1. Introduction
Recent advancements in sensor technology have made it possible to use low-powered, inexpensive sensor-based devices to monitor various physiological parameters related to human health such as heart rate, blood pressure, body temperature etc., [
1,
2]. Numerous applications are becoming reality in the light of wearable sensor technology such as diet monitoring, drug monitoring, activity detection, cardiac health monitoring, etc. [
3]. Among the envisioned applications, designing a robust cardiac health monitoring system to capture early signs of gradually developing cardiac anomalies using low-cost wearable body sensors is a growing interest among medical and research communities as it has serious consequences on human health. According to World Health Organization [
4], CVD is the leading cause of death in people across all age groups. Moreover, the recent report of the American Heart Association [
5] reveals that as high as 30% annual mortality rate is observed due to CVD in the United States and Canada. Under such circumstances, there is a growing need for a reliable and low-cost system that potentially aids in robust cardiac health monitoring.
Several cardiac events take place during successive heartbeats such as opening and closing of the heart valves, blood flow into vessels, contraction–relaxation of ventricular walls, etc. For a healthy person, cardiac events unfold at regular intervals in a predefined order. However, cardiac abnormalities such as myocardial ischemia, infarction, arrhythmias, etc., hinder the normal functioning of cardiac events leading to CVD. Such cardiac abnormalities may result in dizziness, nausea, chest pain, etc., which may lead to severe consequences such as heart attacks if not detected and taken care of in an early stage. In many instances, the irregular heartbeats known as ectopic heartbeats appear intermittently without showing any serious symptoms. Such ectopic heartbeats are becoming quite common among healthy population and are often getting unnoticed.
Nowadays, several clinical practices are used to monitor the cardiac abnormalities such as ECG, magnetic resonance imaging (MRI), computerized tomography scan (CT scan), Echocardiography (Echo), Nuclear myocardial perfusion scan etc., [
6]. Among the mentioned clinical practices, ECG is a well-established and widely adopted practice to monitor physiological activities of the heart. However, ECG provides shallow information on the functioning of various cardiac events. Moreover, a study conducted in [
7] recommends that sole use of ECG is not reliable for the diagnosis of serious cardiac abnormalities. On the other hand, MRI, CT scan, and Echo are reliable practices over ECG, but they are highly expensive, time-consuming, and labor-intensive, which often require expertise to carry out [
6].
Different from the aforementioned clinical practices, non-invasive acquisition of cardiac signals such as seismocardiogram (SCG) and ballistocardiogram (BCG) are also potential low-cost alternatives to monitor cardiac mechanical activities. SCG is an accelerometer sensor based modality that can record ultra-low-frequency vibrations of cardiac cycle mechanics along with timings of corresponding cardiac events [
8]. Although BCG and SCG can record cardiac mechanical activities, diagnosis based on them is still in the premature stage to be considered for clinical practices. However, recent literature [
6,
8,
9,
10,
11] show growing confidence on the applicability of SCG in clinical practices over BCG. Moreover, SCG is a noninvasive as well as an inexpensive accelerometer-based cardiac recording method, which can easily be carried out using wearable sensors. Hence, in this proposal, SCG is explored as an additional measure along with the ECG to design a robust cardiac health monitoring system.
In past, it was difficult to gather a huge amount of continuous cardiac data such as ECG and SCG to carry out comprehensive analysis. However, recent advancements in wearable technology have made it possible to collect cardiac data in an easy and affordable manner via wearable sensors for the longer duration. Although most cardiac abnormalities often appear only intermittently, they need to be registered, tracked and analyzed thoroughly as time goes by. This encourages us to design an early warning system, where continuously generated cardiac data are analyzed thoroughly to capture early signs of cardiac abnormal behavior.
The rest of the paper is organized as follows.
Section 2 describes the related works followed by motivation and goals of the paper. The system model is presented in
Section 3. Cardiological data analysis is described in
Section 4. Performance evaluation is carried out in
Section 5. Results and discussions are made in
Section 6. Concluding remarks and future works are discussed in
Section 7.
2. Related Works
In the past, several efforts are made to remotely monitor various physiological parameters of a user using affordable body sensors in a continuous manner. The
LifeGuard [
12] is one of the early efforts that monitors electrocardiogram, heart rate, respiration rate, temperature and blood pressure to alert the users. The
LifeGuard equipped system transfers the health parameters to the base station via Bluetooth-enabled cell phone and raises the buzzer alarm, whenever abnormalities are observed. Similarly, in [
13], ECG based body networks are proposed that raise an alarm if a heart attack is detected. However, many times due to bad signal quality or intense physical activities by a user, systems may raise false alarms under normal circumstances as well. In [
14], a false arrhythmia alarm reduction framework is proposed using machine learning. In [
15], ECG-based automatic recognition of arrhythmias is proposed for the diagnosis of heart diseases. Usually, remote monitoring of a person becomes a challenging job, if a person is engaged in activities such as motor racing, cycling, car racing and on field military service. The
Smart Helmet is specifically designed to address the aforementioned challenges and proposes a bio-sensors equipped embedded helmet to monitor ECG and respiration of a user in an uninterrupted manner [
16].
The another recent effort to remotely monitor cardiovascular and respiratory variables is a mobile healthcare platform
PlaIMoS [
17]. The conceived
PlaIMoS architectural platform typically deals with data collection, communication, analysis and visualization of various healthcare settings. In continuously generated large data sets of ECG patterns, it is highly challenging to visualize cardiac health information aggregated over the period of time. Under such circumstances, easy-to-interpret visualization facility on cardiac health information is highly desirable to distinguish between healthy and abnormal cardiac settings. The
ECG Clock Generator [
18] is an attempt to provide visualization facility of cardiac activities accumulated over the period of time using ECG patterns.
