Exploring Mechanocardiography as a Tool
to Monitor Systolic Function Improvement
with Resynchronization Pacing
University of Turku
Department of Computing
Master of Science Thesis
Digital Health Technology
March 2024
Fadime Tokmak
The originality of this thesis has been checked in accordance with the University of Turku quality assurance
system using the Turnitin OriginalityCheck service.
UNIVERSITY OF TURKU
Department of Computing
Fadime Tokmak: Exploring Mechanocardiography as a Tool to Monitor Systolic
Function Improvement with Resynchronization Pacing
Master of Science Thesis, 51 p.
Digital Health Technology
March 2024
The thesis explores the utilization of mechanocardiography (MCG) as a novel ap-
proach to assess and quantify improvements in systolic cardiac function resulting
from cardiac resynchronization therapy (CRT). The study focuses on patients with
heart failure and reduced ejection fraction (HFrEF), a population commonly treated
with CRT. The primary objective is to investigate the differences in MCG waveforms
during CRT and single-chamber atrial (AAI) pacing, specifically comparing wave-
form characteristics. 10 patients with heart failure and previously implanted CRT
pacemakers were included in the study. The MCG and ECG signals are recorded
using accelerometers, gyroscopes, and Holter measurement unit placed on the lower
chest and used in the analysis. ECG and MCG recordings were obtained during
both CRT and AAI pacing at a consistent heart rate of 80 beats per minute. The
analysis involved considering six MCG axes and three MCG vectors across vari-
ous frequency ranges to derive key waveform characteristics such as energy, vertical
range, electromechanical systole (QS2), and left ventricular ejection time (LVET).
The results revealed significant differences between CRT and AAI pacing, with CRT
consistently exhibiting higher energy and vertical range during systole across multi-
ple axes. Notably, the study identified optimal differences in SCG-Y, GCG-X, and
GCG-Y axes within the 6–90 Hz frequency range. However, any difference in QS2,
LVET and waveform characteristics around aortic valve closure was not identified
between the pacing modes.
The findings suggest that MCG waveforms can serve as indicators of improved me-
chanical cardiac function during CRT. The use of accelerometers and gyroscopes
may contribute to the development of a non-invasive and potentially predictive
tool for optimizing CRT settings. The promising results underscore the need for
further research to explore the differences in signal characteristics between respon-
ders and nonresponders to CRT. The overall aim is to enhance the clinical applica-
tion of MCG, leveraging wearable technology and micro-electromechanical systems
(MEMS), and ultimately improve the optimization and efficacy of CRT in heart
failure (HF) management.
Keywords: accelerometer, gyroscope, heart failure, resynchronization therapy, smart-
phone
Contents
1 Introduction 1
1.1 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Thesis content summary . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Background 7
2.1 Heart Failure (HF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Cardiac Resynchronization Therapy (CRT) . . . . . . . . . . . 8
2.2 Mechanocardiography (MCG) . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Electrocardiogram (ECG) . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Statistical Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . 13
2.4.1 Parametric Tests . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.2 Nonparametric Tests . . . . . . . . . . . . . . . . . . . . . . . 14
3 Related Work 16
3.1 Filtering of MCG Signals . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1.1 Frequency Filters . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1.2 Wavelet Transform Approach . . . . . . . . . . . . . . . . . . 18
3.1.3 Synchronized Averaging Approach . . . . . . . . . . . . . . . . 19
3.2 Fiducial Point Detection on MCG Signals . . . . . . . . . . . . . . . 19
3.2.1 Search Windows . . . . . . . . . . . . . . . . . . . . . . . . . 19
i
3.2.2 Dynamic Time Warping . . . . . . . . . . . . . . . . . . . . . 20
3.2.3 Hidden Markov Model . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Assessment of Cardiac Function with MCG Signals . . . . . . . . . . 22
3.3.1 Cardiac Performance . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.2 Cardiac Time Intervals . . . . . . . . . . . . . . . . . . . . . . 23
4 MECHANO-HF Dataset 25
5 Proposed Method 28
5.1 Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.1 Signal Resampling . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.2 Noise Removal . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.3 R Peak and Q Wave Detection . . . . . . . . . . . . . . . . . . 29
5.1.4 Aortic Valve Opening & Closing Detection . . . . . . . . . . . 30
5.1.5 Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.6 Vector Creation . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.7 Synchronized Averaging . . . . . . . . . . . . . . . . . . . . . 32
5.2 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2.1 Morphological Features . . . . . . . . . . . . . . . . . . . . . . 33
5.2.2 Timing Features . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3 Statistical Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6 Results 38
6.1 Morphological Features . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.1.1 Systolic Features . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.1.2 Early Diastolic Features . . . . . . . . . . . . . . . . . . . . . 40
6.2 Timing Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7 Discussion 42
ii
8 Conclusion 50
References 52
iii
List of Figures
2.1 Example SCG signal from MECHANO-HF database (Z axis). . . . . 10
2.2 Example ECG signal from MECHANO-HF database. . . . . . . . . . 12
4.1 The device is positioned on the lower part of the patient’s chest,
recording 3-axis acceleration, 3-axis rotation, and a single-lead ECG
while the patient is in a supine position. This illustration has been
adapted from the original figure featured in the article [24] under the
license CC-BY 4.0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1 Example SCG-Z signal from MECHANO-HF database before filtering
(left) and after filtering (right). . . . . . . . . . . . . . . . . . . . . . 29
5.2 Example ECG signal with detected Q waves and R peaks in a heart
failure patient undergoing CRT pacing (left panel) and AAI pacing
mode (right panel). . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3 Example mean beat of SCG vector in a heart failure patient under-
going CRT pacing (left panel) and AAI pacing mode (right panel)
based on AO averaging. . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.4 Definitions of extracted features in a heart failure patient undergoing
CRT pacing (upper panel) and AAI pacing mode (lower panel) based
on AO averaged SCG-Z cycle. This illustration has been altered from
the original figure in the article [34] and is being utilized with explicit
permission. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
iv
5.5 The distribution of segments to various sets involves putting the fea-
tures of the first, second, and third segments of the corresponding
mode into sets 1, 2, and 3, respectively, for each subject. The il-
lustration presented here is a modified version of the original figure
depicted in the article [34], used with permission. . . . . . . . . . . . 36
5.6 Algorithm for data analysis. Each individual paired set underwent
a statistical test, and the null hypothesis was rejected if all parallel
tests rejected the null hypothesis. The illustration shown here has
been adjusted from the original figure in the article [34] and is being
employed with permission. . . . . . . . . . . . . . . . . . . . . . . . 37
6.1 Changes in systolic energy and vertical range for each axis, filtering
frequency and synchronization method. Adapted from the original
figure featured in the article [34], this illustration has been modified
and is utilized with permission. . . . . . . . . . . . . . . . . . . . . . 39
6.2 Mean systolic energy and range among subject during CRT and AAI
mode extracted with AO point synchronization method in the 20-90
Hz range. This illustration originates from the original figure featured
in the article [34] and is being used with permission. . . . . . . . . . . 41
7.1 Mean GCG-X beats extracted with R peak and AO point synchro-
nization for CRT (left column) and AAI (right column) mode with all
individual beats. Individual beats are shown with the lighter colors. 45
v
List of Tables
4.1 Baseline characteristics of patients included in the study. (*After
introduction of optimal medical treatment and CRT) . . . . . . . . . 25
6.1 Mean and standard deviation of QS2 and LVET during CRT and AAI
mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
vi
1 Introduction
Cardiovascular disease (CVD) poses a common and complex public health challenge,
with a prevalence of 48.6% in adults aged 20 years and older and impacting 127.9
million individuals in 2020 in United States only [1], [2]. This prevalence rises
with age in both males and females, and it includes conditions such as coronary
heart disease (CHD), heart failure (HF), stroke, and hypertension [3]. Excluding
hypertension, the overall CVD prevalence is 9.9% [2]. Significantly, long-term fatal
CVD risks are associated with male gender, older age, and smoking, based on current
literature [4].
Heart Failure (HF) is a complex clinical syndrome characterized by the inability
of the heart to pump blood efficiently and leads to insufficient oxygen supply to meet
the body’s needs [5]. According to data from NHANES (2017-2020), approximately
6.7 million Americans aged 20 and above had HF, with a projected increase to over
8 million by 2030 [2]. The prevalence of HF is expected to rise by 46% from 2012
to 2030 in the population, including individuals aged 18 and above [6].
Various factors contribute to the development of HF, including cardiometabolic
factors such as CHD, hypertension, diabetes, obesity, and smoking [7]. The lifetime
risk of HF increases with age [8]. The mortality associated with HF has improved,
primarily due to evidence-based treatments for HF with reduced ejection fraction
(HFrEF) [9]. HF causes a significant economic burden, with high rates of hospital-
izations, emergency department visits, and resulting costs [6]. Improving risk factor
CHAPTER 1. INTRODUCTION 2
management and following evidence-based treatments are crucial in addressing the
growing burden of HF.
Cardiac resynchronization therapy (CRT) devices can be used to manage HF
[5], [10], [11]. They are connected to the heart with three wires. They aim to
enhance ventricular activation, detect heart irregularities, and correct them with
electrical pulses, resulting in improved cardiac function [11], [12]. Supported by
extensive clinical trials, CRT is beneficial for moderately and severely symptomatic
HF patients.
In individuals with sinus rhythm, low ejection fraction (EF ≤ 35%), and QRS
duration of ≥ 120 ms (preferably with left bundle branch block morphology), CRT is
recommended for those who are expected to survive with good functional status for
over a year and aims to reduce the risk of HF hospitalization and premature death
[13]–[15]. It has been shown that it is effective in improving symptoms, quality of
life, and ventricular function and is a valuable option in the management of heart
failure.
Clinical trials evaluating CRT effectiveness in diverse patient groups have shown
significant recovery and reverse remodeling in the majority of heart failure (HF)
patients, but about one-third do not respond positively [16]. Accordingly, the CRT
management process involves crucial follow-up stages to identify factors predicting
positive outcomes and address non-responders [17]. Assessing CRT response in-
cludes measures like left ventricle volumetric changes, quality of life assessments,
and functional capacity evaluations [17]–[19].
Seismocardiography (SCG) was first introduced by Salerno et al. in 1990 as a
noninvasive displacement cardiography technique [20]. SCG is a technique based
on 3-axis accelerometers placed on the sternum, and it aims to capture vibrations
resulted by cardiac contraction, hemodynamic activity, and valve movements [21].
Gyrocardiography (GCG) is another technique invented by Meriheinä et al. [22]
1.1 RESEARCH QUESTIONS 3
and uses a gyroscope to capture three-dimensional angular velocity on the sternum,
providing information about cardiac vibrations [23]. Mechanocardiogram (MCG) is
a methodology that includes both SCG and GCG complementarily to acquire 6-axis
information regarding heart vibrations [24].
The increasing popularity of smart mobile devices, equipped with 3-axis ac-
celerometers and gyroscopes, has made SCG and GCG promising monitoring tech-
niques. Together, they form MCG, which can be gathered in home settings, offering
a potential avenue for remote monitoring of individuals with cardiological conditions.
Literature demonstrates that these vibrations contain crucial information about car-
diac time intervals, cardiac performance, hemodynamic variables, and heart diseases
[25]–[28].
Existing studies have showed the potential usefulness of MCG in identifying
atrial fibrillation, aortic stenosis, and evaluating conditions associated with cardiac
resynchronization therapy, heart failure, and myocardial infarction [26], [29]–[33]. In-
creasing evidence supporting the utility of MCG for heart monitoring holds promise
for its application in patient selection, optimization, and follow-up procedures for
CRT.
1.1 Research questions
This study focuses on the effect of CRT on MCG waveforms by comparing waveform
features during AAI (single chamber) and CRT (dual chamber) pacing. We aim to
address the following research questions in evaluating the waveform characteristics
during CRT and AAI pacing using MCG signals:
• Are there differences in MCG waveform characteristics that are consistently
observable between CRT and AAI pacing modes in patients with HFrEF?
• Which cardiac events show differences following the initialization of CRT
1.2 CONTRIBUTIONS 4
mode?
• Do specific MCG axes and frequency ranges provide differentiation between
CRT and AAI pacing?
• Are there any significant changes in timing parameters, including electrome-
chanical systole (QS2) and left ventricular ejection time (LVET), between CRT
and AAI pacing?
• How do different synchronization methods, such as R-peak and AO point syn-
chronization, impact the analysis of MCG signals in assessing mechanical car-
diac function during CRT and AAI pacing?
The results of this study contribute to understanding the potential of MCG in
detecting and assessing improvements in cardiac mechanical function during CRT.
The findings may have implications for predicting clinical response to CRT and
optimizing pacing settings for patients with HFrEF.
1.2 Contributions
In this study, we present a comprehensive investigation of the waveform character-
istics during CRT and AAI pacing using MCG signals in patients with HFrEF. The
primary contributions of this study are as follows:
• A detailed analysis of waveform features, focusing on energy and vertical range
during systole and early diastole was conducted. It is shown that energy and
range during systole significantly differ between CRT and AAI pacing, reflect-
ing improved systolic performance caused by resynchronization of ventricles.
• QS2 and LVET during CRT and AAI pacing were extracted to compare the
timing parameters associated with these pacing modes.
1.3 THESIS CONTENT SUMMARY 5
• Specific MCG axes and frequency ranges that exhibit optimal differences be-
tween CRT and AAI pacing modes were identified. These findings contribute
to understanding which parameters are most informative in reflecting the im-
provements.
• The impact of different synchronization methods, including R-peak and AO
point synchronization, on the analysis of MCG signals was evaluated. This
provided valuable insights into the robustness and consistency of MCG analysis
under varying synchronization conditions.
• Analysis of the MCG signals while also incorporating the SCG, GCG, and
MCG vectors was conducted. A new algorithm for MCG vector extraction,
accounting for potential range variations between SCG and GCG measurement
units was introduced.
• To address the constraints of a limited patient cohort and account for time-
dependent differences in the analysis, a specialized data analysis algorithm is
introduced.
