ArticlesAll Issue
ArticlesComparative Neural Network Based on Template Cluster for Automated Abnormal Beat Detection in Electrocardiogram Signals
• Seungmin Lee and Daejin Park*

Human-centric Computing and Information Sciences volume 12, Article number: 51 (2022)
https://doi.org/10.22967/HCIS.2022.12.051

Abstract

Automatic abnormal beat detection reduces the time and cost for signal analysis by a cardiologist because abnormal beats rarely occur in an electrocardiogram (ECG) signal. However, the characteristics of normal and abnormal beats vary by individuals, which leads to misdetection. In this study, instead of directly training the input beats, we combine the input beats with corresponding reference normal beats and train the difference within combined beats. We can classify normal and abnormal beats through this approach, even if various types of individual normal and abnormal beats are mixed. The proposed comparative learning has the advantage of being able to classify multiple records using only one neural network. In experiments with learning five records, including premature ventricular contraction beats, we achieved 99.47% sensitivity and 99.28% accuracy for 30 records, with only one comparatively trained neural network. In addition, we confirmed that real-time processing is possible with an average processing time of 20.87 ms per beat.

Keywords

Electrocardiogram, Convolutional Neural Network (CNN), Comparative Learning, Binary Classifier, Abnormal Beat Detection, Template Cluster

Introduction

The recent development of bio-signal monitoring systems has become more important in healthcare to address increases in mortality due to heart disease and loneliness in the aging population [1, 2]. In addition, bio-signals such as electrocardiograms (ECGs), electroencephalography (EEG), and photoplethysmography (PPG) are widely used for security, human activity, and other applications [3, 4]. ECG signal analysis is an important research field because an arrhythmia due to heart disease is represented as an abnormal beat in the ECG signal [513].
In an ECG signal, a normal beat with a similar, periodically repeating waveform occurs throughout most of the signal, whereas an abnormal beat with a deformed waveform or period rarely occurs. Therefore, an accurate diagnosis relies on a signal measured over an extended time, which requires extensive time and cost for diagnosis by a cardiologist.
Recent developments in hardware and software technology have made it possible for ordinary people to measure ECG signals in their daily life. Accordingly, the demand for ECG diagnosis is rapidly increasing, and research is being conducted to automate the steps of ECG signal analysis to reduce the diagnosis time and cost. Machine learning-based classifiers are more effective than knowledge-based classifiers in classifying ambiguous data by better imitating human intuition [1417]. In particular, because ECG signal classification depends significantly on cardiologist intuition, many studies are underway to replace existing knowledge-based classification. However, because the distribution of normal and abnormal beats in ECG signals varies by individual, the accuracy of the classifier is greatly reduced when similar beats are trained in several classes during the training process.
In this study, we propose a binary classification algorithm based on comparative learning by using a corresponding reference normal beat determined by a template cluster.

Motivation
It is difficult to classify given data accurately, but it is relatively easy to determine which data are closer to a given class by comparing two data points [18, 19]. Comparative region convolutional neural networks have been proposed to do this for age classification [2022]. Similarly, it is difficult to determine whether a given beat in an ECG signal is normal or abnormal, but it is relatively easy to compare and determine whether the corresponding reference normal beat is provided. Therefore, instead of training a single input beat, we propose combining the input beat and corresponding reference normal beat and training using it. In contrast to the existing neural network for classifying beats by extracting and analyzing the feature values of the input beats, the proposed neural network learns by comparing the differences within the combined beats. This process makes it possible to imitate human intuition more effectively.

Fig. 1. Comparison between the existing and proposed classifiers.

Contribution
In general, because ECG signals have various types of beats according to the individual, the shape of both normal and abnormal beats may appear mixed. Learning these mixed data reduces the reliability of the neural network. However, generating a separate neural network for each individual would require significant cost and time. Therefore, the proposed comparative learning method trains the differences within the combined beats based on acquiring a reference normal beat determined by a template cluster. Through this, we can generate a classifier applicable to various individuals, even with a single neural network. Fig. 1 shows the comparison between the existing classifier with multiple neural networks and the proposed classifier with a single neural network.>
This paper is organized as follows. Section 2 introduces related studies on ECG signals, deep learning, and template clusters, and Section 3 introduces the proposed comparative learning-based classifier. After describing the experiments to confirm the performance of the method in Section 4, we conclude in Section 5.

Related Work

ECG Signal
2.1.1 Composition of an ECG signal
ECG signals are obtained by measuring electrical signals generated during the contraction and relaxation of the atria and ventricles. The contraction of the atria generates a P wave, and the contraction and relaxation of the ventricle generate a QRS complex and T wave, respectively [23]. When an arrhythmia occurs in the contraction and relaxation of the atrium or ventricle, the shape or period of the corresponding waveform is deformed.
Knowledge-based signal analysis detects the fiducial points, onset (start point), offset (endpoint), and peak. Then, it acquires feature values using the time or amplitude difference between the fiducial points and analyzes the deformation of the shape and period of the waveform.
Fig. 2 shows the fiducial points of the ECG signal and widely used feature values.

