*侵删
*限于博主英语水平，若翻译不当之处恳请批评指正~3Q
Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals
ABSTRACT
The electrocardiogram (ECG) is a useful diagnostic tool to diagnose various cardiovascular diseases (CVDs) such as myocardial infarction (MI).The ECG records the heart’s electrical activity and these signals are able to reflect the abnormal activity of the heart.However,it is challenging to visually interpret the ECG signals due to its small amplitude and duration.Therefore, we propose a novel approach to automatically detect the MI using ECG signals.In this study, we implemented a convolutional neural network (CNN) algorithm for the automated detection of a normal and MI ECG beats (with noise and without noise).We achieved an average accuracy of 93.53% and 95.22% using ECG beats with noise and without noise removal respectively.Further, no feature extraction or selection is performed in this work. Hence, our proposed algorithm can accurately detect the unknown ECG signals even with noise. So, this system can be introduced in clinical settings to aid the clinicians in the diagnosis of MI.
心电图是诊断各种心血管疾病（如心肌梗塞）的一种有用的诊断工具。心电图记录了心电活动，这些信号可以反映出心脏的异常。然而，由于心电图的幅值和周期太小，因此对其进行可视化解读极具挑战。因此我们提出了一种利用心电信号的自动检测心肌梗死的新方法。在这项研究中，我们实现了一个未去噪和去噪下自动检测正常和心肌梗塞心电图的卷积神经网络算法。我们在未去噪和去噪的心电图下分别达到了93.53%和95.22%的平均准确率。此外，在这项工作中我们没有进行特征提取或选择。因此，我们提出的算法可以准确地检测即使带有噪声的未知心电信号。故这个系统可以被用于临床来协助医生进行心肌梗塞的诊断。
1 Introduction
Myocardial infarction (MI) is caused when the blood flow to a segment of the myocardium is disrupted.Coronary arteries are the arteries that supply oxygen-rich blood to the heart muscle. However, if there is a blockage of the coronary artery due to the buildup of plaques, it reduces the blood flow to the heart muscle. That segment of the heart muscle will start to die if blood flow is not restored in time. Fig.1 illustrates the myocardial infarction due to the blockage of a coronary artery. This artery gets blocked with blood clots also known as a thrombus. These blood clots are formed due to the plaque build-up in the artery. The complete blockage of blood flow results in a heart attack as a part of the heart muscle is damaged.
当流向心肌某部分的血流中断时就导致了心肌梗塞。冠状动脉是给心肌供应富氧血液的动脉。然而，如果冠状动脉由于斑块的堆积而阻塞，流向心肌的血流量就会减少。如果血流不能及时恢复，那部分心肌就会开始死亡。图1所示为冠状动脉阻塞引起的心肌梗死。这条动脉被血栓阻塞，也被称为形成血栓。这些血栓是由于动脉中的斑块堆积形成的。血流完全阻塞会导致心脏病发作，因为心肌部分受损。
Furthermore, MI is also often referred to as the silent heart attack. It is because patients are not aware that they are suffering from MI until a heart attack occurs. According to the American Health Association, it is estimated that 750,000 Americans have a heart attack every year. Out of these 750,000 Americans, 210,000 of them have a recurrent heart attack. Hence,approximately 72% of the heart attacks are silent. In other words, 72% of the patients’ heart muscles are damaged but they are not aware of it. As a result, the mortality rate of MI is very high.
再者，心肌梗塞也常被称为无症状心脏病发作。这是因为在心脏病发生之前，病人通常不能意识到自己患有心肌梗塞。据美国健康协会估计，每年约有75万美国人心脏病发作。在这75万美国人中，有21万人是心脏病复发。因此，大约72%的心脏病发作是毫无征兆的。换句话说，72%的病人的心肌受损，但是他们从未意识到这一点。这样的结果是心肌梗死的死亡率很高。
Therefore, an early diagnosis of MI will help patients to get timely treatment, and hence decreasing the prevalence of mortality. The death of the heart muscles is irreversible hence, it is essential to get diagnosed early. The early diagnosis of MI can be conducted with an electrocardiogram (ECG). The ECG is the noninvasive economical primary tool which can be used to diagnose the cardiac abnormalities. Fig. 2 shows the samples of normal and MI ECG signals with and without the removal of noise.
