CN109903852A - User-customized intelligent epilepsy prediction method based on PCA-LDA - Google Patents
User-customized intelligent epilepsy prediction method based on PCA-LDA Download PDFInfo
- Publication number
- CN109903852A CN109903852A CN201910048393.6A CN201910048393A CN109903852A CN 109903852 A CN109903852 A CN 109903852A CN 201910048393 A CN201910048393 A CN 201910048393A CN 109903852 A CN109903852 A CN 109903852A
- Authority
- CN
- China
- Prior art keywords
- lda
- frequency sub
- sample
- coefficient
- band
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 25
- 206010015037 epilepsy Diseases 0.000 title abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 14
- 238000000513 principal component analysis Methods 0.000 claims abstract description 12
- 238000012952 Resampling Methods 0.000 claims abstract description 11
- 239000000284 extract Substances 0.000 claims abstract description 5
- 230000001037 epileptic effect Effects 0.000 claims description 30
- 230000009467 reduction Effects 0.000 claims description 23
- 210000004556 brain Anatomy 0.000 claims description 21
- 230000005611 electricity Effects 0.000 claims description 21
- 238000000354 decomposition reaction Methods 0.000 claims description 20
- 230000004927 fusion Effects 0.000 claims description 16
- 238000000605 extraction Methods 0.000 claims description 9
- 230000002109 interictal effect Effects 0.000 claims description 8
- 238000007637 random forest analysis Methods 0.000 claims description 5
- 238000011946 reduction process Methods 0.000 claims description 4
- 101000617541 Danio rerio Presenilin-2 Proteins 0.000 claims description 3
- 238000003066 decision tree Methods 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 206010010904 Convulsion Diseases 0.000 abstract description 11
- 238000004458 analytical method Methods 0.000 abstract description 4
- 208000028329 epileptic seizure Diseases 0.000 abstract description 3
- 238000012544 monitoring process Methods 0.000 description 5
- 238000013480 data collection Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
Landscapes
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
本发明公开了一种基于PCA‑LDA用户定制的智能癫痫预测方法。本发明步骤如下：步骤1、脑电EEG信号的类别划分及基于SMOTE的样本重采样算法，将样本不均衡分类问题转换成样本均衡分类问题；步骤2、对脑电EEG信号进行6层小波包变换，提取统计学特征和小波包特征；步骤3、针对SFs与WPFs融合的高维特征，采用主成分分析法和线性判别分析法做特征降维；步骤4、应用深度森林算法构建针对不均衡样本数据集的癫痫预测模型。本发明既可以克服临床脑电EEG信号样本不均衡问题，又可以解决高维特征训练时间长效率低的问题，还能够实现癫痫发作期与发作前期的精准预测。
The invention discloses an intelligent epilepsy prediction method based on PCA-LDA user customization. The steps of the present invention are as follows: Step 1, classification of the EEG signal and a sample resampling algorithm based on SMOTE, to convert the sample imbalance classification problem into the sample balanced classification problem; Step 2, perform 6-layer wavelet packet on the EEG signal. Transform, extract statistical features and wavelet packet features; step 3, use principal component analysis and linear discriminant analysis to reduce the feature dimension for the high-dimensional features fused by SFs and WPFs; step 4, apply deep forest algorithm to construct Epilepsy prediction model for a sample dataset. The invention can not only overcome the problem of unbalanced clinical EEG signal samples, but also solve the problem of long time and low efficiency of high-dimensional feature training, and can also achieve accurate prediction of epileptic seizure period and pre-epilepsy period.
Description
Technical field
The invention belongs to EEG Processing and field of wisdom medical treatment, are related to a kind of special based on PCA-LDA fusion statistics
The customized intelligent Epileptic Prediction of (SFs) and wavelet packet character (WPFs) is levied, the present invention is based on unbalanced brain electricity
The intellectual monitoring algorithm of EEG data collection.
Background technique
Traditional epileptics monitoring model is usually the brain electricity EEG signal using a large amount of epileptics to construct, and is built
The premise of mould often assumes that stage of attack, breaking-out early period and interictal sample are balanced.And in practical clinical, by
In the transience of the sporadic and stage of attack of epilepsy, what the epileptic seizure prediction algorithm of personal user's customization almost all suffered from is one
The modeling problem of a imbalanced data sets.For single epileptic, time stage of attack is relative to interictal and breaks out early period,
Its corresponding EEG sample set is that extreme is unbalanced, and traditional prediction algorithm based on the building of equalization data collection can not solve this
Class problem.For this problem, using SMOTE resampling methods, the invention proposes one kind to merge statistics based on PCA-LDA
The customized intelligent Epileptic Prediction of feature (SFs) and wavelet packet character (WPFs)；Meanwhile for breaking-out EEG early period
Sample of signal does classification, to realize the accurate detection of epileptic attack early period and stage of attack.
