Application of stacked convolutional and long short-term memory network for accurate identification of CAD ECG signals
Graphical abstract
Introduction
According to the American College of Cardiology, cardiovascular diseases contribute to one-third of all deaths worldwide [28]. In 2015, more than 400 million people globally were diagnosed with cardiovascular disease. Cardiovascular diseases are significant contributors to global increased mortality rates [28]. For this reason, there is a need to develop a cost-effective and efficacious approach to alleviate this global issue.
Generally, cardiovascular diseases are categorized into three groups: circulatory, electrical, and structural [17]. Coronary artery disease (CAD) (circulatory-related disease involving the coronary circulation) is the most common form of cardiovascular disease [9,28]. CAD occurs through a process called atherosclerosis, which is a disorder caused by build-up of plaque in the inner walls of the coronary arteries [9]. The plaque consists of fatty or cholesterol deposits and its accumulation will constrict blood flow through the arteries. Furthermore, as the plaque builds up, blood flow through the arteries decreases and eventually the heart muscle receives insufficient oxygen-rich blood supply [9]. Over time, the plaque may solidify and rupture. This activates circulating platelets in the blood, resulting in the formation of blood clots on the surface of the plaque That may potentially occluded the vessel entirely, and block off blood flow [12]. An entirely blocked coronary artery may trigger off heart attack (myocardial infarction) [1,4,7]. In addition, chronic CAD may result in heart muscle damage, weakening of heart muscle, and lead to other cardiovascular diseases such as arrhythmias [3,5] and heart failure [12].
The early stage of CAD normally produces no symptoms until the disease progresses into an advanced stage. Pre-symptomatic health check-ups may unearth early disease and avert further progression of the disease with timely treatment. The electrocardiogram (ECG) is a widely accessible diagnostic tool that records the electrical activity of the heart [25]. ECG signals can also be obtained during exercise stress test where the ECG signals are recorded when the subject is undergoing a physical stress [10,11]. In addition, heart rate variability (HRV) signals can be extracted from ECG signals [6,15,21,27,30]. Other common diagnostic tests include echocardiography, which uses sound waves to produce visual images of the heart in order to detect structural heart abnormality [31,32]. Still, the ECG technique is the prime choice for the primary assessment of the heart as it is cost-effective, easy to implement, and noninvasive. Manual diagnosis of ECG signals is however very challenging and tedious as the signals vary morphologically (see Fig. 1). Also, manual analysis of the ECG signals which is currently the clinically standard practice might be subject to inter-observer variability among different clinicians [24].
Thus, a computer-aided diagnosis system is initiated to overcome these limitations of visual inspection of ECG signals. There are numerous works proposed on the computerized decision support using ECG signals to diagnose the different types of heart conditions [2,8,22,26]. Table 1 shows the various studies conducted on the automated detection of CAD using ECG signals. Kumar et al. [22] developed an automated system to differentiate CAD and normal ECG signals with an accuracy of 99.60%. They decomposed the ECG beats with flexible analytic wavelet transform, and then statistically significant features extracted were fed into the classifier for classification.
Acharya et al. [8] compared the performance of higher order bi-spectrum and cumulant features to categorize normal versus CAD ECG beats. They also formulated 2 CAD indexes from the extracted bi-spectrum and cumulant features to numerically characterize the two classes of ECG signals based on a single number. Later, Acharya et al. [2] came up with an improved algorithm. They designed an 11-layer deep convolutional neural network (CNN) to classify ECG signals into normal and CAD classes. Their model attained an accuracy of 94.95% and 95.11% with 2-s and 5-s ECG segments respectively without any hand-engineered features extraction and selection processes.
Oh et al. [26] employed the common spatial pattern technique to extract significant features from decomposed ECG segments. These features were then fed into a k-nearest neighbor classifier for classification. They reported a high diagnostic sensitivity and specificity of 99.64% and 99.71% respectively with an accuracy of 99.65%. Although the proposed technique achieved the maximum performance, this approach used many features. The works by Kumar et al. [22], Oh et al. [26], and Acharya et al. [8] adopted the traditional machine learning process of features extraction, features selection, and classification.
In this work, a stacked long short-term memory (LSTM) [18] network with CNN [23] is proposed to classify normal versus CAD ECG signals. LSTM is known to be well-suited for the processes and prediction of time-series signals. However, the computation performed in LSTM is generally slower.