In summary, this study provides a novel way for mechanical cardiac function
assessment during different pacing modes and offers new insights for future investi-
gations in optimizing CRT for patients with HFrEF. This thesis is connected to the
accepted publication [34].
1.3 Thesis content summary
Later sections of the thesis are structured as follows: Chapter 2 offers the back-
ground, covering topics such as heart failure (HF), cardiac resynchronization ther-
apy (CRT), mechanocardiography (MCG), electrocardiogram (ECG), and statistical
hypothesis testing. Chapter 3 delves into related work, exploring filtering of MCG
1.3 THESIS CONTENT SUMMARY 6
signals, fiducial point detection on MCG signals, and assessment of cardiac function
with MCG signals. Chapter 4 provides information on the MECHANO-HF dataset,
while Chapter 5 outlines our proposed method in detail, including signal processing,
feature extraction, and statistical tests. Chapter 6 presents the results and Chapter
7 provides a thorough discussion of the findings. Finally, Chapter 8 concludes the
thesis.
2 Background
2.1 Heart Failure (HF)
Heart failure (HF) is a leading cause of death worldwide and is a syndrome in
which the heart loses its ability to maintain ventricular filling and expulsion of
blood at the level required for metabolic needs [5]. HF causes physical symptoms
like shortness of breath, fatique, edema and chest pain [35], [36]. The decrease
in the cardiac performance is related with the loss of functional myocardial cells
and etiology of the loss can be various [37]. The severeness of loss in the cardiac
output is important information that affects to the prognosis and treatment of heart
failure patients which are classified according to their left ventricular ejection fraction
(LVEF) as HFrEF, HFimpEF, HFmrEF, HFpEF (from the highest loss to lowest loss)
[5]. Patients with HFrEF have 40% or lower LVEF which means heart is capable to
eject only 40% or less of the blood filled during diastole to the ventricles [38].
In addition to the classification based on LVEF, it is essential to consider the
underlying causes of heart failure for a comprehensive treatment approach. The
etiology of heart failure can be diverse, including ischemic heart disease, hyper-
tension, valvular disorders, and inflammatory cardiomyopathies, among others [37].
Tailoring treatment to address the specific underlying cause is crucial for optimiz-
ing patient outcomes. Furthermore, the presence of accompanying cardiac abnor-
malities, such as mechanical and electrical dyssynchrony, necessitates specialized
2.1 HEART FAILURE (HF) 8
interventions. In cases where dyssynchrony is identified, implantable devices for
synchronization and pacing, such as cardiac resynchronization therapy (CRT) de-
vices, may be employed to improve the coordination of ventricular contractions and
enhance overall cardiac function [5]. Thus, a thorough understanding of both the
classification based on LVEF and the underlying etiology allows healthcare profes-
sionals to formulate personalized treatment strategies for heart failure patients. The
treatment of HFpEF usually considers managing the underlying diseases behind it.
2.1.1 Cardiac Resynchronization Therapy (CRT)
Left ventricular (LV) dyssynchrony is a common pattern observed with HF and ap-
proximately one third of HF patients have left bundle branch block (LBBB) which
causes electrical and mechanical delay between left and right ventricular contrac-
tion [39]–[41]. This impaired contraction decreases the efficiency of contraction and
cardiac output [11]. Cardiac resynchronization therapy (CRT) is an established
strategy that is based on implantable devices and used to deal with HF cases with
impaired cardiac function during the last decades [5], [10], [11].
CRT devices are small devices that are connected to the heart with three wires
and modify the impaired activation of ventricles, sense anomalies in contractions,
and correct irregular patterns by radiating necessary electrical pulses [11], [12]. Re-
versal of heart irregularities provides efficient contraction and improvements in car-
diac function [42].
The effectiveness of CRT on different patient groups has been evaluated in differ-
ent clinical trials [43]–[46]. In these trials, the majority of HF patients have experi-
enced significant recovery and reverse modeling. However, approximately one-third
of patients could not utilize the CRT and achieve a positive response [47].
The CRT management process encompasses a crucial follow-up stage designed
to identify factors predicting positive clinical outcomes and develop approaches for
2.2 MECHANOCARDIOGRAPHY (MCG) 9
addressing non-responders. This includes selecting suitable patients, implementing
effective implantation strategies, and optimizing therapy benefits through program-
ming. However, there is a lack of consensus about the prevalence of non-response
and treatment strategies for CRT non-responders.
In assessing CRT response, various measures are utilized, including volumetric
changes of the left ventricle, quality of life (QOL) assessments, and functional capac-
ity evaluations [17]–[19]. Clinical trials and real-world practices often use different
criteria to evaluate responses, with trials emphasizing event-driven endpoints like
heart failure hospitalizations and mortality, while real-life situations prioritize the
overall well-being of patients [17]. Event-driven endpoints such as heart failure hos-
pitalization are crucial in large-scale trials [13], [48], but may not be as meaningful
in individual clinical practice [17].
2.2 Mechanocardiography (MCG)
Seismocardiography (SCG) is a method that uses accelerometers placed on the chest
to track the linear acceleration sourced by cardiac vibrations, whereas gyrocardio-
graphy (GCG) refers to the use of gyroscopes to track angular acceleration [21],
[23]. Mechanocardiography (MCG) is a joint concept that includes both SCG and
GCG to acquire a 6-axis representation of the mechanical activity on sternum with
combining mutually orthogonal signal sources [24], [49]. The acquired vibrations
thorough these sensors are sourced by contraction of the heart muscles, activies of
heart valves, and the ejection of the blood and it is shown that these vibrations com-
prises information about hemodynamic variables, timing of different cardiac events,
physiological and pathological state of heart in various different aspects [23], [26],
[28], [50], [51]. An example for SCG waveform is represented in Figure 2.1.
In the literature, the waveform annotations on SCG and GCG signals are stud-
ied through simultaneous recordings of these modalities with echocardiography and
2.2 MECHANOCARDIOGRAPHY (MCG) 10
Figure 2.1: Example SCG signal from MECHANO-HF database (Z axis).
electrocardiography. Studies inferred that SCG and MCG show special correlations
with intra-cardiac events and annotated the fiducial points that refer to mitral valve
opening (MO) and closure (MC), aortic valve opening (AO) and closure (AC), rapid
ejection (RE), and isovolumetric contraction (IVC) [20], [23], [52]–[55]. Annotation
of these points enabled the inference for timing of cardiac event and further anal-
ysis of hemodynamics. Left ventricular ejection time (LVET), pre-ejection period
(PEP), isovolumetric contraction time (IVCT) and isovolumetric relaxation time
(IVRT) are some timing properties derived from SCG and GCG in the literature
[23], [54], [56], [57]. These cardiac time intervals are important clinical markers for
the analysis of systolic and diastolic function and therefore the detection of cardiac
abnormalities [58], [59].
In addition to cardiac time intervals, fiducial points provide an opportunity to
analyze the morphological changes in MCG focusing on the cardiac phase/intracar-
diac event of interest. Smaller analysis windows focused on different fiducial points
have been used in the literature for representation of systolic and diastolic features
separately. Some of the common features in the morphologic analysis are kinetic
2.3 ELECTROCARDIOGRAM (ECG) 11
energy, vertical range, i.e. difference between maximum amplitude and minimum
amplitude, mean, median, and maximum amplitude in a selected window. The cor-
relation between these features and cardiac contractility was analyzed in different
settings [26], [29], [60].
MCG signals are affected by different additional sources like motion artifacts,
environmental effects and sensor interference, and this source of artifacts can lead
to errors in the feature calculation and further analysis [61], [62]. Thus, automatic
annotation of fiducial points and temporal analysis requires a well-designed prepro-
cessing pipeline. In the literature, the SCG components below 1 Hz were found to be
correlated with respiration-related movements, while components between 1-20 Hz
and above 20 Hz were correlated with vibrational and acoustic effects resulted from
cardiac activity, respectively [63]–[65]. Also, it has been indicated that S1 and S2
heart sounds occur between 50 and 500 Hz, and S3 sounds lie at lower frequencies,
between 20 and 200 Hz [66]. The available information regarding different frequency
components can be used in the design of the preprocessing pipeline according to the
interested characteristics and activity in the MCG signals.
2.3 Electrocardiogram (ECG)
The electrocardiogram (ECG) is a diagnostic method that displays changes in the
bio-electrical activity of the heart over time [12]. It involves a non-invasive measure-
ment technique, which includes placing electrodes on the patient’s skin to capture
electrical impulses [67]. Depolarization and repolarization events that occur in the
heart chambers manifest with distinct morphologies over time, and the visual as-
sessment of these events is crucial for evaluating and diagnosing cardiac conditions
[68]. In clinical settings, ECGs are crucial for the detection of myocardial infarction,
cardiac ischemia, heart failure, arrhythmias, and a wide range of cardiac conditions
and diseases [69], [70].
2.3 ELECTROCARDIOGRAM (ECG) 12
Figure 2.2: Example ECG signal from MECHANO-HF database.
ECG signals are utilized by visually evaluating the important morphological
structures they contain in a clinical setting. Electrical activity in the heart muscles
generates the PQRST morphology in ECG recordings (Figure 2.2). The P wave
occurs with the initiation of an electrical impulse in the atria, signifying atrial de-
polarization and contraction. It is followed by the QRS complex, which indicates
ventricular depolarization, leading to ventricular contraction. The T wave represents
ventricular repolarization, showing the recovery phase when the electrical balance
returns to its initial state [68]. Evaluating these waveforms can help clinicians ex-
tract information about the rhythmic, structural, and inflammatory aspects of the
heart’s condition [70].
ECG signals are also employed to extract significant time points in heart ac-
tivity, supporting diverse analysis pipelines. The R peaks in ECG recordings are
extensively used in biomedical signal processing applications to segment the signals
coming from different sources into individual heartbeats [55]. In addition, R peaks
are utilized to measure pulse transition time (PTT) in photoplethysmography appli-
cations [71]. The detection and analysis of Q waves are also essential, as they show
2.4 STATISTICAL HYPOTHESIS TESTING 13
the initiation of ventricular systole and are used to calculate QS2 (electromechanical
systole) duration. Various methods for automatically detecting PQRST waveforms
are proposed in the literature [72]–[74].
2.4 Statistical Hypothesis Testing
Statistical hypothesis testing is a fundamental method employed in research and
data analysis to assess the validity of a hypothesis. In this process, researchers
formulate two hypotheses: the null hypothesis (H0) and the alternative hypothesis
(H1). The null hypothesis represents the default assumption and often states the
absence of an effect or no difference, while the alternative hypothesis defines the
desired outcome or the presence of a specific effect. The aim is to determine if there
is enough evidence in the sample data to reject the null hypothesis in favor of the
alternative hypothesis. To make this decision, a p-value is calculated using various
statistical formulas tailored to the specific characteristics of the sample data. The
p-value is then compared to a predetermined significance level (α), usually set at
0.05, to determine statistical significance. If the p-value is less than or equal to α,
the null hypothesis is rejected; otherwise, it is accepted [75], [76]. Statistical tests
can be broadly categorized into two groups based on the assumptions they rely on:
parametric tests and non-parametric tests.
2.4.1 Parametric Tests
Parametric tests are based on the assumption that the given data follows a specific
probability distribution, typically the normal distribution. This assumed distribu-
tion is employed in the calculation of data statistics, such as mean and variance,
and, consequently, the test’s power and precision depend on the accuracy of this
assumption. It is important to note that a violation of the assumption of normality
2.4 STATISTICAL HYPOTHESIS TESTING 14
can effect the validity of parametric tests, and in cases where the data deviates sig-
nificantly from a normal distribution, alternative non-parametric tests may be more
appropriate. Common examples of parametric tests include the t-test and linear
regression [77], [78].
2.4.2 Nonparametric Tests
Nonparametric tests (distribution-free tests) utilize less or no assumption about the
distribution of data. They are used on data that does not meet the assumptions
emerging during the use of parametric tests or has ordinal characteristics. In such
situations, nonparametric tests provide reliable and valid results.
Another key feature of nonparametric tests is their ease of application to small
sample sizes, where parametric tests might be less reliable. This makes them valu-
able in situations where limited data is available or when strict experimental con-
ditions restrict the sample size. Furthermore, nonparametric tests are less sensitive
to the shape of the distribution and, therefore, more robust in scenarios where the
underlying data distribution is unknown. For this reason, they are valuable and
robust, even though they may have less statistical power compared to parametric
tests. Some examples of non-parametric tests are the Mann-Whitney U test, the
Wilcoxon signed-rank test, and the Kruskal-Wallis test [79]–[81].
Wilcoxon Signed-Rank Test
The Wilcoxon signed-rank test is a non-parametric statistical test used to determine
whether the distribution of differences between paired samples is symmetric about
a specified median and assess whether there is a significant shift in the distribution
of differences [80]. This test is particularly useful when analyzing paired data, when
each pair of observations is related in some way (e.g., before-and-after measurements
or control-treatment pairs), parametric assumptions are not met, or data does not
2.4 STATISTICAL HYPOTHESIS TESTING 15
follow a normal distribution.
The test involves ranking the absolute differences between each pair of observa-
tions, considering their signs, and then summing the ranks of the differences. The
test statistics formula for Wilcoxon signed-rank test isW =
∑︁Nr
i=1[sgn(x2,i−x1,i·Ri))]
where W is the test statistic, Nr is the sample size, sgn is the sign function, x2,i, x1,i
are corresponding pairs from two distributions, and Ri is the rank.
3 Related Work
3.1 Filtering of MCG Signals
Filtering of the MCG signals is fundamental for the analysis and interpretation of
the cardiac vibrations. In this chapter, we introduce different filtering methodologies
applied to MCG signals in the literature.
3.1.1 Frequency Filters
Frequency filters play a crucial role in altering the content of signals in electron-
ics and signal processing. Low-pass filters allow lower frequencies to pass through
while reducing higher ones, making them useful for eliminating noise or focusing on
fundamental signal elements. High-pass filters, in contrast, allow higher frequen-
cies to pass while diminishing lower ones, finding application in situations where
low-frequency elements are undesirable. Band-pass filters are designed to enable a
specific range of frequencies to pass through, excluding those outside the designated
band. These filters are useful for isolating desired frequency components. Band-stop
filters, on the other hand, target and reduce a narrow band of frequencies, useful for
removing interference or specific unwanted components [82].
