A myocardial infarction identification method based on position backtracking deep learning network

Abstract

The application belongs to the technical field of medical data recognition, and discloses a myocardial infarction recognition algorithm based on a position backtracking deep learning network. The algorithm proposes a feature selection method based on a position backtracking deep learning network PBDN (Position Backtracking Deep learning Network), adopts the PBDN to obtain a backtracking result of a deep learning model on original features, then maps the result to a slope of a first quadrant y=kx straight line, and uses a sum of slopes of different categories and a sum of angles between each two straight lines to complete final evaluation of the features. Simulation experiments are conducted on a public data set PTB electrocardiogram recognition database, and the results show that, using myocardial infarction data selected by the PBDN for classification, the accuracy rate reaches more than 99.5%, compared with other myocardial infarction recognition algorithms, the recognition rate of the algorithm reaches a satisfactory effect, and the training time and calculation amount are greatly reduced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of myocardial infarction recognition, and particularly relates to a myocardial infarction recognition algorithm based on a position backtracking deep learning network. BACKGROUND

[0002] Cardiovascular diseases have caused serious threats to the quality of human life and health. Although more and more people have begun to pay attention to cardiovascular health, the morbidity and mortality of such diseases are still high. More worrying is that cardiovascular diseases are often accompanied by multiple complications, which not only bring huge burdens to patients themselves, but also cause heavy life pressure to their families. Therefore, prevention and treatment of cardiovascular diseases still need to be solved urgently.

[0003] Electrocardiogram detection is one of the current routine detection methods for myocardial infarction diseases. In the evolution process of myocardial infarction, the electrocardiogram shows ST segment deviation and shape change, pathological Q wave formation, myocardial repolarization extension, and ischemic T wave change. The changes of electrocardiographic features are extremely diverse and complex. For clinical electrocardiogram data, doctors need deep professional knowledge and rich practical experience to accurately identify different types of abnormal electrocardiograms. At the same time, due to the large amount of clinical electrocardiogram data, even if experienced professionals are relied on for manual observation and analysis, the workload is quite large, the efficiency is greatly reduced, and the accuracy of analysis cannot be guaranteed. Therefore, there is an urgent need for a recognition method with fast detection speed and high accuracy.

[0004] Deep learning is a branch of machine learning, aiming to simulate the working way of human brain, and realizing complex pattern recognition and decision-making tasks through building and training multi-level neural networks. Combining deep learning with myocardial infarction recognition can make up for many shortcomings of traditional manual detection of electrocardiogram.

[0005] (1) High automation and accuracy

[0006] Deep learning models can learn features and patterns from a large amount of data, realizing automatic myocardial infarction detection and reducing the influence of human factors. And deep learning models have high accuracy and sensitivity, which can capture subtle features in medical images or signals, helping to more accurately identify myocardial infarction areas.

[0007] (2) Processing ability for complex data

[0008] Deep learning models can process various types of medical data, such as electrocardiogram, ultrasonic image, nuclear magnetic resonance image, etc., and can learn the correlation between different modalities, improving the effect of comprehensive diagnosis.

[0009] (3) Learning ability and generalization ability

[0010] Deep learning models have strong learning capabilities and can perform well on untrained data by training on large-scale data, improving the generalization ability of the model.

[0011] (4) Improve efficiency and speed

[0012] When assisting doctors in diagnosis, deep learning models can quickly process large amounts of data and provide rapid diagnostic results, helping to improve work efficiency and save time.

[0013] In the myocardial infarction recognition algorithm, convolutional neural networks (CNN) are commonly used for research. Acharya processes the original 12-lead ECG signal using discrete wavelet transform, and after QRS detection, the heartbeats are intercepted and imported into the CNN model. Finally, the SoftMax classifier is used to complete the detection and positioning of myocardial infarction. Xue Yanping uses a regression model for classification, and Wang Zhen uses a long short-term memory neural network (LSTM) for research. These studies have achieved satisfactory results in recognition rate.

