CN104523266B - A kind of electrocardiosignal automatic classification method - Google Patents

Abstract

The invention discloses a method for automatic classification of electrocardiographic signals, which is realized according to the following steps: a) acquiring the electrocardiographic signals of the human body, performing filtering processing, and detecting the R wave of the filtered electrocardiographic signals; b) detecting After the R wave, build a data set, the data set is composed of several sets of heart beat data, each set of heart beat data has a label; c) construct a sparse automatic encoding deep learning network; d) train the sparse Automatically encode the deep learning network; e) according to the network weights of the first hidden layer, the network weights of the second hidden layer and the network weights of the softmax classifier obtained in step d), input the heartbeat data to be tested into the Sparse auto-encoding deep learning network to obtain heart beat data for classification output. The present invention applies the sparse auto-encoding deep learning network to the classification of heart beat data, utilizes its self-learning ability and the characteristics of deep feature mining, extracts deeper features of the signal, and classifies the heart beat data.

Description

A method for automatic classification of ECG signals

technical field

The invention relates to an automatic detection and analysis technology of electrocardiographic signals, in particular to a method for automatic classification of electrocardiographic signals.

Background technique

Heart disease is hidden and latent, and it is difficult to show it on the ECG when it does not occur, and it is short-lived when it occurs, and it is too late to observe the ECG. To this end, it is necessary to carry a 24-hour Holter to the patient to collect ECG signals for 24 hours, and then hand over the ECG data to the doctor, who will analyze the data. The amount of data generated at this time is huge, and doctors need a lot of time to find abnormal heart beats. Although Holter's built-in software system can automatically analyze heartbeats and give statistical information, due to the large differences in human body and complex ECG changes, some heartbeats still need to be manually recognized and corrected by doctors. Finding mislabeled heart beats in such a large amount of data is also a very laborious task. In order to greatly save doctors' time and improve diagnosis efficiency, a stable automatic classification algorithm is very necessary.

Contents of the invention

The purpose of the present invention is to provide a method for automatic classification of electrocardiographic signals to solve the unstable problem of classification algorithms for electrocardiographic signals in different human bodies and environments.

The object of the present invention is achieved in that: the electrocardiogram automatic classification method provided by the present invention comprises the following steps:

a) Obtain the ECG signal of the human body, perform filtering processing, and detect the R wave of the filtered ECG signal;

b) After the R wave is detected, a data set is constructed. The data set is composed of several groups of heart beat data, and each set of heart beat data has a label. There are 6 types of labels in total, which are divided into normal heart beat, left beam Branch block, right bundle branch block, ventricular premature beats, atrial premature beats, confluent beats:

Each group of heart beat data includes 270 sampling points, the 270 sampling points are based on the position of the detected R wave, 90 sampling points are selected in front of the peak point of the R wave, and at the peak point of the R wave 179 sampling points are selected behind;

c) Build a sparse auto-encoding deep learning network:

The sparse auto-coding deep learning network has two hidden layers, followed by a softmax classifier, wherein the two hidden layers are respectively the first hidden layer and the second hidden layer;

The input of the sparse automatic encoding deep learning network is 270 sampling points, the number of nodes in the first hidden layer is 130, and the number of nodes in the second hidden layer is 50;

d) step-by-step training of the sparse auto-encoding deep learning network:

d-1) Normalize the heartbeat data of the data set, then input the SAE model, and use the SAE model to train the first hidden layer of the sparse auto-encoding deep learning network to obtain the network of the first hidden layer Weight, and get the shallow features of heart beat data;

d-2) adopting the same SAE model, inputting the shallow features into the SAE model, training the second hidden layer of the sparse auto-encoding deep learning network, obtaining the network weights of the second hidden layer, and Obtaining the high-level features of the several groups of cardiac beat data;

d-3) Input the obtained high-level features into the softmax classifier, train the softmax classifier, and obtain the network weight of the softmax classifier;

e) According to the network weights of the first hidden layer obtained in step d), the network weights of the second hidden layer and the network weights of the softmax classifier, input the heartbeat data to be measured into the sparse automatic encoding deep learning network , to get the classification output of heart beat data.