Butterworth, Chebyshev, and elliptic (Cauer) filters are among the most popular
types of filters, each offering distinct advantages. The Butterworth filter provides a
maximally flat frequency response in the passband, making it suitable for applica-
3.1 FILTERING OF MCG SIGNALS 17
tions where a consistent gain across frequencies is crucial. Chebyshev filters, on the
other hand, allow for a steeper roll-off at the expense of some ripple in either the
passband (Chebyshev Type I) or the stopband (Chebyshev Type II). This flexibility
makes them ideal for scenarios where a sharper transition between the passband and
stopband is required. Elliptic filters offer the steepest roll-off among these filters.
In MCG, filtering methodology and the selected passband depend on the specific
research or diagnostic goals. In seismocardiogram analysis, components below 1 Hz
are associated with respiration-related movements, those between 1-20 Hz and above
20 Hz are linked to vibrational and acoustic effects from cardiac activity, and specific
heart sounds such as S1 and S2 occur between 50-500 Hz, while S3 sounds lie in lower
frequencies between 20-200 Hz [63]–[66]. Based on the desired component to analyze,
different frequency filters can be designed with the corresponding frequency range
and properties. For instance, in [32], it is proposed that most information regarding
cardiac vibration exists above 1 Hz and a bandpass filter with a pass band of 1-45
Hz was used. In [23], a 4th-order Butterworth filter with passbands of 1–20 Hz and
4–45 Hz were used for GCG and SCG, respectively. In [55], a 0.05 Hz highpass
and 90 Hz lowpass filter were used for extracting the mean heartbeat and analyzing
SCG morphology, while a 50–500 Hz bandpass filter was used in the same study
for analyzing heart sounds. For focusing on the mean heart beat morphology, the
baseline and frequency components higher than 90 Hz were eliminated. However,
heart sounds occur in the higher components so that the high-frequency components
were preserved for heart sound analysis.
Moreover, different filtering options can be used in the same analysis pipeline to
reveal the effects of different SCG components. In [62], six different filtering options
(0-1 Hz, 0-20 Hz, 0-40 Hz, 1-20 Hz, 1-40 Hz, 20-40 Hz) on SCG signals were used and
machine learning models were trained for the prediction of the metabolic equivalent
of task (MET) scores using one of the defined filtering options. Analyzing the model
3.1 FILTERING OF MCG SIGNALS 18
performances, the informativeness of different SCG components in the estimation
was discussed and it is found that respiration-related SCG components are critical
for MET estimation.
3.1.2 Wavelet Transform Approach
Wavelet transform is a technique employed in signal and image processing for various
purposes such as denoising, compression, and feature extraction [83]–[85]. Utilizing
small basis functions known as wavelet, the wavelet transform decomposes the signal
into representations that shows different frequency components at different resolu-
tions [84], [86]. This transformative approach is advantageous, especially in heart
sound detection, as it allows a focused examination of distinct frequency compo-
nents, thereby enhancing the clarity of heart sounds while mitigating the impact of
high-pitched noises in MCG signals [87]. Earliest studies [88] applied wavelet trans-
form to phonocardiogram signals, and the literature presents diverse methods for
leveraging wavelet transform in fiducial point detection and heart sound extraction
on SCG signals [89]–[91].
The primary application of wavelet transform involves extracting different de-
tail and approximation waveforms of the MCG signals at various levels and utilizing
these waveforms. The signal processing pipeline preceding and following the wavelet
decomposition stage varies across different studies. Additionally, the wavelet de-
composition step may differ based on the choice of the mother wavelet used in the
decomposition process. In the literature, Morlet wavelet and Daubechies wavelet
families were mostly utilized and proposed as the mother wavelets for especially
heart sound analysis [92]–[95].
3.2 FIDUCIAL POINT DETECTION ON MCG SIGNALS 19
3.1.3 Synchronized Averaging Approach
Synchronized averaging is a signal processing technique designed to enhance the clar-
ity of cardiac vibrations observed in MCG measurements. Synchronized averaging
includes the synchronization of heart beats with the detected reference points and
the averaging of MCG signals over multiple cardiac cycles [64], [65]. One primary
benefit is noise reduction, since the technique reduces the impact of random noise
and motion artifacts in recordings. By highlighting consistent components across
multiple cycles, synchronized averaging increases the signal-to-noise ratio and pro-
vides a cleaner representation of the cardiac vibrations.
3.2 Fiducial Point Detection on MCG Signals
Various cardiac events lead to distinct waveform morphologies at different MCG
axes. Detecting the time points of these cardiac events helps in analyzing the spe-
cific cardiac event and identifying various timing features applicable to cardiac health
assessment. Despite fiducial points being annotated on SCG and GCG signals in
the literature, and approximate morphologies around these events being known,
the presence of noise due to motion artifacts, deviations caused by respiration, or
intersubject variability resulting from body mass distribution, gender, and age dif-
ferences may create differences in these morphologies [32], [96], [97]. Therefore,
detecting fiducial points can be challenging and prone to errors. Various approaches
are proposed in the literature to facilitate the identification of fiducial points on
MCG signals.
3.2.1 Search Windows
This approach is grounded in restricting the search area where the detection al-
gorithm operates. The cardiac time intervals for various events are already docu-
3.2 FIDUCIAL POINT DETECTION ON MCG SIGNALS 20
mented in the literature for both healthy subjects and those with different cardiac
conditions [98]–[101]. These timings can be leveraged during the search for fiducial
points, allowing the search to focus solely on the signal segment known to contain
the fiducial point of interest. This method helps eliminate misleading and similar
morphologies occurring in the heartbeat, ensuring that the detected point is within
the corresponding search window. Once the search window is defined, the detection
algorithm specific to the interested morphology can be executed only for the seg-
ment within the search window. However, it is important to note that this method
necessitates reference points to delineate the search window.
For example, in [102], a 90 ms search window was established to identify aortic
valve opening points on SCG signals, based on the detected R-peaks on the ECG
signals. In [103], AO on SCG was determined as the maxima occurring after 45
ms and before 125 ms following the detected Q wave in the ECG. For aortic valve
closing (AC), AO was utilized as a reference and AC was defined as the maxima
within the interval [AO + 240 ms, AO + 350 ms]. Similarly, in [104], a 250-ms-long
search window was employed starting from 200 ms after S1 and S2 was defined as
the minimum wave in the search window.
3.2.2 Dynamic Time Warping
Dynamic Time Warping (DTW) is a dynamic programming-based algorithm used to
align two sequences that may have different lengths [105]. The aim is to find an opti-
mal alignment between the samples of the sequences that matches one sample from
the first sequence to the other. This matching may not be necessarily one-to-one due
to the possible length differences between sequences. The dynamic programming is
utilized to minimize the pre-defined distance metric that qualifies the quality of align-
ment, and the algorithm aims to find optimal alignment between these sequences
based on specified constraints. These constraints may include monotonicity, conti-
3.2 FIDUCIAL POINT DETECTION ON MCG SIGNALS 21
nuity, boundary, warping window, and slope [106], [107]. DTW is generally used for
aligning sequences that are morphologically similar but occur at different paces, e.g.,
speech recordings contain identical content but are pronounced at different paces.
DTW includes the construction of a cost matrix that includes the cost of align-
ment for each possible sample pair [105]. Dynamic programming recursively finds
a warping path that minimizes the total cost and aligns the sequences according to
the constraints. The cost function may be simply defined as the Euclidean distance
of samples; however, new cost functions can be defined and utilized [108]–[110].
In [111], DTW was applied for fiducial point detection on SCG signals. They
utilized an intermediary reference signal pair consisting of ECG and SCG signals
with a specific RR interval. They proposed a hybrid cost function for DTW that
considers differences in signal values, neighborhood shifting level, signal slope, and
concavity and provided a flexible and adaptable approach for accurate fiducial point
detection. The fiducial point, such as the aortic valve closing event, is detected in
the reference SCG signal and then projected to the nonreference SCG signal using
DTW-based quasi-synchronous alignment.
3.2.3 Hidden Markov Model
A Hidden Markov Model (HMM) is a statistical model used to represent an evolving
system whose state is not directly observable but can be inferred from observed data.
It consists of a set of hidden states, observable outputs, and transition probabilities
between states. The probability of observable outputs depends on the underlying
state chain, and based on observable outputs, the underlying state transition prob-
abilities can be estimated [112].
In the literature, the HMM method is applied to seismocardiogram signals to
estimate heart rate, heart rate variability, and cardiac time intervals [113]. In [113],
SCG heartbeats were defined with N hidden states and modeled a sequential tran-
3.3 ASSESSMENT OF CARDIAC FUNCTION WITH MCG SIGNALS 22
sition between states. The value of N was determined based on the heart rate of
the corresponding subject. Following the modeling, they employed the Baum-Welch
algorithm to estimate the parameters of the defined statistical model and identified
the state of each sampling instance. For detecting cardiac events, they utilized one
manual annotation of the cardiac events for each subject and labeled the states
corresponding to the fiducial points. Using the time points of the same states in
the subsequent cycles, they detected the fiducial points of the following cycles and
measured cardiac time intervals.
3.3 Assessment of Cardiac Function with MCG Sig-
nals
3.3.1 Cardiac Performance
To analyze and measure changes in cardiac function using MCG signals, understand-
ing the correlation between the time and frequency domain properties of MCG and
underlying physiological variables is crucial. Various time and frequency domain
features have been generated and analyzed in the literature concerning the known
pathological effects of different diseases. To acquire information about the overall
trend of magnitude, certain features are defined:
• Amplitude: A = x(t) [29], [60].
• Maximum Amplitude: Amax = max
t
x(t) [114], [115].
• Minimum Amplitude: Amin = min
t
x(t) [114].
• Vertical Range: R = maxt x(t)−mint x(t) [29], [114].
• Power: P =
∑︁N
t=1 x(t)
2 [115].
3.3 ASSESSMENT OF CARDIAC FUNCTION WITH MCG SIGNALS 23
• Root Mean Square Power: Prms =
√︂
1
N
∑︁N
i=1 x
2
i [24], [60], [114], [115].
• Linear Kinetic Energy: KLin(t) = 12m(v
2
x(t) + v
2
y(t) + v
2
z(t)) where vx(t),
vy(t), and vz(t) are the velocities in the x, y, and z directions, respectively
[26].
• Rotational Kinetic Energy: KRot(t) = 12(Ixxω
2
x(t) + Iyyω
2
y(t) + Izzω
2
zz(t))
where Ixx, Iyy, and Izz are the moments of inertia and ωx(t), ωy(t), and ωzz(t)
are the angular velocities in the x, y, and z directions, respectively [26].
These features are associated with the cardiac performance of the heart in dif-
ferent studies. For instance, in [26], it is found that kinetic energy calculated from
MCG measurements is related to cardiac contractility. Similarly, in [29], an increase
in the vertical range was reported when the pacemaker was turned on, indicating
better left ventricular function. In [116], a relationship between left ventricular
contractility and the maximum magnitudes acquired from an accelerometer sensor
placed on the chest of minipigs was proposed. In line with these findings, the mag-
nitudes observed in MCG measurements can be related to the strength of heart
contraction, reflecting the performance of cardiac function.
3.3.2 Cardiac Time Intervals
Cardiac time intervals are crucial parameters for assessing heart function and di-
agnosing cardiovascular disorders. These intervals include left ventricular ejection
time (LVET), pre-ejection period (PEP), and total electromechanical systole period
(QS2). After detecting fiducial points on MCG and ECG signals, these cardiac time
intervals can be measured by calculating the time difference between the fiducial
points of interest. In the literature, these time intervals were derived from MCG
signals and defined as follows:
• PEP: Time difference between the Q wave and AO [28], [117].
3.3 ASSESSMENT OF CARDIAC FUNCTION WITH MCG SIGNALS 24
• LVET: Time difference between AO and AC points [28], [54], [56], [57], [118].
• QS2: Time difference between the Q wave and AC [118].
In the literature, variations in LVET have been proposed to be associated with
conditions such as heart failure [99], [119], [120], hypertension [121], and ischaemic
heart disease [122]. The PEP could serve as an indicator of increased sympathetic
nervous system activity or impaired ventricular contractility [123], [124], while ab-
normalities in the QS2 may suggest an increased risk of mortality, particularly in
cases of coronary artery disease [125]. Therefore, these time intervals are crucial and
important as they allow us to analyze the cardiac health of patients.
4 MECHANO-HF Dataset
MECHANO-HF dataset was collected at Turku University Hospital (TYKS) and it
contains data from 10 CRT patients. All patients were diagnosed as HFrEF and
had previously obtained a CRT pacemaker. Inclusion criteria required participants
to be
• at least 18 years old
• capable of understanding and consenting by signing an informed consent form
• in stable clinical condition without significant valve disease, rhythm irregular-
ities, or significant clinical instability.
Variable
Age (year) Mean (±SD) 73.98 (±7.52)
Sex Males, n 8
Females, n 2
Height (cm) Mean (±SD) 172.20 (±6.97)
Weight (kg) Mean (±SD) 92.80 (±22.95)
BMI (kg/m²) Mean (±SD) 31.00 (±5.59)
Left ventricular ejection
fraction (%)*
Mean (±SD) 39.90 (±8.77)
NYHA class I, n 1
II, n 4
III, n 4
IV, n 1
Table 4.1: Baseline characteristics of patients included in the study. (*After intro-
duction of optimal medical treatment and CRT)
CHAPTER 4. MECHANO-HF DATASET 26
Figure 4.1: The device is positioned on the lower part of the patient’s chest, recording
3-axis acceleration, 3-axis rotation, and a single-lead ECG while the patient is in a
supine position. This illustration has been adapted from the original figure featured
in the article [24] under the license CC-BY 4.0.