In this study our algorithm first slices a 5 s ECG segment (with 1285 data points) into 211 short segments. Each short segment consists of 24 data points. Instead of simply feeding these 211 short segments into layers of LSTM, we perform 2 rounds of 1D convolution-maxpooling to extract the significant features in these segments. The resultant output is a set of 50 short segments, with each segment comprising 32 data points. This process reduces the amount of data points for computation in LSTM (CNN for most of the time runs faster than LSTM). These segments are then fed to 3 layers of LSTM and a fully-connected layer to perform the diagnosis.
To the best of our knowledge, in literature there is no similar structure proposed and applied on the classification of normal and CAD ECG signals. There were deep neural network architectures (which used CNN and LSTM) proposed for the classification of atrial fibrillation [34]. However, in their work they first converted an ECG signal into a logarithmic spectrogram and then followed by a series of 2D convolutions. Whereas in this work, we perform 1D convolution (as defined in Ref. [19]) and 1D maxpooling (as defined in Ref. [19]) before we send features to LSTM.
Section snippets
Data used
The ECG data used in this work were obtained from an open source PhysioNet database [16]. The normal and CAD ECG data were downloaded from Fantasia and St Petersburg Institute of Cardiology Technics 12-leads arrhythmia respectively. Only lead II ECG signals were used for this study. The ECG data used were collected from 7 CAD (6 females and 1 male) and 40 normal (20 females and 20 males) subjects. The ECG signals were cut into segments of 5 s with each segment containing 1285 data points. Fig. 1
Methodology
A stacked CNN-LSTM model is proposed in this study to classify the ECG data into their respective classes.
Results
The stacked convolutional LSTM network is developed and trained using Keras with Theano backend [33]. It was trained on a workstation with two Intel Xeon, 2.20 GHz (E5-2650v4) processor and a 512 GB RAM with a Quadro K4200, 4 GB memory GPU. It took approximately 51 s to run a single epoch.
Table 3 shows the confusion matrix for the classification of CAD and normal ECG segments using first validation approach (non-subject specific). Very small percentage of 0.16% of the normal segments were
Discussion
It is observed that high diagnostic performances were yielded from conventional techniques [8,22,26]. Even though they have obtained remarkably high performance with advanced signal processing techniques, our proposed model does not require R-peak detection as compared to their methods. Furthermore, the published papers recorded in Table 1 have performed the ten-fold cross-validation in their studies whereas, in this work, a blindfold strategy was implemented. Nevertheless, the proposed model
Conclusion
In this study, a state-of-the-art algorithm (stacked CNN-LSTM) is employed to automatically diagnose CAD using ECG signals. The system is fully automatic and requires minimum hand-engineering to train the algorithm. We have obtained the highest diagnostic performance using 8-layer stacked CNN-LSTM network. Our system has the potential to be deployed in clinical settings to assist cardiologists to make objective and reliable diagnosis of ECG signals. Furthermore, the proposed algorithm can be
Conflicts of interest
There is no conflict of interest in this work.
References (34)
- et al.
Automated characterization and classification of coronary artery disease and myocardial infarction by decomposition of ECG signals: a comparative study
Inf. Sci.
(2017) - et al.
Automated detection of coronary artery disease using different durations of ECG segments with convolutional neural network
Knowl. Base Syst.
(2017) - et al.
Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network
Inf. Sci.
(2017) - et al.
Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals
Inf. Sci.
(2017) - et al.
A deep convolutional neural network model to classify heartbeats
Comput. Biol. Med.
(2017) - et al.
Application of empirical mode decomposition (EMD) for automated identification of congestive heart failure using heart rate signals
Neural Comput. Appl.
(2017) - et al.
Automated characterization of coronary artery disease, myocardial infarction, and congestive heart failure using contourlet and shearlet transforms of electrocardiogram signal
Knowl. Base Syst.
(2017) - et al.
Application of higher-order spectra for the characterization of coronary artery disease using electrocardiogram signals
Biomed. Signal Process Contr.
(2017) What Is Cardiovascular Disease?
(2017)- et al.
Effects of principal component analysis on assessment of coronary artery diseases using support vector machine
Expert Syst. Appl.
(2010)