In addition to ECG, in recent times, seismocardiogram (SCG) signals are also used to monitor the vital cardiac health parameters. In [
19], authors have designed seismocardiography using tri-axial accelerometer embedded with electrocardiogram. Recently, Di Rienzo et al. [
20] have designed accelerometer sensors based smart garment to record as well as monitor ECG, SCG, and respiratory variables of an ambulant subject out-of-laboratory setting. SCG also has potential clinical applications such as real-time heart rate monitoring [
21] and left ventricular health monitoring [
22] from wearable SCG measurements.
The raw SCG signals are less informative unless specific peaks are identified and correlated with underlying cardiac activities. A typical SCG cardiac cycle includes a set of nine relevant peaks, also called as feature points such as atrial systole
, closing of mitral valve
, opening of aortic valve
, rapid systolic ejection
, closing of aortic valve
, opening of mitral valve
, rapid diastolic filling
, isovolumic movement
, and isovolumic contraction
. Various hemodynamic parameters can be estimated using SCG feature points. For example, in [
23] systolic time intervals such as pre-ejection time (PEP), left ventricle ejection time (LVET), and electromechanical systole (QS2) are estimated by identifying SCG feature points
,
,
,
, and
. Similarly, in [
8], cardiac time intervals PEP, LVET, systolic time (SYS), and diastolic time (DIA) are estimated by identifying relevant SCG feature points.
In recent years, SCG shows numerous applications in real-time continuous health monitoring. For example, SCG is employed for the detection of respiratory phases in [
9] such as inhale or exhale, which subsequently enable the estimation of systolic time intervals. In [
11], the heart condition is estimated from the learned morphology of seismocardiography. In [
24], SCG based automated detection of atrial fibrillation is demonstrated. In [
25], a comparative study on pulse transit time measurement using seismocardiography, photoplethysmography and acoustic recordings is carried out. In [
10], a cardiac early warning system using wearable ECG and SCG signals is introduced. In [
6], a multichannel SCG is employed and six location-specific feature points are identified in addition to nine SCG feature points described previously.
To the best of our knowledge, the state-of-art recent studies [
6,
8,
23] exclusively focus on separate or combined annotation of ECG and SCG or focus on the annotation based disease-specific cardiac function monitoring [
9,
24]. However, mere annotation of morphologies may not provide the powerful insight unless the comprehensive theoretical analysis is explored. Since ECG and SCG are pseudo-accurate in nature, their exclusive usage for the cardiac health monitoring limits the scope of the improvement. It is known that ECG signals represent only cardiac electrical activities and all heart problems cannot be detected by analyzing these ECG signals. For example, a common heart disorder known as angina may not usually appear in routine ECG. In addition, vulnerable plaque deposition, which is mainly responsible for asymptomatic blockages in heart arteries, is difficult to capture by ECG and requires detailed investigation and tests. The Seismocardiogram (SCG) can be employed to complement the ECG as SCG represents the cardiac mechanical activities and provides in-depth meaningful insights of the mechanical functioning of the heart. However, SCG has also few limitations such as lack of well studied features to determine the heart conditions such as hypertensive heart disease. Hence, feature points based beat-by-beat combined analysis of ECG and SCG is expected to bring additional information to carry out the detail analysis for continuous cardiac health monitoring. It is likely that joint analysis of ECG and SCG can complement to each other to produce more reliable outcomes. The combined analysis of ECG and SCG is beneficial as the outcome of each heartbeat is validated based on the combined knowledge of the cardiac electrical and mechanical activities.
Motivation and Goals
It is increasingly becoming important to record, monitor and investigate the intermittent cardiac anomalies those appear due to the unhealthy lifestyle of people leading to CVD. Usually, cardiac anomalies may occur only intermittently and they go completely unnoticed leading to a sudden death of a cardiac patient. Moreover, ECG signals of cardiac patients also stay entirely normal on regular basis except during those intermittent cardiac anomalies. Under such circumstances, it is highly challenging to foresee potential cardiac health problems in the early stage using traditional in-hospital ECG based diagnosis.
To capture intermittent cardiac anomalies, various cardiac health parameters need to be recorded and monitored for the longer duration of time, which can be accomplished via wearable ECG and SCG sensors. However, raw ECG and SCG signals are less informative unless signals are properly annotated and thoroughly investigated beat by beat. The manual annotation of feature points is a tedious, cumbersome and time-consuming process. In addition, the manual annotation is not a viable approach to design a continuous cardiac health monitoring system. Hence, the computer-assisted automatic annotation of ECG and SCG is highly essential. Additionally, feature points based cardiac health monitoring methods are needed to distinguish the morphology of any cardiac cycle between normal and abnormal. In the proposed study, by use of phrase anomalies, we mean the unusual morphology of ECG and SCG in terms of uncommon recording of an amplitude of peaks, and duration of waves, segments, and intervals. It is to be noted that the mere presence of uncommon recordings in a single cardiac cycle does not mean the cardiac anomalies. For this reason, the set of cardiac cycles is needed to be monitored to help capture developing trends for the cardiac early warnings.
Since cardiac health monitoring is a very broad area, we limit the scope of our investigation only to jointly monitor the normal vs. abnormal behavior of cardiac cycles in simultaneous ECG and SCG recordings. The main objective is to present the viability of joint investigation of ECG and SCG for long-term cardiac health monitoring. In the future, the study can be extended to accommodate the urgency of the abnormality. The goals of our study are listed as follows:
Design a body sensor network to collect the simultaneous ECG and SCG signals;
Design automatic feature point delineation algorithm for annotation of ECG and SCG signals;
Design abnormal morphology detection method for ECG signals;
Design abnormal morphology detection method for SCG signals;
Design combined analysis model for ECG and SCG signals.
