CN109480833A - The pretreatment and recognition methods of epileptic's EEG signals based on artificial intelligence - Google Patents
The pretreatment and recognition methods of epileptic's EEG signals based on artificial intelligence Download PDFInfo
- Publication number
- CN109480833A CN109480833A CN201811569263.9A CN201811569263A CN109480833A CN 109480833 A CN109480833 A CN 109480833A CN 201811569263 A CN201811569263 A CN 201811569263A CN 109480833 A CN109480833 A CN 109480833A
- Authority
- CN
- China
- Prior art keywords
- eeg signals
- epileptic
- parameter
- random forest
- identification
- 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
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/369—Electroencephalography [EEG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7225—Details of analog processing, e.g. isolation amplifier, gain or sensitivity adjustment, filtering, baseline or drift compensation
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/725—Details of waveform analysis using specific filters therefor, e.g. Kalman or adaptive filters
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7253—Details of waveform analysis characterised by using transforms
- A61B5/7257—Details of waveform analysis characterised by using transforms using Fourier transforms
Abstract
The pretreatment and recognition methods of the invention discloses a kind of epileptic's EEG signals based on artificial intelligence, are related to brain science, epileptic attack clinical data identification technology field.The method acquires the EEG signals of epileptic, and pre-processes to the EEG signals；Time frequency analysis is carried out to pretreated signal and extracts time-frequency characteristics, meanwhile, extract the statistics feature of noiseless EEG signals；Then random forest identification model is constructed, the pattern-recognition test of epileptic's EEG signals of the different degrees of patient's condition is carried out.Feature extracting method provided by the invention, can simplify calculating process, improve the accuracy and actual effect of EEG's Recognition.Due to the highly-parallel of random forest recognition methods, operational efficiency is improved, realizes the function of Computer Automatic Recognition epileptic's EEG signals, provides technical support for the diagnosis that medical worker takes time and effort.
Description
Technical field
It is specifically a kind of to be based on artificial intelligence the present invention relates to brain science, epileptic attack clinical data identification technology field
Epileptic's EEG signals pretreatment and recognition methods.
Background technique
The breaking-out of epileptics is the clinical manifestation of intracerebral neuron paroxysm sexual abnormality supersynchronousization electrical activity, is had repeatedly
Property, it is sudden and temporary the features such as.Important tool of the EEG signals as research epileptic condition, the breaking-out that it reflects in real time
Information is that other physiology methods cannot provide.Currently, in the analysis and research of the EEG signals of epileptic, engineering
Habit is the powerful tool of epileptic EEG Signal identification, but most machine learning method identification EEG signals are with more complex
Calculating process not can guarantee the accuracy and actual effect of recognition methods.Computer based mode identification method has very much, such as
K nearest neighbor method, support vector machine method, neural network method etc..What these methods identified on the data set of different cerebral diseases
Accuracy is different, remains unworthiness for the brain power mode recognition methods of computer itself.Further for identification side
The parameter selection of method, which is fixed against, carries out artificial tune ginseng with rich experiences researcher, and the parameter of generation cannot be completely secured to know
The model optimized parameter of other method.
Summary of the invention
The present invention, for the EEG's Recognition problem of epileptic, proposes one to overcome the shortcomings of existing methods
The pretreatment and recognition methods of epileptic EEG signals of the kind based on artificial intelligence.
The pretreatment and recognition methods of a kind of epileptic's EEG signals based on artificial intelligence provided by the invention, including
The following steps:
Step 1: the EEG signals of epileptic are acquired, and the EEG signals are pre-processed.
The EEG signals use the eeg signal acquisition device of 64 leads, and the EEG signals of acquisition include epilepsy sufferer
Stage of attack, intermittent phase and healthy phase EEG signals.
The pretreatment includes filtering processing, rejects bad track and the useless Electrode treatment of removal, superposed average processing, baseline
A variety of artefact processing such as correction and weight reference process, independent component analysis removal eye electricity, electrocardio, myoelectricity, finally obtain noiseless
EEG signals.
The filtering processing uses butterworth high pass filter and low-pass filter；
The noiseless EEG signals are stored, the data of txt format are stored as.
Step 2: carrying out time frequency analysis to the noiseless EEG signals extracts time-frequency characteristics, meanwhile, extract noiseless
The statistics feature of EEG signals；
Specific step is as follows:
Step 201 carries out time frequency analysis with the method for Short Time Fourier Transform, extracts the time-frequency of noiseless EEG signals
Feature obtains γ (25~50Hz), β (12~25Hz), α (6~12Hz), θ (3~6Hz), every segment data in δ (0~3Hz)
Time-frequency characteristics；
Step 202, the average energy value for extracting noiseless EEG signals and energy scale variance are as statistics feature；
Step 203, the processing that the time-frequency characteristics and statistics feature reduce with dimension, to reduce method
Expense is prepared for input automatic identification device.Preferably, dimension-reduction treatment is carried out using Principal Component Analysis.
Step 3: building random forest identification model, knows random forest using the brain electrical feature that step 2 extracts
Other model is trained, and generates optimal random forest identification model；
Step 4: the Random Forest model after optimization to be carried out to epileptic's brain electricity of the different degrees of patient's condition on test set
The pattern-recognition of signal is tested.