The concrete process of described step a) is as follows:

(1) signal collection: gather human body electrocardiogram original signal with the acquisition frequency of 250Hz, and store as the data form of TXT document, then read in the computer with the electrocardiogram original signal data of described TXT document storage with Matlab software;

(2) Carry out filter processing to described ECG original signal data:

(2-1) Carry out wavelet decomposition to described ECG original signal: select DB6 wavelet, carry out 8 layers of decomposition to signal, obtain the wavelet coefficient d _i on each scale;

(2-2) Using the improved calculation threshold method, the threshold value of each scale is obtained, and the wavelet coefficient is thresholded:

Among them, T _i is the improved threshold, _i represents the number of wavelet decomposition layers, e is a natural constant, n represents the number of sampling points, and σi is the mean value of the absolute value of wavelet coefficients:

(2-3) Thresholding the signal using a soft threshold method: selecting different thresholds at different scales for thresholding to obtain filtered ECG signals;

(3) According to the difference in the frequency distribution ranges of QRS complexes, P waves, and T waves, select the third and fourth scales with the least overlapping frequency distribution of QRS complexes, P waves, and T waves for wavelet reconstruction, and obtain the reconstructed ECG signal S';

(4) Carry out energy window transformation on the ECG signal reconstructed by wavelet, and select the maximum value point:

(4-1) Energy window transformation: transform the wavelet reconstructed ECG signal S' from the time domain analysis to the energy domain analysis according to the following formula, and obtain the ECG signal energy curve:

Wherein, E _n represents the energy value of the nth sampling point; N is the selected window length, and takes a value of 26; M is the total number of sampling points; S' _n represents the electrocardiographic signal S' after the wavelet reconstruction nth data;

(4-2) Select the maximum value point: the obtained ECG energy curve is subjected to hard thresholding, namely:

Wherein, T _h is the selected threshold value, taking _{Th h} =0.3*median(E _n ),

Then select the peak position of the ECG energy curve after hard thresholding as the maximum point;

(5) Optimizing the maximum point: set two time thresholds t ₁ and t ₂ , and t ₁ <t ₂ When the time interval between any two maximum points is less than t ₁ , remove these two maximum points The one with the smaller amplitude between the two maximum value points; when the time interval between any two maximum value points is greater than t ₂ , look for another unrecognized extreme value point between these two maximum value points; when If the time interval between any two maximum points is greater than t ₁ and smaller than t ₂ , then the two maximum points are retained, and the optimized maximum points are finally obtained, and each of the optimized The maximum point corresponds to a QRS complex;

(6) According to the time point where each maximum value point is located in step (5), search for the signal within the scope of each 7 sampling points at the corresponding time point on the ECG signal after filtering described in step (2). The point with the largest amplitude is taken as the detected R wave.

The present invention applies the sparse automatic encoding deep learning network to the classification of heart beat data, makes full use of its autonomous learning ability and the characteristics of deep feature mining, extracts deeper features of the signal, and then classifies the heart beat data. The specific sparse automatic coding deep learning network constructed by the present invention can make full use of the big data characteristics of ECG signals, dig out the deep-level features of ECG signals, and make it very useful for the analysis of ECG signals of different individuals in complex environments. Good stability, through the design of a reasonable network structure and appropriate network training methods, the automatic classification of ECG signals in complex individuals and complex environments is realized, which solves the problem of dealing with individual differences and ECG signals in complex environments in the prior art. The signal classification algorithm is unstable, and the precise identification of 6 common arrhythmia beats is realized accurately and stably.