Fig. 2. Composition of an ECG signal.

As shown in Fig. 2, intervals, duration, and segment features are acquired by using the fiducial points of P wave, QRS complex, and T wave. Among features, the RR interval is an important feature used to diagnose abnormal beats related to premature contractions, such as premature ventricular contraction (PVC) and premature atrial contraction (PAC), and the PR interval is used to classify premature junctional contraction (PJC) caused by P wave abnormalities.

2.1.2 Abnormal beat
Arrhythmias are classified by the site of the occurrence and cause, and the characteristics of the corresponding abnormal beat vary. The American Medical Association’s ANSI/AAMI EC57:1998 classified such beats into five categories according to their location: a normal beat (N), upper ventricular arrhythmia beat (S), ventricular arrhythmia beat (V), mixed arrhythmia beat (F), and unclassified beat (Q). PhysioNet provides a variety of arrhythmia databases and categorizes the beats using 19 categories. Table 1 shows the distribution of beats occurring in the 19 beat classifications and 48 measurements from the MIT-BIH arrhythmia database (MIT-BIH ADB) dataset used in this paper [24].

Among the classes, the bundle branch block beat (L, R) and paced beat (/, f) periodically occur in place of a normal beat. In this study, we aim to detect PVC (V), which has the highest rate among all abnormal beats except the bundle branch block beat and paced beat.

Table 1. Classification of beats and distribution in MIT-BIH ADB
 Code Total number Description Class N 75,052 Normal beat Normal L 8,075 Left bundle branch block beat Bundle branch block R 7,259 Right bundle branch block beat B 0 Bundle branch block beat A 2,546 Atrial premature beat Premature a 150 Aberrated atrial premature beat J 83 Junctional premature beat S 2 Supraventricular premature beat V 7,130 Premature ventricular contraction r 0 R-on-T premature ventricular contraction F 803 Fusion of ventricular and normal beat e 16 Atrial escape beat Escape j 229 Junctional escape beat n 0 Supraventricular escape beat E 106 Ventricular escape beat / 7,028 Paced beat Paced F 982 Fusion of paced and normal beat Q 33 Unclassifiable beat Else ? 0 Beat not classified during learning

According to Table 1, about 68% of the beats are normal, and when considering that L, R, and / beats are also normal, about 89% of all the beats are normal. Therefore, we propose a comparative learning method through the acquisition of a representative normal beat using template clustering based on the characteristics of the ECG signal, in which the majority of beats appear as normal beats.

2.1.3 Beat separation
The ECG signal requires long-term measurement, and thus many beats are stored. Each beat is classified based on the detection of the QRS complex or R-peak. This is because the amplitude change of the QRS complex is significant, with the amplitude of the R-peak generally the largest.
Fig. 3 shows the two methods of beat separation.

Fig. 3. Two beat-separation methods: (a) RR interval and (b) centering on the R-peak.

In general, the R-peak-to-R-peak interval is divided into beats based on the detected R-peaks as shown in Fig. 3(a). This has the advantage of not losing any of the sample, but the problem is that the length of each interval is not constant, so information from two beats is mixed in one interval. To improve this, the region from 275 ms before to 375 ms after the R-peak includes the P wave, QRS complex, and T wave, which are mainly used for ECG signal analysis as shown in Fig. 3(b) [25]. Thus, this region is generally used to extract fiducial points based on the R-peak.

Deep Learning in MATLAB
Deep learning has recently been used in various fields. In MATLAB, pre-trained neural network models can be used by the deepNetworkDesigner tool, including SqueezeNet, GoogLeNet, ResNet-50, EfficientNet-b0, DarkNet-53, and others, and various additional models can be installed and used. Fig. 4 shows an example of GoogLeNet generated by the deepNetworkDesigner tool in MATLAB. In general, a fully connected vector of trained data is obtained from the input data through a convolutional neural network (CNN) model, and the loss between them is calculated. Then, the training data with the lowest loss are searched, and the classification is performed.

Fig. 4. GoogLeNet in MATLAB deepNetworkDesigner.

Problem
Fig. 5 shows a schematic for conventional classification.

Fig. 5. Conventional classification scheme.

A feature vector of an input beat is obtained by using a previously generated CNN model. Using the same CNN model, a feature vector and contrastive loss corresponding to N learning beats are calculated, respectively. Then, the class of the input beat (Output) is determined by following the classification result of the beat with the minimized loss.
Fig. 6 depicts an example of the training and classification process in the ECG signal.

Fig. 6. Example of the training and classification process in an ECG signal.

Depending on the class of the abnormal beat, each beat forms different clusters in the feature space. Then, the input beat is classified according to the class of the beat closest to the feature vector of the input beat.
However, in the process of training a neural network, when data of different classes are trained in the same class, similar shapes may cause misdetection, as shown in Fig. 7.
To address this, in this study, we acquired a reference normal beat and combined it with the input beat, and the difference within a combined beat was trained to minimize false detection.

