Electrocardiosignal feature extraction method based on bispectrum analysis and graph Fourier transform
阅读说明：本技术 基于双谱分析和图傅里叶变换的心电信号特征提取方法 (Electrocardiosignal feature extraction method based on bispectrum analysis and graph Fourier transform ) 是由 胡久元 邵杰 刘姝 孔天姣 周凡 于 2020-04-27 设计创作，主要内容包括：本发明公开了一种基于双谱分析和图傅里叶变换的心电信号特征提取方法,属于心电信号特征提取方法技术领域。本方法首先利用高阶谱算法将时域信号转换到高阶谱域,再利用图傅里叶变换,将双谱矩阵转换至特征值域,然后从特征谱矩阵中直接提取图谱特征在得到特征向量后可以根据AAMI标准,对图谱特征进行分类。本方法克服了现有双谱矩阵特征提取方法计算量大的问题,并且准确度高,能够对各种心电信号进行有效分析。(The invention discloses an electrocardiosignal feature extraction method based on bispectrum analysis and graph Fourier transform, and belongs to the technical field of electrocardiosignal feature extraction methods. The method comprises the steps of firstly converting a time domain signal into a high-order spectrum domain by using a high-order spectrum algorithm, then converting a double-spectrum matrix into a characteristic value domain by using graph Fourier transform, and then directly extracting graph features from the characteristic spectrum matrix, and classifying the graph features according to an AAMI standard after obtaining feature vectors. The method solves the problem of large calculation amount of the existing bispectrum matrix characteristic extraction method, has high accuracy, and can effectively analyze various electrocardiosignals.)
1. An electrocardiosignal feature extraction method based on bispectrum analysis and graph Fourier transform is characterized by comprising the following steps:
step A, preprocessing an electrocardiosignal, removing noise in the signal by using a filtering method, and separating to obtain single heartbeat data;
b, performing high-order spectral analysis on the electrocardiosignal, and calculating to obtain a bispectrum matrix S of the signal;
step C, mapping the bispectrum matrix to a characteristic value domain by utilizing two-dimensional graph Fourier transform;
step D, the acquired characteristic value spectrum is subjected toAnd extracting the map features and forming feature vectors including the map flatness, the map brightness and the map roll-off degree.
2. The method for extracting features of electrocardiosignals based on bispectrum analysis and graph Fourier transform as claimed in claim 1, wherein the specific process of step A is as follows:
step (A-1), subtracting the average value from the original electrocardiosignal to eliminate the direct current component;
eliminating baseline drift through a median filter;
eliminating power frequency interference and electromyographic noise by a low-pass filter;
step (A-4) eliminating low-frequency noise through a high-pass filter;
and (A-5) searching the position of a peak R wave in the QRS complex of the signal, taking a peak point as an origin, and taking K data before and after the point to form single heartbeat data h (t) with the length of 2K points.
3. The method for extracting features of electrocardiosignals based on bispectrum analysis and graph Fourier transform as claimed in claim 1, wherein the specific process of step B is as follows:
step (B-1) of calculating the third-order accumulation amount of single heartbeat data h (t)
R3s(τ1,τ2)＝E[h(t)h(t+τ1)h(t+τ2)]，τ1，τ2∈(-∞，∞)
Wherein: h (t) is single beat data, τ1And τ2Is a time delay; h (t + τ)1) And h (t + τ)2) To pass a time delay of tau1，τ2The later time shift signal finally obtains a third-order cumulative quantity R3s(τ1,τ2)；
Step (B-2) of calculating bispectrum of single heartbeat data h (t)
Wherein S is a bispectral matrix of h (t) (. omega.)1，ω2Are two independent frequencies.
4. The method for extracting features of electrocardiosignals based on bispectrum analysis and graph Fourier transform as claimed in claim 1, wherein the specific process of step C is as follows:
step (C-1), calculating a diagonal matrix D, wherein a calculation formula of a point D (m, n) is as follows:
the diagonal matrix D is formed by combining D (m, n), and w (i, j) is the connection relation between the nodes i and j;
step (C-2) of calculating Laplace matrix L
L＝D-W
In the formula: w is a weighted adjacency matrix and is formed by combining elements W (m, n), wherein the W (m, n) shows the connection relation between the node m and the node n, and the size of the W (m, n) is related to the distance between the nodes; l is a real symmetric matrix;
step (C-3) of calculating Laplace eigenvalue λ corresponding to LlAnd laplacian feature matrix Xl，λlAnd XlSatisfies the following conditions:
LXl＝λlXl,l＝0,1,...,N-1
wherein x isl(m, n) is a matrix XlThe m-th row and the n-th column;
step (C-4) calculating Fourier transform of two-dimensional graph of bispectrum matrix S
Obtaining an N-order square matrix Is the element of the m-th row and the n-th column in the matrix, xl(i, j) is a matrix XlAnd S (i, j) is an element in the ith row and the jth column of the dual spectrum matrix S.
