CN104523266A - Automatic classification method for electrocardiogram signals - Google Patents

Abstract

The invention discloses an automatic classification method for electrocardiogram signals. The method is achieved according to the following steps of firstly, obtaining electrocardiogram signals of a human body, conducting filtering on the electrocardiogram signals, and detecting R waves of the electrocardiogram signals where filtering is conducted; secondly, establishing a data set after the R waves are detected, wherein the data set is composed of multiple sets of cardiac beat data, and each set of cardiac beat data has a label; thirdly, establishing a sparse automatic coding deep learning network; fourthly, training the sparse automatic coding deep learning network step by step; fifthly, inputting the to-be-measured cardiac beat data into the sparse automatic coding deep learning network according to the network weight, obtained in the fourth step, of the first hidden layer, the network weight, obtained in the fourth step, of the second hidden layer and the network weight, obtained in the fourth step, of the softmax classifier so as to obtain cardiac data which are output in a classified mode. The sparse automatic coding deep learning network is applied to the classification of the cardiac beat data, and by means of the autonomous leaning capacity and the deep characteristic excavation characteristic of the sparse automatic coding deep learning network, deeper characteristics of signals are extracted, and the cardiac beat data are classified.

Description

A kind of electrocardiosignal automatic classification method

Technical field

The present invention relates to the automatic examination and analysb technology of electrocardiosignal, particularly a kind of electrocardiosignal automatic classification method.

Background technology

Heart disease has disguise and latency, is difficult to show on electrocardiogram when not falling ill, and is again of short duration during morbidity, has little time to observe electrocardiogram.Need to carry 24 hours Holter to patient for this reason, carry out 24 hours ecg signal acquirings, then give doctor electrocardiogram (ECG) data, by doctor to data analysis.The data volume now produced is huge, and doctor needs the plenty of time to find the bat of the abnormal heart.Although the software system energy automatic analysis heart that Holter carries is clapped, and provides statistical information, because human body differs greatly, electrocardio change is complicated, and some heart is clapped still needs doctor's artificial cognition to correct.In so a large amount of data, find the heart of error flag to clap also is a very hard work.In order to greatly save doctor's time, improve diagnosis efficiency, stable Algorithms for Automatic Classification is very necessary.

Summary of the invention

The object of this invention is to provide a kind of electrocardiosignal automatic classification method, to solve existing sorting algorithm to different human body, the instability problem of electrocardiosignal classification under varying environment.

The object of the present invention is achieved like this: electrocardiosignal automatic classification method provided by the present invention, comprises the following steps:

A) obtain the electrocardiosignal of human body, and carry out Filtering Processing, the R ripple of the electrocardiosignal after detection filter;

B) after R ripple being detected, build data set, described data set is made up of some groups of heart beat of data, often organize described heart beat of data all with a kind of label, described label always has 6 kinds, is divided into the bat of the normal heart, left bundle branch block, right bundle branch block, ventricular premature contraction, artrial premature beat, amalgamation heart beating:

Often organize described heart beat of data and comprise 270 sampled points, these 270 sampled points are the positions of R ripple according to detecting, choose 90 sampled points before the wave crest point of described R ripple, described R ripple wave crest point choose 179 sampled points below;

C) sparse automatic encoding degree of deep learning network is built:

Described sparse automatic encoding degree of deep learning network has two hidden layers, after connect softmax grader, wherein, described two hidden layers are respectively the first hidden layer and the second hidden layer;

Described sparse automatic encoding degree of deep learning network be input as 270 sampled points, described first node in hidden layer is 130, and described second node in hidden layer is 50;

D) substep trains described sparse automatic encoding degree of deep learning network:

D-1) heart beat of data of described data set is normalized, then SAE model is inputted, adopt the first hidden layer of sparse automatic encoding degree of deep learning network described in SAE model training, obtain the network weight of the first hidden layer, and obtain the shallow-layer feature of heart beat of data;

D-2) same SAE model is adopted, described shallow-layer feature is inputted described SAE model, train the second hidden layer of described sparse automatic encoding degree of deep learning network, obtain the network weight of the second hidden layer, and obtain the high-level feature of described some groups of heart beat of data;

D-3) the high-level feature obtained is input to softmax grader, training softmax grader, obtains the network weight of softmax grader;

E) according to steps d) network weight of the network weight of the first hidden layer of gained, the network weight of the second hidden layer and softmax grader, to treat the described sparse automatic encoding degree of deep learning network of thought-read beat of data input, the classification obtaining heart beat of data exports.

