Adaptive brain region EEG artifact detection method based on multi-classifier fusion

Abstract

The present invention discloses an adaptive brain region EEG artifact detection method with multi-classifier fusion. The present invention comprises the following steps: 1. filtering multi-channel EEG signals and classifying them into multiple categories of artifacts; 2. analyzing the brain regions and channel correlations corresponding to each category of artifact signals, and obtaining the regional information and channel information of each category of artifacts; 3. extracting features from the channel information of each category of artifacts; 4. performing two-stage feature selection using the ReliefF algorithm and the mRMR algorithm for the regional information and the extracted features, and establishing feature grouping; 5. training a classification model using a machine learning algorithm combined with the selected features; 6. systematically building a plurality of classifiers trained in step 5. The present invention overcomes the tediousness of manual artifact positioning in clinical practice, improves the rapid positioning of abnormal signals, solves the monotony of existing artifact recognition technologies, and can simultaneously realize real-time artifact detection of multi-channel EEG data.

Description

Technical Field

[0001] The present invention belongs to the field of electroencephalogram (EEG) signal processing and intelligent medical treatment. By using a multi-classifier fusion method combined with the proposed pre-selection of brain regions for artifact signals, a system is designed that can realize real-time recognition and detection of multiple types of artifact signals in the collected EEG data. Background Art

[0002] Electroencephalography (EEG) is an essential tool for examining and documenting neurological disorders. However, physiological artifacts are unavoidable in EEG recordings. These artifacts blend into the normal EEG background, making it difficult for medical personnel to identify the normal EEG background data and hindering subsequent EEG signal applications. Therefore, detecting the location and identifying the type of artifacts is crucial.

[0003] However, traditional artifact detection methods have the following limitations:

[0004] 1. Traditional detection methods use independent component analysis (ICA) to process and analyze artifacts. However, brain electrodes are easily affected by external factors. After data mixing, the data of each channel is impure, making the EEG data not meet the requirements of the ICA processing method.

[0005] 2. When distinguishing artifacts in EEG data, traditional methods do not determine the location of the artifacts and use the same method for detection, which is easily interfered by abnormal signals.

[0006] 3. Traditional artifact detection methods can only process certain types of artifact signals in a few channels, lack diversity, and cannot simultaneously detect multiple types of artifacts in EEG data in real time.

[0007] This invention addresses the shortcomings of traditional artifact detection methods by using multi-channel data to reduce the impact of abnormal information in a few channels of EEG data. By analyzing the brain electrodes where artifacts appear, it proposes a method for pre-identifying the brain regions where artifacts appear, extracting corresponding features, and achieving more accurate artifact recognition. Specialized classifiers are trained for different types of artifacts. By fusing multiple classifiers and combining their characteristics, the efficiency and accuracy of recognition are improved, enabling real-time recognition of multiple artifacts. Compared with traditional artifact detection algorithms, this invention has lower complexity, higher recognition speed, and can identify more artifact types. Summary of the Invention

[0008] The present invention comprehensively considers the situation where different types of artifacts in EEG data appear in different brain regions, more comprehensively considers multi-channel data, and screens out several types of artifact signals that most commonly appear in long-term EEG data recordings. It overcomes the situation where traditional artifact detection can only process a few types of artifacts, and proposes a system that pre-analyzes brain regions of artifacts and combines multi-classifier fusion to realize multi-class artifact recognition, thereby realizing the recognition of multiple artifacts in different brain regions.

[0009] The technical solution of the present invention mainly includes the following steps:

[0010] Step 1: Filter the multi-channel EEG signal and classify it into multiple artifact categories;

[0011] Step 2: Perform brain region analysis and channel correlation analysis on each type of artifact signal obtained in step 1 to obtain regional information and channel information of each type of artifact.

[0012] Step 3: Extract features from the channel information of each type of artifact, extracting a total of 28 dimensions of features;

[0013] Step 4: Based on the regional information divided in step 2 and the features extracted in step 3, the ReliefF algorithm and the mRMR algorithm are used to perform two-stage feature selection and establish feature groups.

[0014] Step 5: Use the machine learning algorithm combined with the features selected in step 4 to train the classification model.

[0015] Step 6: Build a system based on the multiple classifiers trained in step 5.

[0016] The specific contents of step 1 are as follows:

[0017] The 19-channel, 1000Hz raw EEG data were filtered, and each channel was subjected to a 50Hz notch filter and a 1Hz-70Hz bandpass filter to obtain EEG data without non-physiological interference; due to the discontinuity of artifacts in the EEG data, the artifact data was segmented using a sliding window method, and the time length of each sliding window was set to 1s; signals containing blink artifacts were recorded as the Eye_Blink class with a category label of 1; signals containing frontal EMG were recorded as the Frontal_EMG class with a category label of 2; signals containing chewing artifacts were recorded as the Chew_EMG class with a category label of 3; signals containing temporal EMG were recorded as the Temp_EMG class with a category label of 4; signals with unconventional EEG background were recorded as the Unknow_Sig class with a category label of 5; and normal EEG background signals were recorded as the Norm_Sig class with a category label of 0. The dataset was divided into training samples (80%) and test samples (20%).