因此，对心肌梗塞的早期诊断有助于帮助患者得到及时的治疗，从而降低高死亡率。心肌梗死是不可逆的，因此必须及早诊断。
心肌梗塞的早期诊断可以通过心电图来进行。心电图是用于诊断心脏异常的无创且经济的主要工具。图2展示了显示正常和心肌梗死心电图信号的样本(去噪和未去噪)。
However, the ECG signals are having a very small amplitude (mV) and small duration (sec). Hence, the interpretation of these long duration of signals may lead to inter and intra-observer variabilities (?). Moreover, it is time-consuming and strenuous to analyze the ECG signals.
然而，心电图信号的幅度(mV)很小和持续时间(sec)也很短。因此，对较长的信号的的分析可能导致检查者间和检查者自身差异（?）。同时，分析心电图信号很耗时且费力。
The limitation of manual inspection of ECG signals can be overcome by using computer-aided diagnosis system. A computer-aided diagnosis (CAD) system is preferred due to its fast, objective, and reliable analysis. Many works have been conducted on the development of CAD for MI.
利用计算机辅助诊断系统可以克服人工检测心电信号的局限性。计算机辅助诊断系统由于其快速、客观和可靠的分析而成为首选。在心肌梗塞的计算机辅助诊断系统上已经有许多工作。
The studies presented in Table 6 have denoised their ECG signals before performing any feature extraction. Nevertheless, denoising is not required in our proposed algorithm. Our algorithm can detect MI ECG signal without filtering any noise present in the ECG signal. Various features extraction techniques have been proposed to automatically detect MI using ECG signals. However, the process of choosing a set of optimal features to classify normal and MI ECG signals is very difficult.
在进行任何特征提取之前，表6中的研究已经对它们的心电信号进行了去噪。然而，我们提出的算法不需要去噪。我们的算法可以在不滤除心电信号中任何噪声的情况下检测出心肌梗塞的心电信号。为了利用心电信号自动检测心肌梗死，已经提出了多种特征提取技术。然而，选择一组最优特征来对正常和心肌梗塞的心电信号进行分类的过程是非常困难的。
Therefore, deep learning technique is introduced in this work to overcome the challenges faced by conventional automated systems. Recently, deep learning techniques have been used by many companies namely Adobe,Apple,Baidu,Facebook,Google, IBM, Microsoft, NEC, Netflix, and NVIDIA. In our work, we have used an eleven layer deep CNN for the classification.
因此，本研究引入深度学习技术，以克服传统自动化系统所面临的挑战。最近，许多公司都使用了深度学习技术，如Adobe、苹果、百度、Facebook、谷歌、IBM、微软、NEC、Netflix和Nvidia。在我们的工作中，我们使用了11层的CNN进行分类。
Deep learning is a representation based learning which consists of an input layer, hidden layers, and an output layer. A representation based learning is a set of systematic procedures that provides a network to be fed with raw data and automatically learns the necessary representations for classification.The term deep describes the multiple stages in the learning process of the network structure. The deep learning neural network is trained using the backpropagation algorithm. The CNN is one of the most popular neural network techniques.
深度学习是一种基于表示的学习，它由输入层、隐藏层和输出层组成。基于表示的学习是一组系统化的过程，它为网络提供原始数据，并自动学习分类所需的表示。deep一词描述了网络结构学习过程中的多个阶段。采用反向传播算法对深度学习神经网络进行训练。CNN是最流行的神经网络技术之一。
CNN has been successfully utilized in computer vision since the early 21st century. It performed well in recognizing
handwritten digits, detecting objects, and speech recognition. It has been used in the medical research field such as
analyzing health informatics, and medical images using computed tomography(CT) images, fundus images, histopathological images, magnetic resonance (MR) images, and X-ray images as well. It is also noted that researchers in the medical analysis field are moving into CNN and obtaining desirable results. Furthermore,we applied CNN in our previous work.Our proposed system achieved the highest accuracy of 92.50% and 94.90% in the detection of arrhythmias with two and five seconds ECG signal. Hence, the CNN has performed well in the biomedical signal and image processing domain. So, in this work, we employed it for the automated diagnosis of MI using ECG signals with and without noise.