Fig. 7. Example of misdetection in conventional classification.

Template Cluster
The basic reference normal beat can be generated using the average and median of the beats acquired over a long time. However, the average is significantly affected by distortion caused by abnormal beats. In addition, the median requires extensive memory and processing time for the sorting operation. Lee and Park [8, 9] proposed a method to generate a template cluster to obtain a reliable template of normal beats.

2.4.1 Pearson similarity
The Pearson similarity is well known in statistics as the Pearson correlation coefficient, which is a value quantified for linear distributions for two distributions X and Y, with a value between 1 and -1. The value indicates perfect positive linear correlation at 1, no linear correlation at 0, and perfect negative linear correlation at -1. The Pearson similarity is the value obtained by the covariance of two distributions X and Y, which is divided by the product of the standard deviations as shown in (1).

(1)

As shown in (1), by subtracting the mean and dividing by the standard deviation, the Pearson similarity is robust to baseline movements and amplitude scale changes. Thus, it is suitable for measuring the similarity of ECG signals.

2.4.2 Template cluster generation
The template cluster uses the Pearson similarity to generate templates, and it updates the matched templates in real time according to the input beat using a weighted mean. In this case, the weight used for the weighted mean is the number of updates. If there is no sufficiently similar template, the input beat is added as a new template to minimize the effect on the existing templates. Because normal beats periodically repeat with a similar shape and occupy most of the signal, the template with the highest number of updates in the template cluster represents the normal beat. Thus, we selected it as the reference normal beat.
Fig. 8 explains the steps of initializing, updating, and determining the reference normal beat of the template cluster.

Fig. 8. Reference normal beat determination scheme.

From the first input signal, the shape of the first beat are input as the first data point of the first cluster of templates, and the counting number of the first template is initialized to 1. The counting number indicates the weight of the weighted mean in the template update process and it is used to determine the reference normal beat.
After entering the initial template, clusters are updated from the sequentially input beats. The input beats measure shape similarity with each template in the template cluster using (1). Then, a template with the highest shape similarity among templates is detected. If shape similarity exceeds the threshold, then the matched template is updated using weighted mean, otherwise, the input beat is added as a new template.
The ECG signal is mostly composed of a normal beat and has a characteristic that similar shapes are repeated. Therefore, the template having the highest count can represent the normal beat, and thus we determined it as the reference normal beat.
Fig. 9 shows an example of a reference normal beat determination process.

Fig. 9. Reference normal beat determination process: (a) initialize, (b) update, and (c) determine the reference normal beat.

In Fig. 9(a), the first input beat was a PVC abnormal beat. In Fig. 9(b), the template cluster of normal beat and abnormal beat was separated through the update process, and the template with the highest count (975) was determined as the reference normal beat as shown in Fig. 9(c). Although the first input beat was an abnormal beat, we confirmed that the normal beat is stably determined as the reference normal beat. In addition, as described above, it is possible to minimize the memory and execution time by using the template cluster with weighted mean to obtain a reliable reference normal beat with minimal distortion.

Proposed Method

Fig. 10. Diagram of the proposed method.

The $i_{th}$ normal beat is similar to the $j_{th}$ abnormal beat, and in the existing method shown in Fig. 7, it is trained in the same class and causes a false detection. However, in the proposed method, the reference normal beat is combined, data of the same class as the reference normal beat are trained as normal, and data of a class not identical to the reference normal beat are trained as abnormal. In this way, the proposed method makes it possible to train different classes of similarly shaped beats correctly into different classes. The deepNetworkDesigner tool provided by MATLAB is used in the proposed method, and training is carried out using GoogLeNet. A column vector consisting of 235 samples of input beats, as shown in Fig. 3, is represented by a combination of a reference normal beat and an input beat. The beats are normalized, combined, and duplicated to a 235×34 matrix for use with the GoogLeNet model. Normalization is performed so that the minimum and maximum amplitude values of the representative normal beat become 0 and 255, respectively. Then, the amplitude value of the input beat is normalized using the weight used to normalize the representative normal beat. If the converted amplitude has a saturated value outside the range 0–255, it is adjusted to 0 or 255. Fig. 11 shows the imaging process of the combined beat. In this way, the reference normal beat is combined and imaged with the input test beat, and classification is performed.

Fig. 11. Example of an imaged combined beat for comparative learning.