The present invention has the advantages that
(1) feature extracting method provided by the invention, can simplify calculating process, improve the accuracy of EEG's Recognition
And actual effect.
(2) a kind of CRT technology method of epileptic's EEG signals of the present invention devises in a kind of ward ICU
Epileptic's EEG signal identification method of the different degrees of patient's condition.Due to the highly-parallel of random forest recognition methods, improve
Operational efficiency, realizes the function of Computer Automatic Recognition epileptic's EEG signals, takes time and effort for medical worker
Diagnosis provides technical support.
(3) a kind of CRT technology method of epileptic's EEG signals of the present invention, random forest recognition methods exist
A large amount of hyper parameter can be generated in training process, the experience for only relying on programmer's long-term adjustment parameter carries out these lifes of manual debugging
At parameter be to be difficult to calculate out the optimal parameter of identification model.Present invention introduces grid search optimization methods can pass through institute
It states computer and repeats the optimal combination that filtration parameter carrys out acceleration search parameter in the form of variable step size.Accelerate Random Forest model
Operational efficiency so that train come random forest identification model be optimal effect.The present invention improve based on optimization with
The CRT technology method of machine forest handles the accuracys rate of epileptic's EEG signals in the ward ICU, for three kinds
The pattern-recognition accuracy rate of the epilepsy state of an illness of the different patient's condition can reach 96% or more.
(4) a kind of CRT technology method of epileptic's EEG signals of the present invention, is optimized using grid search
During method optimizes random forest, while 10 retransposings are carried out to experiment EEG signals data using computer and are tested
Card, avoids that penalty is excessively high and the generation of overfitting state, realizes grid search optimization method and identifies to random forest
While model optimization, the performance of the mode identification method is improved, there are new EEG signals to be input to the mould for realization is subsequent
Also computer identification can be accurately carried out in type, carry out auxiliary support and decision during medical diagnosis to doctor.
Detailed description of the invention
Fig. 1 is the CRT technology method flow diagram of epileptic's EEG signals of the invention；
Fig. 2 is that the present invention is based on the building process schematic diagrames of optimization Random Forest model；
Fig. 3 is the hyper parameter optimizing flow chart the present invention is based on the Random Forest model of improved trellis search method；
Fig. 4 is that the present invention is based on the accurate of the Random Forest model of grid search optimization method and original random forests algorithm
Property index correlation curve；
Fig. 5 is the accuracy result that the present invention carries out 10 retransposings verifying generation during model training using computer；
Fig. 6 is that the present invention is based on the ROC curve of the random forest of improved grid search optimization method and AUC value indexs
Evaluation result.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawings and examples.
The present invention provides a kind of CRT technology method of epileptic's EEG signals, and the method passes through pair first
In the ward ICU epileptic's (epileptic attack phase, epilepsy intermittent phase, healthy phase) of the different degrees of patient's condition carry out long period,
The acquisition of multichannel brain electric signal samples, and the EEG signals of different degrees of patient's condition epileptic are marked and (are labelled)；
The operation such as pretreatment operation and the extraction of brain electrical feature is carried out to EEG signals.The random forest based on machine learning techniques is constructed to know
Other model optimizes processing by the parameter that grid search optimization method generates random forest identification model, meanwhile, it will be pre-
Treated EEG signals imported into building and the random forest identification model that has optimized in, carry out identifying processing.The present invention mentions
The random forest identification model of the optimization of confession is based on machine learning techniques, obtains EEG signals using decision tree related operation layer by layer
The depth of different patient's condition modes is abstract, realizes and carries out automatic identification to the mode of different degrees of patient's condition epileptic EEG signals,
Technical support is provided for follow-up diagnosis of the doctor in the ward ICU to brain electricity.
The present invention provides a kind of CRT technology method of epileptic's EEG signals, and this method is led to for clinic
The EEG signals for identifying epileptic's stage of attack, intermittent phase and healthy phase are crossed, the random forest identification model of optimization are established, by institute
The random forest identification model for the optimization stated is applied in the original EEG signals of epilepsy being recorded in real time in the ward ICU, will know
The identification work of the parameter index auxiliary generated in not and support doctor and other medical workers to different patient's condition epileptics,
Avoid the inefficient work in the stay-at-home ward of the long-time of the medical worker in the ward ICU.
A kind of CRT technology method of epileptic's EEG signals provided by the invention, process as shown in Figure 1, tool
Body includes the following steps:
Step 1: the EEG signals of acquisition epileptic, and the EEG signals of acquisition are pre-processed, obtain noiseless
EEG signals；
Using the acquisition device of the EEG signals of 64 leads, the 10-20 head provided using U.S. clinical nervous physiology association
Skin electrode site distribution map, the eeg signal acquisition object be epilepsy sufferer, acquisition time through epilepsy sufferer stage of attack,
Intermittent phase and healthy phase usually pick up from the epilepsy sufferer in the ward ICU for the EEG signals of epilepsy sufferer stage of attack.
In order to reach EEG signals big data, can epileptic EEG Signal to the stage of attack of the neurology department of various big hospital into
Row is collected, while the EEG signals for acquiring a large amount of Healthy Peoples compare and analyze.Under normal circumstances, the EEG signals of acquisition are main
EEG signals including three types: epilepsy sufferer stage of attack, intermittent phase and healthy phase.
