WO2020061971A1 - Epilepsy brain wave state detection method based on machine learning - Google Patents

Abstract

An epilepsy brain wave state detection method based on machine learning. The method comprises the following steps: input import: importing brain wave data of an epilepsy patient and marking the state thereof; normalized transformation processing: setting a suitable new maximum value and a suitable new minimum value, and mapping brain wave time-domain signal data to a smaller new value interval according to a normalized transformation technique; time domain to frequency domain conversion: carrying out fast Fourier transform on each piece of brain wave time-domain data, and carrying out comprehensive calculation on an amplitude frequency of each piece of data and taking same as a power spectrum thereof; frequency domain range selection: selecting a suitable low-frequency signal to replace an original frequency-domain signal, and removing high-frequency signal noise; linear adaptive dimension reduction of a frequency-domain signal: using a linear adaptive dimension reduction technique to carry out data dimension reduction, so as to effectively carry out classification processing; establishment of a support vector machine classification and prediction model: using a support vector machine classifier to establish a prediction model for a training data set; and epilepsy state classification and prediction: using the established prediction and classification model to carry out state classification and prediction on a brain wave in an unknown state.

Description

Epileptic brain wave state detection method based on machine learning Technical field

The invention belongs to the technical field of machine learning and medical diagnosis, and particularly relates to a method for detecting brain wave state of epilepsy based on machine learning.

Background technique

EEG activity is a spontaneous, rhythmic potential change generated by neurons in the cerebral cortex, and it is the overall reflection of the electrophysiological activity of cerebral nerve cells on the surface of the cerebral cortex or scalp. EEG signals contain a large amount of physiological and disease information; different physiological states have different characteristics of EEG signals. Therefore, EEG signals provide a diagnostic basis for the detection of epilepsy status. With the rapid development of science and technology, human society has begun to enter a revolutionary era based on massive data and the use of information technology for knowledge creation and information mining; the entry of information technology into the medical industry is a need for the development of the times.

In medical diagnosis, the use of data mining technology to detect the state of epilepsy brain waves has become a hot spot in the medical industry. The purpose is to establish a predictive model based on the brain wave data of patients with epilepsy during the onset and intermittent periods to assist doctors in the diagnosis of the state. The analysis methods of epilepsy brain waves include: time-domain analysis, frequency-domain analysis and time-frequency domain analysis. The time-domain analysis method directly extracts characteristic waves similar to EEG for observation; time-frequency analysis is a comprehensive analysis of EEG signals by combining time-domain signals and frequency-domain signals. In the frequency domain analysis method, power spectrum estimation is its main method, and its significance is related to transforming the brain wave with amplitude over time into the spectrum of brain power with frequency domain transformation, so that the distribution of EEG rhythm can be observed intuitively. And changes.

In the frequency dimensionality reduction process of frequency domain analysis, the power spectrum is usually superimposed as a group every 10 consecutive frequencies; if there are 50 groups of frequency data, 1 to 10 groups of frequencies are superimposed as a group, so that the data is superimposed The reduction of 50 columns to 5 columns achieves the purpose of dimensionality reduction, which is conducive to improving the efficiency of subsequent calculations. However, such a rough dimensionality reduction method directly superimposes continuous power spectrum data, ignoring the correlation between frequency data, resulting in the loss of unique information between samples, and eventually leading to a decrease in the prediction performance of epilepsy state detection.

Summary of the Invention

In order to overcome the shortcomings of the prior art, the present invention provides a method for detecting brain wave state of epilepsy based on machine learning, which includes the following main steps:

Step 1. Import data, import brain wave data from patients with epilepsy, and mark their status.

Step 2: Normalize the transformation process, formulate appropriate new maximum and minimum values, and map the EEG time-domain signal data to a smaller new value interval according to the normalized transformation technique.

This step mainly includes:

Step 2.1: Obtain the maximum value max and the minimum value min from the brain wave data set.

Step 2.2. Set the new maximum value new_max and minimum value new_min as required and use the normalized transformation calculation formula:

Normalize transformation calculation (convert to v ') for each data v in the data set, and transform its range from [min, max] to the interval [new_min, new_max].

