CN111870241B

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

Info

Publication number: CN111870241B
Application number: CN202010637325.6A
Authority: CN
Inventors: 丁吉; 刘阳; 汪茜; 张启忠; 高云园; 席旭刚
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2020-07-03
Filing date: 2020-07-03
Publication date: 2022-05-17
Anticipated expiration: 2040-07-03
Also published as: CN111870241A

Abstract

The invention discloses an optimized multidimensional sample entropy-based epileptic seizure signal detection method. In the research, the invention adopts the multi-dimensional sample entropy as the characteristic to distinguish the epileptic seizure state from the normal state, optimizes the epileptic seizure state and the normal state and improves the calculation efficiency. In addition, by combining multidimensional sample entropy feature extraction and Bi-LSTM, a new prediction method is developed to predict epileptic seizures. The result shows that the method obtains good performance, can predict the multi-dimensional sample entropy of the electroencephalogram after 5 minutes, and has the accuracy rate of 80.09 percent and the false alarm rate of 0.26/h. The results of this study show 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 characteristic analysis, and relates to a method for detecting epileptic seizure signals based on optimized multidimensional sample entropy and a bidirectional Long-Term Memory neural network (Bi-LSTM).

Background

Epilepsy is a common brain disease caused by hypersynchronous abnormal discharge of cerebral neurons, and is ranked second among nervous system diseases, and the incidence rate is second to stroke. Seizures are chronic, long-term, repetitive disorders that are paroxysmal and cause temporary loss of brain function during the course of the disease. The prevalence rate of epilepsy is about 3-10 per mill abroad. The disease rate in China is 4-9 per thousand, and about 5 million patients are suffering from pain so far globally. Most epileptics are unconscious, spasmodic and unconscious at the onset. Also, because of the bursty nature of epileptic seizures, patients are susceptible to accidental injury if they are engaged in some dangerous operation (e.g., driving a car). It would be advantageous to a patient if the impending seizure could be predicted prior to the seizure, even for a relatively short period of time, to enable the patient or physician to take the necessary preventative measures in a timely manner. Seizure prediction plays an important role in epilepsy therapy and is thus a hotspot in current epileptics research.

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

Many methods of analyzing brain electrical signals have been proposed in the past decades and can be divided into two categories: linear methods and non-linear methods. Linear methods include time domain analysis, frequency domain analysis and time-frequency domain analysis. Time domain analysis is the first method used to analyze epileptic signals. The advantage of time domain analysis is that the time domain waveform contains all the information of the electroencephalogram, but this method lacks objectivity and has large errors. Frequency domain analysis overcomes the disadvantages of time domain analysis, but it presupposes stationary random signals, whereas EEG signals are non-linear and non-stationary signals, which leads to a number of limitations. The time-frequency domain analysis method comprises short-time Fourier transform, wavelet transform, Hilbert-Huang transform, empirical mode decomposition and the like, is the most common method for researching the electroencephalogram signals at present, and obtains better results.

From the point of view of nonlinear dynamics, a number of studies have shown that the brain's activity has extremely complex dynamics and can be regarded as a nonlinear dynamic system. The epileptic electroencephalogram is identified by extracting electroencephalogram characteristics based on a nonlinear dynamics theory, and the epileptic electroencephalogram becomes one of the leading edge trends of automatic epileptic seizure detection. Biological signals are weak and often carry noise in the environment, while the entropy method has significant advantages in biological signal processing. Another advantage of the entropy-based approach is that it can achieve meaningful results with less data than other non-linear approaches. The sample entropy is a new algorithm researched based on the approximate entropy, and some defects of the approximate entropy can be avoided. The method does not depend on the data length, has better consistency, is not sensitive to the loss of data, and has simpler algorithm. And Mormann et al indicate that bivariate and multivariate measurements have superior performance relative to univariates.

