CN111329445A - Atrial fibrillation identification method based on group convolution residual error network and long-term and short-term memory network - Google Patents

Abstract

The invention discloses an atrial fibrillation identification method based on a group convolution residual network and a long short-term memory network, and relates to the field of atrial fibrillation machine identification. The invention realizes that three channels respectively extract features from the ECG signals of three different frequency bands on the network structure based on the group convolution residual network and the long short-term memory network, and then uses LSTM to perform the feature analysis in the time domain, and finally the ECG signal segments are classified into normal, atrial fibrillation, noisy segments and other beat segments. Using this network model can improve the accuracy of atrial fibrillation identification in the presence of noise interference, shorten the analysis time, and improve the real-time performance of the algorithm. Based on the group convolution residual network structure, the classification accuracy can be improved without increasing the complexity of the parameters. Thanks to the topology of the group convolution block in the residual module, it also reduces the amount of hyperparameter data. .

Description

Atrial fibrillation recognition method based on group convolutional residual network and long short-term memory network

technical field

The invention relates to the technical field of atrial fibrillation machine identification, in particular to an atrial fibrillation identification method based on a group convolution residual network and a long short-term memory network.

Background technique

Atrial fibrillation (AF) is the most common cardiac arrhythmia disease, with a prevalence of around 0.4-1% in the general population, increasing to 8% in people over the age of 80. The appearance of atrial fibrillation symptoms is also closely related to diseases such as coronary heart disease, hypertension and heart failure. Atrial fibrillation itself does not directly threaten the patient's life and health. But if left untreated, atrial fibrillation can cause serious complications, such as heart failure and stroke. Heart failure can seriously affect the quality of life of patients, and stroke is listed as the second leading cause of death in the world by the World Health Organization, both of which seriously threaten people's life and health. Therefore, early detection of atrial fibrillation is essential to prevent the conditions it induces.

The identification methods for atrial fibrillation in the prior art generally adopt the following three methods:

(1) Ruhi Mahajan et al. proposed a feature extraction method combining probabilistic symbol pattern recognition and sample entropy, through which the morphological changes of the ECG signal were characterized, and the atrial fibrillation signal was identified;

(2) Xiaoyan Xu et al. proposed a framework that combines the Modified Frepquency Slice Wavelet Transform (MFSWT) and convolutional neural networks. identify;

(3) Mohamed Limam et al. proposed a convolution-based recurrent neural network, which contains two independent neural networks to extract relevant patterns from ECG and heart rate respectively, and then fuse these patterns into a recurrent neural network, through The recurrent neural network is responsible for extracting the sequence of patterns and then evaluating the final decision through a support vector machine.

The identification method for atrial fibrillation based on the above-mentioned prior art has the defects of long detection time and low real-time monitoring and analysis of atrial fibrillation. In the process of monitoring and detection, relatively clean ECG signals are often required, and it is easy to misreport segments mixed with noise as atrial fibrillation signals, and the recognition accuracy is not high.

SUMMARY OF THE INVENTION

In order to solve the defects of the above technologies, the present invention provides an atrial fibrillation identification method based on a group convolutional residual network and a long short-term memory network.

The technical scheme adopted by the present invention to realize the above technical effect is:

Atrial fibrillation identification method based on group convolutional residual network and long short-term memory network, which includes the following steps:

S1. Data preprocessing, divide the original ECG signal into n segments with a sliding window with a window length of 2500 and a step size of 1, downsample the segmented segments to 250Hz, and then downsample the downsampled signal. Through low-pass filtering, band-pass filtering and high-pass filtering, three frequency band sequences d1, d2 and d3 are obtained;

S2. Input the data of the three frequency band sequences d1, d2 and d3 obtained in the step S1 into the three residual networks for feature extraction in one-to-one correspondence;

S3, the feature signal extracted in the step S2 is input into the LSTM, and its time series feature is extracted in a pattern sequence;

S4. Flatten the output of the LSTM with the fully connected layer, and then use the cross entropy function to calculate the loss;

S5. Determine whether the loss is less than the threshold, and if so, save the data model, otherwise perform backpropagation to continue training the model.