Experiment

Test Data
In this study, we trained and classified normal beats and PVC abnormal beats. The data used in the experiment are from MIT-BIH ADB, provided by PhysioNet. MIT-BIH ADB consists of 48 records, each measuring about 30 minutes in length, and the class of each beat is annotated by a cardiologist. We excluded 13 records in which each bundle branch block beat and the paced beat replaced the normal beat from the experiment: records 102, 104, 107, 109, 111, 118, 124, 207, 212, 214, 217, 231, and 232. In addition, five records, which were difficult to compare to the reference normal beat, were excluded because the shape or amplitude change of the beats was very irregular: records 106, 108, 203, 222, and 230. Thus, out of 48 records, the experiment was conducted on 30 records, excluding 18 records.
We evaluated the classification performance by measuring the sensitivity and specificity according to the true positive (TP), true negative (TN), false positive (FP), and false negative (FN). TP and TN are correct classification results of abnormal and normal beats, respectively. Conversely, FP and FN are misclassification results for normal and abnormal beats, respectively. Sensitivity, specificity, and accuracy measure the non-detection rate, over-detection rate, and correct total beat detection rate, respectively, as shown in (2):

(2)

Experiments
4.2.1 Abnormal beat detection
The experiment was conducted in five stages: (1) classification rate as the size of the imaged signal increases, (2) classification rate as the number of records used for training increases, (3) classification rate with a reduced number of epochs, (4) classification rate based on the amount of training data with a non-uniform ratio, and (5) classification rate when the number of abnormal beats was increased to equal the number of normal beats, and the comprehensive comparison to previous results.

Test 1: From datum 119, 10 normal and 10 abnormal beats were trained. The classification results by image size are shown in Table 2.
This experiment showed that the non-detection rate tends to decrease as the image size increases. However, an excessively large image size significantly increases the over-detection due to overfitting for an abnormal beat.
In subsequent experiments, the experiment was performed after fixing the image size at 34.

Table 2. Classification results by the image size of the combined beat
 Image size Total beats PVC beats TP FN TN FP Sensitivity (%) Specificity (%) Accuracy (%) 34 68,614 5,454 5,375 79 62,112 1,048 98.55 98.34 98.36 84 5,407 47 60,766 2,394 99.14 96.21 96.44 134 5,417 37 60,195 2,965 99.32 95.31 95.62 184 5,428 26 60,448 2,712 99.52 95.71 96.01 234 5,414 40 60,667 2,493 99.27 96.05 96.31

Test 2: The number of records used for training was increased. The first row in Table 3 shows the training and classification results using Record 119. Additional records were added with each table row in the order 200, 208, 221, and 233, including multiple PVCs. Table 3 shows the training and classification results according to the number of training records.
Adding one record greatly reduced the non-detection rate, and adding additional records reduced over-detection by additionally learning the variance in the shape of the normal beat.

Table 3. Classification results according to the number of training records
 # of records Total beats PVC beats TP FN TN FP Sensitivity (%) Specificity (%) Accuracy (%) 1 68,614 5,454 5,375 79 62,112 1,048 98.55 98.34 98.36 2 5,446 8 59,791 3,369 99.85 94.67 95.08 3 5,441 13 60,680 2,480 99.76 96.07 96.37 4 5,442 12 61,284 1,876 99.78 97.03 97.25 5 5,442 12 61,348 1,812 99.78 97.13 97.34

Test 3: In this test, we lowered the number of epochs to eight and compared the results to those of Test 2, which used 20 epochs. Table 4 shows the classification results with a reduced number of epochs.
In this experiment, the non-detection rate decreased, but the over-detection rate increased. In general, non-detection of abnormal beats affects misdiagnosis.
In subsequent experiments, training based on eight epochs was performed using these five records.

Table 4. Classification results with a reduced number of epochs
 # of records Total beats PVC beats TP FN TN FP Sensitivity (%) Specificity (%) Accuracy (%) 1 68,614 5,454 5,417 37 61,215 1,945 99.32 96.92 97.11 2 5,446 8 59,168 3,992 99.85 93.68 94.17 3 5,446 8 59,168 3,992 99.85 93.68 94.17 4 5,444 10 60,742 2,418 99.82 96.17 96.46 5 5,445 9 60,867 2,293 99.83 96.37 96.64
Test 4: This experiment was based on changing the amount of training data. In the previous tests, 10 normal and 10 abnormal beats from each record were used for training. In this test, the experiment was conducted using 100 normal and 10 abnormal beats for training. Conversely, the experiment was also conducted using 10 normal and 100 abnormal beats for training. Tables 5 and 6 show the experimental results for each case.
This experiment showed that as the data distribution in the training class changed, FP decreased and FN increased, and vice versa. In particular, when only Record 119 was used for training, FN was greatly reduced by increasing the number of abnormal beats in the training data.

Table 5. Classification results with 100 normal and 10 abnormal beats used in training
 # of records Total beats PVC beats TP FN TN FP Sensitivity (%) Specificity (%) Accuracy (%) 1 68,614 5,454 5,187 267 62,551 609 95.1 99.04 98.72 2 5,367 87 61,652 1,508 98.4 97.61 97.68 3 5,336 118 62,050 1,110 97.84 98.24 98.21 4 5,265 189 62,897 263 96.53 99.58 99.34 5 5,333 121 62,815 345 97.78 99.45 99.32

