ECG Pattern Recognition Technique for Atrial Fibrillation Detection

Arrhythmia detection system is one of the most common medical applications that are based on the use of ECG. Kiranyaz et al. [8] suggested an ECG beat classification system. It is based on 1D-CNN for real-time patient-specific ECG classification. As CNNs are designed for extracting features from low to high level patterns, this model involves a feature extraction block followed by a classification block. These blocks are fused into a single learning block targeting the speed-up of the classification. ECG analysis represents a fundamental task in clinical ECG workflow [9]. Numerous deep learning techniques have been developed for ECG classification but for different goals [10-14]. For this latter, automated AF detection has a large percentage of concerns from the research community [10, 15-20]. New methods are presented in Refs. [21-23], which present new DL approaches used to detect AF in real time. For example, Andersen et al. [21] used an end-to-end NN cleverly combining CNN with RNN (i.e. Recurrent Neural Network) aiming to collect from input signals their characteristics high level. Eighty-five classes, taken from three different datasets, were used to test this hybrid structure. A particular performance of high processing speed has been demonstrated. Its robustness was tested on the three datasets and reached 98.96% and 86.4% specificity and sensitivity, respectively. Many researchers using different approaches and scenarios have tackled the task of arrhythmia classification based on ECG signals. Rajpurkar et al. [24] have used a DNN algorithm for ECG classification using a raw data of 30 seconds of length. They could classify 14 different classes including AF. AF presents very complex input data for the NN. DNN have shown an excellent performance in learning non-linear input data. As DNN is able to learn complex patterns presenting AF in ECG data, it can widely help scientists on finding the significant parts of ECG data to focus on them during the training phase. Actually, using a CNN, the accuracy reaches 95% [21]. This technique has shown its ability to detect AF in a shorter processing time compared to state-of-the-art techniques. He et al. [17] have proposed an improved architecture. In fact, machine-learning techniques have been used for automatic AF detection. An essential step when starting any classification system is the pre-processing step. It includes in cascade a wavelet transform (WT) and segmentation algorithm, as shown in Figure 1. A stationary WT was introduced to increase the robustness of the analysis by limiting repeatability issues. Finally, a feature extraction step must be performed with the determination of the peak-to-average power ratio and the log-energy entropy. In addition, these parameters were used by SVM to detect the AF. They achieved a 99.5% of ROC and 97.0% and 97.1% for sensitivity and specificity, respectively. In Ref. [16], the use of CNN comes after a pre-processing flow in which signals were filtered by Haar wavelet then segmented into segments of 30 seconds long. In the second stage, the CNN is applied with 11150 parameters passing through its three building blocks: convolution, pooling, and fully connected layers. They used 200 epochs in the training process, and they obtained a prediction accuracy of 98.7%. In the study [17], after segmenting the MIT-BIH dataset, they applied WT to represent the ECG dataset in time and frequency domains. Next, they proposed a CNN, which includes nine layers to perform feature extraction (four convolutional, three dropout and two max-pooling layers), two fully-connected layers to map the extracted features to the final output, and one flattened layer. Finally, the CNN parameters were adaptively optimized using the Gradient Descent (GD) algorithm. In this way, the results of AF detection (in terms of accuracy) have been improved. Two NN architectures have been used such as RNN and CNN in which the area under the receiver operating characteristic curves of the 2 DL classifiers was 0.998% [10]. Ma et al. [20] proposed an improved NN-encoder technique for AF detection. It plays a remarkable role in learning the high level of ECG features. Thanks to the Stack Sparse automatic encoder, they could achieve 98.3% accuracy. Chetan et al. [25] have suggested a low computational cost application to detect AF episodes in ECG signals using Artificial NN (ANN) in a robust way. The model was trained and tested using the MIT-BIH arrhythmia dataset [26]. The system achieved a detection accuracy coming up to 98.3%. It is known that early detection of AF diseases is very helpful for patients to avoid the chances of heart stroke. In this context, Ghosh et al. [27] proposed a novel method for AF pathology detection using the multi-rate cosine filter bank and DL, known as the Hierarchical Extreme Learning Machine (H-ELM). Even though the system achieved a very high detection rate, its use in the real-time application remains in question. Hsieh et al. [28] recommended an AF detection method based on an end-to-end 1D- CNN in order to obtain higher detection accuracy as well as minimize the network complexity. This method achieved an average F1 score of 78.2%. Warrick et al. introduced a new approach, based on a combination of a CNN and long short-term memory (LSTM) unit for ECG recordings classification [29]. They obtained an encouraging F-measure result of 0.83. Kumar et al. [11] have employed a generalized signal pre-processing technique using a multi-layer perceptron network (MLP). The average accuracy of 98.72% has been achieved on the MIT database. Results show that both the combined VMDE and VMSE features, and the radial basis function kernel-based multi-class SVM classifier are suitable to detect Arrhythmia ECG signals with an average accuracy value of 95.60% [30]. In the research [12], the coupling between the collected characteristics and the DBN classifier was able to achieve an AF classification rate of 98.27%, 97.77% and 98.67%, for accuracy, sensitivity and specificity, respectively. The system has shown its readiness for Class AF detection from huge ECG databases, such as major hospitals.