Preferably, in the above-mentioned method for atrial fibrillation identification based on a group convolutional residual network and a long short-term memory network, the residual network is formed by stacking a plurality of residual modules, and the residual module includes a The original single-layer convolution is a group convolution block that performs group convolution.

Preferably, in the above-mentioned method for atrial fibrillation identification based on group convolutional residual network and long short-term memory network, the first part of the residual module is a convolutional layer, a batch normalization layer, and an activation layer; the second part is a group convolution block, batch normalization layer, and activation layer; the third part is a layer of convolution, batch normalization layer, and activation layer, and the data input to the activation layer of the third part is the batch normalization output in the third part The sum of the input data of the data stream and the residual module.

Preferably, in the above-mentioned method for atrial fibrillation identification based on group convolutional residual network and long short-term memory network, the expression of the first part of the residual module is:

Fm _,n =ReLU(BN(conv1D(wm _,n,1 , _xm,n )));

表示残差网络序号，

表示残差模块序号，BN表示批标准化，conv1D(·)表示一维卷积，w_m，n，1表示第m个网络通道第n个残差模块第一部分的卷积核参数，x_m，n表示第m个网络通道第n个残差模块的输入，

ReLU表示激活函数，激活函数的表达式为：

in,

represents the sequence number of the residual network,

Represents the serial number of the residual module, BN represents batch normalization, conv1D( ) represents one-dimensional convolution, w _{m, n, 1} represents the convolution kernel parameter of the first part of the n-th residual module of the m-th network channel, x _{m, n} represents the input of the nth residual module of the mth network channel,

ReLU represents the activation function, and the expression of the activation function is:

Preferably, in the above-mentioned atrial fibrillation identification method based on group convolutional residual network and long short-term memory network, the expression of the second part of the residual module is:

w_i表示每一组的权重参数。where _x =[x ₁ , _{x 2} , . The stacking calculation process of three convolutions in each group in the convolution block, R(x) represents the output result of the group convolution; the group convolution block is expressed as

w _i represents the weight parameter of each group.

三个残差网络的网络通道输出后的拼接向量的表达式为：p＝cat[TH_1，18，TH_2，18，TH_3，18]，

操作表示对向量进行拼接。Preferably, in the above-mentioned method for atrial fibrillation identification based on group convolution residual network and long short-term memory network, the expression of the third part of the residual module is: TH _m,n =ReLU(BN(conv1D( w _m,n,3 , S _m,n ))+x _m,n ) where,

The expression of the splicing vector after the network channel output of the three residual networks is: p=cat[TH _1,18 , TH _2,18 , TH _3,18 ],

An operation means concatenating vectors.

Preferably, in the above-mentioned atrial fibrillation identification method based on group convolutional residual network and long short-term memory network, the vector p after splicing the outputs of the three network channels is input into the forget gate in the LSTM, and the forgetting at time t is obtained. The gate is: f _t =σ(W _f ·cat[h _t-1 , p _t ]+b _f );

Among them, σ is the sigmiod function, W _f is the weight of the forget gate, h _t-1 is the output of the hidden layer of the previous unit, p _t is the input at time t, and b _f is the bias of the forget gate; among them, W _f =d _c ×(d _h +d _p ), d _p is the input dimension, d _h is the hidden layer dimension, and d _c is the dimension of the unit state;

The input gate at time t is: _{i t} =σ(W _i ·cat[h _t-1 , p _t ]+ _bi ); wherein, Wi is the weight of the input gate, and _bi is the bias of the input gate;

其中，w_c为输入单元状态的权重，b_c为输入单元状态的偏置；The cell state of the input at time t is:

Among them, w _c is the weight of the input unit state, and b _c is the bias of the input unit state;

The overall cell state at time t is:

The output gate at time t is: _{o t} =σ(W _o ·cat[h _t-1 , p _t ]+ _bo ; where wo is the weight of the output gate, and b _o is the bias of the output gate;

The overall unit output at time t is: h _t =f _t °tanh(c _t ).