Summary of the invention
The present invention is sporadic for epileptics, and the epileptic attack phase is short and this longer feature of other periods, utilizes SMOTE
Resampling methods handle the imbalanced training sets problem, propose the feature extracting method based on PCA-LDA fusion SFs and WPFs,
And epileptic attack early period and accurate prediction model stage of attack of unbalanced sample are constructed in conjunction with depth forest algorithm.
Technical solution of the present invention mainly includes the following steps:
Step 1, the category division of brain electricity EEG signal and be based on SMOTE (Synthetic Minority
Oversampling Technique) sample resampling methods, imbalanced training sets classification problem is converted into sample equilibrium point
Class problem.
Step 2 carries out 6 layers of wavelet package transforms to brain electricity EEG signal, extracts statistics feature (SFs) and wavelet packet character
(WPFs)。
Step 3, the high dimensional feature merged for SFs with WPFs, using Principal Component Analysis (PCA) and linear discriminant analysis
Method (LDA) does Feature Dimension Reduction.
Step 4 is directed to unequal sample numbers according to the epileptic prediction model of collection using the building of depth forest algorithm.
The step 1, the specific steps are as follows:
The category division of 1-1 brain electricity EEG signal.Seizure class will be denoted as stage of attack；By the data after four hours stages of attack
Four hours data are denoted as interictal Interictal class before stage of attack next time；Breaking-out is divided into 3 one hour early period
A continuous and nonoverlapping period, i.e., breaking-out first 60 minutes to breaking-out first 40 minutes are Pre1 class, are extremely broken out within breaking-out first 40 minutes
First 20 minutes are Pre2 class, and breaking-out first 20 minutes to breaking-out first 0 minute are Pre3 class.
1-2 is since the data time of Seizure class in the brain electricity EEG signal of single epileptic is shorter, and other classes
Other data time is longer, results in imbalanced training sets, to solve this problem, using SMOTE resampling methods to Seizure class
Sample is analyzed, generates new group sample to reach data balancing, and balanced sample classification problem is formed.
The step 2, the specific steps are as follows:
For the every class signal of 2-1 using mono- frame of 2s as input x, taking frame Duplication is 50%.To each of each frame input signal
Channel carries out 6 layers of WAVELET PACKET DECOMPOSITION, and therefore, the last layer after decomposition includes 64 frequency sub-band sm, m=1,2 ..., 64.
In 15 frequency sub-band s of low frequencym′, in m '=1,2 ..., 15, mean value, standard deviation, intermediate value, coefficient of kurtosis are calculated to each frequency range
With coefficient of skew constitutive characteristic SFs.For the brain electricity EEG signal in M channel, this five statistics features constitute 75*M dimension
Feature vector SFs.
2-2 calculates mean value by following formula:
2-3 calculates standard deviation by following formula:
2-4 calculates median by the following method:
By sm′(k) it sorts from small to large ord are as follows:Then when N is odd number, median isWhen N is even number, median is
2-5 calculates coefficient of kurtosis by following formula:
2-6 calculates the coefficient of skew by following formula
Wherein sm′, m '=1,2 ..., 15 indicates the small echo after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band reconstruct of layer 6
Packet coefficient, N indicate the points that each frequency sub-band includes, and μ indicates that a frequency sub-band Coefficient Mean of m ', σ indicate a frequency sub-band of m '
Factor standard is poor.
2-7 characteristic vector W PFs is made of energy ratio and three kinds of entropys, i.e. the energy ratio of 64 sub- frequency extractions of the last layer
With the Shannon entropy, logarithmic energy entropy and norm entropy of 15 sub- frequency extractions of low frequency.The brain electricity EEG for being N for M channel frame length
Signal, energy ratio and three kinds of entropys constitute the characteristic vector W PFs of 109*M dimension.
