KR20230071493A - Arrhythmia Classification Method Using Neural Network and Method Thereof - Google Patents

Abstract

The present invention relates to an arrhythmia classification system and method using a neural network. ) and a long-term memory neural network (LSTM) are serially merged to analyze the plurality of QRS regions and then to an arrhythmia classification system and method using a neural network including a determination unit for determining the presence or absence of an arrhythmia.

Description

Arrhythmia Classification Method Using Neural Network and Method Thereof

The present invention relates to an arrhythmia classification system and method using a neural network, and more specifically, to arrhythmia using a hybrid model combining a convolutional neural network (CNN) and a long short-term memory (LSTM). It relates to an arrhythmia classification system and method using a classifiable neural network.

Arrhythmia is a condition in which the heart beats abnormally fast, slow, or irregularly. A normal heart rate is 60-100 beats per minute, a heart rate of less than 50 beats per minute is called 'bradycardia', and a heart rate of more than 100 beats per minute is called 'tachycardia'. Arrhythmia may cause symptoms immediately, but often disappears spontaneously after appearing intermittently, so if not detected and treated in a timely manner, it can cause stroke, cardiac arrest, etc., leading to death. Therefore, management through early diagnosis is of utmost importance.

In general, as a method of diagnosing arrhythmia, wavelet transform, statistical method, morphological method, etc. may be used using the ECG signal, but noise ( If signal noise) is included, performance degradation may occur in arrhythmia classification.

In addition, in recent years, with the development of computer performance, deep learning (DL) such as machine learning (ML), convolutional neural network (CNN), and recurrent neural network (RNN) are used to solve regression and classification problems. ) was proposed to classify arrhythmias. However, convolutional neural networks (CNNs) have a disadvantage in that they cannot reflect temporal characteristics, and recurrent neural networks (RNNs) can reflect temporal characteristics, but as learning progresses, a long-term dependency problem occurs in which learning ability is significantly reduced.

In determining arrhythmias, Related Document 1 relates to a method and apparatus for classifying cardiac arrhythmias, by segmenting an ECG bit area and converting each ECG bit area into a symbolic sum approximation (SAX) symbol to be an arrhythmia type. can be classified, but this classifies arrhythmias using conventional statistical and morphological methods. As mentioned above, when noise is included in the process of extracting the ECG signal features, the performance decline may occur.

KR 10-1977745

The present invention is to solve the above problems, and noise can be removed in the process of extracting electrocardiogram features, and both spatial and temporal features are reflected to enable more accurate arrhythmia determination than before. An object of the present invention is to obtain an arrhythmia classification system and method using a neural network that determines the presence or absence of an arrhythmia after analyzing the plurality of QRS regions using a hybrid model in which a memory neural network (LSTM) is serially merged.

In order to achieve the above object, the arrhythmia classification system using the neural network of the present invention includes a pre-processing unit for pre-processing after obtaining an ECG signal; an extraction unit extracting a plurality of QRS regions from the preprocessed ECG signal; and a determination unit that analyzes the plurality of QRS regions using a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged, and determines whether or not there is an arrhythmia.

In order to achieve the above object, the arrhythmia classification method using the neural network of the present invention includes a pre-processing step of pre-processing after obtaining an ECG signal by a pre-registration unit; an extraction step of extracting a plurality of QRS regions from the preprocessed ECG signal by an extraction unit; and a determination step of determining, by a determination unit, whether there is an arrhythmia after analyzing the plurality of QRS regions using a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged.

As described above, according to the present invention, a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged is used to analyze the plurality of QRS regions and then determine the presence or absence of arrhythmia, thereby determining the ECG characteristics. In the process of extraction, noise can be removed, and both spatial and temporal characteristics are reflected, so that arrhythmia can be determined more accurately than before.

1 is a configuration diagram of a hybrid model according to an embodiment of the present invention.
2 is a diagram showing a conventional convolutional neural network (CNN) (a) and a conventional long-term memory neural network (LSTM) (b).
3 is a diagram showing an ECG signal (a) and a QRS area (b) according to an embodiment of the present invention.
4 is a diagram showing five classes classified by AAMI, a medical device-related standard publisher, according to the shape of an ECG signal according to an embodiment of the present invention.
5 is a diagram showing learning accuracy (a) and loss (b) when using a conventional LSTM alone and learning accuracy (c) and loss (d) when using a hybrid model according to an embodiment of the present invention .