3. System Model
The architectural view of the proposed ECG and SCG data collection model is shown in
Figure 1. For SCG data collection, an SCG sensing module is placed at a valvular auscultation site called a Tricuspid valve
, and ECG data collection is carried out placing an ECG sensing module (e.g., electrode) at the right arm as shown in
Figure 1. The high quality disposable electrode H135SG Covidien from Bio-Medical Instruments (Clinton Township, MI, USA) [
26] is used as an ECG sensing module. The accelerometer sensor LIS331DLH from STMicroelectronics (Geneva, Switzerland) [
27] is used as a core component of SCG sensing module. The sensing ability, sensing range and gravitational force sensitivity of SCG sensing module is set to 0.5 Hz to 1 kHz, +2 g to –2 g, and 1 mg, respectively. The band pass filter with frequency 0.5 Hz–50 Hz is applied analogically to get the required ECG and SCG signals at sampling frequency of 1000 Hz. The microcontroller system ADuC7020 from Analog Devices Inc. (Cambridge, MA, USA) [
28] is used for the communication from ECG/SCG sensing modules to Analog-to-Digital convert (ADC) circuit and the PowerLab 16/35 from AD Instruments (Dunedin, New Zealand) [
29] is used as the synchronous data logger, which further amplifies and filters the concurrent signals. The class of nonlinear filters also known as filter bank presented in [
30] is employed for the noise reduction and baseline wander removal with minimal signal distortion. It is reported that the nonlinear filters are expected to perform better than other baseline wander removal methods such as adaptive filters, moving average filters, etc. [
30].
The SCG signals acquired from Tricuspid valve and lead
I ECG signals are used for the combined analysis. The conventional Tricuspid valve site is chosen as the interventricular septum is located beneath the Tricuspid valve, which provides more clear signals. During the entire data collection process, the heart rate is monitored using Finger-clip sensor PAH8001EI-2G [
31] and the respiratory rates are monitored manually to ensure the stability and resting position of the subjects. The data collection is performed in three sessions per subject with at least 5 min of break between the sessions. The described data collection procedure is comprehensively verified and approved by Institutional Review Board (IRB) of the Chang Gung Memorial Hospital (CGMH), Taoyuan, Taiwan with IRB license number 104-6615B.
4. Cardiological Data Analysis
In this section, we introduce various feature points of ECG and SCG along with corresponding values to distinguish between normal and abnormal cardiological data, followed by ECG/SCG based feature points delineations and cardiac health monitoring methods. The ECG and SCG cardiac signals with corresponding cardiac electrical and mechanical activities are explained as shown in
Figure 2 using the normal ECG waveform aligned with the normal SCG waveform. The ventricle depolarization, a cardiac electrical activity that takes place during the
complex can be represented by the corresponding cardiac mechanical activities that take place between atrial systole
to the opening of aortic valve
. Similarly, during the ventricle re-polarization (
T wave) of ECG, the cardiac mechanical activity such as rapid ejection of blood flow
takes place until the closing of aortic valve
. Finally, during the atrial depolarization represented as
P wave, cardiac mechanical activity known as rapid diastolic filling
can be observed.
4.1. Differentiation between Normal and Abnormal ECG Morphology
For a normal and healthy heart, each heartbeat reflects an orderly progression of depolarization in ECG tracing, which is helpful to know various heart functionalities. As shown in
Figure 3a, the normal ECG cycle is comprised of several cardiac electrical activities known as depolarization and re-polarization responsible for heart muscular activities. The entire process of depolarization and re-polarization of a cardiac cycle can be explained as follows. The ECG
P wave represents the atrial depolarization spreads from sinoatrial (SA) node throughout the atria followed by brief period of zero voltage isoelectric representing delay at atrioventricular (AV) node. The
complex represents the short duration of ventricular depolarization followed by ventricular re-polarization represented by
T wave. The
segment between
complex and
T wave represents the brief period of zero voltage isoelectric, when both ventricles are completely depolarized. To define the normal and abnormal behavior of depolarization, ECG trace is first divided into a set of heartbeat cycles. Each heartbeat cycle is again sub-divided into various waves, segments, and interval such as
P wave,
wave,
T wave,
segment,
segment,
interval, and
interval. The subdivision of heart beat cycle into waves, segments, and intervals is performed based on the position and order of the ECG cardiac feature points
P,
Q,
R,
S and
T as shown in
Figure 3a. In addition,
interval duration can also be used as a measure to decide between the normal and abnormal behavior of two consecutive heartbeats.
Table 1 shows the notations used in this paper to represent reference maximum and a minimum value of waves, segments, and intervals observed in a normal ECG tracing. Cardiac anomalies such as myocardial infarction, ischemia, sinus arrhythmia, sinus bradycardia, atrial/ventricular fibrillation, etc., disturb the orderly progression of depolarization, and hence morphology of various waves, segments and intervals changes significantly.
Figure 3b shows various prominent ECG anomalies. For example, in
depression, a line at
segment significantly bends downward below the isoelectric line due to stable/unstable angina problem of a patient. On the other hand, in
elevation, a line at
segment bends significantly upward above the isoelectric line due to non-transmural ischemia. Moreover, bradycardia, characterized by a longer
interval, can cause symptoms such as dizziness, fatigue, chest pain etc. Although ECG is widely used to identify various cardiac anomalies, it may lead to the wrong diagnosis and may falsely indicate the presence of CVD in patients with minor symptoms of negligible risk to CVD [
7]. Hence, it is necessary to correlate the abnormal behaviors observed in ECG trace with the corresponding SCG trace to ensure the reliable cardiac health monitoring. The following section describes the process of feature points delineation for ECG and SCG traces.