[0014] The current myocardial infarction recognition algorithm combined with deep learning has the following shortcomings:

[0015] (1) The working mode of the deep learning model is difficult to understand. When processing data, it is based on multi-level neural networks, and the structure of these networks is very complex, containing a large number of parameters and levels, so the decision-making process of the model is not easy for humans to intuitively understand or explain. This lack of interpretability may make it difficult for deep learning to be combined with some fields, such as the medical field, which requires transparency and explainability.

[0016] (2) In recent years, the mobile medical field has developed rapidly, bringing a series of portable ECG detection devices, such as handheld, patch-type products, and even smartwatches have begun to venture into ECG monitoring. However, due to the special nature of cardiovascular diseases, which have a short effective treatment time and require strict recognition speed, the existing recognition algorithm requires a large number of feature dimensions and has high data processing difficulty, making it unsuitable for direct transplantation to portable devices.

[0017] To overcome the problems of combining deep learning with myocardial infarction recognition, the present application proposes a myocardial infarction algorithm based on a position backtracking deep learning network. SUMMARY

[0018] The purpose of the present application is to provide a myocardial infarction recognition algorithm based on a position backtracking deep learning network, which solves the problems raised in the background art.

[0019] To achieve the above object, the application provides the following technical scheme: a myocardial infarction recognition algorithm based on a position backtracking deep learning network, comprising the following steps:

[0020] S1, obtaining original data from a PTB electrocardiogram recognition data set, performing filtering, downsampling, R peak positioning and electrocardiogram interception preprocessing to obtain electrocardiogram data for training, wherein an innovative electrocardiogram data combination method is proposed for evaluation of electrocardiogram feature regions and leads, in the electrocardiogram record of 12 leads, 12 heartbeats with R peaks at the same time are selected to form a 2-dimensional electrocardiogram segment;

[0021] S2, using the feature selection method based on the position backtracking deep learning network PBDN proposed by the application to evaluate the electrocardiogram feature regions and leads of the electrocardiogram data, first, designing the corresponding deep learning network PBDN according to the evaluation dimension, then, dividing the electrocardiogram data into a training set and a test set, the training set is used to iterate the PBDN, after the model training is completed, the test set is put into the PBDN, when the data of different categories are put into the maximum pooling layer in the PBDN, the electrocardiogram feature regions and leads selected are counted, the more the count is, the more effective the recognition of myocardial infarction is, thereby realizing the preliminary evaluation of the original features, making up for the poor explainability of deep learning in the medical field, the electrocardiogram feature regions or leads with good performance should play an important role in the recognition of various categories, therefore, the count is mapped to the slope of the first quadrant y=kx, the sum of the slopes of the straight lines of different categories and the sum of the angles between each two straight lines are used to complete the final evaluation of the electrocardiogram feature regions or leads, wherein, in order to prevent the problem that the optimal electrocardiogram feature subset is discrete and is not conducive to observation by an analyst, the application innovatively divides the electrocardiogram features into continuous electrocardiogram feature regions, and evaluates the electrocardiogram feature regions, which not only retains the significance of feature selection, but also enhances the practical application value;

[0022] S3, obtaining electrocardiogram feature regions and leads with high scores based on the evaluation information of S2, respectively forming an optimal electrocardiogram feature region data set and an optimal lead data set and putting them into an SVM classifier for training to obtain the final recognition result of myocardial infarction.

[0023] As a preferred embodiment of the application, the filtering: the original ECG signal often contains noises such as baseline drift, power frequency interference and electromyographic interference, which can seriously affect the accuracy of the experimental results, in order to solve this problem, a 4th order Butterworth band-pass filter is used, which can effectively filter out the electrocardiogram signal noise below 250Hz and above 2000Hz, thereby improving the quality of the data and the reliability of the experimental results.

[0024] As a preferred embodiment of the present application, the down-sampling: the down-sampling helps to eliminate the redundant information in the signal, reduces the calculation cost, and improves the calculation efficiency.

[0025] As a preferred embodiment of the present application, the R peak positioning: the Pan-Tompkins algorithm is used to position the R peak in the electrocardiogram data set, and the two main operations of the Pan-Tompkins algorithm are R wave enhancement and R wave decision, the R wave enhancement is used to strengthen the characteristics of the R peak, and the R wave decision is used to select the R wave from the R wave and the noise.

