CN102697491B - Identification method and system of characteristic waveform of electrocardiogram - Google Patents

Abstract

The invention discloses identification method and system of a characteristic waveform of an electrocardiogram. The identification system comprises at least two neural network modules and a fuzzy logic mode identification module, wherein a characteristic waveform identification algorithm of each neural network module is different from all characteristic waveform identification algorithms of other neural network modules, the neural network modules operate an input electrocardiogram signal according to the characteristic waveform identification algorithms and output by identifying the characteristic waveforms from operated signals; the fuzzy logic mode identification module is used for receiving the characteristic waveforms output by each neural network module and setting medical history weight values according to the characteristic waveforms, and the fuzzy logic mode identification module carries out weighted operation on the priority levels of the characteristic waveforms and then correspondingly outputs the characteristic waveforms and the weighted priority levels of the characteristic waveforms. In the invention, neural networks with different characteristic waveform identification algorithms are adopted to identify the characteristic waveforms, thereby extending the characteristic waveforms capable of being identified by a single neural network and achieving the purpose of more accurate analysis and classification of the electrocardiogram signals.

Description

Characteristics of electrocardiogram method for waveform identification and system

Technical field

The present invention relates to medical instruments field, particularly relate to the method and system of the Electrocardiographic signature waveform of a kind of aid identification.

Background technology

Mode identification technology is widely used in medical field, such as, carries out pattern recognition, therefrom analyze ECG signal, sorts out typical signature waveform, so that doctor is according to analyzing, sorting out the signature waveform and judge patient.

The main method of pattern recognition has: 1. statistical-simulation spectrometry; 2. syntax pattern distinguishment; 3. Fuzzy Pattern Recognition; 4. logic method; 5. neural network.Wherein, statistical-simulation spectrometry, to baroque pattern extraction feature difficulty, can not reflect the architectural feature of pattern, be difficult to the character of description scheme, is difficult to consider identification problem from overall angle.If syntactic pattern recognition method exists interference and noise when using, then extract feature primitive difficulty, and easily slip up.The accurately rational membership function of Fuzzy Pattern Recognition is often difficult to set up, and limits its application.Logic method has defect at sample, and background is unintelligible, and rule is indefinite when even having an ambiguity, and effect is bad.And neutral net rule can to process some environmental informations very complicated, background knowledge is unclear, the indefinite problem of rule of inference, therefore, is more prone to use neural network carrying out analyzing to ECG signal with when sorting out.But, in the prior art, neural network still has some problems, its major defect is that model is few, the pattern (i.e. signature waveform) that can identify is insufficient, therefore, the neural network of prior art likely occurs identifying some signature waveform when carrying out analysis classification to ECG signal, and causes the accuracy of ECG signal analysis low.

Summary of the invention

Embodiments provide a kind of Characteristics of electrocardiogram method for waveform identification and system, in order to realize analyzing classification more accurately to ECG signal.

According to an aspect of the present invention, provide a kind of Characteristics of electrocardiogram waveform recognition system, comprising:

At least two neural network modules, the signature waveform recognizer of each neural network module is all not identical with the signature waveform recognizer of other arbitrary neural network module, described neural network module is used for carrying out computing to the ECG signal of input according to its signature waveform recognizer, identifies signature waveform and export from the signal calculated;

Fuzzy logic pattern recognition module, for receiving the signature waveform that each neural network module exports, according to the error size between each signature waveform and described ECG signal, calculates the priority of each signature waveform respectively; And according to the medical history weighted value arranged for each signature waveform, after the priority of each signature waveform is computed weighted, signature waveform is carried out corresponding output with the priority after its weighting.

Wherein, described neural network module is specially two; Signature waveform recognizer in two neural network modules is respectively: based on the signature waveform recognizer of incremental learning, the signature waveform recognizer based on batch study.

Described medical history weighted value is arranged according to medical history situation.

Described fuzzy logic pattern recognition module, according to the medical history weighted value arranged for each signature waveform, computes weighted to the priority of each signature waveform and is specially:

Described fuzzy logic pattern recognition module is according to the medical history weighted value arranged for each signature waveform, determine fuzzy matrix: the corresponding each signature waveform respectively of each row in described fuzzy matrix, the value of all elements often in row is the medical history weighted value of this row characteristic of correspondence waveform; The priority of described fuzzy matrix to each signature waveform is used to compute weighted.