自21世纪初以来，CNN已经成功地应用于计算机视觉领域。它在识别手写数字、目标检测和语音识别。它已经被用于医学研究领域，如健康信息分析、计算机断层扫描（CT）、眼底图像、组织病理学图像、磁共振（MR）图像和X射线图像等。另外，医学分析领域的研究人员正在进入CNN领域，并取得了令人满意的结果。此外，我们在之前的工作中应用了CNN，我们提出的系统在用于2秒和5秒的ECG信号检测心律失常时达到了92.50%和94.90%的最高准确率。因此，CNN在生物医学信号和图像处理领域表现良好。因此，在这项工作中，我们使用它来自动诊断心肌梗死的心电图信号有无噪声。
2 Data used
In this work, the ECG signals were obtained from the ECG database (Physikalisch-Technische Bundesanstalt diagnostic
ECG database). This database provides ECG data of 200 subjects (148 MI and 52 healthy subjects). Also, 12 leads signals were recorded from each subject. In our present work, we have used only lead II. Table 1 presents the characteristics of the
ECG data obtained from PTB database.
在这项工作中，ECG信号是从ECG数据库（Physikalisch Technische Bundesanstalt Diagnostic）获得的。该数据库提供200名受试者（148个心肌梗塞和52名健康的受试者）的心电图数据。另外，12线信号记录了每个受试者的情况。在我们目前的工作中，我们只使用了II电极。表1显示了从PTB数据库获得的心电图数据。
Each signal is sampled at 1000 samples per second. We have used a total of 10,546 normal ECG beats and 40,182 MI ECG beats for this study. Each ECG beat consists of 651 samples comprising of one P-QRS-T wave.
每个信号以每秒1000个样本的速度采样。本研究共使用10546个正常心电图的心跳和40182个心肌梗塞心电图的心跳。每一个心电图的心跳由651个样本组成，包括一个P-QRS-T波。
3 Methodology
3.1 pre-processing
In this work, we validate our proposed method with two sets of ECG data. Both datasets consist of the same number of ECG beats. However, in one of the dataset, we denoised and removed the baseline wander from the ECG signal using Daubechies wavelet 6 mother wavelet function. But, in the other dataset, we retained the noises present in the ECG signals. Then, we carried out the R-peak detection on both datasets (with and without noise) using Pan Tompkins algorithm.
在这项工作中，我们用两组心电图数据来验证我们提出的方法。两个数据集包含相同数量的心电图心拍。然而，在其中一个数据集中，我们使用db6母小波函数(Daubechies wavelet 6 mother wavelet function)对心电图信号去除噪声和基线漂移。但是在另一个数据集中，我们保留了心电信号中的噪声。然后我们使用Pan Tompkins算法对两个数据集（有噪声和无噪声）进行了R峰检测。
All the ECG signals are segmented using the detected R-peaks without the inclusion of the first and last beat. Each segment is normalized with Z-score normalization to address the problem of amplitude scaling and eliminate the offset effect before feeding the ECG segments into the 1-dimensional deep learning CNN for training and testing. Each ECG beat consists of 651 samples (250 samples before R-peaks detection and 400 samples after R-peaks detection). Typical ECG beat with and without noise used in this study is shown in Fig. 2.
所有的心电信号都是用检测到的r峰来分割的，不包括第一次和最后一次的搏动。在将心电图片段送入一维CNN进行训练和测试之前，每一分段都进行了Z-score归一化处理，以解决幅度缩放问题，消除偏移效应。每个ECG心拍由651个样本组成(R峰检测前250个样本，R峰检测后400个样本)。本研究中有噪声或无噪声的典型ECG心拍如图2所示。
3.2 The architecture
The standard architecture of a CNN consists of four stages (i) Convolution, (ii) Rectified linear activation function, (iii)Pooling function, and (iv) Fully connected layer. Fig. 3 shows a graphical representation of the architecture of our proposed system. Table 2 summarizes the details of the CNN structure used in this work.
标准的CNN结构包含4个阶段(i)卷积(ii)校正线性激活函数(iii)池化函数以及(iv)全连接层。图3是我们提出的系统架构的图形表示。表2总结了这项工作中使用的CNN结构的细节。
(i) Convolution layer
The convolution layer is the main building block of a CNN. This layer does most of the computational intensive lifting.The prime objective of convolution is to extract features from the input ECG signals. The convolution layers are arranged in feature maps(11 layers of feature maps in total).
