Electrocardiogram electrocardiosignal classification method with multi-scale characteristics combined

[0018] The following describes in detail the fast multi-scale feature fusion technology method provided by the present invention to improve the classification accuracy of heartbeat signals in conjunction with the accompanying drawings.

[0019] figure 1 The flow chart of the heartbeat signal classification method includes the following steps:

[0020] Step S101: Read all the ECG signals in the database.

[0021] Step S102, removing the baseline and high frequency noise of the signal.

[0022] (1) Apply 200ms and 600ms bandwidth median filters to remove ORS complexes and P&T waves, and subtract them from the original signal to obtain a baseline-removed signal.

[0023] (2) Apply a low-pass filter to remove high-frequency noise in the signal.

[0024] Step S103, ECG signal segmentation: first find the reference point (ie R point) of the ECG signal, and use the reference point forward 99 sampling points and backward 100 sampling points as the segmented heartbeat signal for one cycle.

[0025] Step S104, by calculating the wavelet packet decomposition of the ECG signal, the fourth layer decomposition coefficient is obtained as the extracted ECG signal feature.

[0026] Step S105, training the three-dimensional volume data arrangement of the heartbeat signal. Traditional feature fusion and dimensionality reduction methods, such as principal component analysis and independent component analysis, require the input data to be expanded into a column vector before training. The invention adopts a generalized multi-dimensional independent component analysis method to perform feature fusion. After the heartbeat signal is decomposed in four layers of wavelet packet, the coefficients of each group in the last layer can be combined into a second-order tensor, that is, a matrix, so that a series of training heartbeat signals can be arranged into a third-order tensor, and Body data. That is, the first two modes are the number of features and the number of feature components, and the third mode is the number of training heartbeats. Take the four-layer wavelet packet decomposition as an example. There are 16 sets of wavelet packet coefficients in the last layer after decomposition, and each set of wavelet packet coefficients contains m components. As attached figure 2 As shown, the 16 sets of wavelet packet coefficients can be arranged into an m×16 second-order tensor. Therefore, if there are n training signals, the training set is an m×16×n third-order tensor.

[0027] Step S106, through the generalized multi-dimensional independent component analysis method, obtain the solution mixing matrix W of each mode n. First, take the above four-layer wavelet packet decomposition as an example. If Is a series of training samples, It is a series of low-rank kernel tensors, which are the required fusion characteristics of training heartbeat signals. In order to minimize the energy error shown in Equation 1:

[0028] e = arg min W z + , z = 1,2 X i = 1 n | | X i tr - ( S i tr X 1 W 1 + X 2 W 2 + ) | | 2 - - - ( 1 )

[0029] Generalized multi-dimensional independent component analysis method using multi-linear subspace learning algorithm to solve the pseudo-inverse of the solution mixing matrix on each mode Among them, z=1, 2 represents the first two modes of volume data, that is, the number of features and the number of feature components; I 1 = M and I 2 =16 represents the number of each mode; J 1 ≤m and J 2 ≤16 represents the number of each mode after calculation. To get the unmixing matrix

[0030] Step S107, input a heartbeat signal to be tested, and pass steps 2 to 4 to obtain the fourth-level wavelet packet decomposition coefficients, which can be arranged into a second-order tensor X te ∈R m×16. Through step 6, the fusion characteristics of the test heartbeat signal are obtained:

[0031] S te =X te X 1 W 1 X 2 W 2 (2)

[0032] In step S108, a classifier, such as a support vector machine, is used to classify the fusion features of the heartbeat signal.

[0033] In step S109, classification results of normal and multiple abnormal ECG signals are obtained.

[0034] Although the present invention is described with reference to preferred embodiments, the above examples do not constitute a limitation of the scope of protection of the present invention. Any modification, equivalent replacement and improvement within the spirit and principle of the present invention shall be included in the present invention. Within the scope of the claims.