Electrocardiogram morphology recognition method based on hybrid amplitude sequence division network

Abstract

The application discloses a method for recognizing electrocardiogram morphology based on a hybrid amplitude sequence division network. After signal preprocessing, a sequence partition network corresponding to a time sequence and an absolute amplitude sequence partition network are constructed, 10 network features are extracted from the two networks respectively as representations of signal morphology and amplitude level information, and the features are combined into hybrid amplitude sequence division network features. The features are input into a machine learning classification algorithm to classify electrocardiogram morphology, accurate detection of abnormal electrocardiogram is realized, and a flexible morphological abnormality detection framework is provided for other physiological signals.

Description

ECG morphology recognition method based on hybrid amplitude sequence partitioning network Technical Field

[0001] This invention relates to the fields of signal detection and smart healthcare, specifically to a method for electrocardiogram morphology recognition based on a hybrid amplitude sequence partitioning network. Background Technology

[0002] Electrocardiographic (ECG) identification and classification of arrhythmias play a crucial role in the early prevention and diagnosis of cardiovascular diseases. Clinically, subtle changes in ECG morphology and rhythm can indicate cardiovascular pathology. Therefore, many researchers and scientists have dedicated themselves to applying feature engineering and / or artificial intelligence methods to extract morphological and rhythmic information from ECG time series or heartbeats for early arrhythmia estimation. However, the nonlinear nature of the cardiovascular system necessitates new approaches to explore the hidden nonstationary, chaotic, and complex dynamic information within ECGs to achieve more accurate cardiac diagnosis.

[0003] As a nonlinear dynamic time series analysis method, complex networks have been used in various physical and engineering systems, and various network metrics have been used to characterize their underlying dynamics. In the past decade, several graph-based time series characterization methods have been proposed to reveal important information about the sequences themselves, such as visibility graphs, recurrent networks, and transition networks.

[0004] Sequence partitioning networks are a novel transitional network in which nodes represent ordinal patterns associated with the amplitude values ​​of time series segments, and links are assigned based on these patterns. Sequence partitioning networks offer a method for constructing complex networks without explicit embedding processes and outperform amplitude merging methods in capturing information about small amplitude variations. However, the ordinal patterns of segments are permutations defined by the amplitude order between elements, which suppresses amplitude-level information between each window. That is, the ordinal patterns only capture the segment shape, not the amplitude size of the original time series. When two segments have the same amplitude permutation order, the ordinal patterns are identical. Therefore, subsequent network-based features are susceptible to noise under real-world conditions, necessitating a method that preserves overall amplitude information to improve the identification of ECG morphological abnormalities. Summary of the Invention

[0005] Technical issues:

[0006] To better acquire ECG morphology and rhythm information and enhance the interpretability of ECG morphological abnormality identification, this invention proposes an ECG morphology recognition method based on a hybrid amplitude sequence partitioning network. After signal preprocessing, a sequence partitioning network and an absolute amplitude partitioning network corresponding to the ECG signal are constructed. ECG morphology information is encoded into the topology of the complex network. Network features extracted from the two networks are mixed, and the local morphological information obtained from the sequence partitioning network and the amplitude information obtained from the absolute amplitude sequence partitioning network are integrated. Finally, the network features are input into a machine learning model for training and testing.

[0007] Technical solution:

[0008] To achieve the above objectives, the present invention adopts the following technical solution: an electrocardiogram morphology recognition method based on a hybrid amplitude sequence partitioning network, comprising the following steps:

[0009] S1, preprocesses the electrocardiogram signal;

[0010] S2, construct a sequence partitioning network and an absolute amplitude sequence partitioning network corresponding to the time series;

[0011] S3, calculate the feature vectors of the sequence partitioning network and the absolute amplitude sequence partitioning network, and extract ECG morphology and amplitude information. The feature vectors include average degree, average clustering coefficient, graph energy, transitivity, link density, average degree centrality, s-metric, permutation entropy, conditional permutation entropy, and global node entropy.

[0012] S4. Combine the feature vectors from step S3 into a mixed amplitude feature, and classify the mixed amplitude feature into different electrocardiogram morphologies using a machine learning classifier.

[0013] As an improvement of the present invention, in step S1, a Butterworth bandpass filter of 0.1 to 45 Hz is used to remove baseline drift and high-frequency noise, and the filtered signal is normalized to the range of [-1, 1].