(i)卷积层
卷积层是CNN的主要组成部分。这一层完成了大部分计算密集型提升 (?)。卷积的主要目的是从输入的心电信号中提取特征。卷积层分布在特征图中(特征图共有11层)。
(ii) Rectified linear activation function
In general, rectified linear activation serves to map nonlinearity into the data. In this work, the leaky rectifier linear unit (LeakyRelu) is used as an activation function for layers 1, 3, 5, 7, 9, and 10. Also, the softmax function is implemented for layer 11 (last layer).
(ii)线性整流激活函数
一般来说，ReLu激活函数有助于将非线性关系映射到数据中。在这项工作中，带泄露线性整流函数(Leaky ReLu）被用作层1、3、5、7、9和10的激活函数。此外，第11层（最后一层）实现了softmax函数。
(iii) Pooling function
Pooling also referred to as downsampling which is an operation to condense features and computational complexity of the network. The max-pooling operation is employed in this work. Max-pooling outputs only the maximum number in each kernel, thus reducing the feature map size.
Kernel size also refers to the size of the filter which convolves around the feature map while stride controls how the filter convolves around the feature map. The amount by which the filter slides is the stride. In this work, the stride is set at 1. Therefore, the filter convolves around the different layers of feature map by sliding one unit each time.
(iii)池化函数
池化也被称为欠采样，这是一种用于压缩网络的特征和的计算复杂度的操作。这项工作中用了最大池化操作。最大池化输出只包括每个核中最大的数，因此减少了特征图的大小。
核大小即滤波器的大小，该滤波器在特征图上进行卷积，同时stride控制滤波器在特征图上如何卷积。stride控制每次滤波器滑动的改变量。在这项工作中，stride设为1。因此，滤波器通过每次滑动一个单位来对不同层的特征图进行卷积。
(iv) Fully connected layer
The final layer of the fully-connected network is a softmax layer with an output of X dimensional vector where X is the number of classes that we desire to have. In this study, it is a two-class (normal and MI ECG signals) problem, hence, X is set at 2 in this work.
The input layer (layer 0) is convolved with a kernel size of 102 to form the first layer (layer 1). After which, a max-pooling of size 2 is applied to every feature map (layer 2). After performing the max-pooling operation, the number of neurons reduces from 550 × 3 to 275 × 3. Then the feature map from layer 2 is convolved with a kernel (filter of size 24) to form layer 3. A max-pooling is again applied to every feature map (layer 4). After that, a feature map from layer 4 is convolved with a filter of size 11 to produce layer 5. A max-pooling of size 2 is applied to every feature map to reduce the number of neurons to 58 × 10 (layer 6). Subsequently, the feature map in layer 6 is convolved with a kernel (filter of size 9) to form layer 7. A max-pooling is once again performed (layer 8). Finally, in layer 8, the neurons are fully connected to 30 neurons in layer 9. Layer 9 is connected to 10 neurons in layer 10. Layer 10 is connected to the last layer with 2 output neurons.
(iv)全连接层
最后一层全连接网络层是一个输出X维向量的softmax层，X代表我们希望获取的分类的数量。在这项研究中，这是一个二分类（正常和心肌梗塞）问题，因此X在这里被设置为2。
输入层（第0层）经过size为102的卷积核卷积之后形成了第一层。在这之后，size为2的max-pooling操作被用于每个特征映射上（第二层）。在进行最大池化运算后，神经元的数量从550x3降至275x3。然后第二层的特征图进行了卷积核大小为24的卷积运算来形成第三层。Max-pooling操作被再次应用于每个特征映射上（得到第四层）。之后，来自第4层的特征映射与大小为11的卷积核卷积以产生第5层。在每个特征映射上应用大小为2的最大池化操作，以将神经元数量减少到58×10（第6层）。随后，第6层中的特征映射与卷积核（大小为9）进行卷积形成第7层。再次执行最大池化（第8层）。最后，在第8层神经元与第9层的30个神经元全连接。第9层与第10层的10个神经元全连接。第10层与最后一层(2个输出神经元)全连接。
3.2.1 Training
A standard backpropagation with a batch size of 10 is executed in this work. The regularization, momentum, and learning rate parameters are set to 0.2, 3 × 10−4, and 0.7 respectively. These parameters are tuned accordingly to obtain optimum performance. The function of these parameters are as follows :
a. Regularization: To prevent overfitting of the data.
b. Momentum: To control how fast or slow the network learn during training.
c. Learning rate: To help in the convergence of the data.
