An abnormal electroencephalogram automatic detection method based on multi-domain feature fusion

Abstract

The application relates to the technical field of brain waves, and discloses an abnormal brain wave automatic detection method based on multi-domain feature fusion, which collects brain wave signals; the acquired brain wave signals are labeled as abnormal or normal by brain wave experts, that is, if abnormal waveforms exist in the brain wave signals, the brain waves are classified as the abnormal class, a training set and a test set are established, and then, signal pretreatment is realized through methods such as downsampling and signal segmentation. The abnormal brain wave automatic detection method based on multi-domain feature fusion is applied to the binary classification of brain wave signals, solves the limitations of traditional brain wave abnormality detection methods based on single-domain features, and improves detection performance. Meanwhile, the strategy based on multi-scale feature fusion is used to realize feature reduction, which not only effectively reduces information loss caused by feature aggregation, but also reduces feature redundancy, and has strong generalization and robustness.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of brain waves, in particular to an abnormal brain wave automatic detection method based on multi-domain feature fusion. BACKGROUND

[0002] Bioelectricity is one of the basic characteristics of life activities, and brain waves are one of the many bioelectricities. When a human being is thinking, the bioelectricity generated by the brain is the brain wave, which can be measured and studied through a sensor placed on the scalp. At present, the brain wave technology has been widely applied in many fields, including clinical detection, behavior disorder analysis, concentration assessment, etc. Traditionally, brain wave experts realize the diagnosis related to the target task by analyzing some important components in the brain wave signal. In this process, one of the core steps is to judge whether there is an abnormal waveform in the brain wave signal. In practice, an analysis process based on a complex decision tree is often used by brain wave experts to judge whether the brain wave signal belongs to the normal or abnormal class; this is not only time-consuming and laborious, but also susceptible to low consistency between evaluators. Therefore, an intelligent brain wave abnormality automatic detection method without human intervention is urgently needed to assist in analysis.

[0003] At present, researchers have proposed a large number of brain wave abnormality analysis methods based on machine learning. The implementation process of these methods can be roughly divided into two main stages, namely feature extraction and classification based on features. The former plays a crucial role in the classification of abnormal brain wave signals, because the quality of the extracted features will directly affect the performance of the subsequent classification task. Common extraction methods include frequency domain analysis, time domain analysis and time-frequency analysis, etc. Time domain analysis refers to directly using statistical parameters to measure meaningful signal segments from the original brain wave signal as the basic features of brain waves, such as mean, variance, maximum, minimum, peak, etc. Frequency domain analysis technology extracts relevant frequency domain features using the frequency spectrum information of the signal, including discrete Fourier transform, power spectral density, autoregressive model-based frequency spectrum estimation, etc. With the development of signal processing technology, time-frequency analysis technology has been widely concerned since it was proposed, which can effectively provide the amplitude and phase of different frequency components of the brain wave signal with time, and replace the single scale with dual resolution of time and frequency. Representative methods include discrete wavelet transform (DWT), wavelet packet decomposition (WPD), etc.

[0004] The above signal processing methods have their own advantages and limitations. Time domain or frequency domain analysis method is intuitive and easy to understand, and using it can effectively reduce the influence of noise on the original signal, but using a single analysis method to extract the electroencephalogram feature has the problems of single feature and insufficient information description, resulting in poor classification effect. If only the time domain feature of electroencephalogram is considered, the valuable frequency information is ignored. On the contrary, if only the frequency domain feature is considered, the potential correlation between signal frequency and time domain and other beneficial information are ignored. Obviously, the existing electroencephalogram abnormality binary classification method based on time domain or frequency domain analysis ignores the complementarity between different domain feature information. In addition, the extracted electroencephalogram feature has high dimension, and there are redundant and irrelevant features, which causes huge calculation overhead, affects the real-time performance of prediction, and greatly reduces the classification performance. Therefore, an abnormal electroencephalogram automatic detection method based on multi-domain feature fusion is proposed, which is used for identifying the collected electroencephalogram signal as a normal class or an abnormal class. The normal class means that there is no known abnormal waveform in the signal, and the abnormal class is the opposite. The new method effectively solves the problem of low electroencephalogram signal recognition rate caused by single feature extraction, and a new feature aggregation technology is designed to extract significant features, thereby improving the accuracy and efficiency of classification.

