CN102488518B - Electroencephalogram detection method and device by utilizing fluctuation index and training for promotion - Google Patents

Abstract

The invention relates to an electroencephalogram detection method and an electroencephalogram detection device by utilizing a fluctuation index and training for promotion. Electroencephalogram data which is acquired and processed is subjected to feature extraction by using the fluctuation index with a good feature effect, and the extracted feature vector is input into a classifier obtained through a method of training for promotion, so that marks of abnormal electroencephalogram signals are obtained. Therefore, the workload of clinicians for judging large-scale electroencephalogram data is relieved, and the timeliness of abnormal electroencephalogram detection is improved.

Description

A kind of brain electro-detection method and apparatus that utilizes the index of oscillation and training for promotion

Technical field

The invention discloses a kind of brain electro-detection method and apparatus that utilizes the index of oscillation and training for promotion, belong to brain electro-detection technical field.

Technical background

Epilepsy be a kind of with brain neuroblastoma unit repeatedly the intermittent central nervous system function imbalance due to the over-discharge suddenly be the brain disorder of feature.So far, it mainly is that the medical personnel relies on experience that electroencephalogram (EEG) is estimated to finish that epilepsy detects, and checks characteristic waves such as whether containing epileptiform discharge among the EEG, and its workload is big, causes the medical personnel tired and produce erroneous judgement easily.Therefore, in epilepsy detected, the accuracy that automatic checkout system detects the brain electricity had more consequence, and it can greatly improve the detection efficiency to EEG.

From the sixties in last century, the epilepsy detection technique just has been subjected to paying close attention to widely automatically, and numerous scholars in this field have proposed the method for multiple automatic detection brain electricity.Its main flow algorithm has support vector machine (SVM) and neutral net etc.And support vector machine is to find the solution support vector by quadratic programming, finds the solution the calculating that quadratic programming will be referred to high level matrix, and the storage of matrix and calculating will expend a large amount of machine internal memory and operation time.The device of epileptiform discharge has adopted neutral net and feedforward adverse transference to broadcast (BP) learning algorithm in the disclosed a kind of full-automatic detection by quantitative electroencephalogram of CN1253762A (99124032.4).Neutral net must be carried out repeatedly repetitive learning, and training speed is slow, and computational efficiency is low.Because the BP algorithm is a kind of optimization algorithm of Local Search, solve the global extremum of complex nonlinear function with it simultaneously, cause failure to train thereby probably be absorbed in local extremum.

Summary of the invention

At the deficiencies in the prior art, the present invention proposes a kind of brain electro-detection method of utilizing the index of oscillation and training for promotion, this method with the EEG signals index of oscillation that extracts as input parameter, the grader of sending into by training for promotion calculates, obtain the output probability value, output probability value and predetermined threshold value are compared, thereby obtain brain electro-detection result.

The present invention also provides a kind of device that utilizes said method to detect the brain electricity.

Technical scheme of the present invention is as follows:

A kind of brain electro-detection method of utilizing the index of oscillation and training for promotion, step is as follows:

1) utilizes eeg amplifier and data collecting card to gather EEG signals, the EEG signals that collects is changed by A/D, store in the computer;

2) computer carries out filtering and denoising to EEG signals;

3) computer extracts the index of oscillation of each each wavelet layer of passage of EEG signals;

4) index of oscillation input grader that step 3) is extracted calculates, and obtains the output probability value;

5) output probability value and predetermined threshold value are compared, obtain brain electro-detection result and labelling:

The output probability value judges then that greater than predetermined threshold value it is unusual detecting the brain electricity, is labeled as 1;

The output probability value is less than or equal to predetermined threshold value, judges that then it is normal detecting the brain electricity, is labeled as-1.

Preferably, described predetermined threshold value is 0.5.

Eeg amplifier described in the step 1) is Neurofile NT eeg amplifier, and described data collecting card is 16 A/D translation data capture cards, and sample frequency is 256Hz.

Step 2) method step that the computer described in carries out filtering and denoising to EEG signals is as follows:

Gather the EEG signals that a segment length is LEN, utilize the Daubechies-4 small echo to carry out S layer wavelet decomposition, preferred S=5; Subsequently the EEG signals after decomposing is carried out signal reconstruction, extract the 3-30Hz frequency range of reconstruction signal, i.e. the 3rd, 4,5 layer of reconstruction signal a _{J, n}, a _{J, n}Representing length is the EEG signals j channel signal x of LEN _jN layer wavelet reconstruction signal, j=1 wherein, 2 ..., C, n=3,4,5; C is port number, preferred C=6.

Preferred LEN=1024.