Step 3, time-frequency domain conversion, performing fast Fourier transform on the time-domain data of each brainwave, and comprehensively calculating the amplitude frequency of each data as its power spectrum.

Step 4. Select a frequency domain range, select a suitable low frequency signal to replace the original frequency domain signal, and remove high frequency signal noise.

This step mainly includes:

Step 4.1, randomly select d samples in the data set, and record them as {X ₁ , X ₂ , ..., X}, X _i = (x _i1 , x _i2 , ..., x _im ), m represents the data Set of dimensions.

Step 4.2, find the minimum value P, so that for each sample X _t satisfies the condition:

Where R is a user-specified threshold and the recommended range is [0.9,1), which means that the original sample is characterized by a small low-frequency signal, and high-frequency noise in the EEG signal is removed.

In step 4.3, the data obtained by formula (2) is used to intercept the first 1 to P columns of the original sample to replace the original data to achieve the purpose of removing noise.

Step 5. The linear adaptive dimension reduction of the frequency domain signal is performed, and the linear adaptive dimension reduction technology is adopted to perform the data dimension reduction in order to effectively perform the classification processing.

This step mainly includes:

Step 5.1, record the data set X = {X ₁ , X ₂ , ..., X _n }, X _i = (x _i1 , x _i1 , ..., x _iP ), P represents the dimension of the data set, determine each Adaptive final neighborhood points of the sample points X _i :

Step 5.1.1, use K-Means algorithm based on Euclidean distance (K-means algorithm is a distance-based clustering algorithm, using distance as the evaluation index of similarity) to perform cluster analysis on the data set for final Selection and judgment of neighborhood points.

Step 5.1.2, select the initial neighborhood points X _i1 , X _i2 , ..., X _{ik of the} sample points X _i , where k represents the set number of initial neighborhood points.

Step 5.1.3. For the sample point X _i , calculate the average distance between the initial neighborhood points X _i1 , X _i2 , ..., X _ik and the sample X _i , and the calculation formula is:

Among them, D _ik represents the Euclidean distance between the sample point X _i and its neighborhood point X _ik .

Step 5.1.4. In the initial neighborhood points, select the final neighborhood point for each sample point X _i , which must meet the condition: if the distance D _ik between the initial neighborhood point X _ik and the sample point X _i is less than the distance mean DM _i , or if X _i belongs to the same cluster as the initial neighborhood point X _ik , then X _ik is the final neighborhood point of X _i ; otherwise it is not.

In step 5.2, the linear combination of the final neighborhood points determined in step 5.1 is used to reconstruct X _i , and the sample's local embedded weight matrix W is calculated to minimize the reconstructed cost error ε (W). The calculation formula is:

Wherein, upon reconstitution indicates the coefficient ω _ij X _i X _j share of the weight; when not the points in the neighborhood of X _i X _j, ω _ij = 0; and Σ _j ω _ij = 1; n represents the total number of data sets . Obtained by the Lagrangian multiplier method, the embedding weight matrix for the sample X _i is:

Where G _i = (G _jk ), and G _jk = (X _i -X _j ) ^T (X _i -X _k ) (X _j , X _k is a neighborhood point of X _i ); 1 _n is dimension n Column vector of 1, that is, 1 _n = (1,1, ..., 1) ^T ; 1 _n ^T is a transpose of 1 _n ; w _i = (ω _i1 , ω _i2 , ..., ω _in ) ^T , W = (w ₁ , w ₂ , ..., w _n ).

Step 5.3, the optimal mapping in the low-dimensional space is solved by the local embedding weight W, so that the embedding cost error ε (Y) in the p-dimensional space is minimized, that is:

Where Z _i is the embedded representation of X _i in a low-dimensional space, Z = {Z ₁ , Z ₂ , ..., Z _n }, F = (I _n -W) ^T (I _n -W), and tr represents a matrix N, the n number represents the total number of data sets, I _n is an n-th order identity matrix where the elements on the main diagonal are 1 and the rest are all 0. Z needs to satisfy

This problem is actually solving the minimum of the non-zero eigenvalues of the matrix F. Assume that the eigenvectors corresponding to P non-zero eigenvalues of matrix F arranged in ascending order are u ₁ , u ₂ , ..., u _P , and select the first p non-zero eigenvalues from small to large, and finally find the embedding low The sample data set of the dimensional space is U = (u ₁ , u ₂ , ..., u _p ) ^T. In this way, the P-dimensional data set in step 4 becomes a p-dimensional data set through linear adaptive dimensionality reduction.