The key of the epileptic seizure signal detection is not only the extraction of the features, but also the selection of the classifier. Many classification algorithms, such as Support Vector Machines (SVMs), decision trees, Convolutional Neural Networks (CNNs), have been used to classify epileptic features and have achieved good results. The Long Short-Term Memory neural network (LSTM) is a time recurrent neural network, and unlike other classification algorithms, the LSTM can be used for classification and prediction of time series. Wherein the Bi-LSTM is a variant of the LSTM, and is composed of a forward LSTM and a backward LSTM, and is more robust.

Disclosure of Invention

Based on the above discussion, the invention provides an optimized multidimensional sample entropy and Bi-LSTM-based epileptic seizure signal detection method, which combines EEG signals of a plurality of channels to analyze and extract features. Then, the Bi-LSTM is used for predicting the change trend of the multi-dimensional sample entropy, and classifying the predicted multi-dimensional sample entropy so as to distinguish the epileptic seizure period from the normal period to achieve the purpose of detecting epileptic seizure signals.

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

step (1), collecting electroencephalogram data and preprocessing, wherein all signals are obtained by sampling of an international standard 10-20 electrode distribution system. The data preprocessing comprises wavelet denoising, electrocardio and electro-oculogram elimination and the like.

And (2) calculating the optimized multidimensional sample entropy.

A multi-dimensional vector is first constructed. The original data of the optimized multidimensional sample entropy calculation is k-channel EEG signals, each channel of signal has N sampling points, and m is an embedding dimension. Each point in the constructed vector is a k-dimensional vector. The EEG signal collected in each channel has n sampling points, the sampling point of the first channel signal is x₁₁、x₁₂、x₁₃、...、x_1nThe sampling point of the second signal is x₂₁、x₂₂、x₂₃、...、x_2nBy analogy, the sampling point of the k-th channel signal is x_k1、x_k2、x_k3、...、x_kn. Extracting the a-th sampling point of each signal to obtain a multidimensional vector Q (a) ═ x_1a,x_2a,x_3a,...,x_ka). The embedding dimension is m ═ 2, so point x (a) ═ Q (a), Q (a +1) is defined]。

The distance D [ Q (a), Q (a +1) ] between Q (a) and Q (a +1) is calculated. Since Q (a) is a multi-dimensional vector, it is defined herein

Wherein O is_DIs the euclidean distance. All the distances obtained by calculation are stored in a table so as to avoid repeated calculation in the next cycle.

When m is 2, X (a) ═ Q (a), Q (a +1) ], X (a +1) ═ Q (a +1), Q (a +2) ], where D [ Q (a), Q (a +1) ] and D [ Q (a +1), Q (a +2) ], the larger of which is taken as the distance between X (a) and X (a + 1).

Calculating the number x of distances less than R by setting a similar tolerance R₁，

Calculate all averages

When m is 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, D [ Q (a), (a +1) ], D [ Q (a +1), Q (a +2) ] and D [ Q (a +2), Q (a +3) ], the largest being taken as the distance between X (a) and X (a + 1).

Calculating the number x of distances less than R by setting a similar tolerance R₂，

Calculate all

Average value of (2)

Computationally optimized multidimensional sample entropy SampEn ═ -ln [ B^m+1(R)/B^m(R)]。

And (3) detecting epileptic seizure signals by using Bi-LSTM according to the optimized multidimensional sample entropy. And taking the optimized multidimensional sample entropy calculated before the current time as the input of the Bi-LSTM, and outputting the predicted next multidimensional sample entropy by utilizing the function of predicting the time sequence of the Bi-LSTM. And then the predicted multidimensional sample entropy is divided into two categories, namely a seizure period and a normal period, through the classification function of the Bi-LSTM, so as to achieve the purpose of epileptic seizure signal detection.

Compared with the prior art, the invention has the following characteristics:

because the EEG signal is a nonlinear and non-stationary random signal, the EEG characteristics of a patient are researched based on a nonlinear entropy method, and the characteristics of the multi-dimensional sample entropy are extracted by combining a multi-channel EEG signal. Because the EEG signal for more than one channel changes when there is a seizure. The multi-dimensional sample entropy may be combined with all channels of EEG signals to reflect changes occurring in the brain before and after an epileptic seizure.

