CN111870241B - Epileptic seizure signal detection method based on optimized multidimensional sample entropy - Google Patents

Abstract

The invention discloses an epileptic seizure signal detection method based on optimized multi-dimensional sample entropy. In this study, the present invention uses multi-dimensional sample entropy as a feature to distinguish epileptic seizure states from normal states, and optimizes them to improve computational efficiency. Furthermore, by combining multidimensional sample entropy feature extraction and Bi‑LSTM, a new prediction method is developed to predict epileptic seizures. The results show that the method achieves good performance and can predict the multi-dimensional sample entropy of EEG after 5 minutes with an accuracy rate of 80.09% and a false alarm rate of 0.26/h. The results of this study suggest that the proposed prediction scheme is more suitable for actual seizure prediction.

Description

Epileptic seizure signal detection method based on optimized multidimensional sample entropy

technical field

The invention belongs to the field of signal feature analysis, and relates to an epileptic seizure signal detection method based on optimized multi-dimensional sample entropy and bi-directional long short-term memory neural network (Bi-directional Long Short-Term Memory, Bi-LSTM).

Background technique

Epilepsy is a common brain disease caused by abnormal hypersynchronization of brain neurons discharge, ranking second in neurological diseases, and its incidence is second only to stroke. Epilepsy is a chronic, long-term, recurring disease that is sudden and causes temporary loss of brain function during the seizure. The prevalence of epilepsy in foreign countries is about 3‰ to 10‰. , the prevalence rate in my country is 4‰ to 9‰, so far, about 50 million patients worldwide are suffering from pain. Most people with epilepsy experience sudden collapse, body convulsions, and loss of consciousness during a seizure. At the same time, due to the sudden nature of epileptic seizures, if the patient is engaged in some dangerous operation (such as driving a car) at this time, it is easy to suffer accidental injury. If it is possible to predict the onset of epilepsy before the onset of epilepsy, even for a short period of time, the patient or doctor can take necessary preventive and protective measures in time, which is very beneficial to the patient. Seizure prediction plays an important role in epilepsy treatment, so it has become a hot spot in current epilepsy research.

Various techniques have been proposed to solve this problem, among which Electroencephalography (EEG) has multiple advantages, including high temporal resolution, low cost, long-term monitoring and portability, etc. It has been proven to be effective for seizure analysis one of the preferred methods.

Many methods for analyzing EEG signals have been proposed in the past few decades, and these methods can be divided into two types: linear methods and nonlinear methods. Linear methods include time domain analysis, frequency domain analysis and time-frequency domain analysis. Time domain analysis was the first method used to analyze epilepsy signals. The advantage of time-domain analysis is that the time-domain waveform contains all the information of the EEG, but this method lacks objectivity and has large errors. Frequency domain analysis overcomes the shortcomings of time domain analysis, but it is premised on stationary random signals, while EEG signals are nonlinear and non-stationary, which leads to many limitations. Time-frequency domain analysis methods, including short-time Fourier transform, wavelet transform, Hilbert-Huang transform and empirical mode decomposition, are the most commonly used methods for studying EEG signals, and have achieved good results.

From the perspective of nonlinear dynamics, a large number of studies have shown that the activities of the brain have extremely complex dynamic characteristics, and the brain can be regarded as a nonlinear dynamic system. Identifying epileptic EEG by extracting EEG features based on nonlinear dynamics theory has become one of the frontier trends in automatic seizure detection. Biological signals are very weak and often contain noise in the environment, and entropy methods have significant advantages in biological signal processing. Another advantage of the entropy-based method is that it can achieve meaningful results with less data than other nonlinear methods. Among them, sample entropy is a new algorithm based on approximate entropy, which can avoid some defects of approximate entropy. It does not depend on the data length, the consistency is better, it is not sensitive to data loss, and the algorithm is simpler. And Mormann et al. pointed out that bivariate and multivariate measures have better performance than univariate.

