WO2020061987A1

WO2020061987A1 - Multi-lead combined electrocardiogram classification method

Info

Publication number: WO2020061987A1
Application number: PCT/CN2018/108192
Authority: WO
Inventors: 陈彬; 郭维
Original assignee: 深圳大学
Priority date: 2018-09-28
Filing date: 2018-09-28
Publication date: 2020-04-02

Abstract

A multi-lead combined electrocardiogram classification method, comprising: S1: preprocessing electrocardiogram data in a CCDD database; and S2: encoding the preprocessed multi-channel electrocardiogram data into a channel signal, and classifying the channel signal. A computer readable storage medium having a computer program stored thereon. The multi-lead combined electrocardiogram classification method is implemented when the program is executed by a processor. The multi-lead combined electrocardiogram classification method is implemented to automatically detect normal and abnormal electrocardiogram signals, so as to improve the accuracy of classification.

Description

Multi-lead combined ECG classification method

Technical field

The invention relates to the field of data processing, and more particularly, to a multi-lead combined electrocardiogram classification method.

Background technique

At present, 12-lead synchronous recording electrocardiograph has been widely used in clinical practice. The 12-lead synchronous recording electrocardiograph can simultaneously record the ECG signals of the same cardiac cycle on the 12-lead. It can identify and locate single-source or multi-source premature beats, classify arrhythmias, and diagnose indoor conduction block. Instrument has superiority. The 12-lead synchronous recording electrocardiograph can simultaneously and collectively observe the waveforms of the same 12-lead cardiac cycle, which greatly improves the accuracy of various measurements and reduces the variability of existing ECG measurements.

At present, the mainstream clinical ECG on the market is multi-lead multi-channel signals. Therefore, the ECG classification algorithm running on large equipment such as electrocardiographs must be combined with multi-channel signals for auxiliary detection of ECG. It is not logical, and on the one hand, it will lead to a negative impact on accuracy. Therefore, for a 12-lead simultaneous recording of multi-lead ECG signals, a normal and abnormal ECG signal can be automatically detected to improve classification accuracy. Multilead Joint ECG Classification Method.

Summary of the Invention

The technical problem to be solved by the present invention is to provide a multi-lead combined electrocardiogram classification method capable of automatically detecting normal and abnormal ECG signals to improve the classification accuracy in response to the above-mentioned defects of the prior art.

The technical solution adopted by the present invention to solve its technical problems is to construct a multi-lead combined electrocardiogram classification method, including:

S1. Preprocess the ECG data in the CCDD database;

S2. The pre-processed multi-channel electrocardiogram data is encoded into a one-channel signal, and the one-channel signal is classified.

In the multi-lead combined electrocardiogram classification method of the present invention, the step S1 further includes:

S11. Eliminate normal ECG data and duplicate ECG data marked as disease labels in the CCDD database;

S12. Format the remaining ECG data and set the sampling frequency.

S13. Select 8-lead ECG data as experimental data.

In the multi-lead combined electrocardiogram classification method of the present invention, in step S12, the electrocardiogram data in the XML format with a sampling rate of 500 Hz is converted into the electrocardiogram data in the mat format with a sampling rate of 250 Hz.

In the multi-lead combined electrocardiogram classification method of the present invention, the step S2 further includes:

S21: Use the D-LSTM network structure to extract the pre-processed multi-channel ECG data of each lead for feature extraction, and classify the output features through a fully connected layer;

S22. Use a three-layer two-dimensional convolutional network to extract features from the pre-processed multi-channel ECG data of each lead, and classify the output features through a fully connected layer;

S23. Use a three-layer two-dimensional convolutional network to extract features from the pre-processed multi-channel ECG data of each lead, and classify the output features through the D-LSTM network structure; or

S24. The pre-processed multi-channel electrocardiogram data is regarded as eight-channel one-dimensional data and encoded by channel changes, and then the encoded data is classified by a D-LSTM network structure.

In the multi-lead joint ECG classification method of the present invention, the D-LSTM network structure includes a deep LSTM network formed by a plurality of LSTM network units, a first fully connected layer connected to the deep LSTM network, and A second fully connected layer connected to the first fully connected layer.

In the multi-lead combined electrocardiogram classification method of the present invention, the second fully connected layer is a SoftMax layer.

