 53 3.2.1 Siesta data set . . . 54

3.2.2 Eindhoven data set . . . 54

3.3 Methods 55 3.3.1 HRV feature extraction . . . 55

3.3.2 The model . . . 55

3.3.3 Model training and evaluation . . . 57

3.3.4 Analysis . . . 58

3.4 Results 58 3.5 Discussion 59 3.6 Conclusion 64

II Unobtrusive blood pressure monitoring 4

4.1 Introduction 67 4.1.1 Pulse wave velocity . . . 67

4.1.2 The PWV-BP relation . . . 68

4.1.3 Practical issues . . . 69

4.2 Research goals 70

4.3 Materials 71

4.4 Methods 71

4.4.1 Signal Processing . . . 71

4.4.2 Statistical analysis . . . 74

4.5 Results and discussion 75 4.5.1 PPG markers of pulse arrival . . . 75

4.5.2 Site-related differences in PWV and its relation to SBP . . . 75

4.5.3 Prediction of SBP with PWV to different arterial sites . . . 79

4.6 Conclusions 80

5

5.1 Introduction 81 5.2 Background 82 5.2.1 Physiological models . . . 82

5.2.2 Machine learning models . . . 82

5.2.3 Evaluation in free-living context . . . 83

5.2.4 Objective of the current study . . . 83

5.3 Methods 84 5.3.1 Data . . . 84

5.3.2 Feature extraction . . . 85

5.3.3 Models . . . 89

5.3.4 Blood pressure tracking evaluation . . . 90

5.3.5 Blood pressure dip evaluation . . . 90

5.4 Results 90 5.5 Discussion 91 5.5.1 Machine learning models . . . 93

5.5.2 Comparison to earlier work . . . 94

5.5.3 Sources of error and limitations . . . 95

5.5.4 Clinical significance . . . 95

5.6 Conclusions 96

III Lifestyle interventions to improve blood pressure 6

6.1 Introduction and Context 99 6.1.1 Recommender systems . . . 99

6.1.2 Feasibility . . . 99

6.1.3 The Rasch model as a feasibility model . . . 100

6.1.4 Feasibility-based recommendation strategies . . . 101

6.2 Research objectives 102 6.3 Model development 102 6.3.1 Modeled behaviors . . . 103

6.3.2 Population characteristics . . . 104

6.3.3 Model fitting and analysis . . . 104

6.4.1 Recommendation strategies . . . 106

6.4.2 Study design . . . 108

6.4.3 Participants . . . 109

6.4.4 Results . . . 109

6.4.5 Discussion . . . 110

6.5 Conclusions 111

7

7.1 Introduction 113 7.2 Methods 114 7.2.1 Conditions . . . 114

7.2.2 Study design and participants . . . 116

7.2.3 Measures . . . 117

7.2.4 Procedure . . . 118

7.2.5 Statistical analysis . . . 119

7.3 Results 120 7.3.1 Concentric contraction time . . . 120

7.3.2 Eccentric contraction time . . . 121

7.3.3 Concentric endpoint variations . . . 122

7.3.4 Eccentric endpoint variations . . . 122

7.3.5 Respiration . . . 122

7.3.6 Perceived competence . . . 122

7.3.7 Interest/enjoyment . . . 122

7.3.8 Focus on attention . . . 124

7.3.9 Rating of perceived exertion . . . 124

7.3.10 User experience . . . 124

7.4 Discussion 126 7.4.1 Effect on exercise performance . . . 126

7.4.2 Effect on perceived competence and intrinsic motivation . . . 126

7.4.3 Effect on attentional focus and perceived exertion . . . 127

7.4.4 Limitations and future research . . . 127

7.5 Conclusions 127

8

8.1 Unobtrusive sleep stage monitoring 129 8.1.1 Long short-term memory networks for sleep stage classification . . . 130

8.1.2 Transfer learning for PPG-based sleep stage classification . . . 130

8.2 Unobtrusive blood pressure monitoring 131 8.2.1 Pulse arrival time . . . 131

8.2.2 A machine learning approach to blood pressure estimation . . . 131

8.3 Lifestyle interventions 132 8.3.1 Rasch modelling of behavior feasibility . . . 132

8.3.2 Guided home-based resistance exercise . . . 133

8.4 Towards preventing and managing lifestyle disease 133

9

Articles 135 Conference proceedings 155 Books 158 Miscellaneous 159

10

11

A

B

Index . . . 171

Acronyms

A/S ambient light and sonification. 114, 116, 118, 120–122, 124, 126–128

AASM American Association of Sleep Medicine. 20, 22, 35, 49, 50, 52–54, 57, 63 ABPM ambulatory blood pressure monitor. 23, 24, 132

ACSM American College of Sports Medicine. 30, 113

BMI body-mass index. 42, 44, 54

BP blood pressure. 16–19, 22–24, 28–31, 67–69, 71, 74, 79, 81–89, 91, 93–96, 99, 113, 114, 116, 118, 119, 131, 132

DASH dietary approaches to stop hypertension. 103

DBP diastolic blood pressure. 22, 23, 67, 84–86, 88–91, 93, 96, 117

ECG electrocardiography. 16, 17, 20, 21, 28–31, 35–40, 46, 50–55, 57–59, 63, 64, 68, 72–74, 82, 94, 130

EEG electroencephalography. 20, 21, 35, 38, 39, 47, 52, 54 EMG electromyography. 20, 35, 39, 54

EMS engagement maximisation strategy. 102, 106, 108–111 EOG electrooculography. 20, 21, 35, 39, 54

ESC European Society of Cardiology. 23, 99

ESH European Society of Hypertension. 22, 23, 99

HRV heart rate variability. 17–19, 21, 24, 26, 28, 30, 31, 35–38, 40, 46, 47, 52, 91, 129, 130, 132 IBI inter-beat interval. 17, 19, 21, 40, 41, 51

ICC item-characteristics curve. 101, 104

ICD-10 International Classification of Diseases, tenth edition. 39 IRT item-response theory. 100