The terms used in this specification have been selected from general terms that are currently widely used as much as possible while considering the functions in the present invention, but these may vary depending on the intention of a person skilled in the art, precedent, or the emergence of new technologies. In addition, in a specific case, there is also a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, not simply the name of the term.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in this application, it should not be interpreted in an ideal or excessively formal meaning. don't

Arrhythmia classification system using neural network

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings. 1 is a configuration diagram of a hybrid model according to an embodiment of the present invention. 2 is a diagram showing a conventional convolutional neural network (CNN) (a) and a conventional long-term memory neural network (LSTM) (b). 3 is a diagram showing an ECG signal (a) and a QRS area (b) according to an embodiment of the present invention. 4 is a diagram showing five classes classified by AAMI, a medical device-related standard publisher, according to the shape of an ECG signal according to an embodiment of the present invention.

First, the arrhythmia classification system using the neural network of the present invention includes a pre-processing unit 100 that acquires and pre-processes an ECG signal, an extraction unit 200 that extracts a plurality of QRS regions from the pre-processed ECG signal, and a convolutional neural network (CNN). ) and a long-term memory neural network (LSTM) are serially merged using a hybrid model to analyze the plurality of QRS regions and then determine the presence or absence of arrhythmia (300).

More specifically, the ECG signal refers to an electrocardiogram recorded by amplifying the electrical activity of the heart. The pre-processing unit 100 cuts the ECG signal beat by beat and performs zero-padding to adjust the size of data without filling in a specific value so that the length of each beat can be equal. can be done

Next, the extraction unit 200 may detect R waves respectively for a plurality of beats having the same length obtained from the preprocessing unit 100 . This is indicated in (a) of FIG. 3 . The R wave is an upward spike wave among intraventricular excitation waves in the electrocardiogram, and may include an R region including the detected R wave, a Q region before the R region, and an S region after the R region. Here, in order to highlight the characteristics of capturing the spatial characteristics of the shape change more locally, the extraction unit 200 selects a point 0.194 seconds before the R wave in the Q area and a point 0.277 seconds after the R wave. It is most preferable to set it to the S area. Accordingly, the extractor 200 may include the characteristic waveform of the electrocardiogram including the QRS region. That is, as shown in (b) of FIG. 3 , any QRS region may be extracted from the extraction unit 200 .

Next, the hybrid model stored in the determination unit 300 is a convolutional neural network (CNN) followed by a long-term memory neural network (LSTM) to solve problems of each of the conventional convolutional neural network (CNN) or long-term memory neural network (LSTM). This is the serially merged model.

Referring to (a) of FIG. 2, in general, a convolutional neural network (CNN) repeatedly passes through a convolution layer and a pooling layer for an input value, and then outputs an output value through a fully connected layer. It can be output and converted into a probability value through the Softmax function. Here, the convolutional neural network (CNN) has the advantage of being able to learn a lot of information about input values as the number of convolutional layers increases, that is, as the layers deepen, but overlearning that adapts only to the local characteristics of the input values. (Overfitting) phenomenon occurs. In addition, since a convolutional neural network (CNN) is interested only in features common to information whose order is not important, it cannot process the order relation of input information in real time.

To solve this problem, the hybrid model includes n 1D Convolution layers as shown in FIG. 1, and each of the n 1D Convolution layers has a ReLU activation function and batch normalization ( It includes a normalization layer capable of batch normalization, and among the n 1D Convolution layers, the 1st to n-1 1D Convolution layers are capable of Max Pooling. A pooling layer is included, and the n-th 1D convolution layer does not include the pooling layer to provide sufficient information to the long-term memory neural network (LSTM). Here, n is a positive integer.

More specifically, the convolutional neural network (CNN) of the hybrid model may include the 1D convolution layer to extract feature vector values from the plurality of QRS regions, which are 1D data. Most preferably, eight 1D convolution layers are provided so that a convolution operation can be performed in each layer, and spatial features of the plurality of QRS regions can be extracted. Here, in the hybrid model of the present invention, it is most preferable that the kernel size when performing the convolution operation is set to 2, and the stride value for determining the step value moving in the convolution operation is set to 1.

In addition, each of the eight 1D convolution layers may include a ReLU activation function and a normalization layer capable of batch normalization. That is, activation and normalization processes may be performed before moving from an arbitrary 1D convolution layer to the next 1D convolution layer.