4.2. Feature Points Delineation Mechanism
In this subsection, we present two separate mechanisms, one for ECG and another for SCG, to select the corresponding feature points. Five ECG feature points such as
P,
Q,
R,
S, and
T, and nine SCG feature points such as
,
,
,
,
,
,
,
, and
are considered.
Figure 4 shows an example normal/abnormal ECG trace along with five ECG feature points selected by using the proposed ECG feature points delineation algorithm. Similarly,
Figure 5 shows an example SCG trace with nine SCG feature points selected by the proposed SCG feature point delineation mechanism. However, it is expected that the proposed ECG/SCG feature point delineation mechanisms should select the corresponding feature points from normal as well as abnormal ECG/SCG traces, it is to note that cardiac anomalies such as ventricular fibrillation shown in
Figure 3b may result in non-detection of few or all feature points. Here, it is assumed that ECG and SCG traces are collected in the form of vectors of data points represented as
and
with known sampling rate
and mean heart rate
. Each sampled ECG and SCG data point in
and
represents unique amplitude value in terms of millivolts (mV). All of the local maximum and minimum peaks in ECG/SCG cycles are identified using second derivatives with slope value zero to measure the corresponding ECG/SCG amplitude. The measured amplitude represents the sensor value minus the baseline, where the baseline amplitude is computed as zero voltage of ECG/SCG signal. For feature points’ delineation, both methodologies first select the feature points in the first cardiac cycle and continue to select the feature points in subsequent cardiac cycles with a minimum separation distance equivalent to the cardiac cycle length
between the same feature points. The sampling rate
and mean heart rate
are the known input parameters of the experimental data sets used to estimate the cardiac cycle length represented as
. In practice,
is estimated continuously from the
interval duration and updated to continuously estimate the
.
4.2.1. ECG Feature Points Delineation Mechanism
For ECG trace, at first, the feature point
R is selected and then after rest feature points
Q,
R,
S and
T are selected with respect to
R considering the referenced normal values of various waves, intervals, and segments as reported in [
32,
33]. Although normal referenced values may not be the best indicators to detect the very specific heart diseases, they can act as sufficient estimators under quite a bit distorted morphologies for the unobtrusive cardiac health monitoring. Normally, feature point
R exhibits high amplitude, which is easy to detect. The process of feature point
R detection can be described as follows. Firstly, a unique peak
exhibiting the maximum amplitude is identified from all the cardiac cycles under consideration. Later, all the peaks in a cardiac cycle with amplitude greater than
are chosen as candidate
R peaks represented as
. Here,
is a constant between 0 and 1, which must be determined experimentally. In the current study, the
is obtained experimentally, which provides consistently superior performance as shown in
Figure 6a. Finally, the peak exhibiting maximum amplitude among
in a cardiac cycle is designated as feature point
R and is represented by
.
The feature points’ delineation other than R is a non-trivial process, and therefore feature point specific range is formulated in such a way that it maximizes the chances of respective feature point delineation. Four ranges of data points represented as , , , and are formulated with respect to the feature point R to select ECG feature points Q, P, S and T, respectively. The range format for feature points appearing before and after R is formulated as and , respectively. Here, Y represents the set of data points equivalent to the normal wave duration of the corresponding feature point. For example, for feature point Q, . However, to reduce the estimation error of Y, an error margin constant is added as a precautionary measure to get better estimation of feature-point range Y. Here, represents the normal time duration of the wave, and represents the sampling frequency. For each feature point, the corresponding peak is identified and designated as a feature point based on the minima and maxima characteristic. For example, the peak representing the feature point Q in ECG normally appears as minima and therefore the minimum peak from the range is designated as . A similar process is repeated for each feature point except R in each ECG cardiac cycle. The entire process in the form of pseudo-code is given in Algorithm 1.
Besides delineation of ECG feature points, it is also essential to delineate end points of the waves. The ECG end points of waves can be classified into two sets. The set of onset points such as
,
,
, and the set of offset points
,
,
as shown in
Figure 7. In many instances, due to the cardiac abnormalities such as left/right atrial enlargement,
elevation/depression, and
T point raise, the delineation of end points become difficult. In particular, the delineation of end points in case of flatten/inverted
T waves is painful. Hence, a simple end point delineation mechanism is incorporated. At first, using the duration of
P wave,
complex and
T wave observed in normal ECG cycles, the set of onset range
,
,
and the set of offset range
,
,
are derived with respect to the feature points
P,
R and
T, respectively. Finally, the data point with minimum amplitude value is located within the range
,
,
and
and annotated with
,
,
and
, respectively. Similarly,
and
are located as the maximum data points within the range
and
, respectively.
4.2.2. SCG Feature Points Delineation Mechanism
Similar approach of ECG is adopted while selecting various feature points from the SCG trace, where, at first, feature point
is selected and then rest eight feature points
,
,
,
,
,
,
, and
are selected with respect to the position of
. One of the distinguished properties of SCG is the unusual high amplitude of feature point
in SCG morphology as shown in
Figure 5, which makes it easy for the delineation of
in a cardiac cycle. However, this does not hold true in every cardiac cycles and is valid under the specific constraints. For example, in a clear SCG morphology of a healthy subject,
normally exhibits high amplitude in most cycles with few exceptions. However, in distorted SCG morphology of an unhealthy subject, the delineation of
is more complicated and may give poor results. Similar to selecting
R in ECG feature point delineation, maximum amplitude peak
from the set of cardiac cycles is located first. Then, for each individual cardiac cycle, the set of candidate
peaks represented as
are located with amplitude more than
. Here,
is a user-defined constant between 0 and 1, which can be obtained experimentally. In the current study,
is found to give consistently superior performance as shown in
Figure 6b. Finally, from the set
, a unique peak exhibiting maximum amplitude is designated as SCG feature point
represented by
.
