A method for epileptic seizure prediction based on complex brain network and Fiedler value analysis

Abstract

The application discloses a kind of based on complex brain network and Fiedler value analysis's seizure prediction method, belong to electroencephalogram seizure prediction field.Aim at the seizure EEG data is filtered and preprocessed, 23 groups of bipolar lead is regarded as 23 points, and it is divided into six brain regions.Transmission entropy between each electrode signal is calculated, and threshold processing is carried out to generate adjacency matrix, to build brain network and calculate its Fiedler value.The standard curve is obtained by polynomial fitting curve to describe the change of Fiedler value with time, and compared with the actual fitting curve of single patient, dynamic time warping distance DTW is calculated.The results show that there is high similarity between most standard curves and actual curves, which verifies the effectiveness and accuracy of the method.Overall, the method shows high reliability in identifying characteristics of patients with epilepsy, and can provide important reference for subsequent diagnosis and treatment.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram (EEG) epileptic seizure prediction, and adopts a complex brain network construction method based on transfer entropy, combined with Fiedler value analysis and dynamic time warping (DTW) evaluation. Background Art

[0002] Epilepsy is a recurrent neurological disease that seriously affects the quality of life and health of patients. The prediction of epileptic seizures is of great significance for the management and treatment of patients. Existing studies have shown that epileptic seizures are closely related to the abnormal synchronous discharge of neurons in the brain, and electroencephalogram (EEG) signals are one of the main means of studying epileptic seizures. EEG is a non-invasive brain imaging technology that reflects the electrophysiological activity of the brain by recording changes in the postsynaptic potentials of neurons during brain activity. In recent years, with the rapid development of network science and graph theory, researchers have begun to pay attention to epilepsy prediction methods based on complex brain networks. The complex network framework can capture the spatiotemporal characteristics in EEG signals, which helps to reveal potential changes in brain network structure before epileptic seizures.

[0003] Transfer entropy is an effective method for constructing complex brain networks. It can describe the coupling strength and bidirectional coupling relationships between EEG signals, and is independent of established models and nonlinear quantitative analysis. Constructing brain networks based on transfer entropy can more accurately reflect the information transfer between nodes in brain functional networks.

[0004] In brain network analysis, the Fiedler value (also known as algebraic connectivity) serves as an important indicator of network connectivity, reflecting changes in the overall state of the brain network. By analyzing the evolution of the Fiedler value over time, it is possible to capture the precursory characteristics of epileptic seizures. However, due to the complexity and diversity of EEG signals, effectively constructing brain networks and accurately calculating the Fiedler value remain challenges.

[0005] This paper uses a transfer entropy-based approach to construct a brain network. By calculating the effectiveness of information transfer between nodes in the network, the dynamic characteristics of EEG signals are revealed. Furthermore, a theoretical analysis of Fiedler values ​​is constructed to determine a standard curve. Dynamic time warping (DTW) is then used to assess the degree of match between the Fiedler value curves of individual patients and the standard curve, thereby achieving the goal of predicting epileptic seizures. This method aims to improve the accuracy and reliability of epileptic seizure prediction and provide a new technical means for epileptic seizure prediction. Summary of the Invention

[0006] Based on the above, the present invention proposes a method for predicting epileptic seizures based on complex brain networks and Fiedler value analysis. First, the original EEG data is filtered and divided by frequency band. Subsequently, the 23 groups of bipolar leads in the data set are regarded as 23 nodes and divided into six brain regions. Next, the transfer entropy between the electrodes in each brain region is calculated, and threshold processing is performed to generate an adjacency matrix. Then, a brain network is constructed, and its Fiedler value is calculated. The change of Fiedler value over time is described by a polynomial fitting curve to generate a standard curve. Finally, the Fiedler value curve of a single patient is compared with the standard curve, and the degree of matching is evaluated using dynamic time warping (DTW) distance.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for predicting epileptic seizures based on complex brain networks and Fiedler value analysis, comprising the following steps:

[0009] Step A: Filter the EEG data and select the data by stages;

[0010] Butterworth filters and zero-phase filters were used to filter signals of different frequency ranges and time structures. Four frequency bands of EEG signals were obtained after filtering: full-band 0.5-60Hz, alpha band 8-13Hz, beta band 14-29Hz, and gamma band 30-60Hz. The overall data set for each frequency band was then divided into stages according to the temporal sequence of epileptic seizures, resulting in staged data segments at the time of onset, five minutes before onset, 5-10 minutes before onset, 10-15 minutes before onset, 15-20 minutes before onset, and 20-25 minutes before onset. A total of 24 staged data segments were obtained for different frequency bands and stages.