According to another aspect of the present invention, additionally provide a kind of Characteristics of electrocardiogram method for waveform identification, comprising:

The ECG signal of neural network module to input carries out computing according to its signature waveform recognizer, identifies signature waveform and export from the signal calculated; Wherein, described neural network module is at least two, and the signature waveform recognizer of each neural network module is all not identical with the signature waveform recognizer of other arbitrary neural network module;

Fuzzy logic pattern recognition module receives the signature waveform that each neural network module exports, and according to the error size between each signature waveform and described ECG signal, calculates the priority of each signature waveform respectively; And according to the medical history weighted value arranged for each signature waveform, after the priority of each signature waveform is computed weighted, signature waveform is carried out corresponding output with the priority after its weighting.

Wherein, described neural network module is specially two; Signature waveform recognizer in two neural network modules is respectively: based on the signature waveform recognizer of incremental learning, the signature waveform algorithm based on batch study.

Preferably, the training method of the neural network module of the described signature waveform recognizer based on incremental learning is:

By supposing that primary iteration is n '-T during a simulation learning, final iteration is that current iteration n' realizes, wherein, n ' natural number, T is a positive integer constant, and T+1 is hypothesis iteration siding-to-siding block length, if n ' is-T<0, so primary iteration is 0;

When forward direction iteration n', mean square error E _n'calculate according to following formula 11:

$E_{n^{\&prime;}}=\frac{1}{2}{(d_{n^{\&prime;}}-y_{n^{\&prime;}})}^T(d_{n^{\&prime;}}-y_{n^{\&prime;}})$ (formula 11)

D in above-mentioned formula 11 _n'mean Vector, y _n'it is output vector;

Approximate Lagrange's multiplier is calculated by the rear Iterative feedback method of following formula 12

${\widehat{\mathrm{\&lambda;}}}_n=\frac{{\&PartialD;}^{+}{E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_n}=\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_n}+\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{\widehat{\mathrm{\&lambda;}}}_{n+j}\frac{\&PartialD;{\hat{y}}_{n+j}}{\&PartialD;{\hat{y}}_n},$ N=n'-1 ..., n'-T, (formula 12)

In above-mentioned formula 12, L is lead-out terminal number, it is approximate output vector;

Formula 13-15 is used to set up boundary condition:

${\widehat{\mathrm{\&lambda;}}}_{n^{\&prime;}}=-(d_{n^{\&prime;}}-y_{n^{\&prime;}})$ (formula 13)

${\widehat{\mathrm{\&lambda;}}}_{n^{\&prime;}+1},...,{\widehat{\mathrm{\&lambda;}}}_{n^{\&prime;}+L}=0$ (formula 14)

$\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_{n^{\&prime;}-T}},\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_{n^{\&prime;}-T+1}}...,\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_{n^{\&prime;}-1}}=0$ (formula 15)

Error of approximation gradient is obtained according to formula 16:

$\frac{\widehat{{\&PartialD;}^{+}{E}_{n^{\&prime;}}}}{\&PartialD;w\left(n^{\&prime;}\right)}=\underset{n=n^{\&prime;}-T}{\overset{n^{\&prime;}}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}^{n^{\&prime;}-n}{\widehat{\mathrm{\&lambda;}}}_n\frac{\&PartialD;{\hat{y}}_n}{\&PartialD;w\left(n\right)},$ (formula 16)

In above-mentioned formula 16,0< α <1 is the exponential form weight coefficient of setting;

According to formula 17 by n-th ' secondary iteration time the weighted value w (n'+1) of weighted value when calculating the n-th '+1 time iteration:

$w(n^{\&prime;}+1)=w\left(n^{\&prime;}\right)-\mathrm{\&eta;}\frac{\widehat{{\&PartialD;}^{+}{E}_{n^{\&prime;}}}}{\&PartialD;w\left(n^{\&prime;}\right)}$ (formula 17)

In above-mentioned formula 17, η is the adjustment positive number of setting;

In above-mentioned formula 12 determine according to following formula 18,19,20:

$W_0=0.7\&times;s^{\frac{1}{r}}\&times;\mathrm{normr}(2\mathrm{rand}(s,r)-I(s,r))$ (formula 18)

In above-mentioned formula 18, W ₀represent the original state of weight matrix W, s is that network implies the number of plies, and r is the number of the input amendment of neural network module.The equally distributed random number matrix that rand (s, r) arranges for the capable r of s, I (s, r) is all 1's matrix of s capable r row, normr() be the standardization normalizing matrix of matrix;