[0005] Defects and deficiencies of the prior art:

[0006] The existing electroencephalogram signal abnormality classification technology has the following defects:

[0007] (1) Only single domain feature extraction is considered, and the complementarity between different domain features is ignored, which limits the classification accuracy;

[0008] (2) The feature redundancy caused by the increase of the dimension of the extracted electroencephalogram feature is not considered, which reduces the classification performance;

[0009] (3) The electroencephalogram signal has close correlation with the physiological information (such as age and gender) of the patient, and the existing classification method ignores these important information.

[0010] In view of the above problems, an abnormal electroencephalogram automatic detection method based on multi-domain feature fusion is proposed. SUMMARY

[0011] In view of the above defects of the prior art, the present application provides an abnormal electroencephalogram automatic detection method based on multi-domain feature fusion, to determine whether there is an abnormal waveform in the brain activity signal, and to provide accurate pre-analysis for subsequent specific electroencephalogram tasks.

[0012] The present application provides the following technical scheme: an abnormal electroencephalogram automatic detection method based on multi-domain feature fusion, comprising the following steps:

[0013] S1: Collecting electroencephalogram signals; labeling the obtained electroencephalogram signals as abnormal or normal, i.e. if there is an abnormal waveform in the electroencephalogram signal, the electroencephalogram signal is classified as an abnormal class, a training set and a test set are established, and then signal preprocessing is realized through downsampling, signal segmentation and the like;

[0014] S2: Extracting time domain features of each channel of the preprocessed electroencephalogram signal segment by using a time domain signal processing method;

[0015] S3: Extracting frequency domain features of each signal segment after the preprocessed electroencephalogram signal is transformed from the time domain to the frequency domain based on a frequency domain signal processing method;

[0016] S4: Extracting time-frequency domain features of each channel of the preprocessed electroencephalogram signal segment by using a time-frequency signal processing technique;

[0017] S5: Standardizing the extracted electroencephalogram features by using a standardization method;

[0018] S6: Based on the strategy of multi-scale feature fusion, the standardized time domain, frequency domain and time-frequency domain features extracted in steps 2, 3 and 4 are respectively aggregated, and a one-dimensional feature vector is constructed; at the same time, the physiological information of the patient is added to the one-dimensional vector as an additional feature;

[0019] S7: Using the multi-domain fusion feature space obtained in step 6 as the input of the integrated learning classifier, performing classification training to obtain an electroencephalogram abnormality classification model;

[0020] S8: Testing the multi-domain feature space to which the test set established in step 7 belongs by using the trained classification model to obtain a classification result.

[0021] Preferably, in step 2, the time domain features of the preprocessed signal are extracted by using a time domain signal processing method, which specifically comprises:

[0022] I. Using a p-order autoregressive model to fit the electroencephalogram signal, which can be expressed as:

[0023]

[0024] wherein, represents a white noise sequence with a mean of zero and a variance of p is the order of the autoregressive model, is the coefficient of the p-order autoregressive model;

[0025] II. Calculating the autoregressive model coefficient by using the Burg algorithm as the time domain feature vector of the electroencephalogram signal of each channel.

[0026] Preferably, in step 3, the time-frequency domain features of the preprocessed electroencephalogram signal are extracted by using a time-frequency signal processing technique, which specifically comprises:

[0027] I. Decompose the EEG signal into a series of high and low frequency subbands using discrete wavelet transform, and calculate the corresponding discrete wavelet coefficients;

[0028] II. According to the threshold criterion, eliminate the discrete wavelet coefficients less than the threshold, and retain the remaining important discrete wavelet coefficients;

[0029] III. Extract a set of statistical features from each important wavelet coefficient that can significantly improve the detection performance.