SCG morphology is more complex in nature as compared to ECG and therefore it requires significant efforts to correctly locate the feature points other than . Since SCG morphology is not well studied, the normal representative value of an amplitude of various peaks and duration of waves are not yet well defined in the literature. Hence, set of training SCG cardiac cycles from the healthy subjects are used to estimate the distance (i.e., the window size) of various SCG feature points with respect to to formulate feature point specific range. The window size calculated from the training SCG cycles of healthy population has dual advantages: (1) it helps to estimate the normal window size for healthy subjects; and (2) it also helps to identify the significantly varying abnormal SCG features lying outside the normal window size (e.g., outliers), which is observed among non-healthy subjects. For each SCG feature point, a fixed size range consisting of probable data points is formulated at the obtained feature point specific window size , where . Finally, based on the morphological maxima or minima characteristic of each feature point, a unique peak is obtained from the and is designated as a feature point. For example, for SCG feature point , range of probable data points for at distance is formulated with respect to . Subsequently, the peak with maximum amplitude in the range of is designated as SCG feature point . A similar approach is followed for the delineation of all other SCG feature points. Detailed pseudo code for SCG feature points’ delineation is described in Algorithm 2.
Algorithm 1: Delineation of ECG feature points. |
Input: Vector of ECG data points with amplitude, ECG data sampling rate, Heart rate, Set of referenced feature values. Output: ECG feature point set. Notations: : Maximum amplitude data point, : Candidate R points, : Range of points to locate point X, where , : User defined error margin constant, : Constant to detect R points, : ECG feature point set for cardiac cycle. 1 Initialize ; 2 Estimate cardiac cycle length: ; 3 Calculate # of cardiac cycles: ; 4 Select maximum amplitude data point: ; 5 for to do 6 Initialize ; 7 Select candidate R points: , ; 8 ; // Delineation of feature point R.; 9 ; 10 ; 11 , ; // Delineation of feature point Q.; 12 ; 13 ; 14 , ; // Delineation of feature point P.; 15 ; 16 ; 17 , ; // Delineation of feature point S.; 18 ; 19 ; 20 , ; // Delineation of feature point T.; 21 ; 22 ; 23 endfor 24 return ; |
4.3. Data Analysis Methodology
In this subsection, we present mechanisms to differentiate between normal and abnormal morphology of ECG and SCG cardiac cycles.
4.3.1. Abnormal ECG Morphology Detection
The cardiac abnormal morphology in ECG trace either appear in the form of significant variation in amplitude of feature points
P,
Q,
R,
S, and
T, or appear in the form of significant variation in the duration of waves i.e.,
P,
,
T, segments i.e.,
,
, and intervals i.e.,
,
,
, and
. In this paper, the abnormal morphology related to
P wave,
complex, and
T wave is determined with respect to the referenced amplitude and wave durations as suggested in [
32]. It is observed that critical cardiac anomalies in a patient normally last a little longer and affect multiple consecutive heartbeats. Therefore, cardiac abnormalities that arise in an individual heartbeat due to occasional uncommon amplitude and duration are ignored. Instead of monitoring cardiac cycles individually, the group of cardiac cycles is monitored together such as 5-cycles, 10-cycles, 25-cycles, and 50-cycles, etc. The purpose of considering the group of cardiac cycles is to ensure that robust cardiac abnormalities are captured by the system instead of the system being misguided by occasional false abnormalities that arise due to bad signal quality and external noises.
The cardiac abnormal morphology count is maintained, which is increased each time by one, whenever amplitude, wave duration, or both vary substantially with respect to their corresponding reference values. Finally, at the completion of the group of cardiac cycles, a feature point with a maximum number of cardiac abnormalities is considered as the abnormal feature point. The pseudo-code for abnormal morphology detection in ECG trace for the group of k-cycles is shown in Algorithm 3, where k is a user-defined constant. Here, it is assumed that ECG signals are available in the form of vector , where represents set of data points sampled with sampling rate .
Algorithm 2: Delineation of SCG feature points. |
![Sensors 18 00379 i001]() |
Algorithm 3: Abnormal ECG morphology detection |
Input: Vector of ECG data points with amplitude, ECG data sampling rate, Heart rate, Set of referenced feature values, User defined constant to group k-cycles. Output: Detection of abnormal morphologies Notations: : Measured amplitude of X wave, : Measured duration of X wave, : Counter for X wave abnormal morphologies, where , : Counter for interval abnormal morphologies, : Measured duration of interval. 1 Initialize k; 2 Initialize ; 3 Estimate Cardiac cycle Length: ; 4 Calculate # of cardiac cycles: ; 5 for to do 6 j=i; 7 while j≤ do 8 if then 9 ; // P wave abnormal morphology detection 10 end 11 ; 12 if then 13 ; // wave abnormal morphology detection; 14 end 15 if then 16 ; // T wave abnormal morphology detection; 17 end 18 if then 19 ; // interval abnormal morphology detection; 20 end 21 end 22 endfor |
At first, from the past studies [
32,
33], a set of reference values
is formulated. The
consists of reference normal values corresponding to an amplitude of feature points and duration of waves, segments, and intervals. The set
represents the reference normal values within two standard deviations from the mean with percentile range of 2% through 98% of healthy subjects.
Table 1 defines the notations that are used to represent the reference values used in Algorithm 3.