Calculate hidden node according to following formula 19 to export:

$o_j=f(\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}\&times;x_i-q_j)$ (formula 19)

In above-mentioned formula 19, w _ijbe the weighted value connecting i-th node and a jth node in the neutral net of neural network module, i, j are the integer being less than or equal to the total node number of neutral net; q _jfor a jth neural unit threshold value of setting; F is predetermined non-linear action function;

The approximate output of a kth output node when n-th iteration for:

${\hat{y}}_{\mathrm{nk}}=f(\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{jk}}\&times;o_j-q_k)$ (formula 20)

In formula 20, k is the positive integer being less than or equal to output layer interstitial content; During n-th iteration, the approximate output of each output node form described approximate output vector

Described employing carries out computing based on the ECG signal of neural network module to input of the signature waveform recognizer of incremental learning according to its signature waveform recognizer, identifies signature waveform and carry out output and be specially from the signal calculated:

Described employing based on the neural network module of the signature waveform recognizer of incremental learning according to the weight matrix W determined in its signature waveform recognizer and described training process, computing is carried out to the ECG signal of input, from the signal calculated, identifies signature waveform export; Element in described weight matrix W is the weighted value that last iteration obtains.

Described medical history weighted value is arranged according to medical history and health condition.

Described basis is the medical history weighted value that each signature waveform is arranged, and computes weighted, specifically comprise the priority of each signature waveform:

According to the medical history weighted value arranged for each signature waveform, determine fuzzy matrix: the corresponding each signature waveform respectively of each row in described fuzzy matrix, the value of all elements often in row is the medical history weighted value of this row characteristic of correspondence waveform;

The priority of described fuzzy matrix to each signature waveform is used to compute weighted.

The embodiment of the present invention carries out signature waveform identification due to the neutral net that have employed multiple different characteristic waveform recognition algorithm, like this, just extends the signature waveform that single neutral net can identify; That is, Characteristics of electrocardiogram waveform recognition system can be different algorithm, method, angle carry out recognition feature waveform, reach to ECG signal more accurate analysis sort out object, thus Electrocardiographic signature waveform can be provided as much as possible for doctor, judge patient on this basis.

Further, according to the difference between each signature waveform and described ECG signal in the embodiment of the present invention, determine that the priority of each signature waveform knows the degree of the reference reliability of each signature waveform for doctor, for Electrocardiographic analysis provides higher accuracy.

Further, the fuzzy logic pattern recognition module of the embodiment of the present invention, by arranging medical history weighted value for each signature waveform, shows the probability that disease corresponding to each signature waveform or associated situation occur in medical history; Thus after using the priority of medical history weighted value to each signature waveform arranged to compute weighted, obtain the priority after the weighting of each signature waveform, further can reflect the probability that present this signature waveform occurs, thus provide higher accuracy for Electrocardiographic analysis.Supply doctor's reference by corresponding for the priority after each signature waveform and weighting thereof output display to display, the signature waveform that doctor can be higher to the priority after weighting preferentially can be analyzed, consider, more contributes to doctor to Electrocardiographic analysis.

Accompanying drawing explanation

Fig. 1 is the Characteristics of electrocardiogram waveform recognition system structured flowchart of the embodiment of the present invention;

Fig. 2 is the Characteristics of electrocardiogram method for waveform identification flow chart of the embodiment of the present invention;

Fig. 3 is the Characteristics of electrocardiogram waveform recognition system structured flowchart of the embodiment of the present invention.

Detailed description of the invention

For making object of the present invention, technical scheme and advantage clearly understand, enumerate preferred embodiment referring to accompanying drawing, the present invention is described in more detail.But it should be noted that, the many details listed in description are only used to make reader to have a thorough understanding, even if do not have these specific details also can realize these aspects of the present invention to one or more aspect of the present invention.

The term such as " module " used in this application, " system " is intended to comprise the entity relevant to computer, such as but not limited to hardware, firmware, combination thereof, software or executory software.Such as, module can be, but be not limited in: the thread of the process that processor runs, processor, object, executable program, execution, program and/or computer.For example, application program computing equipment run and this computing equipment can be modules.One or more module can be positioned at an executory process and/or thread.