2-8 calculates energy ratio by following formula:
Wherein sm(m=1,2 ..., 65) indicate WAVELET PACKET DECOMPOSITION after layer 6 m-th of frequency sub-band coefficient reconstruct after
Wavelet packet coefficient, N indicate the points that each frequency sub-band includes, and μ indicates that m-th of frequency sub-band Coefficient Mean, σ indicate m-th of son frequency
Section factor standard is poor.
2-9 calculates Shannon entropy by following formula
2-10 calculates logarithmic energy entropy by following formula
2-11 calculates norm entropy by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes, and p >=1 is positive integer.
The step 3, the specific steps are as follows:
Multi-channel feature vector SFs and characteristic vector W PFs fusion is constituted the high dimensional feature vector of (184*M) dimension by 3-1,
Principal component analysis and dimensionality reduction are carried out to it using PCA technology, feature of the variance contribution ratio preceding 90% is retained in reduction process, i.e.,
Lower dimensional feature can be obtained.
3-2 carries out secondary dimensionality reduction to lower dimensional feature obtained in 3-1 using LDA.The dimension of the LDA dimensionality reduction of selection is less than
Equal to classification number.In this classification problem, sample class number is 5, therefore the dimension of settable LDA dimensionality reduction is 4 dimensions.
3-3 is operated by above-mentioned dimensionality reduction twice, and the high dimensional feature vector that the SFs of higher-dimension is merged with WPFs is down to low-dimensional,
The low-dimensional feature based on PCA-LDA fusion SFs and WPFs can be obtained.
3-4 carries out features described above extraction, fusion and dimensionality reduction to all training samples, and training set feature can be obtained.
The step 4, the specific steps are as follows:
4-1 combination depth forest algorithm, three levels of setting join, by adjusting the number of the random forest of knot cascading layers and every
The parameter of a random forest decision tree constructs optimal sorter model.
4-2 is to any test sample, using the feature extracting method above-mentioned based on PCA-LDA fusion SFs and WPFs, benefit
Prediction result is calculated with the sorter model built.
The present invention has the beneficial effect that
The present invention is customized for personal user faces that epileptics is sporadic and transience stage of attack in seizure monitoring system
Problem is constructed using SMOTE resampling methods based on PCA-LDA fusion SFs and WPFs and in conjunction with depth forest classified device
The accurate prediction model of epilepsy can not only overcome the problems, such as clinical brain electricity EEG signal imbalanced training sets, but also can solve high dimensional feature instruction
The problem of practicing time long low efficiency, additionally it is possible to the accurate prediction realizing the epileptic attack phase and breaking out early period.
Detailed description of the invention
Algorithm flow chart Fig. 1 of the invention.
The depth forest cascade structure that Fig. 2 present invention uses.
Specific embodiment
It elaborates with reference to the accompanying drawings and detailed description to the present invention.
As illustrated in fig. 1 and 2, based on the customized intelligent Epileptic Prediction of PCA-LDA, the general epileptic attack phase with
The realization step of breaking-out prediction technique early period has detailed introduction in summary of the invention, i.e. technical solution of the present invention is mainly wrapped
Include following steps:
Step 1, the category division of brain electricity EEG signal and be based on SMOTE (Synthetic Minority
Oversampling Technique) small sample resampling methods, by imbalanced training sets classification problem be converted into sample equilibrium
Classification problem.
Step 2 carries out 6 layers of wavelet package transforms to brain electricity EEG signal, extracts statistics feature (SFs) and wavelet packet character
(WPFs)。
Step 3, the high dimensional feature merged for SFs with WPFs, using Principal Component Analysis (PCA) and linear discriminant analysis
Method (LDA) does Feature Dimension Reduction.
Step 4 is directed to unequal sample numbers according to the epileptic prediction model of collection using the building of depth forest algorithm.
The step 1, the specific steps are as follows:
The category division of 1-1 brain electricity EEG signal.Seizure class will be denoted as stage of attack；By the data after four hours stages of attack
Four hours data are denoted as interictal Interictal class before stage of attack next time；Breaking-out is divided into 3 one hour early period
A continuous and nonoverlapping period, i.e., breaking-out first 60 minutes to breaking-out first 40 minutes are Pre1 class, are extremely broken out within breaking-out first 40 minutes
First 20 minutes are Pre2 class, and breaking-out first 20 minutes to breaking-out first 0 minute are Pre3 class.
1-2 is since the data time of Seizure class in the brain electricity EEG signal of single epileptic is shorter, and other classes
Other data time is longer, results in imbalanced training sets, to solve this problem, using SMOTE resampling methods to Seizure class
Sample is analyzed, generates new group sample to reach data balancing, and balanced sample classification problem is formed.
