KR20190111570A - A system of detecting epileptic seizure waveform based on coefficient in multi-frequency bands from electroencephalogram signals, using feature extraction method with probabilistic model and machine learning - Google Patents

Abstract

The present invention relates to a system for detecting an epilepsy seizure wave based on a multi-frequency band coefficient of an electroencephalogram signal using feature extraction with a probability model and machine learning, which exhibits accurate detection performance. The system of the present invention comprises: an electroencephalogram input unit for receiving an electroencephalogram signal; a signal preprocessing unit; a feature extracting unit for generating a feature vector; a codebook generating unit for generating a codebook; a classified model generation unit for generating at least one classified model; and a classification unit for classifying a seizure state.

Description

A system of detecting epileptic seizure waveform based on coefficient in multi-frequency bands from electroencephalogram signals, using feature extraction method with probabilistic model and machine learning}

The present invention uses an efficient feature extraction method, an optimal machine learning or a probability model based on the electroencephalogram signal obtained from the electroencephalogram measurement device to monitor the seizure wave of the electroencephalogram signal due to epilepsy (epilepsy). The present invention relates to a multi-frequency band coefficient-based epileptic seizure detection system for detecting and diagnosing epileptic seizure waves.

In addition, the present invention monitors the seizure wave of the electroencephalogram signal, provides epilepsy patient information on the timing and number of seizures, and detects seizure waves in real time, multi-frequency band coefficient based epilepsy seizure detection It's about the system.

Epilepsy is a disease that affects about 1-2% of the world's population. Seizures caused by epilepsy are unpredictable and have great difficulty managing in everyday life. In particular, in epilepsy patients, when wearing an oxygen respirator or a serious patient receiving medication, seizures due to epilepsy (hereinafter referred to as epileptic seizures) are difficult to control. In addition, this can be catastrophic, requiring a system for detecting or predicting epileptic seizures.

In particular, epileptic seizures are diagnosed by backtracking using a method of self-awareness by a patient due to an electroencephalogram, a history of the patient, a sudden faint, or the like. If epilepsy is suspected, the clinician or specialist can confirm manually by recording the long-term EEG signal and confirming the time and number of seizures. Epilepsy diagnosis method using such an electroencephalogram requires a lot of time and manpower. Thus, despite the usefulness of the test, there is a limit to its universal use in many hospitals.

Therefore, many conventional techniques for detecting seizure waves have been introduced for epilepsy diagnosis methods using long-time electroencephalogram signals [Patent Documents 1 and 2].

According to the conventional techniques, the feature extraction method and detection classification technique applied for seizure wave monitoring can improve the accuracy of the seizure detection system and increase the efficiency of the system. In addition, the window size required for dividing and using the EEG signal is one of important factors for the detection accuracy and the detection speed, and the performance of the epileptic seizure detection system may vary depending on the window size.

The prior art described in Patent Document 1 provides a method for predicting epileptic seizures before epileptic seizures in an algorithm for predicting epilepsy seizures based on an electroencephalogram signal. The epilepsy seizure prediction method according to the prior art may have a positive effect in practical application to a medical system or a healthcare system by predicting the occurrence of the seizure before the epileptic seizure occurs.

However, the epilepsy seizure prediction algorithm method according to the prior art does not consider the real-time prediction, and there is a limit in interpreting the EEG signal recorded for a long time. In addition, the window size, which is used by dividing the EEG signal, is not considered.

In addition, the prior art described in Patent Document 2 proposes an epilepsy patient monitoring technology using a mobile communication terminal, the epilepsy patient monitoring system according to the prior art monitors the patient's information in real time, detect and determine the emergency situation do. Therefore, it is an effective system that can be applied to the healthcare field in that it can take action quickly.

However, the epileptic seizure monitoring system according to the prior art does not consider optimization of feature extraction methods and detection classification models for maximizing the accuracy or efficiency of the system.

Korean Patent Registration Publication No. 10-0081717 Korean Patent Registration Publication No. 10-0084282

An object of the present invention is to solve the problems described above, from the electroencephalogram signal obtained from the user, by extracting the coefficient by decomposing the frequency band of the electroencephalogram signal, extracting the sophisticated feature value through an iterative calculation method, Performance evaluation based on various machine learning and probabilistic models, through which the epilepsy seizure detection system based on the multi-frequency band coefficient of the electroencephalogram signal can be detected accurately and quickly. .