[0018] The specific contents of step 2 are as follows:

[0019] 2-1. For each artifact signal type obtained in step 1, the average frequency spectral density (PSD) of the artifact signal was extracted. Interpolation plots were drawn and analyzed for each artifact type. Blink artifacts and frontal EMG signals were primarily located in the frontal region's Fp1 and Fp2 channels; chewing artifacts and temporal EMG signals were primarily located in the temporal region's F7, F8, T3, and T4 channels; and ambiguous EEG signals exhibited high amplitude and low frequency characteristics across the scalp, with a wide distribution across multiple channels, including Fp1, Fp2, F7, F8, T3, and T4, and across multiple electrodes. Based on the electrode locations at which these five types of signals (blink artifacts, frontal EMG signals, chewing artifacts, temporal EMG signals, and ambiguous EEG signals) appeared, the brain was divided into three regions: frontal region (Front_Region), temporal region (Temp_Region), and global region (Global_Region).

[0020] 2-2. Perform correlation analysis on the electrode channels of the regional information obtained in step 2-1 for different types of artifact signals to obtain correlations between the channels where different types of artifacts appear.

[0021] In the frontal region, blink artifacts showed high correlation in the Fp1 and Fp2 channels, followed by EMG in the frontal region. In the temporal region, chewing artifacts and EMG artifacts showed high correlation in F7, F8, T3, and T4. Ambiguous EEG signals (unconventional EEG background signals) showed high regional symmetry across all channels, with high correlation between left and right electrodes. Therefore, we focused on the Fp1 and Fp2 channels in the frontal region, the F7 and T4 channels in the temporal region, and the Fp1, F7, T3, T5, O1, F3, C3, P3, Fz, Cz, and Pz channels globally.

[0022] The specific contents of step 3 are as follows:

[0023] 3-1. Calculate the maximum value MAX, minimum value MIN, peak value Peak, peak-to-peak value P2P (peak to peak), mean value Mean, absolute mean AM, root amplitude RA, variance Var, standard deviation Std, root mean square RMS, kurtosis Kurt, skewness Ske, shape factor SF, peak factor PF1, pulse factor PF2, margin factor MF, clearance factor CF, median Median, fractal dimension FD, approximate entropy ApEn, sample entropy SaEn, fuzzy entropy FuEn, spectral entropy SpEn, average frequency AF, barycenter frequency BF, root mean square frequency RMSF frequency), standard deviation of frequency StdF (standard deviation of frequency), average power spectral density APSD (average power spectral density); the above total 28 dimensions of features;

[0024] Assume that each data sample format is S M*N (M: number of channels, N: data length);

[0025] 3-2 Calculate the minimum value MIN of different channel data using the following formula:

[0026] MIN i =min i (S i*N )

[0027] Where min i () represents the minimum value of the i-th channel to be solved.

[0028] 3-3 Calculate the peak value of different channel data using the following formula:

[0029] Peak i =maxi (|S i*N |)

[0030] Where |S i*N | means to find the absolute value of the i-th channel, max i () represents the maximum value of the i-th channel to be solved.

[0031] 3-4 Calculate the fractal dimension FD of different channel data using the following formula:

[0032]

[0033] Among them FD i (S,ε) represents the fractal dimension value, ε represents the grid side length when calculating the box-counting dimension, G(ε) represents the cumulative grid number of signal data, and S represents each sample data.

[0034] 3-5 Calculate the approximate entropy ApEn of different channel data using the following formula:

[0035] ApEn i =φ m (S t*N )-φ m+1 (S t*N )

[0036]

[0037] In the formula Represents the probability measure of the signal sequence, φ m (S t*N ) is expressed as the logarithmic mean of the probability measure of the signal sequence.

[0038] 3-6 Calculate the spectral entropy SpEn of different channel data using the following formula:

[0039]

[0040] Where G x () represents the transfer function that a given data sequence complies with.

[0041] 3-7 Calculate the average frequency AF of different channel data using the following formula:

[0042]

[0043] Where W represents the frequency of the given continuous data signal.

[0044] 3-8 Calculate the average power spectral density APSD of different channel data using the following formula:

[0045] APSD i =PSDi (AF i )

[0046]

[0047] Where PSD i represents the power spectral density of the i-th channel, fft( ) represents the fast Fourier transform of the signal, fs represents the sampling frequency, AF i represents the average frequency of the i-th channel.

