Atrial fibrillation electrocardiogram detection method using distributed entropy network features

Abstract

The present application provides a kind of atrial fibrillation electrocardio detection method using distribution entropy network feature, it is related to atrial fibrillation electrocardio signal processing field, especially it is related to the atrial fibrillation electrocardio detection method using degree distribution entropy, aggregation coefficient distribution entropy, local efficiency distribution entropy network feature.The specific steps include as follows: selecting the data segment of electrocardio signal;Each segment is coarse-grained;Complex network is created for each segment based on visual graph method;Three distribution entropy features of complex network are calculated;Three distribution entropy features, typical time domain features of electrocardio signal are calculated;Machine learning model is trained based on these features;Finally, model is used to detect and identify atrial fibrillation electrocardio;Experimental results show that, compared with traditional network feature, distribution entropy network feature has stronger atrial fibrillation electrocardio identification ability, so the distribution entropy network feature proposed in the present application is superior to traditional network feature.

Description

Technical Field

[0001] This invention relates to the field of atrial fibrillation electrocardiogram signal processing, and particularly to an atrial fibrillation electrocardiogram detection method utilizing network features such as degree distribution entropy, clustering coefficient distribution entropy, and local efficiency distribution entropy. It belongs to the technical field of complex network statistical features and atrial fibrillation electrocardiogram detection and identification. Background Technology

[0002] It is estimated that there are approximately 46.3 million people with atrial fibrillation worldwide. Early detection and treatment of atrial fibrillation in these patients can prevent it from developing into critical cardiovascular disease. Traditional diagnosis of atrial fibrillation relies on doctors observing electrocardiograms (ECGs), which is time-consuming and labor-intensive if the ECG data volume is large.

[0003] The development of computer technology and complex network theory has provided new insights into the detection and identification of atrial fibrillation (AF) ECGs. Computers can not only calculate the time-domain and frequency-domain features of ECG data, but also transform time-domain data into complex networks to calculate network features. Machine learning or deep learning methods can then be used to learn these features and subsequently detect and identify AF ECGs. However, traditional network features have limited ability to identify AF ECGs. This invention proposes three distribution entropy network features that offer significantly stronger AF ECG identification capabilities compared to traditional network features. Summary of the Invention

[0004] The technical problem solved by this invention is to address the issue that traditional network features cannot effectively identify atrial fibrillation ECG signals, and to propose three distribution entropy network features for ECG signals.

[0005] To address the aforementioned technical problems, the technical solution proposed in this invention is: a method for calculating three distribution entropy network features of electrocardiogram (ECG) signals, namely: a method for calculating degree distribution entropy, clustering coefficient distribution entropy, and local efficiency distribution entropy. The calculation steps mainly include: selecting ECG data segments, coarsening the segment data, constructing a complex network of the segment data, calculating the three distribution entropy features of the complex network, and calculating the three distribution entropy features of each ECG signal, which are then used for subsequent atrial fibrillation ECG identification.

[0006] 1. Select ECG data segments. Assume an ECG signal is x, with L points. According to equation (1), extract G segments (each segment with M points) from the center of the signal. The value of G is L divided by M and then rounded down. p represents the segment number, y p Let y represent the p-th segment. p (j) represents the j-th point in the p-th segment.

[0007]

[0008] 2. Coarse-grain each segment. According to equation (2), y p Coarsening, where s is the coarsening scale. This represents the coarsened sequence of the p-th segment, with length N. This represents the i-th point in the sequence.

[0009]

[0010] For the sake of brevity, we will use z to represent it below.

[0011] 3. Create a complex network for each segment sequence based on the visual diagram method.

[0012] Each data point in sequence z is mapped to a node in the network. Whether there is an edge between nodes depends on whether the two data points are visible. If equation (3) is satisfied, then the line between two points z(i) and z(j) is not cut off by the point z(p) between these two points, these two points are visible, and the corresponding network nodes have an edge. If equation (3) is not satisfied, then these two points are not visible, and the corresponding network nodes are not connected.

[0013]

[0014] 4. Calculate the three distribution entropy characteristics of complex networks.