Described step detailed process a) is as follows:

(1) signals collecting: gather human body electrocardio primary signal with the frequency acquisition of 250Hz, and be stored as the data mode of TXT document, then with Matlab software, the electrocardio original signal data that described TXT document stores is read in computer;

(2) Filtering Processing is carried out to described electrocardio original signal data:

(2-1) wavelet decomposition is carried out to described electrocardio primary signal: select DB6 small echo, 8 layers of decomposition are carried out to signal, obtains the wavelet coefficient d on each yardstick _i;

(2-2) adopt the calculated threshold method improved, ask for the threshold value of each yardstick, thresholding process is carried out to wavelet coefficient:

$T_i=\frac{{\mathrm{\σ}}_i\sqrt{2\ln \left(n\right)}}{e^{i-1}},i=\mathrm{1,2},...,8$

Wherein, T _ifor the threshold value improved, i represents the wavelet decomposition number of plies, and e is natural constant, and n represents sampling number, σ _iaverage for wavelet coefficient absolute value: ${\mathrm{\σ}}_i=\frac{\mathrm{median}\left|d_i\right|}{0.6745};$

(2-3) soft threshold method is adopted to carry out thresholding process to signal: to choose different threshold values at different scale and carry out thresholding process, obtain filtered electrocardiosignal;

(3) according to the difference of QRS wave group and P ripple, T wave frequency distribution, select QRS wave group and P ripple, T wave frequency overlapping minimum 3rd, 4 yardsticks that distribute to carry out wavelet reconstruction, obtain the electrocardiosignal S' after reconstruct;

(4) energy window conversion is carried out to the electrocardiosignal through wavelet reconstruction, and chooses maximum point:

(4-1) energy window conversion: by following formula, the electrocardiosignal S' through wavelet reconstruction is transformed to energy domain analysis by time domain analysis, obtains electrocardiosignal energy curve:

$E_n=\underset{n-\frac{N}{2}+1}{\overset{n+\frac{N}{2}}{\mathrm{\Σ}}}{S^{\′2}}_n,n=\frac{N}{2},\frac{N}{2}+1,...,M-\frac{N}{2}-1,M-\frac{N}{2}$

Wherein, E _nrepresent the energy value of the n-th sampled point; N is selected length of window, value 26; M is total sampling number; S' _nrepresent n-th data of the electrocardiosignal S' after described wavelet reconstruction;

(4-2) maximum point is chosen: obtained electrocardiosignal energy curve is carried out hard-threshold process, that is:

$Q_n=\left\{\begin{array}{c}{E}_n,E_n\&GreaterEqual;T_h\\ 0,E_n<T_h\end{array}\right\}$

Wherein, T _hfor selected threshold value, get T _h=0.3*median (E _n),

Then select the crest location of the electrocardiosignal energy curve after hard-threshold process as maximum point;

(5) maximum point is optimized: set 2 time threshold t ₁and t ₂, and t ₁<t ₂when the interval of any two maximum points is less than t ₁time, just remove less that of amplitude between these two maximum points; When any two maximum points interval greater than t ₂time, just between these two maximum points, find another unrecognized extreme point; When the interval of any two maximum points was both greater than t ₁, be less than t again ₂, then these two maximum points all retain, the maximum point through optimizing finally obtained, and the corresponding QRS wave group of maximum point through optimizing described in each;

(6) according to the time point at maximum point place each in step (5), in filtered electrocardiosignal described in step (2) about corresponding time point each 7 sampled points scope in the point of search signal amplitude maximum, as the R ripple detected.