Step 6. Establish a support vector machine classification and prediction model, and use the support vector machine classifier to establish a prediction model on the training data set.

Step 7. Classification and prediction of the status of epilepsy. Use the established prediction classification model to perform state classification prediction on brain waves of unknown states.

The beneficial effects of the present invention are: providing an effective detection method based on machine learning for the state detection of epilepsy brain waves; using a linear adaptive dimensionality reduction algorithm to reduce the dimension of a data set, and introducing a polymorphism into the linear adaptive dimensionality reduction algorithm. The concept of class and mean, through the limitation of these two, allows the neighborhood points of the sample points to be adaptively selected, and enables the low-dimensional embedding of the data set to well maintain the topology and manifold structure of the original data set; the invention It has a good performance in terms of detection efficiency, computing overhead and accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a flowchart of a method for detecting an epilepsy brain wave state based on machine learning according to the present invention.

Figure 2 shows a flowchart of a linear adaptive dimensionality reduction algorithm.

detailed description

The preferred embodiments of the present invention will be further described below with reference to the accompanying drawings and embodiments.

The flowchart shown in Figure 1 gives the specific process of the entire implementation of the present invention:

Step 1. Import data, import brain wave data from patients with epilepsy, and mark their status.

This step includes:

Step 1.1, import brain wave data, each row of data represents the patient's time-domain signal under a sampling period, and each column represents the time-domain signal obtained under a pulse, that is, the data set is M = {m ₁ , m ₂ , ... , m _{r}, r} represents the number of data sets of _{samples, m i = (m i1,} m i2, ..., m is), s represents a number of brain-wave data sample i.e. the number of columns.

In step 1.2, the state corresponding to each sample is recorded through the array y = (y ₁ , y ₂ , ..., y _r ), that is, the state of the intermittent period is recorded as -1 and the state of the onset period is recorded as 1.

Step 2: Normalize the transformation process, formulate appropriate new maximum and minimum values, and map the EEG time-domain signal data to a smaller new value interval according to the normalized transformation technique.

This step includes:

Step 2.1: Obtain the maximum value max and the minimum value min from the brain wave data set.

Step 2.2. Set the new maximum value new_max and minimum value new_min as required and use the normalized transformation calculation formula:

Normalize transformation calculation (convert to v ') for each data v in the data set, and transform its range from [min, max] to the interval [new_min, new_max].

Step 3, time-frequency domain conversion, performing fast Fourier transform on the time-domain data of each brainwave, and comprehensively calculating the amplitude frequency of each data as its power spectrum.

This step includes:

Step 3.1, since the time-domain signal of the brain wave satisfies the continuous multivariate generalized B distribution, that is, the Dirichlet distribution, the function x (t) of the time-domain signal of the brain wave sets Δt as a small time interval in the [0, T] interval. , T = NΔt, then the sampling value of x (t) when t _n = nΔt is x _n , and the sum of each x _{n is} used to replace x (t), then the formula for calculating the fast Fourier transform is:

Due to the symmetry of the fast Fourier transform, each time-domain sample m _t = (m _t1 , m _t2 , ..., m _ts ) is converted to

Dataset dimensions are transformed from s to

Step 3.2: Perform power spectrum calculation according to the amplitude and phase values in the frequency spectrum. For each data such as a _jk + ib _jk , the power spectrum c _{jk is} calculated as:

The power spectrum c _{jk is used} instead of its original data a _jk + ib _jk , and the data set is converted from M to C.

Step 4. Select a frequency domain range, select a suitable low frequency signal to replace the original frequency domain signal, and remove high frequency signal noise.

This step includes:

Step 4.1, randomly select d samples in the data set, and record them as {X ₁ , X ₂ , ..., X _d }, X _i = (x _i1 , x _i2 , ..., x _im ), where m represents The dimensions of the dataset.