[0014] As an improvement of the present invention, step S2, which involves constructing a sequence partitioning network and an absolute amplitude sequence partitioning network corresponding to the time series, specifically includes:

[0015] S21, divide the standardized time series and the coarse-grained time series into a set of embedding vectors;

[0016] S22, mapping the embedded vector to the ordinal pattern according to the amplitude order;

[0017] S23, Construct the network according to the transformation order of the ordinal pattern.

[0018] As an improvement of the present invention, in the step S21, the time series is divided into a set of embedding vectors by using the embedding dimension and the delay time The i-th embedding vector is:[[]]

[0019] v i ={x i , x i+τ , x i+2τ , …, x i+(m-1)τ},

[0020] where m is the embedding dimension, τ is the delay time, x = {x1, x2, x3, …, x N} is a one-dimensional time series with a length of N, and the embedding dimension m determines the fineness of the time series division.[[]]

[0021] As an improvement of the present invention, in the step S21, the standardized time series is coarsened, specifically: by using the equal-interval univariate symbol method, the entire amplitude range of the time series x normalized to the range of [-1, 1] is divided into α regions, and each region is associated with a specific symbol value. According to the region where the data point x(i) is located, each data point is replaced with a specific symbol, and the time series x with a length of N is converted into a symbol sequence y = {y i , i = 1, 2, …, N}, and the conversion formula is:[[]]

[0022]

[0023] where,[[]] is the selected symbol set.[[]]

[0024] As an improvement of the present invention, in the step S22, the elements in each embedding vector are sorted, and according to the element amplitude order the embedding vector v i is mapped to the ordinal pattern s i =(π1, π2, π3, …, π m ), where, π k ∈{1, 2, …, m}, and when k≠l π k≠π l . In order to make the obtained ordinal pattern unique, if the values of two elements in v i are equal, then the ordinals of these two elements are assigned according to the order of their appearance in the embedding vector, that is, when x k =x l , if k<l then π k <π l . For the embedding dimension m, the number of all possible ordinal patterns is m!; the coarsened time series, that is, the symbol sequence y = {y iThe symbol sequence y = {y,i = 1, 2, ..., N} is given by {y,i = 1, 2, ..., N}. i A series of embedding vectors v of i = 1, 2, ..., N i ={y i ,y i+τ ,y i+2τ ,…,y i+(m-1)τ} and ordinal pattern s i =(π1,π2,π3,…,π) m ) directly corresponds to, i.e., y k =π k ,k=i+(m-1)τ.

[0025] As an improvement of the present invention, in step S23, an undirected unweighted sequence partitioning network and an absolute amplitude sequence partitioning network are constructed according to the ordinal pattern. The ordinal pattern of the embedding vector mapping obtained from the normalized time series partitioning is used to construct the sequence partitioning network, and the ordinal pattern of the embedding vector mapping obtained from the coarse-grained time series partitioning is used to construct the absolute amplitude sequence partitioning network. The process of constructing the network according to the ordinal pattern sequence is specifically as follows: the ordinal pattern sequence... The set of all distinct ordinal patterns included in the set is denoted as the set. gather Each distinct ordinal pattern in the network G = {V, E} corresponds to a node in the network. The order of links between nodes in the network depends on whether the node corresponds to an ordinal pattern in the ordinal pattern sequence. The order in which nodes appear in the set, and the edge between two nodes indicates the ordinal pattern of those two nodes in the set. In G, edges are adjacent, E is the set of connected edges, and G is mathematically represented as an adjacency matrix A, whose elements a i,j The value of represents whether there is a link between nodes i and j.

[0026] As an improvement of the present invention, in step S3, the average degree is the average degree of all nodes, and its calculation formula is as follows: Where N is the total number of nodes, d n Let n be the degree of node n, which is the number of neighboring edges that node n has.

[0027] As an improvement of the present invention, in step S3, the ratio of the number of triangles involving a node to the number of connected triples centered on that node is its clustering coefficient, and the average clustering coefficient is the average of the clustering coefficients of all nodes, and its calculation formula is as follows: Among them, e i Represents the k connected to node i. i The number of actual edges between nodes, k i *(k i -1) represents all k nodes of node i. iThe maximum possible number of edges between N neighboring nodes, where V is the set of N nodes.

