CN105105743A - Intelligent electrocardiogram diagnosis method based on deep neural network - Google Patents

Abstract

The invention discloses an intelligent electrocardiogram diagnosis method based on a deep neural network. The intelligent electrocardiogram diagnosis method comprises the steps that 1, signal normalization processing is performed; 2, a training sample space is determined; 3, a verification sample space is determined; 4, a neural network structure is determined; 5, an activation function and an objective function are determined; 6, the neural network is trained; 7, electrocardiogram signals are automatically analyzed. According to the intelligent electrocardiogram diagnosis method based on the deep neural network, data in an MIT-BIH arrhythmia database are taken as samples, the neural network is trained by adopting a Logistic function as the activation function of neurons and adopting a cross entropy cost function as the objective function, an analysis result can be obtained through the trained neural network when the electrocardiogram signals are analyzed, even if a diagnosis doctor does not have very rich clinical experience, a precise diagnosis result can be acquired without needing to consume a large amount of energy of a cardiologist, and therefore the doctor burden is reduced.

Description

Based on the electrocardiogram intelligent diagnosing method of deep neural network

Technical field

The present invention relates to a kind of electrocardiogram intelligent diagnosing method based on deep neural network, in particular, particularly relate to a kind of deep neural network grader based on the training of large sample abnormal electrocardiogram diagram data, abnormal kind autonomic learning can be carried out, construction feature space thus reach the electrocardiogram intelligent diagnosing method based on deep neural network of intelligent diagnostics electrocardiogram object contained by sample.

Background technology

Heart is the vitals of human body, for human recycle system provides power, blood is transported to health everywhere, and whether its health directly affects human body various functions.Electrocardiogram is the visualize of cardiac electric cycle events, and cardiac specialist can obtain bulk information about heart by electrocardiogram, and therefore electrocardiogram has irreplaceable important function clinically.But diagnosis electrocardiogram needs very abundant clinical experience, expends the energy that cardiac specialist is a large amount of simultaneously.The invention of electrocardiogram automatic parsing algorithm can assist cardiac specialist to diagnose heart, alleviates doctor's burden.

Existing ecg analysis algorithm mainly relies on feature identification means, by methods such as template matching method, morphology operations, wavelet analysises, ECG signal is resolved, identify the feature that anomalous ecg shows in electrocardiosignal, then construction feature space, use different algorithm for pattern recognitions to classify to electrocardiogram according to feature space, provide ecg analysis result with this.But to the parsing of electrocardiosignal and the Heuristics of construction feature space requirement cardiac specialist, engineer not only needs to be grasped these profound medical knowledges, also must select feature space, at substantial manpower, and effect is often not ideal enough.

Deep neural network can autonomic learning data, abstract data feature, can set up complicated feature space, can solve the problem in above-mentioned ecg analysis algorithm well.

Deep neural network is a kind of machine learning method grown up on traditional neural network basis in recent years, and it is learnt data by multi-neuron multiprocessing layer, progressively represents data abstractively, thus finds to hide labyrinth in the data.The invention of back-propagating (BP) algorithm greatly improves the learning efficiency of deep neural network, makes it have been widely recognized in business and sphere of learning.Deep neural network has breakthrough progress in the problems such as Object identifying, natural language processing, especially at large data age, ensure that degree of depth study has reliable Data Source.

Deep neural network is made up of three parts, and ground floor is input layer, and last one deck is output layer, and middle each layer is referred to as hidden layer.Except input layer, the input of every one deck all from the output of last layer, data by input layer after each hidden layer process at output layer Output rusults.Each point in network becomes a neuron, and a corresponding activation primitive, adopts traditional Sigmoid function in actual applications usually, as Logistic function, hyperbolic tangent function, and the linear unit R eLU of correction more conventional in recent years.

The object function of neutral net is also known as making cost function, and the object of neural network training is exactly optimize the weight between each neuron, constantly reduces the process of object function.Conventional object function has secondary cost function, cross entropy cost function.Wherein cross entropy cost function is generally better than secondary cost function as the effect of object function.

The optimal solution solving neutral net object function needs to use gradient descent method (GradientDescent), and BP algorithm solves the general-purpose algorithm of object function about the gradient of multilayer neural network weights, by the application to differentiate chain rule, can know that object function can be tried to achieve by the gradient inputted lower one deck for the gradient of certain layer of input.Therefore, from the output layer of network to input layer, recycle above-mentioned rule, just can solve the gradient of object function to the input above every one deck.

Gradient descent method efficiency when training data is excessive reduces, a good solution is stochastic gradient descent method (StochasticGradientDescent), in training process each time, random selecting part sample, can not only raise the efficiency, also can strengthen the generalization ability of whole network simultaneously.。

Summary of the invention

The present invention, in order to overcome the shortcoming of above-mentioned technical problem, provides a kind of electrocardiogram intelligent diagnosing method based on deep neural network.

