Arrhythmia risk grading detection method based on CNN-LSTM palindromic network structure

Abstract

The invention provides an arrhythmia risk grading detection method based on a CNN-LSTM palindromic network structure, and the method comprises the following steps: 1, integrating data of a plurality of electrocardiogram public data sets, extracting complete rhythm segments, and forming a rhythm data set which comprises 16 rhythms; 2, dividing the 16 rhythms in the step 1 into three types of rhythm sections according to criticality, and setting the three types of rhythm sections to be normal, need to be focused and criticality-real-time response levels respectively; and step 3, intercepting specific lengths of the three types of rhythm sections, dividing a training set and a test set by adopting a ten-fold cross validation method, and verifying the detection precision. According to the detection method provided by the invention, multi-level and multi-scale more comprehensive spatial-temporal characteristics can be obtained, and a more accurate prediction result can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and signal processing, and in particular to a method for grading and detecting arrhythmia risk based on a CNN-LSTM palindrome network structure. Background Art

[0002] In clinical practice, electrocardiogram is an important tool for diagnosing and monitoring cardiovascular diseases. Cardiovascular diseases are characterized by rapid onset, high mortality, and high recurrence rates. Therefore, ensuring that emergency cases are handled in a timely manner is a major challenge facing the medical system. However, due to limited medical resources, there is a widespread backlog of electrocardiograms, which not only puts a heavy burden on doctors to read the images, but also affects the timely response to critical cases. In response to the above challenges, the most effective measure is to formulate electrocardiogram grading standards, introduce artificial intelligence to assist diagnosis, and establish an electrocardiogram early warning system.

[0003] In the field of AI-assisted diagnosis, most studies aim to detect specific types of arrhythmias without grading them; a few studies aim to detect critical arrhythmias, which are divided into two levels of critical and non-critical. However, there are many types of arrhythmias with different degrees of criticality, so it is urgent to develop a grading standard that can not only take into account arrhythmia diseases of all critical degrees, but also achieve the optimal allocation of medical resources.

[0004] At present, the most commonly used ECG-assisted diagnosis CNN-LSTM models are mostly based on CNN and LSTM. CNN can effectively extract detailed information and capture local features by using convolution kernels for local perception. By using multiple convolution layers and pooling layers, abstract features can be extracted layer by layer from low to high levels. This hierarchical feature extraction method with spatial invariance can learn spatial features of different frequencies, but its ability to capture long-term dependencies is weak. LSTM can effectively learn the temporal correlation and trend in the sequence through its internal memory unit and gating mechanism (including input gate, forget gate and output gate), so it performs particularly well in processing long time series and long-term dependency tasks, but its spatial feature extraction ability is weak. ECG signal is essentially a time-series bioelectric signal with time dependency and specific waveform characteristics. Multi-level and comprehensive extraction of spatial (i.e. waveform morphology) and temporal (i.e. dynamic characteristics of signal changes over time) features in ECG signals is crucial to improving the performance of CNN-LSTM models in diagnosis, prediction and other applications. This comprehensive feature extraction strategy can more accurately capture the information in the ECG signal, thereby enhancing the generalization ability and accuracy of the CNN-LSTM model. Therefore, it is urgent to design a diagnostic prediction method that can fully and comprehensively learn ECG spatiotemporal features at multiple levels. Summary of the invention

[0005] The purpose of the present invention is to provide a method for grading the risk of arrhythmia based on a CNN-LSTM palindrome network structure in view of the deficiencies of the prior art. The specific scheme is as follows:

[0006] A method for grading arrhythmia risk based on a CNN-LSTM palindrome network structure, characterized by comprising the following steps:

[0007] Step 1: Integrate multiple ECG public data sets, extract complete rhythm segments, and form a rhythm data set, which contains 16 rhythms;

[0008] Step 2: Divide the 16 rhythms in step 1 into three categories according to their severity, and set them at three levels: normal, requiring special attention, and critical-real-time response;