In addition, the present invention selects specific wavelet basis functions and wavelet decomposition layers in the wavelet decomposition process of the signal, and adopts an improved threshold value method in the threshold value processing process of the signal, so that the filtered signal can be used to filter out the electrocardiogram signal. While the electromyographic interference, baseline drift and power frequency interference are mixed in, as much useful information as possible is retained, and the phenomenon of over-smoothing of the general threshold is improved. The invention extracts the QRS wave group by performing wavelet reconstruction, and removes the P wave and T wave as noise, effectively avoiding the false detection caused by the tall P wave and T wave in the detection, and improving the detection accuracy . The invention adopts an energy window transformation method to transform the signal into the energy domain for analysis, thereby solving the problem that the signal is easily affected by high-frequency noise and cannot be completely filtered out in the time domain analysis. In the energy window transformation, the present invention fully considers the time domain characteristics of the ECG signal and the spanning time of the QRS complex, and selects the window length. The experimental results prove that only when the window length is 26, the multiple detection caused by the false peak of the noise and the missed detection caused by the low-amplitude QRS wave can be most effectively avoided.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Figure 2 is the original ECG signal.

Figure 3 is the filtered ECG signal.

FIG. 4 is the reconstructed ECG signal.

Fig. 5 is the electrocardiographic signal of the selected maximum value point after hard thresholding processing.

Fig. 6 is a flow chart of optimizing the maximum point.

Figure 7 shows the detected R wave.

Figure 8 is a structural diagram of the sparse auto-encoding deep learning network.

Figure 9 is a diagram of the deep learning network training process.

detailed description

Example 1:

This embodiment is implemented in a computer with Intel Xeon CPU E5-2697@2.70GHz, internal memory 128.00GB, Win7, 64-bit operating system, and the entire ECG signal automatic classification algorithm is implemented in Matlab language.

Implementation process of the present invention is as shown in Figure 1:

a) Obtain the original ECG signal of the human body, perform filtering processing, and detect the R wave of the filtered ECG signal, which specifically operates according to the following steps:

(1) ECG original signal collection: The present invention utilizes the MedSun 18-lead Holter of Beijing Pengyang Fengye to collect the ECG signal of the human body for a long time, and its sampling output frequency is 250 Hz, and the collected ECG data is stored in the form of TXT. It can be easily read into the Matlab environment for display, and its shape is shown in Figure 2.

(2) Filtering the collected ECG raw signal data:

(2-1) Decompose the original ECG signal by wavelet: select the DB6 wavelet of the Daubechies wavelet series, and perform 8-layer decomposition, as shown in Table 1:

Table 1: 8-layer decomposition of DB6 wavelet at a sampling frequency of 250Hz

Scale 250HZ 1 62.5-125 2 31.25-62.5 3 15.625-31.25 4 7.8125-15.625 5 3.90625-7.8125 6 1.953125-3.90625 7 0.9765625-1.953125 8 0.48828125-0.9765625

Then extract the wavelet coefficients d _i on each scale.

(2-2) Adopt the improved method of calculating the threshold value to obtain the threshold value of each scale to obtain the improved threshold value, namely:

··········Formula (I)

In formula (I), i represents the number of wavelet decomposition layers, T _i is the improved threshold, e is a natural constant, n represents the number of sampling points, σ _i is the mean value of the absolute value of wavelet coefficients, and its expression is

Compared with the existing fixed threshold method and maximin threshold method, after the threshold algorithm is improved according to formula (I), the improved threshold has sideband adaptability and maintains good denoising and reconstruction characteristics .

(2-3) Using the soft threshold method, select the corresponding improved threshold on each scale, and perform threshold processing on the ECG signal according to formula (II), namely:

··········Formula (Ⅱ)

where j = i;

Thus, the filtered ECG signal is obtained, as shown in FIG. 3 .

The filtered ECG signal retains as much useful information as possible, improves the over-smoothing phenomenon of the general threshold, and makes the filtering effect more stable.

