KR102605130B1 - Method for classifying cardiac arrhythmia from standard 12-lead Electrocardiogram signal using artificial neural network and classification device using the same - Google Patents

Abstract

The method of classifying cardiac arrhythmia from a standard 12-lead electrocardiogram signal is to pass the electrocardiogram signal of each of the 12 channels obtained through a standard 12-lead electrocardiograph through a band-pass filter of a predetermined range, detect the R peak, and An R-R peak-based ECG sequence generation step to generate an ECG sequence for each channel by grouping the ECG signal to include three QRS complex waves as a standard, and the entire ECG sequence to extract an ECG sequence containing abnormal signs of single cardiac arrhythmia A K-means clustering step that extracts the representative signal of each channel while dividing it into a predetermined number of groups through K-means clustering, and the representative signal of each channel generated through the K-means clustering step is divided into a predetermined number of groups using STFT (Short time fourier A two-dimensional image extraction step through STFT (Short time Fourier transform), which extracts a two-dimensional frequency map form containing frequency change information over time of a standard 12-lead ECG signal using transform, and a two-dimensional image extraction step. An artificial neural network to classify eight types of cardiac arrhythmia diseases, including atrial fibrillation, type 1 atrioventricular block, atrial premature contraction, ventricular premature contraction, left bundle branch block, right bundle branch block, ST segment elevation, and ST segment depression, using 2D frequency feature images extracted through an artificial neural network. It includes a cardiac arrhythmia classification step using an artificial neural network model that classifies cardiac arrhythmia by using it as input to the model.

Description

Method for classifying cardiac arrhythmia from standard 12-lead Electrocardiogram signal using artificial neural network and classification device using the same}

The present invention relates to a method for classifying cardiac arrhythmia from an electrocardiogram signal, and more specifically, to a method for classifying cardiac arrhythmia from a standard 12-lead electrocardiogram signal using an artificial neural network and a classification device using the same.

An algorithm for diagnosing cardiac arrhythmia from electrocardiogram signals is proposed. These conventional algorithms classify cardiac arrhythmias by extracting time-frequency features from single lead ECG signals. However, since some cardiac arrhythmias are observed only in specific ECG channels, classification accuracy varies depending on the type of cardiac arrhythmia to be classified. can be reduced.

Therefore, in order to accurately diagnose cardiac arrhythmia, it is necessary to comprehensively and closely check the ECG signals acquired from 12 channels. In other words, the standard 12-lead ECG signal has the characteristic of changing the shape of the signal due to changes in electrical and mechanical activity due to various heart diseases, so it is used as an indicator to diagnose cardiac arrhythmia. However, the signs observed in the ECG signal differ depending on the type of cardiac arrhythmia. Unlike persistent cardiac arrhythmia, in which symptoms occur continuously and abnormal signs can be observed continuously in the ECG signal, single cardiac arrhythmia occurs intermittently. In this case, it is difficult to determine whether there is a cardiac arrhythmia with the naked eye from the ECG signal because abnormal signs are observed mixed between normal ECG signals. Additionally, because the ECG signals measured through the standard 12 leads are different depending on the type of arrhythmia, it is necessary to comprehensively analyze the ECG signals obtained from 12 channels in order to accurately determine the patient's cardiac arrhythmia.

In addition, conventional studies did not distinguish between persistent cardiac arrhythmia and single cardiac arrhythmia, but focused on the characteristics of the electrocardiogram itself or classified arrhythmias through a feature engineering process for diagnosis of specific arrhythmias. Since persistent cardiac arrhythmia and single cardiac arrhythmia have different characteristics in ECG signals, it is necessary to extract features considering each characteristic in order to accurately diagnose cardiac arrhythmia.

S. Datta et al., “Identifying normal, AF and other abnormal ECG rhythms using a cascaded binary classifier,” Comput. Cardiol. (2010)., vol. 44, pp. 1-4, 2017, doi: 10.22489/CinC.2017.173-154.

The present invention was proposed to solve the technical problems described above. Atrial fibrillation, type 1 atrioventricular block, atrial premature contraction, and ventricular premature contraction are detected from electrocardiogram signals (standard 12-lead electrocardiogram signals) obtained from 12 channels using an artificial neural network. Provides a method for classifying eight types of cardiac arrhythmia diseases, including constriction, left bundle branch block, right bundle branch block, ST segment elevation, and ST segment depression, and a classification device using the same.