In addition, when the output value of a certain layer becomes the input value of the next layer, the first to seventh layers among the eight 1D Convolution layers can be used to effectively reduce the length of the input value and extract main features. The 1D convolution layer up to may include a pooling layer capable of max pooling. but. The last eighth 1D convolution layer does not include the pooling layer in order to provide sufficient information as an input value of the long-term memory neural network (LSTM), and thus the maximum pooling is applied. It is best not to.

Therefore, the convolutional neural network (CNN) of the hybrid model can extract spatial features by extracting feature vector values from the plurality of QRS regions using eight 1D convolution layers. In addition, the convolutional neural network (CNN) of the hybrid model provides the plurality of QRS regions having spatial and temporal characteristics as input values of the long-term memory neural network (LSTM) disposed after the last 1D convolution layer. There is a clear difference from the conventional simple long-term memory neural network (LSTM) in that it can do this.

Next, the detailed structure of the long-term memory neural network (LSTM) of the hybrid model may have the same structure as the conventional one as shown in (b) of FIG. 2. The long-term memory neural network (LSTM) is proposed to improve the problem of loss and explosion of the gradient that occurs during long-term learning in the conventional recurrent neural network (RNN) model, and includes a forget gate, an input gate, and an output Can contain gates. That is, information can be controlled using the three gates and the calculated value of each gate can be transmitted through the cell-state. The first gate, the forget gate, can determine how much to forget or remember past information using Equation 1 below.

는 0과 1사이의 값으로 출력할 수 있는 시그모이드(Sigmoid) 활성화 함수이다. Here, h _t-1 is the output value at the previous time point (t-1), x _t is the input value at the current time point (t), W _f is the weight of the forget gate, and b _f is the forget ( is the bias of the Forget gate,

is a sigmoid activation function that can output a value between 0 and 1.

Therefore, the Forget gate will forget more information as the output value (f _t ) output from [Equation 1] is closer to 0, and remember more information as it is closer to 1. represent all candidates. In addition, the output value f _t may be multiplied with the previous memory cell C _t-1 by element.

Next, the input gate is a gate that determines whether to store new information in the cell-state using the following [Equation 2].

는 0과 1사이의 값으로 출력할 수 있는 시그모이드(Sigmoid) 활성화 함수이다. Here, h _t-1 is the output value at the previous time point (t-1), x _t is the input value at the current time point (t), W _i is the weight of the Input gate, and b _i is the input ( Input) is the bias of the gate,

is a sigmoid activation function that can output a value between 0 and 1.

는 새로운 기억 셀이고. h_t-1은 이전 시점의 출력값이고, x_t는 현재 시점의 입력값이고, W_C는 상기 새로운 기억 셀의 가중치이고, b_c는 상기 새로운 기억 셀의 편향(bias)이다. 즉, 하이퍼볼릭 탄젠트(Hyperbolic tangent) 연산, 즉, -1과 1 사이의 값을 가지는 함수로 취하여 새로 추가되는 각 원소가 정보로써 얼마큼의 가치가 있는지를 판단할 수 있다.and

is the new memory cell. h _t-1 is an output value at a previous time point, x _t is an input value at a current time point, W _C is a weight of the new memory cell, and b _c is a bias of the new memory cell. That is, it is possible to determine how much value each newly added element is as information by taking it as a hyperbolic tangent operation, that is, a function having a value between -1 and 1.

In addition, the input gate is a cell updated based on the output value (f _t ) output from the forget gate using the following [Equation 3] and the values obtained from the [Equation 2] (C _t ) can be created.

는 새로운 기억 셀이다. Here, f _t is the output value output from [Equation 1], C _t-1 is the previous memory cell, i _t is the output value output from [Equation 2],

is a new memory cell.

)과 상기 [수학식 3]으로부터 출력된 갱신(update)된 셀(C_t)을 하기 [수학식 4]를 이용하여 Cell-state에 얼마나 반영시킬지 결정한 후 최종 출력값(h_t)이 생성될 수 있다.Next, the output gate is a new memory cell output from [Equation 2] (

) and the updated cell (C _t ) output from [Equation 3] is determined by using the following [Equation 4] to reflect the cell-state, and then the final output value (h _t ) can be generated there is.