[0030] Preferably, the multi-scale based feature fusion strategy in step 6 is specifically:

[0031] I. Construct a feature matrix in time sequence according to the EEG signal feature sequence ; wherein, represents the n-th EEG data sample in the N sample data set, C represents the number of channels of the collected EEG signal, E represents the total number of cut EEG segments, and F represents the total number of statistical features extracted from each EEG segment;

[0032] II. Aggregate the EEG features extracted from each segment in the time direction using the mean as the aggregation function, wherein the local aggregation process is: cutting the extracted features into two parts according to the time direction, and then aggregating the EEG features in the time direction and channel by channel; the global aggregation is to aggregate the features in the time direction and channel by channel after taking the extracted features as a whole;

[0033] III. Connect the EEG features obtained by local aggregation and global aggregation, and construct the corresponding feature space , wherein, represents the n-th EEG data sample in the total N sample data set, C represents the number of channels of the collected EEG signal, 3 represents the number of features obtained by executing the proposed feature aggregation method after extracting features from each channel, and F represents the total number of statistical features extracted from each EEG segment;

[0034] IV. Flatten the aggregated feature space into a one-dimensional vector , the size of H is .

[0035] Compared with the prior art, the present application has the following beneficial effects:

[0036] The abnormal electroencephalogram automatic detection method based on multi-domain feature fusion is applied to the binary classification of electroencephalogram signals, solves the limitations of traditional electroencephalogram abnormality detection methods based on single-domain features, and improves the detection performance. Meanwhile, the strategy based on multi-scale feature fusion is used to realize feature reduction, which not only effectively reduces the information loss caused by feature aggregation, but also reduces the redundancy of features, and has strong generalization and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The detection flowchart of the application is shown in the figure.

[0038] Figure 2 The 5th order discrete wavelet decomposition diagram of the application is shown in the figure. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme of the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings of the embodiments of the present disclosure. In order to keep the following description of the embodiments of the present disclosure clear and concise, the present disclosure omits the detailed description of known functions and known components to avoid unnecessary confusion of the concept of the present application.

[0040] Please refer to Figure 1 An abnormal electroencephalogram automatic detection method based on multi-domain feature fusion, characterized by comprising the following steps:

[0041] S1, collecting electroencephalogram signals; the acquired electroencephalogram signals are labeled as abnormal or normal, i.e. if there are abnormal waveforms in the electroencephalogram signals, the electroencephalogram signals are classified as abnormal class, a training set and a test set are established, and then signal preprocessing is realized by downsampling, signal segmentation and other methods; the specific operation is as follows:

[0042] I. Collect electroencephalogram data from different patients using the standard 10 / 20 electrode system, and the duration of collection should be more than 15 minutes.

[0043] II. Resample the retained electroencephalogram signals at a sampling rate of 250 Hz to ensure data consistency.

[0044] III. Divide the electroencephalogram signals of each channel into 100 equal-sized signal segments using a 5-second non-overlapping sliding window, and discard the remaining signal data.

[0045] S2, use time domain signal processing method to extract the time domain features of the preprocessed electroencephalogram signal segments of each channel; the specific operation is as follows:

[0046] I. Use a p-order autoregressive model to fit the electroencephalogram signals, which can be expressed as:

[0047]

[0048] wherein, represents a white noise sequence with zero mean and variance p is the order of the autoregressive model, are the coefficients of the p-th order autoregressive model.

[0049] II. Calculate the autoregressive model coefficients using the Burg algorithm as the time domain feature vector of each channel electroencephalogram signal.

[0050] S3. Convert the preprocessed electroencephalogram signal from the time domain to the frequency domain based on the frequency domain signal processing method, and use the Welch method to extract the frequency domain features of each channel electroencephalogram signal segment, specifically:

[0051] I. Transform the electroencephalogram signal from the time domain to the frequency domain by Fourier transform, and study the frequency spectrum structure and variation law of the signal.