The study’s baseline demographics are detailed in Table 4.1, and ethics approval was
obtained from the Ethics Committee of the Hospital District of Southwest Finland,
aligning with the 2002 Declaration of Helsinki. All participants provided written
informed consent.
A custom-designed device, incorporating a 3-axis accelerometer (ADXL355, Ana-
log Devices Inc., Wilmington, MA, USA), a 3-axis gyroscope (LSM6DS3, STMicro-
electronics, Geneva, Switzerland), and a single-lead ECG, was used for data collec-
tion. The device had accelerometer and gyroscope measurement ranges of ±2 g and
±250 dps, respectively. The accelerometer’s noise density was 25µg/
√
Hz, and the
gyroscope’s rate noise density was 7mdps/
√
Hz. The sensor was placed on the lower
chest (Figure 4.1). Each measurement included signals from one accelerometer and
one gyroscope for each of the three axes, along with a single-lead ECG. Simultaneous
recording of MCG and ECG was conducted.
Recordings took place with patients in a supine position after a 15-minute rest
CHAPTER 4. MECHANO-HF DATASET 27
in a quiet environment. Patients were instructed to breathe normally, avoid talk-
ing, and minimize unnecessary movements. The pacemaker was randomly set to
AAI or CRT mode, with a heart rate of 80 bpm to maintain a consistent heart
rate throughout measurements. In AAI mode, only the atrium is paced, result-
ing in dyssynchronous ventricular function, while the optimized CRT mode aims to
resynchronize ventricular contraction.
5 Proposed Method
5.1 Signal Processing
Respiratory-related variations and baseline measurements are evident in the lower
frequency components of the MCG signal, potentially causing significant fluctua-
tions [126]. Conversely, high-frequency components can introduce ambiguity in the
fiducial point detection process, even though their impact on the waveform is less
pronounced than that of low-frequency artifacts. To enhance the analysis of cardiac
activity, signal processing plays a crucial role in MCG analysis. In this section, we
introduce our signal processing pipeline.
5.1.1 Signal Resampling
The dataset contains ECG and MCG measurements recorded simultaneously for a
duration ranging from 8 minutes to 12 minutes. Before further processing, ECG and
MCG signals of the corresponding measurement were synchronized and resampled
to 400 Hz.
5.1.2 Noise Removal
Initially, segments with high motion interference and low signal quality were manu-
ally excluded. Thirty-second-long segments were chosen from each MCG measure-
ment for analysis. Subsequently, a forward-backward filtering operation was applied
5.1 SIGNAL PROCESSING 29
Figure 5.1: Example SCG-Z signal from MECHANO-HF database before filtering
(left) and after filtering (right).
to SCG-Z signals. A Butterworth bandpass filter with a 20-90 Hz passband was
selected for its desirable characteristics, which include a maximally flat frequency
response within the passband and a smooth roll-off. The application of this filter aids
in refining the MCG waveform. Figure 5.1 illustrates the SCG-Z signal both before
and after the application of the filtering process. Finally, the 30-second segments
were further divided into 10-second segments for the rest of the analysis.
5.1.3 R Peak and Q Wave Detection
Hamilton’s algorithm, as outlined in [72], was utilized for R-peak detection in ECG
signals. The process involves preprocessing to reduce noise, the application of a
band-pass filter to isolate the QRS complex, and the use of techniques such as
differentiation and squaring to identify and enhance the R-peaks. The algorithm
establishes a threshold to select peaks above a certain height, providing accurate
timing for when the heart’s ventricles depolarize in ECG readings. Additionally,
the detection of Q waves in the ECG signals was performed using the wavelet-
based delineation algorithm available in the NeuroKit2 toolbox [127]. Example
ECG recordings with annotated R peaks and Q waves are represented in Figure 5.2.
5.1 SIGNAL PROCESSING 30
Figure 5.2: Example ECG signal with detected Q waves and R peaks in a heart
failure patient undergoing CRT pacing (left panel) and AAI pacing mode (right
panel).
5.1.4 Aortic Valve Opening & Closing Detection
Following the detection of R-peaks, each individual heartbeat has been successfully
identified. To annotate the aortic valve opening (AO) and aortic valve closing (AC)
points, a search window approach was employed. For each heartbeat, the aortic
valve opening point was determined as the maximum peak point of the SCG-Z axis
within the first 125-ms interval after the corresponding R peak.
The left ventricular ejection time (LVET) represents the duration between the
opening and closing of the aortic valve during each cardiac cycle. The search window
for aortic valve closing (AC) was defined based on reported LVET and left ventricular
ejection time index (LVETI) in the literature [99], [128], utilizing the detected AO
points as reference.
Consequently, a 60-ms long window was positioned 240 ms after the correspond-
ing AO, and AC was identified as the maximum peak within the interval between
240-300 ms after the AO. AO and AC detection operations were exclusively con-
5.1 SIGNAL PROCESSING 31
ducted on the SCG-Z signals, with the same locations employed on other axes from
the same segment and synchronized with the same reference as the corresponding
SCG-Z signal.
5.1.5 Filtering
Various frequency components in the MCG measurements were suggested to be
linked to different sources of information. Specifically, SCG components between 1
and 20 Hz are associated with mechanical vibrations, while those above 20 Hz are
linked to acoustic effects caused by cardiac activity [64], [65]. To analyze the effect
of CRT on these components separately, we designed four distinct first-order Butter-
worth filters with a passband chosen from the following options: 20-90 Hz, 6-90 Hz,
1-20 Hz, and >1 Hz (highpass). Each axis of the SCG and GCG in measurements
was initially filtered with these specific filters and subsequently analyzed.
5.1.6 Vector Creation
To represent the total movement of the SCG and GCG sensors, SCG and GCG
vectors were constructed. For SCG and GCG vectors, all three axes of the respective
sensor were combined using the Euclidean formula. Furthermore, MCG vectors were
created for each frequency range (>1 Hz, 20-90 Hz, 6-90 Hz, 1-20 Hz) by combining
SCG and GCG vectors. However, as the GCG and SCG sensors have different
measurement units and ranges, combining them required scaling these measurement
units between [0,1] to ensure a comparable contribution from both sensors before
their combination. To address this range issue, SCG and GCG vectors were scaled
to fall within the [0,1] range. This scaling was achieved by utilizing the maximum
and minimum values calculated for all vectors in the corresponding frequency ranges
from the corresponding measurement units.
Firstly, all SCG and GCG vectors were extracted by filtering the signals with the
5.2 FEATURE EXTRACTION 32
Listing 1 MCG Vector Creation
1: for frequency range in [> 1 Hz, 20-90 Hz, 6-90 Hz, 1-20 Hz] do
2: for measurement unit in [SCG, GCG] do
3: Create vectors
4: Find the maximum and minimum value of all created vectors (1 max and
min value for the whole set)
5: Scale all vectors according to the maximum and minimum value
6: end for
7: Create MCG vectors
8: end for
selected pass band. Then, we identified the minimum and maximum values of the
vectors coming from the sensors. All SCG and GCG vectors were scaled within [0,1]
using the identified maximum and minimum values of the SCG and GCG vectors,
respectively. The operation was applied each pass band option separately. The
pseudocode for extraction of MCG vectors is presented in Listing 1.
5.1.7 Synchronized Averaging
For every signal in a 10-second segment, two mean heartbeats were generated. This
process involved synchronized averaging using both the R peak and AO point as
reference points. Synchronized averaging entails extracting individual heartbeats,
aligning them based on their reference points, and averaging the heartbeats in rela-
tion to this alignment. Two examples for mean SCG vector beats are shown in Fig.
5.3. At the conclusion of this step, there were 72 average cycles for each 10-second
segment, representing each of the 6 axes and 3 vectors (SCG, GCG, and MCG)
filtered with the 4 filtering options, and 2 mean cycles for each.
5.2 Feature Extraction
In this section, we outline how features were generated from the processed MCG
signals to support the comparison of waveform characteristics between CRT and
5.2 FEATURE EXTRACTION 33
Figure 5.3: Example mean beat of SCG vector in a heart failure patient undergoing
CRT pacing (left panel) and AAI pacing mode (right panel) based on AO averaging.
AAI modes.
5.2.1 Morphological Features
To extract features related to systolic and early diastolic characteristics, we initially
identified the systolic and early diastolic regions in the signals by locating AO and
AC points on SCG-Z averages. For each average cycle, two windows of 150 ms
duration were created. One window centered on the detected AO, and the other
window was placed immediately after AC to focus on the early diastolic region.
These windows were termed systolic and early diastolic windows, respectively.
Subsequent to the creation of systolic and early diastolic windows, morphological
features were extracted from these windows. These features included energy (Ex)
and vertical range (Rx), and they were defined as follows:
Ex =
∞∑︂
n=−∞
x[n]2
5.3 STATISTICAL TESTS 34
Rx = maxx[n]−minx[n]
Example average cycles and extracted features are shown in Figure 5.4.
5.2.2 Timing Features
LVET and electromechanical systole (QS2) time intervals were also extracted as
features. We determined the left ventricular ejection time (LVET) and electrome-
chanical systole (QS2) by calculating the time duration between the identified AO
and AC, and between the Q wave and AC, respectively.
5.3 Statistical Tests
Our goal was not only to identify potential differences between CRT and AAI modes
but also to explore specific MCG components that exhibit the most significant dis-
tinctions, providing valuable insights into the optimal features for detecting differ-
ences. Accordingly, in the previous section, we extracted features from various mean
cycles of MCG measurements differing in terms of frequency, axis, and synchronized
averaging method. In this section, we will introduce the statistical analysis method
that we employed in the study.
In our analyses into the differences between CRT and AAI modes in MCG wave-
forms, we conducted paired statistical tests on features extracted from various mean
cycles. These cycles were obtained for all three 10-second segments that belong
to the same measurements, allowing us to analyze distinct segments of the MCG
recordings. Each experiment focused on one frequency component, axis, and syn-
chronized averaging method combination and compared the corresponding MCG
cycles in CRT and AAI modes.
5.3 STATISTICAL TESTS 35
Figure 5.4: Definitions of extracted features in a heart failure patient undergoing
CRT pacing (upper panel) and AAI pacing mode (lower panel) based on AO averaged
SCG-Z cycle. This illustration has been altered from the original figure in the article
[34] and is being utilized with explicit permission.
5.3 STATISTICAL TESTS 36
Figure 5.5: The distribution of segments to various sets involves putting the features
of the first, second, and third segments of the corresponding mode into sets 1, 2,
and 3, respectively, for each subject. The illustration presented here is a modified
version of the original figure depicted in the article [34], used with permission.
The features from the first, second, and third segments were put into set-1, set-
2, and set-3, respectively, creating three individual sets representing separate and
non-overlapping parts of the recordings (Fig. 5.5). To enhance the robustness of
our findings, we applied paired tests to each set, effectively creating a double-check
mechanism and minimizing the risk of type I errors.
For each experiment, statistical comparisons were performed using the Wilcoxon
signed-rank test, with a significance threshold of p < 0.05. We considered differences
between pacing modes as significant only when the null hypothesis was rejected in
all three paired tests conducted on set-1, set-2, and set-3. If any of the tests did
not result in the rejection of the null hypothesis, we did not accept and report the
alternative hypothesis, ensuring a rigorous evaluation of the observed differences
between CRT and AAI modes (Fig. 5.6).
5.3 STATISTICAL TESTS 37
Figure 5.6: Algorithm for data analysis. Each individual paired set underwent a
statistical test, and the null hypothesis was rejected if all parallel tests rejected the
null hypothesis. The illustration shown here has been adjusted from the original
figure in the article [34] and is being employed with permission.
6 Results
6.1 Morphological Features
In this chapter, we report the results of the statistical comparisons of morphological
features extracted from measurements during CRT and AAI mode.
6.1.1 Systolic Features
Vertical range (Rx) and energy (Ex) in the systolic window exhibited statistical dif-
ferences between the two pacing modes. During CRT pacing, systolic features were
higher than those in AAI mode. The changes in systolic energy, vertical range for
each axis, frequency range, and synchronization method were examined during the
transition from CRT to AAI pacemaker mode and represented in Fig. 6.1. Statis-
tically significant differences between the modes were denoted by colored squares.
The percentage within the squares indicates the mean energy and vertical range
change after initiating AAI pacing, relative to the values observed in CRT mode.
The color gradients within each box illustrate the descending order of differences
between all CRT-AAI segment pairs in the corresponding experimental settings.
Every SCG & GCG axis exhibited statistically significant differences in vertical
range across at least one frequency range, resulting in p<0.05 for all three parallel
statistical tests (Fig. 6.1). Additionally, SCG, GCG, and MCG vectors exhibited
differences in systolic range between the two pacing modes. Differences were visible
6.1 MORPHOLOGICAL FEATURES 39
Figure 6.1: Changes in systolic energy and vertical range for each axis, filtering
frequency and synchronization method. Adapted from the original figure featured
in the article [34], this illustration has been modified and is utilized with permission.
6.2 TIMING FEATURES 40
AAI : Mean (SD) CRT : Mean (SD) Difference
QS2 428 ms (± 35) 405 ms (± 46) Not significant
LVET 273 ms (± 14) 269 ms (± 15) Not significant
Table 6.1: Mean and standard deviation of QS2 and LVET during CRT and AAI
mode.
in all frequency components, with the most noticeable changes occurring within the
6-90 Hz frequency range, remaining significant with the extension of the frequency
range. AO Point and R peak synchronization yielded comparable results and showed
differences in a slightly different experiment set. However, significant differences in
the systolic range of the GCG-Z axis were observed in R-peak synchronization, while
the difference in SCG-X was noteworthy only when synchronized with the AO point.
The increase in systolic energy when pacing mode changed from AAI to CRT was
significant in the SCG-Y, SCG-Z, GCG-X, and GCG-Y axes, as well as in the SCG,
GCG, and MCG vectors. Differences in systolic energy were more concordant in the
GCG axes and GCG vector. Moreover, observable differences in systolic energy were
present in the ranges that included higher frequency components, with the difference
not being significant in 1–20 Hz. Systolic energy and vertical range changes in the
GCG-X axis for each subject are shown in Figure 6.2.
