Wearable ECG Device and Machine Learning for Heart Monitoring (2024)

2.1. Hardware Development of Wearable ECG Devices

The ECG data necessary for the study were obtained from a wearable ECG device built by the authors, as shown in Figure 2.

A wearable ECG device is one of the key elements of a cardiology diagnostic system. It facilitates the analog-to-digital conversion and transmission of ECG signals through radio channels to a smartphone.

Figure 3 illustrates the detailed functional flow chart of the proposed ECG device.

A wearable ECG device consists of several main components, among which the most important are microcontrollers.

ADS1298, connected to the microcontroller STM32L151xD through a serial data transmission channel, was used to collect ECG data.

Ensuring an independent power supply for ADS1298 to minimize interference is a main stage of the process. The non-volatile, high-speed memory block, MRAM, was employed to record operative data, while for long-term data storage, an external microCD Flash with a 32 GB capacity was used during data collection. In addition, the ECG device was equipped with a MEMS accelerometer to determine the patient’s heart-attack-induced falls and immobility. Power and battery charge are supplied by an external source of power using a charging device, which tracks the actual battery charge. There is also a power button controller and a residual battery charge detector. A GPRS communication module was used to transfer data to both the medical information system and the ambulance service, while a Bluetooth module was employed to send through a mobile phone. Moreover, a GNSS module was used to identify a patient’s location. The visual control system includes three-color LEDs that display various operating modes and single-color LEDs for signaling about the connection to the Internet and data exchange on the server.

Figure 3 also shows the order of attachment of the ECG electrodes to the patient by cable color, with the specified electrode locations on the torso:

Electrode 1—Red cable—Second intercostal space at the right sternum;

Electrode 2—Yellow cable—Fourth intercostal space at the right sternum;

Electrode 3—Green cable—Fourth intercostal space at the left sternum;

Electrode 4—Black cable—Fifth intercostal space at the right sternum.

2.2. Software Complex

The following tasks were carried out during the development of software components:

A Java Script framework with an open-source Vue.js was chosen to create the user interface for the client module of the cardiac diagnostics system. Vue.js has convenient integration with other Java Script libraries and can be employed to build a single-page application. This framework enables a user interface to be created based on the Model-View-ViewModel architecture template. Vue.js specializes in demonstrating the level of easy integration with other projects and libraries. Websites built with Vue.js can be efficiently uploaded due to their small size (17 Kb).

The free framework Django, written in Python, was chosen for the software part of the service (the backend). Django follows the design pattern MVC (Model-View-Controller) and enables web applications to be created from separate plug-in applications. This framework adheres to the principle of DRY (Do not repeat yourself), which prevents codes from being repeated. One of the features of Django is a URL handler configuration that uses regular expressions. ORM (Object-Relational Mapping) can be used to describe data models in Python and automatically create data base schemes that function in the Django data base. ECG data were retrieved from a server using a REST API (Representational State Transfer) architecture style. All retrieved data was stored in the object-relational database management system PostgreSQL, a flexible and reliable database management system that supports user objects and complex data structures.

Each user has their own unique data, which is sent to the client module in Vue.js with the REST API. The web service collects data every second and visualizes it as a graph that moves in sync with the user’s heart rate. After entering a username and password, the website sends a request to the server, and the requested data appears on the screen.

A JSON Web Token, a token for authentication and user authorization, was used in Python to check user access rights to certain resources. JWT enables the creation of token access to JSON, which confirms the certain privileges of a user. The tokens can be employed to transfer identification data between the client and server.

2.3. ECG Signal Filter

Three-stage noise removal techniques were implemented in the proposed ECG cardio monitoring system, which includes the following flowing processes:

At the first stage, filtering is applied for the analog-to-digital conversion of module 1298:

At the second stage, the Median filter was applied to the hardware on the controller. This method includes defining the filter window, implementing median filtering for each signal count in the window, processing boundary conditions, and repeating the process to improve the filter efficiency. The median filter is effective in removing impulsive and other noises, keeping the main features of signals.