The neural network of the present inventor to prior art is analyzed, and find that the method for neural network training is different, then the pattern that the neutral net trained can identify is not identical yet.When the neural network of prior art carries out pattern recognition, the pattern that can identify is limited, is applied to precision of analysis when analyzing ECG signal low.For the neural network of prior art, main thought of the present invention is when analyzing ECG signal, adopts multiple neutral net, and each neutral net obtains based on different training methodes; The pattern that these neutral nets comprehensive identify, namely these neutral nets comprehensive sort out the signature waveform, thus extend the pattern recognition limitation of single neutral net, thus improve the accuracy analyzed ECG signal.

The concrete technical scheme of the embodiment of the present invention is described in detail below in conjunction with accompanying drawing.The Characteristics of electrocardiogram waveform recognition system that the embodiment of the present invention provides, as shown in Figure 1, comprising: fuzzy logic pattern recognition module 101 and at least two neural network modules 102.

Wherein, the signature waveform recognizer in each neural network module 102 is all not identical with the signature waveform recognizer of other arbitrary neural network module 102.That is, the neural network module of various features waveform recognition algorithm is comprised at Characteristics of electrocardiogram waveform recognition system.Characteristics of electrocardiogram waveform recognition system carries out the method for Characteristics of electrocardiogram waveform recognition, and idiographic flow as shown in Figure 2, comprises the steps:

S201: trained each neural network module 102 receives the ECG signal of input in Characteristics of electrocardiogram waveform recognition system.

Each neural network module 102 in Characteristics of electrocardiogram waveform recognition system all receives ECG signal, and carries out computing to it.

S202: each neural network module 102 carries out computing according to the ECG signal of its signature waveform recognizer to input, obtains the signal calculated.

Different neural network modules 102 takes on a different character waveform recognition algorithm, and each neural network module 102 carries out computing according to the ECG signal of respective adopted signature waveform recognizer to input respectively, obtains the signal calculated respectively.

S203: each neural network module 102 identifies signature waveform and exports from the signal calculated.

Each neural network module 102 carries out computing in the ECG signal of the signature waveform recognizer adopted according to it to input, after obtaining the signal calculated, then from the signal calculated, identify the signal waveform that the signature waveform predetermined with some match export as the signature waveform identified.

Like this, just extend the signature waveform that single neutral net can identify, algorithm that can be different, method, angle carry out recognition feature waveform, more for doctor provides Electrocardiographic signature waveform as much as possible, judge patient on this basis.

In fact, ECG signal is cyclical signal often, the signature waveform recognizer that neural network module 102 adopts according to it carries out real-time operation to ECG signal, the signal calculated obtained also can be periodic, again the signal periodically calculated is analyzed, therefrom identify the signal waveform that the signature waveform predetermined with some match to export as the signature waveform identified, then the signature waveform exported also is periodically.For example, within an ECG signal cycle, neural network module 102 identifies two signature waveforms and exports, and so within the next ECG signal cycle, neural network module 102 also can continue to identify and export this two signature waveforms.

S204: fuzzy logic pattern recognition module 101 receives the signature waveform that each neural network module 102 exports.

S205: each signature waveform received and original ECG signal compare by fuzzy logic pattern recognition module 101, according to the error between each signature waveform and described ECG signal, calculate the priority of each signature waveform respectively.

Although multiple neural network module 102 can identify signature waveform more, but because trained neural network module is subject to the restriction of self algorithm, may identify the signature waveform differed greatly with original electrocardiographicdigital figure signal in signature waveform identifying, and the reliability of this signature waveform reflection electrocardiogram truth is just very low.So, if doctor directly carries out Electrocardiographic judgement according to these signature waveforms may occur erroneous judgement.Therefore, according to the difference between each signature waveform and described ECG signal, determine that the priority of each signature waveform knows the degree of the reference reliability of each signature waveform for doctor, for Electrocardiographic analysis provides higher accuracy.How to judge the difference between signature waveform and ECG signal, and determine what the priority of each signature waveform was well known to those skilled in the art according to difference, repeat no more herein.

Because ECG signal is periodically, therefore, neural network module 102 also can periodicity output characteristic waveform.Fuzzy logic pattern recognition module 101 specifically compares for each signature waveform received in the one-period of ECG signal and original ECG signal, according to the error between each signature waveform and described ECG signal, calculate the priority of each signature waveform respectively.

S206: signature waveform, according to the medical history weighted value arranged for each signature waveform, after computing weighted, is carried out corresponding output with the priority after its weighting to the priority of each signature waveform by fuzzy logic pattern recognition module 101.