1.png

Figure 1. Synoptic flow of the proposed method for AF detection (WT: Wavelet Transform, NN: Neural Networks)

To investigate the effectiveness of the proposed ANN in classifying patient ECG data, we used standard methods such as confusion matrix, histogram error, ROC curve and cross-entropy.

4.1 Confusion matrix (CM)

It is used to measure the performance of classification algorithms by assessing their accuracy for training, validation, and test datasets. The target class represents the desired output. It is an NxN matrix (N represents the number of different classes). It is connected with a classifier indicating the expected and actual classification. Using the CM, the results are divided into four classes; (i) TP (true positive), (ii) TN (true negative), (iii) FP (false positive), and (iv) FN (false negative). Consequently, the last row and the last column of the CM show the FP and the FN results, respectively. It shows a prediction result summary of a classification problem. It responds to the problem of determining the class of each signal. The confusion matrix provides a clear picture on the classifier errors and their types. Classification accuracy alone can lead to confusion when the number of observations in each class is insufficient or the number of classes in the dataset is greater than 2. The results of the CM give a better indication of what a classification model becomes correct and what kind of errors it produces. Additionally, three standard statistical measures were used to quantitatively assess classifier performance: Sensitivity (Se, the total number of model positive hits), Specificity (Sp, the number of negative hits the model returned), the accuracy (ACC, the number of correct predictions made by the model for all types of predictions), which are given by (3) - (5), respectively.

$S e=\frac{T P}{T P+F N}$            (3)

$S p=\frac{T N}{T N+F p}$               (4)

$A c c=\frac{T N+T P}{T N+T P+F N+F P}$              (5)

Figure 5(a) shows the prediction results of experiment 1 with 100%, 80%, and 73.3% accuracy for training, validation, and testing, respectively. The classification accuracy results are acceptable. They partly prove the robustness of the proposed ANN architecture. In the learning phase, the signals are correctly classified with 0% classification error. In the validation phase, the blue background cells in the CM represent the true positive classified ECG signals, however, the white background cells represent the misclassified signals. The first class presents SNR ECG signals, the second class presents AF classified ECG signals and the third class represents noisy signals. We used 142 recordings in the training process, where 69 recordings are classified as SNR signals, 28 recordings are classified as AF signals, and 45 recordings are classified as noisy ECG signals. In experiment 1, 30 records were used for validation, of which 24 records are well classified, reaching an accuracy value of 80%. As a result, the test phase also reached a value of 73.3% where 22 records are indeed assigned to the correct class and 8 records are misclassified (3 records for the first class, 2 records for the second class and 3 records for the third class). Similarly, Figure 5(b) shows the obtained results, which are representative of the precision obtained from the classification of the ECG signal into the three classes. The approach proposed in this experiment is to use an identical number of samples in the 3 classes, aiming to achieve better efficiency and greater accuracy. As shown in the CM, the blue cells represent the true classified signals. The accuracy score obtained for the training, validation, and testing process is 92.6%, 88.5%, and 92.1%, respectively. These results exhibit high agreement, explained by the way the dataset is distributed into three classes, where each class contains all the difficult tasks. Additionally, the CM in experiment 2 shows some misclassified signals, where only 32 ECG recordings are misclassified divided into 18 AF signals and 14 SNR signals. However, for noisy signals, they are all classified the same as the ground truth would be. These results are achieved by our proposed architecture of the following specifications: (i) 10 hidden layers, (ii) 10 neurons for each layer with a sigmoid activation function, (iii) and finally Softmax classifier for system output. However, using few hidden layers (two), the classification accuracy is 82.6% while it reaches 87% using a larger number of layers (about twenty). Therefore, choosing a precise number of hidden layers to obtain a precise number of parameters shows its interest in the ECG classification process.

5a.png

(a)

5b.png

(b)

Figure 5. Confusion matrix results, (a) for experiment 1, and (b) for experiment 2

4.2 Best receiver operating curve in test-process (ROC)

ROC is often used as a graphical tool to measure how well a model fits. In our case, it demonstrates the ability of a test to discriminate between classes. In other words, the closer the curve is to the ideal value of the upper left corner, the better the fit of the model. Figure 6 represents the best ROC in the test process of experiment 2. The blue color represents the first class (AF), the green color represents the second class (NSR signals), and the red color represents the third class (signals noisy). ROC plays an important role in representing TP versus FP rates in the medical field, especially for classification purposes. In fact, in our proposed ANN architecture, we achieved an excellent classification rate reported by the sketches in Figure 6. The AF class (in blue) reaches a true positive (classified AF signals) with a value of 0.81. Regarding the second class, the true positive NSR signals reached a value of 0.95 while the noisy signals reached a ROC value of 1. To conclude, all the ROCs obtained in the test process are close to 1, which proves the agreement of the real classified signals against ground truth.