Table 6. Classification results with 10 normal and 100 abnormal beats used in training
 # of records Total beats PVC beats TP FN TN FP Sensitivity (%) Specificity (%) Accuracy (%) 1 68,614 5,454 5,445 9 60,150 3,010 99.83 95.23 95.6 2 5,451 3 58,906 4,254 99.94 93.26 93.8 3 5,447 7 58,425 4,735 99.87 92.5 93.09 4 5,448 6 59,378 3,782 99.89 94.01 94.48 5 5,447 7 60,008 3,152 99.87 95.01 95.4

Test 5. The same experiment was conducted after adjusting the training data to use half the number of abnormal beats, and the results were compared by synthesizing the previous results (Table 7).
Over-detection was greatly reduced by adjusting the number of abnormal beats used for training to half the number of abnormal beats in the data. Although the non-detection rate increased slightly, the classification was effective when compared with the total number of beats.
As such, we confirmed that the proposed comparative learning can be applied to various records even with one neural network. In particular, in existing experiments to classify beats, 80% of about 100,000 data records from MIT-BIH ADB were used for learning, and 20% were used for testing, with learning carried out with more than 100 epochs
[26, 27]. In contrast, it was confirmed that the proposed method has an excellent detection rate for 68,614 heartbeats, despite learning a total of 100 small beats for up to five records with a low repetition rate of 8 or 20 epochs.

Table 7. Summarized classification results
 Test method Total beats PVC beats TP FN TN FP Sensitivity (%) Specificity (%) Accuracy (%) DB1 68,614 5,454 5,442 12 61,348 1,812 99.78 97.13 97.34 DB2 5,445 9 60,867 2,293 99.83 96.37 96.64 DB3 5,333 121 62,815 345 97.78 99.45 99.32 DB4 5,447 7 60,008 3,152 99.87 95.01 95.4 DB5 5,425 29 62,696 464 99.47 99.27 99.28
DB1–DB4 correspond to Row 5 in Tables 3–6; DB5, all five records using training data with half the number of abnormal beats in each record.

4.2.2 Processing time
Abnormal beat detection is performed in two steps: (1) reference normal beat determination using template clusters and (2) combined beat image generation for input beats with the reference normal beat and classification. We measured the processing time for the 30 records used in Section 4.2.1. The experiments were conducted on 64-bit Windows 10 running on an Intel i5-10400, 2.90-GHz CPU with 32 GB DDR4 RAM and MATLAB R2021a.
Fig. 12 shows the processing time of the reference normal beat determination for 30 records and the average processing time.
When the number of templates generated in the template cluster is large, the number of comparison operations increases, and the processing time likewise increases. However, even in the longest record (Record 200), the processing time was less than 5 seconds. Thus, we confirmed that the result is excellent, considering that the length of each record is about 30 minutes and 2,300 beats. Furthermore, the weighted mean used in the template cluster generation process minimizes memory usage, so real-time processing in low-memory embedded devices is anticipated.

Fig. 12. Processing time to acquire the reference normal beat using template clusters.

Fig. 13 shows the processing time to generate the combined beat with the reference normal beat and classify abnormal beats.
The data in Fig. 13(a) and Fig. 13(b) confirm that the total processing time increases in proportion to the number of beats. Fig. 13(c) shows that the performance is stable and fast when converted to milliseconds per beat, at 20.82 ms. In particular, it can be seen that the average processing time for each record is stable at approximately 1 ms. In future studies, if a lightweight neural network other than GoogLeNet is applied and the process of combining the input data is simplified, we expect that both learning and classification will be possible through lightweight embedded devices.

Fig. 13. Processing time to detect abnormal beats: (a) processing time for whole beats, (b) number of beats, and (c) average processing time per beat.

Discussion of Over-detection
There are records with a high over-detection rate, such as Records 105, 116, and 202. Analyzing their characteristics reveals the following. In Record 105, ventricular fibrillation irregularly distorts the surrounding normal beats, resulting in over-detection. In Record 202, the shape of the normal beat is changed during the measurement, or the shape of the normal beat is significantly changed due to tachycardia and noise. In Record 116, an interval occurs in which the signal is very weak, and as the P and T waves are restored more quickly than the QRS complex when the signal amplitude is restored, the beats are classified as abnormally large P or T wave beats, resulting in over-detection.
As the amount of training data increases, such over-detection can be significantly reduced, as shown in Table 7, by recognizing the wider beat shape changes. However, this also results in abnormal beats with minimal shape changes classified as normal beats. In the proposed comparative learning, because the increase in non-detection is minimal compared to over-detection, we can confirm that a stable classification rate is maintained.

Comparison with Related ECG Classification Methods
A comparison between the proposed ECG classification method and other method is conducted in Table 8 [28–33]. Some existing methods have classification results other than PVC, but we only considered the classification performance of normal beat and PVC.
Most of the papers are using MIT-BIH ADB for testing except support vector machine (SVM) method. By analyzing the experimental results, we confirmed that the performance of the proposed method is the best in all performance except for sensitivity (specificity) of AdaBoost method. In particular, we confirmed that it has better performance on a large number of test data even with a small number of learning compared to other methods.
Also, in the case of the CNN method, by applying the comparative learning scheme, the generation of the neural network is effectively made even with a small number of training, so that more effective detection is possible. Thus, we confirmed that the proposed comparative learning method is a useful approach for grafting lightweight embedded devices in the future works.