LMER linear mixed effects regression. 119

LSTM long short-term memory. 26, 30, 38, 40, 42, 43, 46, 47, 52, 53, 55, 57, 63, 64, 83, 89–94, 96, 129, 130, 132

MMS motivation maximisation strategy. 102, 106, 108–111 PAT pulse arrival time. 51, 67, 73, 74, 76, 79, 82, 94, 131 PLMD Periodic Limb Movement Disorder. 43, 44, 46, 47

PPG photoplethysmography. 16–20, 24, 28–31, 36, 37, 50–55, 57–60, 63, 64, 68, 69, 71, 73–75, 77, 79–88, 91, 94, 95, 130, 131, 133, 134

PSG polysomnography. 20, 21, 35, 39, 49, 50, 52, 53, 130 PWV pulse wave velocity. 28, 67–71, 73–75, 77, 79, 80 R&K Rechtschaffen and Kales. 39, 40, 49, 50, 52–54, 57, 63 RCS random control strategy. 108–111

REM rapid eye movement. 20, 22, 35, 49

RMSE root-mean-squared-error. 83, 84, 90, 91, 93, 94, 96, 168, 169 SBP systolic blood pressure. 17, 22–24, 31, 67–69, 73–75, 79–96, 117, 132 SE sleep efficiency. 51, 54, 58, 59, 64

SOL sleep onset latency. 51, 58, 59, 64 TST total sleep time. 51, 58, 59, 64

WASO wake after sleep onset. 51, 58, 59, 64

1. Introduction

In humanity’s never-ending pursuit of survival, we have continuously reshaped the world around us to be as accommodating as possible to human life. From learning to fight off deadly animal predators up to the advent of vaccinations and eradication of some of the most dangerous infectious diseases, as a species we have come a long way and live a more healthy and peaceful life than ever before. While these efforts are a testimony to the success of human survival, the changes to our environment happen at a pace that our biology cannot keep up with. Our survival instincts and biology are still obsessed over spiders and snakes while modern deadly hazards are more likely to manifest as a driving accident at high speed. A far more subtle category of health hazards has come to light in the last decades: chronic lifestyle diseases. Over the long life spans enjoyed in modern society, an unhealthy lifestyle tends to take its toll on our bodies and severely damage primary bodily functions. For example, sleep deprivation, unhealthy nutrition comprising modern ultra-processed foods, sedentary lifestyle, cognitive stress and a plethora of intoxicating stimulants are common components of a modern lifestyle. Such lifestyle behaviors severely increase the risk of health complications such as cardiovascular, cerebral or autonomic disease. In addition, our biological instincts are not optimised to avoid such lifestyle behaviors. In fact, the opposite is often true in many cases.

Presented with these new challenges, scholars of modern medicine studied the now highly prevalent chronic lifestyle diseases and proposed diagnostic tools and medications to detect lifestyle diseases such as diabetes and hypertension and control the potential damage they can cause upon patients. These tools and medications are in wide-spread use in most parts of the world and have drastically decreased the incidence of life-threatening cardiovascular events. However, to this date there are many who live with a chronic condition for years before its presence is detected, and unfortunately the detection in these cases oftentimes takes place after an acute life-threatening event has happened.

In order to improve prevention and timely detection of lifestyle diseases, governments have dedicated resources to raise awareness about the importance of a healthy lifestyle, which has

proved effective in educating the general public about the negative consequences of unhealthy living. Typical recommendations for a healthy lifestyle to prevent chronic disease are similar throughout the world: a balanced and healthy diet, adequate amounts of physical activity, proper sleep hygiene, avoiding high stress levels and limiting and/or stopping the use of toxic stimulants such as alcohol, drugs and tobacco [167]. These behaviors can keep weight, blood pressure (BP), glucose metabolism and cholesterol at healthy levels, thereby effectively preventing most cardiovascular diseases. The dissemination of such recommendations have already led to strong reductions in for example smoking [323]. However there is still a strong need for better self- management. Cardiovascular disease is still a leading cause of death [1] and health risk behaviors are still prevalent. For example, sleep deprivation has been called an emerging epidemic world-wide [85, 209] while physical inactivity is increasing in developed countries [84].

1.1 Wearable Monitoring

Recently, a unique type of technology has started gaining mainstream attention as a tool in the self-management and self-monitoring of personal health. Through the growing awareness of the importance of a healthy lifestyle, the reduced costs of producing digital technology, and the in- creased availability of connectivity and mobile computing power, these wearable health monitoring technologies have become off-the-shelf consumer devices [214]. Most users of wearables fall within the so-called "quantified self" movement [213]: people who already are very engaged with a healthy lifestyle and want to quantify their achieved positive results as a personal motivator.

However, the promise of wearable health monitoring technology is to provide personalised health data which could assist with diagnosis and the adoption of healthier lifestyle behaviors, provided that concerns regarding safety, reliability and security are addressed [174].

1.1.1 Activity tracking

A main use case for wearable technology is activity tracking. This is usually done with an unobtrusive device such as wrist-worn bracelets containing both an accelerometer and a photo- plethysmography (PPG) sensor, often accompanied by a screen on which readings from the sensors as well as other information aggregated from connected devices can be viewed in real-time. Recent studies are showing that the use of wearables for activity tracking can have a positive effect on weight loss [38] and physical activity promotion [78].

Accelerometers

Accelerometers enjoy widespread use, not only in wearable devices but also in other on-body computers such as smart phones [236] and pendants [31]. Accelerometers have a large variety of uses, such as the recognition of activities of daily life [173], sports activities [119], fall detection [10], sleep-wake pattern detection [112] and step counting [66]. Most of these functions can aid in quantifying one’s lifestyle patterns, however the accelerometer is not a physiological sensor and thus it is by itself not suitable for the measurement of health parameters.