In addition, the frequency band decomposing unit and the feature vector extracting unit included in the feature extraction module decompose the electroencephalogram into various multi-frequency bands to detect seizure waves more precisely. Calculate the wavelet coefficient, extract the feature value by taking advantage of the characteristic of the frequency variation of the wavelet coefficient for each frequency band, and use the multi-frequency band coefficient with the aim of improving the accuracy than the conventional method through iterative operation. Extraction method is included.

In addition, the seizure wave detection unit included in the classification module includes various machine learning methods for predicting and detecting seizure waves, seizure wave timings, seizure wave counts, pause-intermediate seizure units, etc. based on machine learning and probability models. And probabilistic models can be selectively and adaptively determined.

In addition, the seizure wave detection time and frequency output unit included in the output module is an output method for the user and doctor to analyze and observe the EEG signals measured for a long time in the daily life of the patient more efficiently and based on the recorded results. The seizure wave time and the number of times may be provided, and in addition, the real-time seizure wave result unit may receive information on real-time seizure wave detection according to the current state of the patient.

In order to achieve the above object, the present invention relates to a multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal, EEG input unit for receiving an electroencephalogram signal; A signal preprocessor for removing and dividing noise with respect to the electroencephalogram signal; A feature extractor for extracting a signal coefficient for each band by decomposing a plurality of frequency bands for the divided EEG signals, and extracting a feature for each band using the extracted band-specific signal coefficients; A codebook generator for generating a codebook using training data of an electroencephalogram signal, clustering according to each band of the feature vector of the training data and the type of seizure state, and generating a codebook by assigning an index to each cluster; A classification model generator for generating at least one classification model by machine learning using training data of an electroencephalogram signal, and generating at least one classification model by probability using the generated codebook; And a classification unit classifying the seizure state with respect to the input EEG signal by using the classification model by the at least one machine learning and the classification model by the at least one probability.

In addition, the present invention is a multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal, wherein the feature extractor, by applying a discrete wavelet transform to the electroencephalogram signal is decomposed into a plurality of frequency bands, according to the level for each band Generate a threshold to count signal coefficients for each band that is greater than the threshold for each level, and determine whether the level satisfies a predetermined specific level.If not, increase the level by 1 and generate and count the threshold for each level. When the level satisfies a specific level, the rate of change of counting values of all levels is calculated to extract features of each band.

In addition, the present invention is characterized in that in the multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal, the threshold for each level is calculated by the following [Equation 1].

[Equation 1]

T denotes a level, γ ₀ is a predetermined initial threshold, and L denotes a specific level.

In addition, the present invention, in the multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal, regression analysis of counting values of all levels to calculate the slope or rate of change and set the calculated slope or rate of change for each band of the band Characterized in that.

In addition, the present invention is a multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal, the classification model by the machine learning is SVM (Support vector machine), LDA (Linear discriminant analysis), QDA (Quadratic discriminant analysis), Any one of K-NN (K-nearest neighbor) and K-means is used, and the classification model based on the probability is characterized by using a Hidden Markov model (HMM).

In addition, the present invention is a multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal, the system outputs or records the detection result of the seizure state with respect to the input electroencephalogram signal, or the detection time and the frequency of detection of the seizure state It further comprises a result providing unit for providing.

As described above, according to the multi-frequency band coefficient-based epileptic seizure detection system of the electroencephalogram signal according to the present invention, it is possible to solve the problem of physical time and manpower in analyzing and interpreting the electroencephalogram signal recorded for a long time, It is possible to detect faster and more accurately than the existing technology, and the effect is to provide a more efficient seizure monitoring system.

In addition, according to the multi-frequency band coefficient-based epileptic seizure detection system of the electroencephalogram signal according to the present invention, by providing a function of outputting the timing and frequency of occurrence of the epileptic seizure wave, the detection result, to a specialist, guardian, patient, etc. It can be provided to the patient's health care system, through which the effect that can support effective information, such as seizure prevention, treatment methods, first aid methods and the like is obtained.