6.1.2 Early Diastolic Features
Despite the differences found in systolic features, there were no significant differences
observed in early diastolic features.
6.2 Timing Features
Table 6.1 presents the mean and standard deviation concerning measured QS2 and
LVET during CRT and AAI pacing modes. The analysis revealed no notable differ-
ences in QS2 and LVET when comparing CRT and AAI modes.
6.2 TIMING FEATURES 41
Figure 6.2: Mean systolic energy and range among subject during CRT and AAI
mode extracted with AO point synchronization method in the 20-90 Hz range. This
illustration originates from the original figure featured in the article [34] and is being
used with permission.
7 Discussion
With the introduction of CRT pacing, significant changes around the systolic win-
dow were observed in all axes and frequency ranges. During CRT pacing mode,
MCG waveforms in the systolic window exhibited higher energy and vertical range.
Changes in the waveform were more remarkable in the higher frequency components,
with the most concordant differences observed in the 6-90 Hz range. Vertical range
variances were clearer than energy differences, specifically; the differences detected
in the vertical range were significant in a wider range of frequency ranges and axes.
The best axes for observing differences were SCG-Y, GCG-X, and GCG-Y, and the
GCG vector outperformed the SCG vector in showing differences. As shown in Fig.
6.2, all heart failure patients showed a consistent increase in vertical range and en-
ergy when the pacing mode changed to CRT from AAI. R-peak synchronization and
AO point synchronization methods produced similar results while showing differ-
ences in a slightly different experiment set. However, no consistent differences were
observed in the vertical range and energy around the early diastolic window after the
introduction of resynchronization pacing, and there were no significant differences
in the LVET and QS2 parameters measured during CRT and AAI mode.
Cardiac resynchronization therapy (CRT) can enhance cardiac function by syn-
chronizing ventricular contraction, resulting in an elevated left ventricular ejection
fraction, cardiac index, rapid rate of pressure rise, and decreased left ventricular
end-systolic volume index [13], [14], [129]–[132]. Through reverse modeling, CRT
CHAPTER 7. DISCUSSION 43
has been shown to reduce the rate of death and hospitalization in selected heart fail-
ure with reduced ejection fraction (HFrEF) patients [132]–[134]. Existing literature
indicates that CRT pacing mode leads to better cardiac function than AAI pacing
mode [135]. Additionally, signals such as seismocardiography (SCG) and gyrocardio-
graphy (GCG) have been demonstrated to encompass information regarding cardiac
performance. MCG waveforms, specifically, are proposed to reflect left ventricular
contractility, left ventricular stroke volume, and improved cardiac performance [26],
[29], [65], [116], [136], [137].
Building upon methodologies used to analyze SCG and GCG measurements and
the proposed relationships between MCG features and cardiac contractility in the
literature, energy (Ex) and vertical range (Rx) features were defined to reflect con-
traction characteristics in the systolic and early diastolic phases. Consistently, in
the three sets of CRT-AAI segments, energy and vertical range around the systolic
window showed a statistically significant increase in our patient cohort when the
pacemaker mode changed from AAI to CRT. Based on the established relationship
between MCG magnitude characteristics and cardiac contractility, as well as the
link between CRT pacing and improved cardiac function, it can be proposed that
the observed increase in vertical range and energy is related to the improvement of
ventricular function resulting from CRT pacing. This suggests a potential applica-
tion of MCG signals for the assessment and prediction of CRT response in heart
failure patients.
Heart sounds offer crucial insights into heart function and contribute to the di-
agnosis of diverse cardiac conditions [60], [138]–[142]. Specifically, the intensity of
the first heart sound (S1) correlates with cardiac contractility and cardiac reserve,
and a diminished S1 sound serves as a clinical marker for heart failure [143]–[146].
Acoustic vibrations related to heart sounds emerge in the higher frequency compo-
nents of MCG signals, and in our study, the most concordant changes were observed
CHAPTER 7. DISCUSSION 44
by including higher frequency components.
Observed changes in systolic features when pacemaker mode changed to CRT
mode were most prominent in the GCG axes and GCG vector. This is consistent with
the previous observations, which proposed that GCG signals exhibit lower sensitivity
to both inter-subject and intra-subject variations compared to SCG, particularly in
the GCG-X and GCG-Y axes [23], [25], [32], [147], [148]. GCG-Z axis, however,
generally has lower signal quality than other GCG axes and showed less salient
results in our analysis [23]. Among the SCG axes, the SCG-Y axis showed the
most concordant results, including vibrations in the foot-to-head direction of blood
movements.
The vertical range feature exhibited statistical differences in a broader range of
experiments compared to the energy feature. It is important to note that the vertical
range calculation employs only extrema points within the selected window, whereas
the energy calculation encompasses every point, as depicted in Figure 5.4. Utilizing
every sample in the window may render the energy feature more susceptible to noise
and alterations in values. Conversely, the vertical range calculation considers only
extrema points that shows distinctive morphologies and potentially offers greater
reliability.
Synchronized averaging includes the alignment of all heart beats according to
the selected reference and getting the average of the heart beats with respect to this
alignment. Different axes of the MCG measurements follow special shapes (peaks
and valleys) for various hemodynamic events, and these waveform morphologies are
annotated and defined in the literature [23], [55]. It is important to note that the
reference point may have an effect on the extracted mean beat and the analysis by
enhancing some of these shapes or distorting them. This effect of the reference point
on the morphology of the interested fiducial point depends on the variability of the
time difference between these points. In Figure 7.1, mean GCG-X beats extracted
CHAPTER 7. DISCUSSION 45
Figure 7.1: Mean GCG-X beats extracted with R peak and AO point synchroniza-
tion for CRT (left column) and AAI (right column) mode with all individual beats.
Individual beats are shown with the lighter colors.
CHAPTER 7. DISCUSSION 46
with R peak and AO point synchronization for one subject during CRT and AAI
mode are presented with all individual beats, and you can see the differences in
the alignment of individual beats and the waveform differences between the resulted
mean beats. Also, the effect of the selected reference on SCG waveform analysis is
shown in the literature [55]. For this reason, in this study, we employed two different
reference points for synchronized averaging and applied the same analysis pipeline
using both of them. While their results are generally comparable and consistent,
only one of these reference points resulted in significant differences in the SCG-X
and GCG-Z axes.
No statistical difference was observed in the early diastolic features when pacing
mode was changed from AAI to CRT, in contrast with systolic features. In the
literature, it is shown that MCG waveforms can reveal various alterations in cardiac
functions and detect abnormalities like heart failure and myocardial infarction [26],
[149]–[152]. In addition, it was found that patients with ST elevation myocardial
infarction (STEMI) show lower signal strength features in the systolic window than
control patients, consistent with our results and machine learning approaches were
utilized to detect STEMI [60]. On the other hand, the diagnostic precision of these
methodologies is still limited due to inter-subject variability that can be caused by
differences in heart orientation and body mass distribution. These variations consid-
erably effect the separability of the healthy controls and patient groups. However,
comparing the subject with oneself and detecting the changes with a self-similarity-
based approach may be a better approach for detecting abnormalities [24], [26], [153].
Elevated left ventricular filling pressure can be reliably identified with a louder S3
sound [154], and the change in the ventricular filling pressure due to the progression
of heart failure is most concordant in the early diastolic waveforms [155]–[158]. In
our study, early diastolic features did not change significantly after the resychro-
nization started because more time is needed to see compensatory changes in filling
CHAPTER 7. DISCUSSION 47
pressures.
In the literature, it is shown that LVET is shorter in HFrEF patients than in
controls, and extended LVET is observed when the pacemaker mode is switched
from AAI to CRT [128], [159]. On the other hand, the mean difference between
these pacemaker modes was only 6 ms, and this difference was only observable
with impedance cardiography but not with echocardiography [128]. In contrast
to the reported difference, we could not identify any differences in LVET and QS2
between the two modes. Based on the variations in results acquired through different
measurement methods, it is important to analyze our measurement unit. Firstly,
the sampling rate of our sensors was 400 Hz, and a limited sampling rate could be
a potential obstacle for sensitive measurement of timing parameters and detecting
differences in LVET and QS2, considering the low mean difference between the two
groups. Furthermore, studies using SCG at 2000 Hz and 5000 Hz reported that there
is no significant difference in LVET between the CRT on-off settings and CRT–AAI
pacing [29], [160]. Therefore, even though the impact of the CRT mode on systolic
performance is evident in MCG, its influence on timing parameters remains unclear
and necessitates further examination.
The study faced a significant constraint due to the restricted number of patients.
Initially, we had to employ various statistical analyses on the same patient group,
a process typically necessitating p-value correction [161]. However, the limitation
in the number of subjects also constrains the obtained p-values, potentially leading
to a attenuation of the observed differences when applying p-value correction. To
address this issue, we developed a data analysis algorithm that iterates the same
test on three pairs acquired from different time points. The alternative hypothesis
was accepted only if all three pairs exhibited statistically significant results. This
method introduced a double-check mechanism to our tests, effectively mitigating
time-dependent false-positive results.
CHAPTER 7. DISCUSSION 48
Sympathetic activity pertains to the activation of the sympathetic nervous sys-
tem, leading to heightened physiological responses, including an elevated heart rate,
increased blood pressure, enhanced release of adrenaline, and increased alertness.
Variations in sympathetic activity within clinical settings may induce changes in
the features computed from SCG and GCG, influencing the observed findings [162].
Additionally, these effects might introduce alterations even in healthy subjects, and
it is important to distinguish them from the effects of CRT and validate our re-
sults. Therefore, future studies should focus on comparing these variations between
healthy controls and HFrEF patients.
In our study, 1-channel ECG measurements recorded simultaneously with MCG
were utilized to detect individual heartbeats and enhance the accuracy of fiducial
point detection. Although the waveform annotations for aortic valve opening and
closing on SCG and GCG signals were defined in the literature [23], [55], the mor-
phology of MCG measurements can undergo significant alterations due to inter-
personal variabilities, subject motion, or environmental artifacts [55]. These changes
can easily impede the accurate detection of these fiducial points. However, the detec-
tion of the R-peak was more convenient because of the distinctive morphology of the
R-peak, which differentiates itself with a much higher magnitude and steeper slope
from other peaks, and the lower possibility of artifacts in our experimental setting.
In our study, it was crucial to properly detect fiducial points for an accurate evalu-
ation of the systolic and early diastolic regions, and we utilized ECG measurements
for fiducial point detection. On the other hand, for the implementation of wearable
systems that only rely on MCG measurements, it is crucial to develop algorithms
that perform fiducial point detection without depending on ECG.
One strength of our study was maintaining a consistent heart rate of 80 bpm
throughout the MCG recordings. This constancy was important, particularly in the
comparison of LVET and QS2, as changes in heart rate could potentially impact
CHAPTER 7. DISCUSSION 49
these parameters [99]. Additionally, the experimental design incorporated stable
and undisturbed conditions, with quiet breathing, and aimed to minimize signal
interference caused by respiratory sounds and other variations that could potentially
affect the accuracy of hypothesis testing.
8 Conclusion
In this thesis, we conducted an analysis of MCG waveforms in HFrEF patients
undergoing CRT pacing mode and AAI pacing. CRT pacing, involving ventricular
synchronization, demonstrated significantly higher energy and vertical range features
compared to AAI pacing and indicated improved mechanical cardiac function. No
significant differences were found in waveform features around the early diastolic
window, LVET, and QS2 between pacing modes.
Our results suggest that mechanical signals measured on the lower chest wall
through accelerometers and gyroscopes show substantial changes during CRT pacing
and reflects an enhancement in mechanical cardiac function due to resynchronization
therapy.
The advancements in Micro Electro Mechanical Systems (MEMS) have led to
the development of capable wearable devices, such as MCG, based on MEMS ac-
celerometers and gyroscopes. The widespread availability of smart devices with
MEMS units and growing evidence of MCG’s potential utilities make it a promising
technology for clinical applications.
The thesis demonstrates the capability of MCG waveforms in showing improve-
ments in cardiac function and offers promise for predicting clinical responses to
CRT, optimizing CRT procedures, and facilitating post-implementation follow-ups.
However, the study’s limitation in the number of subjects necessitates larger patient
cohorts for validation and a more comprehensive exploration of MCG’s utility for
CHAPTER 8. CONCLUSION 51
heart failure patients.
Furthermore, future studies can focus on MCG signals collected via smartphones,
developing analysis pipelines for these devices to contribute to the advancement of
health monitoring applications in home settings.
References
[1] C. W. Tsao, A. W. Aday, Z. I. Almarzooq, et al., “Heart disease and stroke
statistics—2023 update: A report from the american heart association”, Cir-
culation, vol. 147, no. 8, e93–e621, 2023.
[2] Centers for disease control and prevention and national center for health
statistics, Jan. 2024. [Online]. Available: https://www.cdc.gov/nchs/
nhanes/.
[3] N. C. for Health Statistics et al., “Selected circulatory diseases among adults
aged 18 and over, by selected characteristics: United states, 2018”, vol, vol. 6,
pp. 1–9, 2018.
[4] S. Irawati, R. Wasir, A. Floriaan Schmidt, et al., “Long-term incidence and
risk factors of cardiovascular events in asian populations: Systematic review
and meta-analysis of population-based cohort studies”, Current Medical Re-
search and Opinion, vol. 35, no. 2, pp. 291–299, 2019.
[5] P. A. Heidenreich, B. Bozkurt, D. Aguilar, et al., “2022 aha/acc/hfsa guideline
for the management of heart failure: A report of the american college of
cardiology/american heart association joint committee on clinical practice
guidelines”, Journal of the American College of Cardiology, vol. 79, no. 17,
e263–e421, 2022.
REFERENCES 53
[6] P. A. Heidenreich, N. M. Albert, L. A. Allen, et al., “Forecasting the impact
of heart failure in the united states: A policy statement from the american
heart association”, Circulation: Heart Failure, vol. 6, no. 3, pp. 606–619, 2013.
