A significant value determination method for electrocardiosignal classification problem

Abstract

The application discloses a significant value determination method for electrocardiosignal classification problems, relates to the field of electrocardiosignal classification research, and is characterized in that after filtering and amplitude standardization are performed on an electrocardiosignal, the electrocardiosignal is divided into segments and input into a classification model, then the segment significant value is calculated, the heart beat level significant value result is intercepted and interpolated to a fixed length, and the category level significant value result is obtained through multiple average operations. The significant value determination method for electrocardiosignal classification problems disclosed by the application shows the importance of different regions of a heart beat to a classification model, and improves the interpretability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electrocardiosignal classification, and particularly to a significant value determination method for electrocardiosignal classification. BACKGROUND

[0002] Heart disease has the characteristics of multiple and frequent occurrence, and is one of the most common and most serious human health problems in the world. Electrocardiogram (ECG) records the bioelectric signal of the heart and is an important tool for diagnosing heart disease in the field of clinical medicine. The waveform of electrocardiogram contains information such as heart rhythm and amplitude, reflecting the physiological activity state of each part of the heart, which can assist doctors in diagnosing heart disease.

[0003] In recent years, deep learning-based methods have achieved good performance in electrocardiosignal classification. In order to improve the explainability of deep learning methods, scholars have proposed some significant value calculation methods to intuitively show the classification decision of the model. However, some existing researches do not comprehensively consider all electrocardiosignals, but only show the significant value of a certain segment of electrocardiosignal, which is not representative. Some other researches comprehensively consider all electrocardiosignals, but do not consider the P / Q / R / S / T region misplacement problem caused by RR interval fluctuation. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the present application provides a significant value determination method for electrocardiosignal classification, which effectively shows the importance of different regions of heart beat to the classification model and improves the explainability of the model.

[0005] The technical scheme adopted by the present application to solve its technical problems is:

[0006] A significant value determination method for electrocardiosignal classification, the method comprising the following steps:

[0007] Step 1: Collect electrocardiosignals of multiple individuals of different categories to construct an electrocardiosignal database;

[0008] Step 2: Filter the electrocardiosignal;

[0009] Step 3: Perform amplitude standardization processing on the electrocardiosignal;

[0010] Step 4: Perform R peak detection on the electrocardiosignal to calculate the average RR interval RR interval of the electrocardiosignal;

[0011] Step 5: Divide the electrocardiosignal into segments and input the classification model for classification;

[0012] Step 6: Calculate the segment significant value S c ;

[0013] Step 7: Extracting beat-level saliency results from segment saliency S c ;

[0014] Step 8: Interpolating beat-level saliency results to fixed length;

[0015] Step 9: Calculating individual-level saliency results;

[0016] Step 10: Calculating class-level saliency results.

[0017] Further, in the step 1, using an electrocardiogram signal acquisition device, electrocardiogram signals of multiple individuals of different categories are collected to construct an electrocardiogram signal database.

[0018] Further, in the step 2, using a filter to process the electrocardiogram signal to remove noise in the electrocardiogram signal, the noise in the electrocardiogram signal includes power frequency interference, baseline drift and electromyographic interference.

[0019] Further, in the step 3, the amplitude normalization method is Z-Score, and the calculation formula is as follows:

[0020]

[0021] Where X(t) is the original electrocardiogram signal amplitude, μ is the average value of the original electrocardiogram signal amplitude, σ is the standard deviation of the original electrocardiogram signal amplitude, and Z(t) is the amplitude normalized electrocardiogram signal, whose amplitude average value is 0 and standard deviation is 1.

[0022] In the step 4, the R-peak detection method is used to detect the R-peak position of the electrocardiogram signal, and then the average value of the RR interval, i.e. RR interval , is calculated.

[0023] In the step 5, the electrocardiogram signal is randomly divided into segments of fixed length, and then input into a classification model designed for electrocardiogram signal classification problem for classification; the classification model designed for electrocardiogram signal classification problem is a convolutional neural network model and its variants.

