CN110811609A - Intelligent epileptic spike detection method based on fusion of adaptive template matching and machine learning algorithm - Google Patents

Abstract

The present invention provides an intelligent detection method for epilepsy spike waves based on the fusion of adaptive template matching and machine learning algorithm. (2) data preprocessing: band-pass filtering the collected raw EEG data to obtain standard EEG signals; (3) adaptive template matching spike detection: (4) machine learning-based spike detection Wave detection method: firstly, the EEG signal is divided into 1s-long EEG segments, and then the time domain and frequency domain features in each EEG segment are extracted to construct a spike wave feature vector; using the feature vector to train a random forest classification model, we get Machine learning-based spike detection results. (5) Fusion of detection results: the detection methods of step S3 and step S4 are fused, and if a spike is detected by S3 and S4 at the same time, it is regarded as an epileptic spike.

Description

Intelligent detection of epilepsy spikes based on fusion of adaptive template matching and machine learning algorithm test method

technical field

The invention relates to the field of computers, in particular to an intelligent detection method for epilepsy spikes based on the fusion of adaptive template matching and machine learning algorithms.

Background technique

Epilepsy is a chronic disease in which the sudden abnormal discharge of brain neurons leads to transient brain dysfunction. More than 65 million people worldwide suffer from epilepsy, and there are about 9 million people with epilepsy in China. Seizures are sudden, repetitive, and unpredictable, and can occur at any age.

Electroencephalogram (EEG) is a potential signal generated by the firing of neurons in the brain, which reflects the rhythmic activity of the brain's bioelectricity and contains a large amount of physiological and disease information. In clinical medicine, EEG signal processing can not only provide the basis for the diagnosis of certain brain diseases, but also provide effective treatment methods for certain brain diseases, and play an important role in the detection of epilepsy.

The spike wave is a typical epilepsy characteristic waveform, which is usually recorded in the EEG. Compared with the background waveform, the spike wave waveform is sharp, with high amplitude and transient characteristics. The current clinical epilepsy examination is mainly identified by human eye detection. EEG spikes. At present, the clinical examination of epilepsy EEG is mainly to identify spikes in EEG signals through manual detection, but its efficiency is low and subjectivity is strong, and the accuracy of the results cannot be guaranteed. more and more attention.

There are many methods to identify spikes. The commonly used method is wavelet analysis to decompose the time-frequency features of epilepsy EEG signals, and use wavelet coefficients as the input signal of machine learning classifiers and neural networks to detect spikes. However, due to the difference in the characteristics of mother wavelet and spike wave, in the signal obtained by decomposition and reconstruction, the background EEG suppression is poor, and the effect of spike wave extraction is not ideal. The signal singularity detection method of wavelet transform mode maximum is another common method for spike detection, but this method can only detect positive-phase spikes, and the false-positive rate is high when detecting negative-phase spikes. Morphological filtering is another way to study spike extraction. Based on the geometric features of the signal, it uses pre-defined structural elements to match the signal to extract the signal with similar morphology. It has the characteristics of easy algorithm, clear physical meaning, practical and effective, etc. It can decompose the signal containing complex components into parts with different physical meanings, separate the signal from the background and maintain its global or local main morphological characteristics, but a single Open-close (OC) or closed-open (CO) operations can lead to statistical bias, which makes the detected spikes and the actual spikes deviate in shape and location.

SUMMARY OF THE INVENTION

In view of the problems existing in the prior art, the present invention provides a method for intelligent detection of epilepsy spikes based on the fusion of adaptive template matching and machine learning methods, so as to improve the recognition rate of epilepsy spikes.

In order to achieve the above object, the present invention is realized through the following scheme: a method for intelligent detection of epilepsy spikes based on the fusion of adaptive template matching and machine learning algorithm, the method comprises the following steps:

Step S1: collecting EEG signals, selecting experimental objects, establishing an epilepsy EEG database, and marking the spikes in each channel of the EEG signals;

Step S2: Preprocess the EEG signal, and use a 5th-order Butterworth bandpass filter to remove high-frequency components and artifacts.