[0026] As a preferred embodiment of the present application, the heart beat cutting: the position difference between the two R peaks before and after each lead is calculated, the maximum R peak interval Rmax and the minimum R peak interval Rmin in all leads are found, and according to the position of the R peak in the array, the (Rmax+Rmin) / 2 electrocardiogram data before the R peak and the (Rmax+Rmin) / 2 electrocardiogram data after the R peak are obtained to form a heart beat, and the same operation is performed for each R peak position and each lead.

[0027] As a preferred embodiment of the present application, the training data construction: in the electrocardiogram record of 12 leads, 12 heart beats with R peaks at the same time are selected to form a 2-dimensional electrocardiogram segment, finally, the preprocessed electrocardiogram data set contains 10080 normal electrocardiogram signals and 51402 myocardial infarction electrocardiogram signals, each electrocardiogram signal contains 12 leads of heart beats, and each heart beat contains 140 electrocardiogram features, i.e. 70 electrocardiogram features before the R peak and 70 electrocardiogram features after the R peak.

[0028] Compared with the prior art, the present application has the following advantages:

[0029] 1. The present application carries out simulation experiment on the public data set PTB electrocardiogram recognition database, and the results show that the accuracy rate reaches more than 99.5% using the myocardial infarction data after PBDN feature selection for classification, compared with other myocardial infarction recognition algorithms, the recognition rate of the present application reaches satisfactory effect, and the training time and calculation amount are greatly reduced. The experiment also proves that the PBDN feature selection method can not only evaluate the most effective electrocardiogram feature area for recognizing myocardial infarction, but also can trace back the position of the feature area in the original electrocardiogram signal, and improve the explainability of the deep learning network. In addition, the PBDN feature selection method can also evaluate the best electrocardiogram lead, and can use the best electrocardiogram lead to achieve better myocardial infarction recognition. BRIEF DESCRIPTION OF DRAWINGS

[0030] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the following drawings:

[0031] Figure 1 A flowchart of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application;

[0032] Figure 2 An electrocardiogram data preprocessing flowchart of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application;

[0033] Figure 3 A 12-lead electrocardiogram segment diagram of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application;

[0034] Figure 4 A backtracking flowchart (taking an evaluation electrocardiogram feature region as an example) of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application;

[0035] Figure 5 A No. 1 electrocardiogram feature region feature diagram of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application;

[0036] Figure 6 A No. 5, 6, 7 electrocardiogram feature region diagram of a PBDN feature selection method evaluation highest of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application;

[0037] Figure 7 A V2, V3, V4 lead diagram of a PBDN feature selection method evaluation highest of a myocardial infarction recognition algorithm based on a position backtracking deep learning network of the present application. DETAILED DESCRIPTION

[0038] A myocardial infarction recognition algorithm based on a position backtracking deep learning network, comprising the following steps: S1, obtaining original data from a PTB electrocardiogram recognition data set, performing filtering, downsampling, R peak positioning and electrocardiogram preprocessing, obtaining electrocardiogram data for training, wherein an electrocardiogram data combination method is innovatively proposed for evaluation of electrocardiogram feature regions and leads, in 12-lead electrocardiogram records, 12 heartbeats with R peaks at the same time are selected to form a 2-dimensional electrocardiogram segment;

[0039] S2, the feature selection method based on the position backtracking deep learning network PBDN proposed in the application is used to evaluate the electrocardio feature region and lead of the electrocardio data, first, a corresponding deep learning network PBDN is designed according to the dimension of the evaluation, then, the electrocardio data is divided into a training set and a test set, the training set is used for iterating the PBDN, after the model training is completed, the test set is put into the PBDN, when the data of different categories are put into the maximum pooling layer in the PBDN, the electrocardio feature region selected is counted, and the more the counted times are, the more effective the identification of the myocardial infarction is, so that the preliminary evaluation of the original feature is realized, the problem of poor explainability of deep learning in the medical field is made up, the electrocardio feature region or lead with good performance should play an important role in the identification of various categories, therefore, the counted times are mapped to the slope of the first quadrant y=kx, the sum of the slopes of the straight lines of different categories and the sum of the angles between each two straight lines are used to complete the final evaluation of the electrocardio feature region or lead, wherein, in order to prevent the problem that the optimal electrocardio feature subset is discrete and is not conducive to observation of the analyst, the electrocardio feature is innovatively divided into continuous electrocardio feature regions, the electrocardio feature regions are evaluated, the significance of feature selection is retained, and the practical application value is enhanced.

