CN105105743B - Electrocardiogram intelligent processing method based on deep neural network - Google Patents

Abstract

The electrocardiogram intelligent diagnosing method based on deep neural network of the present invention, including：A). signal normalization process；B). determine training sample space；C). it is determined that checking sample space；D). determine neural network structure；E). determine activation primitive and object function；F). training neutral net；G). ECG signal is automatically analyzed.The electrocardiogram intelligent diagnosing method based on deep neural network of the present invention, with the data in MIT BIH arrhythmia data bases as sample, Logistic functions are adopted to be trained to neutral net for object function as the activation primitive of neuron, cross entropy cost function, when analyzing to ECG signal, analysis result is drawn by using the neutral net for training, even if diagnostician does not have the clinical experience of very abundant, accurate diagnostic result can also be obtained, the substantial amounts of energy of cardiac specialist need not be expended, mitigates doctor's burden.

Description

Electrocardiogram intelligent processing method based on deep neural network

Technical field

The present invention relates to a kind of electrocardiogram intelligent processing method based on deep neural network, in particular, especially relates to And a kind of deep neural network grader trained based on large sample abnormal electrocardiogram diagram data, can be planted according to contained by sample extremely Class carries out autonomic learning, and construction feature space is so as to reaching the electrocardio based on deep neural network of Intelligent treatment electrocardiogram purpose Figure intelligent processing method.

Background technology

Heart is the vitals of human body, provides power for human recycle system, and blood is transported to body everywhere, and which is good for Whether health directly affects human body various functions.Electrocardiogram is the performance directly perceived of cardiac electric cycle events, and cardiac specialist can pass through Electrocardiogram obtains the bulk information about heart, therefore electrocardiogram clinically has irreplaceable important function.But examine Disconnected electrocardiogram needs the clinical experience of very abundant, while expending the substantial amounts of energy of cardiac specialist.Electrocardiogram automatic parsing algorithm Invention can assist in cardiac specialist heart diagnosed, mitigate doctor's burden.

Existing ecg analysis algorithm relies primarily on feature identification means, by template matching method, morphology operations, little The methods such as wave analysiss are parsed to ECG signal, identify the feature that anomalous ecg is shown in electrocardiosignal, then Construction feature space, is classified to electrocardiogram according to feature space using different algorithm for pattern recognitions, provides electrocardio with this Map analysis result.But the parsing to electrocardiosignal and the Heuristicses of construction feature space requirement cardiac specialist, engineer is not Only need to be grasped these profound medical knowledges, it is necessary to select feature space, expend a large amount of manpowers, and effect is often inadequate It is preferable.

Deep neural network with autonomic learning data, abstract data feature, can set up the feature space of complexity, can be with In solving the problems, such as above-mentioned ecg analysis algorithm well.

Deep neural network is a kind of machine learning method for growing up on the basis of traditional neural network in recent years, it Data are learnt by multi-neuron multiprocessing layer, progressively abstractively represent data, so as to find to be hidden in data Labyrinth.The invention of back-propagating (BP) algorithm greatly improves the learning efficiency of deep neural network so as to business and Sphere of learning has been widely recognized.Deep neural network has in the problems such as Object identifying, natural language processing Breakthrough progress, especially in the big data epoch, it is ensured that deep learning has reliable Data Source.

Deep neural network is made up of three parts, and ground floor is input layer, and last layer is output layer, and middle each layer is referred to as For hidden layer.In addition to input layer, each layer of input is all from the output of last layer, and data are hidden through each by input layer In output layer output result after layer process.Each point in network becomes a neuron, one activation primitive of correspondence, in reality Generally using traditional Sigmoid functions, such as Logistic functions, hyperbolic tangent function in the application of border, and in recent years more often Linear unit R eLU of correction.

The object function of neutral net is also referred to as cost function, and the purpose for training neutral net is exactly to optimize each neuron Between weight, constantly reduce the process of object function.Conventional object function has secondary cost function, cross entropy cost function. Wherein cross entropy cost function generally to be preferred over secondary cost function as the effect of object function.

The optimal solution for solving neutral net object function is needed using gradient descent method (Gradient Descent), and BP Algorithm is to solve for general-purpose algorithm of the object function with regard to the gradient of multilayer neural network weights, by answering to derivation chain rule With it is known that object function can be tried to achieve by the gradient to next layer of input for the gradient of certain layer of input.Therefore, from net The output layer of network recycles above-mentioned rule to input layer, it is possible to solve ladder of the object function to the input above each layer Degree.

Gradient descent method efficiency in the case where training data is excessive is reduced, and a preferable solution is stochastic gradient Descent method (Stochastic Gradient Descent), in training process each time, randomly selects part sample, not only Efficiency can be improved, while can also strengthen the generalization ability of whole network..