The key to the detection of epileptic seizure signals is not only the extraction of features, but also the selection of classifiers. Many classification algorithms such as Support Vector Machine (SVM), Decision Tree, Convolutional Neural Networks (CNN) have been used to classify epilepsy features and achieved good results. Long Short-Term Memory (LSTM) is a temporal recurrent neural network. Unlike other classification algorithms, LSTM can be used for both classification and prediction of time series. Among them, Bi-LSTM is a variant of LSTM, which consists of a forward and a backward LSTM, which is more robust.

SUMMARY OF THE INVENTION

Based on the above discussion, the present invention proposes an epileptic seizure signal detection method based on optimized multi-dimensional sample entropy and Bi-LSTM, which combines the EEG signals of multiple channels to analyze and extract features. Then, Bi-LSTM is used to predict the change trend of multi-dimensional sample entropy, and the predicted multi-dimensional sample entropy is classified to distinguish epileptic seizure period and normal period to detect epileptic seizure signals.

In order to achieve the above object, the method of the present invention mainly comprises the following steps:

Step (1). Collecting EEG data and preprocessing, all signals are sampled by the international standard 10-20 electrode distribution system. Data preprocessing includes wavelet denoising, electrocardiogram, and electrooculography.

Step (2). Calculate the optimized multi-dimensional sample entropy.

The first thing to do is to construct a multidimensional vector. The original data of the optimized multi-dimensional sample entropy calculation are k-channel EEG signals, each signal has N sampling points, and m is the embedding dimension. Each point in the constructed vector is a k-dimensional vector. The EEG signal collected by each channel has n sampling points. The sampling points of the first signal are x ₁₁ , x ₁₂ , x ₁₃ ,..., x _1n , and the sampling points of the second signal are x ₂₁ , x ₂₂ , x ₂₃ , ..., x _2n , and so on, the sampling points of the k-th signal are x _k1 , x _k2 , x _k3 , ..., x _kn . Extract the a-th sampling point of each signal to obtain a multi-dimensional vector Q(a)=(x _1a , x _2a , x _3a ,...,x _ka ). The embedding dimension takes m=2, so the definition point X(a)=[Q(a), Q(a+1)].

Compute the distance D[Q(a),Q(a+1)] between Q(a) and Q(a+1). Since Q(a) is a multidimensional vector, this paper defines

where O _D is the Euclidean distance. All the calculated distances are stored in a table to avoid repeated calculation in the next cycle.

When m=2, X(a)=[Q(a), Q(a+1)], X(a+1)=[Q(a+1), Q(a+2)], from the table Take D[Q(a), Q(a+1)] and D[Q(a+1), Q(a+2)], take the larger one as the difference between X(a) and X(a+1) the distance.

Set the similarity tolerance R, calculate the number of distances less than R x ₁ ,

Calculate the average of all

When m=3, X(a)=[Q(a), Q(a+1), Q(a+2)], X(a+1)=[Q(a+1), Q(a+ 2),Q(a+3)] Take D[Q(a),Q(a+1)], D[Q(a+1),Q(a+2)] and D[Q( a+2), Q(a+3)], take the largest one as the distance between X(a) and X(a+1).

Set the similarity tolerance R, calculate the number of distances less than R x ₂ ,

的平均值

calculate all

average of

Calculate the optimized multidimensional sample entropy SampEn=-ln[Bm ⁺¹ (R) ^/ Bm(R)].

Step (3) According to the optimized multi-dimensional sample entropy, use Bi-LSTM to detect epileptic seizure signals. The optimized multi-dimensional sample entropy calculated before the current moment is used as the input of Bi-LSTM, and the function of Bi-LSTM to predict time series is used to output the predicted next multi-dimensional sample entropy. Then, the predicted multi-dimensional sample entropy is divided into two categories through the classification function of Bi-LSTM, namely the seizure period and the normal period, so as to achieve the purpose of epileptic seizure signal detection.

The present invention has the following characteristics with respect to the prior art:

Because EEG signals are nonlinear and non-stationary random signals, the present invention studies the EEG features of patients based on the nonlinear entropy method, and extracts the features of multi-dimensional sample entropy in combination with multi-channel EEG signals. Because when a seizure occurs, the EEG signal of more than one channel changes. Multidimensional sample entropy can combine EEG signals from all channels to reflect changes in the brain before and after seizures.