Electrocardiogram intelligent diagnosing method based on deep neural network of the present invention, its special feature is, realized by following steps: a). signal normalization process, in MIT-BIH arrhythmia data base, the signal chosen in data base after the pretreatment with more high s/n ratio uses as neural network training, utilizes formula (1) to be normalized the signal chosen:

$S^{\′}=\frac{S-Min\left(X\right)}{Max\left(X\right)-Min\left(X\right)}---\left(1\right)$

Wherein, S is primary signal, and S ' is the signal after normalization, and Min (X), for act on sample space, returns the signal that numerical value in sample space is minimum; Max (X), for act on sample space, returns the signal that numerical value in sample space is maximum; B). determine training sample space X, ECG signal is after the normalized of formula (1), all signal numerical value S ∈ [0,1], each signal sequence is divided in order the input vector x that length is 10000 afterwards, give up the signal group of curtailment 10000, be finally met the training sample space X of input requirements:

X＝{x _i|x _i∈[0,1] ^m,i＝1,2,...,m＝10000}；

Each training sample x _icorresponding electrocardiographic abnormality kind vector y _ifor:

$y_i=\left[\begin{array}{c}{p}_1\\ .\\ .\\ .\\ p_j\\ .\\ .\\ .\\ p_{10}\end{array}\right],j=1,2,\mathrm{...},10,$

Wherein, 10 anomalous ecg kinds are as shown in the table:

All y _ijust define the desired output space Y that training sample space X is corresponding; C). determine verify sample space, adopt with step a) with b) identical method, determine inspection checking sample space X ' and correspondence checking sample desired output space Y '; D). determine neural network structure, if L1, Ln are respectively input layer, the output layer of neutral net, the hidden layer between L1 to Ln be respectively L2, L3 ..., L (n-1); L1, L2 ..., neuron number in Ln layer successively decreases successively, and establish its be respectively n1, n2 ..., nn; N1 < n2 < ... < nn; E). determine activation primitive and object function, the Logistic function of selection as shown in formula (3) is as neuronic activation primitive;

$\mathrm{\&sigma;}\left(Z\right)=\frac{1}{1+e^Z}---\left(3\right)$

The cross entropy cost function as shown in formula (4) is selected to be object function;

$C(x,w,b)=-\frac{1}{n}\underset{x}{\&Sigma;}\&lsqb;y\ln a+(1-y)ln(1-a)\&rsqb;---\left(4\right)$

Wherein, x is input ECG signal vector, and w is neural network weight, and b is neutral net bias term, and n is the sample size of training sample, and y is the electrocardiographic abnormality vector that x is corresponding, and a is neutral net output vector; F). neural network training, n training sample is chosen from training sample space X, adopt stochastic gradient descent method compute gradient in the training process, adopt step c) in the checking sample space X ' that determines and corresponding checking sample desired output space Y ' test, stop training when accuracy is more than 98%, now obtain the neutral net that weight w and bias term b determine; Obtain the functional relation that electrocardiographic abnormality kind as shown in formula (5) and ECG signal exist:

y＝Γ(x；w,b)(5)；

G). the automatic analysis of ECG signal, suppose there is an ECG signal to be analyzed, the sequence after format is will bring formula (7) into and can obtain analysis result

Electrocardiogram intelligent diagnosing method based on deep neural network of the present invention, in order to the generalization ability of further strength neural network, improve the accuracy rate to electrocardiographic abnormality classification, solve over-fitting problem, adopt the regularization cross entropy cost function in formula (6), replace step e) middle cross entropy cost function shown in formula (4):

$C(x,w,b)=C_0+\frac{\mathrm{\&lambda;}}{2n}\underset{w}{\&Sigma;}{w}^2---\left(6\right)$

Wherein, λ > 0 is regularization parameter; $C_0==-\frac{1}{n}\underset{x}{\&Sigma;}\&lsqb;y\ln a+(1-y)ln(1-a)\&rsqb;.$

Electrocardiogram intelligent diagnosing method based on deep neural network of the present invention, steps d) described in L1 to Ln between hidden layer number be 9, in input layer, 9 hidden layers and output layer, neuronic number is as shown in the table:

Layer	L1	L2	L3	L4	L5	L6	L7	L8	L9	L10	L11
												Neuron	10000	5000	2000	1000	800	600	400	200	100	50	10

。

The invention has the beneficial effects as follows: the electrocardiogram intelligent diagnosing method based on deep neural network of the present invention, with the data in MIT-BIH arrhythmia data base for sample, set up training sample space X and checking sample space X ', Logistic function is adopted to be that object function is trained neutral net as neuronic activation primitive, cross entropy cost function, and utilize checking sample space X ' to test to the neutral net trained, then think that the neutral net trained meets the demands when success rate exceedes setting threshold value.When the ECG signal to be analyzed gathered is analyzed, utilize the neutral net trained to draw analysis result, even if diagnostician does not have very abundant clinical experience, also can obtain accurate diagnostic result, without the need to expending a large amount of energy of cardiac specialist, alleviate doctor's burden.