[Equation 4]

는 0과 1사이의 값으로 출력할 수 있는 시그모이드(Sigmoid) 활성화 함수이다.Here, h _t-1 is the output value at the previous time point (t-1), x _t is the input value at the current time point (t), W _o is the weight of the Output gate, and b _o is the output ( Output is the bias of the gate,

is a sigmoid activation function that can output a value between 0 and 1.

And the Output gate is a hyperbolic tangent operation for the updated cell (C _t ) output from [Equation 3], that is, a function having a value between -1 and 1 After taking as , the final output value (h _t ) may be finally generated by multiplying the output value (o _t ) output from [Equation 4].

In other words, the long-term memory neural network (LSTM), unlike the conventional recurrent neural network (RNN), has a memory cell and may include three gates to utilize it. The memory cell passes only the addition and multiplication nodes during backpropagation, and since the addition node passes the gradient transmitted during backpropagation as an expectation, gradient loss does not occur, and since the multiplication node used in the memory cell is an element-by-element product rather than a matrix product, each time By using a new gate value, the effect of multiplication is not accumulated, so there is an effect that the loss of the gradient does not occur.

Next, the hybrid model converts the final output value (h _t ) output from the long-term memory neural network (LSTM) back to one-dimensional through a flatten layer disposed after the long-term memory neural network (LSTM) After that, it can be passed to the Fully Connected layer, and the output value output from the Fully Connected layer is the class with the highest probability value among the five classes through the Softmax function as the analysis result. can be derived.

Therefore, the hybrid model of the present invention extracts the spatial features of the plurality of QRS regions through the convolutional neural network (CNN) and then inputs the spatial features to the long-term memory neural network (LSTM) that can consider sequential features of time series data. By using it as data, there is a remarkable effect of extracting and analyzing all of the temporal and spatial features of the plurality of QRS regions.

Also, the determination unit 300 may determine three arbitrary QRS areas among the plurality of QRS areas extracted by the extraction unit 200 as one group. In addition, the determination unit 300 may determine that all QRS regions in the group correspond to normal heartbeats as normal, and may determine arrhythmia if at least one QRS region in the group does not correspond to normal heartbeats. there is.

In other words, the determination unit 300 determines that all analysis results in the group output through the hybrid model are normal when they are all analyzed in the normal heart rate class, and at least one analysis result in the group is normal heart rate. If it does not fall under the class, it can be judged to be arrhythmia.

The above five classes adopt the guidelines proposed by AAMI (Association for the Advancement of Medical Instrumentation), an organization that publishes standards related to medical devices, and may include N class, S class, V class, F class, and Q class. can

The N class has the shape of FIG. 4 (a) and has a normal heartbeat, and the S class has the shape of FIG. 4 (b) and is located in the atrium or atrioventricular node (supraventricular node) ) is a class with an ectopic beating, the V class is a class having an ectopic beating related to the ventricle in the shape of FIG. 4 (c), and the F class is a class having the shape of FIG. Consistent with the normal conduction that arrives, it is a class with fusion beats in which normal and ectopic beats appear as a complex, and the Q class has the shape of FIG. 4 (e) and is a pacemaker that is an artificial pacemaker. A state in which the heart rate is controlled by electrical stimulation of the heart: paced beat, unclassifiable beat, fusion of paced and normal beat beat) is a class that has an unknown beat.

For example, the determination unit 300 may determine that the corresponding ECG signal is normal without arrhythmia if the analysis results of all three QRS areas are in the N class when a certain group is analyzed. On the other hand, if one of the three QRS areas is N class, another is S class, and the other is F class, if the analysis result is derived, it can be determined that the ECG signal has an arrhythmia.

Therefore, according to the present invention, in the process of extracting electrocardiogram features by analyzing the plurality of QRS regions using a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged, by determining the presence or absence of arrhythmias, Since noise can be removed and analysis results are derived by reflecting both spatial and temporal characteristics, there is a remarkable effect of enabling more accurate arrhythmia determination than before.

In addition, the determination unit 300 derives a loss value of the hybrid model using a data set, and uses an Adam function that is an optimization function to update a weight in the hybrid model according to the loss value It may include a learning unit 310 to. Here, the weight is each weight for a plurality of neurons in the convolutional neural network (CNN) to process input values, the weight (W _f ) of the forget gate in the long-term memory neural network (LSTM), the input It may include the weight (W _i ) of the (Input) gate and the weight (W _o ) of the output (Output) gate.

Here, the data set is classified into a learning sample and a verification sample according to a preset ratio, and each sample is labeled as to whether it is normal or arrhythmia.