The method of the index of oscillation of each each wavelet layer of passage of extraction EEG signals described in the step 3) is:

Utilize formula (1) calculation procedure 2) j channel signal x in the EEG signals _jThe index of oscillation wav of n layer wavelet reconstruction signal _{J, n}For

${\mathrm{wav}}_{j,n}=\underset{l=1}{\overset{\mathrm{LEN}-1}{\mathrm{\Σ}}}|a_{j,n}(l+1)-a_{j,n}\left(l\right)|---\left(1\right).$

The method of passing through classifier calculated output probability value described in the step 4) is:

With the index of oscillation wav in the step 3) _{J, n}Send into grader F as characteristic vector w, utilize formula (2)

$p(y=1|w)=\frac{e^{F\left(w\right)}}{e^{F\left(w\right)}+e^{-F\left(w\right)}}---\left(2\right)$

The EEG signals that obtains length LEN is the probability P of unusual brain electricity.

Grader described in the step 4) is to obtain by following training for promotion method, and the specific implementation step is:

A) W is the eeg data that grader is trained used segmentation, W={w _i∈ R ^k, i=1,2, L, N}, K=C * S wherein, C is the port number of EEG, and S is the small echo number of plies, and N is the data hop count, and every segment length is LEN; Y is the correspondence markings amount, Y={y _i∈ 1,1}, and i=1,2, L, N}, labelled amount is-1 expression normal brain activity electricity, labelled amount is the unusual brain of 1 expression; w _iBe the index of oscillation wav of each passage the 3rd of i section EEG signals, 4,5 layers of wavelet reconstruction signal _{J, n}The characteristic vector of forming; F _mThe grader that expression m sets up the step back; The setting iterations is M; Set i section EEG signals characteristic vector w _iThe initial probability that belongs to unusual brain electricity is p ₀(y _i=1|w _i)=0.5, i=1,2, L, N; Set i section EEG signals characteristic vector w _iThe preliminary classification device be F ₀(w _i)=0, i=1,2, L, N; N=900, M=180.

B) m represents the iteration step number, begins to carry out following loop iteration from m=1:

I. calculate grader F _mThe first derivative of likelihood function

P wherein _M-1(y _i=1|w _i) characteristic vector w after expression m-1 step iteration _iThe probit that belongs to unusual brain electricity;

Ii. by method of least square by w _iRight

Match obtains regression coefficient r, with f (w _i) expression i section EEG signals characteristic vector w _iWeak Classifier:

f(w _i)＝r ^Tw _i，i＝1，2，L，N；

Iii. obtain the Weak Classifier f that selects for use after m iteration _m

Iv. the Bernoulli Jacob's regression function L (F that releases by training data _mW Y) can be expressed as:

$L(F_m;W,Y)=\log \left(\underset{i=1}{\overset{N}{\mathrm{\Π}}}{p}_m{(y_i=1|w_i)}^{y_i}{p}_m{(y_i=0|w_i)}^{1-y_i}\right),i=\mathrm{1,2},L,N;$

V. calculate m step back Weak Classifier weight coefficient γ _mFor

${\mathrm{\γ}}_m=\arg \underset{\mathrm{\γ}}{\max }L(F_{m-1}+\mathrm{\γ}{f}_m;W,Y);$

Vi. upgrade grader

F _m＝F _m-1+εγ _mf _m

Wherein ε is a minimum value, ε=0.05;

Vii. by grader F _mCalculated characteristics vector w _iThe probit that belongs to unusual brain electricity:

$p_m(y_i=1|w_i)=\frac{e^{F_m\left(w_i\right)}}{e^{F_m\left(w_i\right)}+e^{-F_m\left(w_i\right)}},i=\mathrm{1,2},L,N$

Wherein, F _m(w _i) the corresponding training data w in expression m step back _iGrader;

Viii. make m=m+1, repeat above-mentioned circulation, if m=M then loop iteration finish, obtain grader F=F _M

A kind of device that utilizes said method to carry out the brain electro-detection, comprise the eeg amplifier, data collecting card and the computer that connect with circuit, be built-in with in the described computer and utilize the index of oscillation and training for promotion method to detect the brain electro-detection module of brain electricity, be transferred in the computer after utilizing eeg amplifier and data collecting card that EEG signals is gathered, the brain electro-detection module of utilizing the index of oscillation and training for promotion method to detect the brain electricity is carried out filtering and denoising to EEG signals; Extract the index of oscillation of every section EEG signals as characteristic vector; Characteristic vector is sent in the grader that is obtained with the training for promotion method, obtained the output probability value; With output probability value and predetermined threshold value relatively, get brain electro-detection result and labelling in addition.