Step 4.2, find the minimum value P, so that for each sample X _t satisfies the condition:

Where R is a user-specified threshold and the recommended range is [0.9,1), which means that the original sample is characterized by a small low-frequency signal, and high-frequency noise in the EEG signal is removed.

In step 4.3, the data obtained by formula (10) is used to intercept the first 1 to P columns of the original sample to replace the original data to achieve the purpose of removing noise.

Step 5. The linear adaptive dimension reduction of the frequency-domain signal is performed, and the linear adaptive dimension reduction technique is adopted to reduce the dimension of the data in order to effectively perform classification processing, as shown in FIG. 2.

This step includes:

Step 5.1, record the data set X = {X ₁ , X ₂ , ..., X _n }, X _i = (x _i1 , x _i1 , ..., x _iP ), P represents the dimension of the data set, determine each The adaptive final neighborhood points of the sample points X _i mainly include:

Step 5.1.1. Use K-Means algorithm based on Euclidean distance to perform cluster analysis on the data set for the selection and judgment of the final neighborhood points. The main steps are as follows: First, k cluster centers are randomly given initially, and the sample selects the nearest cluster center according to the Euclidean distance criterion and classifies it into the corresponding cluster to complete the first assignment. After that, it re-accords to the samples in the cluster. Calculate the centers of the clusters and redistribute the samples into the appropriate clusters according to the Euclidean distance criterion until the cluster centers no longer transform or the number of clusters reaches a set threshold.

Step 5.1.2, select k initial neighborhood points of each sample point of the data set, and record the initial neighborhood points of the sample point X _i as X _i1 , X _i2 , ..., X _ik , where k represents the set Initial neighborhood points.

Step 5.1.3. For the sample point X _i , calculate the average distance between the initial neighborhood points X _i1 , X _i2 , ..., X _ik and the sample X _i , and the calculation formula is:

Among them, D _ik represents the Euclidean distance between the sample point X _i and its neighborhood point X _ik .

Step 5.1.4. In the initial neighborhood points, select the final neighborhood point for each sample point X _i , which must meet the condition: if the distance D _ik between the initial neighborhood point X _ik and the sample point X _i is less than the distance mean DM _i , or if X _i belongs to the same cluster as the initial neighborhood point X _ik , then X _ik is the final neighborhood point of X _i ; otherwise it is not.

In step 5.2, the linear combination of the final neighborhood points determined in step 5.1 is used to reconstruct X _i , and the sample's local embedded weight matrix W is calculated to minimize the reconstructed cost error ε (W). The calculation formula is:

Among them, the coefficient ω _ij represents the weight occupied by X _j when reconstructing X _i ; when X _j is not a neighborhood point of X _i , ω _ij = 0, and

n represents the total number of data sets. Obtained by the Lagrangian multiplier method, the embedding weight matrix for the sample X _i is:

Where G _i = (G _jk ), and G _jk = (X _i -X _j ) ^T (X _i -X _k ) (X _j , X _k is a neighborhood point of X _i ); 1 _n is dimension n Column vector of 1, that is, 1 _n = (1,1, ..., 1) ^T ; 1 _n ^T is a transpose of 1 _n ; w _i = (ω _i1 , ω _i2 , ..., ω _in ) ^T , W = (w ₁ , w ₂ , ..., w _n ).

Step 5.3, the optimal mapping in the low-dimensional space is solved by the local embedding weight W, so that the embedding cost error ε (Y) in the p-dimensional space is minimized, that is:

Where Z _i is the embedded representation of X _i in a low-dimensional space, Z = {Z ₁ , Z ₂ , ..., Z _n }, F = (I _n -W) ^T (I _n -W), and tr represents a matrix N, the n number represents the total number of data sets, I _n is an n-th order identity matrix where the elements on the main diagonal are 1 and the rest are all 0. Z needs to satisfy

This problem is actually solving the minimum of the non-zero eigenvalues of the matrix F. Assume that the eigenvectors corresponding to P non-zero eigenvalues of matrix F arranged in ascending order are u ₁ , u ₂ , ..., u _P , and select the first p non-zero eigenvalues from small to large. The sample data set of the dimensional space is U = (u ₁ , u ₂ , ..., u _p ) ^T. In this way, the P-dimensional data set in step 4 becomes a p-dimensional data set through linear adaptive dimensionality reduction.