[0048] The specific contents of step 4 are as follows:

[0049] 4-1. Use the ReliefF algorithm to perform a stage of feature selection on the 28-dimensional features extracted from each type of artifact, reducing feature redundancy for the next operation, reducing feature reuse, and improving resolution efficiency. The feature weight selection of the ReliefF algorithm is shown in the following formula:

[0050]

[0051] Where p(C) represents the proportion of the class of features in the total population, p(class(R)) is the proportion of a randomly selected sample class, A represents the currently selected feature, R and H represent two different samples in feature A, diff(A, R, H) represents the difference between two different samples R and H on feature A, and M j (C) represents the jth nearest neighbor sample in class C, m represents the number of sampling times, and k represents the number of nearest neighbor samples;

[0052] 4-2. Based on the RelifF algorithm in 4-1, the features of the artifact information in different regions are ranked, and the mRMR method is used for secondary feature selection to construct the most significant feature combination between the different artifact information in different regions. The mRMR algorithm is shown in the following formula:

[0053] Maximum correlation:

[0054]

[0055] Where x i represents the i-th type of feature, c represents the target classification, L represents the feature subset consisting of the selected m features, I(x i ; c) represents the mutual information between the i-th feature and the target category c.

[0056] Minimum redundancy:

[0057]

[0058] mRMR weight combination algorithm:

[0059]

[0060] Among them, X represents the set of features contained in the given data, L m-1 Represents the (m-1) feature subsets selected from a given feature subset.

[0061] The present invention adopts a two-stage feature selection method as follows: (1) the ReliefF algorithm is used to screen out candidate feature sets for different classifiers; (2) the mRMR algorithm is used to directly remove the redundancy of the feature candidate set obtained by the ReliefF algorithm to screen out the best feature grouping for the trained classifier.

[0062] 4-3. Based on the different feature groups analyzed in steps 4-1 and 4-2, establish an N1*M*chs-dimensional feature group DFs for distinguishing brain regions where artifacts occur; establish an N1*M*chs-dimensional feature group EFs for different types of artifacts in the frontal region; establish an N1*M*chs-dimensional feature group TFs for different types of artifacts in the temporal region; and establish an N1*M*chs-dimensional feature group GFs for the entire brain. Where N1 represents the number of samples, M represents the number of selected channels, and chs represents the number of selected channels.

[0063] The specific contents of step 5 are as follows:

[0064] 5-1 Based on the feature groups extracted from different artifacts and brain regions in step 4 and the sample labels they carry, a machine learning algorithm is used to train the initial divided training set to obtain multiple classifier models;

[0065] 5-2 Use the trained classifier model to test the initially divided test set data, calculate the test accuracy, modify the classifier model parameters, and obtain the classifier model with the best performance. Use the feature group DFs to train a classifier for adaptively discriminating brain regions, recorded as Classifier0; use the feature group EFs to train a classifier model for discriminating artifacts in the frontal region, recorded as Classifier1 and Classifier1' respectively; use the feature group TFs to train a classifier model for discriminating artifacts in the temporal region, recorded as Classifier2 and Classifier2' respectively; use the feature group GFs to train a classifier model for discriminating global regions, recorded as Classifier3.

[0066] The specific contents of step 6 are as follows:

[0067] 6-1 Build the final multi-classifier fusion adaptive brain region artifact detection system based on the classifiers trained for different regions in step 5. Since the original EEG data is continuous, a 1s data segment is input each time, and the classifier Classifier0 is used to adaptively determine the brain region of the input data. After obtaining the regional information, the classifier corresponding to each region is used to extract the signal features for artifact detection.

[0068] 6-2 Using unprocessed raw EEG data to test the system, it can accurately identify different types of artifacts in continuous data.

[0069] The beneficial effects of the present invention are as follows:

[0070] The adaptive brain region EEG artifact detection method using multi-classifier fusion is used to process raw EEG data. By pre-dividing the area where the artifact signal appears, the accuracy of artifact detection is improved. At the same time, it can also realize the recognition of multiple types of artifact signals in EEG data, overcoming the shortcomings of traditional recognition methods such as single recognition type, complex process, low recognition efficiency, and channel interference. The multiple classifier models trained are not limited to a single machine learning method. The use of multiple classifier models can combine the advantages of different methods to achieve better recognition results. The present invention can realize the detection of multiple types of artifacts in raw EEG data. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of signal processing and adaptive brain region selection according to the present invention;

[0072] Figure 2 This is a schematic diagram of the classifier fusion system framework of the present invention;

[0073] Figure 3 This is an evaluation of the practical application performance of the present invention. DETAILED DESCRIPTION

[0074] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] like Figure 1 and Figure 2 The following are system framework diagrams for channel data segmentation, brain region division, and multi-classifier fusion. Feature extraction and selection, and the construction of a multi-classifier model have been introduced in detail in the summary of the invention. That is, the technical solution of the present invention mainly includes the following steps:

[0076] Step 1: Filter the multi-channel EEG signal and classify it into multiple artifact categories;

[0077] Step 2: Perform brain region analysis and channel correlation analysis on each type of artifact signal obtained in step 1 to obtain regional information and channel information of each type of artifact.

[0078] Step 3: Extract features from the channel information of each type of artifact, extracting a total of 28 dimensions of features;

[0079] Step 4: Based on the regional information divided in step 2 and the features extracted in step 3, the ReliefF algorithm and the mRMR algorithm are used to perform two-stage feature selection and establish feature groups.