[0009] The normal level includes normal sinus rhythm; the level requiring special attention includes atrial bigeminy, atrial fibrillation, atrial flutter, ventricular bigeminy, 2° conduction block, ventricular escape, atrioventricular node junction rhythm, paced rhythm, preexcitation syndrome, sinus bradycardia, supraventricular tachyarrhythmia, ventricular trigeminy; the critical-real-time response level includes ventricular flutter, ventricular tachycardia and ventricular fibrillation; the normal level indicates that the person under test is in good health; the level requiring special attention indicates that the person under test does not need immediate intervention, but needs special attention; the critical-real-time response level indicates that the person under test needs real-time response and immediate intervention;

[0010] Step 3: Cut the three categories of rhythm segments into specific lengths, and use the ten-fold cross-validation method to divide them into training sets and test sets to verify the detection accuracy.

[0011] Based on the above, the training set obtained in step 3 is input into the CNN-LSTM model of the palindrome structure; in the CNN-LSTM model, the data set is first input into the convolutional neural network CNN to extract spatial features, and the obtained spatial features are then input into the long short-term memory network LSTM to extract temporal features, and the obtained temporal features are again extracted through CNN and LSTM to extract spatial features and temporal features of different scales;

[0012] The extracted time features are input into CNN again to capture high-level information and extract the final feature map. The final features are input into the fully connected layer for feature integration and classification. If the accuracy of the CNN-LSTM model does not meet the requirements after testing, it is necessary to readjust the CNN-LSTM model parameters and repeat the above steps until the accuracy of the CNN-LSTM model meets the requirements.

[0013] Based on the above, the CNN-LSTM model structure is “CNN-LSTM-CNN-LSTM-CNN”.

[0014] Based on the above, the data set in step 3 is time series data, which contains both temporal feature information and spatial feature information, and the data structure is:

[0015]

[0016] Where n is the dimension, the frequency of ECG data is H, and the duration is t, then n = H*t; m is the amount of data in the data set, a ij Represents the amplitude of the i-th data at the j-th time point.

[0017] Based on the above, the core idea of ​​the convolutional neural network CNN is to extract features through convolution operations, including convolution layers and pooling layers; the convolution layer contains multiple convolution kernels, each element of the convolution kernel corresponds to a weight coefficient and a bias, and the area corresponding to the size of the convolution kernel in the input feature is called the receptive field. The convolution kernel is generally square and will regularly scan the input features when working. The input features are multiplied by matrix elements in the receptive field and summed, and then the bias is superimposed, as shown in the following formula, to obtain a pixel value in the output feature.

[0018]

[0019] Where b is the deviation, Z l and Z l+1 Represent the input and output of the l+1th convolutional layer, also known as the feature map, L l+1 Z l+1 The size of Z(i,j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, f, S 0 and p correspond to the size of the convolution kernel, the convolution step, and the number of padding, respectively. In order to better extract the nonlinear characteristics of the data, Z l+1 Introduce the Relu activation function for nonlinear processing.

[0020] The pooling operation is mainly used to reduce the feature dimension, reduce the amount of calculation, and increase the scale invariance and rotation invariance of the CNN-LSTM model. The present invention adopts the maximum pooling method. Where x is the input matrix, which is the result of the nonlinear processing using the activation function mentioned above, k is the size of the pooling kernel, and y is the output matrix, which means that in the pooling process, for each position (i, j) of the input matrix, the corresponding element of the output matrix is ​​generated by selecting the maximum value in the k*k area centered on the position, and the output is the input of the subsequent steps.

[0021] Based on the above, the long short-term memory network LSTM is a time recursive neural network suitable for processing and predicting important events with relatively long intervals and delays in time series; LSTM removes or adds information to the unit state, which is called a gate. Neurons have three gates, the forget gate, the input gate, and the output gate. t The feature information output by CNN in the above step is used as the input of LSTM. The forget gate is used to determine the state C at the previous moment. t-1 The degree of information retention, the output information is f t =σ(W f ·[h t-1 ,x t ]+b f ); the input gate is used to determine the current cell state C t The storage information of i t =σ(W i ·[h t-1 ,x t ]+b i ); the candidate memory cell state is g t =tanh(W c ·[h t-1 ,x t ]+b C ); the historical memory information at the current moment on the long-term memory chain is C t =f t *C t-1 +i t *g t ; The output gate determines the output information, o t =σ(W o [h t-1 ,x t ]+b o ); the output of the hidden layer at the current moment is h t =o t *tanh(C t ). Where C is the cell state, h is the hidden state, W is the corresponding weight coefficient matrix, b is the corresponding bias term, t represents time, x represents input, and σ represents the sigmoid activation function.