Table 8.Comparison with related ECG classification methods
 Method Dataset Train Test Sensitivity (%) Specificity (%) Accuracy (%) Oh et al. [28] U-Net MIT-BIH ADB 90% of dataset 10% of dataset 92.81 98.98 98.49 Chen et al. [29] KNN MIT-BIH ADB Approx. 50,000 Approx. 50,000 99.05 96.01 96.26 He et al. [30] SVM CHN, Incartdb, EDGAR 35,202 506 97.7 96.7 97.2 Krishnan et al. [1] Fuzzy logic MIT-BIH ADB N/A 3,880 95.6 96.8 96.6 Malik et al. [32] AdaBoost MIT-BIH ADB 51,021 49,712 93.82 99.65 99.27 Wang et al. [33] CNN MIT-BIH ADB 148,901 2,187 95.47 97.72 98.25 Proposed method Comparative CNN MIT-BIH ADB 5,454 68,614 99.47 99.27 99.28

Conclusion

Using the proposed comparative learning, it was possible to classify various records with one neural network. A reliable reference normal beat was obtained using the template cluster, resulting in effective comparison and excellent classification results obtained with only a small amount of trained data from a single record. Through various experiments based on the number of records and amount of data used for learning, the recognizable range of beat variations increased, thereby reducing false detections. As such, it was possible to confirm the excellent abnormal beat classification results using comparative learning.
If a reference normal beat was not stably obtained, comparative learning was not performed correctly, resulting in erroneous detection. However, by minimizing non-detection, it is possible for a cardiologist to provide the correct classification, and it is expected that sufficient cost and time savings are possible compared to the total amount of data.
In future research, we will search for CNN models other than GoogLeNet that are effective for comparative learning, and we will also study classification methods other than PVC. As the types of abnormal beats to be classified increases, it is expected that an approach such as fuzzy-set theory will be effective for classification. In addition, through research on local noise suppression and local representative normal beat acquisition, we will research how to adaptively handle changes in the shape of a normal beat during long-term ECG measurement.

Author’s Contributions

Conceptualization, SL, DP. Funding acquisition, DP. Investigation and methodology, SL, DP. Project administration, SL. Resources, SL. Supervision, DP. Writing of the original draft, SL. Writing of the review and editing, SL, DP. Software, SL. Validation, SL, DP. Formal Analysis, SL. Data Curation, SL, DP. Visualization, SL.

Funding

This study was supported by the BK21 FOUR project funded by the Ministry of Education, Korea (4199990113966), and the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2018R1A6A1A03025109, 10%, 2020R1I1A1A01072343, 10%, NRF-2022R1I1A3069260, 20%) and by Ministry of Science and ICT (2020M3H2A1078119). This work was partly supported by an Institute of Information and communications Technology Planning and Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2021-0-00944, Metamorphic approach of unstructured validation/verification for analyzing binary code, 30%) and (No. 2022-0-00816, OpenAPI-based hw/sw platform for edge devices and cloud server, integrated with the on-demand code streaming engine powered by AI, 20%) and (No. 2022-0-01170, PIM Semiconductor Design Research Center, 10%). The EDA tool was supported by the IC Design Education Center (IDEC), Korea.

Competing Interests

The authors declare that they have no competing interests.

Author Biography

Name : Seungmin Lee
Affiliation : School of Electronic and Electrical Engineering, Kyungpook National University, Daegu, Korea
Biography : He received the B.S. and M.S. degrees in mathematics and the Ph.D. degree in electronics engineering from Kyungpook National University (KNU) in 2010, 2012, and 2018, respectively. He expanded his research topics to bioinspired signal processing algorithms and electronics systems. He holds a postdoctoral position in KNU. His research interests include signal processing, image processing, bioinspired signal processing, and compact system implementation. He is focusing his research on the bio-signal processing as the research director of project supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education.

Name : Daejin Park
Affiliation : School of Electronic and Electrical Engineering, Kyungpook National University, Daegu, Korea
Biography : He received the B.S. degree in electronics engineering from Kyungpook National University (KNU) in 2001, the M.S. degree and Ph.D. degree in electrical engineering from the Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea, in 2003, and 2014, respectively. He was a Research Engineer in SK Hynix Semiconductor, Samsung Electronics over 12 years, respectively and have worked on designing low-power embedded processors architecture and implementing fully AI-integrated system-on-chip. Prof. Park is now with School of Electronic and Electrical Engineering as full-time associate professor in KNU, since 2014. He has published over 180 technical papers and 40 patents.