[7] S. M. Dunlay, S. A. Weston, S. J. Jacobsen, and V. L. Roger, “Risk factors for
heart failure: A population-based case-control study”, The American journal
of medicine, vol. 122, no. 11, pp. 1023–1028, 2009.
[8] J. E. Ho, D. Enserro, F. P. Brouwers, et al., “Predicting heart failure with
preserved and reduced ejection fraction: The international collaboration on
heart failure subtypes”, Circulation: Heart Failure, vol. 9, no. 6, e003116,
2016.
[9] M. Merlo, A. Pivetta, B. Pinamonti, et al., “Long-term prognostic impact
of therapeutic strategies in patients with idiopathic dilated cardiomyopathy:
Changing mortality over the last 30 years”, European journal of heart failure,
vol. 16, no. 3, pp. 317–324, 2014.
[10] C. Balla and R. Cappato, “When to choose cardiac resynchronization therapy
in chronic heart failure: Type and duration of the conduction delay”, European
Heart Journal Supplements, vol. 21, no. Supplement_B, B31–B35, 2019.
[11] C. Linde, K. Ellenbogen, and F. A. McAlister, “Cardiac resynchronization
therapy (crt): Clinical trials, guidelines, and target populations”, Heart Rhythm,
vol. 9, no. 8, S3–S13, 2012.
[12] May 2023. [Online]. Available: https : / / www . heart . org / en / health -
topics/heart-attack/diagnosing-a-heart-attack/electrocardiogram-
ecg-or-ekg.
[13] M. R. Bristow, L. A. Saxon, J. Boehmer, et al., “Cardiac-resynchronization
therapy with or without an implantable defibrillator in advanced chronic
REFERENCES 54
heart failure”, New England Journal of Medicine, vol. 350, no. 21, pp. 2140–
2150, 2004.
[14] J. G. Cleland, J.-C. Daubert, E. Erdmann, et al., “The effect of cardiac
resynchronization on morbidity and mortality in heart failure”, New England
Journal of Medicine, vol. 352, no. 15, pp. 1539–1549, 2005.
[15] A. F. Members, J. J. McMurray, S. Adamopoulos, et al., “Esc guidelines for
the diagnosis and treatment of acute and chronic heart failure 2012: The task
force for the diagnosis and treatment of acute and chronic heart failure 2012
of the european society of cardiology. developed in collaboration with the
heart failure association (hfa) of the esc”, European heart journal, vol. 33,
no. 14, pp. 1787–1847, 2012.
[16] W. Heggermont, A. Auricchio, and M. Vanderheyden, “Biomarkers to predict
the response to cardiac resynchronization therapy”, EP Europace, vol. 21,
no. 11, pp. 1609–1620, 2019.
[17] T. F. Chairs, J.-C. Daubert, L. Saxon, et al., “2012 ehra/hrs expert consen-
sus statement on cardiac resynchronization therapy in heart failure: Implant
and follow-up recommendations and management: A registered branch of the
european society of cardiology (esc), and the heart rhythm society; and in
collaboration with the heart failure society of america (hfsa), the american
society of echocardiography (ase), the american heart association (aha), the
european association of echocardiography (eae) of the esc and the heart fail-
ure association of the esc (hfa).”, Europace, vol. 14, no. 9, pp. 1236–1286,
2012.
[18] J. G. Cleland, N. Freemantle, J.-C. Daubert, W. Toff, F. Leisch, and L.
Tavazzi, “Long-term effect of cardiac resynchronisation in patients reporting
mild symptoms of heart failure: A report from the care-hf study”, Heart, 2007.
REFERENCES 55
[19] A. Auricchio and F. W. Prinzen, “Non-responders to cardiac resynchroniza-
tion therapy”, Circulation Journal, vol. 75, 2011.
[20] D. M. Salerno, “Seismocardiography: A new technique for recording cardiac
vibrations. concept, method, and initial observations”, Journal of cardiovas-
cular technology, vol. 9, no. 2, pp. 111–118, 1990.
[21] B. S. Bozhenko, “Seismocardiography–a new method in the study of func-
tional conditions of the heart”, Terapevticheskii arkhiv, vol. 33, pp. 55–64,
1961.
[22] U. Meriheinä, M. Juppo, T. Koivisto, M. Pänäälä, K. Sairanen, and M. Grön-
holm, Heart monitoring system, US Patent 10,178,964, Jan. 2019.
[23] M. Jafari Tadi, E. Lehtonen, A. Saraste, et al., “Gyrocardiography: A new
non-invasive monitoring method for the assessment of cardiac mechanics and
the estimation of hemodynamic variables”, Scientific reports, vol. 7, no. 1,
p. 6823, 2017.
[24] T. Koivisto, O. Lahdenoja, T. Hurnanen, et al., “Mechanocardiography-based
measurement system indicating changes in heart failure patients during hos-
pital admission and discharge”, Sensors, vol. 22, no. 24, p. 9781, 2022.
[25] M. J. Tadi, E. Lehtonen, M. Pankäälä, et al., “Gyrocardiography: A new non-
invasive approach in the study of mechanical motions of the heart. concept,
method and initial observations”, in 2016 38th Annual International Con-
ference of the IEEE engineering in medicine and biology society (EMBC),
IEEE, 2016, pp. 2034–2037.
[26] S. Morra, L. Pitisci, F. Su, et al., “Quantification of cardiac kinetic energy
and its changes during transmural myocardial infarction assessed by multi-
dimensional seismocardiography”, Frontiers in Cardiovascular Medicine, vol. 8,
p. 603 319, 2021.
REFERENCES 56
[27] A. Taebi, B. E. Solar, A. J. Bomar, R. H. Sandler, and H. A. Mansy, “Recent
advances in seismocardiography”, Vibration, vol. 2, no. 1, pp. 64–86, 2019.
[28] P. Dehkordi, F. Khosrow-Khavar, M. Di Rienzo, et al., “Comparison of dif-
ferent methods for estimating cardiac timings: A comprehensive multimodal
echocardiography investigation”, Frontiers in physiology, vol. 10, p. 1057,
2019.
[29] K. Sørensen, P. Søgaard, K. Emerek, A. S. Jensen, J. J. Struijk, and S. E.
Schmidt, “Seismocardiography as a tool for assessment of bi-ventricular pac-
ing”, Physiological Measurement, vol. 43, no. 10, p. 105 007, 2022.
[30] A. Agam, P. Søgaard, K. Kragholm, et al., “Correlation between diastolic seis-
mocardiography variables and echocardiography variables”, European Heart
Journal-Digital Health, vol. 3, no. 3, pp. 465–472, 2022.
[31] T. Hurnanen, E. Lehtonen, M. J. Tadi, et al., “Automated detection of atrial
fibrillation based on time–frequency analysis of seismocardiograms”, IEEE
Journal of Biomedical and Health Informatics, vol. 21, no. 5, pp. 1233–1241,
2016.
[32] O. Lahdenoja, T. Hurnanen, Z. Iftikhar, et al., “Atrial fibrillation detection
via accelerometer and gyroscope of a smartphone”, IEEE Journal of Biomed-
ical and Health Informatics, vol. 22, no. 1, pp. 108–118, 2017.
[33] M. Becker, A. Roehl, U. Siekmann, et al., “Simplified detection of myocardial
ischemia by seismocardiography”, Herz, vol. 39, no. 5, p. 586, 2014.
[34] F. Tokmak, T. Koivisto, O. Lahdenoja, T. Vasankari, S. Jaakkola, and K. J.
Airaksinen, “Mechanocardiography detects improvement of systolic function
caused by resynchronization pacing”, Physiological Measurement, vol. 44,
no. 12, p. 125 009, 2023.
REFERENCES 57
[35] S. Doris, D. T. Lee, J. Woo, and D. R. Thompson, “Correlates of psycho-
logical distress in elderly patients with congestive heart failure”, Journal of
psychosomatic research, vol. 57, no. 6, pp. 573–581, 2004.
[36] I. Ekman, J. G. Cleland, K. Swedberg, A. Charlesworth, M. Metra, and
P. A. Poole-Wilson, “Symptoms in patients with heart failure are prognostic
predictors: Insights from comet”, Journal of cardiac failure, vol. 11, no. 4,
pp. 288–292, 2005.
[37] C. D. Kemp and J. V. Conte, “The pathophysiology of heart failure”, Car-
diovascular Pathology, vol. 21, no. 5, pp. 365–371, 2012.
[38] T. A. Foley, S. V. Mankad, N. S. Anavekar, et al., “Measuring left ventricular
ejection fraction-techniques and potential pitfalls”, Eur Cardiol, vol. 8, no. 2,
pp. 108–114, 2012.
[39] S. Iuliano, S. G. Fisher, P. E. Karasik, R. D. Fletcher, S. N. Singh, D. of Vet-
erans Affairs Survival Trial of Antiarrhythmic Therapy in Congestive Heart
Failure, et al., “Qrs duration and mortality in patients with congestive heart
failure”, American heart journal, vol. 143, no. 6, pp. 1085–1091, 2002.
[40] S. Baldasseroni, C. Opasich, M. Gorini, et al., “Left bundle-branch block
is associated with increased 1-year sudden and total mortality rate in 5517
outpatients with congestive heart failure: A report from the italian network
on congestive heart failure”, American heart journal, vol. 143, no. 3, pp. 398–
405, 2002.
[41] H. J. Shenkman, V. Pampati, A. K. Khandelwal, et al., “Congestive heart
failure and qrs duration: Establishing prognosis study”, Chest, vol. 122, no. 2,
pp. 528–534, 2002.
[42] M. Glikson, J. C. Nielsen, M. B. Kronborg, et al., “2021 esc guidelines on
cardiac pacing and cardiac resynchronization therapy: Developed by the task
REFERENCES 58
force on cardiac pacing and cardiac resynchronization therapy of the european
society of cardiology (esc) with the special contribution of the european heart
rhythm association (ehra)”, EP Europace, vol. 24, no. 1, pp. 71–164, 2022.
[43] W. T. Abraham, J. B. Young, A. R. León, et al., “Effects of cardiac resyn-
chronization on disease progression in patients with left ventricular systolic
dysfunction, an indication for an implantable cardioverter-defibrillator, and
mildly symptomatic chronic heart failure”, Circulation, vol. 110, no. 18, pp. 2864–
2868, 2004.
[44] A. Auricchio, C. Stellbrink, S. Sack, et al., “Long-term clinical effect of hemo-
dynamically optimized cardiac resynchronization therapy in patients with
heart failure and ventricular conduction delay”, Journal of the American Col-
lege of Cardiology, vol. 39, no. 12, pp. 2026–2033, 2002.
[45] J. G. Cleland, J.-C. Daubert, E. Erdmann, et al., “Longer-term effects of
cardiac resynchronization therapy on mortality in heart failure [the cardiac
resynchronization-heart failure (care-hf) trial extension phase]”, European
heart journal, vol. 27, no. 16, pp. 1928–1932, 2006.
[46] C. Linde, W. T. Abraham, M. R. Gold, et al., “Randomized trial of cardiac
resynchronization in mildly symptomatic heart failure patients and in asymp-
tomatic patients with left ventricular dysfunction and previous heart failure
symptoms”, Journal of the American College of Cardiology, vol. 52, no. 23,
pp. 1834–1843, 2008.
[47] E. S. Chung, A. R. Leon, L. Tavazzi, et al., “Results of the predictors of
response to crt (prospect) trial”, Circulation, vol. 117, no. 20, pp. 2608–2616,
2008.
[48] J. Cleland, N. Freemantle, S. Ghio, et al., “Predicting the long-term effects of
cardiac resynchronization therapy on mortality from baseline variables and
REFERENCES 59
the early response: A report from the care-hf (cardiac resynchronization in
heart failure) trial”, Journal of the American College of Cardiology, vol. 52,
no. 6, pp. 438–445, 2008.
[49] Y. D’Mello, J. Skoric, S. Xu, et al., “Real-time cardiac beat detection and
heart rate monitoring from combined seismocardiography and gyrocardiog-
raphy”, Sensors, vol. 19, no. 16, p. 3472, 2019.
[50] A. Taebi, B. E. Solar, A. J. Bomar, R. H. Sandler, and H. A. Mansy, “Re-
cent advances in seismocardiography”, Vibration, vol. 2, no. 1, pp. 64–86,
2019, issn: 2571-631X. doi: 10.3390/vibration2010005. [Online]. Avail-
able: https://www.mdpi.com/2571-631X/2/1/5.
[51] A. Taebi and H. Mansy, “Effect of noise on time-frequency analysis of vibro-
cardiographic signals”, Journal of bioengineering & biomedical science, vol. 6,
no. 4, 2016.
[52] O. T. Inan, P.-F. Migeotte, K.-S. Park, et al., “Ballistocardiography and
seismocardiography: A review of recent advances”, IEEE journal of biomedical
and health informatics, vol. 19, no. 4, pp. 1414–1427, 2014.
[53] V. Gurev, K. Tavakolian, J. Constantino, B. Kaminska, A. P. Blaber, and
N. A. Trayanova, “Mechanisms underlying isovolumic contraction and ejec-
tion peaks in seismocardiogram morphology”, Journal of medical and biolog-
ical engineering, vol. 32, no. 2, p. 103, 2012.
[54] R. S. Crow, P. Hannan, D. Jacobs, L. Hedquist, and D. M. Salerno, “Relation-
ship between seismocardiogram and echocardiogram for events in the cardiac
cycle”, American journal of noninvasive cardiology, vol. 8, no. 1, pp. 39–46,
1994.
REFERENCES 60
[55] K. Sørensen, S. E. Schmidt, A. S. Jensen, P. Søgaard, and J. J. Struijk, “Def-
inition of fiducial points in the normal seismocardiogram”, Scientific reports,
vol. 8, no. 1, p. 15 455, 2018.
[56] M. J. Tadi, T. Koivisto, M. Pänkäälä, et al., “A new algorithm for segmenta-
tion of cardiac quiescent phases and cardiac time intervals using seismocardio-
graphy”, in Sixth International Conference on Graphic and Image Processing
(ICGIP 2014), SPIE, vol. 9443, 2015, pp. 571–577.