[0015] Complex networks can be represented by an adjacency matrix, A = [a ij ] N×N , where a ij A value of 1 indicates that node i is connected to node j, and a ij A value of 0 indicates that node i and node j are not connected, and N represents the number of nodes.

[0016] (1) Degree distribution entropy (DD) e )feature

[0017] The degree D of each network node is calculated according to equation (4). i The degree values ​​are discrete, ranging from 0 to N-1. Based on the degree value of each node, the degree values ​​are arranged from smallest to largest, defining a degree distribution P. D The ratio of the number of nodes corresponding to each degree value to the total number of nodes is shown in equation (5), where D min and D max These are the minimum and maximum degree values, respectively, and N represents the total number of nodes in the network. This represents the number of nodes corresponding to each degree value. Based on the information entropy formula, the degree distribution entropy DD is defined. e As shown in equation (6).

[0018]

[0019]

[0020]

[0021] (2) Clustering coefficient distribution entropy (CD) e )feature

[0022] The clustering coefficient C of each network node is calculated according to equation (7). i Its value is continuous, ranging from 0 to 1, where D i It is the degree of node i, e i It is the number of connections between adjacent nodes of node i. Divide the range from 0 to 1 into K intervals, and arrange the clustering coefficients of each node from smallest to largest, defining the clustering coefficient distribution P. C The ratio of the number of nodes in each interval to the total number of nodes is shown in equation (8), where This represents the number of nodes in the k-th interval. Based on the information entropy formula, the clustering coefficient distribution entropy CD is defined. e As shown in equation (9).

[0023]

[0024]

[0025]

[0026] (3) Local efficiency distribution entropy (ED) e )feature

[0027] The local efficiency E of each network node is calculated according to equation (10). i Its value is continuous, ranging from 0 to 1, where D i It is the degree of node i, l ik It is the shortest distance between node i and node k, N i It is the set of nodes consisting of the adjacent nodes of node i. The range from 0 to 1 is divided into K intervals, and the local efficiency of each node is arranged in ascending order, defining a local efficiency distribution P. E The ratio of the number of nodes in each interval to the total number of nodes is shown in equation (11), where Let represent the number of nodes in the k-th interval. Based on the information entropy formula, define the local efficiency distribution entropy ED. e As shown in equation (12).

[0028]

[0029]

[0030]

[0031] 5. Calculate the three distribution entropy characteristics of the ECG signal.

[0032] Since a single ECG signal can be used to extract G segments (each segment containing M points), each segment needs to be coarsely processed into an N-point sequence. Each N-point sequence can be constructed into a complex network, and each network can solve for three distribution entropy features. Thus, an ECG signal can obtain G sets of distribution entropy features (each set contains three distribution entropy features). Finally, the G sets of distribution entropy features are averaged, and each ECG signal can obtain three distribution entropy features, namely degree distribution entropy feature, clustering coefficient distribution entropy feature, and local efficiency distribution entropy feature.

[0033] For an ECG signal, five typical time-domain characteristics can also be calculated, namely the number of QRS waves per minute (N). QRS ), RR interval mean (R m ), RR interval standard deviation (R) sd ), the mean of the RR interval differences (RD) m ), standard deviation of RR interval difference (RD) sd These five typical time-domain features are existing features, and existing algorithms are used in their calculation.

[0034] 6. Train the machine learning model

[0035] Machine learning models are trained using three types of distribution entropy features and typical time-domain features, such as support vector machines, decision trees, k-nearest neighbors classifiers, and ensemble learning.

[0036] 7. Atrial fibrillation detection

[0037] The trained model was used for the detection and identification of atrial fibrillation electrocardiograms. Attached Figure Description

[0038] The following description, in conjunction with the accompanying drawings, further illustrates the atrial fibrillation electrocardiogram detection method based on the characteristics of distributed entropy networks proposed in this invention.

[0039] Figure 1 This is the overall flowchart of an ECG detection method for atrial fibrillation that utilizes the characteristics of distributed entropy networks.