For the signature waveform that multiple neural network module 102 identifies, although according to the difference between each signature waveform and described ECG signal in above-mentioned steps S205, determine that the priority of each signature waveform can provide higher accuracy for Electrocardiographic analysis; But, in a practical situation, current condition often with its before medical history have phase relation, if with reference to the medical history before it, can further for Electrocardiographic analysis provides higher accuracy.

Therefore, by fuzzy logic pattern recognition module 101, medical history weighted value can be set for each signature waveform, show the probability that disease corresponding to each signature waveform or associated situation occur in medical history.Particularly, described medical history weighted value can be arranged according to medical history and health condition.After using the priority of medical history weighted value to each signature waveform arranged to compute weighted, obtain the priority after the weighting of each signature waveform, further can reflect the probability that in medical history, this signature waveform occurs.By corresponding for the priority after each signature waveform and weighting thereof output display on display, for doctor's reference.And doctor's signature waveform higher to the priority after weighting preferentially can be analyzed, consider, thus provide higher accuracy for Electrocardiographic analysis.For ease of computational analysis, the medical history weighted value of setting can be the value after normalization.

Such as, multiple neural network module 102 identifies 4 signature waveforms a, b, c, d obtaining sorting by priority.The priority that signature waveform a, b, c, d calculate according to error size is respectively: 0.4,0.3,0.25,0.05.Thus form priority matrix A: [0.4 0.3 0.25 0.05].

If according to health condition and the medical history of patient, the probability that b, d signature waveform occurs is 4 times of a, c, and so, arranging normalized medical history weighted value respectively for signature waveform a, b, c, d is: 0.1,0.4,0.1,0.4.According to the medical history weighted value arranged for each signature waveform, the priority of each signature waveform is computed weighted and is specifically as follows:

First build fuzzy matrix B, the corresponding each signature waveform respectively of each row in fuzzy matrix B, the value of all elements often in row is the medical history weighted value of this row characteristic of correspondence waveform.Such as, arranging normalized medical history weighted value respectively for above-mentioned signature waveform a, b, c, d is: 0.1,0.4,0.1,0.4, and constructed fuzzy matrix B is:

$\left[\begin{array}{cccc}0.1& 0.1& 0.1& 0.1\\ 0.4& 0.4& 0.4& 0.4\\ 0.1& 0.1& 0.1& 0.1\\ 0.4& 0.4& 0.4& 0.4\end{array}\right]$

The priority of fuzzy matrix B to each signature waveform is used to compute weighted, obtain the priority after the weighting of each signature waveform: C=B*A=[0.16 0.48 0.1 0.08], in Matrix C, namely each element is respectively the priority after the weighting of signature waveform a, b, c, d.As can be seen from Matrix C, the priority after signature waveform b weighting is the highest, then doctor can pay the utmost attention to according to the priority after its weighting, analyzes this signature waveform.

Preferably, in Characteristics of electrocardiogram waveform recognition system, multiple neural network module can be specifically two, as shown in Figure 3, one of them adopts the neural network module based on the signature waveform recognizer of incremental learning, is referred to as incremental learning neural network module herein; Another is the neural network module of the signature waveform recognizer adopted based on batch study, is referred to as batch learning neural network module herein.

Before incremental learning neural network module and batch learning neural network module are applied in Characteristics of electrocardiogram waveform recognition system and carry out signature waveform identification, first need to carry out learning training.

The training method of batch learning neural network module is:

Terminal (n=n from initial value (n=0) to learning period during a learning training _e) the iterative cycles of discrete series, n _einteger.To upgrade weight after batch learning training terminates between a learning period, and circulate with this.In following formula 1, be batch learning neural network module during a learning training according to the output valve that input iteration goes out; be expectation state amount and predetermined signature waveform, so error E can calculate according to following formula 1:

$E=\underset{n=0}{\overset{n_e}{\mathrm{\&Sigma;}}}\frac{1}{2}{(d_n-y_n)}^T(d_n-y_n)$ (formula 1)

T in above-mentioned formula 1 is matrix transpose operator.Lagrange's multiplier can calculate according to the orderly partial derivative of error E to output valve, namely calculates lagrangian multiplier according to following formula 2 _n:

${\mathrm{\&lambda;}}_n=\frac{{\&PartialD;}^{+}E}{{\&PartialD;y}_n},$ (formula 2)