[0028] As an improvement of the present invention, in step S3, the graph energy is the sum of the absolute values ​​of the real components of the eigenvalues ​​of the adjacency matrix corresponding to the complex network, and its calculation formula is: GE = ∑ i |λ i |, where λ i is the absolute value of the real part of the eigenvalues ​​of the adjacency matrix.

[0029] As an improvement of the present invention, the formula for calculating transitivity in step S3 is as follows: Where adj is the adjacency matrix, G(adj) is the number of connected triples in the graph, and F(adj) is the number of triangles in the graph.

[0030] As an improvement to the present invention, in step S3, the link density is the ratio between the number of edges and the maximum possible number of edges, and its calculation formula is as follows: Among them, e ij This represents the edge connecting node i and node j, where N is the total number of nodes.

[0031] As an improvement of the present invention, in step S3, degree centrality is defined as the ratio between the number of nodes connected to the current node and the total number of nodes in the network, and the formula for calculating the average degree centrality is: Among them, e ij This represents the edge connecting node i and node j, where N is the total number of nodes.

[0032] As an improvement of the present invention, in step S3, the s metric is the sum of the products of the degrees of all two nodes on the edge, and its calculation formula is as follows: Where, d i (d j ) represents the degree of node i(j), that is, the number of neighboring edges that node i(j) has.

[0033] As an improvement of the present invention, in step S3, the permutation entropy is the Shannon entropy of the ordinal pattern S, and its calculation formula is: PE=-∑ i p i logp i , where s i The probability mass function P(s=s) ∈S i ) = p i It is estimated by calculating the relative frequency of each symbol in S.

[0034] As an improvement of the present invention, in step S3, the conditional permutation entropy is the expected value of the average entropy of node i on the stationary distribution of the ordered network, and its calculation formula is: CPE=∑ i (-pi ∑ j p i,j logp i,j ), where p i,j It is estimated from s based on S. i to s j The transition probability.

[0035] As an improvement to the present invention, in step S3, the global node entropy is the transition complexity of S, and its calculation formula is: in, It is the corrected random matrix P of an ordered network that excludes the possibility of self-looping. T elements, elementa i,j This represents the link between nodes i and j.

[0036] As an improvement of the present invention, in step S4, the mixed amplitude features include fragment morphology information from the sequence partitioning network and amplitude information from the absolute amplitude sequence partitioning network, which are input into a machine learning classifier for classification.

[0037] Beneficial effects:

[0038] 1. Complex networks are a nonlinear dynamic analysis method. Changes in the morphology and rhythm of electrocardiogram (ECG) signals are related to changes in the structure and characteristics of complex networks. Different ECG morphologies correspond to different symbol sequence pattern changes in this method. Therefore, the proposed complex network is interpretable and corresponds to clinical diagnostic information.

[0039] 2. The number of transition states between network nodes is related to the width and slope of the original waveform. In the absolute amplitude sequence partition network, nodes composed of the same symbol have a larger degree. These nodes correspond to the characteristic wavelets in the original electrocardiogram. The nodes between them represent the transition between characteristic wavelets, presenting as symbol pairs with increasing and decreasing order. The transition rate depends on the slope of the QRS complex and corresponds to the width of the QRS complex.

[0040] 3. The magnitude of the symbol sequence value of the node is related to the amplitude of the original waveform. The sequence partitioning network captures the local shape information of the time series. As a supplement, the proposed absolute amplitude sequence partitioning network maps the global amplitude level information into the network. The hybrid amplitude sequence partitioning network can obtain more comprehensive information related to the segment shape and amplitude level, reflecting the temporal evolution of the ECG signal and the amplitude-related state changes. Attached Figure Description

[0041] Figure 1 is a schematic diagram of the electrocardiogram morphology recognition method based on a hybrid amplitude sequence partitioning network according to the present invention.

[0042] Figure 2 is a schematic diagram of the electrocardiogram and network construction of normal and premature ventricular contractions in Embodiment 2 of the present invention. The three columns from left to right correspond to normal heartbeat, typical premature ventricular contraction-1 and typical premature ventricular contraction-2, respectively. The three rows from top to bottom are the standardized electrocardiogram, the sequence partitioning network diagram and the absolute amplitude sequence partitioning network diagram, respectively.