在这项工作中我们进行了batch size为10的标准反向传播。正则化参数、动量和学习率分别设置为0.2,3x10-4和0.7。这些参数被相应的微调来获得最佳表现。这些参数的作用如下：
a.正则化：防止数据过拟合。
b.动量：控制训练网络时的学习速度。
c.学习率：帮助数据收敛。
3.2.2. Testing
In this work, we ran a total of 60 epochs of training and testing rounds. At the end of every epoch, our proposed algorithm validates the CNN model. Out of the
training ECG beats, we used
to validate our proposed algorithm. Fig. 4 shows the apportioning of the total ECG beats for training and testing purposes.
在这项工作中，我们一共进行了60轮的训练和测试。在每轮结束时，我们提出的算法验证了CNN模型，在
的ECG节拍中，我们选择其中的
来训练我们提出的算法。图4展示了用于训练和测试的ECG节拍的分配。
3.3 k-fold cross-validation
We have employed a 10-fold cross-validation strategy in this work. We separated our total ECG beats almost equally into 10 segments.
ECG beats are used in the training of CNN while the remainder (
) of the ECG beats are used to validate the performance of our proposed system. This approach is iterated 10 times by shifting the test data. The performances(accuracy, sensitivity, and specificity) are evaluated in each iteration. Finally, the performances recorded in all 10 iterations are averaged and considered as the overall performance of our proposed system.
我们在这项工作中使用了10折交叉验证策略。我们几乎把这些ECG节拍平分成10段。
的ECG节拍被用于CNN的训练。其余的(
)个ECG节拍用于验证系统的性能。该方法通过变换数据进行了10次迭代。在每次迭代中对表现（准确率、敏感性和特异性）进行评估。最后，在10次迭代中记录的表现进行平均并被视作我们提出系统的总体性能。
4 Results
In this study, we trained our algorithm on a workstation with two Intel Xeon 2.40 GHz (E5620) processor and a 24GB RAM. It typically took approximately 2151.055 s to complete an epoch of training for ECG beats data with noise and 2025.178 s for ECG beats data without noise.
在这项研究中，我们使用两个Intel Xeon 2.40 GHz（E5620）处理器和24GB RAM在工作站上训练我们的算法。它通常花费大约2151.055秒来完成具有噪声的ECG节拍数据的训练和2025.178秒的无噪声的ECG节拍数据。
The confusion matrix for ECG beats with noise and without noise are presented in Tables 3 and 4 respectively. It can be observed from Table 3 that, out of 10,546 normal ECG beats, approximately 7.17% of the ECG beats are wrongly classified as MI. Likewise, for MI, a total of 6.29% of ECG beats are wrongly classified as normal ECG beats. Similarly, in Table 4, 94.19% of ECG beats are correctly classified as normal ECG beats and 4.51% are wrongly classified as normal ECG beats.
表3和表4分别列出了有噪声和无噪声的心电图搏动的混淆矩阵。从表3可以看出，在10546个正常ECG节拍中，大约7.17%的ECG节拍被错误地归类为MI。同样，对于心肌梗死，6.29%的心电图被错误地归类为正常心电图。同样，在表4中，94.19%的心电图被正确分类为正常心电图，4.51%被错误分类为正常心电图。
Furthermore, the PPV values for each class (normal and MI) are recorded in Tables 3 and 4. In Table 3, the PPV in the normal class is 79.48% whereas the PPV in the MI class is 98.03%. This shows that the probability of correctly detecting the MI ECG signals from the ECG signals is higher as compared to the correct detection of normal ECG signals. Similarly, in Table 4, the PPV in the normal and MI classes are 84.56% and 98.43% respectively. This also shows that the probability of identifying MI ECG signals is higher than the identification of normal ECG signals in the ECG signals with noise removal.