,
,
, and a set
are considered as input to Algorithm 3 to detect the cardiac abnormal cycles in ECG trace. In each cardiac cycle, the current measured value of amplitude, wave, segment, and interval duration of each feature point is compared with the corresponding reference normal values. As shown in Algorithm 3, in cardiac abnormality detection of
P, the current measured value of amplitude
and wave duration
are compared with the reference minimum and maximum value of amplitude (i.e.,
and
) and wave duration (i.e.,
and
), respectively. The
P wave cardiac abnormal morphology counter
is maintained, which is increased by one in each time, either measured amplitude
or wave duration
lies outside the reference normal values. The similar approach is followed to detect the abnormal morphology of
,
and
.
4.3.2. Abnormal SCG Morphology Detection
As far as our knowledge is concerned, there exists no well-studied method to detect the abnormal morphology of SCG cardiac cycles. The existing literature mainly focuses on the delineation of SCG feature points [
6,
8,
23] or focuses on the applicability of SCG in the diagnosis of cardiac health problems [
20,
23,
24,
25]. In this subsection, a process to identify abnormal morphologies in an SCG cardiac cycle is described. Since SCG is employed as an additional measure to complement the performance of ECG based monitoring, the SCG feature points’ delineation and designing of feature-variables are made independent from the ECG to avoid any sorts of performance influence on SCG. Six SCG feature variables are designed to detect morphological abnormalities in SCG cardiac cycles such as
,
,
,
,
, and
. These six SCG feature-variables represent the duration of various cardiac mechanical activities that take place in a cardiac cycle. The notation of SCG feature-variables along with their corresponding cardiac mechanical activities are described in
Table 2. Due to the sensitive nature of SCG accelerometer sensors, the amplitude behavior of various SCG feature points is observed highly fluctuating. Therefore, the only parameter considered important to design SCG feature-variables is the duration of cardiac mechanical activities ignoring the amplitude parameter. In this paper, an SCG cardiac mechanical activity is considered abnormal, whenever significant variation is observed in the duration of corresponding SCG feature-variable with respect to the normal duration.
Figure 8 shows the derivation of SCG feature-variables from the SCG signal.
In contrast to ECG, SCG does not have predefined referenced values to detect the abnormal behavior of feature-variables. Therefore, the reference value of each SCG feature-variable is first estimated from number of initial SCG cardiac cycles from a set of five subjects. It is to be noted that the estimation of SCG reference values is the mean representation of five subjects. For a given subject and heart rate, these reference time intervals are first normalized and then the customized subject-specific set of reference time intervals is obtained. Later, the estimated value of feature-variables is used to measure the significant variation of feature-variables for a given subject. Here, can be defined experimentally and it varies from one data set to another. The smaller the value of , the more premature the estimation is observed and the higher the value of , the more error propagation is observed. Hence, a trade-off needs to be balanced for optimum estimation of . In this paper, the value of is kept to 20 cardiac cycles, which consistently performs better to estimate the feature-variables with reasonable accuracy. The first number of cardiac cycles is considered as an estimation phase, during which a cardiac abnormality detection process is not initiated; rather, duration of various feature-variables is estimated.
In the estimation phase, time series analysis of data is performed to smoothen out short-term fluctuations and to capture the long-term trend of feature-variables. Moreover, behavioral changes such as respirations and body movements are also accommodated in time-series signal analysis by assigning weights to the cardiac cycles in the decreasing order. The weighted moving average duration and weighted moving standard deviation duration are calculated for each individual feature-variable-i, where . At the end of the number of cardiac cycles, the value of and are used as the decision values to detect the cardiac abnormalities in the duration of feature-variables in subsequent cardiac cycles.
In
estimation phase,
is calculated using Equation (
1) and
is calculated using Equations (
2) and (
3) for
number of cardiac cycles, where
i represents the
feature-variable and
k represents the
cardiac cycle. Here,
represents the measured time duration of
feature-variable in
cardiac cycle, where
:
To calculate
, continuous variance
is first calculated using Equation (
2). The method to calculate
is inspired from B. P. Welford’s method [
34], which is an accurate and guaranteed way to generate the non-negative variance under floating point calculations:
From variance
, weighted moving standard deviation duration
is calculated as shown in Equation (
3):
Once the estimation phase is concluded and decision values
and
are obtained,
evaluation phase is initiated. During each cycle in the
evaluation phase, the value of each feature-variable
, where
and
is inspected against the range
of corresponding feature-variable. If value of any feature-variable
is found lying outside the range
, then the corresponding feature-variable is considered as potential outlier and it is marked as abnormal. Since the distribution of value of feature-variables with respect to the mean of corresponding feature-variable can be considered as Gaussian distribution, we used Chauvenet’s criterion [
35] to identify outliers in each cycle of the evaluation phase. In order to identify the outliers, deviation
and tolerance
for each feature variable-
i is calculated using Equations (
4) and (
5), respectively:
Here, indicates the inverse of standard normal cumulative distribution and represents the total number of cardiac cycles. According to the empirical rule of statistics, in Gaussian distribution, 95% of data lies within two standard deviations from the mean and hence as per the thumb rule, one should consider no more than 5% of data as outliers. Hence, the value of tolerance is calculated in such a way that for those feature-variables whose duration value deviates more than two standard deviations from the mean of the corresponding feature variable is considered as outliers.
4.4. Combined Analysis of ECG and SCG Signals
Once the detection of the various types of abnormal morphologies is concluded from ECG and SCG signals. The next step is to ascertain that the cardiac cycle under consideration is indeed abnormal. Since, mere the detection of abnormal morphologies does not mean the abnormal behavior of the cardiac cycle, a Naïve Bayes probabilistic model is designed to know how likely the cardiac cycle under consideration is to be abnormal. For the probabilistic model design, ECG feature set consisting of seven features is used, which is represented as . Here, , and and represents measured wave amplitude, and duration, respectively. Similarly, SCG feature set consisting of six features is used, which is represented as .