(3) The frequency distribution range of the normal ECG QRS complex is 5-45Hz. It can be seen from Table 1 that it is mainly concentrated on the 3 and 4 scales, while the frequency distribution range of the P wave and T wave is 0.05 From 10Hz to 10Hz, there is no or only a small amount of distribution on the 3 and 4 scales. Therefore, according to the difference in the frequency distribution range of the QRS complex and the P wave and T wave, select the one with the least frequency overlap between the QRS complex and the P wave and T wave distribution. 3, 4 scales perform wavelet reconstruction on the filtered ECG signal, namely:

·········Formula (Ⅲ)

In formula (Ⅲ), with They are the results of thresholding the ECG signals on scales 3 and 4 according to step (2), respectively.

After wavelet reconstruction, the obtained ECG signal is mainly the information of the QRS complex, which plays a role in highlighting the QRS complex, as shown in Figure 4.

(4) Carry out energy window transformation on the ECG signal after wavelet reconstruction, and select the maximum value point:

(4-1) Energy window transformation: According to the following formula (Ⅳ), the ECG signal S' reconstructed by wavelet is transformed from time domain analysis to energy domain analysis, and the energy curve of ECG signal is obtained:

·········Formula (Ⅳ)

Wherein, E _n represents the energy value of the nth sampling point; N is the selected window length (N=26), M is the total sampling point number, and S' _n represents the electrocardiographic signal after step (3) wavelet reconstruction The nth data of S'.

In the energy window transformation, the selection of the window length is a key, which directly determines whether the R-wave detection algorithm is effective. The present invention fully considers the time-domain characteristics of the electrocardiographic signal and the spanning time of the QRS wave group, and the selection of its N value is determined as follows: the sampling frequency of the electrocardiographic signal in this embodiment is 250 Hz, and the normal QRS wave group Generally, it does not exceed 0.1s, which is 25 sampling points. We choose the window length as an even number, which is 26. The experimental results prove that only when the window length is 26, the multiple detection caused by the false peak of the noise and the missed detection caused by the low-amplitude QRS wave can be most effectively avoided.

In time-domain analysis, the signal is easily affected by high-frequency noise and cannot be completely filtered out in filtering. To solve this problem, the method of energy bed transformation is used to transform the time-domain analysis into energy-domain analysis. The energy domain is more robust to noise than the time domain analysis. As shown in Figure 5, after the energy window transformation, the position of the QRS complex of the signal becomes more prominent, the interval between heart beats is more obvious, and the influence of high-frequency noise is correspondingly weakened.

(4-2) Select the maximum value point: perform hard thresholding on the obtained signal energy curve:

········Formula (Ⅴ)

In the formula (Ⅴ), Th is the selected threshold value, and _{Th h =} 0.3*median(E _n ).

Then select the peak position of the energy curve of the ECG signal after hard thresholding as the maximum point, as shown in FIG. 5 .

(5) Optimizing the maximum point: as shown in the flowchart in Figure 6, set two time thresholds t ₁ and t ₂ , and t ₁ <t ₂ when the time interval between any two maximum points is less than t ₁ , remove the one with the smaller amplitude between the two maximum points; when the time interval between any two maximum points is greater than t ₂ , find another one between the two maximum points An unrecognized extreme point; if the time interval between two extreme points is both greater than t ₁ and less than t ₂ , then the two extreme points are kept, so that each optimized extreme point finally obtained The large value points all correspond to a QRS complex.

In Fig. 6, E _t represents the average value of time intervals of all maximum points obtained in step (4-2), t ₁ =0.5×E _t , t ₂ =1.5×E _t .

(6) According to the time point where each maximum value point is determined in step (5), search for the signal in the range of 7 sampling points around the corresponding time point on the ECG signal after filtering in step (2) The point with the largest amplitude is the detected R wave (Figure 7).