[0043] Figure 3 is a schematic diagram of the electrocardiogram of a normal heartbeat and atrial flutter in Embodiment 3 of the present invention. The left figure is a schematic diagram of the electrocardiogram of a normal heartbeat, and the right figure is a schematic diagram of the electrocardiogram of atrial flutter. Detailed Implementation

[0044] To facilitate understanding of the present invention, a more detailed description will be provided below with reference to the accompanying drawings and embodiments. Several application examples of the present invention are shown in the drawings, but the present invention is not limited to the embodiments described herein.

[0045] Example 1

[0046] Figure 1 illustrates the flowchart of the ECG morphology recognition method based on a hybrid amplitude sequence partitioning network, including the following steps:

[0047] S1 preprocesses the ECG signal by using a Butterworth bandpass filter of 0.1–45 Hz to remove baseline drift and high-frequency noise, and normalizes the filtered signal to [-1, 1].

[0048] S2, constructing a sequence partitioning network and an absolute amplitude sequence partitioning network corresponding to the time series, specifically including:

[0049] S21. The standardized time series and the coarse-grained time series are each divided into a set of embedding vectors. Assuming the standardized time series is x = {-0.89, 0.39, 0.29, 0.91, 0.11, 0.25, 0.31, -0.09, 0.18, -0.16}, and setting the embedding dimension m = 3 and the time lag τ = 2, the embedding vector is v = {(-0.89, 0.29, 0.11), (0.39, 0.91, 0.25), (0.29, 0.11, 0.31), (0.91, 0.25, -0.09), (0.11, 0.31, 0.18), (0.25, -0.09, -0.16)}. Using the equal-interval univariate sign method, the time series x is converted into a sign sequence through a unified quantization process. The ten-sign (0-9) is applied to the normalized time series, as shown in the formula.

[0050]

[0051] The coarse-grained symbol sequence of x is obtained as y = {0,6,6,9,5,6,6,4,5,4}. With the embedding dimension m = 3 and the time delay τ = 2, the embedding vector is v = {(0,6,5),(6,9,6),(6,5,6),(9,6,4),(5,6,5),(6,4,4)}.

[0052] S22, Map the embedding vector to the ordinal pattern according to the amplitude order. An embedding vector v1 = (-0.89, 0.29, 0.11) obtained from the normalized time series partitioning satisfies the order x (i-1)+1 ≤x (i-1)+3 ≤x (i-1)+2 (i=1), so its ordinal pattern is s1=(1,3,2). Repeating this step, we get the ordinal patterns S={(1,3,2),(3,1,2),(2,3,1),(3,2,1),(1,3,2),(3,2,1)} corresponding to all the embedding vectors obtained from the standardized time series partitioning. The ordinal pattern S={(0,6,5),(6,9,6),(6,5,6),(9,6,4),(5,6,5),(6,4,4)} corresponding to the embedding vectors obtained from the coarse-grained time series partitioning is the same as its embedding vector.

[0053] S23, construct the network according to the transformation order of the ordinal pattern. Since the embedding dimension m is 3, all nodes in the sequence partitioning network belong to V = {(1,2,3),(1,3,2),(2,1,3),(2,3,1),(3,1,2),(3,2,1)}. In S, the first sequence is {V2,V5}, therefore there is a link between the corresponding nodes V2 and V5, i.e., the adjacency matrix element a. 2,5 =1. In this example, only 4 nodes {V2,V4,V5,V6} appear. The other 2 nodes {V1,V3} do not exist. Therefore, the element a associated with node {V1,V3} is 1. i,j Filled with 0s, the absolute amplitude sequence partitioning network is constructed in the same way. The first continuous sequence in S indicates a link between nodes (0, 6, 5) and (6, 9, 6), and the element a associated with the missing node is... i,j Filled with 0.

[0054] S3, calculate the feature vectors of the sequence partitioning network and the absolute amplitude sequence partitioning network, and extract the ECG morphology and amplitude information. The feature vectors include average degree, average clustering coefficient, graph energy, transitivity, link density, average degree centrality, s-metric, permutation entropy, conditional permutation entropy, and global node entropy.

[0055] S4. The feature vectors from step S3 are combined into a hybrid amplitude feature. The hybrid amplitude feature is then classified into different ECG morphologies using a machine learning classifier. The hybrid amplitude feature includes segment morphology information from the sequence partitioning network and amplitude level information from the absolute amplitude sequence partitioning network. It is used as input to the support vector machine classifier for model training and testing.