[0040] Figure 2 This is a detailed step-by-step diagram illustrating the ECG detection method for atrial fibrillation using the characteristics of distributed entropy networks. Detailed Implementation

[0041] The ECG data used in this embodiment were selected from atrial fibrillation ECGs and normal ECGs from the 2017 PhysioNet / Cinc Challenge, including 758 atrial fibrillation ECGs and 5076 normal ECGs. The sampling frequency of this ECG data was 300Hz, and ECGs with a data length of 10 seconds or more were selected as experimental data, including 757 atrial fibrillation ECGs and 5066 normal ECGs, with ECG data lengths ranging from 3000 to 18300 points.

[0042] 1. Select ECG data segments according to formula (1).

[0043]

[0044] The ECG length L ranges from 3000 to 18300 points. To reduce the size of the subsequent complex network, G segments are selected from the center of each ECG, each segment having a length M of 3000 points. The number of segments extracted varies depending on the ECG length. When 3000 ≤ L < 6000, one segment is selected from the center; when 6000 ≤ L < 9000, two segments are selected; when 9000 ≤ L < 12000, three segments are selected, and so on.

[0045] 2. Coarse-grain the ECG data fragments according to formula (2).

[0046]

[0047] After step 1, the ECG data segment length is 3000 points. To further reduce the size of the subsequent complex network, the data segment needs to be coarsened. The coarsening scale factor s ranges from 5 to 14. When s = 5, every 5 data points are averaged to become 1 point, and the data segment length becomes 600. When s = 6, every 6 data points are averaged to become 1 point, and the data segment length becomes 500. When s = 7, every 7 data points are averaged to become 1 point, and the data segment length becomes 428, and so on.

[0048] 3. Create a complex network for the coarse-grained ECG data fragments according to the visualization criterion (3).

[0049]

[0050] The number of points in a complex network is the same as the number of points in the coarse-grained data segment. Each data point in the ECG segment corresponds to a node in the complex network. Whether any two nodes in the network are connected depends on whether the data points corresponding to these two nodes are visible, i.e., whether the visibility criterion (3) is satisfied. If the visibility criterion (3) is satisfied, that is, any data point p between the i-th data point and the j-th data point does not truncate the connection between these two data points, then the i-th data point and the j-th data point are visible, and the corresponding nodes i and j are connected. If the visibility criterion (3) is not satisfied, nodes i and j are not connected.

[0051] 4. Calculate the three distribution entropy characteristics of complex networks.

[0052] The created complex network can be represented by an adjacency matrix, A = [a ij ] N×N , where a ij A value of 1 indicates that node i is connected to node j, and a ij A value of 0 indicates that node i and node j are not connected, and N represents the number of nodes. The value of s ranges from 5 to 14.

[0053] Calculate the degree D of each node in the complex network according to formulas (4), (5), and (6). i The degree distribution P of complex networks D And the degree distribution entropy DD of complex networks e D min and D max Let these represent the minimum and maximum degree values ​​of a network node, respectively. This represents the number of network nodes corresponding to each degree value.

[0054]

[0055]

[0056]

[0057] According to formulas (7), (8), and (9), calculate the clustering coefficient C of each node in the complex network. i Clustering coefficient distribution P of complex networks C And the clustering coefficient distribution entropy CD of complex networks e , where e i D is the number of connections between the adjacent nodes of node i. i Let be the degree value of node i, and let the clustering coefficient range from 0 to 1. Divide this range into K = 200 intervals. This represents the number of network nodes within each clustering coefficient interval.

[0058]

[0059]

[0060]

[0061] According to formulas (10), (11), and (12), calculate the local efficiency D of each node in the complex network. i The local efficiency distribution P of complex networks D And the local efficiency distribution entropy ED of complex networks e , where l ik D is the shortest distance between node i and node k. i Let be the degree value of node i. The local efficiency range is from 0 to 1. This range is divided into K = 400 intervals. This represents the number of network nodes within each local efficiency interval.

[0062]

[0063]

[0064]

[0065] 5. Calculate the three distribution entropy characteristics of the ECG signal.

[0066] Each ECG signal can be used to extract G segments, each segment containing 3000 points. These segments are then coarsened and converted into point counts. The points are segmented, with the coarse-grained scale s varying from 5 to 14. A complex network is then constructed for each segment, and each network obtains three distribution entropy features. Therefore, an ECG signal can obtain G sets of distribution entropy features. By averaging the G sets of features, the three distribution entropy features of the ECG signal are finally obtained: degree distribution entropy, clustering coefficient distribution entropy, and local efficiency distribution entropy.