In the multi-lead combined ECG classification method of the present invention, the pre-processed multi-channel ECG data of each lead sets the convolution kernel and pooling layer to one through the three-layer two-dimensional convolution network. Dimension, after three layers of convolution, we get 64 * 8 * n features.

Another technical solution adopted by the present invention to solve its technical problem is to construct a computer-readable storage medium on which a computer program is stored, and the multi-lead combined electrocardiogram classification method is implemented when the program is executed by a processor. .

By implementing the multi-lead combined ECG classification method of the present invention, normal and abnormal ECG signals can be automatically detected to improve classification accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be further described below with reference to the accompanying drawings and embodiments. In the drawings:

1 is a method flowchart of a first embodiment of a multi-lead combined electrocardiogram classification method according to the present invention;

2 is a schematic structural diagram of a Multi-RNN model of a multi-lead combined electrocardiogram classification method of the present invention;

3 is a schematic structural diagram of a Vanilla-CNN model of a multi-lead combined electrocardiogram classification method of the present invention;

4 is a schematic structural diagram of a Feature-CNN model of the multi-lead combined electrocardiogram classification method of the present invention;

5 is a schematic structural diagram of a Channel-CNN model of a multi-lead combined electrocardiogram classification method of the present invention;

6 is a schematic structural diagram of a D-LSTM network model of the multi-lead combined electrocardiogram classification method of the present invention;

FIG. 7 shows the accuracy of a multi-lead combined electrocardiogram classification method according to different embodiments of the present invention.

detailed description

In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

The invention relates to a multi-lead combined ECG classification method, which includes: S1, preprocessing ECG data in a CCDD database; S2, encoding the pre-processed multi-channel ECG data into a channel signal, and One channel signal is classified. By implementing the multi-lead combined ECG classification method of the present invention, normal and abnormal ECG signals can be automatically detected to improve classification accuracy.

FIG. 1 is a method flowchart of a first embodiment of a multi-lead combined electrocardiogram classification method of the present invention. As shown in FIG. 1, in step S1, the ECG data in the CCDD database is pre-processed. In a preferred embodiment of the present invention, the step S1 further includes: S11. Excluding normal ECG data and duplicate ECG data marked as disease labels in the CCDD database; S12. Formatting the remaining ECG data and converting Set the sampling frequency; S13. Select 8-lead ECG data as experimental data.

For example, the CCDD database uses a large amount of clinical ECG data, so there are some problems in the data. In order to ensure the accuracy and standard type of the experiment, the database file needs to be preprocessed as follows.

(1) Some normal ECGs in the CCDD database are mixed with disease labels. These contaminated data will cause negative impact on the identification of normal ECG signals in classification experiments. Therefore, the contaminated normal ECG data is removed first, and only samples with normal ECG tags are stored as positive samples, which ensures the purity of the data file.

(2) There are 20,000 completely duplicate samples in the original database file. After verification, all information is the same except for different serial numbers. It should be a database storage error. If these 20,000 samples are mixed into the training set and test set, the classification difficulty will be slightly reduced, and the classification accuracy will be increased. Can't reflect the true ability of the model, so the duplicate samples are eliminated first to ensure the fairness of the experiment.

(3) The raw data file of the CCDD database is in xml format, and the sampling rate is 500Hz. In order to facilitate the use, the selected data is first converted to Mat format which is easily processed by Matlab, and downsampled to 250Hz by Matlab. This can ensure the integrity of the ECG waveform while reducing the amount of calculation. The normal and abnormal ECG data are divided into training set and test set.

(4) Because the ECG is orthogonal, the other four can be deduced using 8 leads, so only 8 leads can be selected for multi-channel classification. The dimension of the input data is reduced while guaranteeing the classification results. Greatly reduces training time and difficulty.

