CN103646188A - Non-invasive diagnostic method of coronary heart disease based on hybrid intelligent algorithm - Google Patents

Abstract

The invention discloses a non-invasive diagnostic method of a coronary heart disease based on a hybrid intelligent algorithm. By adopting the method, a correct non-invasive diagnostic simulation model of the coronary heart disease is built by the following steps: encoding a chromosome; constructing a fitness function; establishing a selection operator, a cross operator and a mutation operator; initializing a BP network weight and a threshold; selecting an input sample; calculating an output value of a hidden layer; calculating a response value of each unit of an output layer; calculating a mean square error sum of each unit of the output layer; calculating the connection weight between the hidden layer and the output layer; calculating the connection weight of the input layer and the hidden layer. By adopting the method, the weakness that a usual method is sensitive to an initial value and easily falls into a local extreme value is overcome on the basis of the hybrid intelligent optimization algorithm; the convergence speed is greatly improved, the dimensionality of the impact factor is reduced, the calculation complexity of the algorithm is reduced, the structure and the parameter of the non-invasive diagnostic method can be regulated according to the concrete training process, and intelligence of the non-invasive diagnosis of the coronary heart disease is achieved.

Description

A kind of coronary heart disease Noninvasive diagnosis method based on hybrid intelligent algorithm

Technical field

The present invention relates to a kind of coronary heart disease Noninvasive diagnosis method based on hybrid intelligent algorithm, particularly a kind of intelligent computation and coronary heart disease Noninvasive diagnosis method, belong to computer utility and medical diagnosis field.

Background technology

Coronary cardiopathy is called for short coronary heart disease.Finger is because lipid-metabolism is undesired, and the lipid calmness in blood, on the endarterium of otherwise smooth, is piled up and formed white patch at the lipid material of some similar congee samples of endarterium, is called atherosclerotic lesion.These patches gradually increase and cause lumen of artery narrow, and blood flow is obstructed, and cause heart ischemia, produce angina pectoris.It is very big and complication is many to harm, becomes society human health one large ' killer '.

Along with the fast development of modern science and technology and Medical Research Work person are to the further investigation of coronary heart disease with probe into, the method for diagnosis of coronary heart disease becomes better and approaching perfection day by day.People come diagnosis of coronary heart disease miocardial infarction and coronary insufficiency according to typical clinical manifestation (comprising sings and symptoms), Myocardial Enzymologic procuratorial work and Characteristics of electrocardiogram the earliest.In recent years, many new procuratorial work method and technology have been developed, as radioactive nuclide procuratorial work, echocardiogram, coronarography, cardiac blood pool imaging etc. are applied to the diagnosis of coronary heart disease.

Coronarography technology is at present uniquely can directly observe the diagnostic method of coronary artery form, and medical circle number is called " goldstandard ".But because it is a kind of traumatic diagnostic techniques that has, diagnosis price is relatively high, high to medical condition requirement, it is even dead to there is severe complication in careless manipulation, and this has just limited extensively carrying out of this diagnostic techniques.Therefore, work out a kind of effectively, noninvasive diagnosis method becomes the most important thing of lot of domestic and foreign scholar's research easily.

At present, the Noninvasive diagnosis method of coronary heart disease, although there is some intellectual technology, also makes some progress, and also exists in actual applications many such as the more difficult problems such as selection, speed of convergence is slow, accuracy rate is low that are difficult to of factor of influence.In order better coronary heart disease to be carried out to Noninvasive diagnosis, BP neural network algorithm is improved, use LM algorithm to replace traditional gradient descent algorithm and carry out iteration, and merge mutually with genetic algorithm, hybrid intelligent algorithm is proposed.