${\mathrm{\&lambda;}}_n=\frac{\&PartialD;E}{{\&PartialD;y}_n}+\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_{n+j}\frac{{\&PartialD;y}_{n+j}}{{\&PartialD;y}_n},$ N=n _e-1 ..., 0, (formula 3)

Wherein L exports number, and its original state is determined according to following formula 4:

${\mathrm{\&lambda;}}_{n_e}=\frac{\&PartialD;E}{{\&PartialD;y}_{n_e}},{\mathrm{\&lambda;}}_{n_e+1},...,{\mathrm{\&lambda;}}_{n_e+L}=0$ (formula 4)

By Lagrange's multiplier, according to following formula 5 error of calculation to the orderly partial derivative of weight:

$\frac{{\&PartialD;}^{+}E}{\&PartialD;w\left(i\right)}=\underset{n=0}{\overset{n_e}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_n\frac{\&PartialD;y_n}{\&PartialD;w_n\left(i\right)}$ (formula 5)

Project output when being n-th iteration is to the partial derivative of weight.Mark i represents the number of times that weight vector has been upgraded; The initial value of i is 0.So far, the gradient descent method upgrading weight shown in following formula 6 can be utilized:

$w(i+1)=w\left(i\right)-\mathrm{\&eta;}\frac{{\&PartialD;}^{+}E}{\&PartialD;w\left(i\right)}$ (formula 6)

In formula 6, η is the learning coefficient of setting.

Batch learning neural network module by output valve and expectation state amount, calculates error E, Lagrange's multiplier, and then calculates error gradient during learning training, utilizes the weight that error gradient is upgraded in batch learning neural network module.After batch learning neural network module learning training terminates, will obtain weight matrix W, the element in weight matrix W is the weighted value w (i+1) that last iteration obtains.Batch learning neural network module, after being applied to Characteristics of electrocardiogram waveform recognition system, will being carried out signature waveform identification according to these weighted values, export the signature waveform identified.After how batch learning neural network module carries out computing according to the ECG signal of the weight matrix W determined in its signature waveform recognizer and above-mentioned training process to input, therefrom identify the technology that signature waveform is well known to those skilled in the art, repeat no more herein.

In the training method of incremental learning neural network module, the upgrading of the weight of current iteration n' only depend on system iteration 0 ... n '-1, the state of n'}, and weight all can change, so Lagrange's multiplier can not by accurate Calculation in iterative process each time; Again because each iteration needs upgrading, so can not accurate Calculation error gradient, thus also just cannot weighted value in accurate Calculation incremental learning neural network module.Therefore, consider that weight all can change in iterative process each time, the incremental learning neural network module of the embodiment of the present invention adopts approximation method to calculate approximate Lagrange's multiplier in learning training process, the approximation upgrading weight of application error partial derivative, thus the weighted value that can comparatively fast obtain in incremental learning neural network module, and the weighted value obtained is also comparatively accurate.The learning training method of the incremental learning neural network module of the embodiment of the present invention is as follows:

The learning training of the incremental learning neural network module of the embodiment of the present invention is based on simulation learning time period concept, and by supposing that primary iteration is n '-T during a simulation learning, final iteration is that current iteration n' realizes.Wherein, n ' is natural number, and T is a positive integer constant, and T+1 is hypothesis iteration siding-to-siding block length.If n ' is-T<0, so primary iteration is 0.

When forward direction iteration n', mean square error E _n'calculate according to following formula 11:

$E_{n^{\&prime;}}=\frac{1}{2}{(d_{n^{\&prime;}}-y_{n^{\&prime;}})}^T(d_{n^{\&prime;}}-y_{n^{\&prime;}})$ (formula 11)

D in formula 11 _n'mean Vector, y _n'it is output vector.

For real-time learning application, because weight all changes in iteration each time, so Lagrange's multiplier by accurate Calculation, therefore, can not can be calculated by the rear Iterative feedback method of following formula 12 and be similar to Lagrange's multiplier

${\widehat{\mathrm{\&lambda;}}}_n=\frac{{\&PartialD;}^{+}{E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_n}=\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_n}+\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{\widehat{\mathrm{\&lambda;}}}_{n+j}\frac{\&PartialD;{\hat{y}}_{n+j}}{\&PartialD;{\hat{y}}_n},$ N=n'-1 ..., n'-T, (formula 12)

In formula 12, L is lead-out terminal number, approximate output.