此外，表3和表4记录了每个类别（正常和mi）的ppv值。在表3中，正常组的ppv为79.48%，而mi组的ppv为98.03%。这表明，从心电信号中正确检测mi-ecg信号的概率比从正常心电信号中正确检测mi-ecg信号的概率高。同样，在表4中，正常组和mi组的ppv分别为84.56%和98.43%。这也表明，在去除噪声的心电信号中，mi-ecg信号的识别概率高于正常心电信号的识别概率。
The performance rate of both ECG beats with and without noise are summarized in Table 5. An average accuracy, sensitivity, and specificity of 93.53%, 93.71%, and 92.83% are achieved using ECG beats with noise introduced respectively. Furthermore, the highest average accuracy of 95.22% sensitivity of 95.49% and specificity of 94.19% is obtained for ECG beat without noise.
表5总结了有噪声和无噪声两种心电图跳动的表现率。引入噪声后的心电图平均准确率为93.53%，敏感性为93.71%，特异性为92.83%。此外，在无噪声条件下，平均准确率为95.22%，敏感性为95.49%，特异性为94.19%。
5 Discussion
Table 6 summarizes the various techniques employed by the researchers for automated detection of MI using ECG signals obtained from the same public database (PTBDB). However, not all studies are performed with lead II ECG signals. The majority of the researchers used 12 leads ECG signals in their studies [2,5,22,24,33,35]. In our previous study, we have used all 12 leads ECG signals to compare the results of the different leads. Banerjee et al.conducted a study with lead III ECG signals. They have used lead III in their work and found morphological differences in MI and normal ECG signals in the QT zone. We have used lead II in this study as it is a commonly used lead for basic cardiac monitoring. Further, lead II can provide good ECG morphological information.
表6总结了研究人员使用从同一公共数据库（PTBDB）获得的ECG信号自动检测心肌梗死的各种技术。然而，并不是所有的研究都是用II导联心电图信号进行的。大多数研究者在研究中使用12导联心电图信号[2,5,22,24,33,35]。在我们之前的研究中，我们使用了所有12导联的心电图信号来比较不同导联的结果。Banerjee等人对三导联心电图信号进行了研究。他们在工作中使用了导联iii，并在QT区发现了mi和正常ecg信号的形态学差异。我们在这项研究中使用了II导联，因为它是基本心脏监测的常用导联。此外，导联ii可提供良好的心电图形态学信息。
It can be noted from Table 6 that the proposed system performed better using the ECG beats without noise. Normally,noise is unwanted information present in the signal. Hence, the noise present in the ECG beats reduces the overall performance of the proposed system. Nevertheless, we achieved comparable results for both with and without noise ECG beats. Therefore, this proves that our proposed method is robust to noise. This also implies that our proposed CNN model can understand the underlying structure of noisy ECG beat. Thus, we might be able to accurately classify the unknown noisy ECG beat with our proposed system.
从表6可以看出，所提出的系统在无噪声的情况下使用ECG节拍表现更好。通常，噪声是存在于信号中的不需要的信息。因此，心电图中存在的噪声降低了系统的整体性能。尽管如此，我们在有和无噪声的心电图上取得了接近的结果。因此，这证明了我们提出的方法对噪声是鲁棒的。这也意味着我们提出的CNN模型可以理解噪声心电节拍的基本结构。因此，我们提出的系统可以准确地对未知的有噪声的心电节拍进行分类。
The performance of our proposed system is comparable to the performances presented in Table 6. In our work, we have used deep learning method. Hence, the CNN need not perform the feature extraction and selection process in signal analysis.This is the advantage of deep learning over the traditional machine learning algorithms. Thus, we need not experiment with different types of feature extraction or feature selection techniques. We are also not required to manually develop an optimum set of features to be fed into the classifiers. Also, the performance of our proposed method will improve with the number of data. Big data is required to train our proposed system for better performance.
我们提出的系统的性能与表6所示的性能相当。在我们的工作中，我们采用了深入学习的方法。因此，CNN在信号分析中不需要进行特征提取和选择过程，这是传统机器学习算法的优势所在。因此，我们不需要试验不同类型的特征提取或特征选择技术。我们也不需要手动开发一组最佳的特征来输入分类器。同时，随着数据量的增加，我们提出的方法的性能也会提高。需要大数据来训练我们提议的系统以获得更好的性能。
The main highlights of our proposed algorithm are as follows:
i. Feature extraction and selection techniques are not needed.
ii. 11-layer deep CNN is implemented.
iii. 10-fold cross-validation is done in this work, hence increasing the robustness of the system.
iv. Denoising is not required.