[57] K. Tavakolian, “Characterization and analysis of seismocardiogram for esti-
mation of hemodynamic parameters”, 2010.
[58] P. Reant, M. Dijos, E. Donal, et al., “Systolic time intervals as simple echocar-
diographic parameters of left ventricular systolic performance: Correlation
with ejection fraction and longitudinal two-dimensional strain”, European
Journal of Echocardiography, vol. 11, no. 10, pp. 834–844, 2010.
[59] H. Boudoulas, “Systolic time intervals”, European heart journal, vol. 11, no. suppl_I,
pp. 93–104, 1990.
[60] T. Koivisto, O. Lahdenoja, T. Hurnanen, et al., “Mechanocardiography in
the detection of acute st elevation myocardial infarction: The mechano-stemi
study”, Sensors, vol. 22, no. 12, p. 4384, 2022.
[61] D. Rai, H. K. Thakkar, S. S. Rajput, J. Santamaria, C. Bhatt, and F. Roca,
“A comprehensive review on seismocardiogram: Current advancements on ac-
quisition, annotation, and applications”, Mathematics, vol. 9, no. 18, p. 2243,
2021.
[62] F. Tokmak and B. Semiz, “Unveiling the relationships between seismocardio-
gram signals, physical activity types and metabolic equivalent of task scores”,
IEEE Transactions on Biomedical Engineering, vol. 70, no. 2, pp. 479–487,
2022.
REFERENCES 61
[63] D. S. Morillo, J. L. R. Ojeda, L. F. C. Foix, and A. L. Jiménez, “An accelerometer-
based device for sleep apnea screening”, IEEE transactions on information
technology in biomedicine, vol. 14, no. 2, pp. 491–499, 2009.
[64] B. Semiz, A. M. Carek, J. C. Johnson, et al., “Non-invasive wearable patch
utilizing seismocardiography for peri-operative use in surgical patients”, IEEE
journal of biomedical and health informatics, vol. 25, no. 5, pp. 1572–1582,
2020.
[65] P. Castiglioni, A. Faini, G. Parati, and M. Di Rienzo, “Wearable seismocar-
diography”, in 2007 29th annual international conference of the IEEE engi-
neering in medicine and biology society, IEEE, 2007, pp. 3954–3957.
[66] C. S. Spencer et al., “Nurses with undiagnosed hearing loss: Implications for
practice”, Online Journal of Issues in Nursing, vol. 20, no. 1, p. 56, 2015.
[67] A. L. Goldberger, Z. D. Goldberger, and A. Shvilkin, “Chapter 4 - ecg
leads”, in Goldberger’s Clinical Electrocardiography (Ninth Edition), A. L.
Goldberger, Z. D. Goldberger, and A. Shvilkin, Eds., Ninth Edition, Else-
vier, 2018, pp. 21–31, isbn: 978-0-323-40169-2. doi: https://doi.org/10.
1016/B978-0-323-40169-2.00004-4. [Online]. Available: https://www.
sciencedirect.com/science/article/pii/B9780323401692000044.
[68] A. L. Goldberger, Z. D. Goldberger, and A. Shvilkin, “Chapter 2 - ecg basics:
Waves, intervals, and segments”, in Goldberger’s Clinical Electrocardiogra-
phy (Ninth Edition), A. L. Goldberger, Z. D. Goldberger, and A. Shvilkin,
Eds., Ninth Edition, Elsevier, 2018, pp. 6–10, isbn: 978-0-323-40169-2. doi:
https://doi.org/10.1016/B978-0-323-40169-2.00002-0. [Online].
Available: https://www.sciencedirect.com/science/article/pii/
B9780323401692000020.
REFERENCES 62
[69] T. Reichlin, R. Abächerli, R. Twerenbold, et al., “Advanced ecg in 2016: Is
there more than just a tracing?”, Swiss medical weekly, vol. 146, no. 1718,
w14303–w14303, 2016.
[70] N. Rafie, A. H. Kashou, and P. A. Noseworthy, “Ecg interpretation: Clinical
relevance, challenges, and advances”, Hearts, vol. 2, no. 4, pp. 505–513, 2021.
[71] Y. Yoon, J. H. Cho, and G. Yoon, “Non-constrained blood pressure monitor-
ing using ecg and ppg for personal healthcare”, Journal of medical systems,
vol. 33, pp. 261–266, 2009.
[72] P. Hamilton, “Open source ecg analysis”, in Computers in cardiology, IEEE,
2002, pp. 101–104.
[73] J. Pan and W. J. Tompkins, “A real-time qrs detection algorithm”, IEEE
transactions on biomedical engineering, no. 3, pp. 230–236, 1985.
[74] J. P. Martínez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna, “A
wavelet-based ecg delineator: Evaluation on standard databases”, IEEE Trans-
actions on biomedical engineering, vol. 51, no. 4, pp. 570–581, 2004.
[75] F. Emmert-Streib and M. Dehmer, “Understanding statistical hypothesis
testing: The logic of statistical inference”, Machine Learning and Knowledge
Extraction, vol. 1, no. 3, pp. 945–962, 2019.
[76] J. Neyman and E. S. Pearson, “Ix. on the problem of the most efficient tests
of statistical hypotheses”, Philosophical Transactions of the Royal Society of
London. Series A, Containing Papers of a Mathematical or Physical Charac-
ter, vol. 231, no. 694-706, pp. 289–337, 1933.
[77] Student, “The probable error of a mean”, Biometrika, vol. 6, no. 1, pp. 1–25,
1908, issn: 00063444. [Online]. Available: http://www.jstor.org/stable/
2331554 (visited on 03/17/2024).
[78] F. Galton, Natural inheritance. Macmillan, 1889, vol. 42.
REFERENCES 63
[79] H. B. Mann and D. R. Whitney, “On a Test of Whether one of Two Random
Variables is Stochastically Larger than the Other”, The Annals of Mathe-
matical Statistics, vol. 18, no. 1, pp. 50–60, 1947. doi: 10 . 1214 / aoms /
1177730491. [Online]. Available: https://doi.org/10.1214/aoms/1177730491.
[80] F. Wilcoxon, “Individual comparisons by ranking methods”, in Breakthroughs
in Statistics: Methodology and Distribution, Springer, 1992, pp. 196–202.
[81] W. H. Kruskal and W. A. Wallis, “Use of ranks in one-criterion variance
analysis”, Journal of the American Statistical Association, vol. 47, no. 260,
pp. 583–621, 1952. doi: 10.1080/01621459.1952.10483441. eprint: https:
//www.tandfonline.com/doi/pdf/10.1080/01621459.1952.10483441.
[Online]. Available: https://www.tandfonline.com/doi/abs/10.1080/
01621459.1952.10483441.
[82] D. Sundararajan, Digital signal processing: theory and practice. World Scien-
tific, 2003.
[83] A. Averbuch, D. Lazar, and M. Israeli, “Image compression using wavelet
transform and multiresolution decomposition”, IEEE Transactions on Image
Processing, vol. 5, no. 1, pp. 4–15, 1996.
[84] Q. Pan, L. Zhang, G. Dai, and H. Zhang, “Two denoising methods by wavelet
transform”, IEEE transactions on signal processing, vol. 47, no. 12, pp. 3401–
3406, 1999.
[85] Q. Zhao and L. Zhang, “Ecg feature extraction and classification using wavelet
transform and support vector machines”, in 2005 International Conference on
Neural Networks and Brain, IEEE, vol. 2, 2005, pp. 1089–1092.
[86] A. N. Akansu, W. A. Serdijn, and I. W. Selesnick, “Emerging applications of
wavelets: A review”, Physical communication, vol. 3, no. 1, pp. 1–18, 2010.
REFERENCES 64
[87] V. N. Varghees and K. Ramachandran, “Effective heart sound segmentation
and murmur classification using empirical wavelet transform and instanta-
neous phase for electronic stethoscope”, IEEE Sensors Journal, vol. 17, no. 12,
pp. 3861–3872, 2017.
[88] L. Huiying, L. Sakari, and H. Iiro, “A heart sound segmentation algorithm
using wavelet decomposition and reconstruction”, in Proceedings of the 19th
Annual International Conference of the IEEE Engineering in Medicine and
Biology Society. ’Magnificent Milestones and Emerging Opportunities in Med-
ical Engineering’ (Cat. No.97CH36136), vol. 4, 1997, 1630–1633 vol.4. doi:
10.1109/IEMBS.1997.757028.
[89] T. Choudhary, L. Sharma, and M. K. Bhuyan, “Automatic detection of aortic
valve opening using seismocardiography in healthy individuals”, IEEE journal
of biomedical and health informatics, vol. 23, no. 3, pp. 1032–1040, 2018.
[90] T. Choudhary, L. Sharma, and M. K. Bhuyan, “Heart sound extraction from
sternal seismocardiographic signal”, IEEE Signal Processing Letters, vol. 25,
no. 4, pp. 482–486, 2018.
[91] M. A. García-González, A. Argelagós-Palau, M. Fernández-Chimeno, and J.
Ramos-Castro, “A comparison of heartbeat detectors for the seismocardio-
gram”, in Computing in Cardiology 2013, 2013, pp. 461–464.
[92] B. Ergen, Y. Tatar, and H. O. Gulcur, “Time–frequency analysis of phono-
cardiogram signals using wavelet transform: A comparative study”, Computer
methods in biomechanics and biomedical engineering, vol. 15, no. 4, pp. 371–
381, 2012.
[93] T. Oskiper and R. Watrous, “Detection of the first heart sound using a time-
delay neural network”, in Computers in Cardiology, IEEE, 2002, pp. 537–
540.
REFERENCES 65
[94] L. Huiying, L. Sakari, and H. Iiro, “A heart sound segmentation algorithm
using wavelet decomposition and reconstruction”, in Proceedings of the 19th
Annual International Conference of the IEEE Engineering in Medicine and
Biology Society.’Magnificent Milestones and Emerging Opportunities in Med-
ical Engineering’(Cat. No. 97CH36136), IEEE, vol. 4, 1997, pp. 1630–1633.
[95] C. N. Gupta, R. Palaniappan, S. Swaminathan, and S. M. Krishnan, “Neural
network classification of homomorphic segmented heart sounds”, Applied soft
computing, vol. 7, no. 1, pp. 286–297, 2007.
[96] K. Pandia, O. T. Inan, G. T. Kovacs, and L. Giovangrandi, “Extracting res-
piratory information from seismocardiogram signals acquired on the chest
using a miniature accelerometer”, Physiological measurement, vol. 33, no. 10,
p. 1643, 2012.
[97] I. Elnaggar, J. Pykäri, T. Hurnanen, et al., “Cardiac time intervals derived
from electrocardiography and seismocardiography in different patient groups”,
in 2022 Computing in Cardiology (CinC), vol. 498, 2022, pp. 1–4. doi: 10.
22489/CinC.2022.370.
[98] J. L. Willems, J. Roelandt, H. DE GEEST, H. Kesteloot, and J. V. JOOSSENS,
“The left ventricular ejection time in elderly subjects”, Circulation, vol. 42,
no. 1, pp. 37–42, 1970.
[99] A. M. Weissler, L. C. Harris, and G. D. White, “Left ventricular ejection
time index in man”, Journal of Applied Physiology, vol. 18, no. 5, pp. 919–
923, 1963.
[100] G. Andelfinger, J.-C. Fouron, S.-E. Sonesson, and F. Proulx, “Reference val-
ues for time intervals between atrial and ventricular contractions of the fetal
heart measured by two doppler techniques”, American Journal of Cardiology,
vol. 88, no. 12, pp. 1433–1436, 2001.
REFERENCES 66
[101] A. S. Alhakak, F. J. Olsen, K. G. Skaarup, et al., “Age-and sex-based normal
reference ranges of the cardiac time intervals: The copenhagen city heart
study”, Clinical Research in Cardiology, pp. 1–13, 2023.
[102] S. Siecinski, E. J. Tkacz, and P. S. Kostka, “Comparison of hrv indices ob-
tained from ecg and scg signals from cebs database”, Biomedical engineering
online, vol. 18, pp. 1–15, 2019.
[103] G. Shafiq, S. Tatinati, and K. C. Veluvolu, “Automatic annotation of peaks in
seismocardiogram for systolic time intervals”, in 2016 38th Annual Interna-
tional Conference of the IEEE Engineering in Medicine and Biology Society
(EMBC), IEEE, 2016, pp. 2672–2675.
[104] M. J. Tadi, E. Lehtonen, T. Hurnanen, et al., “A real-time approach for heart
rate monitoring using a hilbert transform in seismocardiograms”, Physiolog-
ical measurement, vol. 37, no. 11, p. 1885, 2016.
[105] H. Sakoe and S. Chiba, “Dynamic programming algorithm optimization for
spoken word recognition”, IEEE transactions on acoustics, speech, and signal
processing, vol. 26, no. 1, pp. 43–49, 1978.
[106] D. J. Berndt and J. Clifford, “Using dynamic time warping to find patterns in
time series”, in Proceedings of the 3rd international conference on knowledge
discovery and data mining, 1994, pp. 359–370.
[107] K. Wang and T. Gasser, “Alignment of curves by dynamic time warping”,
The annals of Statistics, vol. 25, no. 3, pp. 1251–1276, 1997.
[108] E. J. Keogh and M. J. Pazzani, “Scaling up dynamic time warping to mas-
sive datasets”, in Principles of Data Mining and Knowledge Discovery: Third
European Conference, PKDD’99, Prague, Czech Republic, September 15-18,
1999. Proceedings 3, Springer, 1999, pp. 1–11.
REFERENCES 67
[109] Y.-S. Jeong, M. K. Jeong, and O. A. Omitaomu, “Weighted dynamic time
warping for time series classification”, Pattern recognition, vol. 44, no. 9,
pp. 2231–2240, 2011.
[110] M. Yadav and M. A. Alam, “Dynamic time warping (dtw) algorithm in
speech: A review”, International Journal of Research in Electronics and Com-
puter Engineering, vol. 6, no. 1, pp. 524–528, 2018.