Acquisition device mainly include two work stations, liquid crystal display, amplifier, Ag-Cl electrode brain electricity cap, conductive paste,
Socket, plug corresponding with socket and ground line.Plug includes the pin that several correspondences are plugged on interface；Carry out brain electricity
When signal acquisition, the wiring of amplifier will be connected with computer, while the ground line of brain electricity cap is connected on any one pin,
To connect the circuit where resistance to form closed circuit.In this way, the EEG signals of sensor apparatus then can be with after resistance
It is converted into voltage value.Wherein the system configuration information of two work stations mainly include double-core CPU, memory 4GB, video card 512M and
WinXP professional system.The configuration information and parameter of amplifier include the A/D conversion resolution of 24Bit, input resistance
Anti- to be greater than 10GOhms, amplifier has two kinds of acquisition modes of AC and DC, and the sensitivity of AC mode is 3nV/bit, the spirit of DC mode
Sensitivity is 24nV/bit.
The EEG signals of the epilepsy sufferer of acquisition are pre-processed with the tool box of eeglab this matlab, it is pre- to locate
The method of reason includes: filtering, rejects bad track, superposed average, baseline correction, refers to again, and independent component analysis removes multiple types
Artefact；Finally obtained pretreated EEG signals data are saved.
The filtering processing carries out continuous EEG signals using butterworth high pass filter and low-pass filter
Filtering processing；The threshold value of butterworth high pass filter and low-pass filter is respectively 0.1-55Hz.
Pretreated noiseless EEG signals are stored, the data of txt format are stored as.The present invention uses Bonn
The detailed description that the EEG signals of epilepsy sufferer are followed the steps below as sample data set disclosed in university.
Second step, for noiseless EEG signals, with Short Time Fourier Transform method to noiseless EEG signals into
Row time-frequency characteristics extract, meanwhile, extract the statistics feature of noiseless EEG signals.
Using Short Time Fourier Transform method, obtaining effective frequency respectively is the θ wave of 4~7Hz, the α wave of 8~13Hz, 14
The time-frequency characteristics of the intermediate fast wave of~17Hz, the β wave of 18~30Hz；It is obtained according to the average energy of EEG signals, energy variance
The statistics feature of EEG signals.
It is some to study the automatic nonstationary EEG activity problem monitored and detect for having solved epilepsy.It is most classic to be
Wei Erqi spectral analysis method is introduced into epileptic attack signature analysis.In During Seizures, by using time domain approach pair
The EEG signals of epileptic are analyzed.In addition, the popular spectrum analysis technique based on Fourier is commonly used in analysis brain
Electric signal is in a frequency domain.Fourier transform not can be well solved the brain electricity analytical problem of epileptic.Secondly, this method is
It is very time-consuming.The present invention carries out time frequency analysis to EEG signals with the method for Short Time Fourier Transform, by adjusting difference
Time window, the shortcomings that avoiding Fourier transform.Both available brain is electric for Short Time Fourier Transform method proposed by the present invention
The temporal signatures of signal can obtain the frequency domain character of signal again, greatly excavate the substantive characteristics of EEG signals data.
Specifically:
Step 201 carries out time frequency analysis with the method for Short Time Fourier Transform, extracts the frequency domain character of EEG signals,
Obtain the data characteristics of α wave, β wave, δ wave, every section of θ wave and γ wave；And frequency spectrum is divided according to the wave band of EEG signals, it is raw
At the time-frequency figure of EEG signals.
Step 202, the time-frequency figure according to EEG signals extract the average energy value and energy scale variance conduct of EEG signals
Statistics feature；
Its average energy value and energy scale variance, specific statistics feature are extracted to EEG signals based on computer program
Information is as follows:
Wherein, data set S, F/N and O/Z respectively indicates epileptic stage of attack, intermittent phase and healthy phase.
The time-frequency characteristics and statistics feature of said extracted are all the EEG signals of high dimensional feature, it is therefore desirable to epileptics
The time-frequency characteristics and statistics feature of the EEG signals of trouble carry out reducing dimension processing.The present invention by principal component analytical method into
Row reduces the processing of dimension, to reduce the expense of method, construction feature collection is that input automatic identification device is prepared.
The dimension-reduction treatment principle is as follows: because the maximum variance of data gives the most important information of data.Conversion
When coordinate system, using the maximum direction of variance as change in coordinate axis direction, what first new reference axis selected is variance in initial data
Maximum method, the direction that second new reference axis selected is orthogonal with first new reference axis and variance time is big.Repeating should
Process, number of repetition are the intrinsic dimensionality of initial data.
Step 3: building random forest identification model, largely instructs the brain electrical feature extracted using computer
Practice, generates optimal random forest identification model.