[0056] Example 2

[0057] The ECG morphology recognition method based on a hybrid amplitude sequence partitioning network can be used to distinguish between normal heartbeats and premature ventricular contractions. The specific steps are as follows:

[0058] S1. The ECG signal is preprocessed by using a Butterworth bandpass filter of 0.1 to 45 Hz to remove baseline drift and high-frequency noise. The filtered signal is normalized to [-1,1]. The ECG signal is then segmented according to heart beats and cut into signal segments of equal length. Each signal segment contains a complete QRS wave.

[0059] S2, constructing a sequence partitioning network and an absolute amplitude sequence partitioning network corresponding to the time series, specifically including:

[0060] S21, the normalized time series and the coarse-grained time series are each divided into a set of embedding vectors. Each signal segment extracted in step S1 is a normalized time series x. Using the equal-interval univariate symbol method, the normalized time series x is converted into a symbol sequence through a unified quantization process. The ten symbols (0-9) are applied to the normalized time series, as shown in the formula:

[0061]

[0062] The first row of Figure 2 shows the standardized electrocardiograms of a normal heartbeat and two typical premature ventricular contractions. The left vertical axis represents the standardized time series values, and the right vertical axis represents the coarse-grained 10-symbol sequence values. With a time lag of τ = 1, the standardized time series is divided into a set of embedding vectors according to the embedding dimension m = 5, and the coarse-grained time series is divided into a set of embedding vectors according to the embedding dimension m = 2.

[0063] S22, Map the embedded vector to the ordinal pattern according to the amplitude order, based on the element amplitude order. embed vector v i Mapping to ordinal patterns s i =(π1,π2,π3,…,π) m ), where π k ∈{1,2,…,m}, and π when k≠l k ≠π l To ensure that the obtained ordinal pattern is unique, if vi If the values of two elements in k are equal, then the ordinals of these two elements are assigned according to the order of their appearance in the embedding vector. That is, when x l = x k , if k < l, then π l , for the embedding vectors obtained by dividing the standardized time series according to the embedding dimension m = 5, the number of all possible ordinal patterns is m! = 120; the coarse-grained time series, i.e., the symbol sequence y = {y i , i = 1, 2,..., N}, a series of embedding vectors v i = {y i = {y i , y i+τ , y i+2τ ,..., y i+(m-1)τ} of the symbol sequence y = {y i = (π1, π2, π3,..., π m ) directly correspond, that is, y k = π k , k = i + (m - 1)τ. For the embedding vectors obtained by dividing the coarse-grained time series according to the embedding dimension m = 2, the number of all possible ordinal patterns is 10 m = 100;

[0064] In S23, according to the ordinal patterns, an undirected unweighted sequence partition network and an absolute amplitude sequence partition network are constructed. The sequence partition network is constructed from the ordinal patterns mapped by the embedding vectors obtained from the standardized time series, and the absolute amplitude sequence partition network is constructed from the ordinal patterns mapped by the embedding vectors obtained from the coarse-grained time series. The process of constructing the network according to the ordinal pattern sequence is as follows: Denote all different ordinal pattern sets included in the ordinal pattern sequence as the set Each different ordinal pattern in the set corresponds to a node in the network G = {V, E}. The link order between nodes in the network depends on the order in which the ordinal patterns corresponding to the nodes appear in the ordinal pattern sequence . The edge between two nodes indicates that the ordinal patterns corresponding to these two nodes are adjacent in the set . E is the set of edges. The mathematical representation of G is an adjacency matrix A, and the value of its element a i,j represents whether there is a link between nodes i and j. When the embedding dimension m = 5, there are at most m! = 120 nodes in each sequence partition network, and when the embedding dimension m = 2, there are at most 10 m=100 nodes. The lower half of Figure 2 shows the sequence partitioning network and absolute amplitude sequence partitioning network corresponding to normal heartbeats and two typical premature ventricular contractions.

[0065] S3, calculate the feature vectors of the sequence partitioning network and the absolute amplitude sequence partitioning network, and extract the ECG morphology and amplitude information. The feature vectors include average degree, average clustering coefficient, graph energy, transitivity, link density, average degree centrality, s-metric, permutation entropy, conditional permutation entropy, and global node entropy.