Step 6. Establish a support vector machine classification and prediction model, and use the support vector machine classifier to establish a prediction model on the training data set.

If a given training set _{(X i, y i),} i = 1,2, ..., l, X i ∈R n, y i ∈ {± 1}, referred to as a hyperplane (W · X) + b = 0, w, x ∈ R ⁿ . In order for the classification to correctly classify all samples and have a classification interval, the constraints need to be met:

y _i [(W · X _i ) + b] ≥1, i = 1,2, ..., l (15)

The classification interval can be calculated as 2 / ‖W‖, so the problem of constructing the optimal hyperplane is converted to the minimum classification interval minφ (W) under constraint:

This optimization problem is solved by introducing the Lagrangian multiplier method:

In the formula, a = (a ₁ , a ₂ , ..., a _l ), and any a _i > 0 is a Lagrangian multiplier; the optimal weight vector W and the optimal offset b obtained by the solution are:

and

Where j∈ {j | a _j > 0}, and a satisfies

And

Therefore, the optimal classification function is:

f (X) = sgn {(W · φ (X)) + b) (18)

sgn returns an integer variable that satisfies

Using this method, a support vector machine classification and prediction model is established for the EEG data set and its corresponding epilepsy state for epilepsy state detection.

Step 7. Classification and prediction of the status of epilepsy. Use the established prediction classification model to perform state classification prediction on brain waves of unknown states.

This step includes:

Step 7.1: Import the unlabeled epilepsy brain wave data set.

Step 7.2, transform and process the data set according to steps 2, 3, 4, and 5 in sequence;

Step 7.3: Use the support vector machine classification prediction model established in step 6 to perform state classification prediction on the processed epilepsy brain wave data.

Claims (4)

1. A method for detecting brain wave state of epilepsy based on machine learning, which comprises the following steps:

   Step 1. Import data, import brain wave data from patients with epilepsy, and mark their status.

   Step 2: Normalize the transformation process, formulate appropriate new maximum and minimum values, and map the EEG time-domain signal data to a smaller new value interval according to the normalized transformation technique.

   Step 3, time-frequency domain conversion, performing fast Fourier transform on the time-domain data of each brainwave, and comprehensively calculating the amplitude frequency of each data as its power spectrum.

   Step 4. Select a frequency domain range, select a suitable low frequency signal to replace the original frequency domain signal, and remove high frequency signal noise.

   Step 5. The linear adaptive dimension reduction of the frequency domain signal is performed, and the linear adaptive dimension reduction technology is adopted to perform the data dimension reduction in order to effectively perform the classification processing.

   Step 6. Establish a support vector machine classification and prediction model, and use the support vector machine classifier to establish a prediction model on the training data set.

   Step 7. Classification and prediction of the status of epilepsy. Use the established prediction classification model to perform state classification prediction on brain waves of unknown states.
2. The method for detecting brain wave state of epilepsy based on machine learning according to claim 1, characterized in that, in step 2, an appropriate new maximum value and minimum value are formulated, and the time-domain signal data of the brain wave is calculated according to The normalized transformation technique maps to smaller new value intervals. The step 2 further includes:

   Step 2.1: Obtain the maximum value max and the minimum value min from the brain wave data set.

   Step 2.2. Set the new maximum value new_max and minimum value new_min as required and use the normalized transformation calculation formula:

   

   Normalize transformation calculation (convert to v ') for each data v in the data set, and transform its range from [min, max] to the interval [new_min, new_max].
3. The method for detecting epilepsy brain wave state based on machine learning according to claim 1, characterized in that, in step 4, frequency range selection is performed, and a suitable low-frequency signal is selected instead of the original frequency-domain signal to remove high-frequency signals. Frequency signal noise. The step 4 further includes:

   Step 4.1, randomly select d samples in the data set, and record them as {X ₁ , X ₂ , ..., X _d }, X _i = (x _i1 , x _i2 , ..., x _im ), where m represents The dimensions of the dataset.