1 is a block diagram of a configuration of an entire system for implementing the present invention.
Figure 2 is a block diagram of the configuration of a multi-frequency band coefficient-based epilepsy seizure detection system of the electroencephalogram signal according to an embodiment of the present invention.
Figure 3 is a graph of the EEG signal output of a normal person according to an embodiment of the present invention.
Figure 4 is a graph of the electroencephalogram signal output of epilepsy patients according to an embodiment of the present invention.
FIG. 5 shows MoWavelet functions used for frequency band decomposition according to an embodiment of the present invention, (a) Daubechies third-order function, (b) Daubechies fourth-order function, (c) Coiflet third-order function, (d ) Coiflet fourth-order function, (e) Symlet fourth-order function, (f) is a graph showing a part of the seizure wave.
6 is a flowchart illustrating a feature extraction method according to an embodiment of the present invention.
7 is a graph illustrating a result of decomposing a frequency band using a discrete wavelet transform of an electroencephalogram signal of (a) a normal person and (b) an epilepsy patient according to an embodiment of the present invention.

DETAILED DESCRIPTION Hereinafter, specific contents for carrying out the present invention will be described with reference to the drawings.

In addition, in describing this invention, the same code | symbol is attached | subjected and the repeated description is abbreviate | omitted.

First, examples of the configuration of the entire system for implementing the present invention will be described with reference to FIG.

As shown in FIG. 1, a multi-frequency band coefficient-based epilepsy seizure detection system or method of an electroencephalogram signal according to the present invention receives an electroencephalogram signal 80 measured by an electroencephalogram measuring device 10 to detect an epileptic seizure wave. It may be implemented as a program system on the computer terminal 20. That is, the epileptic seizure wave method or system may be configured as a program and installed and executed in the computer terminal 20. The program installed in the computer terminal 20 may operate like one program system 30.

At this time, the electroencephalogram signal 80 is a signal measuring the electroencephalogram, and is a digital signal for the electroencephalogram waveform.

On the other hand, as another embodiment, the epileptic seizure wave detection system may be implemented as a program and operate in a general-purpose computer, in addition to an electronic circuit such as an ASIC (custom semiconductor). Alternatively, the present invention may be developed as a dedicated terminal 30 or a dedicated device 30 for exclusively processing a task of detecting an epileptic seizure wave by receiving an electroencephalogram signal. This is called an epileptic seizure detection system. Other possible forms may be implemented.

Next, a configuration of a multi-frequency band coefficient-based epilepsy seizure detection system of an electroencephalogram signal according to an embodiment of the present invention will be described in more detail with reference to FIG. 2.

As shown in FIG. 2, the epileptic seizure wave detection system 30 according to an embodiment of the present invention includes an electroencephalogram input unit 31 and an electroencephalogram signal, which receive an electroencephalogram signal 80 measured by the electroencephalogram measurement device 10. A preprocessing unit 32 for removing noise and dividing the signal data, a feature extractor 33 for extracting a signal coefficient by decomposing a frequency band, and clustering feature vectors A codebook generation unit 34 for assigning an index to each index, a classification model generation unit 35 for generating a classification model based on machine learning or probability, and a classification unit 36 for classifying whether a seizure wave is detected by a feature vector. do. In addition, it may be configured to further include a result providing unit 37 for outputting / recording or analyzing the detected results. In addition, it may be configured to further include a storage unit 38 for storing data.

First, the EEG input unit 31 receives an EEG signal measured by the EEG measurement device 10.

The electroencephalogram signal is a signal measured by the electroencephalogram measuring device 10 and is a signal non-invasively measured on the surface of the brain.

The EEG input unit 31 may receive and store the EEG signals recorded by measuring them for a short period or a long time, or may receive EEG signals measured in real time in real time. That is, the electroencephalogram measuring device 10 may measure and store the electroencephalogram signal for a short term or a long term, and the electroencephalogram input unit 31 may receive the stored short and long term electroencephalogram signals collectively. Alternatively, the electroencephalogram measuring apparatus 10 may measure the electroencephalogram in real time, and the electroencephalogram input unit 31 may receive the electroencephalogram signal measured in real time.

3 shows a graph of an EEG signal of a normal person measured by the EEG measuring apparatus 10, and FIG. 4 shows a graph of an EEG signal of an epilepsy patient measured by the EEG measuring apparatus 10.