第二层全连接的输出表达式为

然后使用交叉熵函数对最后一个全连接层的输出进行误差计算；Preferably, in the above-mentioned method for atrial fibrillation recognition based on group convolutional residual network and long short-term memory network, in the step S4, the output classification layer of LSTM includes two fully connected layers and one error calculation layer, each The activation functions of are all ReLU, and the output expression of the first fully connected layer is:

The output expression of the second layer full connection is

Then use the cross entropy function to calculate the error of the output of the last fully connected layer;

Among them, d represents the number of neurons in the fully connected layer, h represents the output vector of LSTM, and Y(h) represents the calculation of h and the weight of the fully connected layer.

Preferably, in the above-mentioned method for atrial fibrillation identification based on group convolution residual network and long short-term memory network, in the data preprocessing process of step S1, the cutoff frequency of low-pass filtering is 5 Hz, and the frequency of band-pass filtering is 5 Hz. The cut-off frequencies are 5 Hz and 13 Hz, respectively, and the cut-off frequency of the high-pass filter is 13 Hz.

The beneficial effects of the present invention are as follows: the present invention can improve the classification accuracy without increasing the parameter complexity, benefit from the topology structure of the group convolution block in the residual module, and also reduce the data volume of hyperparameters. On this network structure, three channels are implemented to extract features of ECG signals in three different frequency bands, and then LSTM is used for feature analysis in the time domain. Finally, the ECG signal segments are classified as normal, atrial fibrillation, and noisy. clips and other beat clips. Using this network model can improve the accuracy of atrial fibrillation identification in the case of noise interference, shorten the analysis time, and improve the real-time performance of the algorithm.

Description of drawings

Fig. 1 is the schematic flow chart of the present invention;

Fig. 2 is the network structure diagram of the present invention;

3 is a structural diagram of the residual network and residual module according to the present invention;

FIG. 4 is a structural diagram of a group of convolution blocks according to the present invention.

Detailed ways

For further understanding of the present invention, the present invention will be further described below with reference to the accompanying drawings and specific embodiments of the description.

In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inside", " The orientation or positional relationship indicated by "outside" is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the indicated device or element must have a specific orientation, so as to The specific orientation configuration and operation are therefore not to be construed as limitations of the present invention.

As shown in FIG. 1 , the method for identifying atrial fibrillation based on a group convolutional residual network and a long short-term memory network proposed by an embodiment of the present invention includes the following steps:

S1. Data preprocessing, divide the original ECG signal into n segments with a sliding window with a window length of 2500 and a step size of 1, and downsample the segmented segments to 250Hz, and then downsample the downsampled signal. Through low-pass filtering, band-pass filtering and high-pass filtering, three frequency band sequences d1, d2 and d3 are obtained;

S2. Input the data of the three frequency band sequences d1, d2 and d3 obtained in the step S1 into the three residual networks for feature extraction in one-to-one correspondence;

S3, the feature signal extracted in the step S2 is input into the LSTM, and its time series feature is extracted in a pattern sequence;

S4. Flatten the output of the LSTM with the fully connected layer, and then use the cross entropy function to calculate the loss;

S5. Determine whether the loss is less than the threshold, and if so, save the data model, otherwise perform backpropagation to continue training the model.

The network structure constructed by the present invention based on the group convolution residual network and the long-term and short-term memory network is shown in Figure 2. The three frequency band sequences d1, d2 and d3 obtained from the data processing of the original ECG signal are respectively input to the three frequency bands in one-to-one correspondence. In the residual network, three residual networks are used for feature extraction, and then the LSTM, that is, a long short-term memory network, is used to extract its time series features in a pattern sequence, and then the output of the LSTM is tiled through two fully connected layers.