Because the data volume used by the multi-dimensional sample entropy is large, and the calculation speed is slow, the multi-dimensional sample entropy is optimized by the method. The calculation is repeated during the point-by-point comparison, as defined by the sample entropy, and the first pattern is continually cycled through while calculating the sample entropy. Combining these loops can avoid a large number of repeated calculations and increase the calculation speed.

LSTM can exploit context information in the mapping between input and output sequences and is therefore well suited to handle time series prediction problems. The invention fully utilizes the characteristics of the LSTM, uses the variant Bi-LSTM with stronger stability to predict the change trend of the multi-dimensional sample entropy, and then carries out classification detection on the epileptic seizure signals.

Drawings

FIG. 1 is a flow chart of an embodiment of the present invention;

FIG. 2 is a diagram of an International 10-20 System EEG electrode nomenclature according to an embodiment of the present invention;

FIG. 3(a) is a line graph before and after an entropy attack of a common sample extracted from an electroencephalogram signal of an epileptic patient according to an embodiment of the present invention;

FIG. 3(b) is a line graph of multi-dimensional sample entropy extracted from electroencephalogram signals of an epileptic patient before and after an epileptic seizure according to an embodiment of the present invention;

FIG. 4 is a comparison graph of the computation time versus signal length for the multi-dimensional sample entropy and the optimized multi-dimensional sample entropy in accordance with an embodiment of the present invention;

FIG. 5(a) is a diagram illustrating the effect of updating the network status according to the predicted value;

FIG. 5(b) is an effect diagram of updating network status according to observation values in the present invention;

FIG. 6 is a graph of Root Mean Square Error (RMSE) versus predicted time length for an implementation of the present invention.

Detailed Description

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings: the present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

As shown in fig. 1, the present embodiment includes the following steps:

step (1), collecting electroencephalogram data. The data used in the present invention is from the Massachusetts institute of technology public database. Selecting 13 patients, collecting 100 samples for each patient, using 1300 samples in total for calculating sample entropy and classifying optimized multidimensional sample entropy, training the first 60% of samples, and testing the last 40% of samples. All data sets were sampled at 256Hz with 16-bit resolution. The international 10-20 system EEG electrode nomenclature is used as shown in fig. 2. Epileptic patients did not take any medication during the experimental data collection process and had no other family genetic history. In most experimental groups, the number of channels used to collect EEG signals was 23, while in a few experiments it was 24 or 26. To facilitate statistical data, the present invention selects data for 23 channels.

And (2) calculating the optimized multidimensional sample entropy. A method of multi-dimensional sample entropy is applied to EEG signal feature extraction, 23 channels of EEG signals are adopted for joint calculation, and each channel of signal has N sampling points to form a 23 multiplied by N matrix. Because of the excessive amount of calculation, it is optimized. Firstly, a method of moving a window is adopted, the data length is reduced to accelerate the calculation, and the window size is set as 5 seconds according to the existing literature. Secondly, from the definition of the sample entropy, the calculation repetition exists during the point-by-point comparison, and the circulation always exists when the sample entropy is calculated, and the circulation is combined, so that a large amount of repeated calculation can be eliminated, and the running speed is improved. The optimized multidimensional sample entropy calculation steps are as follows:

a multi-dimensional vector is first constructed. The original data of the optimized multidimensional sample entropy calculation is k-channel EEG signals, each channel of signal has N sampling points, and m is an embedding dimension. The EEG data selected in the invention are all 23 channels, so k is 23, each channel collects EEG signal with n sampling points, the sampling point of the first channel signal is x₁₁、x₁₂、x₁₃、...、x_1nThe sampling point of the second signal is x₂₁、x₂₂、x₂₃、...、x_2nBy analogy, the sampling point of the k-th channel signal is x_k1、x_k2、x_k3、...、x_kn. Extracting the a-th sampling point of each signal to obtain a multidimensional vector Q (a) ═ x_1a,x_2a,x_3a,...,x_ka). The embedding dimension is m ═ 2, so point x (a) ═ Q (a), Q (a +1) is defined]。