Sparse automatic encoding degree of deep learning network is applied to the classification of heart beat of data by the present invention, makes full use of the characteristic of its independent learning ability and further feature excavation, extracts the deeper feature of signal, and then classify to heart beat of data.Specific sparse automatic encoding degree of deep learning network constructed by the present invention, the large data characteristic of electrocardiosignal can be made full use of, excavate the profound feature of electrocardiosignal, it is made to be provided with good stability for the ECG Signal Analysis of Different Individual under complex environment, by network structure reasonable in design and suitable network training method, achieve the automatic classification of the electrocardiosignal under complicated individuality and complex environment, solve the problem of electrocardiosignal sorting algorithm instability under reply individual variation and complexity bad border in prior art, accurately, the stable accurate identification achieving 6 class common arrhythmia beat.

In addition, the present invention carries out choosing specific wavelet basis function and the wavelet decomposition number of plies in wavelet decomposition process to signal, the threshold method adopting improvement in thresholding processing procedure is being carried out to signal simultaneously, the myoelectricity interference that filtered signal is mingled with in filtering electrocardiosignal, while baseline drift and Hz noise, as much as possiblely remain useful information, improve the phenomenon of generic threshold value excess smoothness.QRS wave group, by carrying out wavelet reconstruction, extracts, and P ripple, T ripple is used as noise eliminating, effectively prevent tall and big P ripple, flase drop that T ripple causes in the detection by the present invention, improves the precision of detection.Present invention employs energy window alternative approach, convert the signal into energy domain and go analyze thus solve in time-domain analysis, signal is subject to the impact of high-frequency noise, and can not by the problem of whole filtering in filtering.In energy window conversion, the present invention has taken into full account the temporal signatures of electrocardiosignal and the leap time of QRS wave group, carries out the selection that window is long.The results show is when only fenestrate length is 26, and the detection leakage phenomenon that the QRS ripple of many inspections that the spurious peaks of noise produces and low amplitude value causes could the most effectively be avoided.

Accompanying drawing explanation

Fig. 1 is flow chart of the present invention.

Fig. 2 is electrocardio primary signal.

Fig. 3 is for carrying out filtered electrocardiosignal.

Fig. 4 is the electrocardiosignal after being reconstructed.

Fig. 5 is the electrocardiosignal choosing maximum point after carrying out hard-threshold process.

Fig. 6 is the flow chart be optimized maximum point.

Fig. 7 is detected R ripple.

Fig. 8 is sparse automatic encoding degree of deep learning network structure chart.

Fig. 9 is degree of deep learning network training process figure.

Detailed description of the invention

Embodiment 1:

The present embodiment, at Intel Xeon CPU E5-2697 2.70GHz, inside saves as 128.00GB, Win7, realizes in the computer of 64 bit manipulation systems, and whole electrocardiosignal Algorithms for Automatic Classification adopts Matlab language to realize.

Implementation process of the present invention is as shown in Figure 1:

A) obtain human body electrocardio primary signal, and carry out Filtering Processing, the R ripple of the electrocardiosignal after detection filter, it specifically operates according to the following steps:

(1) electrocardio primary signal gathers: the present invention utilizes the MedSun18 of Beijing Peng Yangfeng industry Holter that leads to gather the electrocardiosignal of human body for a long time, and its sampling output frequency is 250Hz, gathers electrocardiogram (ECG) data and stores with the form of TXT.It can read easily in Matlab environment and show, and its form is as Fig. 2.

(2) Filtering Processing is carried out to gathered electrocardio original signal data:

(2-1) wavelet decomposition is carried out to electrocardio primary signal: the DB6 small echo selecting Daubechies small echo series, carry out 8 layers of decomposition, as shown in table 1:

Table 1: under the sample frequency of 250Hz, 8 layers of decomposition are carried out to DB6 small echo

Scale	250HZ
		1	62.5-125
2	31.25-62.5
		3	15.625-31.25
4	7.8125-15.625
		5	3.90625-7.8125
6	1.953125-3.90625
		7	0.9765625-1.953125
8	0.48828125-0.9765625

Then the wavelet coefficient d on each yardstick is extracted _i.

(2-2) adopt the method for the calculated threshold improved, ask for the threshold value of each yardstick, to obtain the threshold value through improving, that is:

$T_i=\frac{{\mathrm{\&sigma;}}_i\sqrt{2\ln \left(n\right)}}{e^{i-1}},i=\mathrm{1,2},...,8$ Formula I

In formula I, i represents the wavelet decomposition number of plies, T _ifor the threshold value improved, e is natural constant, and n represents sampling number, σ _ifor the average of wavelet coefficient absolute value, its expression formula is

Compare existing fixed threshold method and minimax threshold method, after improving by the algorithm of formula I to threshold value, the threshold value after improvement has sideband adaptivity, and maintains good denoising reconstruction property.