References

[1] H. Alshammari, S. A. El-Ghany, and A. Shehab, “Big IoT healthcare data analytics framework based on fog and cloud computing,” Journal of Information Processing Systems, vol. 16, no. 6, pp. 1238-1249, 2020. https://doi.org/10.3745/JIPS.04.0193
[2] A. R. Javed, M. U. Sarwar, M. O. Beg, M. Asim, T. Baker, and H. Tawfik, “A collaborative healthcare framework for shared healthcare plan with ambient intelligence,” Human-centric Computing and Information Sciences, vol. 10, article no. 40, 2020. https://doi.org/10.1186/s13673-020-00245-7
[3] S. A. Khowaja, B. N. Yahya, and S. L. Lee, “CAPHAR: context-aware personalized human activity recognition using associative learning in smart environments,” Human-centric Computing and Information Sciences, vol. 10, article no. 35, 2020. https://doi.org/10.1186/s13673-020-00240-y
[4] G. C. Yang, “Next-generation personal authentication scheme based on EEG signal and deep learning,” Journal of Information Processing Systems, vol. 16, no. 5, pp. 1034-1047, 2020. https://doi.org/10.3745/JIPS.03.0147
[5] H. Ozkan, O. Ozhan, Y. Karadana, M. Gulcu, S. Macit, and F. Husain, “A portable wearable tele-ECG monitoring system,” IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 1, pp. 173-182, 2020. https://doi.org/10.1109/TIM.2019.2895484
[6] S. Lee, D. Park, and K. H. Park, “QRS complex detection based on primitive,” Journal of Communications and Networks, vol. 19, no. 5, pp. 442-450, 2017. https://doi.org/10.1109/JCN.2017.000076
[7] S. Lee and D. Park, “Improved dynamic programming in local linear approximation based on a template in a lightweight ECG signal-processing edge device,” Journal of Information Processing Systems, vol. 18, no. 1, pp. 97-114, 2022. https://doi.org/10.3745/JIPS.03.0173
[8] S. Lee and D. Park, “A real-time abnormal beat detection method using template cluster for ECG diagnosis on IoT devices,” Human-centric Computing and Information Sciences, vol. 11, article no. 4, 2021. https://doi.org/10.22967/HCIS.2021.11.004
[9] S. Lee and D. Park, “Efficient template cluster generation for real-time abnormal beat detection in lightweight embedded ECG acquisition devices,” IEEE Access, vol. 9, pp. 70596-70605, 2021. https://doi.org/10.1109/ACCESS.2021.3077628
[10] S. Lee, Y. Jeong, J. Kwak, D. Park, and K. H. Park, “Efficient communication overhead reduction using polygonal approximation-based ECG signal compression,” in Proceedings of 2019 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), Okinawa, Japan, 2019, pp. 58-61. https://doi.org/10.1109/ICAIIC.2019.8668974
[11] S. Lee, Y. Jeong, J. Kwak, D. Park, and K. H. Park, “Advanced real-time dynamic programming in the polygonal approximation of ECG signals for a lightweight embedded device,” IEEE Access, vol. 7, pp. 162850-162861, 2019. https://doi.org/10.1109/ACCESS.2019.2952399
[12] S. Lee and D. Park, “Enhanced dynamic programming for polygonal approximation of ECG signals,” in Proceedings of 2020 IEEE 2nd Global Conference on Life Sciences and Technologies (LifeTech), Kyoto, Japan, 2020, pp. 121-122. https://doi.org/10.1109/LifeTech48969.2020.1570620076
[13] S. Lee and D. Park, “Adaptive ECG signal compression method based on look-ahead linear approximation for ultra long-term operating of healthcare IoT devices,” Human-centric Computing and Information Sciences, vol. 11, article no. 30, 2021. https://doi.org/10.22967/HCIS.2021.11.030
[14] H. Ling, S. Soatto, N. Ramanathan, and D. W. Jacobs, “A study of face recognition as people age,” in Proceedings of 2007 IEEE 11th International Conference on Computer Vision, Rio de Janeiro, Brazil, 2007, pp. 1-8. https://doi.org/10.1109/ICCV.2007.4409069
[15] K. Alkass, B. A. Buchholz, S. Ohtani, T. Yamamoto, H. Druid, and K. L. Spalding, “Age estimation in forensic sciences: application of combined aspartic acid racemization and radiocarbon analysis,” Molecular and Cellular Proteomics, vol. 9, no. 5, pp. 1022-1030, 2010. https://doi.org/10.1074/mcp.M900525-MCP200
[16] H. Han, C. Otto, and A. K. Jain, “Age estimation from face images: Human vs. machine performance,” in Proceedings of 2013 International Conference on Biometrics (ICB), Madrid, Spain, 2013, pp. 1-8. https://doi.org/10.1109/ICB.2013.6613022
[17] M. K. Kim, “Epileptic seizure detection using CNN ensemble models based on overlapping segments of EEG signals,” KIPS Transactions on Software and Data Engineering, vol. 10, no. 12, pp. 587-594, 2021. https://doi.org/10.3745/KTSDE.2021.10.12.587