As shown in Figure 3, it is a structural diagram of the residual network and residual module. In a preferred embodiment of the present invention, as shown in the left figure in Figure 3, the residual network is stacked by a plurality of residual modules made. In a preferred embodiment of the present invention, the number of residual modules is 18, and each residual module includes a group convolution block for performing group convolution on the original single-layer convolution. As shown in the right figure in Figure 3, the first part of the residual module is composed of convolution with kernel size S and channel number F, batch normalization (BN) and activation layer (activation function is ReLU) , the second part of the residual module is composed of group convolution blocks, batch normalization layers and activation layers, and the third part of the residual module is composed of convolution kernel size S, channel number F and batch normalization layer and the activation layer. The data input to the activation layer of the third part is the sum of the data stream of the batch normalized output in the third part and the input data of the residual module.

As shown in FIG. 4, it is the structure diagram of the group of convolution blocks according to the present invention. The group of convolution blocks divides the original single convolution process into 32 groups for convolution respectively, and each group consists of three convolution blocks. The size of the convolution kernel is 1, 3 and 1, respectively, and the convolution stride is 1.

Specifically, in a preferred embodiment of the present invention, the expression of the first part of the residual module is:

Fm _,n =ReLU(BN(conv1D(wm _,n,1 , _xm,n )));

表示残差网络序号，

表示残差模块序号，BN表示批标准化，conv1D(·)表示一维卷积，w_m，n，1表示第m个网络通道第n个残差模块第一部分的卷积核参数，x_m，n表示第m个网络通道第n个残差模块的输入，

ReLU表示激活函数，激活函数的表达式为：

in,

represents the sequence number of the residual network,

Represents the serial number of the residual module, BN represents batch normalization, conv1D( ) represents one-dimensional convolution, w _{m, n, 1} represents the convolution kernel parameter of the first part of the n-th residual module of the m-th network channel, x _{m, n} represents the input of the nth residual module of the mth network channel,

ReLU represents the activation function, and the expression of the activation function is:

组卷积块可以概括性地表示为

The second part of the residual module is a group convolution block, batch normalization layer, and activation layer, which can be expressed as:

The group convolution block can be generally expressed as

将组卷积块的表达式

变换为一般的形式为：

其中，T_i(x)表示组卷积块中每一组中三个卷积的堆叠计算过程，R(x)表示组卷积的输出结果。将公式

中的x替换成F_m，n，即可得到公式

_Among them, _x =[x ₁ , x ₂ , . The weight parameter, that is, the convolution kernel. The final combination of each group of group convolution blocks is used

expression to group convolution blocks

Converted to the general form:

Among them, T _i (x) represents the stacking calculation process of three convolutions in each group in the group convolution block, and R(x) represents the output result of the group convolution. put the formula

Replace x in F _{m, n} to get the formula

操作表示对向量进行拼接。The expression of the third part of the residual module is: TH _m,n =ReLU(BN(conv1D(w _m,n,3 ,S _m,n ))+x _m,n ); a network of three residual networks The expression of the splicing vector after channel output is: p=cat[TH _1,18 , TH _2,18 , TH _3,18 ],

An operation means concatenating vectors.