Summary of the invention

The object of this invention is to provide a kind of coronary heart disease Noninvasive diagnosis method based on hybrid intelligent algorithm, this method, based on mixing intelligent optimizing algorithm, overcomes usual method and initial value sensitivity is easily absorbed in to the weakness of local extremum; Greatly improved speed of convergence; Reduce factor of influence dimension, reduced the computational complexity of algorithm; The structure and parameter of Noninvasive diagnosis method can be adjusted according to concrete training process, has realized the intellectuality of coronary heart disease Noninvasive diagnosis.

The present invention is achieved by the following technical solutions:

(1) in this invention, the actual coronary heart disease data analysis that U.S. Cleveland Clinical Basis institute is announced is summed up.From basic data, influential more than 50 parameters of coronary heart disease are carried out to cluster analysis and principal component analysis (PCA), find out the coronary heart disease conclusive major parameter of having classified.And coronary heart disease is carried out to classification.On pathology, judge that severity of coronary artery disease, often according to the diameter of transversal section, the most serious position of coronary artery stenosis, because coronary artery stenosis causes the narrow area of vessel lumen 5% and following, without coronary heart disease, is slight coronary heart disease 6%～25%; Between 26%～50%, for moderate coronary heart disease, between 51%～75%, for severe coronary heart disease, between 76%～100%, be overweight degree coronary heart disease.Five grades of diagnosis of coronary heart disease, without coronary heart disease, slight coronary heart disease, moderate coronary heart disease, severe coronary heart disease and overweight degree coronary heart disease.

(2) BP algorithm is combined with genetic algorithm, propose hybrid intelligent algorithm.First by genetic algorithm, network is trained, find a more excellent solution, then using this result, the network initial parameter in BP algorithm is trained again.This method can improve the classification capacity of network, avoids result to be absorbed in local optimum.Traditional BP algorithm, when treatment classification problem, is used gradient descent algorithm to carry out iteration, and the slow several thousand step iteration that conventionally need of speed of convergence are even more.Hybrid intelligent algorithm improves traditional BP algorithm, adopt L-M algorithm to substitute gradient descent algorithm and carry out iteration, due to the approximate second derivative information of L-M algorithm utilization, more faster than gradient method, especially when input dimension is lower, L-M optimized algorithm shows higher performance, so can significantly improve network convergence speed.

(3) using 22 main affecting factors as network input parameter, data have been carried out to pre-service, comprised and remove exceptional value and normalization.In the actual coronary heart disease data of announcing in U.S. Cleveland Clinical Basis institute, there is exceptional value in partial data item.In 297 groups of data altogether, after treatment, obtain 282 groups of available data.Then it is normalized.To eliminate the impact of dimension on it.Choose 240 groups as training set, 42 groups as test set.Determine each node layer number of network, carry out simulation training, thereby set up the realistic model of coronary heart disease Noninvasive diagnosis.

Beneficial effect of the present invention:

The present invention is based on mixing intelligent optimizing algorithm, overcome usual method and initial value sensitivity is easily absorbed in to the weakness of local extremum; Greatly improved speed of convergence; Reduce factor of influence dimension, reduced the computational complexity of algorithm; The structure and parameter of Noninvasive diagnosis method can be adjusted according to concrete training process, has realized the intellectuality of coronary heart disease Noninvasive diagnosis.

Accompanying drawing explanation

Fig. 1 is hybrid intelligent algorithm flow schematic diagram of the present invention.

Fig. 2 is hybrid intelligent Algorithm for Training curve map of the present invention.

Embodiment

Concrete steps of the present invention are as follows: as depicted in figs. 1 and 2;

Step (1): chromosomal coding

In chromosomal cataloged procedure, if consider that the intelligent diagnostics problem of coronary heart disease adopts binary coding, can cause coded strings long, can have influence on the study precision of network and the working time of algorithm, so the present invention adopts real coding.

Step (2): the structure of fitness function

Can choosing of fitness function be most important, directly have influence on the speed of convergence of genetic algorithm and find optimum solution.Objective function of the present invention is chosen network error function

wherein, m is output layer neuron number, and P is total sample number, c _kfor reality output, y _kfor target output.As objective function, objective function will be got minimum.Because target function value is contrary with fitness function, fitness function should be got the inverse of objective function, and fitness function is F (E)=1/E.