According to an embodiment of the present invention to solve the above problem, the ECG signals of each of the 12 channels obtained through a standard 12-lead electrocardiograph are passed through a band-pass filter of a predetermined range, the R peak is detected, and the detected R An R-R peak-based ECG sequence generation step to generate an ECG sequence for each channel by grouping the ECG signal to include three QRS complexes based on the peak, and an overall ECG sequence to extract the ECG sequence containing abnormal signs of single cardiac arrhythmia. A K-means clustering step that extracts representative signals of each channel while dividing the ECG sequence into a predetermined number of groups through K-means clustering, and STFT (Short Transformation) of the representative signals of each channel generated through the K-means clustering step. A two-dimensional image extraction step through STFT (Short time Fourier transform), which extracts a two-dimensional frequency map form containing frequency change information over time of a standard 12-lead ECG signal using time fourier transform, and two-dimensional image extraction The two-dimensional frequency feature images extracted through the steps are used to classify eight cardiac arrhythmia diseases: atrial fibrillation, type 1 atrioventricular block, atrial premature contraction, ventricular premature contraction, left bundle branch block, right bundle branch block, ST segment elevation, and ST segment depression. A method for classifying cardiac arrhythmias from a standard 12-lead electrocardiogram signal is provided, including a cardiac arrhythmia classification step using an artificial neural network model to classify cardiac arrhythmias using the artificial neural network model as input.

In addition, in the R-R peak-based ECG sequence generation step included in the present invention, the R peak of each ECG signal is extracted by combining the Pan and Tomkins algorithm and the robust thresholding algorithm.

In addition, in the artificial neural network model of the cardiac arrhythmia classification stage using the artificial neural network model included in the present invention, the 2D feature image is extracted as an artificial intelligence feature while passing through a 2D Convolutional Neural Network layer. Ultimately, it is characterized as being classified as a cardiac arrhythmia disease.

As described above, through the present invention, it is possible to classify cardiac arrhythmia diseases from standard 12-lead electrocardiogram signals, and some arrhythmia signs that occur intermittently rather than continuously and are difficult to determine with the naked eye can be accurately identified through an artificial neural network model. It is possible to classify. In addition, by analyzing the standard 12-lead electrocardiogram signals of patients expected to be suffering from cardiac arrhythmia using the algorithm of the present invention, it can be used as a process for early diagnosis before undergoing detailed examination. The present invention can be used in a medical center using a standard 12-lead electrocardiograph to confirm the presence or absence of heart disease and as a process of early diagnosis before receiving a detailed diagnosis from doctors and experts.

1 is a flowchart of a method for classifying cardiac arrhythmia from a standard 12-lead electrocardiogram signal and a flowchart of a clustering algorithm according to an embodiment of the present invention.
Figure 2 is an example of an RR peak-based sequence signal
Figure 3 is a diagram of a time-frequency feature map and an example of a time-frequency feature map generated by the proposed algorithm.
Figure 4 is a diagram showing the confusion matrix for the proposed model.
Figure 5 is a diagram showing the performance curve for the proposed model

Hereinafter, in order to explain the present invention in detail so that a person skilled in the art can easily implement the technical idea of the present invention, embodiments of the present invention will be described with reference to the accompanying drawings.

- 8 types of cardiac arrhythmias observed on a standard 12-lead electrocardiogram

Electrocardiograms are widely used to diagnose cardiac arrhythmias because the shape of the signal changes due to changes in the electrical and mechanical activity of the heart due to various heart diseases. The standard 12-lead electrocardiogram shows the heart's electrical activity in four directions: inferior (lead II, III, aVF), anterior (lead I, aVL, V1, V2), septal (V3, V4), and lateral (V5, V6). The type of signal in each channel varies depending on the type of cardiac arrhythmia, and abnormal signs are sometimes observed only in specific channels among the 12 channels. Among these cardiac arrhythmias, some are persistent symptoms and abnormal signs are observed continuously on the electrocardiogram, while others are intermittent symptoms and abnormal signs are observed mixed with the normal electrocardiogram signal. Therefore, it is not easy to determine cardiac arrhythmia by looking at the ECG signal with the naked eye, and in order to accurately diagnose cardiac arrhythmia, it is necessary to closely and comprehensively observe the ECG signal acquired from 12 channels.