Step S3: Detecting epilepsy spikes by using an adaptive template matching method.

Step S4: using a machine learning method to detect epilepsy spikes.

Step S5: The detection results of Step S3 and Step S4 are fused to obtain the final spike wave detection result.

According to an embodiment of the present invention, the sampling frequency in step S1 is 500 Hz, and a large amount of EEG data needs to be taken as experimental samples, and the experimental subjects include people of different genders and different ages.

According to an embodiment of the present invention, in the process of adaptive template matching spike detection, morphological features such as rising edge slope, falling edge slope, amplitude and duration of manually marked spikes are counted to establish a general spike template.

According to an embodiment of the present invention, in the process of using the adaptive template matching method to detect epilepsy spike waves, the process includes:

Step S31, count the characteristics of the spike waveform such as the rising edge slope, falling edge slope, amplitude height and duration in the EEG data, and define a general template;

Step S32, setting the window width to 300, and performing a general template matching operation on the EEG signals in chronological order to obtain candidate spike signals;

Step S33, performing K-means clustering on the candidate spikes, and dividing the candidate spikes into different classes according to different waveforms;

Step S34, count the number of candidate spikes in each spike cluster, if the number is less than 5% of the total number of candidate spikes, remove this class, and finally use the centroid of the remaining class as a new template;

Step S35 , use the centroid of each class as a template to perform new template matching, and superimpose the results to obtain a spike detection result.

According to an embodiment of the present invention, performing K-means clustering includes:

Step S331: randomly select k samples in the sample set as the initial cluster centroids;

Step S332: Calculate the distance between each sample and the initial centroid, reclassify according to the minimum distance, and classify each candidate spike into the class with the nearest centroid to obtain a clustering result.

Step S333: Average the obtained samples in each class as the centroid of the next clustering.

Step S334: Repeat steps S332 and S333 until the position of the centroid no longer changes, and the clustering ends.

According to an embodiment of the present invention, the process of using the machine learning method to detect epilepsy spikes includes:

Step S41: Divide the EEG signal into single-channel segments with a length of 1 s, and mark this segment as 1 if there is a spike, and as 2 if there is no spike. The time-domain and frequency-domain features of EEG segments are extracted respectively, and a robust epilepsy spike feature vector is constructed.

Step S42: Randomly divide the feature vector into a training set and a test set, and use multiple EEG signal samples in the training set to train multiple decision trees in the random forest classifier to form a random forest model.

Step S43: Input the data in the test set into the trained random forest model to obtain the spike detection result based on the machine learning method.

According to an embodiment of the present invention, the random forest model training includes:

Step S421: In the training sample set, there is a new training set with the same number of samples extracted as the training set.

Step S422: randomly and without playback sampling in the feature vector set, to form a feature vector set to be selected;

Step S423: According to the candidate feature training set obtained in step S422, calculate the best splitting method of each node and split the node, without pruning, until the impurity of each leaf node reaches the specified requirements, forming a tree. decision tree.

Step S424: Repeat steps S421 to S423 until all decision trees are generated and integrated to obtain a random forest model.

By adopting the technical scheme of the present invention, the detection of epilepsy spikes is carried out through the fusion of adaptive template matching and machine learning, thereby greatly improving the recognition rate of epilepsy spikes.

Description of drawings

FIG. 1 is a general flow chart of the intelligent detection method of epilepsy spikes based on the fusion of adaptive template matching and machine learning algorithm according to the present invention.

FIG. 2 is a flowchart of the adaptive template matching spike detection according to the present invention.

FIG. 3 is a flow chart of K-means clustering according to the present invention.

FIG. 4 is a flow chart of the machine learning spike detection according to the present invention.

FIG. 5 is a flow chart of training a random forest model for machine learning spike detection according to the present invention.

Detailed ways

EEG signals usually contain a lot of physiological information about human diseases, especially playing an important role in the detection of epilepsy. EEG signals contain many epilepsy characteristic waves, among which spike waves are typical waveforms. Therefore, in order to conduct better research, it is necessary to detect spike waves in epilepsy EEG signals. Existing spike wave methods are difficult to completely and accurately determine the location of spike waves, which greatly affects the study of epilepsy. In view of this, this embodiment provides an intelligent detection method for epilepsy spikes based on the fusion of adaptive template matching and machine learning algorithm.