At the third stage, as a software application using digital signal processing based on filtering algorithms, a bandpass filter is implemented. The bandpass filter is adjusted to a certain frequency range containing cardiac activity and blocks frequencies beyond the range. It eliminates not only frequencies that are equal to the electrical activity of the heart, but also blocks interference from other sources, such as electromagnetic interference or network noise, as well as low-frequency artifacts, such as body movement or electric motor activity.

The function H(f) decreases as frequency increases f, reaching half of its maximum value at the cut-off frequency fc. The parameter n identifies the steepness of the frequency response of the filter; the more n, the sharper the transition from passing to suppressing the signal around the cut-off frequency.

2.4. ECG Signal Clustering

In the process of analyzing electrocardiogram (ECG) data, situations may arise where the signals have inverted polarity due to technical aspects of the recording equipment or errors during the recording process. Inverted polarity can lead to an erroneous interpretation of cardiac activity, which is a critical issue in clinical diagnostics.

To address this issue, a K-means algorithm was employed to classify ECG signals based on the similarity of their amplitude characteristics and waveforms. The clustering process began with determining the number of clusters and selecting initial points for the centroids. Each data point was then assigned to the cluster with the closest centroid based on the Euclidean distance. The centroids were recalculated until convergence was achieved, indicating the stabilization of the clusters.

2.5. PQRST Wave Detection

The precise detection of ECG signal features, such as the R peak, QRS complex, and other components of waveforms, is crucial for detecting cardiac abnormalities. Traditional ECG signal analysis means manual interpretation, which takes plenty of time and is prone to human error. Thus, there is a growing interest in designing automated methods for improving the preciseness and effectiveness of ECG signal analysis.

The process of detecting peaks P, Q, R, S, and T in ECG signals involves initially identifying the R peak. This is followed by detecting other significant peaks by analyzing changes in amplitude and temporary intervals between these peaks and other points on the ECG signal.

The Q point is identified in the interval before the R peak. A time interval in the form of a Q interval is used, which is typically approximately 0.04 s (or 14.4 samples at a sampling rate of 360 Hz). Thus, to search for the Q point, the algorithm analyzes the signal in the interval [R−Qinterval,R], where R is the index of the peak R.

The S point is identified after the R peak. Sinterval, with 0.08 s (or 28.8 counts at the same frequency sampling), is used as the S point. The S point search interval is [R, R+Sinterval].

The T point represents elevation, following the S wave, and illustrates a period of repolarization in the ventricles of the heart. The interval after the S point is used for searching for this. The T interval usually starts immediately after the S point and can last until 0.2 s or 72 counts. Detecting T points involves searching for the maximum value of amplitude within this interval.

The P wave illustrates a period of atrial depolarization and precedes the QRS complex. The interval before the Q point is used to detect it. The P interval can vary but is often selected from 0.12 to 0.20 s before the Q point. Searching for the P point involves identifying the maximum amplitude within the interval.

This process introduces the basic methods for detecting important points in the ECG signal. It is worth mentioning that detecting preciseness depends on signal quality, the presence of noise and artifacts, and certain parameters of the algorithm in signal processing.

2.6. Anomaly Detection in ECG Signals

In this study, we adopted the Isolation Forest, a method for detecting abnormalities, to collect data consisting of key point labels in an ECG (waves Q, R, S, T, and P). We aimed to detect potential abnormalities in cardiac rhythms, which could indicate the presence of abnormalities or other irregular cardiac conditions. Isolation Forest is a machine learning algorithm based on ensemble decision trees that is used to detect abnormalities by isolating individual data points.

The algorithm was set using 100 estimators (‘n_estimators = 100’), where parameter ‘contamination’ was set as ‘auto’ for automatically detecting the ratio of abnormalities in the dataset. The arbitrary selection of subsamples and divisions within trees was controlled with the parameter ‘random_state = 42’, ensuring the consistency of the results.