The step 2, the specific steps are as follows:
For the every class signal of 2-1 using mono- frame of 2s as input x, taking frame Duplication is 50%.To each of each frame input signal
Channel carries out 6 layers of WAVELET PACKET DECOMPOSITION, and therefore, the last layer after decomposition includes 64 frequency sub-band sm, m=1,2 ..., 64.
In 15 frequency sub-band s of low frequencym′, in m '=1,2 ..., 15, mean value, standard deviation, intermediate value, coefficient of kurtosis are calculated to each frequency range
With coefficient of skew constitutive characteristic SFs.For the brain electricity EEG signal in M channel, this five statistics features constitute (75*M) dimension
Feature vector SFs.
2-2 calculates mean value by following formula
Wherein sm′(15) m '=1 2 ..., indicates small after the m ' a frequency sub-band reconstruct of layer 6 after WAVELET PACKET DECOMPOSITION
Wave packet coefficient, N indicate the points that each frequency sub-band includes.
2-3 calculates standard deviation by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes, and μ indicates a frequency sub-band Coefficient Mean of m '.
2-4 calculates median by the following method, by sm′(k) it sorts from small to large ord are as follows:Then when N is odd number, median isWhen N is even number, median
For
2-5 calculates coefficient of kurtosis by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes, and μ indicates that a frequency sub-band Coefficient Mean of m ', σ indicate that m ' is a
Frequency sub-band factor standard is poor.
2-6 calculates the coefficient of skew by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes, and μ indicates that a frequency sub-band Coefficient Mean of m ', σ indicate that m ' is a
Frequency sub-band factor standard is poor.
2-7 feature WPFs is made of energy ratio and three kinds of entropys, i.e. the energy ratio of 64 sub- frequency extractions of the last layer and low
Shannon entropy, logarithmic energy entropy and the norm entropy of 15 sub- frequency extractions of frequency.The brain electricity EEG that M channel frame length is N is believed
Number, energy ratio and three kinds of entropys constitute the characteristic vector W PFs of (109*M) dimension.
2-8 calculates energy ratio by following formula
Wherein sm(m=1,2 ..., 64) indicate WAVELET PACKET DECOMPOSITION after layer 6 m-th of frequency sub-band coefficient reconstruct after
Wavelet packet coefficient, N indicate the points that each frequency sub-band includes, and μ indicates that m-th of frequency sub-band Coefficient Mean, σ indicate m-th of son frequency
Section factor standard is poor.
2-9 calculates Shannon entropy by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes.
2-10 calculates logarithmic energy entropy by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes.
2-11 calculates norm entropy by following formula
Wherein sm′(15) m '=1 2 ..., is indicated after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band coefficient reconstruct of layer 6
Wavelet packet coefficient, N indicates the points that each frequency sub-band includes, and p >=1 is positive integer.
The step 3, the specific steps are as follows:
Multichannel SFs and WPFs Fusion Features are constituted the high dimensional feature vector of (184*M) dimension by 3-1, using PCA technology pair
It carries out principal component analysis and dimensionality reduction, and variance contribution ratio is retained in reduction process in preceding 90% feature, can be obtained compared with low-dimensional
Feature.
3-2 carries out secondary dimensionality reduction to lower dimensional feature obtained in 3-1 using LDA.Generally, the LDA dimensionality reduction of selection
Dimension is less than or equal to classification number.In this classification problem, sample class number is 5, therefore the dimension of settable LDA dimensionality reduction is 4 dimensions.
3-3 is operated by above-mentioned dimensionality reduction twice, and the SFs of higher-dimension and WPFs fusion feature are down to low-dimensional, base can be obtained
In the low-dimensional feature of PCA-LDA fusion SFs and WPFs.
3-4 carries out features described above extraction, fusion and dimensionality reduction to all training samples, and training set feature can be obtained.
The step 4, the specific steps are as follows:
4-1 combination depth forest algorithm, three levels of setting join, by adjusting the number of the random forest of knot cascading layers and every
The parameter of a random forest decision tree constructs optimal sorter model.
4-2 is to any test sample, using the feature extracting method above-mentioned based on PCA-LDA fusion SFs and WPFs, benefit
Prediction result is calculated with the sorter model built.