- Electrocardiogram signals due to persistent cardiac arrhythmia; AF, IAV-B, LBBB, RBBB

In the present invention, cardiac arrhythmias in which abnormal signs are continuously observed in electrocardiogram signals are defined as 'persistent cardiac arrhythmias'. Representative persistent cardiac arrhythmias include atrial fibrillation (AF), first-degree atrioventricular block (I-AVB), left bundle branch block (LBBB), and right bundle branch block (RBBB). In the ECG signal of AF patients, an intact QRS wave is observed, but due to multiple irregular electrical stimulation occurring in the atrium, P waves do not appear and irregularly oscillating fibrillatory waves are observed. Additionally, in the case of I-AVB, the PR interval of the electrocardiogram is consistently observed to be longer than 200 ms compared to a normal heart. In the case of LBBB and RBBB, abnormal signs are continuously observed, but the signs are not observed in all 12 channels. In the case of LBBB, an M-shaped signal is often observed in the V6 channel, and in the case of RBBB, abnormal signs are observed in the V1 channel. However, it is not easy to distinguish abnormal signs of persistent cardiac arrhythmia due to motion noise during the acquisition of ECG signals, noise due to contact with electrodes, and noise from the measuring device itself.

- ECG signal due to single cardiac arrhythmia; STD, STE, PAC, PVC

Meanwhile, unlike continuous cardiac arrhythmia, cardiac arrhythmia that occurs intermittently between normal signals was defined as 'episodic cardiac arrhythmia'. ST segment elevation (STE) and depression (STD), which are mainly caused by myocardial ischemia or infarction, are diagnosed when the height difference (slope of the ST segment) between the S peak and PR segment of the electrocardiogram is more than ±0.5 mm and are normal. There are features that are observed intermittently between electrocardiogram waveforms. STD, in which the S peak is approximately 3 mm lower than the PR segment, is generally observed in leads I, V5, V6, and aVL, and normal ECG signals may appear in V2-V3. On the other hand, in the case of STE, the S peak is observed to be about 4 mm higher than the PR segment, and in the ischemic ST segment, the ECG signal in V2-V4 is almost normal, but in the case of the non-ischemic ST segment, abnormalities are observed in V2-V3. Signs may also be observed. Premature atrial contraction (PAC) occurs when the tissue in the atrium temporarily becomes excited beyond the threshold faster than the sinus node. It is a symptom that 80% of people often experience once in their daily lives. P waves are not observed in the ECG signal of patients with premature ventricular contraction (PVC), and the heart rate is increased compared to a normal heart, and the QRS width can be as long as 120 to 210 ms (the QRS width of a normal heart is 60~80ms). However, these isolated cardiac arrhythmias are not easy to diagnose accurately because they are measured intermittently along with a normal electrocardiogram.

The present invention proposes an algorithm for automatically extracting abnormal signs due to single cardiac arrhythmias, including persistent cardiac arrhythmias, from electrocardiogram signals obtained from a standard 12-lead electrocardiograph in order to accurately determine a patient's cardiac arrhythmia, and a cardiac arrhythmia classification device using the same. . In other words, the cardiac arrhythmia classification device receives the electrocardiogram signal from a standard 12-lead electrocardiograph, processes the signal with the proposed algorithm, and classifies the arrhythmia.

In addition, eight cardiac arrhythmias - atrial fibrillation (AF), first-degree atrioventricular block (I-AVB), left bundle branch block (LBBB), and right bundle branch block (Right bundle branch block) RBBB), atrial premature contraction (Premature atrial complex, PAC), premature ventricular contraction (Premature ventricular complex, PVC), ST-segment depression (STD), and ST-segment elevation (STE)- An artificial neural network model for classification was proposed.

Figure 1 is a flowchart showing a flowchart of a method for classifying cardiac arrhythmia from a standard 12-lead electrocardiogram signal and a clustering algorithm according to an embodiment of the present invention.

Referring to Figure 1, eight types of cardiac arrhythmia diseases are classified through the following steps. First, frequency signals between 8 and 16 Hz are extracted from the ECG signal obtained through a standard 12-lead electrocardiograph using a band-pass filter, and then sampled to have two complete ECG signals. Depending on the embodiment, a bandpass filter between 0.5 and 20 Hz may be applied to remove fluctuation noise and minimize data loss.