Photoplethysmography

Heart rate sensors have existed since the 1980s in the form of electrodes placed in chest belts producing a simplified electrocardiography (ECG) signal [127]. The relative bulkiness of the device and its obtrusiveness in comparison to modern wrist-worn wearables have limited the adoption, with the exception of its use for aerobic exercise. Instead, modern wearables come predominantly with a wrist-mounted PPG sensor. The PPG sensor is placed on the skin-facing side of a wrist bracelet, consisting of a light-emitting diode that illuminates the skin with light, while a photo-detector captures the reflection of the light from the skin. The capillaries and arterioles in the skin bed absorb light and thus the reflected light intensity in these wavelengths is attenuated by the volume

1.1 Wearable Monitoring 17 of blood in the skin. This blood volume is modulated by the blood pulses originating from heart beats. Analysing the pulsatile component of PPG can therefore be used to measure heart rate [261]

(see Figure 1.3 for an example). The heart rate provides information about cardiac performance during activity and at rest. In addition, it can be used to improve estimations of energy expenditure [26, 195].

1.1.2 Physiological tracking

wearables, and specifically PPG technology, also have the potential to measure and monitor the user’s physiological state, next to general activity tracking. As PPG sensors have been in existence in clinical setting as blood oxygen saturation monitors, the PPG signal has already been thoroughly investigated for monitoring health. Allen [5] provided an extensive review of the experiments done with PPG to estimate clinical health parameters. In the review, he summarised the results of a large number of studies indicating the potential of PPG for the assessment of BP, cardiac output, autonomic function and vascular disease. Since then, a large number of new studies have been published that also showed the potential of PPG for the monitoring of atrial fibrillation [27, 302, 200], sleep [73], respiration rate [152] and epilepsy [8].

These findings are exciting: the wealth of information that can be derived from a single, ergonomic, low-cost PPG sensor could lead one to think that this device is the swiss army knife of physiological measurement. However these metrics require highly advanced processing of the PPG signal that go beyond average heart rate monitoring. Most of the metrics require the measurement of heart rate variability (HRV) [2] or pulse wave morphology [63], which will be discussed below.

Heart rate variability

The variation of heart rate from one beat to another is called HRV and has been associated with autonomic nervous system activity [2]. HRV is determined by measuring the individual timing of the heart beats, which is called the inter-beat interval (IBI) trace. The detection accuracy of individual heart beats is usually done by comparing this trace with the ECG-derived heart beat trace.

How well PPG can be used as a proxy has been studied widely and many approaches have been proposed for it [255, 199, 284]. Most of the time, in these studies the accuracy of beat localisation is strongly hampered by motion artefacts [255]: the movement of the arm can result in irregular light scattering patterns and leakage of light which results in a disturbed PPG signal on which it is difficult to precisely localise the heart beats. For PPG, HRV is used for sleep stage classification [303, 310], atrial fibrillation [27] and BP monitoring.

Pulse morphology

Another source of physiological markers in PPG are the pulse morphology characteristics [63]. The theory is that the shape of a pulse wave observed in the PPG signal can reveal important information about vascular dynamics [64]. For example, the pulse wave can be decomposed into several components comprising the initial forward traveling wave from the heart and its many reflections across bifurcations and elastic buffers in the aorta. This is used to estimate left-ventricular ejection time [46] and systolic blood pressure (SBP) [210]. Another approach is to differentiate the PPG waveform into its first and second derivatives and study the resulting velocity and acceleration patterns as markers of autonomic activation and arterial stiffness [63]. Pulse morphology can only be performed on very clean signals, as even small artefacts can make pulse decomposition or second derivative computation prone to error.

Limitations of PPG for physiological tracking

As described above, PPG signals potentially contain a wealth of physiological information. However, most of the results that provide evidence for this have been done in a clinical setting where many noise-inducing factors are controlled for. These results may not be straight-forward to replicate in

Figure 1.1: Examples of a traditional transmissive PPG clamp (left) and a modern, connected watch with reflective PPG (right).

daily life with wrist-worn PPG. This context, which will be referred to throughout this manuscript as free-living context, differs significantly from the controlled clinical environment in which studies such as the ones reviewed by Allen [5] are performed.

First of all, body motion is a main source of noise in PPG [238, 239]: the movement of the arm can result in irregular light scattering patterns, leakage of light, movement of blood in the micro-vasculature and changes in orthostatic pressure due to the change in elevation of the wrist relative to the heart. While in most cases heart rate can still be measured despite these artefacts, the measurement of exact beat times and morphological properties of the beats becomes difficult: this is why the day-time coverage of PPG-based health metrics that rely on exact beat localisation tends to be low in comparison to night-time [27].

Another difference between clinical PPG and its application in wearable technology at home is that clinical PPG sensors come in the form of clamps around the finger (see Figure 1.1). These PPG sensors send light through the skin, which is known as transmissive PPG, in contrast to wrist-worn devices which always have reflective PPG as the wrist is too thick to send light through.

Finally, a last complicating factor is the interpretability of the sensor data. Many of the health parameters that could potentially be derived from PPG such as HRV, BP and sleep patterns, have classically been measured under strictly controlled circumstances [241] in clinical practice, and their normative value for what defines a healthy BP level or an adequate sleep pattern are only valid under such circumstances. Even measurements done by the user at home, such as manual cuff-based BP measurements, are recommended to be performed by users in a highly specific way [163] in order to prevent factors that may invalidate normative values such as physical exercise, eating, smoking and so on [144]. When shifting to a continuous monitoring paradigm such as with wearable technology, the requirement of controlled measurements cannot be satisfied and are incompatible with the ubiquitous, unobtrusive nature of wearables.

1.1.3 Wearable monitoring during sleep

As outlined in the previous section, the success of wearable devices has been mainly in the domain of activity tracking while measuring more advanced health parameters has been limited due to the limited reproducibility of results from lab experiments in a free-living context. However there is one part of daily life that might lend itself better to advanced PPG health monitoring: sleep. This is because unlike daytime measurements, measurements made during sleep do not suffer from many

1.1 Wearable Monitoring 19

Figure 1.2: PPG trace over a 24-hour period.

of the factors outlined in the previous section that limit the usefulness of PPG while at the same time providing a unique insight into human physiology and health.

