Computer-aided classification method for heart disease based on statistical manifold curvature

Abstract

The present invention relates to a computer-aided classification method for heart disease based on statistical manifold curvature, and belongs to the field of signal processing technology and computer-aided diagnosis. The method uses delayed embedding and k-nearest neighbors to convert a one-dimensional electrocardiogram time series signal into a point cloud on a positive definite matrix manifold. It is proposed to study the local structural differences of the point cloud distribution by calculating the quantitative curvature of each point and extracting the geometric feature of the curvature discreteness, thereby realizing the classification of the original electrocardiogram. The method has high classification accuracy and strong robustness, reduces information loss and eliminates shortcomings such as poor interpretability, fully ensuring the application effect. The method can accurately classify a large number of electrocardiograms, while greatly reducing the amount of calculation and ensuring high efficiency. The method does not rely on the selection of parameters and has significant advantages. In addition to being applicable to the field of electrocardiogram signal classification, the method also has broad application prospects in fields such as biomedical research and signal processing.

Description

Technical Field

[0001] The present invention relates to a computer-assisted classification method for pre-classifying abnormal heart disease conditions based on statistical manifold curvature, and belongs to the fields of signal processing technology and computer-assisted diagnosis. Background Art

[0002] Heart disease, one of the most common and fatal diseases worldwide, poses a significant threat to human well-being and health. With increasing social pressures, the incidence of heart disease continues to climb annually. Therefore, efficient diagnosis, real-time monitoring, and prediction of heart disease are crucial for its treatment and prevention. Currently, heart disease diagnosis is primarily performed by doctors analyzing electrocardiograms (ECGs). Medical equipment displays the electrical activity of the patient's heart contractions, and doctors provide a diagnosis based on this information. ECG analysis requires specialized medical personnel and detailed medical knowledge, and medical resources are unevenly distributed. In the absence of adequate medical resources, patients' lives and health cannot be fully protected.

[0003] With the rapid development of computer technology, various artificial intelligence technologies are gradually playing a vital role in human life. Using computer-assisted diagnosis to assist in the classification and prediction of electrocardiograms (ECGs) has become an emerging technology to alleviate the imbalance of medical resources and help doctors formulate medical plans.

[0004] Currently, computer-assisted diagnosis (CAD) techniques for ECG analysis primarily focus on single-lead and 12-lead ECGs, focusing on signal processing, dynamical systems, statistical analysis, machine learning, and topological data analysis. ECG classification is achieved through feature extraction and training. However, most existing methods suffer from bottlenecks such as over-reliance on parameters and poor interpretability. Parameter adjustment requires experienced programmers, and improper parameter selection can significantly reduce the accuracy of results. Furthermore, most methods lack a reasonable medical explanation. This significantly hinders the application of CAD. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the existing technology and to solve the technical problems faced by computer-aided diagnosis systems in electrocardiogram classification, such as weak robustness, incomplete signal structured information extraction, poor interpretability, and strong parameter dependence. A computer-aided classification method for heart disease based on statistical manifold curvature is creatively proposed.

[0006] The innovation of this method lies in its use of delayed embedding and k-nearest neighbors to transform a one-dimensional electrocardiogram (ECG) time series signal into a point cloud on a positive definite matrix manifold. Furthermore, it proposes, for the first time, a method that calculates the scalar curvature of each point and extracts the geometric feature of curvature dispersion to investigate the local structural differences in the point cloud distribution, thereby enabling classification of the original ECG.

[0007] The present invention is implemented by adopting the following technical solutions.

[0008] The continuous ECG time series signal can be preprocessed by interpolation and filtering, including signal segmentation and noise reduction.

[0009] Then, a single ECG is transformed into a point cloud on the positive definite matrix manifold by utilizing delayed embedding and k-nearest neighbors.

[0010] Afterwards, the quantitative curvature of each point in the point cloud is calculated and the curvature histogram is obtained.

[0011] Then, the curvature dispersion feature of the curvature histogram is extracted and clustered.

[0012] Finally, based on the clustering results, the system calculates the curvature dispersion of the newly given electrocardiogram and performs auxiliary diagnosis by referring to the classification criteria in the clustering results.

[0013] Beneficial effects

[0014] Compared with the existing technology, this method has the following advantages:

[0015] 1. This novel geometry-based method offers strong interpretability for ECG signal classification. Compared to machine learning and other techniques, this method maintains accuracy while being robust and requiring no parameter tuning. Compared to techniques such as topological data analysis, this method can process larger amounts of data and achieve higher classification accuracy.

[0016] 2. This method achieves high classification accuracy and robustness, reduces information loss, and eliminates shortcomings such as poor interpretability, fully guaranteeing its effectiveness. Local statistics can reflect the local structural differences at each point in the point cloud. The curvature dispersion extracted from the curvature histogram can further highlight these differences in spatial structure.