Useful effect of the present invention is:

Utilize characteristic effect preferably the index of oscillation to gathering and carrying out feature extraction through pretreated eeg data, the characteristic vector of extracting is sent in the grader that is obtained by the training for promotion method, thereby obtain the labelling to unusual EEG signals, not only alleviate the workload that the clinician differentiates extensive eeg data, and improved ageing to unusual brain electro-detection.

Description of drawings

Fig. 1 is the flow chart of brain electro-detection method of the present invention;

Fig. 2 is the hardware connection layout of EEG checking device of the present invention;

Fig. 3 is the index of oscillation of EEG signals described in the embodiment 1, wherein is the unusual brain electricity persistent period between two vertical lines;

Fig. 4 is the classification results of the described EEG signals of Fig. 3: wherein, and 1 expression unusual brain electricity, i.e. epileptiform discharge;-1 expression normal brain activity; It is the unusual brain electricity persistent period between two vertical lines.

The specific embodiment

The present invention will be further described below in conjunction with accompanying drawing and example, and obviously the present invention is not limited to this.

Embodiment 1,

As shown in Figure 1, a kind of brain electro-detection method of utilizing the index of oscillation and training for promotion, step is as follows:

1) utilize Neurofile NT eeg amplifier and 16 A/D translation data capture cards to gather EEG signals, sample frequency is 256Hz, and the EEG signals that collects is changed by A/D, stores in the computer.

2) computer carries out filtering and denoising to EEG signals, and its method step is as follows:

Gather the EEG signals that a segment length is LEN=1024, utilize the Daubechies-4 small echo to carry out S layer wavelet decomposition, preferred S=5; Subsequently the EEG signals after decomposing is carried out signal reconstruction, extract the 3-30Hz frequency range of reconstruction signal, i.e. the 3rd, 4,5 layer of reconstruction signal a _{J, n}, a _{J, n}Representing length is the EEG signals j channel signal x of LEN _jN layer wavelet reconstruction signal, j=1 wherein, 2 ..., C, n=3,4,5; C is port number, C=6.

3) method of the index of oscillation of each each wavelet layer of passage of computer extraction EEG signals is:

Utilize formula (1) calculation procedure 2) j channel signal x in the EEG signals _jThe index of oscillation wav of n layer wavelet reconstruction signal _{J, n}For

${\mathrm{wav}}_{j,n}=\underset{l=1}{\overset{\mathrm{LEN}-1}{\mathrm{\Σ}}}|a_{j,n}(l+1)-a_{j,n}\left(l\right)|---\left(1\right);$

It is the index of oscillation of the 3rd layer of wavelet reconstruction signal of EEG signals the 1st passage shown in Fig. 3.

4) index of oscillation input grader that step 3) is extracted calculates, and obtains the output probability value;

Described method by classifier calculated output probability value is:

With the index of oscillation wav in the step 3) _{J, n}Send into grader F as characteristic vector w, utilize formula (2)

$p(y=1|w)=\frac{e^{F\left(w\right)}}{e^{F\left(w\right)}+e^{-F\left(w\right)}}---\left(2\right)$

The EEG signals that obtains length LEN is the probability P of unusual brain electricity;

Described grader is to obtain by following training for promotion method, and the specific implementation step is:

A) W is the eeg data that grader is trained used segmentation, W={w _i∈ R ^k, i=1,2, L, N}, K=C * S wherein, C is the port number of EEG, and S is the small echo number of plies, and N is the data hop count, and every segment length is LEN; Y is the correspondence markings amount, Y={y _i∈ 1,1}, and i=1,2, L, N}, labelled amount is-1 expression normal brain activity electricity, labelled amount is the unusual brain of 1 expression; w _iBe the index of oscillation wav of each passage the 3rd of i section EEG signals, 4,5 layers of wavelet reconstruction signal _{J, n}The characteristic vector of forming; F _mThe grader that expression m sets up the step back; The setting iterations is M; Set i section EEG signals characteristic vector w _iThe initial probability that belongs to unusual brain electricity is p ₀(y _i=1|w _i)=0.5, i=1,2, L, N; Set i section EEG signals characteristic vector w _iThe preliminary classification device be F ₀(w _i)=0, i=1,2, L, N; N=900, M=180;

B) m represents the iteration step number, begins to carry out following loop iteration from m=1:

I. calculate grader F _mThe first derivative of likelihood function

P wherein _M-1(y _i=1|w _i) characteristic vector w after expression m-1 step iteration _iThe probit that belongs to unusual brain electricity;