Sleep is a period with very limited body motion and thus the presence of artefacts associated with body motion is reduced. In Figure 1.2 a PPG trace is shown over an entire 24-hour period.

In the topmost figure, it is visible that the largest variations in the PPG signal are seen throughout the day, which are motion artefacts. In the bottom part, the signal has been cleaned by applying a band-pass filter, where the artefacts are now more clearly visible on the signal. Zooming in, Figure 1.3 presents two representative samples of PPG from both the day and night periods, where the reduction in noise during sleep is now clearly visible. This reduced level of noise makes it possible to extract features beyond heart rate, which are too noise-sensitive to measure throughout the day. Such features could be related to HRV or morphology. Earlier work has already shown evidence that such features can be effectively extracted during the night from PPG. Bonomi et al.

[27] showed that IBI data during sleep can be used to detect atrial fibrillation while also showing that coverage for IBI data tends to drastically decrease during the day. Fonseca et al. [73] showed that HRV features can be extracted from the sleep IBI series to classify sleep stages, while Gil et al. [82] showed how fluctuations in PPG amplitude can be used to detect apnea events in children. In other work respiratory arousals by combining several features of the raw PPG in sleep apnea patients. Chua et al. [42] showed that the PPG amplitude also correlates with BP during sleep.

Another benefit of monitoring health during sleep is that sleep forms a natural baseline in the otherwise heavily confounded free-living context. This means that unlike in the case of daytime data, for nighttime it is easier to compare measurements of health from night to night and across dif- ferent people. In clinical practice, measurements are always made under certain baseline conditions by controlling for body position, meal intake, physical activity, alcohol and nicotine intake, and so on. During sleep all these factors are naturally controlled for. Sleep therefore provides the ideal conditions for meaningful health measurements from wearables that are hard to find during the day- time. The clinical utility of health measurements at night has already been expressed. For example in ambulatory BP monitoring [227] night-time data can provide unique prognostic parameters (see Section 1.3), even though measurements with ambulatory BP cuffs can be sleep-disturbing [54].

Finally, sleep makes up a third of human life and thus it provides a long window for the measurement of PPG. This allows to go beyond spot measurements: instead of obtaining single measurements such as done at the doctor’s office in current clinical practice, it is possible to track

0 2 4 6 8 10 0.3

0.2 0.1 0.0 0.1 0.2

Night-time

Normalized PPG during day and night

0 2 4 6 8 10

Time (s)

0.10 0.05 0.00 0.05 0.10

Day-time

Figure 1.3: Typical PPG signal during day and night. Day-time PPG is more prone to artefacts.

the dynamics and variabilities of the person’s health, which can reveal insights that are hard to get with short measurements. For example, night-time holter ECG can be used to detect the presence of paroxysmal arrhythmias which might not show up during a doctor’s visit. Until now however, the sleep period is rarely monitored, as most medical measurement equipment is obtrusive, expensive and may easily disrupt sleep. This is where wearable devices, being light-weight, inexpensive and unobtrusive, can make a difference in this area.

1.2 Sleep stage monitoring

The objective measurement of sleep in adult humans involves sleep staging: the process of segment- ing a sleep period into epochs, typically 30 seconds long, and assigning a sleep stage to each epoch.

The American Association of Sleep Medicine (AASM) [19] distinguishes [305] between two main types of sleep, rapid eye movement (REM) sleep and non-REM sleep, the latter is subdivided into three stages non-REM1 up to non-REM3. Sleep staging is usually performed by visually scoring the electric activity in the brain, eye movement and chin muscles, measured respectively with electroencephalography (EEG), electrooculography (EOG) and electromyography (EMG).

Together with sensors measuring cardiac and respiratory activity, this setup is collectively referred to as polysomnography (PSG).

The sleep hypnogram can reveal many insights about health. For example, the synaptic down- scaling that is part of the homeostatic sleep drive process [225] occurs during non-REM sleep. It has more specifically been linked to low-frequency oscillations in the EEG called slow waves [225]

which occur most frequently in non-REM3. In general however, the hypnogram is one of the main tools used for the diagnosis of sleep disorders [192].

As the measurement of the sleep hypnogram currently requires expensive and obtrusive equip- ment in the form of PSG and an expensive manual annotation procedure by a trained expert, this has led to its potential being under-utilised. PSG is only prescribed when there already is a high suspicion of sleep problems and not as a tool for general screening. In the NIHANES study of

1.2 Sleep stage monitoring 21 2005 [178] a survey was used to estimate the prevalence of sleep disorders in the United States, and the experimenters concluded that diagnosis rates were much lower than these estimates. More recently, evidence has been accumulating that sleep disorders may often be an underlying cause of cardiovascular disease: an under-diagnosis of sleep disorders was found in cardiology patients [43].

1.2.1 Automated sleep stage classification using EEG

A means to simplify the process of sleep hypnogram measurement could potentially close the diagnosis gap for sleep disorders, which is why this has been the goal of numerous studies in the last decades. Automation of EEG-based sleep stage scoring has been studied intensely. Many types of algorithms have been proposed, starting with rule-based approaches that aimed to mimic the rules of manual scoring from PSG [131] but eventually advancing to machine learning algorithms that automatically approximate the relation between the EEG and the sleep stage using for example artificial neural networks [103] or random forest networks [75]. Such approaches only use a single EEG channel rather than the many sensors included in a full PSG montage. Others attempted to only use EOG as this sensor is easier to wear [235]. Recently, using advanced deep learning algorithms that will be explained in detail in Chapter 2, the performance of such automated methods has become comparable to manual scoring [207, 211], making automated scoring very close to becoming a part of clinical practice.