Following formula 13-15 is used to set up boundary condition:

${\widehat{\mathrm{\&lambda;}}}_{n^{\&prime;}}=-(d_{n^{\&prime;}}-y_{n^{\&prime;}})$ (formula 13)

${\widehat{\mathrm{\&lambda;}}}_{n^{\&prime;}+1},...,{\widehat{\mathrm{\&lambda;}}}_{n^{\&prime;}+L}=0$ (formula 14)

$\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_{n^{\&prime;}-T}},\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_{n^{\&prime;}-T+1}}...,\frac{{\&PartialD;E}_{n^{\&prime;}}}{\&PartialD;{\hat{y}}_{n^{\&prime;}-1}}=0$ (formula 15)

Because each iteration all will be upgraded, so can not accurate Calculation error gradient.Alternative method is the approximation upgrading weight of application error partial derivative.With error of calculation partial derivative during the simulation learning of length T+1 by the approximation of a generation true partial derivative.So, weight can be upgraded with following error of approximation gradient formula 16, namely obtains error of approximation gradient according to formula 16:

$\frac{\widehat{{\&PartialD;}^{+}{E}_{n^{\&prime;}}}}{\&PartialD;w\left(n^{\&prime;}\right)}=\underset{n=n^{\&prime;}-T}{\overset{n^{\&prime;}}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}^{n^{\&prime;}-n}{\widehat{\mathrm{\&lambda;}}}_n\frac{\&PartialD;{\hat{y}}_n}{\&PartialD;w\left(n\right)},$ (formula 16)

In formula 16,0< α <1 is the exponential form weight coefficient of setting; represent error of approximation gradient.

So far, incremental learning pattern can be obtained according to following formula 17 and upgrade in the weight of iteration n', namely in formula 17 can according to n-th ' secondary iteration time the weighted value w (n'+1) of weighted value when calculating the n-th '+1 iteration:

$w(n^{\&prime;}+1)=w\left(n^{\&prime;}\right)-\mathrm{\&eta;}\frac{\widehat{{\&PartialD;}^{+}{E}_{n^{\&prime;}}}}{\&PartialD;w\left(n^{\&prime;}\right)}$ (formula 17)

In formula 17, η is the adjustment positive number of setting.

In above-mentioned formula 12 determine according to following formula 18,19,20:

$W_0=0.7\&times;s^{\frac{1}{r}}\&times;\mathrm{normr}(2\mathrm{rand}(s,r)-I(s,r))$ (formula 18)

In above-mentioned formula 18, W ₀represent the original state of weight matrix W, s is that network implies the number of plies, and r is the number of the input amendment of neural network module.Network implies the number that the number of plies is hidden layer in the multiple structure of neural network module, and network optimum structure (comprising the number of plies) obtains according to structural optimization methods such as genetic algorithms.Network implies that the computational methods of the number of plies are well known to those skilled in the art, and repeats no more herein.The equally distributed random number matrix that rand (s, r) arranges for the capable r of s, I (s, r) is all 1's matrix of s capable r row, norm r() be the standardization normalizing matrix of matrix;

Calculate hidden node according to following formula 19 to export:

$o_j=f(\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}\&times;x_i-q_j)$ (formula 19)

In above-mentioned formula 19, w _ijbe the weighted value connecting i-th node and a jth node in the neutral net of neural network module, i, j are the integer being less than or equal to the total node number of neutral net.Namely in an iterative process, and as n-th ' in secondary iterative process, w _ijit is the weighted value connecting i-th node and a jth node in the neutral net in weight matrix W.Q _jfor a jth neural unit threshold value of setting; F is predetermined non-linear action function;

The approximate output of a kth output node when n-th iteration for:

${\hat{y}}_{\mathrm{nk}}=f(\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{jk}}\&times;o_j-q_k)$ (formula 20)

In formula 20, k is the positive integer being less than or equal to output layer interstitial content; During n-th iteration, each output node the approximate output vector of composition

Incremental learning neural network module is by obtaining the weight matrix W of incremental learning neural network module after above-mentioned learning training method, the element in weight matrix W is w (n'+1) weighted value that last iteration obtains.Incremental learning neural network module, after being applied to Characteristics of electrocardiogram waveform recognition system, carries out signature waveform identification by according to the weighted value in these weight matrix W, exports the signature waveform identified.Namely incremental learning neural network module is according to the weight matrix W determined in its signature waveform recognizer and above-mentioned training process, carries out computing, identify signature waveform and export from the signal calculated the ECG signal of input.How incremental learning neural network module carries out computing according to its signature waveform recognizer and weight matrix W thereof to the ECG signal inputted, and identifies the technology that signature waveform is well known to those skilled in the art, repeat no more herein from the signal calculated.

