阅读说明：本技术 一种结合希尔伯特曲线和集成学习的心电磁信号分类方法 (Electrocardiogram and magnetic signal classification method combining Hilbert curve and integrated learning ) 是由 马辛 付幸文 曹一荻 于 2021-01-25 设计创作，主要内容包括：本发明一种结合希尔伯特曲线和集成学习的心电磁信号分类方法,属于心电磁信号分类领域,具有训练简单、分类精度高、检测速度快、适应性好、可靠性高的特点。本发明包括以下步骤：(1)获取心电磁信号并进行预处理再拆分为多段心拍信号；(2)采用希尔伯特曲线将每一段心拍信号填充为图像信号并重整得到数据集；(3)使用EasyEnsemble算法对数据集进行类别平衡；(4)采用集成学习方法以及Stacking结合策略得到分类模型,最后对分类模型进行评估。(The invention discloses an electrocardio-magnetic signal classification method combining Hilbert curve and integrated learning, belongs to the field of electrocardio-magnetic signal classification, and has the characteristics of simplicity in training, high classification precision, high detection speed, good adaptability and high reliability. The invention comprises the following steps: (1) acquiring a heart electromagnetic signal, preprocessing the heart electromagnetic signal and splitting the heart electromagnetic signal into a plurality of sections of heart beat signals; (2) filling each section of heart beat signal into an image signal by using a Hilbert curve and reforming to obtain a data set; (3) carrying out category balance on the data set by using an easy Ensemble algorithm; (4) and obtaining a classification model by adopting an ensemble learning method and a Stacking combination strategy, and finally evaluating the classification model.)

1. An electrocardio-magnetic signal classification method combining Hilbert curve and integrated learning is characterized by comprising the following steps:

(1) preprocessing the acquired heart electromagnetic signals, splitting the heart electromagnetic signals into a plurality of sections of heart beat signals, and determining the category label of each section of heart beat signals after fixed-length sampling;

(2) filling each section of heart beat signal into an image signal by using a Hilbert curve and reforming to obtain a data set;

(3) carrying out class balance and splitting on the data set by using an easy Ensemble algorithm to obtain a plurality of training subsets, a test set and a verification set;

(4) the method comprises the steps of adopting an ensemble learning method and a Stacking combination strategy to construct a classification model, wherein the classification model comprises a plurality of primary learners and a secondary learner, using a plurality of training subsets to respectively train the plurality of primary learners, after the training is finished, using a test set to test each primary learner, using an obtained prediction result as the input of the secondary learner, training the secondary learner by using a class mark of the test set as a target, obtaining a classification model by combining the plurality of primary learners and the secondary learner, and using a verification set to evaluate the classification model to obtain a classification precision index.

2. The classification method for the electrocardiographic and electromagnetic signals combining Hilbert curve and integrated learning according to claim 1, which is characterized in that: the step (1) comprises the following steps:

(11) acquiring an electrocardio-magnetic signal by using an electrocardio-magnetic measuring device;

(12) eliminating power frequency noise, baseline drift and motion noise of the electrocardio-magnetic signals by adopting a filtering algorithm;

(13) finding the position of each heart beat of the filtered heart electromagnetic signals by adopting a QRS waveform positioning algorithm and splitting the heart beats into a plurality of sections of heart beat signals;

(14) and carrying out fixed-length sampling on each section of heart beat signal in the plurality of sections of heart beat signals and determining the category mark of each section of heart beat signal.

3. The classification method of the electromagnetic signals of heart combining hilbert curve and integrated learning as claimed in claim 2, wherein: in the step (12), the filtering algorithm adopts one or more of band-pass filtering, Kalman filtering and Gaussian filtering.

4. The classification method of the electromagnetic signals of heart combining hilbert curve and integrated learning according to claim 2, characterized in that: and (13) adopting one or more of a difference threshold method, a double-threshold detection algorithm and a wavelet transform method for the QRS waveform positioning algorithm.