Random forest identification model is a kind of effective ensemble machine learning method combined by decision tree, and random forest is known
Other model is suitable for high dimensional data and the speed of service is fast.But a large amount of hyper parameter can be generated in operational process, in order to obtain
The precision of higher identification epileptic EEG Signal, needs the parameter to random forest identification model to optimize.Currently, to random
The method that parameter in Forest mapping model is in optimized selection is relatively fewer, usually carries out artificial parameter choosing by experience
It selects.Influence of the decision tree quantity k to random forest identification model performance is bigger especially in random forest identification model, and
For the data of different classifications, the k value when the performance of random forest identification model is optimal is different.Only
The parameter that selected random forest identification model is only carried out by experience is not typically available the random forest identification mould of best performance
Type.The present invention utilizes improved trellis search method, to decision tree quantity k, the disruptive features number in random forest identification model
M, the minimum sample number of leaf node and punishment parameter carry out parameter optimization, while utilizing the side of 10 retransposings verifying in machine learning
Method more efficiently avoids the overfitting problem of the random forest identification model trained.By original insane in the ward ICU
The simulation result of epilepsy EEG signals shows that the parameter that the present invention is optimized using trellis search method can be such that random forest knows
The recognition effect of other model is significantly improved.
Specifically, the generation step of the optimal random forest identification model is as follows:
The brain electrical feature extracted in step 1 is normalized in step 301；
Brain electrical feature after normalization is divided into three groups by step 302, respectively training set, test set and for preventing
The verifying collection of random forest identification model over-fitting；
Step 303, the framing that time synchronization is carried out to the brain electrical feature of the epileptic of the different degrees of patient's condition, obtain lead more
The brain electrical feature signal of connection；
Step 304, being marked to the EEG signals of epileptic stage of attack, intermittent phase and healthy phase: stage of attack
" 1 ", intermittent phase " 0 ", healthy phase " -1 "；
Step 305, process as shown in Figure 2 randomly select a bootstrap training sample from the data M of training set
Mi, and be to carry out k extraction with putting back to, to obtain the training set M generated at random*={ M1,M2,…,Mk, i=
1,2,…,k；
Step 306, the training set M to be generated at random in above-mentioned 305 step*For training data, k decision tree T is establishedj, j
=1,2 ..., k randomly choose the subset comprising k feature, from institute from the characteristic set of each node of decision tree
The differentiation that m optimal characteristics are randomly selected in subset is stated as identification feature, k here controls the size of random degree；
Enter step 307；
Step 307 recursively repeats step 306 above-mentioned steps by each terminal node to decision tree, will be random gloomy
Woods decision tree is grown to enhancing binding data, until the decision tree can be accurately to training set M*It is identified, is reached simultaneously
Minimum node size.
During model training, using identification regression tree CART method split vertexes, gini index GINI value conduct
The foundation of split vertexes.Training set M*Containing the different features of k, then this training set M*Gini index are as follows:
Wherein piFor the probability of occurrence of ith feature.{M1,M2,…,MkCorresponding characteristic is respectively { n1,n2,…,
nk, then the gini index divided are as follows:
Wherein, n indicates n1+n2+…+nk, i=1,2 ..., k.GINI(Mi) indicate sample MiGINI index.
Step 308, all decision trees of set, for an input sample Mi, k decision tree have k recognition result, at random
Forest inherits all identification voting results；
Step 309 is predicted that the most identification of number of votes is output on new node, enters step 310；
Random forest method is as follows:
Step 310 optimizes Random Forest model using the optimization method of grid search；
Grid search optimization method, which refers to, carries out gridding for variable region, traverses all mesh points, solves and meet about
The target function value of beam condition, selects optimum value.It traverses all parameters on grid and needs a large amount of training times, in order to improve instruction
Practice speed, the invention proposes a kind of random forest parameter optimization methods based on improved grid search.
Step 311, first coarse search hyper parameter: decision tree quantity k, disruptive features number m, the minimum sample number of leaf node with
And punishment parameter；Enter step 312；
Firstly, it is 10 that step-size in search, which is arranged, in a big way with big step-length grid division, carries out coarse search and select one
Suboptimum parameter.
Then step 312 utilizes the long grid division of small step near an optimized parameter, make grid dividing more crypto set,
It scans for selecting Quadratic Optimum parameter again；The small step-length (when secondary small step-length refers to the half of last step-length) is
Relative quantity for big step-length, it is smaller, generally choose the step-length of 5~10 ranges.
Step 313 repeats step 312, the long grid division of small step and search parameter is utilized near Quadratic Optimum parameter, directly
It is less than given value to grid spacing or objective function variable quantity；
In order to improve the pattern-recognition performance of computer program, need to consider simultaneously single decision tree recognition correct rate and
The diversity of decision tree, however there is also certain relationships between the two.The present invention is for decision tree in random forest method
The characteristics of number k and candidate disruptive features number m is discrete value carries out parameter optimization using trellis search method.The present invention proposes
The target function value of the random forest parameter optimization based on improved grid search select the training data i.e. bag do not drawn
The identification accuracy of outer data estimation.Due to randomness of the random forest in building process, identify that accuracy may be one
Determine fluctuation in range, therefore to reduce the uncertain influence generated to parameter selection, present invention choosing when seeking identification accuracy
With the average value of multiple Random Forest models identification accuracy.