For this purpose, the R peak is detected based on the “Robust thresholding” algorithm and then extracted from the first detected R peak to the fourth detected R peak (3 cycles of PQRST signal). Among the extracted signals, the K-means clustering algorithm is used to find representative signals that are judged to have signs of cardiac arrhythmia. One representative signal is extracted by repeating the process of grouping n signals extracted through the K-means clustering algorithm into two partitions.

The 12 signals finally detected in each of the 12 channels are converted into two-dimensional images through short time Fourier transform (STFT). The converted two-dimensional image contains information on the frequency change over time of the standard 12-lead electrocardiogram signal, and also includes information on specific cardiac arrhythmia, so it is classified as one of eight types of cardiac arrhythmia through an artificial neural network model.

Hereinafter, we will look in detail at the method of classifying cardiac arrhythmia from standard 12-lead electrocardiogram signals using an artificial neural network and the cardiac arrhythmia classification device using the same.

First, the data used was the standard 12-lead ECG signal of 6,877 people (men: 3,699, women: 3,178) released in the Computing in Cardiology 2020 Physionet challenge. The ECG signal data used includes normal sinus rhythm without any cardiac arrhythmia, ECG signals from patients with AF, ECG from patients with I-AVB, ECG from patients with LBBB and RBBB, and patients with PAC and PVC. Includes electrocardiogram, STD and STE patient electrocardiogram. Each signal was sampled every 500 Hz, and each ECG signal was measured for a minimum of 6 seconds and a maximum of 60 seconds.

A band-pass filter was applied to remove signal noise and baseline fluctuation noise from the original ECG signal, which included operation sounds and signal noise during measurements using the ECG device. In the present invention, a band-pass filter between 8 and 16 Hz or 0.5 and 20 Hz was applied to remove fluctuation noise and minimize data loss.

- R-R peak based sampling step (S10)

Let's take a closer look at the R-R peak-based sampling step (S10), that is, the R-R peak-based ECG sequence generation step.

In the present invention, the R peak is detected to classify eight types of cardiac arrhythmia by extracting the features of the sequence signal that includes both the change in heart rate due to persistent cardiac arrhythmia and the shape information of the ECG signal, and then based on the detected R peak. The ECG signals were grouped (generated ECG sequence) to include three QRS complexes. The R peak of each ECG signal was extracted by combining the Pan and Tomkins algorithm and the robust thresholding algorithm.

In other words, the Pan and Tompkins algorithm is used to extract the R peak from the 8 to 16 Hz signal extracted through a band-pass filter. The slope information of the signal is extracted by differentiating the filtered signal and then goes through a normalization process (d[n]). To increase the amplitude of the R peak signal, the normalized signal is squared (d ² [n]), and then the R peak is detected based on the “Robust thresholding” algorithm.

The “Robust thresholding” algorithm detects R peak by replacing it with 0 when the signal energy is less than the threshold parameter (η). At this time, the threshold value is adjusted according to each ECG segment.

At this time, the problem of squared signals distorting the length of the QS interval of the ECG occurs. To prevent this, Shannon energy transformation is used to improve the detection accuracy of the QRS wave. Shannon energy is calculated by normalizing the previously obtained threshold energy signal.

The Shannon energy signal is trimmed through a moving average filter, and then positive R peak and negative R peak are detected through a first-order Gaussian filter.

Based on the R peak detected in the entire signal, the ECG signal is sampled (ECG sequence signal generation) from the first detected R peak (r1) to the fourth detected R peak (r4) so that the two ECG cycles are maintained intact. . Accordingly, the ECG sequence signal includes information on the shape of the PQRST waveform of the ECG and information on the heart rate calculated as the reciprocal of the R-R interval.

Figure 2 is an example diagram of an RR peak-based sequence signal, which is an example of an electrocardiogram signal sampled through this algorithm. Depending on the type of ventricular arrhythmia and patient characteristics, the sampled ECG signals do not all have the same number of samples. Therefore, to make each sample have the same number of sample data, linear interpolation is used to augment the short sample data based on the longest sample data.