5. The classification method for the electrocardiographic and electromagnetic signals combining Hilbert curve and integrated learning according to claim 1, which is characterized in that: the step (2) comprises the following steps:

(21) processing each cardiac electromagnetic signal by using the step (1) to obtain a plurality of sections of cardiac beat signals, and filling each section of cardiac beat signal into an image signal by using a Hilbert curve;

(22) all image signals are combined and reformed into a 4-dimensional data set, and each dimension represents the number of samples, the number of channels, the length of the image signal and the width of the image signal.

6. The classification method for the electrocardiographic and electromagnetic signals combining Hilbert curve and integrated learning according to claim 1, which is characterized in that: the step (3) comprises the following steps:

(31) splitting the data set obtained in the step (2) into a majority type sample set and a minority type sample set according to the number of samples;

(32) and performing multiple undersampling on a majority sample set according to an EasyEnsemble algorithm, performing multiple oversampling on a minority sample set, wherein the number of samples obtained by each undersampling and oversampling is approximately equal, combining the samples pairwise to obtain a plurality of training subsets, and splitting the samples which are not extracted into 1 verification set and 1 test set.

7. The classification method for category-unbalanced cardiac electromagnetic signals based on Hilbert curve and integrated learning according to claim 1, which is characterized by comprising the following steps: the step (4) comprises the following steps:

(41) training the training subsets obtained in the step (3) to a plurality of primary learners respectively;

(42) after the training of the primary learners is finished, testing each primary learner by using a test set, using the output of the test set obtained by the primary learners as the input of a secondary learner according to a Stacking combination strategy, training by using class marks of the test set as learning targets, and after the training of the secondary learners is finished, combining a plurality of primary learners and the secondary learners to be used as a complete classification model;

(43) and evaluating the classification model by using the verification set to obtain a model classification precision index, and measuring the classification effect of the classification model according to the model classification precision index.

8. The classification method for category-unbalanced cardiac electromagnetic signals based on Hilbert curve and ensemble learning according to claim 7, comprising the following steps: in the steps (41) and (42), the primary learner uses a two-dimensional convolutional neural network (2D-CNN).

9. The classification method for category-unbalanced cardiac electromagnetic signals based on Hilbert curve and ensemble learning according to claim 7, comprising the following steps: in the step (43), the secondary learner is a Feedforward Neural Network (FNN) trained by a BP algorithm.

Technical Field

The invention relates to the field of classification of an electrocardio-magnetic signal, in particular to an electrocardio-magnetic signal classification method combining a Hilbert curve and integrated learning.

Background

Coronary heart disease, also known as ischemic heart disease, is a leading cause of death in the world according to a report by the World Health Organization (WHO) in 2016. The world health organization studies have shown that more than 1770 million die each year from cardiovascular disease, 80% of which are caused by heart disease. Myocardial infarction, which is a coronary heart disease, is the result of partial or complete occlusion of coronary arteries leading to insufficient blood flow to the heart. Patients with myocardial infarction can be diagnosed by methods such as changes in cardiac biomarkers such as electrocardiogram, echocardiogram, Magnetic Resonance Imaging (MRI), creatine kinase MB (CK-MB), troponin, and myoglobin. In practice, because the diagnosis of myocardial infarction has high requirements on timeliness, patients who have urgent needs have a diagnosis method of electrocardiogram as the first choice. An electrocardiograph is generally provided in an emergency room, and if it can be identified accurately, myocardial infarction can be diagnosed quickly and with high accuracy. In addition, cardiac magnetic signal detection equipment is also rapidly developed at present, and compared with cardiac electric signals, the cardiac magnetic signals have vectority, so that theoretically more information can be carried. Some researchers find that the cardiac magnetic signal has higher specificity and stability in the diagnosis of heart diseases such as myocardial infarction. Although there are some problems with the current use of cardiac magnetic signals in clinical applications, it cannot be denied that they have many potential clinical applications. Therefore, the intelligent classification of myocardial infarction diseases based on the electrocardio-electromagnetic signals is a work with great significance and prospect in clinical medical treatment.