6.png

Figure 6. Best ROC in testing phase: Experiment 2

4.3 Cross entropy results

Table 4 shows the cross-entropy and resulting errors throughout the process of training, validation, and testing the neural network. Certainly, it is essential to have low cross-entropy values to obtain decent results classification. As presented in Table 4, the cross-entropy value is low and approximately, close to zero, which proves the efficiency of our proposed ANN technique. Therefore, the error (in %) indicates the percentage of misclassified data. Values of 0 and 100 indicate no classification error and maximum classification error, respectively.

Table 4. Cross-entropy results

4.4 Kappa-index and balanced accuracy results

The Kappa index is defined as the proportional reduction of the error obtained by a classification compared to the error obtained by a completely random classification. It indicates the percentage of true classes classified as non-hazardous.

$Kappa Coefficient=\frac{\left.(T S * T C S)-\sum(\text { Column Total } * \text { Row Total })\right)}{T S^{n}-\sum(\text { Column Total } * \text { Row Total })} * 100$              (6)

The kappa coefficient results obtained with 77.4% for experiment 1 and 88.11% for experiment 2 (see Table 5) demonstrate perfect agreement. Experiment 2 shows a value greater than 80%, which presents a perfect agreement between the real classified samples and the ground truth. This phenomenon of agreement is explained by the fact of the balanced accuracy. It is a metric that one can use to assess the quality of a binary classifier. It is useful, in particular; when the classes are unbalanced, i.e. one of the two classes appears much more often than the other. This usually occurs in many contexts such as the detection of abnormalities and the presence of disease,

$Balanced Accuracy =\frac{\text { Sensitivity }+\text { Specificty }}{2}$              (7)

Table 5. Accuracy assignment results (experiments 1 and 2)

 The Kappa coefficient and the balanced accuracy are respectively 77.4% and 91% for experiment 1, and 88.11% and 91.9% for experiment 2.

4.5 Discussion

The state of having 100% accuracy (i.e. 0% error) is hard to reach. Regardless of this problem, ANN algorithm succeeds to outperform the related techniques in terms of accuracy. This is because the deeper you go into the layers, the more accurate results you get. We obtained the best accuracy for the classification of MIT-BIH ECG signals using the ANN model with an accuracy value of 93.1%, where CNN comes third with 92.7%. In fact, we exceeded the results of the study [31], where an echo-state NN has been proposed, with an accuracy of 92.7%. Moreover, we outperform the state-of-the-art in Ref. [32] with more than 3% accuracy, and we have near-zero error in histogram error detection with a small cross-entropy, which evidences the robustness of our classifier model. Our results outperform most state-of-the-art results [31, 33] in terms of sensitivity and specificity. However, they are close to the results reported in Refs. [34, 35] with less complexity. Indeed, the results reported by Ghosh et al. [27] surpass our work. This is explained by the remarkable role played in the pre-processing stage, where the analysis of the ECG signal is carried out with a large number of coefficients extracted from the input ECG signal. The associated algorithm is very complex and time-consuming, making it inefficient for real-time healthcare systems. A multi-rate cosine filter bank architecture is adopted for the evaluation of the ECG signal coefficients. Thus, the more thorough the pre-processing, the greater the accuracy, sensitivity, and specificity of the DNN. Moreover, in Ref. [36], the achieved result is 99% better than our work of 9% using CNN with the cost of processing time. This is explained by the deep feature extraction performed by the CNN convolutional layers and the key role of the Adadelta network optimizer. However, the disadvantage of these methods which require manual feature selection and prior domain information in addition to high computational complexity. Therefore, their use in a real-time application is not straightforward. Nevertheless, our proposed ANN is characterized by low architecture complexity and low execution time, which make it suitable for real-time applications such as embedded healthcare systems. The obtained results described in Table 6 are more or less similar to the state of the art in a fully automatic ECG classifier. Moreover, they clearly outperform complex feature selection approaches based on ECG classifiers. Table 6 also shows an achieved result of 93%, within the same quantitative range of the discussed techniques, but with much less computational effort. Therefore, we can say that our suggested ANN architecture managed to show a better balance between classification accuracy and network parameter complexity. On one hand, this is very important in real life where most systems are modeled as complex systems, especially for the elderly, where health care is a primary problem to be solved. On the other hand, our proposed ANN model can be generalized to predict with high accuracy on unseen datasets. Moreover, the proposed ANN model speeds up the classification of ECG signals given its architecture where the hidden layers identify the most significant information from the input layer and transmit it to the next layer, living the redundant information. However, there is no particular rule for determining the optimum ANN structure, and the suitable one is obtained by trial and error method [37]. It should be mentioned that the implementation, of different algorithms, was on GPU. It consists of hundreds of ALUS, many control units, many cache memories and DRAM memory [38]. Thus, the algorithm takes less time to train and run compared to the CPU implementation.

Table 6. Comparative study with the state-of-the-art