[0040] S3, the electrocardio feature region and lead with high evaluation information acquisition scores obtained based on S2 are respectively composed of an optimal electrocardio feature region data set and an optimal lead data set and are put into an SVM classifier for training, so that the final identification result of the myocardial infarction is obtained.

[0041] The application proposes a method of enhancing explainability and enabling the deep learning network to backtrack: a deep learning network PBDN specially used for evaluating original features is designed, the electrocardio data is divided into a training set and a test set, the training set is used for iterating the PBDN, after the training is completed, the test set is used to analyze the feature selection of the maximum pooling layer, and the backtracking of the electrocardio data is obtained.

[0042] The PBDN is composed of a convolution layer, a pooling layer and a fully connected layer. The PBDN can evaluate features of any dimension by skillfully designing the size, step and window size of the convolution kernel of the convolution layer and the maximum pooling layer, that is, a PBDN network can be generated for the features of each dimension. In actual situations, only a specific PBDN needs to be selected according to the need.

[0043] The PBDN is described in detail as follows:

[0044] X=(S1, S2, S3,..., S m

[0045] Wherein, X represents the structure of each data in the data set, S i ​Different dimensions (e.g. a color picture as input, containing 3 dimensions of height, width, and channel number), i = 1, 2, 3,..., m, m represents the number of dimensions.

[0046] (1) Input of the model:

[0047] In = (N, F1, F2, F3,..., Fm) (2.1) m ) (2.2)

[0048] where In represents the input of the deep learning model, N represents the number of samples, F i represents the number of features contained in different dimensions, i = 1, 2, 3,..., m.

[0049] (2) Design of convolutional layer:

[0050] First, some symbols are explained: S i , F i have the same meaning as formula (2.1-2); for any integer z, CvzD represents that this is a z-dimensional convolutional layer; kz represents the size, and sd represents the stride; represents the size and stride of the z-dimensional convolutional kernel in the PBDN used to evaluate the dimension S i .

[0051] When m = 1, the data set has only one dimension S1, and the number of features is F1. One PBDN can be constructed.

[0052]

[0053]

[0054] When m = 2, the data set has two dimensions S1, S2, and the number of features is F1, F2 respectively. Two PBDNs can be constructed.

[0055]

[0056]

[0057] When m = 3, the data set has three dimensions S1, S2, S3, and the number of features is F1, F2, F3 respectively. Three PBDNs can be constructed.

[0058]

[0059]

[0060] When m ≥ 3, the data set has m dimensions S1, S2, S3,..., S m , and the number of features is F1, F2, F3,..., Fm respectively.m m PBDNs can be constructed.

[0061]

[0062]

[0063] (3) Design of the pooling layer:

[0064] First, some symbols are explained: S i , F i have the same meaning as formula (2.1-2); ps represents the size of the window; represents the window size of the largest pooling layer in the PBDN for evaluating the dimension S i .

[0065] When m≥1, the data set has m dimensions S1, S2, S3,..., S m , and the number of features is F1, F2, F3,..., F m , respectively.

[0066]

[0067]

[0068]

[0069] The meaning of backtracking is to intuitively evaluate the features of the original data according to the training parameters of the deep learning network. Specifically, for the myocardial infarction recognition of the present application, the meaning of backtracking is to evaluate the features in the original electrocardiogram signal based on the training parameters of the PBDN, which solves the problem of weak interpretability of the deep learning network.

[0070] The specific process of backtracking will be described below taking the evaluation of the electrocardiogram feature region as an example, and the flowchart is shown in Figure 4 .

[0071] The structure of the electrocardiogram data after preprocessing of the PTB data set is:

[0072] In=(N,D,L) (2.12)

[0073] Where D represents the number of leads, L represents the number of electrocardiogram features, and the remaining symbols refer to formula (2.2). The electrocardiogram data contains two dimensions, the first dimension is the lead, and the second dimension is the electrocardiogram feature.