The intelligent classification method of myocardial infarction diseases based on the electrocardio-electromagnetic signals is mainly divided into 2 types. The characteristic of the classic classification method is that the data is preprocessed to extract the features, and then a shallow classifier is used to classify the extracted features. The feature extraction part and the classification part of the method are mutually independent in structure and are only coupled through features. Therefore, the overall performance is mainly determined by the quality of the features, so the method is excessively dependent on feature engineering, requires a lot of manpower, and is limited by the level of human knowledge and experience. Another is an end-to-end deep neural network classification method that allows the original signal to be fed directly into the neural network after a simple noise reduction pre-processing. The advantage of using the neural network as the classifier is that it can automatically learn to extract features and synthesize the features to give classification results only by providing training samples. Because the extraction of the features is automatically carried out according to the training algorithm, the method can avoid the limitation of human knowledge and experience level so as to obtain higher accuracy. However, the classification of the cardiac electromagnetic signals by the existing deep neural network mainly has the following 2 problems. One is that because the ecg signal is a typical time series signal, although the deep neural network is well developed in computer vision and speech recognition, it is difficult to construct a classification model when encountering the time series signal, because the reasons include: the recurrent neural network is difficult to train, and the research results in the aspect of partial computer vision are difficult to apply to time series signals. The other is that the category imbalance problem exists in the electrocardio-magnetic signal data inevitably. At present, some processing methods at home and abroad mainly comprise 2 methods, one method is to directly combine most samples and few samples in an electrocardio-magnetic signal data set to form a training set to train a network, so that data is not wasted, but the method is far from the actual situation, so that the classification effect of a classification model obtained by an experiment is good, and the practical value is general. The other method is to undersample a large number of myocardial infarction signals and then combine the undersampled signals with healthy control signals, so that although the problem of unbalanced categories is solved, partial myocardial infarction signals are wasted, and partial characteristics of the myocardial infarction signals are lost.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the problem of difficulty in training and the problem of unbalanced category of an electromagnetic signal classification model serving as a time sequence is solved, the electromagnetic signal classification method combining the Hilbert curve and integrated learning is provided for myocardial infarction real-time detection, and the classification model obtained by the method has the characteristics of simplicity in training, high classification precision, high detection speed, good adaptability and high reliability.

The technical scheme adopted by the invention for solving the technical problems is as follows:

an electrocardio-magnetic signal classification method combining Hilbert curve and integrated learning comprises the following steps:

(1) preprocessing the acquired heart electromagnetic signals, splitting the heart electromagnetic signals into a plurality of sections of heart beat signals, and determining the category label of each section of heart beat signals after fixed-length sampling;

(2) filling each section of heart beat signal into an image signal by using a Hilbert curve and reforming to obtain a data set;

(3) carrying out class balance and splitting on the data set by using an easy Ensemble algorithm to obtain a plurality of training subsets, a test set and a verification set;

(4) the method comprises the steps of adopting an ensemble learning method and a Stacking combination strategy to construct a classification model, wherein the classification model comprises a plurality of primary learners and a secondary learner, using a plurality of training subsets to respectively train the plurality of primary learners, after the training is finished, using a test set to test each primary learner, using an obtained prediction result as the input of the secondary learner, training the secondary learner by using a class mark of the test set as a target, obtaining a classification model by combining the plurality of primary learners and the secondary learner, and using a verification set to evaluate the classification model to obtain a classification precision index.

The step (1) comprises the following steps:

(11) acquiring an electrocardio-magnetic signal by using an electrocardio-magnetic measuring device;

(12) eliminating power frequency noise, baseline drift and motion noise of the electrocardio-magnetic signals by adopting a filtering algorithm;

(13) finding the position of each heart beat of the filtered heart electromagnetic signals by adopting a QRS waveform positioning algorithm and splitting the heart beats into a plurality of sections of heart beat signals;

(14) and carrying out fixed-length sampling on each section of heart beat signal in the plurality of sections of heart beat signals and determining the category mark of each section of heart beat signal.

In the step (12), the filtering algorithm adopts one or more of band-pass filtering, Kalman filtering and Gaussian filtering.

And (13) adopting one or more of a difference threshold method, a double-threshold detection algorithm and a wavelet transform method for the QRS waveform positioning algorithm.

The step (2) comprises the following steps:

(21) processing each cardiac electromagnetic signal by using the step (1) to obtain a plurality of sections of cardiac beat signals, and filling each section of cardiac beat signal into an image signal by using a Hilbert curve;

(22) all image signals are combined and reformed into a 4-dimensional data set, and each dimension represents the number of samples, the number of channels, the length of the image signal and the width of the image signal.