[0022] Based on the above, the loss function is used to measure the difference between the predicted value and the true value of the CNN-LSTM model. By minimizing the loss function, the CNN-LSTM model can better fit the training data and improve the prediction accuracy. The loss function formula is as follows:

[0023]

[0024] Where m is the number of samples, k is the number of classification categories, y represents the true label of the test sample, and y represents the predicted label of the test sample.

[0025] Based on the above, the accuracy Acc and Kappa coefficient kap are used to measure the performance of the CNN-LSTM model. The calculation of Acc and Kap is based on the confusion matrix, and the formula is as follows:

[0026]

[0027] Where TP represents the number of positive examples that are correctly detected as positive examples; FP represents the number of negative examples that are incorrectly detected as positive examples; TN represents the number of negative examples that are correctly detected as negative examples; FN represents the number of positive examples that are incorrectly detected as negative examples;

[0028]

[0029] where p o is the overall classification accuracy, that is, the number of correctly classified samples divided by the total number of samples, a 1 ,a 2 ,...,a C Represents the number of real samples of each class, b 1 ,b 2 ,...,b C Represents the number of predicted samples for each class, n represents the total number of samples, kap value is [-1,1], and in practical applications the value range is generally [0,1].

[0030] The present invention has outstanding substantive features and significant progress compared to the prior art. Specifically, the present invention has the following advantages:

[0031] Since CNN and LSTM each have their own advantages and limitations in ECG-assisted diagnosis, for example, CNN has significant advantages in local perception and spatial hierarchical feature extraction, while LSTM has its unique advantages in processing sequence data and long-term dependencies. Therefore, the CNN-LSTM model provided by the present invention can extract features from input data through the CNN convolution layer to achieve spatial feature learning, and then input the extracted feature sequence into the LSTM for further sequence modeling to achieve temporal feature learning, wherein the multiple "CNN-LSTM" layers in the palindrome structure can extract spatial features and temporal features of different levels and scales of data, thereby enhancing the model's ability to understand and predict data. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is an implementation flow chart of the present invention.

[0033] Figure 2 This is the CNN-LSTM structure diagram.

[0034] Figure 3This is a performance comparison chart of the present invention, a pure CNN model, and a pure LSTM model. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described in detail below through specific implementation methods.

[0036] Example

[0037] like Figure 1-3 As shown, the present invention provides a method for grading arrhythmia risk based on a CNN-LSTM palindrome network structure, which is characterized by comprising the following steps:

[0038] Step 1: Integrate multiple ECG public data sets, extract complete rhythm segments, and form a rhythm data set, which contains 16 rhythms;

[0039] Step 2: Divide the 16 rhythms in step 1 into three categories according to their severity, and set them at three levels: normal, requiring special attention, and critical-real-time response;

[0040] The normal level includes normal sinus rhythm; the level requiring special attention includes atrial bigeminy, atrial fibrillation, atrial flutter, ventricular bigeminy, 2° conduction block, ventricular escape, atrioventricular node junction rhythm, paced rhythm, preexcitation syndrome, sinus bradycardia, supraventricular tachyarrhythmia, ventricular trigeminy; the critical-real-time response level includes ventricular flutter, ventricular tachycardia and ventricular fibrillation; the normal level indicates that the person under test is in good health; the level requiring special attention indicates that the person under test does not need immediate intervention, but needs special attention; the critical-real-time response level indicates that the person under test needs real-time response and immediate intervention;

[0041] Step 3: Cut the three categories of rhythm segments into specific lengths, and use the ten-fold cross-validation method to divide them into training sets and test sets to verify the detection accuracy.