1.2.2 Sleep stage classification from PPG

While automated scoring of PSG data already can increase speed and lower cost of sleep testing, it still requires an EEG system to be worn during sleep. Such a system is difficult to operate by untrained individuals, bulky and obtrusive, making it unattractive for prolonged use at home by people who are not severely suffering from sleep problems. A potential surrogate to PSG-based sleep stage classification is to use HRV as input signal. It has been known for decades that HRV markers of autonomic nervous system activity correlate with sleep stages [257, 274]. These markers are the well-known time domain and frequency domain features of the IBI series, such as:

• standard deviation of the IBI sequence

• standard deviation of successive difference of the IBI sequence

• high-frequency band power (0.15-0.4 Hz) of the IBI sequence

• low-frequency band power (0.04-0.15 Hz) of the IBI sequence.

These and similar other features are widely used as surrogate markers of autonomic nervous system activity in different applications [2]. A complete overview of features used in this thesis including references is presented in Table 2.3. It took some time until a first approach was pro- posed that classified sleep staging using only autonomic nervous activity. Redmond et al. [181, 182] proposed such an approach by combining several HRV features (extracted from ECG) and

respiratory features (extracted from a respiratory inductance plethsmopgrahy sensor in a machine learning model. Since then, many new features have been proposed and new models have been tested, leading to significant improvements in the performance, as reviewed in Section 2.1.

Even with all the improvements achieved in the past decade in HRV sleep stage classification, the methodology is still lacking behind in performance in comparison to EEG-based automated sleep stage classification. One of the main challenges is that the expression of sleep in HRV is different from the expression of sleep in EEG: while there are strong correlations between several frequencies in EEG and HRV, sometimes the correlations decrease. For example, autonomic disorders may alter HRV patterns [226] and a sleep stage classification algorithm should be flexible enough to deal with such alterations. Furthermore, the sleep stages are a rough discretisation of continuous sleep processes [28]. Sometimes, measurement noise may lead to transients in the sleep

Table 1.1: Reference values for the diagnosis of hypertension for office BP measurements[145].

Category SBP DBP

Optimal < 120 and < 80

Normal 120-129 and/or 80-84

High normal 130-139 and/or 85-89

Grade 1 Hypertension 140-159 and/or 90-99 Grade 2 Hypertension 160-179 and/or 100-109 Grade 3 Hypertension ≥ 180 and/or ≥ 100 Isolated Systolic Hypertension ≥ 140 and < 90

hypnogram. The AASM [305] annotation standard prescribes a set of post-processing rules to remove such transients on the basis of neighbouring epochs: for example, any REM epoch before the appearance of N2 sleep is re-classified as N1 [305]. Thus, a machine learning model should also have the capability to infer the sleep stage of a segment based on neighbouring segments. Next to that, a sleep stage dependent time delay has been observed between autonomic and cortical activity [230], increasing the need for proper temporal modelling.

1.3 Blood pressure monitoring

BP is one of the most frequently measured vital signs. Together with heart-rate and core body temperature, it forms the set of vital signs that are most telling of a human’s health. Typically, BP is measured through so-called spot measurement. By standard, spot measurements are taken at the brachial artery with an oscillometric cuff while the person is comfortably seated, at rest, not speaking, with uncrossed legs, supported back and arm, and with no clothing covering the upper arm at which the measurement is taken. Spot measurements are taken by either a clinician in the medical environment (the so-called office BP measurement), or by the patient at home (the home BP measurement). For special cases, the measurement is not done in the upper arm but on other locations. For example, in morbidly obese patients the upper arm is inaccessible for cuff measurement and a wrist cuff is used instead. The wrist cuff in turn introduces extra noise in the form of hydrostatic pressure variations (as a result of the wrist’s relative positioning to the heart) [172].

The spot measurement results in three values:

• SBP: this is the pressure after systole (i.e. the ejection of blood by the left ventricle of the heart into the circulatory system)

• Diastolic blood pressure (DBP): this is the pressure as the left atrium passively fills with returning blood

• Mean arterial pressure: calculated as DBP + (SBP-DBP)/3.

These values, when measured correctly, have a highly recognised prognostic value and define the diagnosis of hypertension. See Table 1.1 for an overview of how a measured BP value corre- sponds to a certain diagnosis. However the spot measurement is far from perfect.

1.3.1 Blood pressure as a continuous variable

External factors influence the spot measurement, such as room temperature, exercise, alcohol or nicotine consumption, positioning of the arm, muscle tension, bladder distention, talking and background noise [269]. The length and the width of a BP cuff also affects the measured pressure [172]. These factors influence the readings and lead to misdiagnosis to such an extent that the Euro-

1.3 Blood pressure monitoring 23 pean Society of Hypertension (ESH) and the European Society of Cardiology (ESC) recommend clinicians to formulate their diagnosis only after having repeatedly acquired consistent and clear results at two different visits and having checked the BP in both arms.

While confounding factors can all be adequately controlled by strict procedures of measurement in the clinic, there are also factors that cannot be controlled for. The white-coat effect [234] is a psychological effect on vital signs readings that occurs when the patient is seeing a doctor (in a white coat). The increased stress level of being under examination introduces a significant increase in BP values measured at the office. This is why the thresholds on what healthy BP levels are have been set slightly lower for home measurements than for office measurements. Another major limitation of the spot measurement is due to the highly normalised circumstances under which the spot measurement is taken. In practice, the patient might be exposed continuously to factors that increase the BP. This is referred to as masked hypertension [22].

Advances in BP measurement techniques have enabled the creation of automatic BP cuffs that periodically measure the BP by automatically inflating, deflating and processing the pressure values.

Technologies that rely on this principle are collectively referred to as ambulatory blood pressure monitor (ABPM) [145, 172]. The use of ABPM has led to the realisation that BP is not a constant but a continuously changing variable. In modern clinical practice, the 24-hour assessment of BP is frequently used to aid in diagnosis of hypertension and assessment of vascular health. Several parameters have been derived from the 24-hour assessment:

• 24-hour average of SBP and DBP

• daytime average of SBP and DBP

• nighttime average of SBP and DBP.