我们提出的算法的主要亮点如下：
i.不需要特征提取和选择技术。
ii.实现了11层深度的CNN
iii.在这项工作中进行了10折交叉验证，从而提高了系统的鲁棒性。
iv.不需要去噪。
The drawbacks of our proposed algorithm are as follows:
i. It is computationally intensive to learn the features.
ii. It requires a huge diverse of data.
我们提出的算法的缺点如下：
i.学习这些特征需要大量的计算。
ii.它需要大量多样的数据。
In fact, the long training time is secondary, if our proposed system can classify normal and MI classes accurately. Furthermore, once our proposed system is trained, the system can identify an unknown ECG beat immediately. Moreover, given that CNNs are concurrently-based algorithms, training the CNNs with graphics processing unit (GPU) will help to reduce the complexity and power consumption due to computation.
事实上，如果我们的系统能够准确地分类正常类和心肌梗塞类，那么长的训练时间是次要的。此外，一旦我们提出的系统被训练，该系统可以立即识别一个未知的心电节拍。此外，考虑到CNNs是基于并发的算法，利用图形处理单元（GPU）来训练CNNs将有助于降低由于计算引起的复杂性和功耗。
As part of our future study, we intend to boost the performance and reliability of our proposed system by applying bagging algorithm in our next work and to obtain more ECG data from other open source databases. We also intend to extend this approach to other cardiovascular diseases such as heart failure, hypertensive heart disease, cardiomyopathy.
作为我们未来研究的一部分，我们打算在下一步工作中应用bagging算法来提高系统的性能和可靠性，并从其他开源数据库中获取更多的心电数据。我们还打算将这种方法推广到其他心血管疾病，如心力衰竭、高血压性心脏病、心肌病。
Conclusion
The early diagnosis of MI can save life and can help to provide timely treatment. Thus, it is necessary to go for annual health checkups. The ECG is the primary tool to diagnose the electrical activity of the heart. Any abnormalities present in the heart activity is reflected in the ECG signals. However, it is challenging and time-consuming to visually assess the ECG signals. Therefore, implementing a CAD system in clinical settings will ensure an objective and fast diagnosis of MI. In this work, we proposed a novel method to automatically diagnose MI using 11-layer deep CNN. We have used two different datasets (with and without noise) to evaluate the effectiveness of our proposed method. We have achieved an average accuracy, sensitivity, and specificity of 93.53%, 93.71%, and 92.83% respectively for ECG beats with noise. Our proposed system attained high-performance results even though there are noises present in the ECG beats. This suggests that our system can recognize the class of the ECG signals even with the presence of noise in the signal. Also, we obtained an average accuracy,sensitivity, and specificity of 95.22%, 95.49%, and 94.19% for ECG beats without noise. This shows that the overall performance of our proposed system is good enough and hence, can be introduced in clinical settings. Our proposed system can assist doctors in their diagnosis.
早期诊断心肌梗死可挽救生命，有助于及时治疗。因此，每年进行一次健康检查是必要的。心电图是诊断心脏电活动的主要工具。心脏活动中出现的任何异常都反映在心电图信号中。然而，对心电信号进行视觉评估是一项具有挑战性和耗时性的工作。因此，在临床环境下实施计算机辅助诊断系统，可以保证心肌梗死的客观、快速诊断。本文提出了一种基于11层的CNN的心肌梗死自动诊断方法。我们使用了两个不同的数据集（有噪声和无噪声）来评估我们提出的方法的有效性。对于有噪声的心电图，平均准确率为93.53%，敏感性为93.71%，特异性为92.83%。我们提出的系统，即使在心电图跳动中有杂讯，也能取得高性能的结果。这说明我们的系统即使在有噪声的情况下也能识别出心电信号的类别。同时，我们获得了无噪声心电图的平均准确性、敏感性和特异性分别为95.22%、95.49%和94.19%。这表明我们提出的系统的整体性能是足够好的，因此，可以引入临床设置。我们提出的系统可以帮助医生进行诊断。
词汇
myocardium n.心肌
heart muscle 心肌
artery n.动脉 （pl. arteries）
coronary artery 冠状动脉
buildup of plaques 斑块堆积
blood clots 血块/血栓
essential 极其重要的
noninvasive 无创的
strenuous 费力的
arrhythmias 心律失常
baseline wander 基线漂移
r-peak r峰