[111] C.-H. Chen, W.-Y. Lin, and M.-Y. Lee, “Computer-aided detection of fiducial
points in seismocardiography through dynamic time warping”, Biosensors,
vol. 12, no. 6, p. 374, 2022.
[112] B.-J. Yoon, “Hidden markov models and their applications in biological se-
quence analysis”, Current genomics, vol. 10, no. 6, pp. 402–415, 2009.
[113] J. Wahlström, I. Skog, P. Händel, et al., “A hidden markov model for seismo-
cardiography”, IEEE Transactions on Biomedical Engineering, vol. 64, no. 10,
pp. 2361–2372, 2017.
[114] B. Semiz, A. M. Carek, J. C. Johnson, et al., “Non-invasive wearable patch
utilizing seismocardiography for peri-operative use in surgical patients”, IEEE
Journal of Biomedical and Health Informatics, vol. 25, no. 5, pp. 1572–1582,
2021. doi: 10.1109/JBHI.2020.3032938.
[115] F. Tokmak and B. Semiz, “Unveiling the relationships between seismocardio-
gram signals, physical activity types and metabolic equivalent of task scores”,
IEEE Transactions on Biomedical Engineering, vol. 70, no. 2, pp. 479–487,
2023. doi: 10.1109/TBME.2022.3194594.
[116] V. Gemignani, E. Bianchini, F. Faita, et al., “Transthoracic sensor for non-
invasive assessment of left ventricular contractility: Validation in a minipig
model of chronic heart failure”, Pacing and clinical electrophysiology, vol. 33,
no. 7, pp. 795–803, 2010.
REFERENCES 68
[117] Z. M. Işilay Zeybek, V. Racca, A. Pezzano, M. Tavanelli, and M. Di Rienzo,
“Can seismocardiogram fiducial points be used for the routine estimation of
cardiac time intervals in cardiac patients?”, Frontiers in Physiology, vol. 13,
p. 372, 2022.
[118] M. Di Rienzo, E. Vaini, P. Castiglioni, P. Meriggi, and F. Rizzo, “Beat-to-beat
estimation of lvet and qs2 indices of cardiac mechanics from wearable seismo-
cardiography in ambulant subjects”, in 2013 35th Annual International Con-
ference of the IEEE Engineering in Medicine and Biology Society (EMBC),
IEEE, 2013, pp. 7017–7020.
[119] A. S. Alhakak, J. R. Teerlink, J. Lindenfeld, M. Böhm, G. M. Rosano, and
T. Biering-Sørensen, “The significance of left ventricular ejection time in
heart failure with reduced ejection fraction”, European journal of heart fail-
ure, vol. 23, no. 4, pp. 541–551, 2021.
[120] T. Biering-Sørensen, G. Querejeta Roca, S. M. Hegde, et al., “Left ventric-
ular ejection time is an independent predictor of incident heart failure in a
community-based cohort”, European journal of heart failure, vol. 20, no. 7,
pp. 1106–1114, 2018.
[121] A. Dodek, J. R. Burg, and F. E. Kloster, “Systolic time intervals in chronic
hypertension: Alterations and response to treatment”, Chest, vol. 68, no. 1,
pp. 51–55, 1975.
[122] R. P. Lewis, H. Boudoulas, T. G. Welch, and W. F. Forester, “Usefulness of
systolic time intervals in coronary artery disease”, The American journal of
cardiology, vol. 37, no. 5, pp. 787–796, 1976.
[123] W. Chen and D. Gibson, “Mechanisms of prolongation of pre-ejection period
in patients with left ventricular disease.”, Heart, vol. 42, no. 3, pp. 304–310,
1979.
REFERENCES 69
[124] H. Schächinger, M. Weinbacher, A. Kiss, R. Ritz, and W. Langewitz, “Car-
diovascular indices of peripheral and central sympathetic activation”, Psy-
chosomatic medicine, vol. 63, no. 5, pp. 788–796, 2001.
[125] H. Boudoulas, Y. H. Sohn, W. O’neill, R. Brown, and A. M. Weissler, “The
qt> qs2 syndrome: A new mortality risk indicator in coronary artery disease”,
The American Journal of Cardiology, vol. 50, no. 6, pp. 1229–1235, 1982.
[126] P. Bifulco, G. Gargiulo, G. d’Angelo, et al., “Monitoring of respiration, seis-
mocardiogram and heart sounds by a pvdf piezo film sensor”, Measurement,
vol. 11, pp. 786–789, 2014.
[127] D. Makowski, T. Pham, Z. J. Lau, et al., “NeuroKit2: A python toolbox for
neurophysiological signal processing”, Behavior Research Methods, vol. 53,
no. 4, pp. 1689–1696, Feb. 2021. doi: 10.3758/s13428- 020- 01516- y.
[Online]. Available: https://doi.org/10.3758%2Fs13428-020-01516-y.
[128] K. Noda, H. Endo, T. Kadosaka, et al., “Comparison of the measured pre-
ejection periods and left ventricular ejection times between echocardiography
and impedance cardiography for optimizing cardiac resynchronization ther-
apy”, Journal of arrhythmia, vol. 33, no. 2, pp. 130–133, 2017.
[129] D. J. Friedman, S. M. Al-Khatib, F. Dalgaard, et al., “Cardiac resynchroniza-
tion therapy improves outcomes in patients with intraventricular conduction
delay but not right bundle branch block: A patient-level meta-analysis of ran-
domized controlled trials”, Circulation, vol. 147, no. 10, pp. 812–823, 2023.
[130] A. S. Tang, G. A. Wells, M. Talajic, et al., “Cardiac-resynchronization ther-
apy for mild-to-moderate heart failure”, New England Journal of Medicine,
vol. 363, no. 25, pp. 2385–2395, 2010.
REFERENCES 70
[131] A. J. Moss, W. J. Hall, D. S. Cannom, et al., “Cardiac-resynchronization
therapy for the prevention of heart-failure events”, New England Journal of
Medicine, vol. 361, no. 14, pp. 1329–1338, 2009.
[132] C. Linde, M. R. Gold, W. T. Abraham, et al., “Long-term impact of car-
diac resynchronization therapy in mild heart failure: 5-year results from the
resynchronization reverses remodeling in systolic left ventricular dysfunction
(reverse) study”, European heart journal, vol. 34, no. 33, pp. 2592–2599, 2013.
[133] P. Steendijk, S. A. Tulner, J. J. Bax, et al., “Hemodynamic effects of long-
term cardiac resynchronization therapy: Analysis by pressure-volume loops”,
Circulation, vol. 113, no. 10, pp. 1295–1304, 2006.
[134] C.-M. Yu, E. Chau, J. E. Sanderson, et al., “Tissue doppler echocardiographic
evidence of reverse remodeling and improved synchronicity by simultaneously
delaying regional contraction after biventricular pacing therapy in heart fail-
ure”, Circulation, vol. 105, no. 4, pp. 438–445, 2002.
[135] C. Leclercq, S. Cazeau, H. Le Breton, et al., “Acute hemodynamic effects of
biventricular ddd pacing in patients with end-stage heart failure”, Journal of
the American College of Cardiology, vol. 32, no. 7, pp. 1825–1831, 1998.
[136] P. Bordachar, L. Labrousse, S. Ploux, et al., “Validation of a new nonin-
vasive device for the monitoring of peak endocardial acceleration in pigs:
Implications for optimization of pacing site and configuration”, Journal of
cardiovascular electrophysiology, vol. 19, no. 7, pp. 725–729, 2008.
[137] K. Tavakolian, A. P. Blaber, B. Ngai, and B. Kaminska, “Estimation of hemo-
dynamic parameters from seismocardiogram”, in 2010 Computing in Cardi-
ology, IEEE, 2010, pp. 1055–1058.
[138] S. Dornbush and A. E. Turnquest, “Physiology, heart sounds”, 2019.
REFERENCES 71
[139] W. P. Harvey, “Heart sounds and murmurs”, Circulation, vol. 30, no. 2,
pp. 262–271, 1964.
[140] S. Schmidt, C. Holst-Hansen, C. Graff, E. Toft, and J. Struijk, “Detection of
coronary artery disease with an electronic stethoscope”, in 2007 Computers
in Cardiology, IEEE, 2007, pp. 757–760.
[141] E. D. Folland, B. J. Kriegel, W. G. Henderson, K. E. Hammermeister, G. K.
Sethi, and P. in the Veterans Affairs Cooperative Study on Valvular Heart
Disease*, “Implications of third heart sounds in patients with valvular heart
disease”, New England Journal of Medicine, vol. 327, no. 7, pp. 458–462, 1992.
[142] M. Akay, J. Semmlow, W. Welkowitz, M. Bauer, and J. Kostis, “Detection of
coronary occlusions using autoregressive modeling of diastolic heart sounds”,
IEEE Transactions on Biomedical Engineering, vol. 37, no. 4, pp. 366–373,
1990.
[143] P. H. Thakur, Q. An, L. Swanson, Y. Zhang, and R. S. Gardner, “Haemo-
dynamic monitoring of cardiac status using heart sounds from an implanted
cardiac device”, ESC Heart Failure, vol. 4, no. 4, pp. 605–613, 2017.
[144] P. B. Hansen, A. A. Luisada, D. J. Miletich, and R. F. Albrecht, “Phonocar-
diography as a monitor of cardiac performance during anesthesia”, Anesthesia
& Analgesia, vol. 68, no. 3, pp. 385–387, 1989.
[145] S. Xiao, X. Guo, X. Sun, and Z. Xiao, “A relative value method for measuring
and evaluating cardiac reserve”, Biomedical engineering online, vol. 1, no. 1,
pp. 1–5, 2002.
[146] Y. Zheng, X. Guo, J. Qin, and S. Xiao, “Computer-assisted diagnosis for
chronic heart failure by the analysis of their cardiac reserve and heart sound
characteristics”, Computer methods and programs in biomedicine, vol. 122,
no. 3, pp. 372–383, 2015.
REFERENCES 72
[147] M. J. Tadi, E. Lehtonen, J. Teuho, et al., “A miniaturized mems motion pro-
cessing system for nuclear medicine imaging applications”, in 2016 Computing
in Cardiology Conference (CinC), IEEE, 2016, pp. 133–136.
[148] O. Lahdenoja, T. Humanen, M. J. Tadi, M. Pänkäälä, and T. Koivisto, “Heart
rate variability estimation with joint accelerometer and gyroscope sensing”,
in 2016 Computing in Cardiology Conference (CinC), IEEE, 2016, pp. 717–
720.
[149] O. T. Inan, M. Baran Pouyan, A. Q. Javaid, et al., “Novel wearable seis-
mocardiography and machine learning algorithms can assess clinical status
of heart failure patients”, Circulation: Heart Failure, vol. 11, no. 1, e004313,
2018.
[150] J. Blomster, O. Lahdenoja, K. Jafarian Dehkordi, et al., “Cardiosignal smart-
phone application detects atrial fibrillation in heart failure population”, Cir-
culation, vol. 144, no. Suppl_1, A12517–A12517, 2021.
[151] S. Mehrang, O. Lahdenoja, M. Kaisti, et al., “Classification of atrial fibrilla-
tion and acute decompensated heart failure using smartphone mechanocar-
diography: A multilabel learning approach”, IEEE Sensors Journal, vol. 20,
no. 14, pp. 7957–7968, 2020.
[152] A. Hossein, D. C. Mirica, J. Rabineau, et al., “Accurate detection of dobutamine-
induced haemodynamic changes by kino-cardiography: A randomised double-
blind placebo-controlled validation study”, Scientific reports, vol. 9, no. 1,
p. 10 479, 2019.
[153] M. M. H. Shandhi, J. Fan, J. A. Heller, M. Etemadi, L. Klein, and O. T.
Inan, “Estimation of changes in intracardiac hemodynamics using wearable
seismocardiography and machine learning in patients with heart failure: A
REFERENCES 73
feasibility study”, IEEE Transactions on Biomedical Engineering, vol. 69,
no. 8, pp. 2443–2455, 2022.
[154] G. M. Marcus, I. L. Gerber, B. H. McKeown, et al., “Association between
phonocardiographic third and fourth heart sounds and objective measures of
left ventricular function”, Jama, vol. 293, no. 18, pp. 2238–2244, 2005.
[155] A. Capucci, L. Santini, S. Favale, et al., “Preliminary experience with the
multisensor heartlogic algorithm for heart failure monitoring: A retrospective
case series report”, ESC heart failure, vol. 6, no. 2, pp. 308–318, 2019.
[156] R. Sandler, P. Gamage, M. K. Azad, et al., “Potential scg predictors of heart
failure readmission”, Journal of cardiac failure, vol. 26, no. 10, S87, 2020.
[157] N. Joshi, “The third heart sound.”, Southern medical journal, vol. 92, no. 8,
pp. 756–761, 1999.
[158] N. J. Mehta and I. A. Khan, “Third heart sound: Genesis and clinical im-
portance”, International journal of cardiology, vol. 97, no. 2, pp. 183–186,
2004.
[159] W. Am, “Systolic time intervals in heart failure in man.”, Circulation, vol. 37,
pp. 149–152, 1968.
[160] F. I. Marcus, V. Sorrell, J. Zanetti, et al., “Accelerometer-derived time inter-
vals during various pacing modes in patients with biventricular pacemakers:
Comparison with normals”, Pacing and Clinical Electrophysiology, vol. 30,
no. 12, pp. 1476–1481, 2007.
[161] M. Jafari and N. Ansari-Pour, “Why, when and how to adjust your p values?”,
Cell Journal (Yakhteh), vol. 20, no. 4, p. 604, 2019.
[162] S. Morra, A. Gauthey, A. Hossein, et al., “Influence of sympathetic activa-
tion on myocardial contractility measured with ballistocardiography and seis-
mocardiography during sustained end-expiratory apnea”, American Journal
REFERENCES 74
of Physiology-Regulatory, Integrative and Comparative Physiology, vol. 319,
no. 4, R497–R506, 2020.