Once the model was trained on an ECG dataset, the abnormal condition of cardiac rhythm was predicted. Algorithm predictions categorized the dataset into normal (designated as “1”) and anomalous means (designated as “−1”). These results were integrated back into the initial dataset, enhancing it with abnormal labels for each ECG record.

2.7. Predicting Heart Disease

The research aims to analyze relationships between various factors (age, sex, chest pain, blood pressure, cholesterol, alcohol, diabetes, ECG change, smoking) and a patient’s health condition.

To carry out the study, 132 participants were chosen from the cardiology center of the Central Clinical Hospital of Almaty, Kazakhstan, and their data were recorded. The study included both men (98 participants) and women (34 participants). Their ages varied from 28 to 68. The participants’ average blood pressure was 129.93 mmHg, with a minimum value of 94 mmHg and a maximum value of 200 mmHg; their cholesterol level was within the minimum level at 126 mg/dL and the maximum at 341 mg/dL. In total, 43.18% of participants reported using alcohol, 31.06% of participants had diabetes, and 55.30% of them were smokers, while changes in ECG data were observed in 37.12% of participants. Data were collected using conventional cardiac methods of assessment, including measuring blood pressure and cholesterol level, as well as conducting an ECG study using a wearable system built by the authors. This study aimed to analyze the relationship between lifestyle, physiological parameters, and the risk of developing cardiac diseases. All participants were requested to abstain from alcohol, coffee, or tea for 12 h prior to the experiment. Before the experiment, each subject underwent a checkup, where we measured their blood pressure and consulted their medical history. Furthermore, during the experiment, we employed a wearable ECG device for recording the ECG signals. The results were verified by an experienced cardiologist, and he made notes to collect data for further research.

The participants’ data were divided into 10 parameters. Information about dataset attributes is illustrated in Table 1.

The aim of this study was to assess 10 parameters: the presence of cardiovascular disease (CVD) in patients and the first 9 parameters shown in Table 1.

During the study, various machine learning methods were applied, including logical regression, decision trees, random trees, support vector machines (SVM), XGBoost, and convolutional neuron networks (CNN), followed by a comparative analysis and investigation of the efficiency of each method.

3.1. ECG Signal Filter

As mentioned in Section 2.3, the process of filtering ECG signals with the aim of suppressing noise was implemented concurrently in three stages; two of them were executed at the hardware level. Particularly, the analog-to-digital conversion was carried out by applying an RC filter, high- and low-frequency filters, digital thinning filters, and network filters. In the microcontroller, the median filter was involved in removing impulse interference. As a result, the ECG signal was obtained, as illustrated in Figure 4. Artifacts were detected that hindered subsequent signal analysis.

Signal filtering is crucial for preparing ECG signals for further analysis, as it improves visualization, which is important for observing components such as the R peak, and facilitates automatic detection. A bandpass filter was used to improve the precision and reliability of diagnostic outcomes.

The following parameters were set for filtering the ECG signals:

A low-cut range of 0.5 Hz, which removes low-frequency oscillations that are not related to cardiac activity.

A high cut range of 100.0 Hz, which facilitates the suppression of high-frequency noises while maintaining the important components of ECG signals.

The sampling frequency (fs) of the ECG signal is 360 Hz, a typical value for medical ECG data, ensuring the overall quality of signal recovery.

Segments after ECG signal filtering are illustrated in Figure 5, which shows the effectiveness of the implemented ECG data processing.

Pre-filtering the signals enabled the removal of some artifacts and highlighted the main artifacts for analyzing signal functions.

3.2. Clustering ECG Signals

Clustering enables signal points to be classified by amplitude, facilitating visualization and data interpretation. This algorithm requires the number of clusters to be specified (in our case, they are 5). It scales well to large numbers of samples and is used in a wide range of ECG signal processing applications.

For the analysis of ECG data, we employed the K-means method with five clusters. This choice was based on a preliminary analysis of signal characteristics and experiments conducted with varying numbers of clusters. We found that using five clusters allows for the most adequate division of data, highlighting significant signal features. Such an approach ensures an accurate representation of key ECG components and enhances subsequent data interpretation.