The step (3) comprises the following steps:

(31) splitting the data set obtained in the step (2) into a majority type sample set and a minority type sample set according to the number of samples;

(32) and performing multiple undersampling on a majority sample set according to an EasyEnsemble algorithm, performing multiple oversampling on a minority sample set, wherein the number of samples obtained by each undersampling and oversampling is approximately equal, combining the samples pairwise to obtain a plurality of training subsets, and splitting the samples which are not extracted into 1 verification set and 1 test set.

The step (4) comprises the following steps:

(41) training the training subsets obtained in the step (3) to a plurality of primary learners respectively;

(42) after the training of the primary learners is finished, testing each primary learner by using a test set, using the output of the test set obtained by the primary learners as the input of a secondary learner according to a Stacking combination strategy, training by using class marks of the test set as learning targets, and after the training of the secondary learners is finished, combining a plurality of primary learners and the secondary learners to be used as a complete classification model;

(43) and evaluating the classification model by using the verification set to obtain a model classification precision index, and measuring the classification effect of the classification model according to the model classification precision index.

In the steps (41) and (42), the primary learner uses a two-dimensional convolutional neural network (2D-CNN).

In the step (43), the secondary learner is a Feedforward Neural Network (FNN) trained by a BP algorithm.

The principle of the invention is as follows: the hilbert curve is a well-known space filling curve, and can traverse each pixel point in a two-dimensional image by using a one-dimensional curve, and ensure that two adjacent points in the curve are also adjacent in the image, and when the order of the curve is increased, the position of each point in the curve cannot change greatly in the image. The electrocardio-magnetic signals are typical time sequence signals, if a Hilbert curve is adopted to fill the electrocardio-magnetic signals into an image matrix, the time dependency of the time sequence signals can be converted into the position dependency of the image signals, and a two-dimensional convolutional neural network (2D-CNN) is extremely sensitive to the shape and the position of the image, so that the classification precision and the stability of a classification model can be improved by utilizing the visual advantage of the existing computer on the premise of not losing any information of the electrocardio-magnetic signals, and the training difficulty of the classification model is reduced. In addition, the easynesemble algorithm is a classic class imbalance data modeling algorithm, a plurality of class samples are randomly divided into a plurality of subsets, each subset is respectively combined with a few class samples to obtain a plurality of new data subsets, and a primary classifier is trained by utilizing each data subset. And then training a secondary classifier according to a Stacking combination strategy, and finally combining the primary classifier and the secondary classifier to obtain a classification model. Therefore, the problem of class imbalance can be solved on the premise of not losing most sample features, and the classification precision and the adaptability of a classification model are further improved.

Compared with the prior art, the invention has the advantages that:

(1) according to the invention, the time sequence signal is filled into the image signal through the Hilbert curve, and the classification precision and stability of the classification model are improved by utilizing the existing computer vision advantages, so that the training difficulty of the classification model is reduced.

(2) The invention solves the problem of class imbalance by the EasyEnsemble algorithm and the Stacking combined strategy on the premise of not losing most sample characteristics, and further improves the classification precision and the adaptability of a classification model.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a raw ECG signal;

FIG. 3 is a pre-processed ECG signal;

FIG. 4 is a heartbeat signal;

FIG. 5 is an image signal after filling via a Hilbert curve;

FIG. 6 is a schematic view of a data set;

FIG. 7 is a schematic diagram of the EasyEnsemble algorithm and Stacking combined strategy;

fig. 8 is a hilbert plot for different orders.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, exemplary embodiments of the present invention are described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other. The classification targets involved in the foregoing technical solutions may be cardiac electrical signals (ECG) and cardiac magnetic signals (MCG), and the following describes a specific implementation process of the present invention by taking ECG signals as an example

As shown in fig. 1, the method of the present invention specifically includes the following steps:

(1) the method comprises the following steps of acquiring an ECG signal, preprocessing the ECG signal, splitting the ECG signal into multiple segments of heartbeat signals, and determining the category label of each segment of heartbeat signal after fixed-length sampling, wherein the method specifically comprises the following substeps:

(1-1) acquiring a raw ECG signal from The PTB Diagnostic ECG Database (PTB), wherein fig. 2 shows The raw ECG signal of 4.096 seconds, which is specifically a time sequence signal, The sampling frequency is 1kHz, 4096 data points are included, The potential change of The heart corresponding to time is represented by a voltage amplitude value, and certain power frequency noise, baseline drift and motion noise are carried;

and (1-2) eliminating power frequency noise, baseline drift and motion noise of the ECG signal by adopting a filtering algorithm, wherein the filtering algorithm is one or more of band-pass filtering, Kalman filtering and Gaussian filtering. FIG. 3 shows the results obtained by filtering the original ECG signal shown in FIG. 2 with a band-pass filter of 5-200 Hz, which suppresses the power frequency noise, baseline drift and motion noise of the original ECG signal within the allowable range.

(1-3) finding the position of each heart beat of the filtered ECG signal by adopting a QRS waveform positioning algorithm and splitting the position into multiple sections of heart beat signals, wherein the QRS waveform positioning algorithm adopts one or more of a differential threshold method, a double-threshold detection algorithm and a wavelet transformation method. Fig. 4 shows the heart beat signals obtained by QRS positioning and splitting by using the differential threshold method, each heart beat signal only contains one complete heart beat process, so the time length changes with the change of each heart beat time, but the sampling frequency is still 1 kHz.

(1-4) sampling each heart beat signal in a fixed length mode and determining a classification mark of the heart beat signal, wherein each obtained heart beat has a fixed length of 256 data points, the mark of a myocardial infarction patient is 1, and the mark of a healthy control is 0.

(2) Filling each segment of heart beat signal into an image signal by using a Hilbert curve and reforming to obtain a data set, and specifically comprising the following substeps:

(2-1) processing each ECG signal using the step (1) to obtain a cardiac signal, and then filling the cardiac signal with a hilbert curve to obtain an image signal, as shown in fig. 5, where a picture has 16 × 16 to 256 pixels, each pixel corresponds to each data point of the cardiac signal, a darker pixel indicates that the voltage amplitude of the cardiac signal at the point is larger, the hilbert curve is shown in fig. 8, which is an image filling curve and is also a recursive parting curve, a first-order graph is obtained by rotating and reconnecting 4 lower-order graphs, and a 1-order hilbert curve is a minimum unit, where a 4-order hilbert curve is used in this example.

(2-2) all the image signals are combined and reformed into a 4-dimensional data set, as shown in fig. 6, each dimension represents the number of samples, the number of channels, the length of the image signal, and the width of the image signal, each cube is a sample, 72 patients with myocardial infarction, 52 healthy controls, and each person contains 20 samples, so that the data set contains 2480 samples, each sample has 12 channels, and each channel is a picture with 16 × 16 ═ 256 pixels.

(3) Carrying out class balance and splitting on the data set by using an easy Ensemble algorithm to obtain a plurality of training subsets, a test set and a verification set, and specifically comprising the following substeps:

(3-1) splitting the data set obtained in the step (2) into a majority class sample set and a minority class sample set according to the number of samples.

(3-2) performing 5 times of undersampling on a majority sample set and 5 times of oversampling on a minority sample set according to an easy Ensemble algorithm, wherein the number of samples obtained by each undersampling and oversampling is approximately equal, combining the samples two by two to obtain 5 training subsets, and splitting the samples which are not extracted into 1 verification set and 1 test set, wherein the training subsets, the test sets and the verification sets are 5:2: 3.

(4) The method comprises the following steps of constructing a classification model by adopting an ensemble learning method and a Stacking combination strategy, wherein the classification model comprises a plurality of primary learners and a secondary learner, respectively training the plurality of primary learners by using a plurality of training subsets, testing each primary learner by using a test set after the training is finished, taking an obtained prediction result as the input of the secondary learner, training the secondary learner by taking a class mark of the test set as a target, obtaining a classification model by combining the plurality of primary learners and the secondary learner, and evaluating the classification model by using a verification set to obtain a classification precision index, and specifically comprises the following sub-steps:

(4-1) training the training subsets obtained from the step (3) to a plurality of primary learners, respectively.