Ii. by method of least square by w _iRight

Match obtains regression coefficient r, with f (w _i) expression i section EEG signals characteristic vector w _iWeak Classifier:

f(w _i)＝r ^Tw _i，i＝1，2，L，N；

Iii. obtain the Weak Classifier f that selects for use after m iteration _m

Iv. the Bernoulli Jacob's regression function L (F that releases by training data _mW Y) can be expressed as:

$L(F_m;W,Y)=\log \left(\underset{i=1}{\overset{N}{\mathrm{\Π}}}{p}_m{(y_i=1|w_i)}^{y_i}{p}_m{(y_i=0|w_i)}^{1-y_i}\right),i=\mathrm{1,2},L,N;$

V. calculate m step back Weak Classifier weight coefficient γ _mFor

${\mathrm{\γ}}_m=\arg \underset{\mathrm{\γ}}{\max }L(F_{m-1}+\mathrm{\γ}{f}_m;W,Y);$

Vi. upgrade grader

F _m＝F _m-1+εγ _mf _m

Wherein ε is a minimum value, ε=0.05;

Vii. by grader F _mCalculated characteristics vector w _iThe probit that belongs to unusual brain electricity:

$p_m(y_i=1|w_i)=\frac{e^{F_m\left(w_i\right)}}{e^{F_m\left(w_i\right)}+e^{-F_m\left(w_i\right)}},i=\mathrm{1,2},L,N$

Wherein, F _m(w _i) the corresponding training data w in expression m step back _iGrader;

Viii. make m=m+1, repeat above-mentioned circulation, if m=M then loop iteration finish, obtain grader F=F _M

5) output probability value and predetermined threshold value are compared, obtain brain electro-detection result and labelling:

The output probability value judges then that greater than predetermined threshold value it is unusual detecting the brain electricity, is labeled as 1;

The output probability value is less than or equal to predetermined threshold value, judges that then it is normal detecting the brain electricity, is labeled as-1; Described predetermined threshold value is 0.5.

Fig. 4 is the labelling result of the described EEG signals of Fig. 3.

Embodiment 2,

A kind of embodiment of utilization 1 described method is carried out the device of brain electro-detection, comprise the eeg amplifier, data collecting card and the computer that connect with circuit, be built-in with in the described computer and utilize the index of oscillation and training for promotion method to detect the brain electro-detection module of brain electricity, be transferred in the computer after utilizing eeg amplifier and data collecting card that EEG signals is gathered, the brain electro-detection module of utilizing the index of oscillation and training for promotion method to detect the brain electricity is carried out filtering and denoising to EEG signals; Extract the index of oscillation of every section EEG signals as characteristic vector; Characteristic vector is sent in the grader that is obtained with the training for promotion method, obtained the output probability value; With output probability value and predetermined threshold value relatively, get brain electro-detection result and labelling in addition.

Utilize the present invention that 21 routine epileptics' brain electricity is detected, to the rate of accuracy reached 94% that the epileptic paradoxical discharge detects, per hour the error detection number of times is 0.2 time.

Claims (2)

1. a brain electro-detection method of utilizing the index of oscillation and training for promotion is carried out the device of brain electro-detection, it is characterized in that, described device comprises eeg amplifier, data collecting card and the computer that connects with circuit, be built-in with in the described computer and utilize the index of oscillation and training for promotion method to detect the brain electro-detection module of brain electricity, be transferred in the computer after utilizing eeg amplifier and data collecting card that EEG signals is gathered, the brain electro-detection module of utilizing the index of oscillation and training for promotion method to detect the brain electricity is carried out filtering and denoising to EEG signals; Extract the index of oscillation of every section EEG signals as characteristic vector; Characteristic vector is sent in the grader that is obtained with the training for promotion method, obtained the output probability value; With output probability value and predetermined threshold value relatively, get brain electro-detection result and labelling in addition;

Wherein said eeg amplifier and the data collecting card of utilizing gathered EEG signals, and the EEG signals that collects is changed by A/D, stores in the computer; Described eeg amplifier is Neurofile NT eeg amplifier, and described data collecting card is 16 A/D translation data capture cards, and sample frequency is 256Hz;

The brain electro-detection module of the wherein said index of oscillation and training for promotion method detection brain electricity is as follows to the method step that EEG signals is carried out filtering and denoising:

Gather the EEG signals that a segment length is LEN, utilize the Daubechies-4 small echo to carry out S layer wavelet decomposition; Subsequently the EEG signals after decomposing is carried out signal reconstruction, extract the 3-30Hz frequency range of reconstruction signal, i.e. the 3rd, 4,5 layer of reconstruction signal a _{J, n}, a _{J, n}Representing length is the EEG signals j channel signal x of LEN _jN layer wavelet reconstruction signal, j=1 wherein, 2 ..., C, n=3,4,5; C is port number; Described port number C=6; Described LEN=1024;