Step (3): select the establishment of operator

Selection strategy is used ratio conventional in genetic algorithm to select operator, and establishing population size is M, and the fitness of individual i is F _i, the selected probability P of individual i _ifor

Step (4): the improvement design of crossover operator

What adopt due to the present invention is real coding, so crossover operator adopts arithmetic Crossover Strategy, if through selection after operation, to the V choosing individuality with Probability p _ccarry out interlace operation, suppose that V is even number, in (0,1), evenly generate independently random number the V a choosing individuality is done to following intersection:

$\left\{\begin{array}{c}{x}_{2i}^{t+1}=x_{2i}^t+{\mathrm{\λ}}_i^1(x_{2i+1}^t-x_{2i}^t)\\ x_{2i+1}^{t+1}=x_{2i+1}^t+{\mathrm{\λ}}_i^2(x_{2i+1}^t-x_{2i}^t)\end{array},,\right.$ Wherein

with

that before intersecting, parent individuality is right,

with it is offspring individual pair after intersecting.

Step (5): the design of mutation operator

Mutation operator adopts Gaussian mutation strategy, supposes to have an individual X=x of being ₁x ₂... x _k... x _iif, x _kfor change point, its span is

at this point, individual X is carried out after Gaussian mutation operation, can obtain a new individual X=x ₁x ₂... x' _k... x _iwherein the new genic value of change point is x' _k=x _k+ N (0,1), wherein, N (0,1) average is 0, the random number that meets Gaussian distribution that variance is 1.

Step (6): BP network weight and threshold value initialization.

After training finishes through genetic algorithm, finding out the individuality of adaptive value maximum, each component of this individuality is decoded into corresponding parameter value, is exactly initial weight and the threshold value of network.

Step (7): input choosing of sample.

Choose at random one group of input sample $X_k=(x_1^k,x_2^k,...,x_n^k)$ And target sample $Y_k=(y_1^k,y_2^k,...,y_m^k),$ They are offered to network.

Step (8): calculate hidden layer output valve

By

j=1,2 ..., q calculates the input value of each unit of hidden layer, wherein w _ijfor input layer is to the connection weights between hidden layer, θ _jfor the threshold value of hidden layer, q is hidden neuron number, then passes through B _j=f ₁(S _j) obtain the output valve B of hidden layer _j, wherein, f ₁(x) transport function.

Step (9): the response of calculating each unit of output layer

By

t=1,2 ..., m calculates the input value of each unit of output layer, then by transport function C _t=f ₂(L _t), the response C of calculating each unit of output layer _t.Wherein, f ₂(x) be transport function.

Step (10): calculate each unit of output layer square error and

By

calculate each unit of output layer square error and.

Choose other samples, repeating step (8)～(10), after all samples are all inputted once, global error is

$E=\frac{1}{2}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{(c_k^p-y_k^p)}^2=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}{E}_p$

Step (11): calculate the connection weight between hidden layer and output layer.

Due to the approximate second derivative information of L-M algorithm utilization, more faster than gradient method, especially, when input dimension is lower, L-M optimized algorithm shows higher performance, so in the present invention, uses L-M algorithm to replace original gradient descent algorithm.For making global error E minimalization, by

$\left\{\begin{array}{c}\mathrm{\Δ}{v}_{\mathrm{jt}}=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\η}(c_k^p-y_k^p)\·f_2^{\′}\left(L_j\right)\·B_j\\ v_{\mathrm{jt}}^{\′}=v_{\mathrm{jt}}+\mathrm{\Δ}{v}_{\mathrm{jt}}\end{array}\right.$

Adjust the connection weight v' of hidden layer and output layer _jt, wherein η is learning rate, Δ v _jtfor connection weight regulation.With L-M algorithm, further optimizing, new weights regulation is Δ v=(J ^tj+ μ J) ^-1.J ^te, wherein: e is error vector; J is that network error is to weights v' _jtthe Jacobi matrix of derivative; μ is scalar, and when μ is very large, Δ v is close to gradient method, and when μ is very little, Δ v has become Gauss-Newton method, and therefore, μ is also self-adaptation adjustment.Thereby again by v _jt=v' _jt+ Δ v obtains the connection weight v of hidden layer and output layer _jt.