Calculating the distance D [ Q (a), Q (a +1) ] between Q (a) and Q (a +1), as Q (a) is a multidimensional vector, defined herein

Calculate all

Average value of (2)

When m is 3, X (a) ═ Q (a), Q (a +1), Q (a +2) ], X (a +1) ═ Q (a +1), Q (a +2), Q (a +3) ] are shown in the table

Taking D [ Q (a), Q (a +1) ], D [ Q (a +1), Q (a +2) ], D [ Q (a +2), Q (a +3) ], and taking the maximum as the distance between X (a) and X (a + 1).

Calculate all

Average value of (2)

Calculating sample entropy SampEn ═ ln [ B^m+1(R)/B^m(R)]。

To compare the performance difference between the sample entropy and the multi-dimensional sample entropy, 13 patients were selected. Both the sample entropy and the multi-dimensional sample entropy for each patient were calculated at the same EEG data interval before and after the seizure. It can be seen from fig. 3(a), (b) that the sample entropy at the time of the epileptic seizure tends to increase in value, but the difference between before and after the epileptic seizure is still not significant; the multi-dimensional sample entropy value also shows a trend of increasing during the epileptic seizure, and the difference of the multi-dimensional sample entropy value before and after the epileptic seizure is more obvious than the sample entropy value. As shown in table 1, the accuracy of multidimensional sample entropy (ACC), recall, Specificity (SPC) and Positive Predictive Value (PPV) were all higher than the sample entropy, showing superior performance.

TABLE 1 comparison of Performance of sample entropy and multidimensional sample entropy

Five sets of EEG signals with

sample points

256, 512, 1024, 2048, and 4096 were truncated for comparison to find the advantage of optimized multi-dimensional sample entropy in computational efficiency. And respectively calculating the multidimensional sample entropy and the optimized multidimensional sample entropy of each segment, and recording the required time. Table 2 lists the average times for calculating the sample entropy and the optimized multi-dimensional sample entropy. It can be seen that the optimization algorithm reduces the computation time. Fig. 4 shows a curve of the calculation time as a function of the signal length, and the optimized multidimensional sample entropy has more prominent advantages in terms of computational efficiency as the signal length increases, and is therefore more suitable for clinical diagnosis.

TABLE 2 comparison of computation time between multi-dimensional sample entropy and optimized multi-dimensional sample entropy

And (3) detecting epileptic seizure signals by using Bi-LSTM according to the optimized multidimensional sample entropy. And taking the optimized multidimensional sample entropy calculated before the current time as the input of the Bi-LSTM, and outputting the predicted next multidimensional sample entropy by utilizing the function of predicting the time sequence of the Bi-LSTM. And then the predicted multidimensional sample entropy is divided into two categories, namely attack period and normal period, through the classification function of the Bi-LSTM.

There are two ways to update the state of the Bi-LSTM network. One is to update the predicted network state using the predicted values and use the previous predicted values as inputs to the function; the other is to update the predicted network state using the observations and predict the next time using the observations at the previous time. RMSE is used to calculate the deviation between the observed and true values, as shown in fig. 5(a), (b), and the prediction results are more accurate when updating the network state using the observed values. The invention uses the observations to update the Bi-LSTM network state and performs 250 rounds of training using the "Adam" optimizer. To prevent gradient explosions, the gradient threshold is set to 1. The initial learning rate was designated as 0.05 and multiplied by 0.2 after 125 rounds of training to reduce the learning rate. If the learning rate is too small, the convergence rate will be very slow; if too large, the loss function will oscillate, even deviating from a minimum. Therefore, a larger learning rate is set first, and when the variation between two iterations is below a threshold, the learning rate will decrease.