[18] H. O. Song, Y. Xiang, S. Jegelka, and S. Savarese, “Deep metric learning via lifted structured feature embedding,” 2015 [Online]. Available: https://arxiv.org/abs/1511.06452.
[19] Z. Cheng, S. Wang, P. Zhang, S. Wang, X. Liu, and E. Zhu, “Improved autoencoder for unsupervised anomaly detection,” International Journal of Intelligent Systems, vol. 36, no. 12, pp. 7103-7125, 2021. https://doi.org/10.1002/int.22582
[20] F. S. Abousaleh, T. Lim, W. H. Cheng, N. H. Yu, M. A. Hossain, and M. F. Alhamid, “A novel comparative deep learning framework for facial age estimation,” EURASIP Journal on Image and Video Processing, vol. 2016, article no. 47, 2016. https://doi.org/10.1186/s13640-016-0151-4
[21] Y. Jeong, S. Lee, D. Park, and K. H. Park, “Accurate age estimation using multi-task Siamese network-based deep metric learning for frontal face images,” Symmetry, vol. 10, no. 9, article no. 385, 2018. https://doi.org/10.3390/sym10090385
[22] A. E. Ba Alawi and A. Y. A. Saeed, “Facial age estimation using convolution neural networks,” in Proceedings of 2021 International Congress of Advanced Technology and Engineering (ICOTEN), Taiz, Yemen, 2021, pp. 1-4. https://doi.org/10.1109/ICOTEN52080.2021.9493508
[23] R. J. Huszar, Basic Dysrhythmias: Interpretation and Management. Maryland Heights, MO: Mosby, 2007.
[24] G. B. Moody and R. G. Mark, “The MIT-BIH arrhythmia database on CD-ROM and software for use with it,” in Proceedings Computers in Cardiology, Chicago, IL, 1990, pp. 185-188. https://doi.org/10.1109/CIC.1990.144205
[25] A. Li, S. Wang, H. Zheng, L. Ji, and J. Wu, “A novel abnormal ECG beats detection method,” in Proceedings of 2010 the 2nd International Conference on Computer and Automation Engineering (ICCAE), Singapore, 2010, pp. 47-51. https://doi.org/10.1109/ICCAE.2010.5452002
[26] X. Hua, J. Han, C. Zhao, H. Tang, Z. He, Q. Chen, S. Tang, J. Tang, and W. Zhou, “A novel method for ECG signal classification via one-dimensional convolutional neural network,” Multimedia Systems, vol. 28, pp. 1387-1399, 2022. https://doi.org/10.1007/s00530-020-00713-1.5452002
[27] C. Sarvan and N. Ozkurt, “ECG beat arrhythmia classification by using 1-D CNN in case of class imbalance,” in Proceedings of 2019 Medical Technologies Congress (TIPTEKNO), Izmir, Turkey, 2019, pp. 1-4. https://doi.org/10.1109/TIPTEKNO.2019.8895014
[28] S. L. Oh, E. Y. K. Ng, R. S. Tan, and U. R. Acharya, “Automated beat-wise arrhythmia diagnosis using modified U-Net on extended electrocardiographic recordings with heterogeneous arrhythmia types,” Computers in Biology and Medicine, vol. 105, pp. 92-101, 2019. https://doi.org/10.1016/j.compbiomed.2018.12.012
[29] J. Chen, A. Valehi, and A. Razi, “Smart heart monitoring: early prediction of heart problems through predictive analysis of ECG signals,” IEEE Access, vol. 7, pp. 120831-120839, 2019. https://doi.org/10.1109/ACCESS.2019.2937875
[30] K. He, G. Zhong, X. Ding, and C. Yang, “Recognition of premature ventricular contraction beat from 12Lead ECG based on a novel detection function of QRS onset,” in Proceedings of 2020 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Montreal, Canada, 2020, pp. 349-352. https://doi.org/10.1109/embc44109.2020.9175775
[31] P. Krishnan, V. Rajagopalan, and B. I. Morshed, “ANovel severity index of heart disease from beat-wise analysis of ECG using fuzzy logic for smart-health,” in Proceedings of 2020 IEEE International Conference on Consumer Electronics (ICCE), Las Vegas, NV, 2020, pp. 1-5. https://doi.org/10.1109/ICCE46568.2020.9042984
[32] J. Malik, Z. Loring, J. P. Piccini, and H. T. Wu, “Interpretable morphological features for efficient single-lead automatic ventricular ectopy detection,” Journal of Electrocardiology, vol. 65, pp. 55-63, 2021. https://doi.org/10.1016/j.jelectrocard.2020.11.014
[33] L. H. Wang, L. J. Ding, C. X. Xie, S. Y. Jiang, I. C. Kuo, X. K. Wang, J. Gao, P. C. Huang, and P. A. R. Abu, “Automated classification model with OTSU and CNN method for premature ventricular contraction detection,” IEEE Access, vol. 9, pp. 156581-156591, 2021. https://doi.org/10.1109/ACCESS.2021.3128736

Seungmin Lee and Daejin Park*, Comparative Neural Network Based on Template Cluster for Automated Abnormal Beat Detection in Electrocardiogram Signals, Article number: 12:51 (2022) Cite this article 2 Accesses