Join in order to reach better epileptic seizure phase and breaking-out accurate prediction effect early period, when below will be from practical application
Several selections and design aspect expansion are introduced, the reference to be used for other application as the invention:
For this method when handling epileptic electroencephalogram (eeg) EEG signal, it is 2 seconds that each frame input, which is arranged, and the Duplication of frame is set as 50%.
Aiming at the problem that clinically personal user customizes and faces sporadic epileptics and transience stage of attack in monitoring system, present invention benefit
Imbalanced training sets problem is handled with SMOTE resampling methods.The statistics feature and 2-9 to 2-11 that step 2-1 to 2-6 is extracted mention
The three kinds of entropys taken only extract on 15 low frequency frequency sub-band, this is because containing the most main of epileptic EEG Signal in low frequency band
Information is wanted, and the energy ratio that step 2-8 is extracted is extracted on 64 frequency sub-band, this is to guarantee the complete of energy information
Property.
In PCA reduction process, retain feature of the variance contribution ratio preceding 90%；In bis- dimensionality reductions of LDA, after dimensionality reduction is set
Dimension be 4 dimension.The depth forest classified algorithm finally used be provided with that three levels hold examination jointly worry to be if that cascading layers are very few can lead
Cause classifying quality poor, if excessively to will lead to the training time too long for cascading layers.
It is proposed of the invention can be used for solving clinically customizing in seizure monitoring system for personal user facing epileptics
The problem of sporadic and transience stage of attack, by the artificial synthesized new minority class sample of analysis to minority class sample, and adopts
Low-dimensional feature vector is obtained with the method based on PCA-LDA fusion SFs and WPFs, can be reached in conjunction with depth forest classified model
Accurately epileptic attack phase and breaking-out prediction early period.
Claims (5)
1. based on the customized intelligent Epileptic Prediction of PCA-LDA, it is characterised in that include the following steps:
Step 1, the category division of brain electricity EEG signal and the sample resampling methods based on SMOTE ask imbalanced training sets classification
Topic is converted into sample equilibrium classification problem；
Step 2 carries out 6 layers of wavelet package transforms to brain electricity EEG signal, extracts statistics feature and wavelet packet character；
Step 3, the high dimensional feature merged for SFs with WPFs, do feature using Principal Component Analysis and Fisher face
Dimensionality reduction；
Step 4 is directed to unequal sample numbers according to the epileptic prediction model of collection using the building of depth forest algorithm.
2. according to claim 1 based on the customized intelligent Epileptic Prediction of PCA-LDA, it is characterised in that described
Step 1 is implemented as follows:
The category division of 1-1 brain electricity EEG signal；Seizure class will be denoted as stage of attack；By the data after four hours stages of attack under
Four hours data are denoted as interictal Interictal class before stage of attack；Breaking-out is divided into three companies one hour early period
Continuous and nonoverlapping period, i.e., breaking-out first 60 minutes to breaking-out first 40 minutes are Pre1 class, breaking-out extremely breaking-out preceding 20 in first 40 minutes
Minute is Pre2 class, and breaking-out first 20 minutes to breaking-out first 0 minute are Pre3 class；
1-2 is since the data time of Seizure class in the brain electricity EEG signal of single epileptic is shorter, and other classification numbers
It is longer according to the time, imbalanced training sets are resulted in, to solve this problem, using SMOTE resampling methods to Seizure class sample
It analyzed, generate new group sample to reach data balancing, form balanced sample classification problem.