By averaging the value over a longer timeframe, most of the above-mentioned factors can be averaged out, resulting in a more reliable indicator of health risks [145]. Also the variation in BP throughout a 24-hour period has been associated with health risks and receives significant attention from the epidemiologic communities [68]. The general consensus is that blood pressure variability needs to be studied further with the help of new technologies for continuous measurement [68, 144].

1.3.2 The blood pressure dip

The ESH and ESC acknowledge the importance of BP variability [145] but while they do not recommend clinical use of most of the measures of BP variability such as BP standard deviation [87], the morning surge [117, 118] or the hyperbaric load [91, 92, 105], there is one parameter with strong enough evidence to be recommended for clinical use: the nocturnal BP dip.

The nocturnal BP dip is a relative decrease in both SBP and DBP during the night. The BP dipping pattern characterizes the circadian blood pressure variation profile [97] that is modulated by the circadian oscillations originating from the suprachiasmatic nucleus. The BP dipping pattern is expected in humans and its absence may indicate health risks as explained in the rest of this section.

The abscence of the BP dip may be caused by circadian misalignment [177], which happens when the circadian sleep drive does not align with the homeostatic sleep drive such as during shift work [155]. It could also be caused by conflicting time-of-day cues (e.g. changes in day-night pattern after travel over many time zones) that disrupt the circadian pacemakers in the suprachiasmatic nucleus, causing mixed signals to de-regulate circadian functions [61]. Finally, it could also be caused by arousals as a consequence of disrupted sleep, for example apnea events [139].

Relative BP dipping

Normative values for the BP dip have been established as fractional decreases compared to daytime, with BP values of 80% to 90% being regarded as optimal. Nighttime BP values that are higher are referred to as mild dipping (90%-100%) or non-dipping (≥100%), while stronger dipping patterns

< 80% are referred to as extreme dipping or over-dipping. It has also been regarded as an absolute decrease in mmHg of SBP [23, 95].

Non-dipping has been repeatedly associated with an increased risk of cardiovascular disease.

A non-dipping profile has been associated with obstructive sleep apnoea, obesity, high salt intake, orthostatic hypothension, autonomic dysfunction, chronic kidney disease, diabetic neuropathy and old age [145].

Absolute BP dipping

The absolute decrease in BP during the night has also been found to be a predictor of both cardiovascular and non-cardiovascular mortality [23], independent of other measures of BP such as the spot measurement or 24-hour average. More recently, it was shown that an absolute dip in SBP of 5 mmHg in the night is associated with a 17% reduction of cardiovascular risk [95]. Moreover, it was shown that administering anti-hypertensive drugs before sleep in people who do not elicit a sufficient level of nocturnal BP dipping reduces risk of cardiovascular disease and chronic kidney disease [94].

1.3.3 The measurement of BP dipping with PPG

The measurement of BP during sleep requires the ABPM procedure, which is highly obtrusive and sleep-disturbing [54]. This is why it is underused as a clinical parameter, despite official guidelines recommending its use due to the strong supporting evidence [145]. Fortunately, the BP dipping pattern is an ideal candidate for PPG-based monitoring. This is because the BP dip manifests in the PPG in two different ways: first of all, the dip during sleep is predominantly steered by sympathetic nervous system activity [196]. The sympathetic nervous system is one of the two parts of the autonomic nervous system, which as mentioned earlier is often measured with HRV features.

Second of all, morphological analysis of the PPG pulse [63] can be done to quantify hemodynamics that are correlated with BP such as vasoconstriction, arterial bifurcation and pulse reflections.

1.4 Machine learning

In the previous sections, an overview was given of the type of physiological characteristics present in wearable sensors that may be used to monitor health parameters such as the sleep hypnogram and the circadian blood pressure dip. For example, body movements from accelerometry, HRV from PPG as well as PPG pulse morphology. However, how exactly these characteristics map to certain sleep stages or blood pressure values has not been discussed. It is possible to study the correlation between individuals characteristics and the target variable and establish such models, however when there is a large number of characteristic features available it becomes difficult to manually construct a function that maps all the characteristics to the target. Luckily, these functions can be approximated computationally by examining the patterns in example data sets. Algorithms that do this are known as supervised machine learning algorithms.

1.4.1 Supervised machine learning

A classic supervised machine learning problem can be formulated in terms of a domain and a prediction task.

1.4 Machine learning 25

Figure 1.4: Simple artificial neural network. Source: Wikimedia Commons.

The domainD, which for example could be the data that is collected from the wearable sensor, is a collection of N features. These features make up an N-dimensional feature space denoted as X . The features can be considered as a set of stochastic variables with a marginal probability distribution P(X ) where X = X₁, · · · , XN∈X . The domain is then expressed as a combination of bothX and P(X) as D = {X ,P(X)}.

The classification taskT is the task that needs to be performed on D, e.g. the prediction of sleep stages or blood pressure. The target variables make up the label space, denoted asY . The labels can also be denoted as stochastic variables that are conditioned on the feature space: P(Y |X ), where Y = Y1, · · · ,YL∈Y where L is the number of labels to be predicted. The prediction task is then expressed asT = {Y ,P(Y|X)}. The task T is learned from example data. Example data consists of observed value pairs xi and yiwhere xi∈ X and yi∈ Y and i denotes the i^thpair in the training data set.

Classic machine learning algorithms include:

• decision trees [260] that represent the machine learning task as a tree-like model of decisions based on xito produce a prediction ˆyi of the corresponding label.

• support vector machines [212] that represent the task as a number of hyperplanes inX that separate samples from different labels.

• artificial neural networks that model the label prediction task as an inter-connected group of nodes that are inspired by simplified models of biological neurons in a brain.

An example of an elementary artificial neural network with a perceptron node is shown in Figure 1.4. The figure shows how a set of input features are multiplied by weights wi j where i denotes the feature number while j denotes the perceptron number (multiple perceptrons can be trained simultaneously and later aggregated). The result is aggregated through summation into netj, corresponding to the net input of perceptron j. Finally, the linear output can undergo an additional transformation known as an activation function. Example activation functions are sigmoid or rectified linear functions. This is done to enable the perceptron to approximate non-linear relations.