Step (12): calculate input layer and hidden layer connection weight.

According to the connection weight v of output layer _jt, by

$\left\{\begin{array}{c}\mathrm{\Δ}{w}_{\mathrm{ij}}=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\η}{v}_{\mathrm{jt}}{x}_i(c_k^p-y_k^p)f_2^{\′}\left(L_j\right)f_1^{\′}\left(S_k\right)\\ w_{\mathrm{ij}}^{\′}=\mathrm{\Δ}{w}_{\mathrm{ij}}+w_{\mathrm{ij}}\end{array}\right.$

Adjust input layer and hidden layer connection weight w ' _ij, more further optimize with L-M algorithm, by Δ w=(J ^tj+ μ J) ^-1.J ^te and wi _j=w ' _ij+ Δ w calculates input layer and hidden layer connection weight w _ij.

Above-mentioned steps is the process of a hybrid intelligent Algorithm Learning generation, more again from m learning sample, choose one group input and target sample learn, until global error E is less than predefined permissible error, claim that this result is network convergence.If study number of times is greater than predefined study number of times, network cannot be restrained.

Claims (3)

1. the coronary heart disease Noninvasive diagnosis method based on hybrid intelligent algorithm, the method comprises the following steps:

Step (1): chromosomal coding

In chromosomal cataloged procedure, if consider that the intelligent diagnostics problem of coronary heart disease adopts binary coding, can cause coded strings long, can have influence on the study precision of network and the working time of algorithm, so literary grace real coding, i.e. x=(w, θ, v, γ);

Step (2): the structure of fitness function

Can choosing of fitness function be most important, directly have influence on the speed of convergence of genetic algorithm and find optimum solution, and objective function is chosen network error function

c _kfor reality output, y _kfor target output; As objective function, objective function will be got minimumly, and because target function value is contrary with fitness function, fitness function should be got the inverse of objective function, and fitness function is F (E)=1/E;

Step (3): select the establishment of operator

Selection strategy is used ratio conventional in genetic algorithm to select operator, and establishing population size is M, and the fitness of individual i is F _i, the selected probability of individual i is P _ifor

Step (4): the design of crossover operator

Due to what adopt, be real coding, so crossover operator adopts arithmetic Crossover Strategy, be provided with individuality

the two carries out arithmetic intersection, and two new individualities that produce after crossing operation are $\left\{\begin{array}{c}{x}_A^{i+1}={\mathrm{\αx}}_B^i+(1-\mathrm{\α})x_A^i\\ x_B^{i+1}=\mathrm{\α}{x}_A^i+(1-\mathrm{\α})x_B^i\end{array},,\right.$ Wherein, α is the equally distributed random number that meets between (0,1);

Step (5): the design of mutation operator

Mutation operator adopts even Mutation Strategy, supposes to have an individual X=x of being ₁x ₂... x _k... x _iif, x _kfor change point, its span is

at this point, individual X is carried out after even mutation operation, can obtain a new individual X=x ₁x ₂... x _'k...x _iwherein the new genic value of change point is

wherein, r meets equally distributed random number between (0,1);

Step (6): BP network weight and threshold value initialization;

After training finishes through genetic algorithm, finding out the individuality of adaptive value maximum, each component of this individuality is decoded into corresponding parameter value, is exactly initial weight and the threshold value of network;

Step (7): input choosing of sample;