Specifically, in a preferred embodiment of the present invention, the vector p after splicing the outputs of the three network channels is input into the forget gate in the LSTM, and the forget gate at time t is obtained as: f _t =σ(W _f ·cat[ h _t-1 , p _t ]+b _f );

Among them, σ is the sigmiod function, W _f is the weight of the forget gate, h _t-1 is the output of the hidden layer of the previous unit, p _t is the input at time t, and b _f is the bias of the forget gate; among them, W _f =d ^c ×(d ^h +d ^p ), d ^p is the input dimension, d _h is the hidden layer dimension, and d _c is the dimension of the unit state;

The input gate at time t is: _{i t} =σ(W _i ·cat[h _t-1 , p _t ]+ _bi ); wherein, Wi is the weight of the input gate, and _bi is the bias of the input gate;

其中，w_c为输入单元状态的权重，b_c为输入单元状态的偏置；The cell state of the input at time t is:

Among them, w _c is the weight of the input unit state, and b _c is the bias of the input unit state;

The overall cell state at time t is:

The output gate at time t is: o _t =σ(W _o ·cat[h _t-1 , p _t ]+ _{bo ); among them, wo is the weight of the output gate, and b o is} the bias of the output gate;

The overall unit output at time t is: h _t =f _t °tanh(c _t ).

第二层全连接的输出表达式为

然后使用交叉熵函数对最后一个全连接层的输出进行误差计算。In the step S4, the output classification layer of the LSTM includes two fully connected layers and one error calculation layer, the respective activation functions are ReLU, and the output expression of the first fully connected layer is:

The output expression of the second layer full connection is

The error calculation is then performed on the output of the last fully connected layer using the cross-entropy function.

Among them, d represents the number of neurons in the fully connected layer, h represents the output vector of LSTM, and Y(h) represents the calculation of h and the weight of the fully connected layer.

Specifically, in a preferred embodiment of the present invention, in the data preprocessing process of step S1, the cut-off frequency of low-pass filtering is 5 Hz, the cut-off frequency of band-pass filtering is 5 Hz, 13 Hz, and the cut-off frequency of high-pass filtering is 13 Hz. .

The experimental results show that the ECG monitor and ECG machine implanted in the algorithm model can monitor the ECG state of the subject in real time. After the subject has atrial fibrillation signal, the monitor can accurately alarm in about 5 seconds at the fastest. Real-time performance has been effectively improved. At the same time, the false alarm rate is effectively reduced when the ECG signal interference is large.

To sum up, the present invention can improve the classification accuracy without increasing the parameter complexity, benefit from the topology of the group convolution block in the residual module, and also reduce the data volume of hyperparameters. On this network structure, three channels are implemented to extract features from the ECG signals of three different frequency bands, and then LSTM is used to analyze the features in the time domain. Finally, the ECG signal segments are classified as normal, atrial fibrillation, and noisy. clips and other beat clips. Using this network model can improve the accuracy of atrial fibrillation identification in the case of noise interference, shorten the analysis time, and improve the real-time performance of the algorithm.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above-mentioned embodiments. The above-mentioned embodiments and descriptions describe only the principles of the present invention. Without departing from the spirit and scope of the present invention, there are various Variations and improvements all fall within the scope of the claimed invention, which is defined by the appended claims and their equivalents.

Claims (9)

1. The atrial fibrillation identification method based on the group convolution residual error network and the long-short term memory network is characterized by comprising the following steps of:

s1, preprocessing data, namely segmenting the original electrocardiosignals into n segments by using a sliding window with the window length of 2500 and the step length of 1, down-sampling the segmented segments to 250Hz, and then performing low-pass filtering, band-pass filtering and high-pass filtering on the down-sampled signals to obtain three frequency band sequences d1, d2 and d 3;

s2, correspondingly inputting the data of the three frequency band sequences d1, d2 and d3 obtained in the step S1 into three residual error networks one by one for feature extraction;

s3, inputting the characteristic signal extracted in step S2 to LSTM, and extracting the time-series characteristic thereof in a pattern sequence;

s4, tiling the output of the LSTM by using a full connection layer, and then calculating the loss by using a cross entropy function;

and S5, judging whether the loss is less than a threshold value, if so, saving the data model, and otherwise, performing back propagation to continue training the model.

2. The method according to claim 1, wherein the residual network is formed by stacking a plurality of residual modules, and the residual modules comprise a group convolution block for performing a group convolution on the original single-layer convolution.