The content of the invention

The present invention is in order to overcome the shortcoming of above-mentioned technical problem, there is provided a kind of electrocardiogram intelligence based on deep neural network Can processing method.

The electrocardiogram intelligent processing method based on deep neural network of the present invention, which is particular in that, by following Step is realizing：A). signal normalization process, in MIT-BIH arrhythmia data bases, there is higher letter in choosing data base Make an uproar than pretreatment after signal as training neutral net uses, using formula (1) to selection signal be normalized Process：

Wherein, S is primary signal, and S ' is the signal after normalization, and Min (X) is that sample space is acted on, and is returned The minimum signal of numerical value in sample space；Max (X) is that sample space is acted on, and returns numerical value maximum in sample space Signal；B). determine training sample space X, ECG signal after the normalized of formula (1), all signal numerical value S ∈ Each signal sequence is divided into the input vector x that length is 10000, gives up curtailment 10000 by [0,1] afterwards in order Signal group, finally obtain the training sample space X for meeting input requirements:

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

Each training sample x_iCorresponding electrocardiographic abnormality species vector y_iFor：

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

All of y_iIt is formed the corresponding desired output space Y of training sample space X；C). it is determined that checking sample space, Using with step a) and b) identical method, determine that the checking sample space X ' of inspection and corresponding checking sample expectation are defeated Go out space Y '；D). determine neural network structure, if L1, Ln are respectively the input layer of neutral net, output layer, between L1 to Ln Hidden layer be respectively L2, L3 ..., L (n-1)；L1, L2 ..., the neuron number in Ln layers successively decreases successively, and sets its difference For n1, n2 ..., nn；N1 ＜ n2 ＜ ... ＜ nn；E). determine activation primitive and object function, select as shown in formula (3) Activation primitive of the Logistic functions as neuron；

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

Wherein, x is input ECG signal vector, and w is neural network weight, and b is neutral net bias term, and n is training The sample size of sample, y are the corresponding electrocardiographic abnormality vectors of x, and a is neutral net output vector；F). training neutral net, N training sample is chosen from training sample space X, gradient is calculated using stochastic gradient descent method in the training process, adopted The checking sample space X ' and corresponding checking sample desired output space Y ' determined in step c) tests, and works as accuracy Terminate training during more than 98%, now obtain the neutral net that weight w and bias term b determine；Obtain the heart as shown in formula (5) The functional relation that electrograph exception species is present with ECG signal：

Y=Γ (x；w,b) (5)；

G). ECG signal is automatically analyzed, it is assumed that have an ECG signal to be analyzed, and the sequence after formatting is x ～, by x～bring into formula (7) be obtained analysis result y～.

The electrocardiogram intelligent processing method based on deep neural network of the present invention, in order to further enhance neutral net Generalization ability, improves the accuracy rate to electrocardiographic abnormality classification, solves the problems, such as over-fitting, is handed over using the regularization in formula (6) Fork entropy cost function, the cross entropy cost function in replacement step e) shown in formula (4)：

Wherein, λ ＞ 0, are regularization parameters；

The electrocardiogram intelligent processing method based on deep neural network of the present invention, between the L1 to Ln described in step d) Hiding number of layers be 9, in input layer, 9 hidden layers and output layer, the number of neuron is as shown in the table：

The invention has the beneficial effects as follows：The electrocardiogram intelligent processing method based on deep neural network of the present invention, with Data in MIT-BIH arrhythmia data bases are sample, set up training sample space X and checking sample space X ', adopt Logistic functions are trained to neutral net for object function as the activation primitive of neuron, cross entropy cost function, And the neutral net for training is tested using checking sample space X ', instruction is then thought when success rate exceedes given threshold The neutral net practised has met requirement.When ECG signal to be analyzed to gathering is analyzed, using the neutral net for training Analysis result can be drawn, even if diagnostician does not have the clinical experience of very abundant, can also obtain accurate diagnostic result, The substantial amounts of energy of cardiac specialist need not be expended, mitigates doctor's burden.

Description of the drawings

Fig. 1 is the structure chart of the deep neural network employed in the present invention；

Fig. 2 is the flow chart of training neutral net in the present invention；

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

Specific embodiment

The invention will be further described with embodiment below in conjunction with the accompanying drawings.

The present invention be embodied as be related to two parts, Part I carries out pretreatment to ECG signal, mainly returns One change is processed, and Part II is deep neural network to be trained using the data in MIT-BIH arrhythmia data bases, Jing The neutral net for crossing training can be analyzed to electrocardiogram to be analyzed as grader.

Part I：Pretreatment ECG signal, formats.