[0067] Simultaneously, five typical time-domain characteristics of each ECG signal were calculated, namely the number of QRS waves per minute (N). QRS ), RR interval mean (R m ), RR interval standard deviation (R) sd ), the mean of the RR interval differences (RD) m ), standard deviation of RR interval difference (RD) sd ).

[0068] 6. Train the machine learning model

[0069] The experimental data consisted of 757 atrial fibrillation ECGs and 5066 normal ECGs from the 2017 PhysioNet / Cinc Challenge. 80% of the ECGs were used as the training set, and 20% as the test set. Five machine learning models were employed: Support Vector Machine (SVM), Decision Tree (DT), Russboost Tree (RBT) from ensemble learning, k-Nearest Neighbors (KNN) classifier, and Bayesian classifier (NB). The machine learning models were trained using the distribution entropy features and temporal features of the ECG signals.

[0070] 7. Atrial fibrillation detection

[0071] For a 20% ECG test set, five pre-trained machine learning models were used to detect atrial fibrillation ECGs. For experimental data and results analysis, please refer to 9 (Comparative Analysis of Atrial Fibrillation ECG Recognition Accuracy).

[0072] 8. Compare the characteristics of three distribution entropy networks with those of traditional networks based on the T-test method.

[0073] Table 1 lists 15 network features, the first three of which are the three distribution entropy network features proposed in this invention, and the last 12 are traditional network features.

[0074] Table 1. 15 Network Characteristics of ECG

[0075]

[0076]

[0077] The T-test was used to study the ability of 15 network features to identify atrial fibrillation ECGs from normal ECGs. Table 2 records the p-values ​​of the T-tests; the smaller the p-value, the stronger the feature's ability to identify atrial fibrillation ECGs. It can be seen that the minimum p-value for degree distribution entropy is 1.16E-94 (when s=5), the minimum p-value for clustering coefficient distribution entropy is 3.47E-76 (when s=10), and the minimum p-value for local efficiency distribution entropy is 2.16E-72 (when s=10). Except for the transitivity feature, the three distribution entropy features outperform all other network features.

[0078] Table 2. T-test p-values ​​for 15 network features

[0079] <![CDATA[DD e ]]> 1.16E-94 2.27E-92 6.41E-92 6.00E-92 3.07E-91 3.89E-91 1.66E-91 2.36E-92 5.10E-90 6.84E-90 <![CDATA[CD e ]]> 6.61E-23 2.10E-35 1.71E-47 1.16E-59 1.07E-68 3.47E-76 4.74E-75 7.19E-76 9.49E-72 7.54E-70 <![CDATA[ED e ]]> 2.41E-21 1.04E-33 7.88E-46 3.78E-57 2.26E-65 2.16E-72 3.43E-70 1.38E-70 1.42E-68 1.84E-66 <![CDATA[D m ]]> 1.58E-70 5.27E-70 4.27E-71 2.98E-70 4.20E-71 7.19E-71 2.10E-70 1.34E-71 1.11E-70 1.41E-70 <![CDATA[D sd ]]> 2.61E-21 4.46E-17 2.57E-15 3.47E-12 1.08E-11 7.94E-10 1.23E-09 3.03E-09 2.18E-08 5.62E-08 <![CDATA[C m ]]> 0.6029 0.7725 0.1157 0.0079 2.13E-04 2.02E-05 7.62E-08 3.38E-09 5.91E-11 1.08E-12 <![CDATA[C sd ]]> 3.10E-44 2.72E-47 1.94E-49 3.53E-47 1.15E-49 5.63E-51 1.12E-49 3.09E-48 4.15E-49 8.05E-50 <![CDATA[T r ]]> 4.34E-46 2.67E-53 8.29E-62 9.64E-70 7.55E-76 1.84E-82 4.75E-86 2.32E-93 8.72E-95 1.26E-98 <![CDATA[L m ]]> 5.62E-15 7.21E-12 4.10E-10 4.91E-10 4.78E-08 1.43E-07 1.54E-07 4.18E-07 1.20E-07 3.38E-07 <![CDATA[L sd ]]> 2.33E-04 0.0063 0.0198 0.0041 0.0567 0.0472 0.0607 0.0539 0.0071 0.0556 <![CDATA[E g ]]> 7.36E-19 1.70E-15 7.86E-14 3.10E-13 1.65E-11 5.14E-11 5.82E-11 1.18E-10 1.02E-10 5.00E-11 <![CDATA[E m ]]> 0.5493 0.1072 0.0016 1.59E-05 2.61E-08 5.30E-10 1.39E-13 5.41E-15 3.33E-17 6.73E-20 <![CDATA[E sd ]]> 8.16E-42 4.86E-40 6.76E-42 1.36E-38 2.00E-41 7.12E-43 8.64E-43 3.35E-39 1.20E-38 1.50E-39 <![CDATA[E e ]]> 3.09E-26 3.10E-25 8.21E-29 1.41E-26 1.35E-29 4.33E-32 8.33E-32 1.72E-28 6.99E-28 1.50E-29 <![CDATA[C e ]]> 2.52E-35 3.33E-40 3.08E-45 2.33E-45 5.31E-49 1.95E-51 1.48E-51 6.92E-50 1.22E-50 1.52E-52