3. according to claim 2 based on the customized intelligent Epileptic Prediction of PCA-LDA, it is characterised in that described
Step 2 is implemented as follows:
For the every class signal of 2-1 using mono- frame of 2s as input x, taking frame Duplication is 50%；To each channel of each frame input signal
6 layers of WAVELET PACKET DECOMPOSITION are carried out, therefore, the last layer after decomposition includes 64 frequency sub-band sm, m=1,2 ..., 64；Low
15 frequency sub-band s of frequencym′, in m '=1,2 ..., 15, mean value, standard deviation, intermediate value, coefficient of kurtosis and partially are calculated to each frequency range
State coefficient constitutive characteristic SFs；For the brain electricity EEG signal in M channel, this five statistics features constitute the feature of 75*M dimension
Vector SFs；
2-2 calculates mean value by following formula:
2-3 calculates standard deviation by following formula:
2-4 calculates median by the following method:
By sm′(k) it sorts from small to large ord are as follows:Then when N is odd number, median isWhen N is even number, median is
2-5 calculates coefficient of kurtosis by following formula:
2-6 calculates the coefficient of skew by following formula
Wherein sm′, m '=1,2 ..., 15 indicates the wavelet packet system after WAVELET PACKET DECOMPOSITION after the m ' a frequency sub-band reconstruct of layer 6
Number, N indicate the points that each frequency sub-band includes, and μ indicates that a frequency sub-band Coefficient Mean of m ', σ indicate a frequency sub-band coefficient of m '
Standard deviation；
2-7 characteristic vector W PFs is made of energy ratio and three kinds of entropys, i.e. the energy ratio of 64 sub- frequency extractions of the last layer and low
Shannon entropy, logarithmic energy entropy and the norm entropy of 15 sub- frequency extractions of frequency；The brain electricity EEG that M channel frame length is N is believed
Number, energy ratio and three kinds of entropys constitute the characteristic vector W PFs of 109*M dimension；
2-8 calculates energy ratio by following formula:
Wherein sm(m=1,2 ..., 64) indicate WAVELET PACKET DECOMPOSITION after layer 6 m-th of frequency sub-band coefficient reconstruct after wavelet packet
Coefficient, N indicate the points that each frequency sub-band includes, and μ indicates that m-th of frequency sub-band Coefficient Mean, σ indicate m-th of frequency sub-band coefficient
Standard deviation；
2-9 calculates Shannon entropy by following formula
2-10 calculates logarithmic energy entropy by following formula
2-11 calculates norm entropy by following formula
Wherein sm′(15) m '=1 2 ..., indicates small after the m ' a frequency sub-band coefficient reconstruct of layer 6 after WAVELET PACKET DECOMPOSITION
Wave packet coefficient, N indicate the points that each frequency sub-band includes, and p >=1 is positive integer.
4. according to claim 3 based on the customized intelligent Epileptic Prediction of PCA-LDA, it is characterised in that described
Step 3 is implemented as follows:
Multi-channel feature vector SFs and characteristic vector W PFs fusion is constituted the high dimensional feature vector of 184*M dimension by 3-1, using PCA
Technology carries out principal component analysis and dimensionality reduction to it, and variance contribution ratio is retained in reduction process in preceding 90% feature, can be obtained
Lower dimensional feature；
3-2 carries out secondary dimensionality reduction to lower dimensional feature obtained in 3-1 using LDA；The dimension of the LDA dimensionality reduction of selection is less than or equal to
Classification number；In this classification problem, sample class number is 5, therefore the dimension of settable LDA dimensionality reduction is 4 dimensions；
3-3 is operated by above-mentioned dimensionality reduction twice, and the high dimensional feature vector that the SFs of higher-dimension is merged with WPFs is down to low-dimensional
Obtain the low-dimensional feature that SFs and WPFs is merged based on PCA-LDA；
3-4 carries out features described above extraction, fusion and dimensionality reduction to all training samples, and training set feature can be obtained.
5. according to claim 4 based on the customized intelligent Epileptic Prediction of PCA-LDA, it is characterised in that described
Step 4 is implemented as follows:
4-1 combination depth forest algorithm, setting three levels join, by adjust knot cascading layers random forest number and each with
The parameter of machine forest decision tree constructs optimal sorter model；
4-2 is to any test sample, using the feature extracting method above-mentioned based on PCA-LDA fusion SFs and WPFs, using
The sorter model built calculates prediction result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910048393.6A CN109903852A (en) | 2019-01-18 | 2019-01-18 | User-customized intelligent epilepsy prediction method based on PCA-LDA |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910048393.6A CN109903852A (en) | 2019-01-18 | 2019-01-18 | User-customized intelligent epilepsy prediction method based on PCA-LDA |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109903852A true CN109903852A (en) | 2019-06-18 |
Family
ID=66943795