The conditional Naïve Bayes probability model is designed to classify each cardiac cycle into class
normal or
abnormal. Let us say that for the
ECG cardiac cycle with feature set
, the conditional probability for a cardiac cycle to be
normal or
abnormal can be defined as shown in Equation (
6):
Here,
and
represents output class
and
, respectively. The
and
represents the probability of
cardiac cycle to be
and
, respectively, for a given ECG feature set
. The
can be rewritten as shown in Equation (
7):
Under the Naïve Bayes conditional independence assumption, each feature
is assumed conditionally independent to every other features
for
. Hence, Equation (
7) can be simplified as follows:
The Naïve Bayes conditional probability model defined in Equation (
8) can be transformed into classifier using the maximum a posteriori decision rule as follows:
Similarly, the Naïve Bayes conditional probability classifier as shown in Equation (
10) can be constructed for SCG using a set of seven SCG features:
Here,
and
is assigned with class label
for some
l based on the maximum a posteriori probability. Using the probabilistic outcome of
and
, each ECG and SCG cardiac cycle is marked as normal (i.e., binary ’0’) or abnormal (i.e., binary ’1’):
As mentioned in earlier sections, abnormalities in ECG cycles do not necessarily indicate the underlying abnormal cardiac activities and therefore we look for the abnormalities in corresponding SCG cardiac cycles as well. For an ECG and corresponding SCG cardiac cycle, if only one of them is detected abnormal (e.g., binary ’1’) by proposed Naïve Bayes probabilistic classifier, then a conclusion is drawn that the cardiac cycle under consideration is more likely to be normal in nature (e.g., binary ’0’) and observed abnormal morphology in either ECG or SCG cardiac cycle is due to the external reasons such as noise. However, if both ECG as well as corresponding SCG cardiac cycles are simultaneously detected with abnormal morphology, then a conclusion is drawn that the cardiac cycle under consideration is indeed abnormal. Equation (
11) acts as an additional measure to generate reliable outcomes in the presence of signal artifacts. For example, if one of the modalities (let us say ECG) outputs the abnormal cardiac cycle (e.g., binary 1) due to external signal artifacts, and the other one (let us say SCG) outputs as entirely normal (e.g., binary 0), then the concerned cardiac cycle is treated as normal ( e.g., 1 ∧ 0 = 0) to reduce the false positives and to avoid the mis-interpretation of results. The outcome of the combined analysis of ECG and SCG cardiac cycle can be calculated as shown in Equation (
11). Here,
and
represent individual outcomes of
cardiac cycle of ECG and SCG, respectively. The
represents the combined outcome. In addition,
Table 3 shows all possibilities that arise in Equation (
11). A new parameter called Cardiac Abnormality Index (CAI) is defined to represent the intensity of cardiac abnormal behavior. The
is defined as the number of abnormal cardiac cycles out of the total number of cardiac cycles in a group as shown in Equation (
12). The higher the value of
, the higher the risk of CVD and vice versa. If
increases gradually over the period of time and crosses the predefined threshold value
, an alert warning may be issued to the user to consult the cardiologist. It is to note that the value of
can be determined in consultation with the cardiologists.
6. Results and Discussion
We have experimented with ECG/SCG traces of three normal and two abnormal subjects. The experiments are performed considering cardiac cycles in the group of 5-cycles, 10-cycles, 25-cycles and 50-cycles to verify the effectiveness of
. Based on our observations,
performs effectively well for the small sized group of cycles such as 5-cycles and 10-cycles for patients having chronic heart disease. In patients with chronic heart disease, abnormal cardiac cycles appear more frequently and the distance between two abnormal cardiac cycles is expected to be less as shown in
Figure 9b,c. However, the effectiveness of
decreases drastically for small sized groups of cycles, in patients with occasional heart problems. In the case of occasional heart problems, it is less likely that an abnormal cardiac cycle appears frequently and together. Hence, most small sized groups of cardiac cycles output lower
leading to conclude that patient is normal, which is misleading. On the other hand, increasing the group size to 100-cycles and 200-cycles keeps the ratio of the number of abnormal cycles to the total number of cycles in a group very small, which also indicates the normal functioning of heart leading to the wrong diagnosis. Hence, in order to balance the effectiveness of
between the small sized groups and large sized groups, we experimented with medium sized groups such as 25-cycles and 50-cycles.
The sample output of combined cardiac anomaly detection for ECG and SCG considering the group of five cycles is shown in
Figure 9. Three separate cases i.e., two for the normal subjects and one for an abnormal subject are shown for the group size of 5-cycles each. The first case for the normal subject is shown in
Figure 9a, which shows the normal morphology detection in ECG as well as SCG cycles (
) and therefore we may conclude that there are no cardiac abnormalities in patient’s heart functioning, and he/she can be considered as normal. The second case for the normal subject is shown in
Figure 9b, where two out of five ECG cardiac cycles are detected abnormal, one due to
R amplitude abnormality and another due to
P amplitude abnormality with
. However, the corresponding SCG cardiac cycles show no abnormalities with
and therefore we may conclude that detection of abnormalities in ECG might be due to external noises. The third case for an abnormal subject is shown in
Figure 9c, which shows the abnormal functioning of heart since all of the five ECG, as well as SCG cardiac cycles, have been recorded abnormal with
. Here, for ECG, the amplitude of
R shows the significantly abnormal behavior in each cardiac cycle leading to the abnormality detection. On the other hand, corresponding SCG cardiac cycles also recorded as abnormal. Since cardiac cycles of both traces have simultaneously resulted in abnormal behavior, we may conclude that the patient has serious heart malfunctioning and he/she should immediately take the advice of cardiac experts.