(2-3) adopt soft threshold method, each yardstick being chosen the corresponding threshold value through improving, by formula II, thresholding process being carried out to electrocardiosignal, that is:

${\hat{d}}_j=\left\{\begin{array}{c}\mathrm{sgn}\left(d_j\right)\left(\right|d_j|-T_j),\left|d_j\right|\&GreaterEqual;T_j\\ 0,\left|d_j\right|<T_j\end{array}\right.$ Formula II

Wherein j=i;

Thus obtain filtered electrocardiosignal, as shown in Figure 3.

Filtered electrocardiosignal, as much as possiblely remains useful information, improves the phenomenon of generic threshold value excess smoothness, makes filter effect more stable.

(3) the frequency distribution scope of normal electrocardiosignal QRS wave group is 5-45Hz, as can be seen from Table 1, it mainly concentrates on 3,4 yardsticks, and P ripple and T wave frequency distribution are 0.05 to 10Hz, 3,4 yardsticks do not have or only has a small amount of distribution, therefore, according to the difference of QRS wave group and P ripple, T wave frequency distribution, 3,4 yardsticks selecting QRS wave group and P ripple, T wavelength-division cloth frequency overlap minimum carry out wavelet reconstruction to through filtered electrocardiosignal, that is:

$S^{\&prime;}={\hat{d}}_3+{\hat{d}}_4$ Formula III

In formula III, with be respectively by step (2) the result of electrocardiosignal after thresholding process on 3,4 yardsticks.

After wavelet reconstruction, the electrocardiosignal obtained is mainly the information of QRS wave group, serves the effect highlighting QRS wave group, as shown in Figure 4.

(4) energy window conversion is carried out to the electrocardiosignal after wavelet reconstruction, and chooses maximum point:

(4-1) energy window conversion: by following formula IV, the electrocardiosignal S' through wavelet reconstruction is transformed to energy domain analysis by time domain analysis, obtains electrocardiosignal energy curve:

$E_n=\underset{n-\frac{N}{2}+1}{\overset{n+\frac{N}{2}}{\mathrm{\&Sigma;}}}{S^{\&prime;2}}_n,n=\frac{N}{2},\frac{N}{2}+1,...,M-\frac{N}{2}-1,M-\frac{N}{2}$ Formula IV

Wherein, E _nrepresent the energy value of the n-th sampled point; N is selected length of window (N=26), and M is total sampling number, S' _nrepresent n-th data of the electrocardiosignal S' after step (3) wavelet reconstruction.

In energy window conversion, choosing of window length is a key, and it directly determines that whether R ripple detection algorithm is effective.The present invention has taken into full account the temporal signatures of electrocardiosignal and the leap time of QRS wave group, the selection of its N value is determined as follows: the sample frequency of the present embodiment electrocardiosignal is 250Hz, and normal QRS wave group is generally no more than 0.1s, be 25 sampled points, we choose window long is even number, is 26.The results show is when only fenestrate length is 26, and the detection leakage phenomenon that the QRS ripple of many inspections that the spurious peaks of noise produces and low amplitude value causes could the most effectively be avoided.

In time-domain analysis, signal is subject to the impact of high-frequency noise, and can not, by whole filtering, for this problem, adopt the method for energy bed conversion that time domain analysis is transformed to energy domain analysis in filtering.Energy domain, compared to time domain analysis, has better robustness to noise.As Fig. 5, after energy window conversion, the QRS wave group location point of signal becomes more outstanding, the heart is clapped and the heart claps between interval more obviously, the impact of high-frequency noise also dies down accordingly.

(4-2) maximum point is chosen: the signal energy curve obtained is carried out hard-threshold process:

$Q_n=\left\{\begin{array}{c}{E}_n,E_n\&GreaterEqual;T_h\\ 0,E_n<T_h\end{array}\right\}$ Formula (V)

In formula (V), T _hfor selected threshold value, get T _h=0.3*median (E _n).