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910048393.6A Pending CN109903852A (en) | 2019-01-18 | 2019-01-18 | User-customized intelligent epilepsy prediction method based on PCA-LDA |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109903852A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112861903A (en) * | 2020-12-03 | 2021-05-28 | 南京航空航天大学 | Gearbox fault diagnosis method based on improved deep forest |
| CN114093501A (en) * | 2021-10-19 | 2022-02-25 | 杭州电子科技大学 | Intelligent auxiliary analysis method for children's motor epilepsy based on synchronous video and electroencephalogram |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096810A (en) * | 2011-01-26 | 2011-06-15 | 北京中星微电子有限公司 | Method and device for detecting fatigue state of user before computer |
| WO2015102733A1 (en) * | 2013-10-20 | 2015-07-09 | Massachusetts Institute Of Technology | Using correlation structure of speech dynamics to detect neurological changes |
| CN105160352A (en) * | 2015-08-18 | 2015-12-16 | 中南大学 | High-dimensional data subspace clustering projection effect optimization method based on dimension reconstitution |
| CN108962391A (en) * | 2018-05-02 | 2018-12-07 | 杭州电子科技大学 | Epileptics prediction technique early period based on wavelet packet character and random forest |
-
2019
- 2019-01-18 CN CN201910048393.6A patent/CN109903852A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096810A (en) * | 2011-01-26 | 2011-06-15 | 北京中星微电子有限公司 | Method and device for detecting fatigue state of user before computer |
| WO2015102733A1 (en) * | 2013-10-20 | 2015-07-09 | Massachusetts Institute Of Technology | Using correlation structure of speech dynamics to detect neurological changes |
| CN105160352A (en) * | 2015-08-18 | 2015-12-16 | 中南大学 | High-dimensional data subspace clustering projection effect optimization method based on dimension reconstitution |
| CN108962391A (en) * | 2018-05-02 | 2018-12-07 | 杭州电子科技大学 | Epileptics prediction technique early period based on wavelet packet character and random forest |
Non-Patent Citations (2)
| Title |
|---|
| 尤婷婷: "健康大数据预处理技术及其应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 陈群等: "基于深度森林的脑电注意力识别研究", 《电子设计工程》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112861903A (en) * | 2020-12-03 | 2021-05-28 | 南京航空航天大学 | Gearbox fault diagnosis method based on improved deep forest |
| CN114093501A (en) * | 2021-10-19 | 2022-02-25 | 杭州电子科技大学 | Intelligent auxiliary analysis method for children's motor epilepsy based on synchronous video and electroencephalogram |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sharma et al. | Environment sound classification using multiple feature channels and attention based deep convolutional neural network | |
| CN103544963B (en) | A kind of speech-emotion recognition method based on core semi-supervised discrimination and analysis | |
| CN105956624B (en) | Mental imagery brain electricity classification method based on empty time-frequency optimization feature rarefaction representation | |
| CN109378010A (en) | Neural network model training method, voice denoising method and device | |
| CN108363493B (en) | User characteristic model establishing method and system based on brain-computer interface and storage medium | |
| Zhao et al. | Deep CNN model based on serial-parallel structure optimization for four-class motor imagery EEG classification | |
| CN111461201B (en) | Sensor data classification method based on phase space reconstruction | |
| CN104586387A (en) | Method for extracting and fusing time, frequency and space domain multi-parameter electroencephalogram characters | |
| CN109599120A (en) | One kind being based on large-scale farming field factory mammal abnormal sound monitoring method | |
| CN113076878B (en) | Constitution Identification Method Based on Convolutional Network Structure of Attention Mechanism | |
| CN114091529B (en) | A method of EEG emotion recognition based on generative adversarial network data enhancement | |
| CN109497997A (en) | Based on majority according to the seizure detection and early warning system of acquisition | |
| CN104091602A (en) | Speech emotion recognition method based on fuzzy support vector machine | |
| CN107766781A (en) | A kind of method and its system of quick electrocardio identification | |
| CN110575141A (en) | A method for epilepsy detection based on generative adversarial networks | |
| CN113180659B (en) | Electroencephalogram emotion recognition method based on three-dimensional feature and cavity full convolution network | |
| CN109903852A (en) | User-customized intelligent epilepsy prediction method based on PCA-LDA | |
| CN109685071A (en) | Brain electricity classification method based on the study of common space pattern feature width | |
| CN107174209A (en) | Sleep stage based on nonlinear kinetics method by stages | |
| CN112382311A (en) | Infant crying intention identification method and device based on hybrid neural network | |
| CN108962391A (en) | Epileptics prediction technique early period based on wavelet packet character and random forest | |
| CN113349801A (en) | Imaginary speech electroencephalogram signal decoding method based on convolutional neural network | |
| CN108806725A (en) | Speech differentiation method, apparatus, computer equipment and storage medium | |
| CN108078563A (en) | A kind of EEG signal analysis method of integrated classifier | |
| Jia | [Retracted] Music Emotion Classification Method Based on Deep Learning and Explicit Sparse Attention Network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190618 |