The same experiments are repeated with the same set of input data considering the group sizes of 10-cycles, 25-cycles, 50-cycles, and 200-cycles. The experimental results for the group size of 10-cycles for normal subjects are shown in
Figure 10, which shows the lower
values 0.1 and 0.2. In addition, we have also recorded the Timestamp instances of various feature points of ECG as well as the SCG traces.
Table 11 shows the example Timestamp instances for the ECG feature points
P,
Q,
R,
S and
T, whereas
Table 12 shows the example Timestamp instances for the SCG feature points
,
,
,
,
,
,
,
and
across two cardiac cycles for reference purposes only. From the two example cardiac cycles shown in
Table 11, it is observed that the difference between successive Timestamp instances of ECG feature points are more consistent and regular for the normal subjects
,
and
than those observed for the abnormal subjects
and
. Similarly,
Table 12 reports the consistent behavior of SCG feature points for the normal subjects
,
and
than the abnormal subjects
and
. The aforementioned observations do not specify the general conclusions, but state with respect to the randomly selected two examples of cardiac cycles. It is to be noted that one or more SCG feature points may not be repeated between two heartbeats due to external noise. For example, for subject-1
in
Table 5, the reported number of cardiac cycles (defined as OCCs) is 825 and therefore the number of times
and
should appear is 825 times each. However, as reported in
Table 7, the total number of manual feature points
and
those are retrieved for subject-1 is 791 and 801 times, respectively. This confirms the non-consistent repetition of
and
. The similar trend is observed for the rest of the subjects except abnormal subject 4
due to its highly distorted morphology. Due to the aforementioned limitations, the current study has considered analyzing the group of cardiac cycles together to capture the abnormal behavior of heart instead of a single cardiac cycle at a time.
The performance of any cardiac health monitoring system depends on its ability to accurately distinguish between normal and abnormal cardiac cycles. Hence, the performance of the proposed combined analysis of ECG and SCG is compared with the standalone use of ECG and SCG. All of the 12,000 cardiac cycles are chosen and equally divided into 12 sets of 1000 cardiac cycles each. For each set, the ability of the system to detect the number of normal and abnormal cardiac cycles is measured. At first, the baseline reference performance output is generated by recruiting expert cardiologist, who are tasked to manually distinguish the cardiac cycles into normal and abnormal for each set. Later, the performance of systems such as ECG only, SCG only, and ECG and SCG are compared with the baseline output.
Figure 11 shows the output of each of the aforementioned systems with respect to the baseline performance.
Figure 11a,b shows the comparison with respect to the ability of the proposed mechanisms to correctly detect the percentage of
normal and
abnormal cardiac cycles, respectively. As shown in
Figure 11a,b, the combined analysis of ECG and SCG consistently outperforms the
ECG only and
SCG only mechanism. In addition, the performance of
ECG only and
SCG only is nearly identical to each other, with
ECG only marginally performs better. The potential rational behind marginal improvement of ECG over SCG is due to the preprocessing of ECG signals for baseline wandering removal and noise reduction. The main reason behind the better performance of a combined analysis of
ECG and SCG lies in its ability to validate and re-validate the
normal and
abnormal behavior of the cardiac cycle using cardiac electrical and mechanical information altogether.
Figure 11c presents the performance of the systems with respect to the overall accuracy. As shown in
Figure 11c, the combined analysis of
ECG and SCG gives better overall accuracy as compared to the
ECG only and
SCG only.
7. Conclusions
In this paper, a combined cardiac health monitoring approach is presented using concurrent Electrocardiogram (ECG) and Seismocardiogram (SCG) signals. Each ECG and SCG cardiac cycle is tracked, analyzed and jointly investigated to distinguish the Normal and Abnormal behavior. A separate delineation mechanism is designed for the automatic delineation of specific feature points of ECG and SCG from the corresponding signal morphology followed by feature points based abnormal morphology detection mechanisms. For the reliable beat by beat analysis, individually detected abnormal morphologies in ECG and SCG are jointly investigated by designing a novel Naïve Bayes conditional probability model. The effectiveness of the proposed feature point delineation mechanisms is extensively verified against the performance metrics such as , , and by acquiring 12,000 concurrent ECG and SCG cardiac cycles from 20 real subjects approved by Institutional Review Board (IRB) of Chang Gung Memorial Hospital, Taiwan. Experimental results show that combined analysis of ECG and SCG can mitigate the disadvantages of both modalities up to some extent and can improve the cardiac health monitoring accuracy subject to the accurate feature points delineation. The main objective of the current study was to show that there is a scope for further improvement in continuous cardiac health monitoring, which can be accomplished by jointly analyzing the mechanical and electrical aspects of the cardiac activities obtained from SCG and ECG, respectively. The study is a step forward and needs to be continued for realizing the ultimate goal. The results presented in this study are encouraging and provide a solid foundation to explore the benefits of the combined investigation of ECG and SCG for future works. However, there are few limitations of this preliminary study. The current study only covers the basic ECG and SCG features. It could be possible that the better set of ECG and SCG features may further improve the results. In addition, the naive approach for the ECG delineation is employed, which can be further improved by using the proprietary algorithms designed to specifically delineate the ECG signal. The current combined analysis of one-lead ECG signals with concurrent SCG can be further extended and improved in the future work by engaging the two-lead ECG signals with concurrent SCG, as the two-lead ECG may provide more redundant information. In addition, the research can also be extended in the future by considering the non-detection of certain existing features as new features and incorporating them in machine learning algorithms to improve the accuracy of cardiac monitoring. The future work will include more ambulatory monitoring to test the algorithms in real conditions, i.e., with more noise and range of activities. It is to be noted that the results are obtained under a specific in-laboratory environment for subjects with specified demographic condition, and, therefore, results should be interpreted accordingly.