[0066] The formula for calculating the average degree is as follows: Where N is the total number of nodes, d n Let n be the degree of node n, which is the number of neighboring edges that node n has;

[0067] The formula for calculating the average clustering coefficient is as follows: Among them, e i Represents the k connected to node i. i The number of actual edges between nodes, k i *(k i -1) represents all k nodes of node i. i The maximum possible number of edges between N neighboring nodes, where V is the set of N nodes;

[0068] The formula for calculating graph energy is: GE = ∑ i |λ i |, where λ i is the absolute value of the real part of the eigenvalues ​​of the adjacency matrix;

[0069] The formula for calculating transitivity is as follows: Where adj is the adjacency matrix, G(adj) is the number of connected triples in the graph, and F(adj) is the number of triangles in the graph;

[0070] The formula for calculating link density is as follows: Among them, e ij This represents the edge connecting node i and node j, where N is the total number of nodes;

[0071] The formula for calculating mean degree centrality is: Among them, e ij This represents the edge connecting node i and node j, where N is the total number of nodes;

[0072] The formula for calculating the s metric is as follows: Where, d i (d j ) represents the degree of node i(j), that is, the number of neighboring edges that node i(j) has;

[0073] The formula for calculating the permutation entropy is: PE = -∑i p i logp i , where s i The probability mass function P(s=s) ∈S i ) = p i It is estimated by calculating the relative frequency of each symbol in S;

[0074] The conditional permutation entropy is calculated using the formula: CPE = ∑ i (-p i ∑ j p i,j logp i,j ), where p i,j It is estimated from s based on S. i to s j The transition probability;

[0075] The formula for calculating global node entropy is as follows: in, It is the corrected random matrix P of an ordered network that excludes the possibility of self-looping. T elements, elementa i,j This represents the link between nodes i and j.

[0076] S4. Combine the feature vectors from step S3 into a mixed amplitude feature. Classify the mixed amplitude feature using a machine learning classifier to classify it into different ECG morphologies. The mixed amplitude feature includes segment morphology information from the sequence partitioning network and amplitude level information from the absolute amplitude sequence partitioning network. The ECG data is divided into a training set and a test set. The mixed amplitude features of normal and premature ventricular contractions in the training set are used as input to the support vector machine classifier to train the model. The classification effect of the model is tested using normal and premature ventricular contractions in the test set.

[0077] Example 3

[0078] The ECG morphology recognition method based on a hybrid amplitude sequence partitioning network can be used to distinguish between normal heartbeats and atrial flutter heartbeats. The specific steps are as follows:

[0079] S1. The ECG signal is preprocessed by using a Butterworth bandpass filter of 0.1 to 45 Hz to remove baseline drift and high-frequency noise. The filtered signal is normalized to [-1,1]. The ECG signal is then segmented according to heart beats and cut into signal segments of equal length. Each signal segment contains a complete QRS wave.

[0080] S2, constructing a sequence partitioning network and an absolute amplitude sequence partitioning network corresponding to the time series, specifically including:

[0081] S21. Divide the standardized time series and the coarse-grained time series into groups of embedding vectors respectively. Each signal segment intercepted in step S1 is a standardized time series x. Using the equidistant univariate symbol method, the standardized time series x is converted into a symbol sequence through a unified quantization process. Apply the ten symbols (0 - 9) to the normalized time series. The formula is

[0082]

[0083] Figure 3 shows the standardized electrocardiograms of normal heartbeats and atrial flutter heartbeats. The left vertical axis is the standardized time series value, and the right vertical axis is the coarse-grained ten-symbol sequence value. Take the time delay τ = 1. Divide the standardized time series into a group of embedding vectors according to the embedding dimension m = 6, and divide the coarse-grained time series into a group of embedding vectors according to the embedding dimension m = 3;

[0084] S22. Map the embedding vectors to ordinal patterns according to the amplitude order. According to the element amplitude order Map the embedding vector v i to the ordinal pattern s i =(π1,π2,π3,…,π m ), where π k ∈{1,2,…,m}, and when k≠l, π k ≠π l . To make the obtained ordinal pattern unique, if the values of two elements in v i are equal, then assign the ordinals of these two elements according to the order of their appearance in the embedding vector. That is, when x k =x l , if k < l, then π k <π l . For the embedding vectors divided from the standardized time series according to the embedding dimension m = 6, the number of all possible ordinal patterns is m! = 720; the coarse-grained time series, that is, the symbol sequence y={y i ,i = 1,2,…,N}, a series of embedding vectors v i ={y i ={y i ,y i+τ ,y i+2τ ) directly correspond to the ordinal pattern s i =(π_{1},π_{2},π_{3},…,π m ), that is, y [[ID=],i = 1,2,…,N}, a series of embedding vectors v i ={y i ,y i+τ ,y i+2τ ,…,y i+(m-1)τ} directly correspond to the ordinal pattern s i =(π1,π2,π3,…,π m ), that is, y k =π kFor an embedding vector derived from a coarse-grained time series partitioned by an embedding dimension m = 2, the number of all possible ordinal patterns is 10. m =1000;