The method that the brain electro-detection module of the described index of oscillation and training for promotion method detection brain electricity is extracted the index of oscillation of each each wavelet layer of passage of EEG signals is:

Utilize j channel signal x in formula (1) the calculation procedure EEG signals _jThe index of oscillation wav of n layer wavelet reconstruction signal _{J, n}For

${\mathrm{wav}}_{j,n}=\underset{l=1}{\overset{\mathrm{LEN}-1}{\mathrm{\Σ}}}|a_{j,n}(l+1)-a_{j,n}\left(l\right)|---\left(1\right);$

Described method by classifier calculated output probability value is:

With index of oscillation wav _{J, n}Send into grader F as characteristic vector w, utilize formula (2)

$p(y=1|w)=\frac{e^{F\left(w\right)}}{e^{F\left(w\right)}+e^{-F\left(w\right)}}---\left(2\right)$

The EEG signals that obtains length LEN is the probability P of unusual brain electricity;

Described grader is to obtain by following training for promotion method, and the specific implementation step is:

A) W is the eeg data that grader is trained used segmentation, W={w _i∈ R ^k, i=1,2 ..., N}, K=C * S wherein, C is the port number of EEG, and S is the small echo number of plies, and N is the data hop count, and every segment length is LEN; Y is the correspondence markings amount, Y={y _i∈ 1,1}, and i=1,2 ..., N}, labelled amount is-1 expression normal brain activity electricity, labelled amount is the unusual brain of 1 expression; w _iBe the index of oscillation wav of each passage the 3rd of i section EEG signals, 4,5 layers of wavelet reconstruction signal _{J, n}The characteristic vector of forming; F _mThe grader that expression m sets up the step back; The setting iterations is M; Set i section EEG signals characteristic vector w _iThe initial probability that belongs to unusual brain electricity is p ₀(y _i=1|w _i)=0.5, i=1,2 ..., N; Set i section EEG signals characteristic vector w _iThe preliminary classification device be F ₀(w _i)=0, i=1,2 ..., N; N=900, M=180;

B) m represents the iteration step number, begins to carry out following loop iteration from m=1:

I. calculate grader F _mThe first derivative of likelihood function

${\tilde{y}}_i=2(y_i-p_{m-1}(y_i=1|w_i)),i=\mathrm{1,2},\·\·\·,N$

P wherein _M-1(y _i=1|w _i) characteristic vector w after expression m-1 step iteration _iThe probit that belongs to unusual brain electricity;

Ii. by method of least square by w _iRight

Match obtains regression coefficient r, with f (w _i) expression i section EEG signals characteristic vector w _iWeak Classifier:

f(w _i)＝r ^Tw _i,i＝1,2,…,N；

Iii. obtain the Weak Classifier f that selects for use after m iteration _m

$f_m=\arg \underset{f}{\min }\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{({\tilde{y}}_i-f\left(w_i\right))}^2;$

Iv. the Bernoulli Jacob's regression function L (F that releases by training data _m; W Y) can be expressed as:

$L(F_m;W,Y)=\log \left(\underset{i=1}{\overset{N}{\mathrm{\Π}}}{p}_m{(y_i=1|w_i)}^{y_i}{p}_m{(y_i=0|w_i)}^{1-y_i}\right),i=\mathrm{1,2},\·\·\·,N;$

V. calculate m step back Weak Classifier weight coefficient γ _mFor

${\mathrm{\γ}}_m=\arg \underset{\mathrm{\γ}}{\max }L(F_{m-1}+\mathrm{\γ}{f}_m;W,Y);$

Vi. upgrade grader

F _m＝F _m-1+εγ _mf _m

Wherein ε is a minimum value, ε=0.05;

Vii. by grader F _mCalculated characteristics vector w _iThe probit that belongs to unusual brain electricity:

$p_m(y_i=1|w_i)=\frac{e^{F_m\left(w_i\right)}}{e^{F_m\left(w_i\right)}+e^{-F_m\left(w_i\right)}},i=\mathrm{1,2},\·\·\·,N$

Wherein, F _m(w _i) the corresponding training data w in expression m step back _iGrader;

Viii. make m=m+1, repeat above-mentioned circulation, if m=M then loop iteration finish, obtain grader F=F _MDescribed predetermined threshold value is 0.5.

2. a kind of brain electro-detection method of utilizing the index of oscillation and training for promotion according to claim 1 is carried out the device of brain electro-detection, it is characterized in that, utilizes the Daubechies-4 small echo to carry out S layer wavelet decomposition, described S=5.

Images