3. The method for identifying atrial fibrillation according to claim 2, wherein the first part of the residual error module is a convolutional layer, a batch normalization layer and an activation layer; the second part is a group rolling block, a batch standardization layer and an activation layer; the third part is a layer of convolution, batch normalization layer and activation layer, and the data input into the activation layer of the third part is the sum of the data stream output by batch normalization in the third part and the input data of the residual module.

4. The method of claim 3, wherein the expression of the first part of the residual error module is:

F_m，n＝ReLU(BN(conv1D(w_m，n，1，x_m，n)))；

wherein ,

a residual network sequence number is indicated,

denotes the residual module number, BN denotes batch normalization, conv1D (. cndot.) denotes one-dimensional convolution, w_m，n，1Convolution kernel parameter, x, representing the first part of the nth residual block of the mth network channel_m，nRepresenting the input of the nth residual block of the mth network channel,

ReLU represents the activation function, whose expression is:

5. the method of claim 4, wherein the expression of the second part of the residual error module is:

wherein x is [ x ]₁，x₂，...，x_g]Representing the input of each group in the group convolution block, g representing the number of groups convolved by the group, i representing the ordinal number, T_i(x) A stack computation process representing the three convolutions in each of the groups of convolution blocks, R (x) represents the output of the group convolution; the group volume block is represented as

w_iThe weight parameter of each group is represented.

6. The method of claim 5, wherein the expression of the third part of the residual error module is: TH_m，n＝ReLU(BU(conv1D(w_m，n，3，S_m，n))+x_m，n), wherein ,

the expression of the splicing vector after the network channels of the three residual error networks output is as follows: p ═ cat [ TH ]_1，18，TH_2，18，TH_3，18]，

The operation represents stitching the vectors.

7. The atrial fibrillation recognition method based on the group convolution residual error network and the long-short term memory network as claimed in claim 6, wherein the vector p obtained by splicing the outputs of the three network channels is input to a forgetting gate in the LSTM, and the forgetting gate at the time t is obtained by: f. of_t＝σ(W_f·cat[h_t-1，p_t]+b_f)；

Wherein σ is a sigmiod function, W_fWeight of forgetting gate, h_t-1For the hidden layer output of the last cell, p_tIs an input at time t, b_fA bias for a forgetting gate; wherein, W_f＝d_cx(d_h+d_p)，d_pTo input dimension, d_hTo hide the layer dimension, d_cDimension that is the cell state;

the input gates at time t are: i.e. i_t＝σ(W_i·cat[h_t-1，p_t]+b_i)； wherein ,W_iAs the weight of the input gate, b_iIs the bias of the input gate;

the input cell state at time t is:

wherein ,w_cAs weights of the states of the input cells, b_cA bias that is an input cell state;

the overall unit state at time t is:

the output gate at time t is: o_t＝σ(W_o·cat[h_t-1，p_t]+b_o； wherein ,w_oAs weights of output gates, b_oIs the offset of the output gate;

the overall unit output at time t is:

8. the method for atrial fibrillation recognition based on the group convolution residual error network and the long-short term memory network, wherein in step S4, the output classification layer of the LSTM includes two fully-connected layers and an error calculation layer, the respective activation functions are ReLU, and the output expression of the first fully-connected layer is:the output expression of the second layer full connection is

Then, performing error calculation on the output of the last full-connection layer by using a cross entropy function;

where d represents the number of neurons in the fully connected layer, h represents the output vector of LSTM, and y (h) represents the calculation of the weights of h and the fully connected layer.

9. The method for atrial fibrillation recognition based on the group convolution residual error network and the long-short term memory network of claim 1, wherein during the data preprocessing of step S1, the cut-off frequency of the low-pass filtering is 5Hz, the cut-off frequency of the band-pass filtering is 5Hz and 13Hz respectively, and the cut-off frequency of the high-pass filtering is 13 Hz.

Images