As further shown in FIG. 1, in step S2, the pre-processed multi-channel electrocardiogram data is encoded into a channel signal, and the one channel signal is classified. In a preferred embodiment of the present invention, a D-LSTM network structure is used for feature extraction of pre-processed multi-channel ECG data of each lead, and the output features are classified by a fully connected layer, hereinafter referred to as " Multi-RNN "network model. In yet another preferred embodiment of the present invention, a three-layer two-dimensional convolutional network is used to extract features from the pre-processed multi-channel ECG data of each lead, and classify the output features through a fully connected layer; the following; Called "Vanilla-CNN" network model. In yet another preferred embodiment of the present invention, a three-layer two-dimensional convolutional network is used to extract features from the pre-processed multi-channel ECG data of each lead, and classify the output features through the D-LSTM network structure. ; Hereinafter referred to as the "Feature-CNN" network model. In another preferred embodiment of the present invention, pre-processed multi-channel electrocardiogram data is taken as eight-channel one-dimensional data and encoded by channel changes, and then the encoded data is classified by a D-LSTM network structure, This is called the "Channel-CNN" network model.

FIG. 2 is a schematic structural diagram of a Multi-RNN model of a multi-lead combined electrocardiogram classification method of the present invention. As shown in Figure 2, in the Multi-RNN model, since each lead of the electrocardiogram is a one-dimensional signal, it can be processed by the RNN. What is obtained is the characteristics of the timing signal, and then the output features are fully connected to obtain the classification result. The RNN model may adopt the D-LSTM network model shown in FIG. 6. As shown in FIG. 6, it includes a deep LSTM network composed of multiple LSTM network units, a fully connected layer connected to the deep LSTM network, and a SoftMax layer connected to the fully connected layer. The network uses a many-to-one model of sequence-to-label, that is, a time series uses only the output value of the last unit to match the label. The back end of the model uses two fully connected layers to complete the final classification task, and the final output is the classification result. The ECG sample contains multiple time states. The multiple time states of the ECG signal are separated into t vectors of x1, x2 ... xt, and sent to the LSTM network to obtain the coding vector containing the sequence connection. After the full connection and the SoftMax layer are converted into probability values, the output classification results are obtained. In a preferred embodiment of the present invention, a four-layer D-LSTM network model is preferably used.

FIG. 3 is a schematic structural diagram of a Vanilla-CNN model of the multi-lead combined electrocardiogram classification method of the present invention. As shown in Figure 3, in the Vanilla-CNN model, a two-dimensional convolution network is used. The convolution kernel and pooling layer are set to one dimension. After three layers of convolution, the final result is a 64-channel 8 * n Features are equivalent to extracting features independently for each lead, and each lead will not affect each other.

FIG. 4 is a schematic structural diagram of a Feature-CNN model of the multi-lead combined electrocardiogram classification method of the present invention. As shown in Figure 4, in the Feature-CNN model, the CNN processing part is the same as Vanilla-CNN, and the multi-channel independent 8-lead feature is obtained. Feature end did fusion coding operation.

FIG. 5 is a schematic structural diagram of a Channel-CNN model of the multi-lead combined electrocardiogram classification method of the present invention. As shown in Figure 5, in the Channel-CNN model, the ECG data is treated as a multi-channel one-dimensional signal instead of an independent multi-conductor signal. Therefore, the CNN network chooses to use one-dimensional convolution and performs channel operations. The ECG signal is regarded as one-dimensional data of 8 channels, which is encoded by the change of the channel. Also access D-LSTM finally for classification.

FIG. 7 shows the accuracy of a multi-lead combined electrocardiogram classification method according to different embodiments of the present invention. As shown in Figure 7, four different coding models achieved 81.64%, 81.5%, 82.5%, and 83.9% accuracy, respectively. Multi-RNN and Vanilla-CNN, a model in which each lead uses RNN or CNN to perform independent coding classification and then make classification decisions, has obtained similar classification accuracy. This result is better than the optimal V5 in a single lead classification The 80.36% accuracy of the lead is higher. From this result, we can see that the multi-lead joint classification has certain advantages over the classification using only a single lead. The increase in the amount of original information makes the classification more accurate. Feature-CNN uses a model of CNN encoding and then RNN classification. After experimental results, it is found that it reaches about 82.5%, and the accuracy rate is 1% higher than the previous two models. This result shows that the fusion of CNN and RNN is effective in the classification of ECG signals. CNN is used for front-end coding fusion, and RNN is used for classification at the back end. Good results have been achieved on multi-channel timing signals such as ECG. Channel-CNN uses the best class accuracy. Unlike most studies, multi-lead ECG signals are treated as two-dimensional matrices. This model treats ECG signals as one-dimensional signals from multiple channels. When performing the convolution operation, a one-dimensional convolution layer is used, and then the channel fusion operation is performed. The data of all channels is weighted at the beginning, instead of each lead's independent coding and finally fusion. The back end of the model still uses the RNN structure for classification. The experimental results prove that using this model structure is the most suitable classification method for ECG signals.