[0080] Step 5: Use the machine learning algorithm combined with the features selected in step 4 to train the classification model.

[0081] Step 6: Build a system based on the multiple classifiers trained in step 5.

[0082] The specific contents of step 1 are as follows:

[0083] The 19-channel, 1000Hz raw EEG data were filtered, and all channels were subjected to 50Hz notch filtering and 1Hz-70Hz bandpass filtering to obtain EEG data without non-physiological interference; due to the discontinuity of artifacts in the EEG data, the artifact data were segmented using a sliding window method, and the time length of each sliding window was set to 1s; signals containing blink artifacts were recorded as Eye_Blink class with category label 1; signals containing frontal EMG were recorded as Frontal_EMG class with category label 2; signals containing chewing artifacts were recorded as Chew_EMG class with category label 3; signals containing temporal EMG were recorded as Temp_EMG class with category label 4; signals with unconventional EEG background were recorded as Unknow_Sig class with category label 5; normal EEG background signals were recorded as Norm_Sig class with category label 0; and all samples were divided into training samples (80%) and testing samples (20%).

[0084] The specific contents of step 2 are as follows:

[0085] 2-1. For each artifact signal type obtained in step 1, the average frequency spectral density (PSD) of the artifact signal was extracted. Interpolation plots were drawn and analyzed for each artifact type. Blink artifacts and frontal EMG signals were primarily located in the frontal region's Fp1 and Fp2 channels; chewing artifacts and temporal EMG signals were primarily located in the temporal region's F7, F8, T3, and T4 channels; and ambiguous EEG signals exhibited high amplitude and low frequency characteristics across the scalp, with a wide distribution across multiple channels, including Fp1, Fp2, F7, F8, T3, and T4, and across multiple electrodes. Based on the electrode locations at which these five types of signals (blink artifacts, frontal EMG signals, chewing artifacts, temporal EMG signals, and ambiguous EEG signals) appeared, the brain was divided into three regions: frontal region (Front_Region), temporal region (Temp_Region), and global region (Global_Region).

[0086] 2-2. Perform correlation analysis on the electrode channels of the regional information obtained in step 2-1 for different types of artifact signals to obtain correlations between the channels where different types of artifacts appear.

[0087] In the frontal region, blink artifacts showed high correlation in the Fp1 and Fp2 channels, followed by EMG in the frontal region. In the temporal region, chewing artifacts and EMG artifacts showed high correlation in F7, F8, T3, and T4. Ambiguous EEG signals (unconventional EEG background signals) showed high regional symmetry across all channels, with high correlation between left and right electrodes. Therefore, we focused on the Fp1 and Fp2 channels in the frontal region, the F7 and T4 channels in the temporal region, and the Fp1, F7, T3, T5, O1, F3, C3, P3, Fz, Cz, and Pz channels globally.

[0088] The specific contents of step 3 are as follows:

[0089] 3-1 Calculate the maximum value MAX, minimum value MIN, peak value Peak, peak-to-peak value P2P (peak to peak), mean value Mean, absolute mean AM, root amplitude RA, variance Var, standard deviation Std, root mean square RMS, kurtosis Kurt, skewness Ske, shape factor SF, peak factor PF1, pulse factor PF2, margin factor MF, clearance factor CF, median Median, fractal dimension FD, approximate entropy ApEn, sample entropy SaEn, fuzzy entropy FuEn, spectral entropy SpEn, average frequency AF, barycenter frequency BF, root mean square frequency RMSF frequency), standard deviation of frequency StdF (standard deviation of frequency), average power spectral density APSD (average power spectral density); the above total 28 dimensions of features;

[0090] Assume that each data sample format is S M*N (M: number of channels, N: data length);

[0091] 3-2 Calculate the maximum value MAX of different channel data using the following formula:

[0092] MAX i =max i (S i*N )

[0093] Where max i () represents the maximum value of the i-th channel to be solved.

[0094] 3-3 Calculate the minimum value MIN of different channel data using the following formula:

[0095] MIN i =min i(S i*N )

[0096] Where min i () represents the minimum value of the i-th channel to be solved.

[0097] 3-4 Calculate the peak value of different channel data using the following formula:

[0098] Peak i =max i (|S i*N |)

[0099] Where |S i*N | means finding the absolute value of the i-th channel.

[0100] 3-5 Calculate the peak-to-peak value P2P of different channel data using the following formula:

[0101] P2P i =max i (S i*N )-min i (S i*N )

[0102] 3-6 Calculate the mean of different channel data using the following formula:

[0103]

[0104] 3-7 Calculate the absolute average value AM of data from different channels using the following formula:

[0105]

[0106] 3-8 Calculate the root square amplitude RA of different channel data using the following formula:

[0107]

[0108] 3-9 Calculate the variance Var of different channel data using the following formula:

[0109]

[0110] In the formula Represents the expected value of all data in the i-th channel.

[0111] 3-10 Calculate the standard deviation Std of data from different channels using the following formula:

[0112]

[0113] Where Var i Represents the variance of the i-th channel calculated in 3-9.