[0085] In S23, an undirected, unweighted sequence partitioning network and an absolute amplitude sequence partitioning network are constructed based on ordinal patterns. The ordinal patterns of the embedding vector maps obtained from standardized time series partitioning are used to construct the sequence partitioning network, and the ordinal patterns of the embedding vector maps obtained from coarse-grained time series partitioning are used to construct the absolute amplitude sequence partitioning network. The specific process of constructing the network based on the ordinal pattern sequence is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The set of all distinct ordinal patterns included in the set is denoted as the set. gather Each distinct ordinal pattern in the network G = {V, E} corresponds to a node in the network. The order of links between nodes in the network depends on whether the node corresponds to an ordinal pattern in the ordinal pattern sequence. The order in which nodes appear in the set, and the edge between two nodes indicates the ordinal pattern of those two nodes in the set. In G, edges are adjacent, E is the set of connected edges, and G is mathematically represented as an adjacency matrix A, whose elements a i,j The value of represents whether there is a link between nodes i and j. When the embedding dimension m = 6, each sequence partitioning network has a maximum of m! = 720 nodes. When the embedding dimension m = 3, each absolute amplitude sequence partitioning network has a maximum of 10 m = 1000 nodes.

[0086] S3, calculate the feature vectors of the sequence partitioning network and the absolute amplitude sequence partitioning network, and extract the ECG morphology and amplitude information. The feature vectors include average degree, average clustering coefficient, graph energy, transitivity, link density, average degree centrality, s-metric, permutation entropy, conditional permutation entropy, and global node entropy.

[0087] The formula for calculating the average degree is as follows: Where N is the total number of nodes, d n Let n be the degree of node n, which is the number of neighboring edges that node n has;

[0088] The formula for calculating the average clustering coefficient is as follows: Among them, e i Represents the k connected to node i. i The number of actual edges between nodes, k i *(k i -1) represents all k nodes of node i. i The maximum possible number of edges between N neighboring nodes, where V is the set of N nodes;

[0089] The formula for calculating graph energy is: GE = ∑i |λ i |, where λ i is the absolute value of the real part of the eigenvalues ​​of the adjacency matrix;

[0090] The formula for calculating transitivity is as follows: Where adj is the adjacency matrix, G(adj) is the number of connected triples in the graph, and F(adj) is the number of triangles in the graph;

[0091] The formula for calculating link density is as follows: Among them, e ij This represents the edge connecting node i and node j, where N is the total number of nodes;

[0092] The formula for calculating mean degree centrality is: Among them, e ij This represents the edge connecting node i and node j, where N is the total number of nodes;

[0093] The formula for calculating the s metric is as follows: Where, d i (d j ) represents the degree of node i(j), that is, the number of neighboring edges that node i(j) has;

[0094] The formula for calculating the permutation entropy is: PE = -∑ i p i logp i , where s i The probability mass function P(s=s) ∈S i ) = p i It is estimated by calculating the relative frequency of each symbol in S;

[0095] The conditional permutation entropy is calculated using the formula: CPE = ∑ i (-p i ∑ j p i,j logp i,j ), where p i,j It is estimated from s based on S. i to s j The transition probability;

[0096] The formula for calculating global node entropy is as follows: in, It is the corrected random matrix P of an ordered network that excludes the possibility of self-looping. T elements, elementa i,j This represents the link between nodes i and j.

[0097] S4. Combine the feature vectors from step S3 into a mixed amplitude feature. Classify the mixed amplitude feature using a machine learning classifier to classify it into different ECG morphologies. The mixed amplitude feature includes segment morphology information from the sequence partitioning network and amplitude level information from the absolute amplitude sequence partitioning network. The ECG data is divided into a training set and a test set. The mixed amplitude features of normal and atrial flutter beats in the training set are used as input to the support vector machine classifier to train the model. The classification effect of the model is tested using normal and atrial flutter beats in the test set.