[0011] Step B: Divide brain regions;

[0012] The dataset used a bipolar lead method to record EEG, with each subject using 23 bipolar channels for recording. These 23 bipolar channels were used as nodes of the brain network and divided into six brain regions: the left region: FP1-F7, F7-T7, T7-P7, P7-O1; the left middle region: FP1-F3, F3-C3, C3-P3, P3-O1; the right middle region: FP2-F4, F4-C4, C4-P4, P4-O2; the right region: FP2-F8, F8-T8, T8-P8, P8-O2; the central region: FZ-CZ, CZ-PZ; the peripheral region: P7-T7, T7-FT9, FT9-FT10, FT10-T8, T8-P8;

[0013] Step C: Calculate the transfer entropy between nodes in each brain region corresponding to each subject and each epileptic seizure stage data segment to obtain the corresponding transfer entropy matrix;

[0014] Step D: Perform threshold processing on the transfer entropy matrix, generate the adjacency matrix and construct the network;

[0015] First, the transfer entropy matrices obtained from each stage data segment of all subjects and all epileptic seizures were processed with corresponding means, resulting in a total of 24 mean matrices. Then, the data in each mean matrix was sorted according to numerical size, and the percentile was selected as the threshold as needed. The data in the mean matrix was compared with the threshold. Data below the threshold was set to 0, and other data remained unchanged to obtain an adjacency matrix. Next, the adjacency matrix was converted into a graph to form a brain network. The edges between nodes in the brain network graph are directional, with arrows indicating the direction of the connection. The thickness of the edge reflects the strength of the connection between nodes, and the thicker the edge, the stronger the connection relationship.

[0016] Step E: Calculate the Fiedler value in each stage data segment corresponding to the six brain regions;

[0017] Step F: For each frequency band and each brain region, a polynomial fitting method is used to draw a fitting curve of the Fiedler value changes over time. For each frequency band, a standard curve for each of the six brain regions is obtained. All frequency bands are processed in the same way to obtain a total of 24 standard curves. Then, the data of a single patient are tested to obtain the corresponding 24 curves for each patient as the actual curves.

[0018] Step G: Use dynamic time warping (DTW) to measure the similarity between the patient's actual curve and the corresponding standard curve;

[0019] Step H: Determine weights and integrate dynamic time warping distance values;

[0020] Based on the data in the 24 adjacency matrices obtained in step D, weights are assigned to different frequency bands and different brain regions. The dynamic time warping distance value of each brain region and each frequency band is then multiplied by the weight and summed up to obtain the dynamic time warping distance value for each frequency band. The calculation formula is as follows:

[0021]

[0022] Among them, Total_sum is the sum of all values ​​in the same frequency band in all adjacency matrices, TE k is the data in the adjacency matrix of the kth brain region, Weight k is the weight of each brain region, is the integrated dynamic time warping distance value in each frequency band, is the dynamic time warping distance value of the kth brain area in frequency band f.

[0023] Furthermore, the calculation method of the transfer entropy in step C is:

[0024] For two EEG signal time series x(t) and y(t) that can be approximated by a Markov process, x(t) is represented by X and y(t) is represented by Y, then a causal relationship measure can be used to calculate the deviation of the probability represented by the following formula from the generalized Markov condition;

[0025]

[0026] in and are the mth and nth order Markov processes in x and y, respectively; the right side of the equation is the probability of its value calculated given its history of the first n steps, while the left side estimates this probability considering the history of x(t) and y(t); when the transition probability of y(t) is independent of the past of x(t), that is, when there is no causal relationship between x(t) and y(t), the above equation is fully satisfied, but in order to calculate the causal relationship between x(t) and y(t), the Kullback-Leibler divergence between the two probability distributions is used to define the transfer entropy from x(t) to y(t):

[0027]

[0028] This formula calculates the directional information flow from x(t) to y(t); based on the above definition, the transfer entropy of the time series x(t) to y(t) can be calculated by the following formula:

[0029]

[0030] Where t is the time index of the discrete value in the sequence, u represents the prediction time, which is a discrete value time interval; and d y and d x dimensional delay embedding vector, details are as follows:

[0031]

[0032] Here, τ represents the delay time.

[0033] Furthermore, the calculation process of the Fiedler value in step E is as follows:

[0034] L=DA

[0035] |λI-E|=0

[0036] Where L is the Laplacian matrix, D is the degree matrix, A is the adjacency matrix, λ is the eigenvalue, and I is the identity matrix of the same size as L; the eigenvalues ​​of the Laplacian matrix are arranged in ascending order, and the smallest eigenvalue is always 0, corresponding to the eigenvector with the same elements; the Fiedler value is the second smallest eigenvalue, that is: λ2.