Step 314, if there is multiple groups parameter is optimal the method for random forest, then selected from these group of parameter
So that the smallest that group of parameter of penalty is optimized parameter；
Step 315 carries out precise search to the parameter that random forest method obtains, and reduces search range, i.e. reduction grid
The step-size in search of chess game optimization method, step-length is traditionally arranged to be 0.1 at this time, can also be according to the reality of random forest identification model
Situation, adjusting step size, so that punishment parameter minimum is optimal parameter group；Specific step is as follows:
It determines the range of decision tree quantity k and disruptive features number m, step-length is set, in decision tree quantity k and disruptive features number
Two-dimensional grid is established on m coordinate system, grid node is exactly the parameter pair of corresponding decision tree quantity k and disruptive features number m；
Random forest identification model is constructed to each group of parameter on grid node, and utilizes the outer data assessment identification of bag
Accuracy；
The selection identification highest parameter k and m of accuracy is exported optimal if identification accuracy or step-length are met the requirements
Parameter and recognition accuracy；Otherwise, step-length is reduced, precise search is continued.It is above-mentioned based on the random of improved trellis search method
Forest parameters optimizing flow chart is as shown in Figure 3.
Step 316 is obtained to the parameter after random forest progress grid search optimization, uses 10 retransposings with computer is crossed
Verifying goes to identify the accuracy of the random forest identification model；Specimen sample is carried out for given training set, produces 10
Different subsets, then trained from each subset and carry out the random forest identification submodel for belonging to the subset, pass through in this way
Different subsets, which train the random forest identification submodel come, has bigger difference, to be effectively prevented from random forest mould
The overfitting problem of type.It is simultaneously to obtain more preferably random forest identification model, it is also desirable to which each random forest identifies submodel
Cannot be too poor, if the subset differentiated is excessively more, can only be trained by fraction data, such model is not enough into
Random forest preferably identifies submodel, thus the present invention consider to use by it is a kind of it is mutually overlapping in a manner of training set is adopted
Sample.
Step 317, concept transfer number, then upset data at random, select a completely new verifying collection to carry out random forest knowledge
Other normatron identification；
Step 318 repeats step 317, until cross validation accuracy highest；
Step 319, number of nodes at this time are considered as optimal number of nodes；The parameter trained is random forest identification model
Optimized parameter, model determine；
Step 4: the random forest identification model after optimization to be carried out to the epileptic of the different degrees of patient's condition on test set
The pattern-recognition of EEG signals is tested；Random forest this computer automatic identification method is judged according to a variety of test indexs
Performance；Assess the generalization ability of Random Forest model.
Firstly, determining all parameters and value of random forest CRT technology algorithm；
Then, coarse search is carried out to all parameters that random forest identification model generates by grid search optimization algorithm,
Limit the subrange of parameter；The secondary optimization that small step-length is carried out on the basis of this subrange, obtains Random Forest model
Optimized parameter；
Finally, optimized parameter is input in random forest CRT technology model, optimal identification model is generated,
Computer identification is carried out for the EEG signals to epileptic, so that carrying out the diagnosis of the epilepsy patient's condition for doctor provides technical support
And auxiliary.
Experimental result
The eeg data of use derives from Bonn, Germany epilepsy laboratory.The data are divided into O, Z, F, N, S totally 5 groups of brain telecommunications
Number, it include 100 samples in every group of data, subject is 5 people, and each sample includes 4097 sampled points, and signal record is adopted
With standard 10-20 system, sample frequency 173.61Hz, sampling time 23.6s.The specifying information of 5 groups of eeg data collection is such as
Under.
The present invention is divided into O/Z, F/N, S this 3 group data set and identifies to EEG signals.Wherein data set O/Z is health
People is in the EEG signals of scalp surface under awake eyes-open state, and data set F/N does not break out for epileptic to cause a disease in interphase
The encephalic EEG signal of stove inner region, data set S are the encephalic EEG signals that epileptic is in stove inner region of causing a disease stage of attack,
It is denoted as healthy phase, breaking-out intermittent phase and EEG signals stage of attack respectively.
Experimental result is assessed using accuracy ACC curve index.To experimental result using true positive rate and vacation sun
Property rate index is assessed, and the evaluations of estimate such as the Receiver operating curve ROC of generation and AUC value index is according to obtained identification
The assessment of the random forest method of result optimizing.
This experiment uses altogether four evaluation indexes, accuracy including epileptic EEG Signal mode identification method, true
Positive probability (True positive rate, TPR) and false positive probability (False positive rate, FPR), false positive
Probability is horizontal axis, and true positives probability is coordinate diagram composed by the longitudinal axis, thus the Receiver operating curve (ROC) that generates and
Two evaluation indexes of area AUC value under the region ROC.
Wherein TP, FP, TN, FN respectively indicate kidney-Yang number, false positive number, Kidney-Yin number and false negative number.
Accuracy result such as Fig. 4 of the random forest recognition methods of front and back is improved, it is random after being optimized by grid search
The accuracy rate of forest model has reached 96.7%, improves nearly 10 percentage point than random forest method accuracy rate is used alone.
To training set do 10 retransposings verifying model accuracy variation as shown in figure 5, application enhancements random forest optimization method it is another
The result of two evaluation indexes ROC and AUC are as shown in fig. 6, wherein AUC achieves 99% high-accuracy.
Improved random forest Computer Identification is applied to the automatic knowledge of the EEG signals of epileptic by the present invention
During not, the identification of different degrees of patient's condition epileptic EEG signals is realized, while devising the optimization method of grid search
Random forest recognition methods is optimized, promotion and the program operation process of computer operation calculated result accuracy rate are realized
Acceleration, assist vast medical worker for epilepsy sufferer diagnosis during technical support and service are provided, it is latent to having
Incidence probability is eliminated and controlled in the epileptic of morbidity.