[0042] The training set obtained in step 3 is input into the CNN-LSTM model of the palindrome structure; in the CNN-LSTM model, the data set is first input into the convolutional neural network CNN to extract spatial features, and the obtained spatial features are then input into the long short-term memory network LSTM to extract temporal features, and the obtained temporal features are again extracted through CNN and LSTM to extract spatial features and temporal features of different scales; the extracted temporal features are input into CNN again to capture high-level information to extract the final feature map, and the final features are input into the fully connected layer for feature integration and classification; if the accuracy of the CNN-LSTM model does not meet the requirements after testing, it is necessary to readjust the CNN-LSTM model parameters and repeat the above steps until the accuracy of the CNN-LSTM model meets the requirements.

[0043] The above CNN-LSTM model structure is "CNN-LSTM-CNN-LSTM-CNN".

[0044] The data set in step 3 above is time series data, which contains both temporal feature information and spatial feature information. The data structure is:

[0045]

[0046] Where n is the dimension, the frequency of ECG data is H, and the duration is t, then n = H*t; m is the amount of data in the data set, where a ij Represents the amplitude of the i-th data at the j-th time point.

[0047] The core idea of ​​the above convolutional neural network (CNN) is to extract features through convolution operations, such as Figure 2 As shown, it includes a convolution layer and a pooling layer; the convolution layer contains multiple convolution kernels, each element of the convolution kernel corresponds to a weight coefficient and a bias, and the area corresponding to the size of the convolution kernel in the input feature is called the receptive field. The convolution kernel is generally square and will regularly scan the input features when working. The input features are multiplied by matrix elements in the receptive field and summed, and then the bias is superimposed, as shown in the following formula, to obtain a pixel value in the output feature.

[0048]

[0049] Where b is the deviation, Z l and Z l+1 Represent the input and output of the l+1th convolutional layer, also known as the feature map, L l+1 Z l+1 The size of Z(i,j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, f, S 0 and p correspond to the size of the convolution kernel, the convolution step, and the number of padding, respectively. In order to better extract the nonlinear characteristics of the data, Z l+1 Introduce the Relu activation function for nonlinear processing.

[0050]

[0051] The pooling operation is mainly used to reduce the feature dimension, reduce the amount of calculation, and increase the scale invariance and rotation invariance of the CNN-LSTM model. The present invention adopts the maximum pooling method. Where x is the input matrix, which is the result of the nonlinear processing using the activation function mentioned above, k is the size of the pooling kernel, and y is the output matrix, which means that in the pooling process, for each position (i, j) of the input matrix, the corresponding element of the output matrix is ​​generated by selecting the maximum value in the k*k area centered on the position, and the output is the input of the subsequent steps.

[0052] The above-mentioned long short-term memory network LSTM (Long Short Term Memory) is a time recurrent neural network, which is suitable for processing and predicting important events with relatively long intervals and delays in time series; Figure 2 As shown, LSTM removes or adds information to the cell state, which is called a gate. Neurons have three gates: forget gate, input gate, and output gate. t The feature information output by CNN in the above step is used as the input of LSTM. The forget gate is used to determine the state C at the previous moment. t-1 The degree of information retention, the output information is f t =σ(W f ·[h t-1 ,x t ]+b f ); the input gate is used to determine the current cell state C t The storage information of i t =σ(W i ·[h t-1 ,x t ]+b i ); the candidate memory cell state is g t =tanh(W c ·[h t-1 ,x t ]+b C ); the historical memory information at the current moment on the long-term memory chain is C t =f t *C t-1 +i t *g t ; The output gate determines the output information, o t =σ(W o [h t-1 ,x t ]+b o ); the output of the hidden layer at the current moment is h t =o t *tanh(C t ). Where C is the cell state, h is the hidden state, W is the corresponding weight coefficient matrix, b is the corresponding bias term, t represents time, x represents input, and σ represents the sigmoid activation function.