On the other hand, the electroencephalogram measuring device 10 measures the electroencephalogram signal by using a non-invasive method for non-invasive measurement of the electroencephalogram on the surface of the brain.

In addition, the EEG measuring apparatus 10 may include an EEG signal sensor composed of a plurality of channels. The plurality of electroencephalogram signal sensors are individually attached to the surface of the user's head according to the international standard law for measuring the electroencephalogram signal, and the position of the attachment and the number of electrodes may vary according to the purpose of use.

Electrode placement method for electroencephalogram recording can be attached to the frontal head, central body, head, laryngeal head, temporal head, etc. according to 10-20 system of the International Association of Brain Science Standards. It may include. In addition, in order to acquire additional biometric information, the eye may include an eye electrocardiogram (EOG), an electrocardiogram (ECG), and an electromygoram (EMG).

Next, the signal preprocessor 32 removes noise with respect to the electroencephalogram signal and divides the signal data.

That is, the signal preprocessor 32 removes noise caused by power supply noise or movement from the electroencephalogram signal by applying a signal filter to the electroencephalogram signal. Signal filters used include low pass filters, high pass filters, band pass filters, notch filters, and the like.

In addition, the signal processor 32 divides the data of the electroencephalogram signal according to a predetermined window size. That is, since the EEG signal is a continuous signal waveform, it is divided into a predetermined size by the window size for analysis.

The divided EEG signal is decomposed in the frequency band by the feature extractor 33 and a feature vector is generated therefrom. That is, the electroencephalogram signal is a continuous signal, and when it is divided, a plurality of consecutive segmentation electroencephalogram signals are generated. For example, a divided signal is generated such as S ₁ , S ₂ ,..., S _n, and a feature vector is extracted for each divided signal.

Next, the feature extractor 33 extracts a signal coefficient for each band of the decomposed band by decomposing a frequency band with respect to the divided EEG signal, and extracts a feature using the extracted signal coefficient for each band. In particular, the feature extractor 33 extracts the features of the electroencephalogram signal at a plurality of levels, and generates a feature vector from the features for each level.

First, the feature extractor 33 decomposes the frequency band of the electroencephalogram signal in order to analyze various frequency bands.

Preferably, the feature extractor 33 applies a discrete wavelet transform (DWT) to the divided EEG signal to decompose the EEG signal into a plurality of frequency bands. In the analysis of EEG signals using the discrete wavelet transform, it is very important to select the level of decomposition and the appropriate mother wavelet function.

As shown in FIG. 5, the MoWavelet function used for frequency band decomposition includes wavelet functions such as Dobechies, Coiflet, Symlet, and Harar. do. Also, the order according to its resolution can be selected.

Referring to FIG. 5, FIG. 5 (f) is a graph of a portion of an electroencephalogram signal representing an epileptic seizure wave. In addition, by checking the correlation between the electroencephalogram signal of FIG. 5 (f) and various Mowaveret functions, which have been exemplified to FIGS. 5 (a)-(e), a mowaveret function to be used may be selected. In particular, in Figure 5, (a) Daubechies cubic function, (b) Daubechies cubic function, (c) Coiflet cubic function, (d) Coiflet cubic function, (e) Symlet quaternary function.

The feature extractor 33 extracts the detail coefficient and the approximation coefficient by decomposing the frequency band of the electroencephalogram signal by using the Mowavelet function.

In addition, the feature extractor 33 extracts a feature for each band for the extracted signal coefficient for each band. That is, if the number of decomposed bands is K, a feature (band-specific feature) is extracted from signal coefficients of K bands, respectively. Therefore, K band-specific features are extracted and a feature vector composed of these K features is generated.

In particular, the feature extractor 33 extracts the wavelet coefficients from the given n electroencephalogram signals, and performs a repetitive operation using the extracted wavelet coefficients to extract the feature vectors.

The specific method performed by the feature extraction unit 33 will be described below.

Next, the codebook generator 34 clusters the feature vectors of the training data (learning data) and assigns an index to each cluster.

Codebooks are used for classification models based on probability models.

Training data is data on the electroencephalogram signal and its consequences (epileptic seizure status). The feature vectors of the EEG signals of the training data are clustered according to the result. Clustering refers to clustering feature values having similar values with each other, and assigning a number (index) to the group to generate a codebook.