In the actual application process, you can choose other frequencies to collect ECG signals according to your needs, and choose any wavelet from Haar, Daubechies or Symlets to decompose the appropriate number of layers, and determine the subsequent wavelet reconstruction according to the actual situation. Select the scale and the selection of the window length when transforming the energy window. For different application objects, the structure of the deep learning network and the corresponding number of hidden layer nodes should be slightly different. The training method should also be flexible and change, not stick to a fixed form.

b) After obtaining the R wave position, construct a data set. The data set consists of 33,950 sets of heart beat data. Each set of heart beat data has a label. There are 6 labels in total, representing normal heart beat, left bundle branch block, Right bundle branch block, ventricular premature beats, atrial premature beats, confluent beats:

Each group of cardiac beat data includes 270 sampling points, the 270 sampling points are the position of the R wave obtained according to step a), and 90 sampling points are selected in front of the R wave peak point in the filtered electrocardiogram, and 179 sampling points are selected in the back sampling points, that is, each set of heart beat data contains 270 sampling points.

The data set in this embodiment contains 33950 sets of heart beat data from multiple people and the same person in different periods. as a test data set. Both the training dataset and the testing dataset contain six types of labels.

c) Construct a sparse auto-encoding deep learning network (referred to as learning network)

The sparse auto-encoding deep learning network structure is shown in Figure 8, which has two hidden layers (the first hidden layer and the second hidden layer), and a softmax classifier is connected behind the second hidden layer.

The input of the learning network is 270 sampling points, the number of nodes in the first hidden layer is 130 (130 shallow features can be obtained through the first hidden layer), and the number of nodes in the second hidden layer is 50 (through the second hidden layer layer can get 50 high-level features).

d) Step-by-step training of sparse autoencoding deep learning networks

d-1) Normalize the 22350 sets of heart beat data in the training data set:

in x _min and x _max are the maximum and minimum values of the input heart beat amplitude respectively.

Input the normalized cardiac beat data into the SAE model, and use the SAE model to train and learn the first hidden layer of the learning network. The process is shown in Figure 9.

Specifically, the input of the SAE model is a set of heartbeat data (270 sampling points), and the number of hidden layer nodes is 130. Select the sigmoid function f(z)=1/(1+exp(-z)) as the activation function of the neuron; the activation value function of the first hidden layer node is h(g)=f(W ₁ g+b ₁ ) ,in is the weight matrix, is the bias vector. The output of SAE is is the weight matrix, is the bias vector.

In order to make the output Infinitely close to g, introduce a cost function, train the network by minimizing the cost function, get the network weights W ₁ and W ₂ , and get 130 shallow features of the heart beat data, specifically:

When the number of training samples (a normalized heartbeat in the training data set is a training sample) is q, the cost function of SAE is expressed as:

in, Is the sparse penalty factor, β controls the weight of the sparse penalty factor, is the average activation degree of the jth neuron in the hidden layer on q training samples, ρ is the sparsity parameter, we choose λ=3×10 ^-10 , β=3, ρ=0.2;

Then we use the L-BFGS optimization method to minimize the cost function J _W,b (g), and the maximum number of iterations is set to 400, so that we get 130 shallow features of the heartbeat data.

d-2) Input the obtained 130 shallow features into the same SAE model, use the same method to train and learn the second hidden layer of the network, obtain the weight of the second layer network, and obtain 50 high-level features H of heart beat data .

Specifically, the number of nodes in the second hidden layer is 50, and the sigmoid function is also selected as the activation function of neurons, and the activation value function of the nodes in the second hidden layer is H(g)=f(W ₃ h(g)+b ₃ ),in is the weight matrix, is the bias vector. At this time, the output of SAE is is the weight matrix, is the bias vector. In order to make the output Infinitely close to h, introduce the same cost function, train the network by minimizing the cost function to obtain weights W ₃ and W ₄ , and obtain 50 high-level features of heartbeat data, specifically:

When the number of training samples is q, the cost function of SAE is expressed as:

in, Is the sparse penalty factor, β controls the weight of the sparse penalty factor, is the average activation of the jth neuron in the hidden layer on q training samples, and ρ is the sparsity parameter. We choose λ=3×10 ^-10 , β=3, ρ=0.2;

Then we use the L-BFGS optimization method to minimize the cost function J _W,b (g), and the maximum number of iteration steps is set to 400, thus obtaining 50 high-level features of the heart beat data.