By stacking neurons it is possible to create very versatile function approximators. Very deep architectures are referred to as deep artificial neural networks. Perceptrons as the one shown in Figure 1.4 are just an example of a neural node. In the imaging field for example, convolutional

nodes are very popular due to their unprecedented capacity to recognise visual patterns [128].

1.4.2 Sequence-to-sequence modelling

In Section 1.2.2 it was highlighted that state-of-the-art approaches to sleep stage classification from HRV are not capable of capturing temporal dynamics. A promising way to model temporal dynamics adequately is through sequence-to-sequence modelling, in which a sequence of inputs are mapped to a sequence of outputs. In such models, observed values are not shaped as a vector of features, but as a matrix of both features and time steps, the output is shaped as a matrix of both labels and time steps. If the total number of time steps in a sequence is T , then the input and output variables can be represented as:







X_1,1 · · · X_1,T ... . .. ... X_N,1 · · · X_N,T





 ∈ X







Y_1,1 · · · Y_1,T ... . .. ... Y_L,1 · · · Y_L,T





 ∈ Y

This allows the model to make use of the entire input series when predicting the output series, regardless of the time step. This way, temporal relationships in the sequences are exploited. There are many types of sequence-to-sequence models, however most of them are recurrent algorithms that can be embedded as nodes in artificial neural networks, which allow the model to infer over sequential data [286].

One of the most flexible recurrent neural nodes are the long short-term memory (LSTM) cells.

LSTM cells generate an output ht based on its input xt, its last output ht−1(short-term recurrence) and its internal cell state Ct (long-term recurrence). The internal memory state C has dedicated variables to store each of the inputs x as well as its own prediction h with an input gate it to control input to the memory and a forget gate ft to clear the memory. The behavior of all these variables is trained through weight vectors. An overview of an LSTM cell is shown in Figure 1.5. At each time step the LSTM computes the gate values:

ft= σ (Wf· [h_t−1, xt] + bf) (1.1)

it= σ (Wi· [h_t−1, xt] + bi) (1.2)

where σ () denotes the sigmoid activation function and bf, bi, bC denoting a trainable bias term.

Then new candidate values ˜Ct are proposed for the cell state and the cells are updated through:

C˜_t= tanh(WC· [h_t−1, xt] + bC) (1.3)

Ct= ft∗C_t−1+ it∗ ˜Ct (1.4)

The final output is determined as:

h_t = (Wo· [h_t−1, xt] + bo) ∗ Ct (1.5)

The trainable parameters of an LSTM cell are the weight vectors Wo,Wi,WCand Wf together with the respective bias terms. In this thesis, LSTM based neural networks will form an important part of the proposed solutions for continuous monitoring of sleep and blood pressure.

1.4 Machine learning 27

Figure 1.5: Long short-term memory cell. Source: Wikimedia Commons.

Figure 1.6: General scheme for training machine learning models. Source: Wikimedia Commons.

Model training

Training supervised machine learning models is done on labeled observations. The general scheme is shown in Figure 1.6. A data set of labeled observations is split into two parts: a training set and a testing set. The training set is used by the machine learning algorithm to learn the model’s structure (typically as model weights). Optimisation is done by minimising a loss function. With artificial neural networks, training is an iterative process in which the data set is presented multiple times to the learner algorithm. Each time, the algorithm optimises the model’s weights further through back-propagation. Back-propagation computes the gradient of the loss function with respect to the weights of the network, and then updates the weights to minimise the loss function, typically through a gradient descent algorithm.

The resulting model is tested by predicting the labels for data in the testing set. It is important that the testing set has not been seen before by the learner function to ensure that the test results gen- eralise to new data. In addition, when dealing with data collected from humans (such as wearable sensor data), it is important that all data points of a single individual belong either exclusively to the training set or the test set, to ensure that the results represent the scenario in which a new individual

Figure 1.7: Illustration of the cross-validation scheme. Source: Wikimedia Commons.

of whom no data was seen before will use the sensor. The quality of these predictions is quantified to determine the model’s effectiveness at the machine learning task. Quantifying the model’s prediction quality is done through evaluation metrics. These differ from one problem domain to another. An overview of common evaluation metrics that are used throughout this manuscript is given in Appendix B.

The drawback of the train-test split approach is that the amount of data left to evaluate the model’s performance is limited, making it prohibitive to draw strong conclusions or perform formal statistical testing on the results. To alleviate this problem, cross-validation can be performed. An overview of cross-validation is presented in Figure 1.7. In n-fold cross-validation, the data is divided into n subgroups, and the model training procedure is repeated n times, each time using a different subgroup as testing set. This allows testing the performance of the model on the entire available set of observations.

1.4.3 Transfer learning

A challenge with training deep neural networks with a large number of weights is that a very large amount of data is needed to make the algorithm converge to a weight configuration with a high accuracy and generalisability to unseen data. This is especially a challenge in domains where data needs to be collected prospectively. This is the case with wearable PPG data, since such sensors are relatively new and their signals depend heavily on the specific make. Luckily, features derived from PPG are comparable to the ones derived from more well-established sensors. For example, HRV measures can also be derived from ECG. This enables the use of older ECG data sets to train such a model. However, when extracted from ECG, the features will not be equivalent: heart beats take time to arrive from the heart to the location where the sensor is worn. This time is known as pulse wave velocity (PWV) [30] and is known to vary under the influence of BP and arterial vessel properties (see Chapter 4 for more background and experimental results on PWV). This variable time delay causes HRV features to be different when compared across ECG and PPG. In other words, if the data containing labeled ECG observations is regarded as the source data set S and the labeled PPG observations are regarded as the target T , then P(XS) 6= P(XT) as a consequence of a shift in the domain’s feature values [162]. In this case a domain adaptation is needed.