[0037] Furthermore, the process of obtaining the fitting curve in step F is as follows:

[0038] Before fitting the curve, it is necessary to use linear interpolation to expand the Fiedler values ​​in the six phase data segments of each frequency band and each brain region into N points. The linear interpolation formula is:

[0039]

[0040] in is the interpolation point, and is the horizontal coordinate of the known data point, and is the ordinate of the known data point;

[0041] Use the polynomial fitting method to fit the N points after interpolation with a fifth-order polynomial. The fitted polynomial P(x) formula is:

[0042]

[0043] The fitting coefficients a5, a4, a3, a2, a1, and a0 are determined by the least squares method.

[0044] Furthermore, the calculation method of DTW in step G is:

[0045] Suppose we have two time series and They are

[0046]

[0047] Calculate the distance matrix D between two sequences using the Euclidean distance:

[0048]

[0049] Construct the cumulative distance matrix C, where C(i,j) represents the minimum cumulative distance from (1,1) to (o,j); the initial conditions and recursive formula are as follows

[0050] C(1,1)=D(1,1)

[0051] C(o,1)=D(o,1)+C(o-1,1),o=2,…,w

[0052] C(1,j)=D(1,j)+C(1,j-1),j=2,…,q

[0053] C(o,j)=D(o,j)+min{C(o-1,j),C(o,j-1),C(o-1,j-1)},o=2,…,w

[0054]

[0055] in is the final DTW distance.

[0056] The present invention has the following beneficial effects:

[0057] (1) The present invention uses transfer entropy to construct brain networks. Transfer entropy can accurately describe the bidirectional coupling relationship and information flow between EEG signals, overcoming the limitations of traditional methods such as coherence and phase-locking value in reflecting the mutual influence between brain network nodes, thereby more accurately capturing the complex dynamic characteristics of brain networks.

[0058] (2) The present invention uses Fiedler values ​​to analyze brain networks. As the second smallest eigenvalue of a graph, Fiedler values ​​can effectively reflect the connectivity of brain networks. By fitting and analyzing the changes in Fiedler values ​​over time, changes in network structure before epileptic seizures can be identified, providing early warning signals.

[0059] (3) The present invention uses dynamic time warping (DTW) to assess the degree of match between the patient and the standard curve. DTW can flexibly handle the nonlinear alignment problem of time series data, allowing for precise matching of the Fiedler value curve of a single patient with the standard curve, thereby improving the accuracy of epileptic seizure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the overall framework diagram of the present invention;

[0061] Figure 2 It is the distribution map of brain electrodes;

[0062] Figure 3 The adjacency matrix at the onset of the attack (left) and the adjacency matrix five minutes before the onset of the attack (right) in all frequency bands.

[0063] Figure 4 This is an image of brain network connectivity during a seizure (left) and five minutes before the seizure (right).

[0064] Figure 5is the standard curve of polynomial fitting of the full frequency band and the actual curve. DETAILED DESCRIPTION

[0065] In order to describe the present application more specifically, the technical solutions of the present application are described in detail below in combination with the drawings and specific embodiments.

[0066] Step A: Filtering and staging of EEG data. The public dataset CHB-MIT dataset was used, which contains scalp electroencephalogram (EEG) data collected from 23 epilepsy patients after discontinuing antiepileptic drugs for a period of time. Butterworth filter and zero-phase filter were used to filter signals of different frequency ranges and time structures. Butterworth filter is a low-pass, high-pass or band-pass filter that can select the frequency range to be passed by setting the cutoff frequency. Zero-phase filter is a linear phase filter that does not introduce additional phase delay. It is often used in situations where the signal time structure needs to be maintained, such as in EEG data processing, which can maintain the time structure of event-related potentials. The filtered signals of four frequency bands are full frequency band (0.5-60Hz), alpha band (8-13Hz), beta band (14-29Hz) and gamma band (30-60Hz). Then the overall data set of each frequency band is staged according to the time sequence of seizures, and the staging data segments of seizure, five minutes before seizure, 5-10 minutes before seizure, 10-15 minutes before seizure, 15-20 minutes before seizure, 20-25 minutes before seizure are obtained, a total of 24 staging data segments of different frequency bands and different stages are obtained.

[0067] Step B: Division of brain regions. The dataset uses bipolar lead method to record EEG, which uses two electrodes to record the voltage difference between the two electrodes to reflect the neural activity in the brain region of interest. The distribution of the brain electrodes is shown in Figure 2 Each subject uses 23 bipolar channels for recording. These 23 bipolar channels are used as nodes of brain network and divided into six brain regions, namely left region: FP1-F7, F7-T7, T7-P7, P7-O1, left middle region: FP1-F3, F3-C3, C3-P3, P3-O1, right middle region: FP2-F4, F4-C4, C4-P4, P4-O2, right region: FP2-F8, F8-T8, T8-P8, P8-O2, central region: FZ-CZ, CZ-PZ, peripheral region: P7-T7, T7-FT9, FT9-FT10, FT10-T8, T8-P8.