- K-means clustering (cluster analysis) step (S20)

Signs of ECG signals due to cardiac arrhythmia may be observed continuously or may be observed temporarily between normal ECG signals. Therefore, in order to extract sample data containing abnormal signs due to cardiac arrhythmia, the ECG data is divided into several groups through K-means clustering. At this time, the process of dividing the entire signal into two groups and dividing the smaller clusters into two groups again is repeated until finally only one sample data signal remains. K-means clustering is performed for each of the 12 channels, and representative signals are extracted from lead I, lead II, lead III, aVR, aVF, aVL, V1, V2, V3, V4, V5, and V6 channels.

For example, the clustering algorithm for extracting representative signals containing shape information due to single cardiac arrhythmia from grouped signals is shown in the workflow diagram in Figure 1(b). First, the continuous signal group containing the three PQRST signals is grouped into two groups (A and B). In order to find a group of signals with abnormal signs due to single cardiac arrhythmia, if the ratio of the small group to the large group is less than 0.35, one signal is randomly selected from the small group of signals, and if the ratio of the small group to the large group is less than 0.35, one signal is randomly selected from the large group. One signal is randomly selected from the group of signals. If the number of divided groups is the same, clustering is performed once more in each group and grouped into four groups (SA1, SA2, SB1, SB2), and one signal is randomly selected from the group with the smallest number among the four groups. pull out The extracted signal was used as a representative signal (S#) containing abnormalities due to single cardiac arrhythmia and continuous cardiac arrhythmia.

- 2D image extraction step through STFT (S30)

The signals extracted from each channel are extracted as a two-dimensional array (in the form of a two-dimensional frequency map) through STFT (Short time Fourier transform) analysis to check the change in frequency over time. At this time, linear interpolation or zero padding algorithm is used to equalize the length of sample data extracted from each channel through K-means clustering.

Figure 3 is a diagram of a time-frequency feature map and an example of a time-frequency feature map generated by the proposed algorithm.

Referring to FIG. 3, the signals extracted through STFT from each channel are extracted only from the 0 to 50 Hz band, which is the main frequency band of the ECG signal, and then added to the bottom as shown in FIG. 3(a) to form one image. Figures 3(a) and 3(b) are examples of time-dependent frequency feature images obtained from this algorithm. The finally extracted frequency feature image includes information on the change in frequency over time in each channel and the change in signal frequency band due to a specific cardiac arrhythmia.

In other words, the representative signals extracted from each of the 12 channels are extracted as a time-dependent frequency map through local Fourier transform (Short-time Fourier transform, STFT), and because the ECG signal contains frequency signals between 8 and 16 Hz, the STFT In the results, only the frequency band information between 0 and 50 Hz was used (frequency characteristics due to AF were observed in the 2.5-25 Hz range), and the time-frequency maps extracted from 12 channels were sequentially used (Figure 3a). In the present invention, the frequency window was set to 300 Hz to observe the change in frequency over time with a frequency resolution of 1.667 Hz, and the overlap ratio was set to 270 Hz to minimize the decrease in time resolution due to the increase in frequency resolution. Because the length of the R-R grouped signal varies depending on the type of cardiac arrhythmia, the time-frequency map extracted as a result of STFT was converted to a 120 × 120 time-frequency map using linear interpolation. Examples of time-frequency maps extracted through the algorithm used in the present invention are shown in Figures 3(b) and 3(c), which were used as input for a two-dimensional convolution artificial intelligence model.

- Cardiac arrhythmia classification step using artificial neural network model (S40)

The frequency feature image extracted through the above-mentioned steps shows eight cardiac arrhythmia diseases: atrial fibrillation, type 1 atrioventricular block, atrial premature contraction, ventricular premature contraction, left bundle branch block, right bundle branch block, ST segment elevation, and ST segment depression. It is used as input to an artificial neural network model for classification. In the artificial neural network model, the 2D feature image passes through a 2D convolutional neural network layer, is extracted as an artificial intelligence feature, and is finally classified as a cardiac arrhythmia disease. This series of processes is divided into several It is possible to classify cardiac arrhythmia more accurately through learning, and the classification results can be confirmed relatively faster than checking cardiac arrhythmia by visually comparing 12 channel electrocardiograms.