d-3) input the deep feature H of the heartbeat data extracted in d-2) into the Softmax classifier, train the softmax classifier, and obtain the network weight of the softmax classifier, as follows:

Softmax classifier can be expressed as:

Each component in r _θ (H ⁽ⁱ⁾ ) p(y ⁽ⁱ⁾ =j|H ⁽ⁱ⁾ ; θ) represents the probability that H ⁽ⁱ⁾ belongs to the jth class, i=1,2,...50 , j is 1,2,...6, θ is the network weight matrix,

The chosen cost function is defined as:

Wherein 1{y ⁽ⁱ⁾ =j} is an indicator function, and the expression in curly braces is true if the indicator function value is 1, otherwise the indicator function value is 0. The part after the plus sign in the above formula is the weight attenuation item added to prevent the model from overfitting, and is selected as α=6×10 ^-7 .

We use the L-BFGS optimization method to fine-tune the network weights of the softmax classifier, minimize the cost function J(θ) by selecting the L-BFGS optimization method, set the maximum number of iterations to 400, and when the algorithm converges, the maximum The label corresponding to the probability value is the beat category predicted by the automatic arrhythmia recognition algorithm.

e) Verification: According to the network weights of the first hidden layer, the network weights of the second hidden layer and the network weights of the softmax classifier obtained in step d), the test data set is input into the learning network to obtain the classification output Heart beat data, to achieve automatic classification of heart beats.

The result of using the method of the present invention to classify the cardiac beat data is compared with the self-label of the cardiac beat data. The classification effect of 6 common heartbeats, premature beats, atrial premature beats, and fusion heartbeats is very good, and all of them have achieved very high classification accuracy.

Table 1 Classification result test table

Label Number of samples Number of correct classifications Accuracy normal beat 3600 3598 99.94% left bundle branch block 2400 2387 99.46% right bundle branch block 2400 2398 99.92% premature ventricular beats 1150 1129 98.17% atrial premature beat 850 830 97.65% confluent heartbeat 1200 1200 100% total 11600 11542 99.50%

Claims (1)

1. a method for building a deep learning network for electrocardiographic signal classification, characterized in that it may further comprise the steps: a) Obtain the ECG signal of the human body, perform filtering processing, and detect the R wave of the filtered ECG signal; b) After the R wave is detected, a data set is constructed. The data set is composed of several sets of heart beat data. Each set of heart beat data has a label. There are 6 types of labels in total, which are normal heart beat, left beam Branch block, right bundle branch block, ventricular premature beats, atrial premature beats, confluent beats: Each group of heartbeat data includes 270 sampling points, the 270 sampling points are based on the position of the detected R wave, 90 sampling points are selected in front of the peak point of the R wave, and at the peak point of the R wave 179 sampling points are selected behind; c) Build a sparse autoencoding deep learning network: The sparse automatic encoding deep learning network has two hidden layers, the two hidden layers are respectively the first hidden layer and the second hidden layer, and a softmax classifier is connected behind the second hidden layer ; The input of the sparse automatic encoding deep learning network is 270 sampling points, the first hidden layer node number is 130, and the second hidden layer node number is 50; d) step-by-step training of the sparse auto-encoding deep learning network: d-1) normalize the beat data of the data set, then input the SAE model, and use the SAE model to train the first hidden layer of the sparse auto-encoding deep learning network to obtain the network of the first hidden layer Weight, and get the shallow features of heart beat data; d-2) adopting the same SAE model, inputting the shallow features into the SAE model, training the second hidden layer of the sparse auto-encoding deep learning network, obtaining the network weights of the second hidden layer, and Obtaining the high-level features of the several groups of cardiac beat data; d-3) Input the obtained high-level features into the softmax classifier, train the softmax classifier, and obtain the network weight of the softmax classifier; e) The sparse auto-encoding deep learning network constructed in step c) is combined with the network weights of the first hidden layer, the network weights of the second hidden layer and the network weights of the softmax classifier obtained in step d), namely A deep learning network for ECG signal classification.