Since the multi-dimensional sample entropy uses a large amount of data and the calculation speed is very slow, the present invention optimizes the multi-dimensional sample entropy. According to the definition of sample entropy, the calculation is repeated during the point-by-point comparison, and the first mode is continuously looped as the sample entropy is calculated. Combining these loops avoids a lot of repeated computations and increases computation speed.

LSTMs can utilize contextual information in the mapping between input and output sequences, making them ideal for dealing with time series forecasting problems. The invention makes full use of the characteristics of LSTM, and uses a variant Bi-LSTM with stronger stability to predict the change trend of multi-dimensional sample entropy, and then classify and detect epileptic seizure signals.

Description of drawings

Fig. 1 is the implementation flow chart of the present invention;

Fig. 2 is the nomenclature diagram of the EEG electrode of the international 10-20 system according to the embodiment of the present invention;

Fig. 3 (a) is the broken-line graph before and after the entropy of common sample is extracted from the EEG signal of epilepsy patient according to the embodiment of the present invention;

Fig. 3(b) is a broken line graph of the multi-dimensional sample entropy extracted from the EEG signal of epilepsy patients before and after epileptic seizure according to the embodiment of the present invention;

4 is a comparison diagram of the relationship between calculation time and signal length of multi-dimensional sample entropy and optimized multi-dimensional sample entropy according to an embodiment of the present invention;

Figure 5(a) is an effect diagram of the present invention updating the network state through the predicted value;

Figure 5(b) is an effect diagram of the present invention updating the network state through the observation value;

FIG. 6 is a graph showing the relationship between the root mean square error (RMSE) and the prediction time length in the implementation of the present invention.

Detailed ways

Below in conjunction with the accompanying drawings, the embodiments of the present invention are described in detail: the present embodiment is implemented on the premise of the technical solution of the present invention, and provides detailed embodiments and specific operation processes, but the protection scope of the present invention is not limited to the following described embodiment.

As shown in Figure 1, this embodiment includes the following steps:

Step (1), collecting EEG data. The data used in the present invention were obtained from the Massachusetts Institute of Technology public database. 13 patients were selected, 100 samples were collected from each patient, a total of 1300 samples were used to calculate the sample entropy and the optimized multi-dimensional sample entropy for classification, the first 60% of the samples were trained, and the last 40% of the samples were tested . All datasets are sampled at 256Hz with 16-bit resolution. The International 10-20 system of EEG electrode nomenclature was used, as shown in Figure 2. The epilepsy patients were not taking any medications and had no other family history of genetic disease during the experimental data collection. The number of channels used to collect EEG signals was 23 in most experimental groups and 24 or 26 in a few experiments. For the convenience of statistical data, the present invention selects the data of 23 channels.

Step (2), calculate the optimized multi-dimensional sample entropy. In the feature extraction of EEG signal, the method of multi-dimensional sample entropy is used, and 23 EEG signals are used for joint calculation. Each signal has N sampling points, forming a 23×N matrix. Optimize it due to excessive computation. First, the method of moving the window is adopted to reduce the data length to speed up the calculation. According to the existing literature, the window size is set to 5 seconds. Secondly, from the definition of sample entropy, there are calculation repetitions during the point-by-point comparison, and there are always loops when calculating sample entropy. These loops can be combined to remove a large number of repeated calculations and improve the running speed. Therefore, the optimized multi-dimensional sample entropy calculation steps are as follows:

The first thing to do is to construct a multidimensional vector. The original data of the optimized multi-dimensional sample entropy calculation are k-channel EEG signals, each signal has N sampling points, and m is the embedding dimension. The EEG data selected in the present invention are all 23 channels, so k=23, the EEG signal collected by each channel has n sampling points, and the sampling points of the first signal are x ₁₁ , x ₁₂ , x ₁₃ , ... , x _1n , the sampling points of the second signal are x ₂₁ , x ₂₂ , x ₂₃ , ..., x _2n , and so on, the sampling points of the k-th signal are x _k1 , x _k2 , x _k3 , . .., x _kn . Extract the a-th sampling point of each signal to obtain a multi-dimensional vector Q(a)=(x _1a , x _2a , x _3a ,...,x _ka ). The embedding dimension takes m=2, so the definition point X(a)=[Q(a), Q(a+1)].