[0080] 9. Comparative Analysis of Atrial Fibrillation ECG Recognition Accuracy

[0081] Select the smallest p-value less than 10 from Table 2. -30 Better network features are used for atrial fibrillation ECG identification. These features include: three distribution entropy features (DD... e CD e ED e), transitivity (T) r ), degree mean (D) m ), clustering coefficient standard deviation (C sd ), local efficiency standard deviation (E) sd ), local efficiency entropy (E) e ), clustering coefficient entropy (C e The coarsening factor s was chosen to be 10. The number of QRS waves per minute (N) was calculated based on five time-domain characteristics of the electrocardiogram signal. QRS ), RR interval mean (R m ), RR interval standard deviation (R) sd ), the mean of the RR interval differences (RD) m ), standard deviation of RR interval difference (RD) sd Based on these five features, different network features are added to each feature to compare the accuracy of various network features in identifying atrial fibrillation ECGs.

[0082] The top three precision values ​​for each model in Table 3 are shown in bold. It can be seen that, except for the third-place model in the RBT model, which is due to the addition of the mean degree (D... m The top three accuracies of other models and the top two accuracies of the RBT model were obtained by adding degree distribution entropy (DD). e ), or clustering coefficient distribution entropy (CD) e ), or local efficiency distribution entropy (ED) e As can be seen, the three distribution entropy network features proposed in this invention are superior to other traditional network features in their ability to identify atrial fibrillation ECGs.

[0083] Table 3 shows the ECG classification accuracy of each model after adding different network features.

[0084] <![CDATA[F time (N QRS ,R m ,R sd ,RD m ,RD sd )]]> 94.42 92.59 93.08 92.99 89.06 <![CDATA[F time +ED e ]]> 95.16 94.92 95.51 94.87 89.96 <![CDATA[F time +CD e ]]> 95.11 94.72 94.72 94.77 90.06 <![CDATA[F time +DD e ]]> 95.01 94.72 94.08 94.87 90.16 <![CDATA[F time +T r ]]> 94.77 93.33 93.53 93.93 89.25 <![CDATA[F time +D m ]]> 94.72 93.53 94.57 93.18 89.71 <![CDATA[F time +C sd ]]> 94.77 93.13 93.38 94.03 88.76 <![CDATA[F time +E sd ]]> 94.17 93.33 93.93 93.08 89.31 <![CDATA[F time +E e ]]> 94.42 92.59 93.08 93.44 89.01 <![CDATA[F time +C e ]]> 94.82 93.53 93.73 93.08 88.96

[0085] The present invention is not limited to the specific technical solutions described in the above embodiments. All technical solutions formed by equivalent substitutions are within the scope of protection claimed by the present invention.