Then the crest location of the electrocardiosignal energy curve after hard-threshold process is chosen as maximum point, as shown in Figure 5.

(5) maximum point is optimized: the flow chart given by Fig. 6, sets 2 time threshold t ₁and t ₂, and t ₁<t ₂when the interval of any two maximum points is less than t ₁time, just remove less that of amplitude between these two maximum points; When any two maximum points interval greater than t ₂time, just between these two maximum points, find another unrecognized extreme point; Interval as two maximum points was both greater than t ₁, be less than t again ₂, then these two maximum points all retain, the corresponding QRS wave group of each maximum point through optimizing so finally obtained.

In Fig. 6, E _trepresent the meansigma methods of the interval of all maximum points that step (4-2) obtains, t ₁=0.5 × E _t, t ₂=1.5 × E _t.

(6) according to the time point at each maximum point place determined in step (5), in the step (2) after filtering in electrocardiosignal about corresponding time point each 7 sampled points scope in the point of search signal amplitude maximum, be the R ripple (Fig. 7) detected.

In actual application, other frequency can be selected as required to carry out the collection of electrocardiosignal, and select any one small echo in Haar, Daubechies or Symlets to carry out the decomposition of the suitable number of plies, and according to practical situation determine voluntarily follow-up carry out yardstick selected by wavelet reconstruction and energy window conversion time window long selection.Should be slightly different for the structure of different application degree of deep learning networks and corresponding node in hidden layer.Training method also will change flexibly, is not limited to fixed form.

B) after obtaining R ripple position, data set is built, this data set is made up of 33950 groups of heart beat of data, often organize heart beat of data all with a kind of label, label always has 6 kinds, represents the bat of the normal heart, left bundle branch block, right bundle branch block, ventricular premature contraction, artrial premature beat, amalgamation heart beating respectively:

Often organize heart beat of data and comprise 270 sampled points, these 270 sampled points are the positions of R ripple a) obtained according to step, choose 90 sampled points before this R wave crest point in electrocardiosignal figure after the filtering, after choose 179 sampled points, namely each group heart beat of data comprises 270 sampled points.

Data set in the present embodiment contains 33950 groups of heart beat of data from multiple people and same person different times, data centralization comprises various types of heart and claps, and by wherein 22350 groups of heart beat of data are as training dataset, remaining 11600 groups of heart beat of data are as test data set.Training dataset and test data concentrate the label all comprising six types.

C) sparse automatic encoding degree of deep learning network (abbreviation learning network) is built

As shown in Figure 8, it has two hidden layers (the first hidden layer and the second hidden layer) to this sparse automatic encoding degree of deep learning network structure, at the softmax of the connection below grader of the second hidden layer.

Learning network be input as 270 sampled points, first node in hidden layer is 130 (can obtain 130 shallow-layer features through the first hidden layer), and the second node in hidden layer is 50 (can obtain 50 high-level features through the second hidden layer).

D) substep trains sparse automatic encoding degree of deep learning network

D-1) 22350 of training dataset groups of heart beat of data are normalized:

$g_i=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }},i=\mathrm{1,2},...270$

Wherein x _minand x _maxbe respectively the minimum and maximum value of inputted heart amplitude of beat value.

By the heart beat of data input SAE model after normalized, adopt SAE model training learning network first hidden layer, its flow process as shown in Figure 9.

Concrete, SAE model be input as one group of heart beat of data (270 sampled point), hidden layer node number is 130.Select sigmoid function f (z)=1/ (1+exp (-z)) as neuronic activation primitive; First hidden layer node activation value function is h (g)=f (W ₁g+b ₁), wherein for weight matrix, for bias vector.The output of SAE is $\tilde{g}=f(W_{2}h\left(g\right)+b_2),$ for weight matrix, for bias vector.

In order to make output infinite approach g, introducing cost function, by minimizing cost function training network, obtaining network weight W ₁and W ₂, and obtain 130 shallow-layer features of heart beat of data, concrete:

When the number of training sample (training data is concentrated after normalized heart to clap and is a training sample) is q, the cost function of SAE is expressed as:

Wherein, be sparse penalty factor, β controls the weight of sparse penalty factor, be the average activity of this hidden layer jth neuron on q training sample, ρ is openness parameter, and we choose λ=3 × 10 ^-10, β=3, ρ=0.2;

Then we adopt L-BFGS optimization method to minimize cost function J _w,bg (), greatest iteration step number is set to 400, and we obtain 130 shallow-layer features of heart beat of data like this.