In order to make the purpose, realization scheme and innovation point of the present invention more prominent, the present invention will be further described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

Fig. 1 is the general flow chart of the intelligent detection method of epilepsy spike waves based on the fusion of adaptive template matching and machine learning algorithm of the present invention, including:

Step S1: EEG signal acquisition: select experimental objects, use EEG acquisition equipment to collect EEG data of epilepsy patients, and establish an experimental database;

Step S2: data preprocessing: perform Butterworth bandpass filtering on the collected raw EEG data to obtain a standard EEG signal;

Step S3: Adaptive template matching spike detection: first, define a general template according to the waveform characteristics of epilepsy spikes, perform general template matching, and obtain candidate spike signals; then use the K-means algorithm to cluster the candidate spikes to obtain several Classes; count the number of candidate spikes in each class, if the number of spikes is less than 5% of the total number of spikes, remove this class; use the filtered class center as a new template for adaptive template matching, and all The matching results are added to obtain the spike detection result.

Step S4: Machine learning spike detection: first segment the EEG signal into 1s-long EEG segments, then extract the time-domain and frequency-domain features in each EEG segment to construct a spike-wave feature vector; use the feature vector to train random Forest classification model to obtain spike detection results based on machine learning.

Step S5 : fusion of detection results: the S3 spike detection method and the S4 spike detection method are fused. If a spike is detected by S3 and S4 at the same time, it is regarded as an epileptic spike.

The intelligent detection method for epilepsy spikes based on the fusion of adaptive template matching and machine learning algorithm proposed in this embodiment is described in detail below with reference to FIG. 1 to FIG. 5 .

The intelligent detection method for epilepsy spike waves based on the fusion of adaptive template matching and machine learning algorithm provided in this embodiment starts from step S1, in which a polyconductive electroencephalograph is used to collect long-range monitoring EEG signals of patients, and the sampling frequency is 500 Hz , The electrode distribution adopts the international 10-20 EEG acquisition standard, a total of 19 channels of EEG data are collected, and a large number of EEG signals of different genders and different ages are collected to obtain multiple EEG signal samples. Professional doctors mark multiple EEG signal samples, and mark the spike waveform in each channel of the EEG signal.

Then, step S2 is executed to perform preprocessing operation on the EEG. A 5th-order Butterworth band-pass filter is used to filter out frequency components above 32Hz and below 0.5Hz to reduce the interference of noise and artifacts.

Step S3 performs adaptive template matching on the pre-processed EEG signal to obtain a spike detection result. The adaptive template matching spike detection method will be described in detail below with reference to FIG. 2 .

First, perform statistical analysis on the marked spike waveforms in the EEG signal, and obtain the average value of the rising edge slope, falling edge slope, peak value and duration of all marked spike waves as a standard to establish a general template (step S31 ). Then, the window width is set to 300, and a general template matching operation is performed on the EEG signals in time sequence to obtain candidate spike signals (step S32). K-means clustering is performed on the candidate spikes, and the candidate spikes are divided into different classes according to different waveforms (step S33). The number of candidate spikes in each spike cluster is counted, if the number is less than 5% of the total number of candidate spikes, this class is eliminated, and finally the centroid of the remaining class is used as a new template (step S34). New template matching is performed using the centroid of each class as a template, and the results are superimposed to obtain a spike detection result (step S35).

As shown in Figure 3, the process of K-means clustering is as follows:

Among the n candidate spikes, k samples are randomly selected as the initial cluster centroids (step S331). Calculate the distance between each candidate spike and each initial centroid, reclassify according to the minimum distance, and classify each candidate spike into the class with the nearest centroid to obtain a clustering result (step S332). The candidate spikes in each cluster are averaged as the centroid of the next cluster (step S333). Steps S332 and S333 are repeated until the position of the centroid no longer changes or the number of clustering times meets the requirement, the clustering ends, and the clustering result is obtained (step S334). In this embodiment, the number of initial centroids is k=n, that is, each candidate spike is used as a centroid for clustering, and finally n clustering results are obtained.