Another embodiment of the present invention provides a machine-readable memory and / or storage medium, and the machine code and / or computer program stored therein includes at least one code segment, which is executed by a machine and / or a computer such that the machine and / or Or the computer executes each step of the multi-lead combined electrocardiogram classification method described in this application.

Therefore, the present invention can be implemented by hardware, software, or a combination of software and hardware. The invention may be implemented in a centralized manner in at least one computer system or in a decentralized manner by different parts distributed among several interconnected computer systems. Any computer system or other device that can implement the method of the present invention is applicable. The combination of commonly used software and hardware can be a general-purpose computer system with a computer program installed, and the computer system can be controlled to run according to the method of the present invention by installing and executing the program.

The present invention can also be implemented by a computer program product. The program includes all the features capable of implementing the method of the present invention. When installed in a computer system, the method of the present invention can be implemented. The computer program in this document refers to: any expression that can use a set of instructions written in any programming language, code, or symbol, which enables the system to have information processing capabilities to directly implement specific functions, or to perform A specific function is achieved after describing one or two steps: a) conversion to other languages, codes or symbols; b) reproduction in different formats.

Although the present invention is described through specific embodiments, those skilled in the art should understand that various changes and equivalent substitutions can be made to the present invention without departing from the scope of the present invention. In addition, for the specific situation or material, various modifications can be made to the invention without departing from the scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed, but should include all implementations falling within the scope of the claims of the present invention.

Claims

A multi-lead combined electrocardiogram classification method includes:

S1. Preprocess the ECG data in the CCDD database;

S2. The pre-processed multi-channel electrocardiogram data is encoded into a one-channel signal, and the one-channel signal is classified.
The multi-lead combined electrocardiogram classification method according to claim 1, wherein the step S1 further comprises:

S11. Eliminate normal ECG data and duplicate ECG data marked as disease labels in the CCDD database;

S12. Format the remaining ECG data and set the sampling frequency.

S13. Select 8-lead ECG data as experimental data.
The multi-lead combined electrocardiogram classification method according to claim 2, wherein in step S12, the electrocardiogram data in the xml format with a sampling rate of 500 Hz is converted into the electrocardiogram data in the mat format with a sampling rate of 250 Hz.
The multi-lead combined electrocardiogram classification method according to claim 1, wherein the step S2 further comprises:

S21: Use the D-LSTM network structure to extract the pre-processed multi-channel ECG data of each lead for feature extraction, and classify the output features through a fully connected layer;

S22. Use a three-layer two-dimensional convolutional network to extract features from the pre-processed multi-channel ECG data of each lead, and classify the output features through a fully connected layer;

S23. Use a three-layer two-dimensional convolutional network to extract features from the pre-processed multi-channel ECG data of each lead, and classify the output features through the D-LSTM network structure; or

S24. The pre-processed multi-channel electrocardiogram data is regarded as eight-channel one-dimensional data and encoded by channel changes, and then the encoded data is classified by a D-LSTM network structure.
The multi-lead joint ECG classification method according to claim 4, wherein the D-LSTM network structure comprises: a deep LSTM network composed of a plurality of LSTM network units, and a first full-length LSTM network connected to the deep LSTM network. A connection layer, and a second fully connected layer connected to the first fully connected layer.
The method according to claim 5, wherein the second fully connected layer is a SoftMax layer.
The multi-lead combined ECG classification method according to claim 4, characterized in that the pre-processed multi-channel ECG data of each lead is used to convolve and pool the convolution kernel through the three-layer two-dimensional convolution network. The layers are set to one dimension. After three layers of convolution, the 64-channel 8 * n features are obtained.
A computer-readable storage medium having stored thereon a computer program, characterized in that when the program is executed by a processor, the multilead combined electrocardiogram classification method according to any one of claims 1-7 is implemented .