[0053] The loss function is used to measure the difference between the predicted value and the true value of the CNN-LSTM model. By minimizing the loss function, the CNN-LSTM model can better fit the training data and improve the prediction accuracy. The loss function formula is as follows:

[0054]

[0055] Where m is the number of samples, k is the number of classification categories, y represents the true label of the test sample, and y represents the predicted label of the test sample.

[0056] The accuracy Acc and Kappa coefficient kap are used to measure the performance of the CNN-LSTM model. The calculation of Acc and Kap is based on the confusion matrix. The formula is as follows:

[0057]

[0058] Where TP represents the number of positive examples that are correctly detected as positive examples; FP represents the number of negative examples that are incorrectly detected as positive examples; TN represents the number of negative examples that are correctly detected as negative examples; FN represents the number of positive examples that are incorrectly detected as negative examples;

[0059]

[0060] where p o is the overall classification accuracy, that is, the number of correctly classified samples divided by the total number of samples, a 1 ,a 2 ,...,a C Represents the number of real samples of each class, b 1 ,b 2 ,...,b C Represent the number of predicted samples for each class, n represents the total number of samples, and kap takes a value of [-1, 1]. It should be noted that in practical applications, the value range is generally [0, 1], and the higher the value, the higher the classification performance of the CNN-LSTM model.

[0061] From the test results, it can be seen that Figure 3 As shown in the figure, the arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure provided by the present invention has achieved good results on four data sets composed of three databases, and has a certain improvement compared with the pure CNN model and the pure LSTM model. It shows that the CNN-LSTM palindrome network structure model of the present invention can obtain more comprehensive spatiotemporal features at multiple levels and scales, and can obtain more accurate prediction results.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention, which should be included in the scope of the technical solution for protection of the present invention.

Claims

1. A method for grading arrhythmia risk based on a CNN-LSTM palindrome network structure, characterized in that The following steps are involved: Step 1: Integrate multiple ECG public data sets, extract complete rhythm segments, and form a rhythm data set, which contains 16 rhythms; Step 2: Divide the 16 rhythms in step 1 into three categories according to their severity, and set them at three levels: normal, requiring special attention, and critical-real-time response; The normal level includes normal sinus rhythm; the level requiring special attention includes atrial bigeminy, atrial fibrillation, atrial flutter, ventricular bigeminy, 2° conduction block, ventricular escape, atrioventricular node junction rhythm, paced rhythm, preexcitation syndrome, sinus bradycardia, supraventricular tachyarrhythmia, ventricular trigeminy; the critical-real-time response level includes ventricular flutter, ventricular tachycardia and ventricular fibrillation; the normal level indicates that the person under test is in good health; the level requiring special attention indicates that the person under test does not need immediate intervention, but needs special attention; the critical-real-time response level indicates that the person under test needs real-time response and immediate intervention; Step 3: Cut the three categories of rhythm segments into specific lengths, and use the ten-fold cross-validation method to divide them into training sets and test sets to verify the detection accuracy.

2. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 1 is characterized in that: The training set obtained in step 3 is input into the CNN-LSTM model of the palindrome structure; in the CNN-LSTM model, the data set is first input into the convolutional neural network CNN to extract spatial features, and the obtained spatial features are then input into the long short-term memory network LSTM to extract temporal features, and the obtained temporal features are again extracted through CNN and LSTM to extract spatial features and temporal features of different scales; The extracted time features are input into CNN again to capture high-level information and extract the final feature map. The final features are input into the fully connected layer for feature integration and classification. If the accuracy of the CNN-LSTM model does not meet the requirements after testing, it is necessary to readjust the CNN-LSTM model parameters and repeat the above steps until the accuracy of the CNN-LSTM model meets the requirements.

3. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 2 is characterized in that: The CNN-LSTM model structure is "CNN-LSTM-CNN-LSTM-CNN".

4. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 3 is characterized in that: The data set in step 3 is time series data, which contains both temporal feature information and spatial feature information, and the data structure is: Where n is the dimension, the frequency of ECG data is H, and the duration is t, then n = H*t; m is the amount of data in the data set, where a ij Represents the amplitude of the i-th data at the j-th time point.

5. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 4 is characterized in that: The core idea of ​​the convolutional neural network CNN is to extract features through convolution operations, including convolution layers and pooling layers; the convolution layer contains multiple convolution kernels, each element of the convolution kernel corresponds to a weight coefficient and a bias, and the area corresponding to the size of the convolution kernel in the input feature is called the receptive field. The convolution kernel is generally square and will regularly scan the input features when working. The input features are multiplied by matrix elements in the receptive field and summed, and then the bias is superimposed, as shown in the following formula, to obtain a pixel value in the output feature. (i,j)∈{0,1,...L l+1 }, Where b is the deviation, Z l and Z l+1 Represent the input and output of the l+1th convolutional layer, also known as the feature map, L l+1 Z l+1 The size of Z(i,j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, f, S0 and p correspond to the size of the convolution kernel, the convolution step and the number of padding, respectively. In order to better extract the nonlinear features of the data, Z l+1 Introduce the Relu activation function for nonlinear processing. The pooling operation is mainly used to reduce the feature dimension, reduce the amount of calculation, and increase the scale invariance and rotation invariance of the CNN-LSTM model. The present invention adopts the maximum pooling method. Where x is the input matrix, which is the result of the nonlinear processing using the activation function mentioned above, k is the size of the pooling kernel, and y is the output matrix, which means that in the pooling process, for each position (i, j) of the input matrix, the corresponding element of the output matrix is ​​generated by selecting the maximum value in the k*k area centered on the position, and the output is the input of the subsequent steps.

6. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 5 is characterized in that: The long short-term memory network LSTM is a time recurrent neural network suitable for processing and predicting important events with relatively long intervals and delays in time series; LSTM removes or adds information to the unit state, which is called a gate. Neurons have three gates, the forget gate, the input gate, and the output gate. t The feature information output by CNN in the above step is used as the input of LSTM. The forget gate is used to determine the state C at the previous moment. t-1 The degree of information retention, the output information is f t =σ(W f ·[h t-1 ,x t ]+b f ); the input gate is used to determine the current cell state C t The storage information of i t =σ(W i ·[h t-1, x t ]+b i ); the candidate memory cell state is g t =tanh(W c ·[h t-1 ,x t ]+b C ); the historical memory information at the current moment on the long-term memory chain is C t =f t *C t-1 +i t *g t ; The output gate determines the output information, o t =σ(W o [h t-1 ,x t ]+b o ); the output of the hidden layer at the current moment is h t =o t *tanh(C t ). Where C is the cell state, h is the hidden state, W is the corresponding weight coefficient matrix, b is the corresponding bias term, t represents time, x represents input, and σ represents the sigmoid activation function.

7. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 6 is characterized in that: The loss function is used to measure the difference between the predicted value and the true value of the CNN-LSTM model. By minimizing the loss function, the CNN-LSTM model can better fit the training data and improve the prediction accuracy. The loss function formula is as follows: Where m is the number of samples, k is the number of classification categories, y represents the true label of the test sample, and y represents the predicted label of the test sample.

8. The arrhythmia risk grading detection method based on the CNN-LSTM palindrome network structure according to claim 7 is characterized in that: The accuracy Acc and Kappa coefficient kap are used to measure the performance of the CNN-LSTM model. The calculation of Acc and Kap is based on the confusion matrix. The formula is as follows: Where TP represents the number of positive examples that are correctly detected as positive examples; FP represents the number of negative examples that are mistakenly detected as positive examples; TN represents the number of negative examples that are correctly detected as negative examples; FN represents the number of positive examples that are mistakenly detected as negative examples; where p o is the overall classification accuracy, that is, the number of correctly classified samples divided by the total number of samples, a1, a2, ..., a C Represents the number of real samples of each class, b1, b2, ..., b C Represents the number of predicted samples for each class, n represents the total number of samples, and the kap value is [-1,1].

Images