[0024] In the step 6, the segment saliency S c characterizes the importance of each point in the segment to the category c, and the calculation method of the segment saliency S c is as follows: inputting the segment to obtain a feature vector A, , the feature vector corresponding to the channel k is denoted as A k ; performing global pooling and full connection layer classification on the feature vector A to obtain the output y c corresponding to the category c; calculating y c , the feature vector A kThe gradient of , get the weight of channel k It represents the importance of the features of channel k to classification. The calculation formula is as follows:

[0025]

[0026] The weight of channel k and the feature vector A of channel k k Multiply, then sum channel by channel, and then process through the ReLU function to get W c , W c The calculation formula is as follows:

[0027]

[0028] W c Linearly interpolate to the same dimension as the input fragment, that is, obtain the fragment saliency value S c .

[0029] In step 7, the position of the R peak is taken as the center, and the segment significance value S is obtained. c The interception length is RR interval The significant value fragment of is called the beat-level significant value result.

[0030] In step 8, the fixed length is the average value of the RR interval lengths of all ECG signals in the ECG signal database.

[0031] In step 9, all heart rate-level significance value results of the same individual are averaged to obtain an individual-level significance value result.

[0032] In step 10, the individual-level significance value results of all individuals in the same category are averaged to obtain the category-level significance value result.

[0033] The beneficial effects of the present invention are mainly manifested in:

[0034] (1) The present invention provides a method for determining a saliency value for the electrocardiogram (ECG) signal classification problem. The method is highly versatile and applicable to various problems in the field of ECG signal classification.

[0035] (2) The significance value results obtained by the present invention calculate the average significance value results of multiple individuals in different categories, comprehensively considering the characteristics of all individuals, and are more convincing.

[0036] (3) The saliency value results obtained by the present invention show the importance of different areas of the heartbeat to the model classification, thereby improving the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Fig. 1A flowchart of a significant value determination method for electrocardiosignal classification problems;

[0038] Fig. 2 A structural diagram of a classification model;

[0039] Fig. 3 A class-level significant result. DETAILED DESCRIPTION

[0040] The application will be further described below with reference to the accompanying drawings.

[0041] Reference Figs. 1-3 A significant value determination method for electrocardiosignal classification problems, comprising the following steps:

[0042] Step 1: using an electrocardiosignal acquisition device, acquiring electrocardiosignals of multiple individuals of different categories, and constructing an electrocardiosignal database;

[0043] Step 2: using a filter to process the electrocardiosignal, removing noise such as power frequency interference, baseline drift and electromyographic interference in the electrocardiosignal;

[0044] Step 3: performing Z-Score amplitude standardization processing on the electrocardiosignal, and the calculation formula of Z-Score is as follows:

[0045]

[0046] wherein X(t) is the original electrocardiosignal amplitude, μ is the average value of the original electrocardiosignal amplitude, σ is the standard deviation of the original electrocardiosignal amplitude, Z(t) is the amplitude standardized electrocardiosignal, and the average value of the amplitude is 0 and the standard deviation is 1;

[0047] Step 4: using the R-peak detection method proposed by Pan-Tompkins, detecting the R-peak position of the electrocardiosignal, and then calculating the average value of the RR interval, i.e. RR interval ;

[0048] Step 5: randomly dividing the electrocardiosignal into 10-second segments, and then inputting them into the classification model shown in Fig. 2 The classification model comprises an initial convolutional neural network layer, four residual blocks, a global average pooling layer and a fully connected layer, uses a SoftMax activation function to calculate the output category probability; all convolutional layers in the network are one-dimensional convolution, the convolution kernel size is 11 and the step is 1; the size of the max pooling layer is 2 and the step is 2; the drop-out ratio is 0.5; the channel number of the first residual block is 32, and the channel number of each residual block is increased by 32 thereafter, finally reaching 128;

[0049] Step 6: calculating the segment significant value S cwhich represents the importance of each point in the segment to the class c, segment saliency S c The calculation process is as follows:

[0050] Step 6.1: The input segment is subjected to feature extraction to obtain a feature vector A, The feature vector corresponding to channel k is denoted as A k ;

[0051] Step 6.2: Global average pooling and fully connected layer classification are performed on the feature vector A to obtain the output y c corresponding to the class c;

[0052] Step 6.3: The gradient of y c with respect to the feature vector A k of channel k is calculated to obtain the weight of channel k which represents the importance of the features of channel k to the classification, The calculation formula is as follows:

[0053]

[0054] Step 6.4: The weight of channel k is multiplied by the feature vector A k of channel k, then summed channel by channel, and then processed through a ReLU function to obtain W c , the calculation formula of W c is as follows:

[0055]

[0056] Step 6.5: W c is linearly interpolated to the same dimension as the input segment, i.e., the segment saliency S c is obtained;

[0057] Step 7: The position of R peak is taken as the center, and a saliency segment with a length of RR c is cut from the segment saliency S interval , which is called the beat-level saliency result;

[0058] Step 8: The beat-level saliency result is interpolated to a length of 425, which is the average RR interval size of all individuals in the electrocardiogram signal database under a sampling frequency of 512 Hz;

[0059] Step 9: All beat-level saliency results of the same individual are averaged to obtain the individual-level saliency result, which shows the average importance of different positions in the individual's heartbeat to the classification problem;

[0060] Step 10: Average the individual-level saliency results of all individuals under the same category to obtain the category-level saliency results, which show the global importance of different positions in the heartbeat to the classification problem. The area with greater saliency is more important to the classification decision of the model. Fig. 3 An example of the category-level saliency result is shown. It can be seen that the saliency around the S peak is greater, indicating that the waveform features around the S peak are more important to the classification decision of the model.

[0061] The details of the present application are known. The above examples are only to illustrate the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made in accordance with the spirit and essence of the present application shall be covered within the protection scope of the present application.

Claims

1. A method for determining a significant value for an electrocardiogram signal classification problem, characterized in that: The method comprises the following steps: Step 1: Collect ECG signals of multiple individuals of different categories and build an ECG signal database; Step 2: Filter the ECG signal; Step 3: Perform amplitude normalization on the ECG signal. The amplitude normalization method is Z-Score, which is calculated as follows: ； in is the original ECG signal amplitude, is the average value of the original ECG signal amplitude, is the standard deviation of the original ECG signal amplitude, is the ECG signal after amplitude standardization, with an average amplitude of 0 and a standard deviation of 1; Step 4: Detect the R peak of the ECG signal and calculate the average RR interval of the ECG signal ; Step 5: Divide the ECG signal into segments and input them into the classification model for classification; Step 6: Calculate the segment saliency value , the fragment significance value Represents the categories of each point pair in the segment The importance of the fragment The calculation method is as follows: the input fragment is extracted to obtain the feature vector , , called channel The corresponding eigenvector is ; For the eigenvector Perform global pooling and fully connected layer classification to obtain the corresponding category Output ;calculate Channel The eigenvector of The gradient of the channel Weight , which characterizes the channel The importance of the features to classification, The calculation formula is as follows: ； The channel Weight With channel The eigenvector of Multiply, then sum channel by channel, and then process it through the ReLU function to get , The calculation formula is as follows: ； Will Linearly interpolate to the same dimension as the input fragment, that is, obtain the fragment saliency value ; Step 7: Obtain significant values ​​from the fragment The results of significant value of heart rate level in the middle interception; Step 8: Interpolate the heart rate level significance value results to a fixed length; Step 9: Calculate the individual-level significance value results; Step 10: Calculate the category-level significance value results.

2. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1, wherein: In step 1, an electrocardiogram signal acquisition device is used to acquire electrocardiogram signals of multiple individuals of different categories to construct an electrocardiogram signal database.

3. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1 or 2, wherein: In step 2, a filter is used to process the ECG signal to remove noise in the ECG signal, where the noise in the ECG signal includes power frequency interference, baseline drift, and myoelectric interference.

4. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1 or 2, wherein: In step 4, the R peak position of the ECG signal is detected using the R peak detection method, and then the average value of the RR interval is calculated, that is, .

5. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1 or 2, wherein: In step 5, the ECG signal is randomly divided into segments of fixed length, and then input into a classification model designed for the ECG signal classification problem for classification; the classification model designed for the ECG signal classification problem is a convolutional neural network model and its variants.

6. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1 or 2, wherein: In step 7, the R peak position is taken as the center, and the fragment significance value is The intercept length is The significant value fragment of is called the beat-level significant value result.

7. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1 or 2, wherein: In step 8, the fixed length is the average value of the RR interval lengths of all ECG signals in the ECG signal database.

8. The method for determining a saliency value for an electrocardiogram signal classification problem according to claim 1 or 2, wherein: In step 9, all heart rate level significance value results of the same individual are averaged to obtain the individual level significance value result; In step 10, the individual-level significance value results of all individuals in the same category are averaged to obtain the category-level significance value result.

Images