Step S4 adopts a machine learning method to extract spike waves from the preprocessed EEG signal. The machine learning spike detection method will be described in detail below with reference to FIG. 3 .

First, the EEG signal of each channel is divided into 1s-long segments, and multiple feature parameters of each segment are extracted, including time-domain feature parameters and frequency-domain feature parameters, and a feature vector corresponding to each EEG segment is constructed ( Step S41). Then, the feature vector is divided into a training set and a test set, and a random forest classification model is trained using the data in the training set (step S42). The data in the test set is input into the random forest model, and the output result obtained after voting by each decision tree is the spike detection result, which can detect whether there is a spike in this segment (step S43).

The EEG segment obtained by segmentation in step S41 is denoted as x(n), n=1, 2, . . . , N, where N is the length of the EEG segment. =500. Rhythm waves are extracted by wavelet packet transform before feature extraction. Because the frequency range of spike waves is above 14 Hz, the "db6" wavelet basis function is used to decompose and reconstruct the signal to obtain β waves and γ waves, which are denoted as x ₁ (n ) and x ₂ (n).

The time-domain characteristic parameters extracted in the step S41 include the minimum value, maximum value, average value, standard deviation, peak value of the original EEG signal x(n) and the two rhythmic wave signals x ₁ (n) and x ₂ (n) Degree, Skewness, and Hjorth parameters. The minimum value Min and the maximum value Max are the maximum value of the signal amplitude, respectively, and the average value Mean is the trend of the EEG signal amplitude. The formula is as follows:

The standard deviation SD reflects the difference between the amplitude of each sampling point and the mean, and the formula is as follows:

where x(n) is the EEG signal, and N is the number of sampling points of x(n), which is the average value of the amplitudes of all sampling points in x(n).

The kurtosis Kur represents the peak level of the data frequency distribution curve, and the formula is as follows:

Skew represents the characteristics of the degree of asymmetry in the amplitude of the EEG signal, and the formula is as follows:

The Hjorth parameter contains Hjorth mobility and Hjorth complexity:

Hjorth mobility can be expressed by the following formula:

The Hjorth complexity can be expressed by the following formula:

dnf_n＝x(n)-x(n-1)。in

dnf _n =x(n)-x(n-1).

The frequency domain characteristic parameters extracted in the step S41 include the energy E _i of the two rhythmic waves, and the ratio of the energy of the two rhythmic waves to the total signal energy R _i .

The rhythm wave is extracted by wavelet packet transform, and the EEG signal is decomposed by five layers of wavelet using the "db6" wavelet function to obtain β wave and γ wave, which are denoted as x ₁ (n) and x ₂ (n) respectively.

The two-rhythm wave energy E _i is obtained by the following formula:

The formula for the total signal energy E _all is as follows:

Furthermore, the energy ratio of the rhythmic wave can be calculated, R _i =E _i /E _all , i=1,2.

The step S42 is the training process of the random forest model. The random forest classifier includes a plurality of decision trees, and its output category is determined by the result of all the trees with the most votes. Through the bootstrap resampling technology, m samples are randomly selected from the original training sample set M samples repeatedly to generate a new training sample set, and then a random forest is generated by m individual decision tree classifiers. The essence of the random forest classifier is an improvement of the decision tree algorithm, which combines multiple decision trees, and the establishment of each tree relies on independently randomly selected samples. Every tree in the forest has the same distribution, and the classification error depends on the classification ability of each tree and its correlation.

This paper divides the data set into training set and test set. The following describes the training process of the random forest model in conjunction with Figure 5:

Step S421: First, M samples with replacement are taken from all feature vector sets to form a feature set to be selected. The number of samples in the feature set to be selected is the same as the number of samples in the original feature vector set.

Step S422: Second, randomly select a certain number of feature vectors from the features to be selected, and select the optimal feature among them.

Step S423: According to the candidate feature training set obtained in step S422, calculate the best splitting method of each node and split the node, without pruning, until the impurity of each leaf node reaches the specified requirements, forming a tree. decision tree.