[0052] II. Calculate the power spectral density of the electroencephalogram signal using the Welch method, and process each channel electroencephalogram segment with a 500 millisecond window and a 25% overlap using a Hanning window.

[0053] III. Use the average power, mean, standard deviation, skewness, kurtosis, and energy ratio of the power spectral density of each channel to measure the characteristics of each frequency component, i.e. the frequency domain feature vector of each channel electroencephalogram signal segment is 6-dimensional.

[0054] S4. Extract the time-frequency domain features of the electroencephalogram signal using the discrete wavelet transform with Sym6 as the mother wavelet. Specifically:

[0055] I. Use a complementary low-high pass filter bank to perform multi-level decomposition on the input signal. The low pass filter and high pass filter output of each level correspond to the approximation coefficients and detail coefficients, respectively; the generated approximation coefficients are further divided into approximation coefficients and detail coefficients of the next level. Repeat this process until the preset decomposition level is reached.

[0056] II. Since the sampling rate of the preprocessed signal is 250 Hz, in order to filter out irrelevant information, the number of decomposition levels in this example is set to 5. (As shown in Figure 2 The final 5 detail coefficients D1, D2, D3, D4, D5 are obtained. Their frequency ranges are 62.5-125Hz, 31.25-62.5Hz, 15.525-31.25Hz, 7.8125-15.625Hz, 3.90625Hz-7.8125Hz, respectively, and one approximation coefficient A5, see Figure 2 Since the Delta, Theta, Alpha and Beta rhythms contain significant discriminative information, the detail coefficients D3, D4, D5 and the approximation coefficient A5 are retained.

[0057] III, Statistics of each selected wavelet coefficient, average value, average value of absolute deviation, standard deviation, average value of absolute value, skewness and kurtosis, as the time-frequency domain features of the decomposition signal. Here, the time-frequency domain features of each channel signal segment are 6x4=24 dimensions; 6 is the number of statistical features extracted from each coefficient, and 4 is the total number of selected wavelet coefficients.

[0058] S5, Standardize the time domain, frequency domain, and time-frequency domain feature vectors using the Z-Score standardization method, so that each feature vector has zero mean and unit variance, thereby obtaining a set of dimension-consistent features.

[0059] S6, Based on the strategy of multi-scale feature fusion, the time domain, frequency domain, and time-frequency domain features of the EEG signal are fused. The specific process is:

[0060] I, For each channel of the EEG signal segment, the 7-dimensional time domain features extracted in the second step, the 6-dimensional frequency domain features in the third step, and the 24-dimensional time-frequency domain features extracted in the fourth step are constructed into a new feature matrix in time order ; Where, n represents the nth sample in the collected data set, C represents the number of EEG channels, S represents the total number of segments, and 37 represents the total number of features extracted from each signal segment.

[0061] II, Aggregate the EEG features extracted from each segment in the time direction using the mean as the aggregation function, where the local aggregation process is: cut the extracted features into two parts in the time direction, and then aggregate the EEG features in each channel in the time direction; Global aggregation is to aggregate the features in each channel in the time direction after considering the extracted features as a whole.

[0062] III, Connect the EEG features obtained by local aggregation and global aggregation to construct the corresponding feature space , where C represents the number of channels of the EEG signal, 37 represents the total number of statistical features extracted from each EEG segment, and 3 represents the number of aggregated features, which consists of two local features and one global feature.

[0063] S7, Flatten the feature space into a one-dimensional vector and input it into the integrated learning classification method for training to obtain an EEG abnormality classification model.

[0064] S8, Test the multi-domain feature space belonging to the test set established in step 7 using the trained classification model to obtain the classification result.

[0065] The above examples are only exemplary embodiments of the present application, and are not intended to limit the present application, and the protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements are also considered to fall within the protection scope of the present application.