[0114] 3-11 Calculate the root mean square (RSM) of data from different channels using the following formula:

[0115]

[0116] 3-11 Calculate the kurtosis of different channel data using the following formula:

[0117]

[0118] Where μ i represents the mean of all data in the i-th channel, σ i Indicates the standard deviation of all data in the i-th channel.

[0119] 3-12 Calculate the skewness Ske of different channel data using the following formula:

[0120]

[0121] 3-13 Calculate the waveform factor SF of different channel data using the following formula:

[0122]

[0123] RMS in the formula i Represents the root mean square of the i-th channel calculated in 3-11.

[0124] 3-14 Calculate the kurtosis factor PF1 of different channel data using the following formula:

[0125]

[0126] Peak in the formula i Indicates the peak value of the i-th channel calculated in 3-4.

[0127] 3-15 Calculate the pulse factor PF2 of different channel data using the following formula:

[0128]

[0129] AM in the formula i represents the absolute mean value of the i-th channel calculated in 3-7.

[0130] 3-16 Calculate the margin factor MF of different channel data using the following formula:

[0131]

[0132] RA in the formula i Indicates the square root amplitude of the i-th channel calculated in 3-8.

[0133] 3-17 Calculate the clearance factor CF of different channel data using the following formula:

[0134]

[0135] 3-18 Calculate the median of different channel data using the following formula:

[0136]

[0137] In the formula, sort(Si*N) means sorting all data of the i-th channel.

[0138] 3-19 The fractal dimension FD of different channel data is calculated using the following formula:

[0139]

[0140] Among them FD i (S,ε) represents the fractal dimension value, ε represents the grid side length when calculating the box-counting dimension, and G(ε) represents the cumulative number of grids of signal data.

[0141] 3-20 The approximate entropy ApEn of different channel data is calculated using the following formula:

[0142] ApEn i =φ m (S t*N )-φ m+1 (S t*N )

[0143]

[0144] In the formula Represents the probability measure of the signal sequence, φ m (S t*N ) is expressed as the logarithmic mean of the probability measure of the model sequence.

[0145] 3-21 Calculate the sample entropy SaEn of different channel data using the following formula:

[0146]

[0147] Where m represents the dimension of the reconstructed signal, δ represents the time delay, r represents the predefined threshold parameter, and n() represents the number of matching m-dimensional vector pairs.

[0148] 3-22 The fuzzy entropy FuEn of different channel data is calculated using the following formula:

[0149]

[0150] Where O n(m,r) represents the reconstructed n-1 dimensional vector, It represents the similarity of the defined fuzzy function relationship, m and r represent the gradient and width of the exponential function boundary respectively.

[0151] 3-23 Calculate the spectral entropy SpEn of different channel data using the following formula:

[0152]

[0153] Where G x () represents the transfer function that a given data sequence complies with.

[0154] 3-24 Calculate the average frequency AF of different channel data using the following formula:

[0155]

[0156] Where W represents the frequency of the given continuous data signal.

[0157] 3-25 Calculate the center of gravity frequency BF of different channel data using the following formula:

[0158]

[0159] 3-26 Calculate the root mean square frequency (RMSF) of data from different channels using the following formula:

[0160]

[0161] 3-27 The frequency standard deviation StdF of the data in different channels is calculated using the following formula:

[0162]

[0163] Where BF i represents the centroid frequency of the ith channel.

[0164] 3-28 The average power spectral density APSD of different channel data is calculated using the following formula:

[0165] APSD i =PSD i (AF i )

[0166]

[0167] Where PSD i represents the power spectral density of the i-th channel, fft(*) represents the fast Fourier transform of the signal, fs represents the sampling frequency, AF i represents the average frequency of the i-th channel.

[0168] The specific contents of step 4 are as follows:

[0169] 4-1 Use the ReliefF algorithm to select the 28-dimensional features extracted from each type of artifact to reduce feature redundancy for the next step, reduce feature reuse, and improve the model's resolution efficiency. The feature weight selection of the ReliefF algorithm is shown in the following formula:

[0170]

[0171] Where P(C) represents the proportion of the class of features in the total population, P(class(R)) represents the proportion of a randomly selected sample class, A represents the currently selected feature, R and H represent two different samples in feature A, diff(A, R, H) represents the difference between the two samples in feature A, and M j (C) represents the jth nearest neighbor sample in class C, m represents the number of sampling times, and k represents the number of nearest neighbor samples;

[0172] 4-2 The mRMR method is used to perform secondary feature selection on the artifact features contained in the regions divided by different artifacts, construct the most significant features between regions, and filter out the most significant features of different types of artifacts within the region. The mRMR algorithm is shown in the following formula:

[0173] Maximum correlation:

[0174]

[0175] Where x i represents the i-th type of feature, c represents the target classification, L represents the feature subset consisting of the selected m features, I(x i ; c) represents the mutual information between the i-th feature and the target category c.

[0176] Minimum redundancy:

[0177]

[0178] MrMr weight combination algorithm:

[0179]

[0180] Among them, X represents the set of features contained in the given data, L m-1 Represents the (m-1) feature subsets selected from a given feature subset.