Further, the learning unit 310 may derive a loss value by comparing the analysis result, which is the class having the largest output value among the five classes mentioned above, with the actual correct answer of the corresponding sample. In addition, the learning unit 310 has a remarkable effect of improving the accuracy of the hybrid model by itself by updating the weight of the hybrid model according to the loss value using the Adam function.

Arrhythmia classification method using neural network

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings. It relates to a corrective vein classification method using a neural network of the present invention, and more specifically, a preprocessing step (S100) of acquiring and then preprocessing an ECG signal by the pretreatment unit 100, and preprocessing by the extraction unit 200 By the extraction step (S200) of extracting a plurality of QRS regions from the ECG signal and the determination unit 300, a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged is used to obtain the plurality of QRS regions. After analyzing the QRS region, a determination step (S300) of determining whether or not there is an arrhythmia is included.

More specifically, the preprocessing step (S100) is zero padding in which the size of data is adjusted without filling in a specific value so that the length of each beat can be equally matched after the ECG signal is cut by beat ( Zero-Padding) may be performed.

Next, in the extraction step (S200), each R wave may be detected for a plurality of beats having the same length obtained from the preprocessing step (S100). The R wave is an upward spike wave among intraventricular excitation waves in the electrocardiogram, and may include an R region including the detected R wave, a Q region before the R region, and an S region after the R region. Most preferably, referring to FIG. 3, in the extraction step (S200), a point 0.194 seconds before the location of the R wave may be set as a Q area and a point 0.277 seconds after the location of the R wave may be set as an S area.

Next, in the determining step (S300), three arbitrary QRS areas among the plurality of QRS areas extracted from the extraction unit 200 may be determined as one group. In the determination step (S300), if the analysis result derived using the hybrid model corresponds to a normal heartbeat for all QRS regions in the group, it is determined that the result is normal, and for at least one QRS region in the group If it does not correspond to a normal heartbeat, it can be judged as an arrhythmia.

For example, in the determination step (S300), when an arbitrary group is analyzed and all three QRS regions are analyzed as N-class, the corresponding ECG signal may be determined to be normal without arrhythmia. On the other hand, if one of the three QRS areas is N class, another is S class, and the other is F class, if the analysis result is derived, it can be determined that there is arrhythmia in the corresponding ECG signal.

Meanwhile, in the determining step (S300), the data set is used to derive the loss value of the hybrid model, and the Adam function, which is an optimization function, is used to update the weight in the hybrid model according to the loss value It may include a learning step (S310) to be. That is, in the learning step (S310), the loss value may be updated by comparing the class having the largest output value among the five classes mentioned above with the actual correct answer of the corresponding sample.

Here, the weight is each weight for a plurality of neurons in the convolutional neural network (CNN) to process input values, the weight (W _f ) of the forget gate in the long-term memory neural network (LSTM), the input The weight (W _i ) of the (Input) gate and the weight (W _o ) of the output (Output) gate may be included.

Also, the data set may be classified into a learning sample and a verification sample according to a predetermined ratio, and each sample is labeled as to whether it is normal or arrhythmia. In the learning step (S310), a loss value is derived a predetermined number of times, and the weight of the hybrid model according to the loss value may be repeatedly updated by using the Adam function. Therefore, the determining step (S300) has a remarkable effect that the classes to which the plurality of QRS regions correspond can be more accurately analyzed by using the updated hybrid model.

Therefore, according to the present invention, in the process of extracting electrocardiogram features by analyzing the plurality of QRS regions using a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged, by determining the presence or absence of arrhythmias, Noise can be removed, and both spatial and temporal characteristics are reflected, so there is a remarkable effect that arrhythmia can be determined more accurately than before.

Example 1 (performance comparison with conventional LSTM)

Experimental parameters and experimental conditions

In order to learn and verify the hybrid model according to an embodiment of the present invention, the data set classified into a training sample and a verification sample according to a ratio of 8:2 was used. At this time. Each sample is in a state of being labeled for normal or arrhythmia.

In learning a hybrid model equipped with eight 1D convolution layers according to an embodiment of the present invention, each layer has a convolution filter having parameters as shown in Table 1 below. ) through which the output value was derived. And, Binary Cross Entropy (BCE) used for binary classification was used as a loss function to derive a loss value. An Adam function was used as an optimization function to update the weight in the hybrid model through the loss value.