Claims

1. An automatic detection method for abnormal electroencephalograms based on multi-domain feature fusion, characterized in that, Includes the following steps: S1: Collect EEG signals; label the acquired EEG signals as abnormal or normal, that is, if there are abnormal waveforms in the EEG signals, the EEG signals are classified as abnormal, establish training sets and test sets, and then perform signal preprocessing through downsampling and signal segmentation methods; S2: Use time-domain signal processing methods to extract the time-domain features of the EEG signal segments of each channel after preprocessing; S3: Based on frequency domain signal processing methods, the preprocessed EEG signal is transformed from the time domain to the frequency domain, and the frequency domain features of each signal segment are extracted; S4: Use time-frequency signal processing techniques to extract the time-frequency domain features of the EEG signal segments from each channel after preprocessing; S5: Standardize the extracted EEG features using a standardization method; S6: Based on the multi-scale feature fusion strategy, the standardized time-domain, frequency-domain, and time-frequency-domain features extracted in steps two, three, and four are aggregated and a one-dimensional feature vector is constructed. At the same time, the patient's physiological information is added as an additional feature to the one-dimensional vector. S7: Use the multi-domain fusion feature space obtained in step 6 as the input of the ensemble learning classifier to perform classification training and obtain the EEG abnormality classification model. S8: Use the trained classification model to test the multi-domain feature space to which the test set established in step 7 belongs, and obtain the classification result.

2. The automatic detection method for abnormal EEG based on multi-domain feature fusion according to claim 1, characterized in that: In step 2, the extraction of the preprocessed time-domain features of the signal using time-domain signal processing methods is specifically as follows: I. Using a p-order autoregressive model to fit the EEG signal, it can be expressed as: ； in, This indicates that the mean is zero and the variance is... The white noise sequence; p is the order of the autoregressive model. The coefficients of the p-order autoregressive model; II. Calculating Autoregressive Model Coefficients Using the Burg Algorithm As the temporal feature vector of the EEG signal in each channel.

3. The automatic detection method for abnormal EEG based on multi-domain feature fusion according to claim 1, characterized in that: Step 4 involves using time-frequency signal processing techniques to extract the time-frequency domain features of the preprocessed EEG signal, specifically as follows: I. The EEG signal is decomposed into a series of high and low frequency sub-bands using discrete wavelet transform, and the corresponding discrete wavelet coefficients are calculated. II. Based on the threshold criterion, remove discrete wavelet coefficients that are less than the threshold and retain the remaining important discrete wavelet coefficients. III. Extract a set of statistical features from each important wavelet coefficient that can significantly improve detection performance.

4. The automatic detection method for abnormal EEG based on multi-domain feature fusion according to claim 1, characterized in that: The multi-scale feature fusion strategy in step 6 is as follows: I. Construct a feature matrix from the EEG signal feature sequences in chronological order. ;in, This represents the nth EEG data sample in the N samples of the dataset, C represents the number of channels for acquiring EEG signals, E represents the total number of EEG segments after segmentation, and F represents the total number of statistical features extracted from each EEG segment. II. The EEG features extracted from each segment are aggregated according to the time direction using the mean as the aggregation function. The local aggregation process is as follows: the extracted features are cut into two parts in half according to the time direction, and then the EEG features are aggregated channel by channel according to the time direction. The global aggregation process is to treat the extracted features as a whole and then aggregate the features channel by channel according to the time direction. III. Connect the EEG features obtained from local and global aggregation, and construct the corresponding feature space. ,in, This represents the nth EEG data sample out of the total N samples in the dataset. C represents the number of channels for acquiring EEG signals. 3 represents the number of features obtained after performing the proposed feature aggregation method after feature extraction for each channel. F represents the total number of statistical features extracted from each EEG segment. IV. Flatten the aggregated feature space into a one-dimensional vector. The size of H is .

Images