Fig. 6 shows RMSE for the predicted times of 2 minutes, 5 minutes and 10 minutes, respectively, and it can be seen that the RMSE gradually increases as the predicted time increases. This indicates that the longer the prediction time, the larger the error. Since there is not enough time to prevent or control a seizure, an alarm generated within 2 minutes before the seizure (intervention time) can be ignored. The prediction time of the present invention was chosen to be 5 minutes, combining the values of the multidimensional sample entropy and the RMSE.

In this study, a new characteristic multidimensional sample entropy was proposed to distinguish epileptic seizures from normal states, and a method for predicting epileptic seizures using Bi-LSTM was proposed. Firstly, the characteristics of the multi-dimensional sample entropy are extracted from 23 channels of electroencephalogram signals, and the algorithm is optimized to improve the calculation efficiency. The next multi-dimensional sample entropy is then predicted using Bi-LSTM. And classifying the predicted multi-dimensional sample entropy to determine whether the epilepsy is about to attack. The method obtains good results, has higher accuracy and lower false alarm rate, and enables doctors and patients to have more sufficient time to take countermeasures.

Claims

1. The epileptic seizure signal detection system based on the optimized multidimensional sample entropy comprises a signal acquisition and preprocessing module, a feature extraction module and a prediction module, and is characterized in that the system operates through the following steps:

step (1), acquiring and preprocessing electroencephalogram data through a signal acquisition and preprocessing module;

calculating optimized multidimensional sample entropy through a feature extraction module;

firstly, constructing a multi-dimensional vector;

the method comprises the steps that initial data calculated by optimized multi-dimensional sample entropy are k-channel EEG signals, each channel of signal is provided with N sampling points, and m is an embedding dimension; each point in the constructed vector is a k-dimensional vector; the EEG signal collected in each channel has n sampling points, the sampling point of the first channel signal is x₁₁、x₁₂、x₁₃、…、x_1nThe sampling point of the second signal is x₂₁、x₂₂、x₂₃、…、x_2nAnd so on, the sampling point of the k-th channel signal is x_k1、x_k2、x_k3、…、x_kn(ii) a Extracting the a-th sampling point of each signal to obtain a multidimensional vector Q (a) ═ x_1a,x_2a,x_3a,…,x_ka) (ii) a The embedding dimension is m ═ 2, so point x (a) ═ Q (a), Q (a +1) is defined]；

Calculating the distance D [ Q (a), Q (a +1) ] between Q (a) and Q (a + 1); since Q (a) is a multi-dimensional vector, it is a multi-dimensional vector

Wherein O is_DIs the Euclidean distance; all the distances obtained by calculation are stored in a table so as to avoid repeated calculation in the next cycle;

when m is 2, X (a) ═ Q (a), Q (a +1) ], X (a +1) ═ Q (a +1), Q (a +2) ], where D [ Q (a), Q (a +1) ] and D [ Q (a +1), Q (a +2) ], the larger of which is taken as the distance between X (a) and X (a + 1);

Calculate all

Average value of (2)

When m is 3, X (a) ═ Q (a), Q (a +1), Q (a +2) ], X (a +1) ═ Q (a +1), Q (a +2), Q (a +3) ], where D [ Q (a), Q (a +1) ], D [ Q (a +1), Q (a +2) ], and D [ Q (a +2), Q (a +3) ], the largest being taken as the distance between X (a) and X (a + 1);

calculating the number X of distances between X (a) and X (a +1) less than R, assuming a similar tolerance R₂，

Calculate all

Average value of (2)

Computationally optimized multidimensional sample entropy SampEn ═ -ln [ B^m+1(R)/B^m(R)]；

Step (3), detecting epileptic seizure signals by using the Bi-LSTM through a prediction module according to the optimized multidimensional sample entropy obtained by the feature extraction module; taking the optimized multidimensional sample entropy calculated before the current moment as the input of the Bi-LSTM, and outputting the predicted next multidimensional sample entropy by utilizing the function of predicting the time sequence of the Bi-LSTM; and then the predicted multidimensional sample entropy is divided into two categories, namely a seizure period and a normal period, through the classification function of the Bi-LSTM, so as to achieve the purpose of epileptic seizure signal detection.