D-2) obtain 130 shallow-layer features are inputted same SAE model, adopt training study network second hidden layer that uses the same method, obtain second layer network weight, and obtain 50 high-level feature H of heart beat of data.

Concrete, the second node in hidden layer is 50, and same sigmoid function of selecting is as neuronic activation primitive, and the second hidden layer node activation value function is H (g)=f (W ₃h (g)+b ₃), wherein for weight matrix, for bias vector.The now output of SAE is $\tilde{h}=f(W_{4}H\left(g\right)+b_4),$ for weight matrix, for bias vector.In order to make output infinite approach h, introduces identical cost function, obtains weights W by minimizing cost function training network ₃and W ₄, and obtain 50 high-level features of heart beat of data, concrete:

When training sample number is q, the cost function of SAE is expressed as:

$J_{W,b}\left(g\right)=\frac{1}{2q}\underset{i=1}{\overset{q}{\mathrm{\&Sigma;}}}{\left|\right|h^{\left(i\right)}-{\tilde{h}}^{\left(i\right)}\left|\right|}_2^2+\frac{\mathrm{\&lambda;}}{2}({\left|\right|W_3\left|\right|}_2^2+{\left|\right|W_4\left|\right|}_2^2)+\mathrm{\&beta;}\underset{j=1}{\overset{50}{\mathrm{\&Sigma;}}}\mathrm{KL}\left(\mathrm{\&rho;}\right|\left|{\widetilde{\mathrm{\&rho;}}}_j\right)$

Wherein, be sparse penalty factor, β controls the weight of sparse penalty factor, be the average activity of this hidden layer jth neuron on q training sample, ρ is openness parameter.We choose λ=3 × 10 ^-10, β=3, ρ=0.2;

Then we adopt L-BFGS optimization method to minimize cost function J _w,bg (), greatest iteration step number is set to 400, so just obtains 50 high-level features of heart beat of data.

D-3) by d-2) in the further feature H of heart beat of data that extracts input Softmax grader, training softmax grader, obtains the network weight of softmax grader, specific as follows:

Softmax grader can be expressed as:

$r_{\mathrm{\&theta;}}\left(H^{\left(i\right)}\right)=\left[\begin{array}{c}p(y^{\left(i\right)}=1|H^{\left(i\right)};\mathrm{\&theta;})\\ p(y^{\left(i\right)}=2|H^{\left(i\right)};\mathrm{\&theta;})\\ .\\ .\\ .\\ p(y^{\left(i\right)}=6|H^{\left(i\right)};\mathrm{\&theta;})\end{array}\right]=\frac{1}{\underset{j=1}{\overset{6}{\mathrm{\&Sigma;}}}{e}^{{\mathrm{\&theta;}}_j^T{H}^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\&theta;}}_1^T{H}^{\left(i\right)}}\\ e^{{\mathrm{\&theta;}}_2^T{H}^{\left(i\right)}}\\ .\\ .\\ .\\ e^{{\mathrm{\&theta;}}_6^T{H}^{\left(i\right)}}\end{array}\right]$

Wherein r _θ(H ⁽ⁱ⁾) in each component p (y ⁽ⁱ⁾=j|H ⁽ⁱ⁾; θ) represent H ⁽ⁱ⁾belong to the probability of jth class, i=1,2 ... 50, j is 1,2 ... 6, θ is network weight matrix, $\mathrm{\&theta;}={\left[\begin{array}{cccc}{\mathrm{\&theta;}}_1^T& {\mathrm{\&theta;}}_2^T& ...& {\mathrm{\&theta;}}_6^T\end{array}\right]}^T,$

Selected cost function is defined as:

$J\left(\mathrm{\&theta;}\right)=-\frac{1}{q}\left[\underset{i=1}{\overset{q}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{6}{\mathrm{\&Sigma;}}}1\right\{y^{\left(i\right)}=j\left\}\ln \frac{e^{{\mathrm{\&theta;}}_j^T{h}^{\left(i\right)}}}{\underset{l=1}{\overset{6}{\mathrm{\&Sigma;}}}{e}^{{\mathrm{\&theta;}}_l^T{h}^{\left(i\right)}}}\right]+\frac{\mathrm{\&alpha;}}{2}\underset{i=1}{\overset{6}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{130}{\mathrm{\&Sigma;}}}{\mathrm{\&theta;}}_{\mathrm{ij}}^2$

Wherein 1{y ⁽ⁱ⁾=j} is indicator function, and the expression formula among brace is that very then indicator function value is 1, otherwise indicator function value is 0.In above formula, plus sige part is below the weights attenuation term in order to prevent model generation over-fitting from adding, and elects α=6 × 10 as ^-7.

We adopt L-BFGS optimization method to finely tune the network weight of softmax grader, cost function J (θ) is minimized by selecting L-BFGS optimization method, arranging greatest iteration step number is 400, and the heart that the label that the most probable value of Softmax regression Calculation is corresponding when algorithmic statement is the prediction of arrhythmia automatic identification algorithm claps classification.

E) verify: according to steps d) network weight of the network weight of the first hidden layer of gained, the network weight of the second hidden layer and softmax grader, test data set is inputted learning network, obtain the heart beat of data exported of classifying, realize the automatic classification that the heart is clapped.

The result adopting method of the present invention to classify to heart beat of data and self label of heart beat of data contrast, result is as shown in table 1, the classifying quality that the heart that visible method of the present invention is clapped the normal heart, left bundle branch block, right bundle branch block, ventricular premature contraction, artrial premature beat, amalgamation heart beating 6 kinds are common is clapped is very good, the very high nicety of grading all reached.

Table 1 classification results test chart

Label	Sample number	Correct number of categories	Accuracy rate
				The normal heart is clapped	3600	3598	99.94％
Left bundle branch block	2400	2387	99.46％
				Right bundle branch block	2400	2398	99.92％
Ventricular premature contraction	1150	1129	98.17％
				Artrial premature beat	850	830	97.65％
Amalgamation heart beating	1200	1200	100％
				Amount to	11600	11542	99.50％

Claims (1)

1. an electrocardiosignal automatic classification method, is characterized in that, comprises the following steps:

A) obtain the electrocardiosignal of human body, and carry out Filtering Processing, the R ripple of the electrocardiosignal after detection filter;

B) after R ripple being detected, build data set, described data set is made up of some groups of heart beat of data, often organize described heart beat of data all with a kind of label, described label always has 6 kinds, is respectively the bat of the normal heart, left bundle branch block, right bundle branch block, ventricular premature contraction, artrial premature beat, amalgamation heart beating:

Often organize described heart beat of data and comprise 270 sampled points, these 270 sampled points are the positions of R ripple according to detecting, choose 90 sampled points before the wave crest point of described R ripple, described R ripple wave crest point choose 179 sampled points below;

C) sparse automatic encoding degree of deep learning network is built:

Described sparse automatic encoding degree of deep learning network has two hidden layers, after connect softmax grader, wherein, described two hidden layers are respectively the first hidden layer and the second hidden layer;

Described sparse automatic encoding degree of deep learning network be input as 270 sampled points, described first node in hidden layer is 130, and described second node in hidden layer is 50;

D) substep trains described sparse automatic encoding degree of deep learning network:

D-1) heart beat of data of described data set is normalized, then SAE model is inputted, adopt the first hidden layer of sparse automatic encoding degree of deep learning network described in SAE model training, obtain the network weight of the first hidden layer, and obtain the shallow-layer feature of heart beat of data;

D-2) same SAE model is adopted, described shallow-layer feature is inputted described SAE model, train the second hidden layer of described sparse automatic encoding degree of deep learning network, obtain the network weight of the second hidden layer, and obtain the high-level feature of described some groups of heart beat of data;

D-3) the high-level feature obtained is input to softmax grader, training softmax grader, obtains the network weight of softmax grader;

E) according to the network weight of the network weight of the first hidden layer of step d) gained, the network weight of the second hidden layer and softmax grader, will treat the described sparse automatic encoding degree of deep learning network of thought-read beat of data input, the classification obtaining heart beat of data exports.