[0074] Since there are many ECG features selected before and after the R-peak, evaluating each ECG feature individually would result in a discrete subset of optimal ECG features, which would be detrimental to the analyst's observation of the ECG. Therefore, this invention adopts a method of dividing ECG features into ECG feature regions. For example, L ECG features are divided into A consecutive regions, each containing L / A ECG features. This design retains the significance of feature selection while enhancing its practical application value.

[0075] 1. Design PBDN: Loc (PBDN: Loc represents the PBDN used to evaluate the characteristic region of an electrocardiogram)

[0076] From Algorithm 1, we know that:

[0077]

[0078]

[0079]

[0080] The meanings of the symbols can be found in formulas (2.5, 2.6, 2.11).

[0081] The preprocessed PTB dataset is divided into a training set and a test set. The training set is then placed into PBDN:Loc for training.

[0082] 2. When the loss function in PBDN converges, training is complete. The test set is then placed into the trained model, and the pooling layers are analyzed.

[0083] The window size of the max pooling layer is equal to the number of ECG feature regions; that is, each channel selects the largest feature from A ECG feature regions as its channel feature. The selected ECG region for each channel is recorded.

[0084]

[0085]

[0086] in This indicates the selected ECG feature regions recorded across all channels for data categorized as normal. This represents the selected ECG feature regions recorded across all channels for data categorized as myocardial infarction. This indicates the number of times the i-th ECG feature region was selected by the max pooling layer when the category is normal.

[0087] This represents the number of times the i-th electrocardiogram feature region of category myocardial infarction is selected by the max pooling layer, i = 1, 2, 3, ..., A.

[0088] Through the above two steps, the number of times that the selected electrocardio feature region is put into the pooling layer when the test data of different categories is obtained. The more the number of times is counted, the more effective the identification of a certain category is, and the electrocardio feature region can be found in the original electrocardio signal, so that the function of backtracking is realized. Through comparison of the counted number, the preliminary evaluation of the electrocardio feature region is also completed.

[0089] Evaluation feature

[0090] The electrocardio feature region or lead with good characterization ability should have a large weight in the judgment of all categories. The number of times that the feature region recorded by each category is selected is converted into a weight and then assigned to the slope of y=kx in the present application. The feature image of each feature region contains the same number of straight lines as the categories, and whether the position feature occupies a large weight in the identification of all categories is judged by the sum of the angles between each two straight lines.

[0091] Specific steps:

[0092] 1. Convert the record number into a weight and then assign it to the slope of y=kx, the formula is as follows:

[0093]

[0094] where c represents the category, l represents a electrocardio feature region to be evaluated, and j represents the total number of electrocardio feature regions to be evaluated, represents the number of times that the electrocardio feature region l is selected by the maximum pooling layer when the category is c, represents the slope of the straight line corresponding to the electrocardio feature region l when the category is c, represents the weight corresponding to the electrocardio feature region l when the category is c.

[0095] 2. Draw the feature image of each position according to the slope and calculate the sum of the slopes and the sum of the angles

[0096] Take the evaluation of the No. 1 electrocardio feature region as an example:

[0097] The feature image of the No. 1 electrocardio feature region is shown in Figure 5 , (wherein two straight lines represent the mapping of the No. 1 electrocardio feature region in the normal and myocardial infarction two categories respectively):

[0098] The calculation method of the sum of the slopes and the sum of the angles is as follows:

[0099]

[0100]

[0101] wherein k sum1 represents the sum of the slopes of the No. 1 electrocardio feature image, and θsum1 represents the sum of the angles between every two straight lines in the No. 1 electrocardio feature map. The rest of the symbols refer to formula (2.18).

[0102] 3. Final evaluation of features

[0103] The final evaluation of electrocardio feature regions is obtained by the product of the sum of slopes and the cosine of the sum of angles:

[0104] P g = k sumg cosθ sumg (2.21)

[0105] where g = 1, 2, 3,..., j, j is the total number of electrocardio feature regions to be evaluated, P g is the final score of electrocardio feature region g, and the rest of the parameters refer to formula (2.19-2.20).