[0181] The present invention adopts a two-stage feature selection method as follows: (1) the ReliefF algorithm is used to screen out candidate feature sets for different classifiers; (2) the mRMR algorithm is used to directly remove the redundancy of the feature candidate set obtained by the ReliefF algorithm to screen out the best feature grouping for the trained classifier.

[0182] 4-3. Based on the different feature groups analyzed in steps 4-1 and 4-2, establish an N1*M*chs-dimensional feature group DFs for distinguishing brain regions where artifacts occur; establish an N1*M*chs-dimensional feature group EFs for different types of artifacts in the frontal region; establish an N1*M*chs-dimensional feature group TFs for different types of artifacts in the temporal region; and establish an N1*M*chs-dimensional feature group GFs for the entire brain. Where N1 represents the number of samples, M represents the number of selected channels, and chs represents the number of selected channels.

[0183] The specific contents of step 5 are as follows:

[0184] 5-1 Based on the feature combinations and sample labels extracted from different artifacts and brain regions in step 4, a machine learning algorithm is used to train the initial divided training set to obtain multiple classifier models;

[0185] 5-2 Use the trained classifier model to test the initially divided test set data, calculate the test accuracy, modify the classifier model parameters, and obtain the classifier model with the best performance. Use the feature group DFs to train a classifier for adaptively discriminating brain regions, recorded as Classifier0; use the feature group EFs to train a classifier model for discriminating artifacts in the frontal region, recorded as Classifier1 and Classifier1' respectively; use the feature group TFs to train a classifier model for discriminating artifacts in the temporal region, recorded as Classifier2 and Classifier2' respectively; use the feature group GFs to train a classifier model for discriminating global regions, recorded as Classifier3.

[0186] The specific contents of step 6 are as follows:

[0187] 6-1 Build the final multi-classifier fusion adaptive brain region artifact detection system based on the classifiers trained for different regions in step 5. Since the original EEG data is continuous, a 1s data segment is input each time, and the classifier Classifier0 is used to adaptively determine the brain region of the input data. After obtaining the regional information, the classifier corresponding to each region is used to extract the signal features for artifact detection.

[0188] 6-2 Using unprocessed raw EEG data to test the system, it can accurately identify different types of artifacts in continuous data.

[0189] In order to achieve better EEG data multi-artifact recognition results, the following will introduce the parameter settings and design aspects of actual applications, which can serve as a reference for other applications of this invention.

[0190] When using this method to process multi-channel EEG data, the collected EEG data needs to be filtered with a 50Hz notch filter and a 1Hz-70Hz bandpass filter. Because artifacts occur intermittently, the combined duration of each artifact of different types does not exceed 1s. Therefore, when inputting data, each segment is 1s long. To prevent missed detections, a 50% overlap rate is used for data reading to reduce the missed detection rate. In step 3, a total of 28 dimensional features are extracted from the EEG data, including time domain, frequency domain, and nonlinear features. By analyzing these features, the significance ranking of different types of artifacts is obtained, and the appropriate feature group is selected as the main basis for training the classifier model.

[0191] During feature selection, both the ReliefF algorithm and the mRMR algorithm were used for two-stage feature selection and processing, establishing a feature set with high sensitivity to different types of artifacts. Different classifier models were compared, and the model with the highest accuracy and best stability was selected as the final system classifier. The parameters of the classifiers used in the experiment were dynamically adjusted during training. The number of clusters, K, in the K-nearest neighbor classification model (KNN) did not exceed 20, and was not an even number. The decision tree of the random forest classification model (RF) was dynamically adjusted to achieve optimal accuracy. The naive Bayes classification model (NB) used internal default adjustment parameters, and the support vector machine classification model (SVM) established standard template parameters. During model training, internal parameters were dynamically adjusted based on the input data to achieve optimal results.

[0192] Among the models trained in step 5, Classifier0 uses a support vector machine (SVM) model with a test set accuracy of 95.83%, Classifier1 uses a support vector machine (SVM) model with a test set accuracy of 99.83%, Classifier1' uses a random forest (RF) model with a test set accuracy of 96.93%, Classifier2 uses a random forest (RF) model with a test set accuracy of 93.12%, Classifier2' uses a naive Bayes (NB) model with a test set accuracy of 99.49%, and Classifier3 uses a naive Bayes (NB) model with a test set accuracy of 98.15%. Each classifier has high discrimination performance and can achieve relatively accurate multi-artifact recognition.

[0193] like Figure 3 As shown, the obtained classifier is Figure 2 The final system model is obtained after fusing multiple classifiers in a fusion manner. A two-minute EEG data of a real patient is tested on this model, wherein the actual labels are marked by professional EEG technicians, and the predicted number is the number of different artifacts obtained using the model of the present invention. The final recognition accuracy is obtained by comparing the manually marked and model-predicted label data. It can be seen that different types of artifacts have a high recognition rate, and the processing efficiency can meet the requirements of real-time analysis.