5. The method for extracting features of electrocardiosignals based on bispectrum analysis and graph Fourier transform as claimed in claim 1, wherein the specific process of step D is as follows:
step (D-1) calculating map flatness GSF
Where N is the number of points in the characteristic spectrum matrix,is the element of the p row and q column in the matrix G;
step (D-2) calculating map brightness GSB
F is a given boundary characteristic value;
step (D-3) calculating the map roll-off degree GSR
Where β is a coefficient;
finally, feature vectors are formed by the feature values, and the atlas features are classified according to the AAMI standard.
Technical Field
The invention relates to an electrocardiosignal feature extraction method based on bispectrum analysis and graph Fourier transform, and belongs to the technical field of electrocardiosignal feature extraction methods.
Background
Cardiovascular diseases have gradually become one of the most common diseases endangering human life, so how to diagnose and prevent such diseases has become an important problem facing the medical community today. The electrocardiosignals can be used for analyzing and identifying various diseases such as arrhythmia, myocardial infarction and the like, can reflect the damage degree and the development process of myocardial cells, the functional structures of atria and ventricles and the like, and become a simple and effective tool in cardiovascular disease diagnosis.
The high-order spectrum of the signal is a non-stationary signal analysis tool, and the bispectrum, which is the lowest order in the high-order spectrum, not only contains all the characteristics of the high-order spectrum, but also has the simplest calculation, so the bispectrum is widely applied. In actual operation, the bispectrum matrix is relatively complex, and most of the feature extraction methods need further dimension reduction processing such as contour integration, bispectrum slicing, principal component analysis, independent component analysis, kernel principal component analysis and the like on the basis of obtaining the bispectrum matrix.
As a very important research direction in algebraic graph theory, graph theory has begun to develop as early as the fifties of the last century. The map theory is a basic idea of the map theory, which is simply to establish a corresponding relationship between a map and a matrix. Thus, after the problem of the graph is transferred to the memory matrix, the problem of the graph can be researched by researching the relevant attributes of the matrix spectrum. The memory matrix mainly includes a adjacency matrix and a laplacian matrix. The study of adjacency matrices has a long history and is a relatively mature discipline. More studies have emerged in recent years on the laplacian matrix spectra of the graphs than on the adjacency matrices. And graph signal processing derived from graph theory is also an application of rapid development in recent years. The graph fourier transform is the expansion of the graph signal with respect to the characteristic function of the graph laplace matrix, and is also the basis of the graph signal processing.
The bispectrum matrix established on the bispectrum basis is converted into a characteristic value domain through graph Fourier transform, characteristics are extracted from statistic of the characteristic value, and the method for constructing the characteristic vector by utilizing the graph characteristics is simple in calculation and high in distinguishability.
Disclosure of Invention
In order to solve the defects of the prior art, the invention provides the electrocardiosignal feature extraction method based on the bispectrum analysis and graph Fourier transform, which not only solves the problem of large calculation amount of the traditional bispectrum matrix feature extraction method, but also has high accuracy, can effectively analyze various electrocardiosignals and realize the effective diagnosis of the cardiovascular diseases such as arrhythmia and the like.
The invention adopts the following technical scheme for solving the technical problems:
an electrocardiosignal feature extraction method based on bispectrum analysis and graph Fourier transform comprises the following steps:
step A, preprocessing an electrocardiosignal, removing noise in the signal by using a filtering method, and separating to obtain single heartbeat data;
b, performing high-order spectral analysis on the electrocardiosignal, and calculating to obtain a bispectrum matrix S of the signal;
step C, mapping the bispectrum matrix to a characteristic value domain by utilizing two-dimensional graph Fourier transform;
step D, the acquired characteristic value spectrum is subjected toAnd extracting the map features and forming feature vectors including the map flatness, the map brightness and the map roll-off degree.
The specific process of the step A is as follows:
step (A-1), subtracting the average value from the original electrocardiosignal to eliminate the direct current component;
eliminating baseline drift through a median filter;
eliminating power frequency interference and electromyographic noise by a low-pass filter;
step (A-4) eliminating low-frequency noise through a high-pass filter;
and (A-5) searching the position of a peak R wave in the QRS complex of the signal, taking a peak point as an origin, and taking K data before and after the point to form single heartbeat data h (t) with the length of 2K points.