Claims (5)
1. a kind of pretreatment and recognition methods of epileptic's EEG signals based on artificial intelligence, it is characterised in that: including such as
Lower step,
Step 1: the EEG signals of epileptic are acquired, and the EEG signals are pre-processed；
The pretreatment includes filtering processing, rejects bad track and the useless Electrode treatment of removal, superposed average processing, baseline correction
And weight reference process, independent component analysis removal eye electricity, electrocardio and myoelectricity processing, finally obtain noiseless EEG signals；
Step 2: carrying out time frequency analysis to the noiseless EEG signals extracts time-frequency characteristics, meanwhile, extract noiseless brain electricity
The statistics feature of signal；
Specific step is as follows:
Step 201 carries out time frequency analysis with the method for Short Time Fourier Transform, and the time-frequency for extracting noiseless EEG signals is special
Sign, obtain γ (25~50Hz), β (12~25Hz), α (6~12Hz), θ (3~6Hz), in δ (0~3Hz) every segment data when
Frequency feature；
Step 202, the average energy value for extracting noiseless EEG signals and energy scale variance are as statistics feature；
Step 203, the processing that the time-frequency characteristics and statistics feature are carried out with principal component analysis reduction characteristic dimension, thus
The expense in computer operational process is reduced, is prepared for input automatic identification device；
Step 3: building random forest identification model, identifies mould to random forest using the brain electrical feature that step 2 extracts
Type is trained, and generates optimal random forest identification model；
Step 4: the Random Forest model after optimization to be carried out to epileptic's EEG signals of the different degrees of patient's condition on test set
Pattern-recognition test.
2. pretreatment and the identification side of a kind of epileptic's EEG signals based on artificial intelligence according to claim 1
Method, it is characterised in that: the filtering processing uses butterworth high pass filter and low-pass filter.
3. pretreatment and the identification side of a kind of epileptic's EEG signals based on artificial intelligence according to claim 1
Method, it is characterised in that: specific step is as follows for step 3:
Brain electrical feature is normalized in step 301；
Brain electrical feature after normalization is divided into three groups by step 302, respectively training set, test set and random for preventing
The verifying collection of Forest mapping model over-fitting；
Step 303, the framing that time synchronization is carried out to the brain electrical feature of the epileptic of the different degrees of patient's condition, obtain multi-lead
Brain electrical feature signal；
Step 304, being marked to the EEG signals of epileptic stage of attack, intermittent phase and healthy phase: stage of attack " 1 ",
It has a rest the phase " 0 ", healthy phase " -1 "；
Step 305 randomly selects a training sample M from training seti, and be to carry out k extraction with putting back to, to obtain
The training set M generated at random to one*={ M1,M2,…,Mk, i=1,2 ..., k；
Step 306, the training set M to be generated at random in above-mentioned 305 step*For training data, k decision tree T is establishedj, j=1,
2 ..., k randomly choose the subset comprising k feature from the characteristic set of each node of decision tree；
Step 307 recursively repeats step 306 above-mentioned steps by each terminal node to decision tree, and random forest is determined
Plan tree is grown to enhancing binding data, until the decision tree can be accurately to training set M*It is identified, while reaching minimum
Node size；
Step 308, all decision trees of set, for an input sample Mi, k decision tree have k recognition result, random forest
Inherit all identification voting results；
Step 309 is predicted that the most identification of number of votes is output on new node, enters step 310；
Step 310 optimizes Random Forest model using the optimization method of grid search；
Step 311 is obtained to the parameter after random forest progress grid search optimization, is verified with computer is crossed using 10 retransposings
Go to identify the accuracy of the random forest identification model；
Step 312, concept transfer number, then upset data at random, it selects a completely new verifying collection to carry out random forest and identifies mould
The identification of type computer；
Step 313 repeats step 312, until cross validation accuracy highest；
Step 314, number of nodes at this time are considered optimal number of nodes；The parameter trained is the optimal of random forest identification model
Parameter, model determine.
4. a kind of CRT technology method of epileptic's EEG signals according to claim 3, it is characterised in that:
The grid search includes coarse search and precise search two parts；
(A) the coarse search hyper parameter described in specifically:
Firstly, setting step-size in search is 10 with big step-length grid division, carries out coarse search and select optimized parameter；Once most
Excellent parameter nearby utilizes the long grid division of small step, makes grid dividing more crypto set, scans for selecting Quadratic Optimum ginseng again
Number；Grid dividing and search ... ... are carried out again near Quadratic Optimum parameter, until grid spacing or objective function variable quantity
Less than given value；If there is multiple groups parameter is optimal the method for random forest, then selected from these group of parameter so that
The smallest that group of parameter of penalty is optimized parameter；
(B) precise search specifically: setting step-length is 0.1, according to the actual conditions of Random Forest model, adjusting step size,
So that punishment parameter minimum is optimal parameter group；Using the hyper parameter in optimal parameter group as final hyper parameter.