That is, the codebook generator 34 generates clusters by clustering according to each band of the feature vector of the training data and the type of seizure state, and generates a codebook with an index assigned to each cluster.

Next, the classification model generator 35 learns and generates a classification model for detecting an epileptic seizure wave. That is, the classification model generator 35 generates a classification model for detecting seizure waves and classifying seizure interpolators / seizure machines. At this time, the classification model includes both a classification model by machine learning and a classification model by probability.

Machine learning is a process of improving technology that obtains results by repeatedly performing tasks in order to learn information using data acquired in the learning stage and to efficiently use the learned information. The classification model or classifier by machine learning includes methods such as support vector machine (SVM), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), K-nearest neighbor (K-NN), and K-means. The present invention includes at least one machine learning method.

The classification model by machine learning is learned by training data. As described above, the training data is data on the electroencephalogram signal and its result (epilepsy seizure state). The training data of the machine learning and the training data at the time of generating the coding book may be the same.

On the other hand, training is performed using the divided EEG signals. That is, as training data for classification learning, if 100 consecutively divided EEG signals are generated, the feature vector is 100 × K (K = a split level of the frequency band to be divided, if the wavelet is transformed to the fifth order level, Divided into five frequency bands). Later, as test data for classification learning, if 100 consecutively divided EEG signals are generated, as in the previous classification learning step, feature values of 100 × K are calculated, and feature values of training and test data are calculated. The final classification.

In addition, a classification model (or a probability model) based on probability is a model in which a phenomenon generated by chance is formulated and a new input data is trained according to a statistical probability distribution. Preferably, HMM (Hidden Markov) and at least one or more probabilistic models.

On the other hand, the classification model generator 35 may evaluate the accuracy, precision, specificity, etc. of seizure waves, seizure time points, and seizure times detected by machine learning and probability model generation. The optimal classification model can be selected from a plurality of classification models (or classifiers) generated.

If the classification accuracy is calculated from the test data input based on the learned data using the classifier A, the detection accuracy of the classifier A is the same as the detection accuracy when the classifier B is used in the same manner. After calculating all the accuracy of using other classifier, select the classifier with the highest accuracy. It is important to use a classifier suitable for each object to be used and for each type of seizure wave according to the content of detecting the seizure wave. Therefore, the training data is acquired based on the electroencephalogram signal of the user and pre-learned by various machine learning and probability model techniques, and the test data obtained offline or online (real time) can be used to capture seizure timing, seizure frequency, seizure status, Seizure prediction is to be made.

Next, the classification unit 36 detects an epileptic seizure wave using various machine learning and probability-based classification models using the feature vector and the codebook.

Preferably, the classification model used may include at least one classification model of machine learning and a classification model based on probability.

According to the present invention, the reason for using two methods, the classification model by machine learning and the classification model by probability, is as follows. Binary classification, which detects seizure waves that show ideal signs unlike background waves of EEG signals, is more efficient in machine learning and predicts time-dependent predictions of seizure pauses, seizure intervals, and seizure waves over time. Probability models are more advantageous for detection.

Therefore, in the present invention, a classification model of machine learning and probability method can be selectively and adaptively used.

The results provided by the present invention include seizure timing, seizure frequency, and real-time seizure prediction. In this case, since the seizure time and the number of times are to record seizure waves after recording an EEG signal for a long time, the machine learning method can be used. On the other hand, real-time seizure prediction uses a probabilistic model for predicting seizure waves of EEG signals input in real time after learning previously recorded signals. Since the probability model is a model for predicting the previous state, the current state, and the next state, the probability model can obtain more accurate results as the EEG signals input in real time are transferred in the order of the resting period, the seizure period, and the seizure.

Next, the result providing unit 37 outputs / records or analyzes the seizure wave detection result for the input EEG signal.

In particular, when the electroencephalogram signal is measured and input in real time by the electroencephalogram measurement device 10, the electroencephalogram signal measured in real time is classified by the classification unit 36. The result providing unit 37 detects the seizure wave in real time according to the classification result of the classification unit 36 and outputs the result.

In addition, when the measured electroencephalogram signal is a batch of past measurement data, the result provider 37 may output a seizure wave detection result for the entire measurement data. Preferably, it is possible to output recordings, detection statistics, and the like for seizure wave detection such as the seizure wave detection time and the number of detections. Alternatively, finally, a detection result such as a seizure wave time point and the number of times may be provided to diagnose epilepsy.