The specific process of the step B is as follows:
step (B-1) of calculating the third-order accumulation amount of single heartbeat data h (t)
R3s(τ1,τ2)＝E[h(t)h(t+τ1)h(t+τ2)]，τ1，τ2∈(-∞，∞)
Wherein: h (t) is single beat data, τ1And τ2Is a time delay; h (t + τ)1) And h (t + τ)2) To pass a time delay of tau1，τ2The later time shift signal finally obtains a third-order cumulative quantity R3s(τ1,τ2)；
Step (B-2) of calculating bispectrum of single heartbeat data h (t)
Wherein S is a bispectral matrix of h (t) (. omega.)1，ω2Are two independent frequencies.
The specific process of the step C is as follows:
step (C-1), calculating a diagonal matrix D, wherein a calculation formula of a point D (m, n) is as follows:
the diagonal matrix D is formed by combining D (m, n), and w (i, j) is the connection relation between the nodes i and j;
step (C-2) of calculating Laplace matrix L
L＝D-W
In the formula: w is a weighted adjacency matrix and is formed by combining elements W (m, n), wherein the W (m, n) shows the connection relation between the node m and the node n, and the size of the W (m, n) is related to the distance between the nodes; l is a real symmetric matrix;
step (C-3) of calculating Laplace eigenvalue λ corresponding to LlAnd laplacian feature matrix Xl，λlAnd XlSatisfies the following conditions:
LXl＝λlXl,l＝0,1,...,N-1
wherein x isl(m, n) is a matrix XlThe m-th row and the n-th column;
step (C-4) calculating Fourier transform of two-dimensional graph of bispectrum matrix S
Obtaining an N-order square matrixIs the element of the m-th row and the n-th column in the matrix, xl(i, j) is a matrix XlAnd S (i, j) is an element in the ith row and the jth column of the dual spectrum matrix S.
The specific process of the step D is as follows:
step (D-1) calculating map flatness GSF
Where N is the number of points in the characteristic spectrum matrix,is the element of the p row and q column in the matrix G;
step (D-2) calculating map brightness GSB
F is a given boundary characteristic value;
step (D-3) calculating the map roll-off degree GSR
Where β is a coefficient.
Finally, feature vectors are formed by the feature values, and the atlas features are classified according to the AAMI standard.
The invention has the following beneficial effects:
1. on the basis of preprocessing, the invention adopts bispectrum as the first step of feature extraction, and theoretically eliminates the influence of Gaussian noise by utilizing the excellent properties of the bispectrum.
2. And Fourier transform of the two-dimensional graph is used as a second step, the dual-spectrum matrix is converted into a characteristic value spectrum, the difference of various characteristics is increased, and the distinguishability is improved. The features are directly extracted on the feature value spectrum, so that errors and calculated amount caused by secondary extraction of the features are avoided.
Drawings
Fig. 1 is a block diagram of an electrocardiographic signal feature extraction step.
Fig. 2(a) is a time domain waveform diagram of record No. 100 before the single heartbeat signal preprocessing, and fig. 2(b) is a time domain waveform diagram of record No. 100 after the single heartbeat signal preprocessing.
Fig. 3 is a single beat dual spectrum of the cardiac signal recorded No. 100.
Fig. 4 is a feature value map of a single heartbeat recorded at No. 100 after fourier transform of a two-dimensional map.
Fig. 5(a) is a box diagram of features extracted after N-class electrocardiographic signals are subjected to fourier transform of a two-dimensional graph and converted into a feature value domain, fig. 5(b) is a box diagram of features extracted after S-class electrocardiographic signals are subjected to fourier transform of a two-dimensional graph and converted into a feature value domain, and fig. 5(c) is a box diagram of features extracted after V-class electrocardiographic signals are subjected to fourier transform of a two-dimensional graph and converted into a feature value domain.
Detailed Description
The invention is further described with reference to the accompanying drawings. Taking the data in the MIT-BIH database as an example, the embodiment comprises the following steps:
and step A, preprocessing the electrocardiosignal, removing noise in the signal by using various filter methods, and separating to obtain single heartbeat data.
And (A-1) subtracting the average value from the original electrocardiosignal to eliminate the direct current component.
And (A-2) eliminating the baseline drift through a median filter.
And (A-3) eliminating power frequency interference and electromyographic noise by a low-pass filter.
And (A-4) eliminating low-frequency noise through a high-pass filter.