Accompanying drawing explanation

The structure chart of deep neural network of Fig. 1 for adopting in the present invention;

Fig. 2 is the flow chart of neural network training in the present invention;

Fig. 3 is the schematic diagram using housebroken neutral net in the present invention.

Detailed description of the invention

Below in conjunction with accompanying drawing and embodiment, the invention will be further described.

Specific embodiment of the invention relates to two parts, Part I carries out pretreatment to ECG signal, mainly normalized, Part II uses the data in MIT-BIH arrhythmia data base to train deep neural network, and trained neutral net can treat analyzing ecg analysis as grader.

Part I: pretreatment ECG signal, format.

According to formula (1), do normalized, make all signal numerical value S ∈ [0,1];

$S^{\&prime;}=\frac{S-Min\left(X\right)}{Max\left(X\right)-Min\left(X\right)}---\left(1\right)$

Afterwards signal sequence is divided in order the input vector x that length is 10000, gives up the signal group of curtailment 10000, be finally met the training sample space X of input requirements:

X＝{x _i|x _i∈[0,1] ^m,i＝1,2,...,m＝10000}；

Suppose electrocardiographic abnormality kind and ECG signal existence function relation Γ:

y＝Γ(x)(7)；

Wherein, x is ECG signal vector, then x is the electrocardiographic abnormality kind vector that x is corresponding.According to actual needs, the electrocardiographic abnormality kind that we are concerned about is defined as ten kinds of abnormal kinds in table 2.Namely

Each training sample x _icorresponding electrocardiographic abnormality kind vector y _ifor:

$y_i=\left[\begin{array}{c}{p}_1\\ .\\ .\\ .\\ p_j\\ .\\ .\\ .\\ p_{10}\end{array}\right],j=1,2,\mathrm{...},10,$

Wherein, 10 anomalous ecg kinds are as shown in the table:

All y _ijust define the desired output space Y that training sample space X is corresponding;

Adopt above-mentioned identical method, determine the checking sample space X ' of inspection and corresponding checking sample desired output space Y '; Desired output space Y corresponding to sample space X is built by above-mentioned functional relationship, and the checking sample desired output space Y of checking sample space X ' and correspondence '.Like this format analysis processing is carried out to meet practical situation to the data needed for neural network training.

As shown in Figure 1, give the structure chart of the deep neural network adopted in the present invention, input layer L1 comprises 10000 neurons, input data are ECG signal vector, output layer L11 comprises 10 neurons, output content is anomalous ecg kind vector, and hidden layer totally 9 layers, specifically every layer of neuronal quantity is in table 1.The activation primitive of neural unit selects Logistic function:

$\mathrm{\&sigma;}\left(Z\right)=\frac{1}{1+e^Z}---\left(3\right)$

Wherein, z is each neuronic input, then σ (z) is neuronic output.

Table 1

Because Logistic function is Sigmoid function, when initial error is comparatively large, the learning efficiency can significantly reduce.Therefore the object function in the present invention selects the cross entropy cost function do not affected by this:

$C(x,w,b)=-\frac{1}{n}\underset{x}{\&Sigma;}\&lsqb;y\ln a+(1-y)ln(1-a)\&rsqb;---\left(4\right)$

Wherein, x is input ECG signal vector, and w is neural network weight, and b is neutral net bias term, and n is the sample size of training sample, and the electrocardiographic abnormality vector of y corresponding to x, a is neutral net output vector.

In order to the generalization ability of further strength neural network, improve the accuracy rate to electrocardiographic abnormality classification, solve over-fitting problem, use L ₂regularization method, namely adds a regular terms to above-mentioned cross entropy cost function, is referred to as regularization cross entropy cost function:

$C(x,w,b)=C_0+\frac{\mathrm{\&lambda;}}{2n}\underset{w}{\&Sigma;}{w}^2---\left(6\right)$

Wherein, λ > 0 is regularization parameter; $C_0==-\frac{1}{n}\underset{x}{\&Sigma;}\&lsqb;y\ln a+(1-y)ln(1-a)\&rsqb;.$

Part II: neural network training.