[0106] Proof:

[0107] cosθ is a monotonically decreasing function in the interval [0, π / 2] and its value range is between [0, 1], the smaller the angle, the larger the value of cosθ. When the angle exceeds π / 2, it means that the feature plays a great difference in the evaluation of different categories, which is not conducive to feature evaluation, so it is proved that formula (2.21) can evaluate the feature. Through the above three steps, the final evaluation of electrocardio feature regions can be realized. According to the evaluation results, the electrocardio data set is processed into the optimal electrocardio feature region subset.

[0108] Through the above three steps, the final evaluation of electrocardio feature regions can be realized. According to the evaluation results, the electrocardio data set is processed into the optimal electrocardio feature region subset.

[0109] OLoc = (N, D, OL) (2.22)

[0110] where OLoc represents the optimal electrocardio feature region subset after PBDN feature selection, N represents the number of samples, D represents the number of leads, and OL represents the number of electrocardio features contained in the optimal electrocardio feature region subset. Using these three steps, PBDN: Ld can also be constructed to evaluate the leads and find the optimal lead subset.

[0111] 2.2.4 PBDN feature selection method pseudo code

[0112]

[0113]

[0114] Introduction of experimental data

[0115] ​PTB Diagnostic ECG Database is an electrocardiogram (ECG) dataset commonly used in the field of medical research and ECG signal processing. This dataset was created and maintained by the Department of Cardiology, University of Bonn, Germany. The dataset contains ECG data of 290 patients in the University of Bonn Hospital. Each patient has 15-lead (12 leads + 3 additional leads) ECG signal recordings. These signal recordings include resting ECG and some recordings during exercise and stress tests. PTB Diagnostic ECG Database is commonly used in the fields of arrhythmia detection, cardiac pathology detection, evaluation and verification of ECG signal processing algorithms, etc.

[0116] Evaluation of ECG feature regions and leads using PBDN feature selection method

[0117] Firstly, according to Algorithm 1, PBDN: Loc and PBDN: Ld are constructed. PBDN: Loc is used to evaluate ECG feature regions, and the model structure is as shown in Figure 4 PBDN: Ld is used to evaluate leads, and the model structure is similar to PBDN: Loc, which is not described here. The preprocessed ECG data is divided into 8:2 training set and test set. The PBDN network is iterated using the training set, where the batch_size size is set to 128, one round of training is 100 epochs, and the basic learning rate is set to 0.001. After training, the test set is put into PBDN, the number of features selected is recorded, and it is converted into weight. The experimental results are shown in Tables 3.1, 3.2, 3.3, and 3.4.

[0118] Table 3.1 Record table of ECG feature regions retraced in normal ECG signals

[0119]

[0120]

[0121] Table 3.2 Record table of ECG feature regions retraced in myocardial infarction ECG signals

[0122]

[0123] Table 3.3 Record table of leads retraced in normal ECG signals

[0124]

[0125] Table 3.4 Record table of leads retraced in myocardial infarction ECG signals

[0126]

[0127]

[0128] Tables 3.1-3.4 record the number of times each type of data in the test set is selected as a feature after being put into the max pooling layer, where the weight is proportional to the number of times, and only the weight is observed to obtain the backtracking result.

[0129] Specifically, the weights of the 5th, 6th, and 7th electrocardiogram feature regions in Table 3.1 are 0.196, 0.336, and 0.152, respectively, which have obvious numerical advantages compared to other electrocardiogram feature regions. This indicates that the 5th, 6th, and 7th electrocardiogram feature regions are more effective for recognition when the model judges whether the test set data belongs to the normal type.

[0130] The weights of the 5th, 6th, and 7th electrocardiogram feature regions in Table 3.2 are 0.179, 0.330, and 0.228, respectively, which have obvious numerical advantages compared to other electrocardiogram feature regions. This indicates that the 5th, 6th, and 7th electrocardiogram feature regions are more effective for recognition when the model judges whether the test set data belongs to the myocardial infarction type.

[0131] The weights of V2, V3, V4, AVR, and V5 leads in Table 3.3 are 0.1249, 0.1257, 0.1741, 0.1714, and 0.1226, respectively, which have obvious numerical advantages compared to other leads. This indicates that V2, V3, V4, AVR, and V5 leads are more effective for recognition when the model judges whether the test set data belongs to the normal type.