In addition, when the result providing unit 37 detects and outputs the seizure wave in real time, the result providing unit 37 may provide the user with information about the presence and degree of epileptic seizure as an alarm or the like. In particular, it may be configured to send an alarm or emergency to the guardian's cell phone, a nearby medical institution, a disaster management center.

Based on the results provided by the result providing unit 37, it can be provided as feedback data for the diagnosis and treatment of epilepsy to patients, specialists, carers, and the like.

The result provided by the result provider 37 may include a seizure time point and the number of times. That is, when a seizure detected from a long recorded EEG signal occurs (ex. 20xx-xx-xx day, one seizure occurs at xx am xx minute xx seconds), and the number of seizures (ex. One day, one week, one month) X seizure of the liver). The specialist may determine the patient's condition, treat the patient with the drug or treatment, and as a result, feed back the detection of the seizure wave to alleviate or enhance the treated drug or treatment.

Next, a method of extracting a feature vector by the feature extractor 34 according to an embodiment of the present invention will be described in more detail with reference to FIG. 6.

As shown in FIG. 6, the feature extractor 34 decomposes the frequency band by applying a discrete wavelet transform to the EEG signal (S31). When the electroencephalogram signal is decomposed into a plurality of bands, each of the decomposed frequency bands has a wavelet coefficient (hereinafter, referred to as a wavelet coefficient for each band). In this case, the number of bands is determined as K.

At this time, the feature value is obtained using the wavelet coefficient for each band. That is, the counting value is repeatedly obtained while changing the threshold value, and the feature value is obtained by the rate of change of the counting value. At this time, each process is repeated as a level. The first level is set to 1, incremented by 1, and repeated up to level L. That is, it repeats L times to obtain the counting value by changing the threshold.

Hereinafter, the repeated steps S32 to S35 will be described. Each of the following steps is performed for each wavelet coefficient for each band.

First, a threshold for each level is generated (S32). The equation for obtaining the threshold γ _t here is as follows.

[Equation 1]

Here, t means the level of counting, and γ ₀ means the initial threshold value.

Preferably, the initial threshold value is designated as the maximum value of the wavelet coefficient extracted by frequency band decomposition.

That is, the threshold for each level is generated by decreasing by 1/2, 1/4, 1/8, ... (1/2) ^t of the magnitude of the input signal.

The threshold is for observing the frequency variation of the input signal, that is, for analyzing the frequency variation of the input signal on the time series axis by counting the number of signals exceeding a specific reference point among the input signals. The specific reference point mentioned herein may be designated as 1 / h times based on the largest value among the input signals, and in the present invention, h is designated as 2 to reduce the threshold by 1/2 times.

Next, the signal coefficient for each band is counted using the threshold for each level (S33).

That is, a counting value is obtained by counting wavelet coefficient values larger than the threshold for each level in the wavelet coefficient. At this time, the signal coefficient is counted for each decomposed band.

On the other hand, the counted value is referred to as the counting value, δ. That is, it is a feature extraction method for extracting a change according to a frequency variation of an input signal based on a specific threshold value.

As mentioned above, a method of counting the values of wavelet coefficients larger than the threshold for each level in the wavelet coefficient is as follows. That is, it is a formula for generating the counting value (δ).

[Equation 2]

Here, N means the number of samples. That is, N means the number of wavelet coefficients for each band. In addition, t denotes a level, a _i denotes an input signal (here wavelet coefficient), and γ _t denotes a threshold of the level t.

On the other hand, the counting value δ (t) is obtained for each band. Since the number of bands is K, K counting values are calculated.

Next, it is determined whether the level satisfies a predetermined specific level (or a predetermined threshold level) (S34). If the specific level is not satisfied, the level is increased to 1 (S35), and the steps of counting (S32 to S33) are repeated.

Therefore, the counting value is extracted for each level. Counting values are extracted at all levels, from the initial level to a specific level. In addition, at each level, each counting value is obtained for each band.

In addition, when the specific level is reached, the repeating structure ends. In this case, the specific level may be arbitrarily designated in advance. If the particular level is L, L counting values are extracted. The counting value of each of the K bands at each level is calculated.