In other words, the CNN model for classifying cardiac arrhythmia consists of three 2D CNN layers, a BatchNormaliza-tion (BN) layer, and a 2D Maxpooling (2DMaxP) layer (2DCNN-BN-2DMaxP). A dropout rate of 0.5 was applied to the first and second 2DCNN-BN-2DMaxP layers, respectively, and a dropout rate of 0.3 was applied to the last 2DCNN-BN-2DMaxP layer. Kernel normalization and bias normalization with loss_lambda of 0.00015 were applied to each 2D CNN layer to prevent the model from overfitting the training data. The activation function of each layer used ReLU, and the activation function of the output layer used softmax.

To learn the model, 80% of the total data was used as training data (5,484 items) and 20% was used as test data (1,372 items). The model was trained 500 times using a sparse categorical crossentropy loss function and the Adam optimization function with a learning rate of 0.01, and the optimal model was verified using 5-fold cross-validation. The classification performance of the final model was evaluated through accuracy, recall, precision, F1 score, confusion matrix, ROC curve, and precision-recall curve.

Figure 4 is a diagram showing the confusion matrix for the proposed model, and Figure 5 is a diagram showing the performance curve for the proposed model.

Referring to Figures 4 and 5, the classification results of eight types of cardiac arrhythmia using the proposed algorithm are shown. The proposed model showed an F1 score of over 0.8 in the classification of AF and persistent cardiac arrhythmias such as I-AVB, LBBB, and RBBB. The F1 score was highest in LBBB at 0.89, and the classification F1 scores for AF and RBBB were 0.86 and 0.85, respectively. The F1 score of I-AVB was 0.80. On the other hand, the classification performance of single cardiac arrhythmias such as PAC, PVC, and STE was relatively low. In particular, the classification of STE, which had the smallest number of data classes, had the lowest F1 score of 0.52 (the F1 score of 0.53 for PAC and PVC, respectively). and 0.64). In addition, the classification of Normal and STD showed moderate classification performance with F1 scores of 0.77 and 0.76. There was a difference in the number of data used for each class. Accordingly, the final macro F1 score of the proposed model was 0.74, but the weighted F1 score considering the difference in the number of data for each class was 0.78.

To quantitatively evaluate the classification performance of the proposed model, two performance curves were used to compare (Figure 5). Receiver operating characteristics (ROC) curves show the true positive rate according to the false positive rate of classified data, and the classification accuracy of each class can be confirmed through the area under the curve (AUC). The AUCs of the persistent cardiac arrhythmia groups such as AF, I-AVB, LBBB, and RBBB were 0.90, 0.90, 0.91, and 0.90, respectively, and the AUCs of Normal and STD were 0.87 and 0.89. However, the AUC of the single cardiac arrhythmia group such as PAC, PVC, and STE showed moderate accuracy at 0.72, 0.79, and 0.78 (Figure 5a - Receiver operating characteristics (ROC) curves -).

The number of cardiac arrhythmia data used in model learning and classification differed depending on each class. Accordingly, considering the difference in the number of data for each class, the classification performance of the model was evaluated through the Precision-Recall curve. The resulting classification accuracy showed excellent performance with AUCs of over 0.90 for AF, LBBB, and RBBB among the persistent cardiac arrhythmia groups (AUCs; 0.93 for AF, 0.94 for I-AVB, and 0.93 for RBBB), and I-AVB and Normal , STD showed a classification accuracy of over 0.80 (AUCs; 0.88 for I-AVB, 0.85 for Normal, and 0.82 for STD). However, as in the ROC curve, the AUC of PAC, PVC, and STE showed low classification accuracy at 0.52, 0.68, and 0.55, respectively (Figure 5b - Precision-recall curves - ).

As described above, the proposed method for classifying cardiac arrhythmias from standard 12-lead electrocardiogram signals is:

First, a band-pass filter of 8-16 Hz is applied to remove meaningless noise signals and baseline wandering noise of the ECG signal from the ECG signal measured using a standard 12-lead electrocardiograph, and to extract only meaningful ECG signals. The filtered time series signal is divided into a certain data length to learn an artificial intelligence algorithm using short-term ECG signals and derive cardiac arrhythmia classification results. At this time, the R peak is detected using the Pan and Tompkins algorithm and the robust thres holding algorithm so that the divided ECG signal includes the PQRST waveform information and heart rate information of the intact ECG, and starts from the first detected R peak. The signal is divided up to the fourth detected R peak to generate several ECG sequence signals. Meanwhile, as mentioned earlier, the form observed in the ECG signal varies depending on the type of cardiac arrhythmia disease. In particular, in the case of event cardiac arrhythmia disease in which symptoms occur intermittently, abnormal signals are observed only in a portion of the entire time series signal. . Therefore, the K-means algorithm is used to extract only ECG signal sequences with abnormal signs of single cardiac arrhythmia diseases, including abnormal signs of persistent cardiac arrhythmia diseases whose symptoms can be continuously observed in the ECG signal, and sequence signals with abnormal signs. The normal sequence signal is divided, and the ECG sequence that is judged to have abnormalities is selected as the representative signal of a specific ECG channel. The 12 signals finally selected from each of the 12 channels are converted into a two-dimensional image containing frequency change information over time through short rune Fourier transform (STFI) analysis. The converted 2D image is classified as one of eight cardiac arrhythmias through a 2D convolutional artificial intelligence model.

Through the present invention, it is possible to classify cardiac arrhythmia diseases from standard 12-lead electrocardiogram signals, and it is possible to accurately classify some arrhythmia signs that occur intermittently rather than continuously and are difficult to determine with the naked eye through an artificial neural network model. . In addition, by analyzing the standard 12-lead electrocardiogram signals of patients expected to be suffering from cardiac arrhythmia using the algorithm of the present invention, it can be used as a process for early diagnosis before undergoing detailed examination. The present invention can be used in a medical center using a standard 12-lead electrocardiograph to confirm the presence or absence of heart disease and as a process of early diagnosis before receiving a detailed diagnosis from doctors and experts.

As such, a person skilled in the art to which the present invention pertains will understand that the present invention can be implemented in other specific forms without changing its technical idea or essential features. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive. The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

Claims (4)

In the cardiac arrhythmia classification device that receives the electrocardiogram signals of each of the 12 channels obtained through a standard 12-lead electrocardiograph, processes the signals using a predetermined algorithm and classifies the arrhythmia, the algorithm includes:
After passing the ECG signals of each of the 12 channels obtained through a standard 12-lead electrocardiograph through a band-pass filter of a predetermined range, the R peak is detected, and the ECG signal is converted to include three QRS complexes based on the detected R peak. RR peak-based ECG sequence generation step of grouping and generating ECG sequences for each channel;
In order to extract an ECG sequence containing abnormal signs of single cardiac arrhythmia, the entire ECG sequence is divided into a predetermined number of groups through K-means clustering, and in extracting representative signals of each channel, the entire ECG sequence is divided through K-means clustering. A K-means clustering step of extracting representative signals containing abnormal signals by repeating K-means clustering hierarchically based on the number of signals belonging to two groups;
In extracting a two-dimensional frequency map containing frequency change information over time of a standard 12-lead ECG signal using STFT (Short time Fourier transform) for the representative signal of each channel generated through the K-means clustering step, , a two-dimensional image is extracted by applying STFT (Short time Fourier transform) to the ECG signal of each channel, and then only images in the 0 to 50 Hz band where abnormal signals of cardiac arrhythmia to be detected are observed are extracted, and the vertically stacked images are finalized. Two-dimensional image extraction step through STFT (Short time Fourier transform) set as a feature; and
The 2D frequency feature images extracted through the 2D image extraction step are used for eight types of cardiac arrhythmias: atrial fibrillation, type 1 atrioventricular block, atrial premature contraction, ventricular premature contraction, left bundle branch block, right bundle branch block, ST segment elevation, and ST segment depression. A cardiac arrhythmia classification device using an artificial neural network model to classify cardiac arrhythmia by using it as an input to an artificial neural network model for classifying diseases. A cardiac arrhythmia classification device characterized by processing.
According to paragraph 1,
In the RR peak-based ECG sequence generation step,
A cardiac arrhythmia classification device that extracts the R peak of each ECG signal by combining the Pan and Tomkins algorithm and the robust thresholding algorithm.
According to paragraph 1,
In the artificial neural network model of the cardiac arrhythmia classification stage using the artificial neural network model, the 2D feature image is extracted as an artificial intelligence feature while passing through a 2D convolutional neural network layer, and is finally classified as a cardiac arrhythmia disease. A cardiac arrhythmia classification device characterized by classification. delete