5. a kind of CRT technology method of epileptic's EEG signals according to claim 1 or 2, feature exist
In: the hyper parameter includes decision tree quantity k, disruptive features number m, the minimum sample number of leaf node and punishment parameter.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811001414 | 2018-08-30 | ||
| CN2018110014140 | 2018-08-30 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109480833A true CN109480833A (en) | 2019-03-19 |
Family
ID=65711228
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811569263.9A Pending CN109480833A (en) | 2018-08-30 | 2018-12-21 | The pretreatment and recognition methods of epileptic's EEG signals based on artificial intelligence |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109480833A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109933539A (en) * | 2019-04-15 | 2019-06-25 | 燕山大学 | A kind of Software Defects Predict Methods based on principal component analysis and combination sampling |
| CN109998530A (en) * | 2019-04-15 | 2019-07-12 | 杭州妞诺科技有限公司 | Portable brain pyroelectric monitor system based on VR glasses |
| CN110338786A (en) * | 2019-06-28 | 2019-10-18 | 北京师范大学 | A kind of identification of epileptiform discharges and classification method, system, device and medium |
| CN111000556A (en) * | 2019-11-29 | 2020-04-14 | 上海师范大学 | Emotion recognition method based on deep fuzzy forest |
| CN111105305A (en) * | 2019-12-06 | 2020-05-05 | 安徽海汇金融投资集团有限公司 | Machine learning-based receivable and receivable cash cashing risk control method and system |
| CN112057068A (en) * | 2020-08-27 | 2020-12-11 | 武汉工程大学 | Epilepsia pathological data classification method and device and storage medium |
| CN112336310A (en) * | 2020-11-04 | 2021-02-09 | 吾征智能技术(北京)有限公司 | Heart disease diagnosis system based on FCBF and SVM fusion |
| CN112656431A (en) * | 2020-12-15 | 2021-04-16 | 中国科学院深圳先进技术研究院 | Electroencephalogram-based attention recognition method and device, terminal equipment and storage medium |
| CN112741638A (en) * | 2021-01-05 | 2021-05-04 | 北京航空航天大学 | Medical diagnosis auxiliary system based on EEG signal |
| CN114224300A (en) * | 2022-02-23 | 2022-03-25 | 广东工业大学 | Epilepsy classification detection system and method based on three-dimensional quaternion graph convolution neural network |
| CN114376522A (en) * | 2021-12-29 | 2022-04-22 | 四川大学华西医院 | Method for constructing computer recognition model for recognizing juvenile myoclonus epilepsy |
| WO2022135449A1 (en) * | 2020-12-22 | 2022-06-30 | 北京航空航天大学 | Interictal epileptiform discharge activity detection apparatus and method for epileptic patient |
| CN116849684A (en) * | 2023-08-29 | 2023-10-10 | 苏州唯理创新科技有限公司 | Signal source space positioning method of multichannel sEMG based on independent component analysis |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2002082999A1 (en) * | 2001-04-10 | 2002-10-24 | Battelle Memorial Institute | Image analysis system and method for discriminating movements of an individual |
| CN102058413A (en) * | 2010-12-03 | 2011-05-18 | 上海交通大学 | Method for detecting EEG (electroencephalogram) alertness based on continuous wavelet transform |
| CN104887224A (en) * | 2015-05-29 | 2015-09-09 | 北京航空航天大学 | Epileptic feature extraction and automatic identification method based on electroencephalogram signal |
| CN108420429A (en) * | 2018-04-19 | 2018-08-21 | 北京工业大学 | A kind of brain electricity epilepsy automatic identifying method based on the fusion of various visual angles depth characteristic |
-
2018
- 2018-12-21 CN CN201811569263.9A patent/CN109480833A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2002082999A1 (en) * | 2001-04-10 | 2002-10-24 | Battelle Memorial Institute | Image analysis system and method for discriminating movements of an individual |
| CN102058413A (en) * | 2010-12-03 | 2011-05-18 | 上海交通大学 | Method for detecting EEG (electroencephalogram) alertness based on continuous wavelet transform |
| CN104887224A (en) * | 2015-05-29 | 2015-09-09 | 北京航空航天大学 | Epileptic feature extraction and automatic identification method based on electroencephalogram signal |
| CN108420429A (en) * | 2018-04-19 | 2018-08-21 | 北京工业大学 | A kind of brain electricity epilepsy automatic identifying method based on the fusion of various visual angles depth characteristic |
Non-Patent Citations (3)
| Title |
|---|
| 张雨烟等: "基于非负矩阵分解的癫痫脑电自动检测", 《吉林大学学报》 * |
| 李冬梅等: "基于PCA多导联的癫痫脑电信号分类及致痫灶定位研究", 《生物医学工程研究》 * |
| 温博文等: "基于改进网格搜索算法的随机森林参数优化", 《计算机工程与应用》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109998530A (en) * | 2019-04-15 | 2019-07-12 | 杭州妞诺科技有限公司 | Portable brain pyroelectric monitor system based on VR glasses |
| CN109933539A (en) * | 2019-04-15 | 2019-06-25 | 燕山大学 | A kind of Software Defects Predict Methods based on principal component analysis and combination sampling |
| CN110338786A (en) * | 2019-06-28 | 2019-10-18 | 北京师范大学 | A kind of identification of epileptiform discharges and classification method, system, device and medium |
| CN110338786B (en) * | 2019-06-28 | 2020-10-02 | 北京师范大学 | Epileptic discharge identification and classification method, system, device and medium |
| CN111000556A (en) * | 2019-11-29 | 2020-04-14 | 上海师范大学 | Emotion recognition method based on deep fuzzy forest |