[0194] The fusion method of the classifiers used in the classifier fusion system proposed in the present invention also helps to reduce the complexity of classification and improve the accuracy of classification, so as to realize real-time detection of artifact signals in EEG data, which can help doctors quickly determine the artifact type of the recorded EEG data and reduce the interference of artifacts on diagnosis in clinical medicine.

Claims

1. The adaptive brain EEG artifact detection method based on multi-classifier fusion is characterized by The steps include: Step 1: Filter the multi-channel EEG signals and classify them into multiple artifact categories; the multiple artifact categories include blink artifact signals, frontal EEG artifact signals, chewing artifact signals, temporal EEG artifact signals, and EEG signals with unclear meanings; Step 2: Analyze the brain regions and channel correlations corresponding to each type of artifact signal obtained in step 1 to obtain the regional information and channel information of each type of artifact; Step 3: Extract features from the channel information of each type of artifact, extracting a total of 28 dimensions of features; Step 4: Based on the regional information divided in step 2 and the features extracted in step 3, the ReliefF algorithm and the mRMR algorithm are used to perform two-stage feature selection and establish feature groups; Step 5: Use the machine learning algorithm combined with the features selected in step 4 to train the classification model; Step 6: Build a system based on the multiple classifiers trained in step 5.

2. The adaptive brain region EEG artifact detection method based on multi-classifier fusion according to claim 1 is characterized in that The specific contents of step 1 are as follows: The 19-channel, 1000Hz raw EEG data were filtered, and each channel was subjected to a 50Hz notch filter and a 1Hz-70Hz bandpass filter to obtain EEG data without non-physiological interference. Due to the discontinuity of artifacts in the EEG data, the artifact data was segmented using a sliding window method, and the time length of each sliding window was set to 1s. Signals containing blink artifacts were recorded as the Eye_Blink class with a category label of 1; signals containing frontal EMG were recorded as the Frontal_EMG class with a category label of 2; signals containing chewing artifacts were recorded as the Chew_EMG class with a category label of 3; signals containing temporal EMG were recorded as the Temp_EMG class with a category label of 4; signals with unconventional EEG background were recorded as the Unknow_Sig class with a category label of 5; and normal EEG background signals were recorded as the Norm_Sig class with a category label of 0. The dataset was divided into training samples and test samples.

3. The method for detecting EEG artifacts in an adaptive manner using multiple classifiers according to claim 1 or 2, characterized in that The specific contents of step 2 are as follows: 2-1. For each type of artifact signal obtained in step 1, the average frequency spectral density of the artifact signal is extracted. By drawing an interpolation graph and analyzing the interpolation graph of each type of artifact, it is found that the blink artifact signal and the frontal EMG signal mainly appear in the Fp1 and Fp2 channels of the frontal region; the chewing artifact signal and the temporal EMG signal mainly appear in the F7, F8, T3, and T4 channels of the temporal region; the ambiguous EEG signal exhibits high amplitude and low frequency characteristics across the entire scalp and is widely distributed, covering multiple channels including Fp1, Fp2, F7, F8, T3, and T4 channels, and affecting multiple electrodes; the brain is divided into three regions based on the electrode locations where the five types of signals, namely, the blink artifact signal, the frontal EMG signal, the chewing artifact signal, the temporal EMG signal, and the ambiguous EEG signal, appear: the frontal region, the temporal region, and the global region; 2-2. Performing correlation analysis on the electrode channels of the regional information obtained in step 2-1 for different types of artifact signals to obtain correlations between the channels where different types of artifacts appear; In the frontal area, blink artifacts showed high correlation in the Fp1 and Fp2 channels, and the correlation in the frontal area electromyography in the Fp1 and Fp2 channels was second; in the temporal area, chewing artifacts and electromyography artifacts showed high correlation between F7 and F8, T3 and T4; the EEG signals with unclear meaning showed high brain area symmetry in all channels, and both the left and right brain electrodes showed high correlation; therefore, Fp1 and Fp2 channels were selected in the frontal area; F7 and T4 channels were selected in the temporal area; and Fp1, F7, T3, T5, O1, F3, C3, P3, Fz, Cz, and Pz channels were selected in the global area as the focus of the study.