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

Signal sequence is divided into input vector x that length is 10000 afterwards in order, gives up curtailment 10000 Signal group, finally obtains the training sample space X for meeting input requirements：

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

Assume electrocardiographic abnormality species and ECG signal existence function relation Γ：

Y=Γ (x) (7)；

Wherein, x is ECG signal vector, then x is the corresponding electrocardiographic abnormality species vectors of x.According to actual needs, Ten kinds of abnormal species that the electrocardiographic abnormality species that we are concerned about is defined as in table 2.I.e.

Each training sample x_iCorresponding electrocardiographic abnormality species vector y_iFor：

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

All of y_iIt is formed the corresponding desired output space Y of training sample space X；

Using above-mentioned identical method, determine that the checking sample space X ' and corresponding checking sample of inspection expect defeated Go out space Y '；The corresponding desired output space Ys of sample space X, and checking sample space X ' are built by above-mentioned functional relationship With corresponding checking sample desired output space Y '.So to training the data needed for neutral net to carry out format analysis processing with full Sufficient practical situation.

As shown in figure 1, giving the structure chart of the deep neural network employed in the present invention, input layer L1 is included 10000 neurons, input data are that ECG signal is vectorial, and output layer L11 includes 10 neurons, and output content is electrocardio Abnormal species vector, totally 9 layers of hidden layer, concrete every layer of neuronal quantity are shown in Table 1.The activation primitive of neural unit is selected Logistic functions：

Wherein, z is the input of each neuron, then outputs of the σ (z) for neuron.

Table 1

As Logistic functions are Sigmoid functions, when initial error is larger, the learning efficiency can be significantly reduced.Therefore Object function in the present invention is from the cross entropy cost function not suffered from this：

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

In order to further enhance the generalization ability of neutral net, the accuracy rate to electrocardiographic abnormality classification is improved, was solved Fitting problems, using L₂Regularization method, i.e., add a regular terms to above-mentioned cross entropy cost function, and referred to as regularization is handed over Fork entropy cost function：

Wherein, λ ＞ 0, are regularization parameters；

Part II：Training neutral net.

As shown in Fig. 2 the flow chart that neutral net is trained in giving the present invention, input sample space X and desired output Neutral net of the space Y in Fig. 1, carries out propagated forward using Logistic functions as neuron activation functions, using public affairs The cross entropy cost function of formula (5) or formula (6) regularization as object function, in the training process using under stochastic gradient Drop method calculates gradient, terminates training when accuracy is more than 98%, now obtains the neutral net that weight w and bias term b determine. Now function gamma is

Y=Γ (x；w,b) (5)

Hypothesis has an ECG signal to be analyzed, and the sequence after formatting is, bring formula (7) into and can be analyzed As a result, categorizing process is as shown in Figure 3.

Claims (3)

1. a kind of electrocardiogram intelligent processing method based on deep neural network, it is characterised in that realized by following steps：

A). signal normalization process, in MIT-BIH arrhythmia data bases, there is more high s/n ratio in choosing data base Signal after pretreatment is used as training neutral net, the signal chosen is normalized using formula (1)：

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

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

B). determine training sample space X, ECG signal after the normalized of formula (1), all signal numerical value S ∈ Each signal sequence is divided into the input vector x that length is 10000, gives up curtailment 10000 by [0,1] afterwards in order Signal group, finally obtain the training sample space X for meeting input requirements:

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

Each training sample x_iCorresponding electrocardiographic abnormality species vector y_iFor：

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

All of y_iIt is formed the corresponding desired output space Y of training sample space X；

C). it is determined that checking sample space, using with step a) and b) identical method, determine the checking sample space of inspection X ' and corresponding checking sample desired output space Y '；

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

E). determine activation primitive and object function, the Logistic functions as shown in formula (3) are selected as the activation of neuron Function；

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

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}{\Σ}\[y\ln a+(1-y)ln(1-a)\]---\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 training sample Sample size, y is x corresponding electrocardiographic abnormalities vector, and a is neutral net output vector；

F). training neutral net, n training sample is chosen from training sample space X, in the training process using stochastic gradient Descent method calculates gradient, using the checking sample space X ' determined in step c) and corresponding checking sample desired output space Y ' tests, and terminates training when accuracy is more than 98%, now obtains the neutral net that weight w and bias term b determine.

2. the electrocardiogram intelligent processing method based on deep neural network according to claim 1, it is characterised in that：In order to The generalization ability of neutral net is further enhanced, the accuracy rate to electrocardiographic abnormality classification is improved, is solved the problems, such as over-fitting, adopted Regularization cross entropy cost function in formula (6), the cross entropy cost function in replacement step e) shown in formula (4)：

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

Wherein, λ ＞ 0, are regularization parameters；

3. the electrocardiogram intelligent processing method based on deep neural network according to claim 1 and 2, it is characterised in that： Hiding number of layers between L1 to Ln described in step d) is 9, neuron in input layer, 9 hidden layers and output layer Number is as shown in the table：