And (4-2) after the training of the primary learners is finished, testing each primary learner by using the test set, taking the output of the test set obtained by the primary learners as the input of the secondary learner according to a Stacking combination strategy, training by taking class marks of the test set as learning targets, and after the training of the secondary learners is finished, combining a plurality of primary learners and the secondary learners to be used as a complete classification model.

The primary learners in the steps (4-1) and (4-2) adopt two-dimensional convolutional neural networks (2D-CNN), and each two-dimensional convolutional neural network (2D-CNN) is consistent in structure and comprises 3 convolutional layers, 3 pooling layers, 1 full-connection layer and 1 output layer. The 1 st convolutional layer has 10 convolution kernels of size 2 × 2, the 2 nd convolutional layer has 20 convolution kernels of size 2 × 2, and the 3 rd convolutional layer has 50 convolution kernels of size 2 × 2. Each convolutional layer is followed by a 2 x 2 max pooling layer. The fully-connected layer has 50 neurons, the activation functions of the convolutional layer and the fully-connected layer are both relu, and the activation function of the output layer is softmax. The optimization method is Adam, an L2 regularization method is adopted, the attenuation coefficient is 1e-3, and the learning rate is 1 e-3. The secondary learner is a Feedforward Neural Network (FNN) trained by adopting a BP algorithm, and comprises 1 full-connection layer and 1 output layer, wherein the full-connection layer is provided with 25 neurons, the activation function is relu, the activation function of the output layer is softmax, the optimization method is SGD, an L2 regularization method is adopted, the attenuation coefficient is 1e-5, and the learning rate is 1 e-2.

And (4-3) evaluating the classification model by using the verification set to obtain a model classification precision index, and measuring the classification effect of the classification model according to the model classification precision index.

The schematic diagram of the easy Ensemble algorithm and the Stacking combining strategy based on the data stream in the step (3) and the step (4) is shown in FIG. 7.

The present invention adopts the following indexes to evaluate the performance of the classification model, as shown in formulas (1), (2), (3),

wherein ACC is classification accuracy, Sn is sensitivity, Sp is specificity, TP, FP, TN, FN represent true positive, false positive, true negative, false negative, respectively, and the confusion matrix of the classification results is shown in Table 1.

TABLE 1 confusion matrix of classification results

TABLE 2 model indices of the invention

TABLE 3 model indices in the field

The final results obtained on the validation set are shown in table 2, wherein 260 myocardial infarction samples and 240 healthy control samples are obtained in the validation set. From comparison on the model indexes of the invention, the performance of the integrated classification model is far superior to that of any primary learner because the integrated classification model fuses all primary learners by adopting a proper method. In addition, table 3 also lists some model indexes in the field, wherein a traditional classification model needing feature extraction is more ideal in classification accuracy, but depends on feature engineering and is limited by the level of human knowledge and experience; although the long-time memory network (LSTM) belonging to the recurrent neural network is relatively suitable for time series signals, the model training is difficult, and a high classification precision is difficult to achieve; the classification precision of the one-dimensional convolutional neural network (1D-CNN) is poor, the classification precision of the 13-layer Deep Neural Network (DNN) is superior, but the model is too complex.

Based on the comparison, the time sequence signals are filled into image signals by using the Hilbert curve, and the advantages of computer vision are utilized, so that the classification precision of the model is greatly improved, the network structure is simpler, the training is facilitated, and the time required by classification is shortened. The invention adopts the easy Ensemble algorithm to carry out class balance on training data and integrates learning and Stacking combined strategies, so that the adaptability of the model is stronger, and the classification precision of the whole classification model is further improved. The method has great application significance in the field of classification of cardiac electromagnetic signals with unbalanced classes.

Those skilled in the art will appreciate that the invention may be practiced without these specific details.

The above examples are provided only for the purpose of describing the present invention, and are not intended to limit the scope of the present invention. The scope of the invention is defined by the appended claims. Various equivalent substitutions and modifications can be made without departing from the spirit and principles of the invention, and are intended to be within the scope of the invention.