Step S424: Repeat steps S421 to S423 until all decision trees stop growing, and generate a random forest.

The step S43 is to input the EEG data in the test set into the random forest model, and after voting by the decision tree, the spike wave detection result can be obtained, the EEG segment where the spike wave is located, and then the EEG channel where the spike wave is located can be determined. and time point.

The step S5 first performs adaptive template matching on the data in the test set to obtain a spike detection result. At the same time, the test set is input into the random forest model for classification to obtain spike detection results. Then the results of the two methods are fused and compared, and if they are detected as spikes by both methods at the same time, they are regarded as epileptic spikes.

The descriptions of the above embodiments are only used to help understand the method and the core idea of the present invention. It should be pointed out that for those skilled in the art, without departing from the principle of the present invention, several improvements and modifications can also be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (3)

1. the intelligent detection method of epilepsy spikes based on adaptive template matching and machine learning algorithm fusion, is characterized in that, this method may further comprise the steps: Step S1: EEG signal acquisition; select experimental objects, use EEG acquisition equipment to collect EEG data of epilepsy patients, and establish an experimental database; Step S2: data preprocessing; Butterworth bandpass filtering is performed on the collected original EEG data to obtain a standard EEG signal; Step S3: Adaptive template matching spike detection; first define a general template according to the waveform characteristics of epilepsy spikes, perform general template matching to obtain candidate spike signals; then use the K-means algorithm to cluster the candidate spikes to obtain several Classes; count the number of candidate spikes in each class, if the number of spikes is less than 5% of the total number of spikes, remove this class; use the filtered class center as a new template for adaptive template matching, and all The matching results are added to obtain the spike detection result; Step S4: Machine learning spike detection; first, the EEG signal is divided into 1s-long EEG segments, and then the time domain and frequency domain features in each EEG segment are extracted to construct a spike feature vector; the feature vector is used to train random Forest classification model to obtain spike detection results based on machine learning; Step S5: fusion of detection results; fusion of the detection result of the spike wave in step S3 and the detection result of the spike wave in step S4, if a spike wave signal is detected in the same segment in steps S3 and S4, the final result is recorded as the segment with a spike wave signal. If it exists, it is regarded as an epileptic spike; Wherein, the step S3 further includes: Step S31, count the characteristics of the spike waveform such as the rising edge slope, falling edge slope, amplitude height and duration in the EEG data, and define a general template; Step S32, setting the window width to 300, and performing a general template matching operation on the EEG signals in chronological order to obtain candidate spike signals; Step S33, performing K-means clustering on the candidate spikes, and dividing the candidate spikes into different classes according to different waveforms; Step S34, count the number of candidate spikes in each spike cluster, if the number is less than 5% of the total number of candidate spikes, remove this class, and finally use the centroid of the remaining class as a new template; Step S35, respectively use the centroid of each class as a template to perform new template matching, and superimpose the results to obtain a spike detection result; The step S4 further includes: Step S41, dividing each channel of the EEG signal into 1s long segments, extracting the time domain feature and frequency domain feature of each segment, and constructing a feature vector of each EEG segment; Step S42, dividing the feature vector into a training set and a test set, and using the data in the training set to train a random forest classification model; In step S43, the data in the test set is input into the random forest model, and the obtained output result is the detection result of the spike wave, which can detect whether there is a spike wave in this segment. 2. the epilepsy spike intelligent detection method based on adaptive template matching and machine learning algorithm fusion as claimed in claim 1, is characterized in that, when carrying out data preprocessing, using frequency range is 5th order IIR Bart of 0.5-32Hz Worth bandpass filter to remove noise and artifacts from EEG signals. 3. the epilepsy spike wave intelligent detection method based on self-adaptive template matching and machine learning algorithm fusion as claimed in claim 1 or 2, it is characterized in that, spike wave detection result is compared with the detection result of adjacent channel, if detect If there is a "tit-for-tat" phenomenon in the adjacent channel, it will be regarded as a spike wave, otherwise the result will be discarded.

Images