| CN111105305A (en) * | 2019-12-06 | 2020-05-05 | 安徽海汇金融投资集团有限公司 | Machine learning-based receivable and receivable cash cashing risk control method and system |
| CN112057068A (en) * | 2020-08-27 | 2020-12-11 | 武汉工程大学 | Epilepsia pathological data classification method and device and storage medium |
| CN112336310B (en) * | 2020-11-04 | 2024-03-08 | 吾征智能技术(北京)有限公司 | FCBF and SVM fusion-based heart disease diagnosis system |
| CN112336310A (en) * | 2020-11-04 | 2021-02-09 | 吾征智能技术(北京)有限公司 | Heart disease diagnosis system based on FCBF and SVM fusion |
| CN112656431A (en) * | 2020-12-15 | 2021-04-16 | 中国科学院深圳先进技术研究院 | Electroencephalogram-based attention recognition method and device, terminal equipment and storage medium |
| WO2022135449A1 (en) * | 2020-12-22 | 2022-06-30 | 北京航空航天大学 | Interictal epileptiform discharge activity detection apparatus and method for epileptic patient |
| CN112741638B (en) * | 2021-01-05 | 2022-04-08 | 北京航空航天大学 | Medical diagnosis auxiliary system based on EEG signal |
| CN112741638A (en) * | 2021-01-05 | 2021-05-04 | 北京航空航天大学 | Medical diagnosis auxiliary system based on EEG signal |
| CN114376522A (en) * | 2021-12-29 | 2022-04-22 | 四川大学华西医院 | Method for constructing computer recognition model for recognizing juvenile myoclonus epilepsy |
| CN114376522B (en) * | 2021-12-29 | 2023-09-05 | 四川大学华西医院 | Method for constructing computer identification model for identifying juvenile myoclonus epilepsy |
| CN114224300A (en) * | 2022-02-23 | 2022-03-25 | 广东工业大学 | Epilepsy classification detection system and method based on three-dimensional quaternion graph convolution neural network |
| CN114224300B (en) * | 2022-02-23 | 2022-07-12 | 广东工业大学 | Epilepsy classification detection system and method based on three-dimensional quaternion graph convolution neural network |
| CN116849684A (en) * | 2023-08-29 | 2023-10-10 | 苏州唯理创新科技有限公司 | Signal source space positioning method of multichannel sEMG based on independent component analysis |
| CN116849684B (en) * | 2023-08-29 | 2023-11-03 | 苏州唯理创新科技有限公司 | Signal source space positioning method of multichannel sEMG based on independent component analysis |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109480833A (en) | The pretreatment and recognition methods of epileptic's EEG signals based on artificial intelligence | |
| CN109645990A (en) | A kind of CRT technology method of epileptic's EEG signals | |
| US20200367800A1 (en) | Method for identifying driving fatigue based on cnn-lstm deep learning model | |
| CN105654063B (en) | Mental imagery brain power mode recognition methods based on the optimization of artificial bee colony time and frequency parameter | |
| US6988056B2 (en) | Signal interpretation engine | |
| WO2019161611A1 (en) | Ecg information processing method and ecg workstation | |
| CN108511055B (en) | Ventricular premature beat recognition system and method based on classifier fusion and diagnosis rules | |
| CN104586387A (en) | Method for extracting and fusing time, frequency and space domain multi-parameter electroencephalogram characters | |
| Yu et al. | Epileptic seizure prediction using deep neural networks via transfer learning and multi-feature fusion | |
| CN115081486B (en) | System and method for positioning epileptic focus by using intracranial brain electrical network in early stage of epileptic seizure | |
| CN113274037B (en) | Method, system and equipment for generating dynamic brain function network | |
| CN113208593A (en) | Multi-modal physiological signal emotion classification method based on correlation dynamic fusion | |
| CN112674782B (en) | Device and method for detecting epileptic-like electrical activity of epileptic during inter-seizure period | |
| CN115563484A (en) | Street greening quality detection method based on physiological awakening identification | |
| CN114532993A (en) | Automatic detection method for electroencephalogram high-frequency oscillation signals of epileptic | |
| CN113158964A (en) | Sleep staging method based on residual learning and multi-granularity feature fusion | |
| CN114159079B (en) | Multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model | |
| CN112698720B (en) | Movement imagery identification method based on mutual information feature extraction and multi-person fusion | |
| CN115414050A (en) | EEG brain network maximum clique detection method and system for realizing emotion recognition | |
| CN108078563A (en) | A kind of EEG signal analysis method of integrated classifier | |
| CN111931578A (en) | Electroencephalogram identification method based on minimum spanning tree and regional double-layer network | |
| Hou et al. | Epilepsy detection using random forest classification based on locally linear embedding algorithm | |
| CN113768464A (en) | Epileptic seizure time period detection system and method based on empirical mode decomposition | |
| CN111012337A (en) | Brain network and regularized discriminant analysis-based electroencephalogram analysis method | |
| CN111476295A (en) | GS-L SSVM-based multichannel physiological signal somatosensory gesture recognition method |
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: 20190319 |