Next, a rate of change (or slope) of counting values of all levels for each band is calculated (S36). The counting value change rate (or slope) of each band is set as a feature value of the corresponding band, and a feature vector is generated from feature values of each band (S37).

That is, the feature value is calculated by counting the value according to the level, the rate of change of δ. Specifically, L counting values are obtained by counting the signal coefficients (N sample sizes) of the k-th band for each level. The rate of change of the L counting values is obtained, and the rate of change is defined as a feature value of the k-th band. Preferably, the L counting values are regressed to obtain the slope or rate of change.

In this way, when the feature value (or change rate) is obtained for each band, since the number of bands is K, K change rate or feature values are obtained. Each of the K change rates is a feature value, and these are represented by vectors to form feature vectors. Therefore, the size of the feature vector is K, the number of bands.

That is, the feature vectors are decomposed into frequency bands and extracted, and each feature value is calculated as a counted value corresponding to a level, a change rate of δ.

In summary, the EEG signal obtained from the EEG device 10 is divided into n divided EEG signals and divided into K frequency bands through wavelet transform. In each K frequency bands, each N wavelet coefficients are generated. At this time, for each of the N wavelet coefficients corresponding to each of the K frequency bands, one in each of the K frequency bands through the feature extraction method, that is, as many as K number each time the feature extraction algorithm is finally performed. The feature vector is calculated (K × 1). Each time the number of algorithm runs increases by P, this feature vector is generated as (K × P).

The feature vector is obtained in the feature extraction step, and represents the rate of change of the counted value of each level. In other words, as the level increases, the counted value changes linearly. The feature value is obtained using this feature. For example, if any wavelet coefficient of an EEG signal is counted at level 1 at 10, counted at level 2 is 20, .. at 50, counted at level 5, finally at level 1 When increasing to 5, the rate of change of the counted value is calculated as the final feature value.

FIG. 7 illustrates the results of resolving frequency bands of the electroencephalogram signals of (a) normal persons and (b) epilepsy patients using a discrete wavelet transform.

S shown in FIG. 7 is an electroencephalogram signal that is raw data. In addition, D is a detail coefficient, A is an approximate coefficient, and means a wavelet coefficient obtained by wavelet transforming an electroencephalogram signal. In addition, the numbers in D and A represent the level of wavelet transform. That is, 1 means level 1, 2 means level 2, and 5 means level is divided into 5 levels.

Next, the code generation method of the codebook generation unit 34 according to an embodiment of the present invention will be described in more detail.

The codebook generator 34 uses a plurality of training data. Training data consists of the EEG signal and the result of the seizure state. A feature vector is obtained from the EEG signal of the training data, and clustering is performed on the obtained feature vector and the seizure state of the training data.

If the size of the feature vector is K and the number of types of seizure states is M, then K × M clustering is generated. Feature vector based clustering clusters feature values of the generated feature vectors. Therefore, if K is decomposed into bands and the M seizure states are classified, K × M clustering is generated.

As a specific example, a case in which the seizure state is classified into two states, which has been decomposed into six frequency bands and is seizure and seizure intermittent states, will be described. At this time, 12 classes of 6x2 in total are clustered. This is a result of clustering feature values according to frequency band bands for each seizure state.

Clustering is performed on the basis of feature vectors to generate cluster numbers. If K × M clustering is generated, numbers are assigned from 1 to K × M. Numbers are assigned sequentially, randomly.

The codebook generated as described above is used when generating a classification model based on probability.

In other words, feature vectors should be clustered before training the probability model. The reason is that the probabilistic model does not classify the feature vectors based on the reference point in the state where the feature vectors are diffused in the multidimensional space, like the machine learning method. Therefore, the feature vector value cannot be used as it is.

For example, if the weather is sunny today, the probability of rain tomorrow is 0.2, if the weather is sunny today, the probability is sunny tomorrow, if the weather is rain 0.8, the probability is sunny tomorrow. 0.9 can be assumed. At this time, if yesterday's weather rains and today's weather is clear, when the probability of tomorrow's weather rain is calculated, the actual value of the feature vectors when the state is transferred from the previous state to the current state and the next state does not exist. The final probability is obtained using the probability data generated from the actual values. And detecting the most probable state is the basic concept of probability model.