[0132] The weights of V2, V3, V4, and AVR leads in Table 3.4 are 0.1672, 0.1444, 0.0878, and 0.0871, respectively, which have obvious numerical advantages compared to other leads. This indicates that V2, V3, V4, and AVR leads are more effective for recognition when the model judges whether the test set data belongs to the myocardial infarction type.

[0133] In order to obtain features with greater weight in all categories, the weight is assigned to the slope of the straight line y=kx, the sum of the slopes of all straight lines and the sum of the angles between each two straight lines are calculated, and finally the feature score P is calculated using these two values to complete the final evaluation of the features. The experimental results are shown in Tables 3.5 and 3.6.

[0134] Table 3.5 Evaluation Table of Electrocardiogram Feature Regions

[0135]

[0136] Table 3.6 Evaluation Table of Leads

[0137]

[0138] It can be seen from Tables 3.5-3.6 that the scores P of the 5th, 6th and 7th electrocardiogram feature regions are 0.3749, 0.6659 and 0.3699 respectively, which are obviously higher than those of other electrocardiogram feature regions. The scores P of the V2, V3 and V4 leads are 0.2918, 0.2700 and 0.2609 respectively, which are obviously higher than those of other leads. Therefore, the 5th, 6th and 7th electrocardiogram feature regions are selected to form an optimal electrocardiogram feature region subset, and the V2, V3 and V4 leads are selected to form an optimal lead subset. The positions of the selected electrocardiogram feature regions in the original electrocardiogram signal are shown in Fig. 3.1 (the selected electrocardiogram feature regions are marked by black circles) and Fig. 3.2 (the selected leads are marked by black squares). The experiment proves that the present application has the function of backtracking. Figure 6

[0139] The optimal feature subset is put into the SVM classifier for training

[0140] In order to prove whether the PBDN feature selection algorithm is effective, two experiments are designed in this subsection.

[0141] Experiment 1: All electrocardiogram data are put into the classifier for training, and the optimal electrocardiogram feature region subset is selected for training.

[0142] Experiment 2: All electrocardiogram data are put into the classifier for training, and the optimal lead subset is selected for training.

[0143] The results of the two experiments are shown in Tables 3.7 and 3.8.

[0144] Table 3.7 results of Experiment 1

[0145]

[0146] Table 3.8 results of Experiment 2

[0147]

[0148] From Tables 3.7 and 3.8, it can be seen that in Experiment 1, the number of electrocardiogram feature regions is reduced from 10 to 3 (reduced by 70%), and the recognition rate is only reduced by 0.41%, which can still reach 99.59%. In Experiment 2, the number of leads is reduced from 12 to 3 (reduced by 75%), and the recognition rate is only reduced by 0.07%, which can still reach 99.93%. Both experiments save more than 70% of computing power, but the recognition result is still as high as 99.5 or above, which achieves the expected research purpose and gets satisfactory results.

[0149] Through Experiments 1 and 2, the effectiveness of the PBDN feature selection method is proved, that is, the PBDN algorithm can select more effective leads and electrocardiogram feature regions for myocardial infarction recognition.

[0150] ​The myocardial infarction recognition algorithm proposed in the application is compared with other myocardial infarction recognition algorithms, and the results are shown in Table 3.9.

[0151] Table 3.9 Comparison table of myocardial infarction recognition methods

[0152]

[0153]

[0154] As can be seen from the comparison in Table 3.9, the myocardial infarction recognition method proposed in the application uses only 3 leads and only 42 features for each lead, saves more than 90% of computing power, greatly reduces the training time, and still achieves a recognition rate of 99.65%, achieving satisfactory results. The problems of poor interpretability and feature dimension expansion existing in other myocardial infarction recognition algorithms are solved.

[0155] IV. Conclusion

[0156] A new feature selection method based on PBDN is proposed in this paper, which helps to enhance the interpretability of deep learning models. In addition, the application also proposes a new myocardial infarction recognition algorithm, which uses the PBDN feature selection method to select the optimal feature subset and put it into the SVM model for training. The results show that the PBDN feature selection method is effective, and the new myocardial infarction recognition algorithm has excellent recognition rate while greatly reducing the computational cost.