And (A-5) searching the position of a peak R wave in a signal QRS (QRS wave represents potential change of two ventricular excitation propagation processes) complex. Taking the peak point as the origin, K before and after the point is 100 data, and forming a single heartbeat data h (t) with the length of 200 points.
Fig. 1 is a block diagram illustrating a signal feature classification process. The method comprises four steps, wherein signals of a single heart beat are obtained through preprocessing, then a corresponding bispectrum matrix is obtained through bispectrum calculation, then two-dimensional graph Fourier transform is adopted, and finally, graph features are extracted.
As shown in FIG. 2, the electrocardiosignal pre-processing electrocardiosignal pre-processing recorded in the database with the sampling frequency of 360 Hz.
And B, performing high-order spectral analysis on the electrocardiosignal, and calculating to obtain a bispectrum matrix S of the signal, wherein the mth row and the nth column of elements in the matrix are S (m, n).
And (B-1) calculating the third-order accumulation amount of the single heartbeat data h (t).
R3s(τ1,τ2)＝E[h(t)h(t+τ1)h(t+τ2)]，τ1，τ2∈(-∞，∞)
Wherein: h (t) is single beat data, τ1And τ2Is a time delay; h (t + τ)1) And h (t + τ)2) To pass a time delay of tau1，τ2The later time shift signal finally obtains a third-order cumulative quantity R3s(τ1,τ2)；
And (B-2) calculating the bispectrum of the single heartbeat data h (t).
Wherein S is a bispectral matrix of h (t) (. omega.)1，ω2Are two independent frequencies.
A result of the bispectrum analysis of the single heartbeat data signal is obtained, and as shown in fig. 3, a single heartbeat bispectrum of the electrocardiographic signal recorded in the 100 th record in the database is obtained.
And C, mapping the bispectral matrix to a characteristic value domain by utilizing two-dimensional graph Fourier transform. In the two-dimensional graph Fourier transform, an undirected graph G (V, E, W) is shown, where V represents a set of vertices and E represents a set of edges. W is a weighted adjacency matrix and is formed by combining elements W (m, n), wherein the W (m, n) shows the connection relation between the node m and the node n, and the size is related to the distance between the nodes. Here, the following are selected:
wherein m, N is 1, 2.
And (C-1) calculating a diagonal matrix D. The calculation formula for the point d (m, n) is:
the diagonal matrix D is formed by combining D (m, n), and w (i, j) is the connection relation between the nodes i and j.
And (C-2) calculating a Laplace matrix L.
L＝D-W
Where L is a real symmetric matrix.
Step (C-3) of calculating a Laplacian eigenvalue λ corresponding to LlAnd laplacian feature matrix Xl，λlAnd XlSatisfies the following conditions:
LXl＝λlXl,l＝0,1,...,N-1
xl(m, n) is a matrix XlRow m and column n.
And (C-4) calculating the Fourier transform of the two-dimensional graph of the bispectrum matrix S.
Obtaining an N-order characteristic spectrum matrixxl(i, j) is a matrix XlAnd S (i, j) is an element in the ith row and the jth column of the dual spectrum matrix S.
A graph fourier transform result of the single heart beat data double spectrogram is obtained, and as shown in fig. 4, the graph is a characteristic value spectrogram of the single heart beat data double spectrogram recorded in the 100 th record in the database after the graph fourier transform.
Step D, the acquired characteristic value spectrum is subjected toAnd extracting the feature of the map, and forming a feature vector.
The method comprises the following steps: map flatness (GSF), map lightness (GSB), and map roll-off (GSR). Step (D-1) of calculating the flatness of the map
Where N is the number of points in the feature spectrum matrix, where N equals 100,is the element in the p-th row and q-th column of the matrix G.
Step (D-2) of calculating the map brightness
F is the given boundary feature value, here taken to be 25.
Step (D-3) calculating the map roll-off degree
Where β is a coefficient, selected to be 0.3.
Obtaining the atlas characteristics of a single heartbeat data signal in a characteristic value range, and fig. 5 is a box-type distribution diagram of atlas flatness, atlas brightness and atlas roll-off degree of four types of electrocardiosignals obtained by dividing an MIT-BIH database according to the Association for the Advancement of medical instrumentation (AAMI) standard, wherein: class N (normal or bundle branch block beats) 67994 in total; 2577S (supraventricular abnormal beats) in total; 4249 in total for class V (ventricular abnormal beats); 784 total F classes (fusion beats). As can be seen from the figure, the graph flatness and the graph roll-off degree characteristics of the four types of signals are obviously distinguished. Finally, feature vectors are formed by the feature values, and the atlas features are classified according to the AAMI standard.