Therefore, if yesterday's weather is rainy and today's weather is clear, if you can find the probability of tomorrow's weather, you can predict tomorrow's weather just by knowing what the weather was yesterday. Here, the clustering technique may be used in a method of knowing whether the weather is rainy or clear yesterday. Assuming that there is a characteristic value of humidity, observing the humidity of the rainy day, the analysis results that the humidity is more than 80%, the sunny day is less than 30%. At this time, if the humidity continuously shows values such as 80, 81, 92, 42, etc. in yesterday's weather, clustering these values causes the rainy clusters to gather. When this cluster is designated as 1, codebook 1 is generated.

Preferably, the codebook generation unit 34 uses a Hidden Markov model (HMM) as a classification model to predict the seizure when it is transferred to the resting, seizure, and seizure state. For example, when the state of the seizure is not known, the function vector is generated by receiving an electroencephalogram signal in a certain state, and the feature vectors are clustered. The codebook K is entered into the Hidden Markov model to calculate the posterior probability based on prior probabilities. Finally, breaks, seizures, and seizures are classified.

As mentioned above, although the invention made by this inventor was demonstrated concretely according to the said Example, this invention is not limited to the said Example and can be variously changed in the range which does not deviate from the summary.

10: EEG measuring device 20: Computer terminal
30: program system 80: electroencephalogram signal
31: EEG input unit 32: Signal preprocessor
33: feature extraction unit 34: codebook generation unit
35: classification model generation unit 36: classification unit
37: result provider 38: storage unit

Claims (6)

In the epilepsy seizure detection system based on the multi-frequency band coefficient of the electroencephalogram signal,
An electroencephalogram input unit for receiving an electroencephalogram signal;
A signal preprocessor for removing and dividing noise with respect to the electroencephalogram signal;
A feature extractor for extracting a signal coefficient for each band by decomposing a plurality of frequency bands for the divided EEG signals, and extracting a feature for each band using the extracted band-specific signal coefficients;
A codebook generator for generating a codebook using training data of an electroencephalogram signal, clustering according to each band of the feature vector of the training data and the type of seizure state, and generating a codebook by assigning an index to each cluster;
A classification model generator for generating at least one classification model by machine learning using training data of an electroencephalogram signal, and generating at least one classification model by probability using the generated codebook; And,
An electroencephalogram signal comprising a classification unit for classifying seizure states with respect to an input electroencephalogram signal using at least one of the classification model by the at least one machine learning and the classification model by the at least one probability Multi-frequency band coefficient based epileptic seismic detection system.
The method of claim 1,
The feature extracting unit applies a discrete wavelet transform to an electroencephalogram signal, decomposes it into a plurality of frequency bands, generates a threshold value according to a level for each band, and counts a signal coefficient for each band that is greater than or equal to the threshold value for each level, and the level is previously If it is not satisfied, it is determined whether the level is satisfied, and if it is not satisfied, the level is increased by 1 and a threshold value for each level is generated and counted, and when the level is satisfied, the rate of change of counting values of all levels is calculated to calculate each band. An interstitial seizure detection system based on multi-frequency band coefficients of an electroencephalogram signal, characterized by extracting the band-specific features.
The method of claim 2,
An epilepsy seizure detection system based on multi-frequency band coefficients of an electroencephalogram signal, characterized in that the threshold for each level is calculated by the following [Equation 1].
[Equation 1]

T denotes a level, γ ₀ is a predetermined initial threshold, and L denotes a specific level.
The method of claim 2,
A multi-frequency band coefficient-based epileptic seizure detection system of an electroencephalogram signal, characterized by regression analysis of counting values of all levels to calculate a slope or rate of change and to set the calculated slope or rate of change for each band of each band.
The method of claim 1,
The classification model by machine learning uses any one of a support vector machine (SVM), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), K-nearest neighbor (K-NN), and K-means. Probability classification model using a hidden Markov model (HMM, Hidden Markov model) multi-frequency band coefficient-based epilepsy seizure detection system characterized in that the EEG signal.
The method of claim 1,
The system may further include a result providing unit for outputting or recording a detection result of a seizure state with respect to an input EEG signal, or providing a time and number of detections of the seizure state. Seizure detection system.

Images