[0017] 3. This method can accurately classify a large number of ECGs while significantly reducing computational effort and ensuring high efficiency. Because curvature is a geometric invariant, this method is independent of parameter selection, offering significant advantages.

[0018] 4. In addition to being applied in the field of ECG signal classification, this method also has broad application prospects in biomedical research and signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is an overall flow chart of the method of the present invention;

[0020] Figure 2 These are examples of the seven types of ECGs classified by this method;

[0021] Figure 3 is a schematic diagram of the noise reduction process in the present invention;

[0022] Figures 4 to 7 They are respectively the image of a normal electrocardiogram, the Euclidean space point cloud image, the grayscale image and the curvature histogram of the Wasserstein distance matrix of the point cloud on the positive definite matrix manifold in the present invention;

[0023] Figures 8 to 11 They are respectively the image of the left bundle branch block electrocardiogram, the Euclidean space point cloud image, the grayscale image and the curvature histogram of the point cloud Wasserstein distance matrix on the positive definite matrix manifold in the present invention;

[0024] Figures 12 to 15 They are respectively an image of the right bundle branch block electrocardiogram, a Euclidean space point cloud image, a grayscale image of the Wasserstein distance matrix of the point cloud on the positive definite matrix manifold, and a curvature histogram in the present invention;

[0025] Figures 16 to 19 They are respectively an image of atrial premature beat electrocardiogram, a Euclidean space point cloud image, a grayscale image and a curvature histogram of the point cloud Wasserstein distance matrix on the positive definite matrix manifold in the present invention;

[0026] Figures 20 to 23 They are respectively the image of the ventricular premature beat electrocardiogram, the Euclidean space point cloud image, the grayscale image and the curvature histogram of the point cloud Wasserstein distance matrix on the positive definite matrix manifold in the present invention;

[0027] Figures 24 to 27 They are respectively an image of the ventricular fusion heartbeat electrocardiogram of the present invention, a Euclidean space point cloud image, a grayscale image and a curvature histogram of the Wasserstein distance matrix of the point cloud on the positive definite matrix manifold;

[0028] Figures 28 to 31 They are respectively an image of the electrocardiogram of ventricular flutter in the present invention, a Euclidean space point cloud image, a grayscale image and a curvature histogram of the Wasserstein distance matrix of the point cloud on the positive definite matrix manifold;

[0029] Figure 32 is a schematic diagram of the curvature dispersion distribution of the original electrocardiogram data set in the present invention;

[0030] Figure 33 It is a schematic diagram of the segmentation of curvature discreteness by the support vector machine SVM in the present invention. DETAILED DESCRIPTION

[0031] The method of the present invention is further described in detail below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, a computer-aided classification method for heart disease based on statistical manifold curvature includes the following steps:

[0033] Step 1: If Figure 2 As shown, in this embodiment, main ECG signals including auxiliary classification of healthy, left bundle branch block, right bundle branch block, atrial premature beats, ventricular premature contractions, ventricular flutter and ventricular fusion heartbeat are included.

[0034] First, the continuous ECG signal is preprocessed by interpolation and filtering, such as noise reduction and segmentation.

[0035] For example, since the frequency of AC power used in my country's power supply equipment is 50Hz, in order to reduce the impact of noise, we can first use a Butterworth filter to filter and reduce noise on ECG frequencies above 50Hz.

[0036] Then, the R wave position of each heartbeat is detected, and a fixed ratio of points before and after the R wave (for example, 1 / 3 before the wave and 2 / 3 after the wave) are selected to form a single heartbeat. Interpolation is further used to uniformly perform scaling transformation to ensure that the amount of information for each heartbeat is consistent (for example, 300 points per heartbeat). Figure 3 shown.

[0037] Finally, continuous heartbeats are segmented according to the position of the R wave.

[0038] Step 2: Convert the ECG signal into a point cloud in Euclidean space using delayed embedding.

[0039] Specifically, the embedding step size τ and the embedding dimension d are set to transform each heartbeat E with N electrocardiogram signal points into a point cloud in a d-dimensional Euclidean space with N-(d-1)τ points.

[0040] Step 3: Use k-nearest neighbors to transform the point cloud in the d-dimensional Euclidean space obtained in step 2 into a point cloud on the positive definite matrix manifold.

[0041] Specifically, the number of neighbors k of each point in the point cloud is set, and the mean and covariance of each point and the nearest k points around it are calculated, thereby obtaining a point cloud in a d(d+1) / 2-dimensional positive definite matrix manifold.

[0042] Step 4: Assign the Wasserstein metric to the positive definite matrix manifold, calculate the quantitative curvature of each point in the point cloud in the positive definite matrix manifold, and obtain the curvature histogram.