In other words, in the hybrid model, each neuron in the convolutional neural network (CNN) has a weight for processing an input value, a weight (W _f ) of the forget gate, and a weight (W _i ) of the input gate. ), and the weight (W _o ) of the output gate. In addition, each weight may be updated based on the loss value by using the Adam function. Here, the Adam function is a function devised to solve the problem of deriving an inaccurate optimization value by falling to a local minimum value according to an initial value in gradient descent, and is an algorithm that combines the advantages of RMS Prop and Momentum.

Convolution Filter Value 1st convolution filter 16 2nd convolution filter 16 3rd convolution filter 32 4th convolution filter 32 5th convolution filter 64 6th convolution filter 64 7th convolution filter 128 8th convolution filter 256

In addition, in this embodiment, a learning rate of 0.001 is applied. If the learning rate exceeds 0.001, it may not converge to the lowest value, and if it is less than 0.001, a lot of learning time is required.

In addition, in this embodiment, the batch size (Batch Size) was set to 64. Here, the batch size is a hyperparameter that determines whether to compare the results of actual learning samples after only ‘how many’ samples in the middle of the learning sample derive the analysis result. If the batch size is less than 64, the weight update of the long-term memory neural network (LSTM) may occur frequently, resulting in unstable learning. there is. Finally, the hyperparameter Epoch, which determines how many times to perform training for all training samples in the data set, is set to 10. That is, the learning of the hybrid model is performed 10 times for all training samples.

That is, the learning process through the learning sample aims to update the weights of the long-term memory neural network (LSTM), and the verification process through the verification sample aims to confirm the performance of the updated weights. In the example, weights are updated every epoch through learning samples, and the accuracy of the updated weights is confirmed through verification samples.

Experiment result 1

5 is a diagram showing learning accuracy (a) and loss (b) when using a conventional LSTM alone and learning accuracy (c) and loss (d) when using a hybrid model according to an embodiment of the present invention .

More specifically, (a) of FIG. 5 shows the accuracy (Accuracy) for epochs 1 to 10 of the training process (Training) through the learning sample in the conventional LSTM and the validation process (Validation) through the verification sample in the conventional LSTM ), and the accuracy of each process in the conventional LSTM converged to a value between 0.9 and 0.95. In addition, it was confirmed that the accuracy did not change at all over 10 epochs.

5(c) shows the accuracy of epochs 1 to 10 of the training process through training samples in the hybrid model according to an embodiment of the present invention and an embodiment of the present invention Accuracy for epochs 1 to 10 of the validation process (validation) through the validation sample in the hybrid model according to is displayed, Accuracy of the training process (Accuracy) in 10 epochs converged to 0.987 and the accuracy of the validation process converged to 0.983. That is, it was confirmed that the accuracy was higher than that of the conventional LSTM in all epochs of all processes. In addition, it was confirmed that the accuracy did not change at all over 10 epochs.

Next, (b) of FIG. 5 shows loss values for epochs 1 to 10 of the training process through training samples in conventional LSTM and validation process through verification samples in conventional LSTM. ), and the loss value of each process in the conventional LSTM converged to a value between 0.15 and 0.2. And it was confirmed that all loss values did not change over 10 epochs.

5(d) shows loss values for epochs 1 to 10 of the training process through training samples in the hybrid model according to an embodiment of the present invention and an embodiment of the present invention In the hybrid model according to the example, loss values for epochs 1 to 10 of validation through validation samples are displayed, and loss of training at 10 epochs ( The loss value converged to 0.035 and the loss value of the validation process converged to 0.038. That is, it was confirmed that the hybrid model according to an embodiment of the present invention has a loss value lower than that of the conventional LSTM in all epochs of all processes. And it was confirmed that all loss values did not change over 10 epochs.

Therefore, the QRS region, which is the input value of the hybrid model according to an embodiment of the present invention, has very distinct regional characteristics of the electrocardiogram, and the hybrid model according to one embodiment of the present invention has very distinct spatial characteristics of the QRS region, unlike the conventional LSTM. Since can be extracted, it has been proven that the hybrid model according to an embodiment of the present invention converges to a higher accuracy and lower loss value than the conventional LSTM.