[0098] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. An electrocardiogram morphology recognition method based on a hybrid amplitude sequence partitioning network, characterized in that, Includes the following steps: S1, preprocess the electrocardiogram signal; S2, construct the sequence partitioning network and absolute amplitude sequence partitioning network corresponding to the time series; S3, calculate the feature vectors of the sequence partitioning network and the absolute amplitude sequence partitioning network, and extract ECG morphology and amplitude information. The feature vectors include average degree, average clustering coefficient, graph energy, transitivity, link density, average degree centrality, s-metric, permutation entropy, conditional permutation entropy, and global node entropy; S4, combine the feature vectors from step S3 into a hybrid amplitude feature, and classify the feature matrix into different ECG morphologies using a machine learning classifier; In step S2, constructing the sequence partitioning network and the absolute amplitude sequence partitioning network corresponding to the time series specifically includes: S21, dividing the standardized time series and the coarse-grained time series into a set of embedding vectors; S22, mapping the embedding vectors to ordinal patterns according to the amplitude order; S23, constructing the network according to the conversion order of the ordinal patterns; In step S21, the time series is converted into a set of embedding vectors using the embedding dimension and delay time, specifically: ,in For embedded dimensions, To delay time, For length is The one-dimensional time series; in step S21, the standardized time series is coarsened, specifically by using the equal-interval univariate sign method to normalize the time series to the range [-1,1]. The entire amplitude range is divided into Each region is associated with a specific symbolic value, based on the data points. In the region where it is located, each data point is replaced with a specific symbol, and the length is... time series Convert to symbolic sequence The conversion formula is: ,in, It is the selected set of symbols; in step S22, according to the element magnitude order embed vector Mapping to ordinal pattern ,in, And when hour ,if If two elements have the same value, then the rank of these elements is assigned according to their order of appearance in the vector. hour, Coarse-grained time series, i.e., symbolic sequences Embedded vector Directly corresponds to ordinal pattern ,Right now In step S23, an undirected, unweighted sequence partitioning network and an absolute amplitude sequence partitioning network are constructed based on the ordinal pattern. Each element represents a network In a network, G is a node, E is the set of edges, and the ordinal pattern transformation determines the links between nodes. G is mathematically represented as an adjacency matrix. Its elements Representative node and Links between them.

2. The ECG morphology recognition method based on a hybrid amplitude sequence partitioning network according to claim 1, characterized in that, In step S1, a Butterworth bandpass filter with a frequency of 0.1~45Hz is used to remove baseline drift and high-frequency noise, and the filtered signal is normalized to [-1,1].

3. The ECG morphology recognition method based on a hybrid amplitude sequence partitioning network according to claim 1, characterized in that, The formula for calculating the average degree in step S3 is as follows: Where N is the total number of nodes. For nodes The degree; the formula for calculating the average clustering coefficient in step S3 is: ,in, Represents nodes Connected The actual number of edges between nodes It is a node all The maximum possible number of edges between N neighboring nodes, where V is the set of N nodes; the formula for calculating graph energy in step S3 is: ,in, The value is the absolute value of the real part of the eigenvalues ​​of the adjacency matrix; the formula for calculating transitivity in step S3 is: ,in, It is an adjacency matrix. This represents the number of connected triples in the graph. The number of triangles in the diagram; the formula for calculating the link density in step S3 is: ,in, Represents a node and nodes The edges connecting the nodes are N, where N is the total number of nodes; the formula for calculating the average degree centrality in step S3 is: ,in, Represents a node and nodes The edges connecting the nodes, where N is the total number of nodes; the formula for calculating the s metric in step S3 is: ,in, ( ) represents a node ( The degree of permutation entropy in step S3 is calculated using the following formula: ,in, probability mass function Through calculation It is estimated by the relative frequency of each symbol; the formula for calculating the conditional permutation entropy in step S3 is: ,in, It is based on The estimated from arrive The conversion probability; the formula for calculating the global node entropy in step S3 is: ,in, It is a corrected random matrix of an ordered network that excludes the possibility of self-looping. elements, ,element Representative node and Links between them.

4. The ECG morphology recognition method based on a hybrid amplitude sequence partitioning network according to claim 1, characterized in that, In step S4, the network-based hybrid amplitude features include fragment morphology information from the sequence partitioning network and amplitude information from the absolute amplitude sequence partitioning network, which are input into a machine learning classifier for classification.

Images