Claims

1. A method for detecting atrial fibrillation using distributed entropy network characteristics, characterized in that, include: S1: Select a data segment of the electrocardiogram (ECG) signal; Assuming x is an ECG signal at point L, G segments are extracted from the center of the signal according to equation (1). Each segment has M points, and the value of G is L divided by M and then rounded down. g represents the segment number. This represents the g-th segment. This represents the m-th point in the g-th segment; , , (1)； S2: Coarse-grained processing is performed on each segment selected in S1 to obtain the coarse-grained sequence z; According to formula (2) Coarsening, where s is the coarsening scale. This represents the coarsened sequence of the g-th segment, with length N. , This represents the nth point in the coarsened sequence of g fragments. （2） For the sake of brevity, we will use z to represent it below. ; S3: Create a complex network for each segment based on the visual diagram method; Specifically, this includes mapping each data point in sequence z to a node in the network. Whether there is a connection between nodes depends on whether the two data points are visible. If equation (3) is satisfied, then the connection between two points z(i) and z(j) is not cut off by the point z(p) between these two points, these two points are visible, and the corresponding network nodes have a connection. If equation (3) is not satisfied, then these two points are not visible, and the corresponding network nodes are not connected. (3)； S4: Calculate the three distribution entropy characteristics of complex networks; Specifically, this includes: complex networks can be represented by adjacency matrices. , where a ij A value of 1 indicates that node i is connected to node j, and a ij A value of 0 indicates that node i and node j are not connected, and N represents the number of nodes. Define the degree distribution entropy respectively. Features, clustering coefficient distribution entropy Features, local efficiency distribution entropy feature; S5: Calculate the three distribution entropy characteristics of the electrocardiogram signal; S6: Feed the features into the machine learning model for training; S7: Use the trained model for atrial fibrillation ECG recognition; Calculate the degree D of each network node according to equation (4). i The degree value is discrete, ranging from 0 to N-1; based on the degree value of each node, the degree values ​​are arranged from smallest to largest, defining the degree distribution. The ratio of the number of nodes corresponding to each degree value to the total number of nodes is shown in equation (5), where D min and D max These are the minimum and maximum degree values, respectively, and N represents the total number of nodes in the network. This represents the number of nodes corresponding to each degree value; based on the information entropy formula, the degree distribution entropy is defined. As shown in equation (6): (4) (5) (6)； The clustering coefficient C of each network node is calculated according to equation (7). i Its value is continuous, ranging from 0 to 1, where D i It is the degree of node i, e i It is the number of connections between adjacent nodes of node i; divide the range from 0 to 1 into K intervals, and arrange the clustering coefficients of each node from smallest to largest, defining the clustering coefficient distribution. The ratio of the number of nodes in each interval to the total number of nodes is shown in equation (8), where This represents the number of nodes in the k-th interval; based on the information entropy formula, the clustering coefficient distribution entropy is defined. As shown in equation (9): （7） (8) (9)； The local efficiency E of each network node is calculated according to equation (10). i Its value is continuous, ranging from 0 to 1, where D i It is the degree of node i, l jq It is the shortest distance between node j and node q, N i It is the set of nodes consisting of the adjacent nodes of node i; the range from 0 to 1 is divided into K intervals, and the local efficiency of each node is arranged from smallest to largest, defining the local efficiency distribution. The ratio of the number of nodes in each interval to the total number of nodes is shown in equation (11), where This represents the number of nodes in the k-th interval; based on the information entropy formula, the local efficiency distribution entropy is defined. As shown in equation (12): (10) (11) (12)。 2. The method for detecting atrial fibrillation using distributed entropy network characteristics according to claim 1, characterized in that: S5 specifically includes: in S1, each ECG signal can be extracted into G segments; in S3, a complex network is constructed for each segment; each ECG signal constructs G complex networks, which have G sets of distribution entropy features, namely G degree distribution entropy, G clustering coefficient distribution entropy, and G local efficiency distribution entropy; the G sets of features are averaged respectively to finally obtain the degree distribution entropy, clustering coefficient distribution entropy, and local efficiency distribution entropy features of each ECG signal.

3. The method for detecting atrial fibrillation using distributed entropy network characteristics according to claim 1, characterized in that, S5 also includes: calculating five typical time-domain characteristics; that is, calculating the number N of QRS waves per minute. QRS RR interval mean R m RR interval standard deviation R sd The mean of the RR interval differences RD m RR interval difference standard deviation RD sd .

Images