Calculate the distance D[Q(a), Q(a+1)] between Q(a) and Q(a+1), since Q(a) is a multidimensional vector, this paper defines

where O _D is the Euclidean distance. All the calculated distances are stored in a table to avoid repeated calculation in the next cycle.

When m=2, X(a)=[Q(a), Q(a+1)], X(a+1)=[Q(a+1), Q(a+2)], from the table Take D[Q(a), Q(a+1)] and D[Q(a+1), Q(a+2)], take the larger one as the difference between X(a) and X(a+1) the distance.

Set the similarity tolerance R, calculate the number of distances less than R x ₁ ,

的平均值calculate all

average of

When m=3, X(a)=[Q(a), Q(a+1), Q(a+2)], X(a+1)=[Q(a+1), Q(a+ 2), Q(a+3)] from the table

Take D[Q(a), Q(a+1)], D[Q(a+1), Q(a+2)] and D[Q(a+2), Q(a+3)] , take the largest one as the distance between X(a) and X(a+1).

Set the similarity tolerance R, calculate the number of distances less than R x ₂ ,

的平均值calculate all

average of

Calculate the sample entropy SampEn=-ln[Bm ⁺¹ (R) ^/ Bm(R)].

To compare performance differences between sample entropy and multidimensional sample entropy, 13 patients were selected. Both sample entropy and multidimensional sample entropy for each patient were calculated at the same interval of EEG data before and after seizures. From Figure 3(a), (b), it can be seen that the value of sample entropy tends to increase during epileptic seizures, but the difference before and after seizures is still not obvious; and the multidimensional sample entropy values also show an increasing trend during epileptic seizures, and epilepsy The differences of multidimensional sample entropy values before and after seizure were more obvious than the sample entropy values. As shown in Table 1, the accuracy (ACC), recall, specificity (SPC) and positive predictive value (PPV) of multi-dimensional sample entropy are all higher than sample entropy, showing better performance.

Table 1. Performance comparison of sample entropy and multidimensional sample entropy

In order to compare the advantages of the optimized multi-dimensional sample entropy in terms of computational efficiency, five sets of EEG signals with sampling points of 256, 512, 1024, 2048 and 4096 were intercepted. Calculate the multidimensional sample entropy and the optimized multidimensional sample entropy for each segment separately, and record the time required. Table 2 lists the average time to compute the sample entropy and the optimized multidimensional sample entropy. It can be seen that the optimization algorithm reduces the computation time. Figure 4 shows the curve of computation time as a function of signal length. With the increase of signal length, the optimized multi-dimensional sample entropy has more prominent advantages in terms of computational efficiency and is therefore more suitable for clinical diagnosis.

Table 2 Computational time comparison between multidimensional sample entropy and optimized multidimensional sample entropy

Step (3), according to the optimized multi-dimensional sample entropy, use Bi-LSTM to detect epileptic seizure signals. The optimized multi-dimensional sample entropy calculated before the current moment is used as the input of Bi-LSTM, and the function of Bi-LSTM to predict time series is used to output the predicted next multi-dimensional sample entropy. Then, the predicted multi-dimensional sample entropy is divided into two categories through the classification function of Bi-LSTM, namely the seizure period and the normal period.

There are two ways to update the state of the Bi-LSTM network. One is to use the predicted value to update the predicted network state and use the previous predicted value as the input to the function; the other is to use the observed value to update the predicted network state and use the previous moment's observation to predict the next moment. . RMSE is used to calculate the deviation between the observed value and the true value, as shown in Fig. 5(a), (b), when the network state is updated with the observed value, the prediction result is more accurate. So the present invention uses the observations to update the Bi-LSTM network state, and uses the "Adam" optimizer for 250 rounds of training. To prevent exploding gradients, set the gradient threshold to 1. Specify the initial learning rate as 0.05, and reduce the learning rate by multiplying by 0.2 after 125 epochs of training. If the learning rate is too small, the convergence rate will be very slow; if it is too large, the loss function will oscillate and even deviate from the minimum value. Therefore, a large learning rate is set first, and when the change between iterations is below a threshold, the learning rate will decrease.