   Step 4.2, find the minimum value P, so that for each sample X _t satisfies the condition:

   

   Where R is a user-specified threshold and the recommended range is [0.9,1), which means that the original sample is characterized by a small low-frequency signal, and high-frequency noise in the EEG signal is removed.

   In step 4.3, the data obtained by formula (2) is used to intercept the first 1 to P columns of the original sample to replace the original data to achieve the purpose of removing noise.
4. The method for detecting an epilepsy brain wave state based on machine learning according to claim 1, characterized in that, in step 5, the linear adaptive dimension reduction of the frequency domain signal is performed in order to effectively perform classification processing. The step 5 further includes:

   Step 5.1, record the data set X = {X ₁ , X ₂ , ..., X _n }, X _i = (x _i1 , x _i1 , ..., x _iP ), P represents the dimension of the data set, determine each Adaptive final neighborhood points of the sample points X _i . The step 5.1 further includes:

   Step 5.1.1, use K-Means algorithm based on Euclidean distance (K-means algorithm is a distance-based clustering algorithm, using distance as the evaluation index of similarity) to perform cluster analysis on the data set for final Selection and judgment of neighborhood points.

   Step 5.1.2, select the initial neighborhood points X _i1 , X _i2 , ..., X _{ik of the} sample points X _i , where k represents the set number of initial neighborhood points.

   Step 5.1.3. For the sample point X _i , calculate the average distance between the initial neighborhood points X _i1 , X _i2 , ..., X _ik and the sample X _i , and the calculation formula is:

   

   Among them, D _ik represents the Euclidean distance between the sample point X _i and its neighborhood point X _ik .

   Step 5.1.4. In the initial neighborhood points, select the final neighborhood point for each sample point X _i , which must meet the condition: if the distance D _ik between the initial neighborhood point X _ik and the sample point X _i is less than the distance mean DM _i , or if X _i belongs to the same cluster as the initial neighborhood point X _ik , then X _ik is the final neighborhood point of X _i ; otherwise it is not.

   In step 5.2, the linear combination of the final neighborhood points determined in step 5.1 is used to reconstruct X _i , and the sample's local embedded weight matrix W is calculated to minimize the reconstructed cost error ε (W). The calculation formula is:

   

   Wherein, upon reconstitution indicates the coefficient ω _ij X _i X _j share of the weight; when not the points in the neighborhood of X _i X _j, ω _ij = 0; and Σ _j ω _ij = 1; n represents the total number of data sets . Obtained by the Lagrangian multiplier method, the embedding weight matrix for the sample X _i is:

   

   Where G _i = (G _jk ), and G _jk = (X _i -X _j ) ^T (X _i -X _k ) (X _j , X _k is a neighborhood point of X _i ); 1 _n is dimension n Column vector of 1, that is, 1 _n = (1,1, ..., 1) ^T ; 1 _n ^T is a transpose of 1 _n ; w _i = (ω _i1 , ω _i2 , ..., ω _in ) ^T , W = (w ₁ , w ₂ , ..., w _n ).

   Step 5.3, the optimal mapping in the low-dimensional space is solved by the local embedding weight W, so that the embedding cost error ε (Y) in the p-dimensional space is minimized, that is:

   

   Where Z _i is the embedded representation of X _i in a low-dimensional space, Z = {Z ₁ , Z ₂ , ..., Z _n }, F = (I _n -W) ^T (I _n -W), and tr represents a matrix N, the n number represents the total number of data sets, I _n is an n-th order identity matrix where the elements on the main diagonal are 1 and the rest are all 0. Z needs to satisfy

   

   This problem is actually solving the minimum of the non-zero eigenvalues of the matrix F. Assume that the eigenvectors corresponding to P non-zero eigenvalues of matrix F arranged in ascending order are u ₁ , u ₂ , ..., u _P , and select the first p non-zero eigenvalues from small to large, and finally find the embedding low The sample data set of the dimensional space is U = (u ₁ , u ₂ , ..., u _p ) ^T. In this way, the P-dimensional data set in step 4 becomes a p-dimensional data set through linear adaptive dimensionality reduction.

Images