As shown in Figure 2, give the flow chart of neural network training in the present invention, input amendment space X and desired output space Y are to the neutral net in Fig. 1, Logistic function is used to carry out propagated forward as neuron activation functions, use the cross entropy cost function of formula (5) or formula (6) regularization as object function, adopt stochastic gradient descent method compute gradient in the training process, stop training when accuracy is more than 98%, now obtain the neutral net that weight w and bias term b determine.Now function gamma is

y＝Γ(x；w,b)(5)

Suppose there is an ECG signal to be analyzed, the sequence after format is bring formula (7) into and can obtain analysis result categorizing process as shown in Figure 3.

Claims (3)

1. based on an electrocardiogram intelligent diagnosing method for deep neural network, it is characterized in that, realized by following steps:

A). signal normalization process, in MIT-BIH arrhythmia data base, the signal chosen in data base after the pretreatment with more high s/n ratio uses as neural network training, utilizes formula (1) to be normalized the signal chosen:

$S^{\&prime;}=\frac{S-Min\left(X\right)}{Max\left(X\right)-Min\left(X\right)}---\left(1\right)$

Wherein, S is primary signal, and S ' is the signal after normalization, and Min (X), for act on sample space, returns the signal that numerical value in sample space is minimum; Max (X), for act on sample space, returns the signal that numerical value in sample space is maximum;

B). determine training sample space X, ECG signal is after the normalized of formula (1), all signal numerical value S ∈ [0,1], each signal sequence is divided in order the input vector x that length is 10000 afterwards, give up the signal group of curtailment 10000, be finally met the training sample space X of input requirements:

X＝{x _i|x _i∈[0,1] ^m,i＝1,2,...,m＝10000}；

Each training sample x _icorresponding electrocardiographic abnormality kind vector y _ifor:

$y_i=\left[\begin{array}{c}{p}_1\\ .\\ .\\ .\\ p_j\\ .\\ .\\ .\\ p_{10}\end{array}\right],j=1,2,\mathrm{...},10,$

Wherein, 10 anomalous ecg kinds are as shown in the table:

All y _ijust define the desired output space Y that training sample space X is corresponding;

C). determine verify sample space, adopt with step a) with b) identical method, determine inspection checking sample space X ' and correspondence checking sample desired output space Y ';

D). determine neural network structure, if L1, Ln are respectively input layer, the output layer of neutral net, the hidden layer between L1 to Ln be respectively L2, L3 ..., L (n-1); L1, L2 ..., neuron number in Ln layer successively decreases successively, and establish its be respectively n1, n2 ..., nn; N1 < n2 < ... < nn;

E). determine activation primitive and object function, the Logistic function of selection as shown in formula (3) is as neuronic activation primitive;

$\mathrm{\&sigma;}\left(Z\right)=\frac{1}{1+e^Z}---\left(3\right)$

The cross entropy cost function as shown in formula (4) is selected to be object function;

$C(x,w,b)=-\frac{1}{n}\underset{x}{\&Sigma;}\&lsqb;y\ln a+(1-y)ln(1-a)\&rsqb;---\left(4\right)$

Wherein, x is input ECG signal vector, and w is neural network weight, and b is neutral net bias term, and n is the sample size of training sample, and y is the electrocardiographic abnormality vector that x is corresponding, and a is neutral net output vector;

F). neural network training, n training sample is chosen from training sample space X, adopt stochastic gradient descent method compute gradient in the training process, adopt step c) in the checking sample space X ' that determines and corresponding checking sample desired output space Y ' test, stop training when accuracy is more than 98%, now obtain the neutral net that weight w and bias term b determine; Obtain the functional relation that electrocardiographic abnormality kind as shown in formula (5) and ECG signal exist:

y＝Γ(x；w,b)(5)；

G). the automatic analysis of ECG signal, suppose there is an ECG signal to be analyzed, the sequence after format is will bring formula (7) into and can obtain analysis result

2. the electrocardiogram intelligent diagnosing method based on deep neural network according to claim 1, it is characterized in that: in order to the generalization ability of further strength neural network, improve the accuracy rate to electrocardiographic abnormality classification, solve over-fitting problem, adopt the regularization cross entropy cost function in formula (6), replace step e) middle cross entropy cost function shown in formula (4):

$C(x,w,b)=C_0+\frac{\mathrm{\&lambda;}}{2n}\underset{w}{\&Sigma;}{w}^2---\left(6\right)$

Wherein, λ > 0 is regularization parameter; $C_0==-\frac{1}{n}\underset{x}{\&Sigma;}\&lsqb;y\ln a+(1-y)ln(1-a)\&rsqb;.$

3. the electrocardiogram intelligent diagnosing method based on deep neural network according to claim 1 and 2, it is characterized in that: steps d) described in L1 to Ln between hidden layer number be 9, in input layer, 9 hidden layers and output layer, neuronic number is as shown in the table:

Layer L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 Neuron 10000 5000 2000 1000 800 600 400 200 100 50 10

。