1.5 Lifestyle change 29 In addition, the machine learning task may change itself, for example when the demographics or annotation of the older ECG dataset is different from the new PPG dataset, in other words, also P(YS|X_S) 6= P(YT|X_T) [162]. In this case an adaptation of the decision boundaries is needed.

Such adaptations can be achieved by a technique known as transfer learning [162]. With this technique, the entire neural network is trained using observations from a source dataset, and later an adaptation is performed using observations from the target dataset. Since the number of observations in the target dataset is limited in comparison to the source dataset, usually the majority of the model’s weights are frozen, and tuning is only done on a small subset of the model’s weights.

By limiting the degrees of freedom, it is possible to make the model converge for the target problem.

1.5 Lifestyle change

While the majority of the work presented in this thesis concerns wearable continuous monitoring of physiology and personal health, by itself the monitoring technology is only the first step towards lifestyle change. When risk factors are identified, the next step is for individuals to take action by making appropriate changes to lifestyle. For hypertension, lifestyle interventions are associated with the reduction or even normalisation of elevated BP [57, 145]. Of course not all elevation in BP is attributable to lifestyle. There are genetic factors that can explain elevated BP [122] and possibly can even mediate the efficacy of lifestyle interventions [86, 218]. However there is a wide consensus in guidelines of hypertension management that lifestyle interventions can play a strong role in the management of the disease. These interventions are diverse, comprising many behavioral categories: salt restriction, moderation of alcohol consumption, a diet rich in vegetables, fruits and low-fat dairy products, weight reduction, regular exercise and smoking cessation. When individuals engage in such interventions, they can change their habits [65] and reduce their BP [124], lasting for up to four years after the intervention was delivered.

Even though the benefits of lifestyle interventions are undeniable given previous research, the cost of behavior interventions and lifestyle counselling is still high and healthcare institutions are not well-equipped to deliver such interventions at the scale to serve all who need them. As such, internet technology and mobile devices have been welcome platforms for computer-assisted lifestyle interventions that aim to provide virtual support in the lifestyle change process. Several reviews exist outlining the effectiveness and potential of e-counseling [135], mobile and internet interventions [4, 17] and self-care applications [136, 146]. The category of self-care in particular can be highly cost-efficient. While self-care apps cannot substitute human counselling, they may be used in situations where people have limited access to such counselling. In general, lifestyle interventions are regarded as low-risk and physicians encourage self-care [111]. There are a few issues however that need to be addressed. For once, some exceptions could be associated with health risk, such as for example the risk of injury in high intensity training [222] and the risk of the valsalva maneouver in resistance exercise training [314]. Next to that, some interventions are difficult to perform, e.g. following a correct and balanced diet [13, 311, 313]. Finally, individuals may also be unwilling to change habits [111].

1.5.1 Feasibility

The number of lifestyle interventions for hypertension management is high and it can be daunting for a patient or an at-risk person to take in all the advice and effectively change all their habits at once, especially given the high likelihood that they have led a markedly unhealthy lifestyle for a long time to cause the elevated BP. One way to approach this problem is by selecting and proposing

interventions that have a higher chance of being adopted by the user. In classic (e-)counseling this is done based on domain knowledge of the counsellor. However when there is no counsellor available, such tailored recommendations are not accessible. It has been repeatedly suggested that the difficulty of the lifestyle intervention is an important determinant for the willingness of people to change their lifestyle [13, 311, 313].

1.5.2 Safety

Physical activity remains one of the key behaviors to prevent and combat lifestyle-related disease, in particular hypertension [316, 169]. Wearable technology has been well-known for tracking aerobic physical activity (e.g. Jakicic et al. [110]), however aerobic exercise is only half the story:

the American College of Sports Medicine (ACSM) [292] distinguishes between different types of physical exercise, of which the main two are aerobic and resistance exercise. For hypertension and other cardiovascular disease, the ACSM recommends a combination of both types of exercise [301, 314]. The prescription of the ACSM is to combine daily aerobic exercise with three or more sessions of resistance exercise per week [292]. Recent evidence [316, 170] suggests to combine resistance and aerobic exercise in a single session, which is referred to as combined exercise, however this has not been incorporated in current recommendations yet. A particular challenge for hypertensive individuals is exercise safety [292, 222]: static exercise [316] may cause a strong increase in BP due to the valsalva manoeuvre [140]: an attempted exhalation against a closed airway during the exercise. Thus, there is not only a need for tools that help hypertensive individuals meet recommended amounts of resistance exercise, but also tools that can simultaneously guide the user in performing the exercise safely.

1.6 Overall project aims

The overall aim of this project is to explore how data-driven algorithms can be used to build a physiological, mental and behavioral model of an individual that can be used to provide continuous personal monitoring information to the user about their health, provide personalised lifestyle recommendations and guide patients in executing lifestyle interventions. An overview of this overall aim and how the different chapters contribute to the aim is shown in Figure 1.8.

The majority of the work addresses the continuous personal health theme, in which methods are proposed and evaluated for the monitoring of the sleep stages throughout the night (Part 1) and the monitoring of blood pressure variations throughout day and night (Part 2) using a wearable PPG sensor technology. Both parts build upon a vast amount of prior art concerning the modelling of physiology from HRV, PPG morphology and accelerometry modalities and propose to improve upon the state-of-the-art by introducing deep learning algorithms, specifically LSTM neural net- works, to exploit the temporal patterns in these modalities to more accurately model and predict these physiological processes.

Part I splits up in two chapters. In Chapter 2 the problems with existing machine learning methods to perform temporal modelling for sleep stage classification from HRV are elaborated and it is proposed to solve these challenges with the LSTM model. The approach is evaluated in a large data set and it is shown how this modelling technique achieves state-of-the-art performance using HRV features derived from ECG data. In Chapter 3 a transfer learning technique is evaluated to adapt the model from the earlier chapter to PPG data. The method is compared to baselines such as only pre-training the model on ECG, as well as training the model directly on the (limited) PPG dataset without pre-training on ECG, and it is shown that the transfer learning approach outperforms