Example 2 (Hybrid model performance verification)

Experimental parameters and experimental conditions

Experimental parameters and experimental conditions are the same as in Example 1. In this embodiment, performance evaluation was performed for class classification, which is an analysis result output from the hybrid model of the present invention. That is, the performance of the hybrid model learned through the learning sample in the data set was evaluated through the validation sample in the data set. The performance indicators include Accuracy, which represents the proportion of correctly classified positives and negatives among all analysis results, Precision, which represents the proportion of correctly classified positives among those classified as positives, and Correctly classified both positives and negatives. Among the classified analysis results, there is recall, which indicates the proportion of correctly classified into aspects, and finally, F1 score, which indicates the harmonic average of recall and precision.

Experiment result 2

Table 2 below is a table comparing the conventional LSTM and the hybrid model according to an embodiment of the present invention for each performance index. Referring to [Table 2], it can be seen that the hybrid model according to an embodiment of the present invention has a higher percentage than the conventional LSTM in all performance indicators.

Accuracy Precision Recall F1 Conventional LSTM 90.57% 89.20% 88.74% 88.81% hybrid model 92.3% 90.98% 92.20% 90.72%

Therefore, the QRS region, which is the input value of the hybrid model according to an embodiment of the present invention, has very distinct regional characteristics of the electrocardiogram, and the hybrid model according to one embodiment of the present invention has very distinct spatial characteristics of the QRS region, unlike the conventional LSTM. Since can be extracted, it has been proven that the hybrid model according to an embodiment of the present invention shows higher values than the conventional LSTM in all performance indicators.

Embodiments may be implemented by hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. When implemented in software, firmware, middleware or microcode, the program code or code segments that perform necessary tasks may be stored on a computer readable storage medium and executed by one or more processors.

And aspects of the subject matter described herein may be described in the general context of computer-executable instructions, such as a program module or component executed by a computer. Generally, program modules or components include routines, programs, objects, and data structures that perform particular tasks or implement particular data types. Aspects of the subject matter described herein may be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

As described above, although the embodiments have been described with limited examples and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques are performed in a different order than the described method, and/or the components of the system, structure, device, circuit, etc. described in are combined or combined in a different form than the described method, or in a different configuration. Appropriate results can be achieved even when substituted or substituted by elements or equivalents.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

100.. pre-processing unit
200.. extraction unit
300.. Judgment
310.. Department of Learning

Claims (5)

a pre-processing unit that pre-processes after obtaining the ECG signal;
an extraction unit extracting a plurality of QRS regions from the preprocessed ECG signal; and
An arrhythmia classification system using a neural network comprising: a determination unit that analyzes the plurality of QRS regions using a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged and determines whether or not there is an arrhythmia. According to claim 1,
The QRS area,
An R region including the detected R wave, a Q region before the R region, and an S region after the R region,
The judge,
3 QRS areas are determined as one group, and if all QRS areas in the group correspond to normal heartbeats, it is determined to be normal, and if at least one QRS area in the group does not correspond to normal heartbeats, arrhythmia is determined. Arrhythmia classification system using a neural network, characterized in that for determining that. According to claim 1,
The judge,
A learning unit for deriving a loss value of the hybrid model using a data set and updating a weight in the hybrid model according to the loss value using an Adam function, which is an optimization function;
The weight is
Each weight for the plurality of neurons in the convolutional neural network (CNN) to process input values, the weight (W _f ) of the forget gate in the long-term memory neural network (LSTM), the weight of the input gate ( An arrhythmia classification system using a neural network, characterized in that it includes W _i ) and a weight (W _o ) of an output gate. According to claim 1,
The hybrid model,
Including n 1D convolution layers,
Each of the n 1D convolution layers includes a ReLU activation function and a normalization layer capable of batch normalization,
Among the n 1D Convolution layers, the 1st to n−1th 1D Convolution layers include pooling layers capable of max pooling, and Arrhythmia classification system using a neural network, characterized in that the 1D convolution layer does not include the pooling layer to provide sufficient information to the long-term memory neural network (LSTM).
Here, n is a positive integer. a pre-processing step of acquiring and then pre-processing the ECG signal by the pre-treatment unit;
an extraction step of extracting a plurality of QRS regions from the preprocessed ECG signal by an extraction unit; and
Using a neural network comprising: a determination step of analyzing the plurality of QRS regions by a determination unit using a hybrid model in which a convolutional neural network (CNN) and a long-term memory neural network (LSTM) are serially merged and then determining whether or not there is an arrhythmia; Arrhythmia classification method.

Images