4. The method for adaptive brain region EEG artifact detection based on multi-classifier fusion according to claim 3, characterized in that The specific contents of step 3 are as follows: 3-1. Calculate the maximum value MAX, minimum value MIN, peak value Peak, peak-to-peak value P2P, mean value Mean, absolute mean value AM, root mean amplitude RA, variance Var, standard deviation Std, root mean square RMS, kurtosis Kurt, skewness Ske, form factor SF, peak factor PF1, pulse factor PF2 (pulse factor), margin factor MF, clearance factor CF, median Median, fractal dimension FD, approximate entropy ApEn, sample entropy SaEn, fuzzy entropy FuEn, spectral entropy SpEn, average frequency AF, center of gravity frequency BF, root mean square frequency RMSF, frequency standard deviation StdF, average frequency spectral density APSD of all channels of the sample; the above totals 28 dimensions of features; Assume that each data sample format is S M*N , M: number of channels, N: data length; 3-2 Calculate the minimum value MIN of different channel data using the following formula: MY i =my i (S i*N ) Where min i () represents the minimum value of the i-th channel to be solved; 3-3 Calculate the peak value of different channel data using the following formula: Peak i =max i (|S i*N |) Where |S i*N | means to find the absolute value of the i-th channel, max i () represents the maximum value of the i-th channel; 3-4 Calculate the fractal dimension FD of different channel data using the following formula: Among them FD i (S,ε) represents the fractal dimension value, ε represents the grid side length when calculating the box-counting dimension, G(ε) represents the cumulative grid number of signal data, and S represents each sample data; 3-5 Calculate the approximate entropy ApEn of different channel data using the following formula: ApEn i =φ m (S t*N )-φ m+1 (S t*N ) In the formula Represents the probability measure of the signal sequence, φ m (S t*N ) is expressed as the logarithmic mean of the probability measure of the signal sequence; 3-6 Calculate the spectral entropy SpEn of different channel data using the following formula: Where G x () represents the transfer function that a given data sequence complies with; 3-7 Calculate the average frequency AF of different channel data using the following formula: Where W represents the frequency of the given continuous data signal; 3-8 Calculate the average power spectral density APSD of different channel data using the following formula: NIV i =PSD i (AF i ) Where PSD i represents the power spectral density of the i-th channel, fft( ) represents the fast Fourier transform of the signal, fs represents the sampling frequency, AF i represents the average frequency of the i-th channel.

5. The method for adaptive brain region EEG artifact detection based on multi-classifier fusion according to claim 4 is characterized in that The specific contents of step 4 are as follows: 4-1. Use the ReliefF algorithm to perform a one-stage feature selection on the 28-dimensional features extracted for each type of artifact. The feature weight selection of the ReliefF algorithm is shown in the following formula: Where p(C) represents the proportion of the class of features in the total population, p(class(R)) is the proportion of a randomly selected sample class, A represents the currently selected feature, R and H represent two different samples in feature A, diff(A, R, H) represents the difference between two different samples R and H on feature A, and M j (C) represents the jth nearest neighbor sample in class C, m represents the number of sampling times, k represents the number of nearest neighbor samples, and C≠class(R); 4-2. Using the RelifF algorithm in 4-1, we rank the features of artifact information in different regions and perform secondary feature selection using the mRMR method to construct the most significant feature combination between different artifact information in different regions. The mRMR algorithm is shown in the following formula: Maximum correlation: Where x i represents the i-th type of feature, c represents the target classification, L represents the feature subset consisting of the selected m features, I(x i ; c) represents the mutual information between the i-th feature and the target category c; Minimum redundancy: mRMR weight combination algorithm: Among them, x represents the set of features contained in the given data, L m-1 Represents the (m-1) feature subsets selected from a given feature subset; 4-3. Based on the different feature groups analyzed in steps 4-1 and 4-2, establish a feature group DFs of N1*M*chs dimensions to distinguish the brain regions where artifacts appear; establish a feature group EFs of N1*M*chs dimensions for different types of artifacts in the frontal region; establish a feature group TFs of N1*M*chs dimensions for different types of artifacts in the temporal region; and establish a feature group GFs of N1*M*chs dimensions for the entire brain; where N1 represents the number of samples, M represents the number of selected channels, and chs represents the number of selected channels.

6. The method for adaptive brain region EEG artifact detection based on multi-classifier fusion according to claim 5, characterized in that The specific contents of step 5 are as follows: 5-1 Based on the feature groups extracted from different artifacts and brain regions in step 4 and the sample labels they carry, a machine learning algorithm is used to train the initial divided training set to obtain multiple classifier models; 5-2 Use the trained classifier model to test the initially divided test set data, calculate the test accuracy, modify the parameters of the classifier model, and obtain the classifier model with the best performance; The feature group DFs was used to train a classifier for adaptively discriminating brain regions, which was recorded as Classifier0. The feature group EFs was used to train a classifier model for discriminating artifacts in the frontal region, which were recorded as Classifier1 and Classifier1' respectively. The feature group TFs was used to train a classifier model for discriminating artifacts in the temporal region, which were recorded as Classifier2 and Classifier2' respectively. The feature group GFs was used to train a classifier model for discriminating global regions, which was recorded as Classifier3.

7. The method for adaptive brain region EEG artifact detection based on multi-classifier fusion according to claim 6, characterized in that The specific contents of step 6 are as follows: 6-1 Build the final multi-classifier fusion adaptive brain region artifact detection system based on the classifiers trained for different regions in step 5. Since the original EEG data is continuous, each time a 1-second data segment is input, classifier Classifier0 is used to adaptively determine the brain region of the input data. After obtaining the regional information, the classifier corresponding to each region is used to extract the signal features for artifact detection; 6-2 Using unprocessed raw EEG data to test the system, it can accurately identify different types of artifacts in continuous data.

Images