Figure 6 shows a graph containing data from the electrocardiogram and the processing results from the clustering algorithm. The X axis is the duration of cardiac activity, while the Y axis represents the amplitude of the electrical signal.

The amplitudes were inverted to improve visualization and analyze individual clusters of signals (Figure 7). This step enabled the polarity of signals to be corrected and improved the detection of key points in the ECG.

As seen in Table 2, there are 5 clusters, and in Figure 6 and Figure 7, they are depicted in different colors.

The algorithm grouped the data points of ECG signals detected into features of the presented signals based on their similarities. Clusterization highlights the differences in heart rate characteristics identified in ECG signals.

As a result of clustering using the K-means algorithm, an actual visualized ECG signal was obtained, as illustrated in Figure 8.

The data obtained through the filtering and clustering processes were processed to determine the PQRST wave in the ECG signal.

3.3. PQRST Wave Detection

The process of identifying points P, Q, R, S, and T on an ECG involves detecting R peaks and identifying other points using amplitude analysis and temporary intervals within them. The Q point was identified prior to the R peak by applying the temporary interval, Qinterval. The S point was identified after the R-peak by applying the interval, Sinterval. The T point was identified after the S waves. The P point was detected prior to the QRS complex.

The results of detecting PQRST waves can be seen in Figure 9.

As seen in Figure 9, the peaks of points P, Q, R, S, and T are depicted in certain colors: yellow (P point); green (Q point); and red (R point). Peaks designated by blue are the S point, and purple peaks are the T point.

The forms of PQRST waves are almost the same as described in Section 2.5, meaning that the wearable system was effective enough to detect waves in an ECG signal.

3.4. Anomaly Detection in ECG Signals

The results obtained from this study reveal the importance of a complex approach to analyzing ECG signals, including methods of digital signal processing and clustering. In Table 3, we kept the identified labels with data for further analysis, utilizing machine learning methods. Table 3 presents a sample of the labeled data, including indices of various key points (Q, R, S, T, and P) in the ECG data.

These data are used to analyze the characteristics of the ECG signal and classify different phases of the cardiac cycle.

The analysis of the results demonstrated that a certain number of records were classified as “anomalous”, while others were regarded as “normal”. A qualitative assessment revealed a distribution between anomalous and normal states (Table 4) that indicates the potential presence of irregular patterns in the dataset retrieved from the ECG signals. Table 4 contains readings of the detected anomalous and normal data.

These data help assess the effectiveness of algorithms in classifying normal and anomalous ECG records, which is important for improving diagnostic accuracy and predicting patients’ health conditions.

We applied the Isolation Forest method to classify 56 cardiac rhythm records without prior anomaly labels. As a result, the model correctly classified 20 out of 21 actual anomalous records, which accounts for 97% of the total anomalous records. This yields the following metrics: accuracy of 97%, true positive rate of 95%, and false positive rate of 5%. The remaining 36 records were classified as normal (Figure 10).

To configure the Isolation Forest model, we selected parameters that balanced accuracy, performance, and reproducibility of results. The number of trees in the ensemble was set at a level that ensured stable predictions. The proportion of data considered anomalous was adjusted so that the model could automatically assess this fraction based on the data. Additionally, an initial value was set for the random number generator to ensure the reproducibility of results in the future. These parameters were chosen based on common practices and experiments to achieve optimal model performance in ECG data analysis.

Further analysis demonstrated that the average values of Q, R, S, T, and P waves in the index of anomalous records differ from the average values of normal records. In particular, anomalous records manifested the early presence of all waves in cardiac rhythm compared with normal records. It can indicate certain changes in cardiac rhythm associated with abnormalities.

In Figure 11, the y-axis represents the amplitude of the ECG signal at each time index. This index facilitates the visualization of amplitude changes in the signal over time.

Rectangular graphs may appear similar for different points, such as R peaks and points P, Q, S, and T, due to the normalization process applied before clustering. During normalization, each segment of the ECG signal is scaled to a comparable range, enabling sequential clustering based on the waveform rather than the signal amplitude.