[0157] The basic principles and main features of the application and the advantages of the application are shown and described above. It is obvious for those skilled in the art that the application is not limited to the details of the above exemplary embodiments, and the application can be implemented in other specific forms without departing from the spirit or essential characteristics of the application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the application. Any reference signs in the claims should not be considered as limiting the claims involved.

[0158] In addition, it should be understood that although the present specification is described in terms of embodiments, each embodiment does not contain only one independent technical solution, and the description manner of the specification is only for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be properly combined to form other embodiments that those skilled in the art can understand.

Claims

1. A method for myocardial infarction identification based on location backtracking deep learning network, characterized in that, The method comprises the following steps: S1, obtaining original data from a PTB electrocardio recognition data set, performing filtering, downsampling, R peak positioning and heart beat cutting preprocessing to obtain electrocardio data for training, in 12-lead electrocardio records, 12 heart beats with R peaks at the same time are selected to form a 2-dimensional electrocardio segment; S2, a feature selection method based on a position backtracking deep learning network PBDN is used to evaluate the electrocardio feature region and lead of the electrocardio data, first, a corresponding deep learning network PBDN is designed according to the evaluation dimension, the deep learning network PBDN is composed of a convolution layer, a pooling layer and a full connection layer, then the electrocardio data is divided into a training set and a test set, the training set is used to iterate the PBDN, after the model training is completed, the test set is put into the PBDN, the data of the normal and myocardial infarction two categories are put into the maximum pooling layer in the PBDN, the selected electrocardio feature region is counted, and the more the counted times are, the more effective the recognition of the myocardial infarction is, so as to realize the preliminary evaluation of the original feature, therefore the record times are converted into weights and then assigned to the slope of y=kx, the feature image of each position is drawn according to the slope and the sum of the slopes and the sum of the angles are calculated, two straight lines respectively represent the mapping of the normal and myocardial infarction two categories in the electrocardio feature region, the final evaluation of the electrocardio feature region is obtained through the product of the sum of the slopes and the cosine of the sum of the angles, through the above steps, the final evaluation of the electrocardio feature region is realized; S3, based on the evaluation information of S2, the electrocardio feature region and lead with high scores are obtained, respectively forming an optimal electrocardio feature region data set and an optimal lead data set and put into an SVM classifier for training to obtain the final recognition result of the myocardial infarction. 2.The myocardial infarction identification method based on location backtracking deep learning network according to claim 1, wherein, The filtering: a fourth-order Butterworth band-pass filter is used to filter out electrocardio signal noises below 250Hz and above 2000Hz. 3.The myocardial infarction identification method based on location backtracking deep learning network according to claim 1, characterized in that, The R peak positioning: the Pan-Tompkins algorithm is used to position the R peak in the electrocardio data set, the two main operations of the Pan-Tompkins algorithm are R wave enhancement and R wave decision, the R wave enhancement is used to strengthen the characteristics of the R peak, and the R wave decision is used to select the R wave from the R wave and the noise. 4.The myocardial infarction identification method based on location backtracking deep learning network according to claim 1, characterized in that, The heart beat cutting: the position difference between the R peaks before and after each lead is calculated, the maximum R peak interval Rmax and the minimum R peak interval Rmin in all leads are found, according to the position of the R peak in the array, (Rmax+Rmin) / 2 electrocardio data before the R peak and (Rmax+Rmin) / 2 electrocardio data after the R peak are obtained to form a heart beat, the same operation is performed on each R peak position and each lead. 5.The myocardial infarction identification method based on location backtracking deep learning network according to claim 1, wherein, The S1 includes 12 ECG records, 12 heartbeats with R peaks at the same time are selected to form a 2D ECG segment, finally, the preprocessed ECG dataset contains 10080 normal ECG signals and 51402 myocardial infarction ECG signals, each ECG signal contains 12 leads of heartbeats, and each heartbeat contains 140 ECG features, i.e. 70 ECG features before R peak and 70 ECG features after R peak.

Images