[0043] Specifically, for a positive definite matrix A, its scalar curvature is given by:

[0044] 3tr(UD(U+U T )+(U+U T )DU+(U+U T )DUD(U+U T ))

[0045] Where tr represents the trace of the matrix, that is, the sum of the diagonal elements of the matrix, D = diag(λ1,…,λ n ) is orthogonal to A,λ n represents the nth eigenvalue of matrix A; U is an upper triangular matrix, the position of its i-th row and j-th column is 1 / (λ i +λ j ), T represents matrix transpose.

[0046] like Figures 4 to 31 As shown in the figure, the 28 figures respectively represent the images of 7 types of electrocardiograms, the point cloud images in Euclidean space, the grayscale images of the Wasserstein distance matrix of the point cloud on the positive definite matrix manifold, and their respective curvature histograms.

[0047] Step 5: Extract the curvature dispersion from the curvature map and use support vector machines (SVM) to extract the segmentation line for segmentation.

[0048] Specifically, the first component of the curvature dispersion can reflect the distribution of curvature in the curvature histogram, and the second component can reflect the difference in the number of each curvature in the curvature histogram. The curvature dispersion of the original electrocardiogram is shown as follows: Figure 32 shown.

[0049] Finally, the support vector machine with linear kernel is used to obtain the secant line, which can be used to classify the new ECG signal.

[0050] Example verification

[0051] This method was applied to a dataset of 19,926 heartbeats from the MIT-BIH heartbeat database, including 16,102 normal heartbeats, 1,312 left bundle branch block heartbeats, 614 right bundle branch block heartbeats, 244 atrial premature beats, 1,235 ventricular premature beats, 303 ventricular fusion heartbeats, and 116 ventricular flutter heartbeats. Figure 33 As shown, the accuracy of each category reaches more than 99%.

Claims

1. A computer-assisted classification method for heart disease based on statistical manifold curvature, characterized by: Firstly, for the continuous ECG time series signal, delayed embedding and k-nearest neighbor are used to transform a single ECG into a point cloud on the positive definite matrix manifold; Specifically, it uses delayed embedding to convert the ECG signal into a point cloud in Euclidean space; Set the embedding step size τ and embedding dimension d, and transform each heartbeat E with N electrocardiogram signal points into a point cloud in d-dimensional Euclidean space with N-(d-1)τ points. Using k-nearest neighbors, the point cloud in the d-dimensional Euclidean space is converted into a point cloud on a positive definite matrix manifold. The number of neighbors of each point in the point cloud is set to k, and the mean and covariance of each point and its nearest k points are calculated to obtain a point cloud on a d(d+1) / 2-dimensional positive definite matrix manifold. After that, the quantitative curvature of each point in the point cloud is calculated and the curvature histogram is obtained; By assigning the Wasserstein metric to the positive definite matrix manifold, the quantitative curvature of each point in the point cloud in the positive definite matrix manifold is calculated, and then the curvature histogram is obtained, as follows: For a positive definite matrix a, its scalar curvature is given by: 3tr(UD(U+U T )+(U+U T )DU+(U+U T )DUD(U+U T )) Where tr represents the trace of the matrix, that is, the sum of the diagonal elements of the matrix, D = diag(λ1,…,λ n ) is orthogonal to A,λ n represents the nth eigenvalue of matrix A; U is an upper triangular matrix, the position of its i-th row and j-th column is 1 / (λ i +λ j ), T represents matrix transpose; Then, the curvature dispersion features of the curvature histogram are extracted and clustered; Specifically, the curvature discreteness is extracted from the curvature map, and the segmentation line is extracted using the support vector machine (SVM) for segmentation. The method is as follows: The first component of the curvature dispersion can reflect the distribution of curvature in the curvature histogram, and the second component can reflect the difference in the number of each curvature in the curvature histogram; the secant line is obtained by using the support vector machine with a linear kernel, and the secant line is used to classify the new ECG signal; Finally, based on the clustering results, the system calculates the curvature dispersion of the newly given electrocardiogram and performs auxiliary diagnosis by referring to the classification criteria in the clustering results.

2. The computer-aided classification method for heart disease based on statistical manifold curvature according to claim 1, characterized in that: Through interpolation and filtering, the continuous ECG signal is first preprocessed, including noise reduction and segmentation.

3. The computer-aided classification method for heart disease based on statistical manifold curvature according to claim 2, wherein: The ECG signal segmentation method is to detect the position of the R wave of each heartbeat, select a fixed ratio of points before and after the R wave to form a single heartbeat, and further use interpolation to uniformly perform scaling transformation to ensure that the amount of information of each heartbeat is consistent; finally, the continuous heartbeats are segmented according to the position of the R wave.

Images