Choose at random one group of input sample $X_k=(x_1^k,x_2^k,...,x_n^k)$ And target sample $Y_k=(y_1^k,y_2^k,...,y_m^k),$ They are offered to network;

Step (8): calculate hidden layer output valve

By

j=1,2 ..., p calculates the input value of each unit of hidden layer, wherein w _ijfor input layer is to the connection weights between hidden layer, θ _jfor the threshold value of hidden layer, then pass through B _j=f ₁(s _j) obtain the output valve B of hidden layer _j, j=1,2 ..., p, wherein, f ₁(x) be transport function;

Step (9): the response of calculating each unit of output layer

By

t=1,2 ..., q calculates the input value of each unit of output layer, then by transport function C _t=f ₂(L _t), the response C of calculating each unit of output layer _t, t=1,2 ..., q; Wherein, f ₂(x) be transport function;

Step (10): calculate each unit of output layer square error and

By

calculate each unit of output layer square error and;

Choose other samples, repeating step (8)～(10), after all samples are all inputted once, global error is

$E=\frac{1}{2}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{(c_k^p-y_k^p)}^2=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}{E}_p$

Step (11): calculate the connection weight between hidden layer and output layer;

Due to the approximate second derivative information of L-M algorithm utilization, more faster than gradient method, especially, when input dimension is lower, L-M optimized algorithm shows higher performance, uses L-M algorithm to replace original gradient descent algorithm; For making global error E minimalization, by

$\left\{\begin{array}{c}\mathrm{\Δ}{v}_{\mathrm{jt}}=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\η}(c_k^p-y_k^p)\·f_2^{\′}\left(L_j\right)\·B_j\\ v_{\mathrm{jt}}^{\′}=v_{\mathrm{jt}}+\mathrm{\Δ}{v}_{\mathrm{jt}}\end{array}\right.$

Adjust the connection weight v' of hidden layer and output layer _jt, wherein η is learning rate, Δ v _jtfor connection weight regulation; With L-M algorithm, further optimizing, new weights regulation is Δ v=(J ^tj+ μ J) ^-1.J ^te, wherein: e is error vector; J is that network error is to weights v' _jtthe Jacobi matrix of derivative; μ is scalar, and when μ is very large, Δ v is close to gradient method, and when μ is very little, Δ v has become Gauss-Newton method, and therefore, μ is also self-adaptation adjustment; Thereby again by v _jt=v' _jt+ Δ v obtains the connection weight v of hidden layer and output layer _jt;

Step (12): calculate input layer and hidden layer connection weight;

According to the connection weight v of output layer _jt, by

$\left\{\begin{array}{c}\mathrm{\Δ}{w}_{\mathrm{ij}}=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\η}{v}_{\mathrm{jt}}{x}_i(c_k^p-y_k^p)f_2^{\′}\left(L_j\right)f_1^{\′}\left(S_k\right)\\ w_{\mathrm{ij}}^{\′}=\mathrm{\Δ}{w}_{\mathrm{ij}}+w_{\mathrm{ij}}\end{array}\right.$

Adjust input layer and hidden layer connection weight w ' _ij, more further optimize with L-M algorithm, by Δ w=(J ^tj+ μ J) ^-1.J ^te and w _ij=w ' _ij+ Δ w calculates input layer and hidden layer connection weight w _ij.

2. a kind of coronary heart disease Noninvasive diagnosis method based on hybrid intelligent algorithm according to claim 1, is characterized in that:

(1) use cluster analysis and principal component analysis (PCA) to filter out the main affecting factors of coronary heart disease Noninvasive diagnosis;

(2) use hybrid intelligent algorithm to carry out non-invasive intelligent diagnostics to coronary heart disease.

3. a kind of coronary heart disease Noninvasive diagnosis method based on hybrid intelligent algorithm according to claim 1, is characterized in that: by genetic algorithm, come initialization weights and threshold value, avoid network to be absorbed in locally optimal solution; Use improvement BP neural network to calculate and improved network convergence speed.

Images