Although the raw amplitudes of R peaks are typically much higher than those of points P, Q, S, and T, normalization is necessary to prevent amplitude bias during clustering. This approach allows segments with similar shapes to be grouped, which is crucial for the accurate identification and analysis of various components of the ECG signals.

These results are important for identifying potential abnormalities in ECG data, which can offer a justification for further research into cardiovascular diseases. Additional analysis, including comparison with medical data and initial diagnoses, is required to confirm the clinical importance of detected abnormalities.

3.5. Predicting Heart Disease

Our data set includes information about the group of patients, explained thoroughly about their demographic characteristics, lifestyles, and indicators that show the health status of the subject. This creates a basis for studying the relationship between “decommunization” and lifestyle choice and the consequences of this for health. We used unprocessed data from 132 patients, which we collected as a result of using our ECG device. The data fraction is given in Table 5.

Prior to conducting a more detailed analysis of our data, a statistical analysis was conducted to evaluate the direction of parameter interrelations. Correlations between each pair of features were analyzed and identified.

• Age—patient’s age.

• Sex—gender of the patient.

• Chest pain—is there pain in the patient’s heart or not?

• Blood pressure—the patient’s blood pressure during the examination.

• Cholesterol—cholesterol levels.

• Alcohol—does the patient drink alcohol or not?

• Diabetes—does the patient have diabetes?

• ECG Change—a change in the patient’s ECG.

• Smoking—cigarette use?

• Condition—sick or not?

The coefficient of correlation between variables is illustrated in Figure 12. Each table cell represents a correlation between two parameters in our dataset. Correlational analysis shows the significance of the relationship between some features and the health conditions of patients.

Figure 12 shows the correlation coefficient between each pair of features. However, as noted, correlation coefficients below 0.6 may not provide convincing evidence of significant relationships.

For a more comprehensive analysis, we included boxplots, which visualize the distribution of data and help identify potential dependencies between variables and states. These graphs clearly demonstrate how various factors correlate with patient conditions, enhancing our understanding of the data and confirming the findings of the correlation analysis.

Figure 13 displays boxplots for the following variables: age, gender, chest pain, blood pressure, cholesterol, alcohol consumption, diabetes, ECG changes, and smoking. These graphs visually assess the distribution of each variable’s values depending on the condition (0 or 1). For instance, it can be observed that age, cholesterol level, and blood pressure exhibit a wider range of values among patients with condition 1. Similarly, ECG changes and diabetes are more prevalent among patients with condition 1.

The use of boxplots allows for a clearer visualization of the data and complements the results of the correlation and regression analysis, confirming the significance of various factors for the patients’ condition.

The researchers applied statistical and machine learning methods, including logical regression, decision trees, random forests, support vector machines (SVM), and convolutional neural networks (CNN), for analysis. The efficiency of the model was assessed using an accuracy label and the percentage of accurate responses from the model in relation to all predictions, enabling the overall effectiveness of the model to be evaluated.

Multiple random divisions of preprocessed data into training and evaluation subsets were applied to evaluate the machine learning method. Furthermore, features were scaled to improve the training process and accuracy of the model. Scaling gives each feature a standardized form with an average value of 0 and a standard deviation of 1.

The graph shows the accuracy and F1 score values for each model.

The graph was illustrated to compare the accuracy and F1 score values for each model and analyze the retrieved data (Figure 14).

The results of training and assessing the performance of convolutional neuron networks (CNN) for binary classification tasks demonstrate the dynamics of model training and its ability to be classified based on the given data.

Applying various models of machine learning indicated their various degrees of efficiency in predicting patients’ health conditions. Convolutional neuron networks demonstrated remarkable performance, reaching an accuracy of 0.926, which indicates their potential for processing complex data patterns.

This research highlights the importance of a comprehensive analysis of medical data in predicting health conditions. The efficacy of integrating traditional machine learning with deep learning techniques opens new perspectives on designing prediction models in medicine.