Figure 6 shows the RMSE for the prediction time of 2 minutes, 5 minutes and 10 minutes, and it can be seen that as the prediction time increases, the RMSE also increases gradually. This shows that the longer the prediction time, the larger the error. Since there is not enough time to prevent or control seizures, alerts generated within 2 minutes before the seizure (intervention time) can be ignored. Combined with the values of multidimensional sample entropy and RMSE, the prediction time of the present invention is chosen to be 5 minutes.

In this study, a novel feature multidimensional sample entropy is proposed for distinguishing seizure states from normal states, and a method for predicting seizures with Bi-LSTM is proposed. Firstly, the features of multi-dimensional sample entropy are extracted from 23 EEG signals, and the algorithm is optimized to improve computational efficiency. Then use Bi-LSTM to predict the entropy of the next multi-dimensional samples. The predicted multidimensional sample entropy is then classified to determine whether an epilepsy is about to occur. The method achieved good results, with high accuracy and low false positive rate, giving doctors and patients more time to respond.

Claims (1)

1. the epileptic seizure signal detection system based on optimized multidimensional sample entropy, comprises signal acquisition and preprocessing module, feature extraction module and prediction module, it is characterized in that this system runs through the following steps: Step (1), collecting EEG data and preprocessing through the signal acquisition and preprocessing module; Step (2), calculate the optimized multi-dimensional sample entropy by the feature extraction module; ①, construct multi-dimensional vector; The original data of the optimized multi-dimensional sample entropy calculation is k-channel EEG signals, each signal has N sampling points, and m is the embedding dimension; each point in the constructed vector is a k-dimensional vector; The EEG signal has n sampling points, the sampling points of the first signal are x ₁₁ , x ₁₂ , x ₁₃ , ..., x _1n , and the sampling points of the second signal are x ₂₁ , x ₂₂ , x ₂₃ , ..., x _2n , and so on, the sampling points of the k-th signal are x _k1 , x _k2 , x _k3 , ..., x _kn ; extract the a-th sampling point of each signal to obtain a multi-dimensional vector Q(a)=(x _1a , x _2a , x _3a ,...,x _ka ); the embedding dimension takes m=2, so the definition point X(a)=[Q(a), Q(a+1)]; Calculate the distance D[Q(a), Q(a+1)] between Q(a) and Q(a+1); since Q(a) is a multidimensional vector, so

Among them, O _D is the Euclidean distance; all the calculated distances are stored in a table to avoid repeated calculation in the next cycle; When m=2, X(a)=[Q(a), Q(a+1)], X(a+1)=[Q(a+1), Q(a+2)], from the table Take D[Q(a), Q(a+1)] and D[Q(a+1), Q(a+2)], take the larger one as the difference between X(a) and X(a+1) the distance; Set the similarity tolerance R, calculate the number of distances less than R x ₁ ,

calculate all

average of

When m=3, X(a)=[Q(a), Q(a+1), Q(a+2)], X(a+1)=[Q(a+1), Q(a+ 2), Q(a+3)], take D[Q(a), Q(a+1)], D[Q(a+1), Q(a+2)] and D[Q from the table (a+2), Q(a+3)], take the largest one as the distance between X(a) and X(a+1); Set the similarity tolerance R, calculate the number of distances x ₂ where the distance between X(a) and X(a+1) is less than R,

calculate all

average of

Calculate the optimized multi-dimensional sample entropy SampEn=-ln[B ^m+1 (R)/B ^m (R)]; Step (3). Through the prediction module, according to the optimized multi-dimensional sample entropy obtained by the feature extraction module, use Bi-LSTM to detect epileptic seizure signals; use the optimized multi-dimensional sample entropy calculated before the current moment as the input of Bi-LSTM, use Bi-LSTM predicts the time series function, and outputs the predicted entropy of the next multi-dimensional sample; and then divides the predicted multi-dimensional sample entropy into two categories through the classification function of Bi-LSTM, that is, the seizure period and the normal period, so as to achieve the epileptic seizure signal. purpose of detection.

Images