The embodiment of the present invention carries out signature waveform identification due to the neutral net that have employed multiple different characteristic waveform recognition algorithm, extends the signature waveform that single neutral net can identify.That is, Characteristics of electrocardiogram waveform recognition system can be different algorithm, method, angle carry out recognition feature waveform, reach the object realizing analyzing more accurately classification to ECG signal, thus Electrocardiographic signature waveform can be provided as much as possible for doctor, judge patient on this basis.

Further, according to the difference between each signature waveform and described ECG signal in the embodiment of the present invention, determine that the priority of each signature waveform knows the degree of the reference reliability of each signature waveform for doctor, for Electrocardiographic analysis provides higher accuracy.

Further, the fuzzy logic pattern recognition module of the embodiment of the present invention, by arranging medical history weighted value for each signature waveform, shows the probability that disease corresponding to each signature waveform or associated situation occur in medical history; Thus after using the priority of medical history weighted value to each signature waveform arranged to compute weighted, obtain the priority after the weighting of each signature waveform, further can reflect the probability that this signature waveform in medical history occurs, thus provide higher accuracy for Electrocardiographic analysis.Supply doctor's reference by corresponding for the priority after each signature waveform and weighting thereof output display to display, the signature waveform that doctor can be higher to the priority after weighting preferentially can be analyzed, consider, more contributes to doctor to Electrocardiographic analysis.

One of ordinary skill in the art will appreciate that all or part of step realized in above-described embodiment method is that the hardware that can carry out instruction relevant by program has come, this program can be stored in computer read/write memory medium, as: ROM/RAM, magnetic disc, CD etc.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (6)

1. a Characteristics of electrocardiogram waveform recognition system, comprising:

At least two neural network modules, the signature waveform recognizer of each neural network module is all not identical with the signature waveform recognizer of other arbitrary neural network module, described neural network module is used for carrying out computing to the ECG signal of input according to its signature waveform recognizer, identifies signature waveform and export from the signal calculated;

Fuzzy logic pattern recognition module, for receiving the signature waveform that each neural network module exports, according to the difference between each signature waveform and described ECG signal, calculates the priority of each signature waveform respectively; And according to the medical history weighted value arranged for each signature waveform, after the priority of each signature waveform is computed weighted, signature waveform is carried out corresponding output with the priority after its weighting.

2. the system as claimed in claim 1, is characterized in that, described neural network module is specially two; Signature waveform recognizer in two neural network modules is respectively: based on the signature waveform recognizer of incremental learning, the signature waveform recognizer based on batch study.

3. system as claimed in claim 2, it is characterized in that, described medical history weighted value is arranged according to medical history situation.

4. system as claimed in claim 3, is characterized in that,

Described fuzzy logic pattern recognition module, according to the medical history weighted value arranged for each signature waveform, computes weighted to the priority of each signature waveform and is specially:

Described fuzzy logic pattern recognition module is according to the medical history weighted value arranged for each signature waveform, determine fuzzy matrix: the corresponding each signature waveform respectively of each row in described fuzzy matrix, the value of all elements often in row is the medical history weighted value of this row characteristic of correspondence waveform; The priority of described fuzzy matrix to each signature waveform is used to compute weighted.

5. a Characteristics of electrocardiogram method for waveform identification, comprising:

Neural network module carries out computing according to the ECG signal of its signature waveform recognizer to input, identifies signature waveform and export from the signal calculated; Wherein, described neural network module is at least two, and the signature waveform recognizer of each neural network module is all not identical with the signature waveform recognizer of other arbitrary neural network module;

Fuzzy logic pattern recognition module receives the signature waveform that each neural network module exports, and according to the difference between each signature waveform and described ECG signal, calculates the priority of each signature waveform respectively; And according to the medical history weighted value arranged for each signature waveform, after the priority of each signature waveform is computed weighted, signature waveform is carried out corresponding output with the priority after its weighting.

6. method as claimed in claim 5, it is characterized in that, described neural network module is specially two; Signature waveform recognizer in two neural network modules is respectively: based on the signature waveform recognizer of incremental learning, the signature waveform recognizer based on batch study.