[0068] Step C: Calculate the transfer entropy between nodes within each brain region for each subject and each epileptic seizure phase data segment to obtain the corresponding transfer entropy matrix. For two EEG signal time series x(t) and y(t) that can be approximated by a Markov process, with x(t) represented by X and y(t) represented by Y, a causal relationship measure can be used to calculate the deviation of the probability represented by the following formula from the generalized Markov condition.

[0069]

[0070] in and are the mth and nth order Markov processes in x and y, respectively. The right side of the equation is the probability of its value calculated given its history of the previous n steps, while the left side estimates this probability considering the history of x(t) and y(t). When the transition probability of y(t) is independent of the past of x(t), that is, when there is no causal relationship between x(t) and y(t), the above equation is fully satisfied. However, in order to calculate the causal relationship between x(t) and y(t), the Kullback-Leibler divergence between the two probability distributions is used to define the transfer entropy from x(t) to y(t):

[0071]

[0072] This formula calculates the directional information flow from x(t) to y(t). Based on the above definition, the transfer entropy of the time series x(t) to y(t) can be calculated by the following formula:

[0073]

[0074] Where t is the time index of the discrete value in the sequence, and u represents the prediction time, which is a discrete value time interval. and d y and d x dimensional delay embedding vector, details are as follows:

[0075]

[0076] Here, τ represents the delay time.

[0077] Step D: Threshold processing is performed on the transfer entropy matrix for each time period to generate an adjacency matrix and construct a network. First, the transfer entropy matrices obtained from each stage data segment of all subjects and all epileptic seizures are processed with corresponding mean values, resulting in a total of 24 mean matrices. Then, the data in each mean matrix is ​​sorted according to the size of the value, and the percentile is selected as the threshold as needed, such as the 70th percentile, which means that 70% of the values ​​in the data are less than or equal to this number. The data in the mean matrix is ​​compared with the threshold, and the data less than the threshold is set to 0, while the other data remains unchanged to obtain the adjacency matrix. Next, the adjacency matrix is ​​converted into a graph to form a brain network. The edges between nodes in the brain network graph are directional, and the arrows indicate the direction of the connection. The thickness of the edge reflects the strength of the connection between nodes. The thicker the edge, the stronger the connection relationship.

[0078] The adjacency matrix at the onset of the full frequency band (left) and the adjacency matrix five minutes before the onset (right) are as follows Figure 3 As shown in the figure. The brain network connection during the attack (left) and the brain network connection five minutes before the attack (right) are shown in the figure. Figure 4 shown.

[0079] Step E: Calculate the Fiedler value for each data segment corresponding to the six brain regions. The size of the Fiedler value affects the speed of propagation of information or disturbances in the network. A network with a large Fiedler value means that changes on any node will propagate quickly to other nodes, which helps the network reach a consistent state, that is, synchronization. On the contrary, a network with a low Fiedler value may have long paths or weak connections, which will hinder the synchronization process. The calculation process of the Fiedler value is as follows

[0080] L=DA

[0081] |λI-E|=0

[0082] Where L is the Laplacian matrix, D is the degree matrix, A is the adjacency matrix, λ is the eigenvalue, and I is the identity matrix of the same size as L. The eigenvalues ​​of a Laplacian matrix are usually arranged in ascending order. The smallest eigenvalue is always 0, corresponding to an eigenvector with all elements being identical. The Fiedler value is the second smallest eigenvalue, i.e., λ².

[0083] As shown in Table 1, the Fiedler value is low before a seizure and increases significantly during a seizure. A network with a high Fiedler value means that changes at any node propagate rapidly to other nodes, which helps the network achieve a consistent state, known as synchronization. Epilepsy is believed to be caused by abnormally synchronized neuronal discharges. Conversely, a network with a low Fiedler value may have long paths or weak connections, which can hinder synchronization.

[0084] Table 1 Full-band Fiedler values ​​(0.5-60Hz)

[0085]

[0086]

[0087] Step F: For the Fiedler values ​​obtained in the six staged data segments for each frequency band and each brain region, a fitting curve of the Fiedler values ​​changing over time is drawn using a polynomial fitting method. Standard curves for the six brain regions will be obtained for each frequency band. The same processing is performed on all frequency bands to obtain a total of 24 standard curves. The data of a single patient are then subjected to experiments to obtain the corresponding 24 curves for each patient as the actual curves.

[0088] Before fitting the curve, it is necessary to use linear interpolation to expand the Fiedler values ​​in the six phase data segments of each frequency band and each brain region to 50 points. The formula for linear interpolation is:

[0089]

[0090] in is the interpolation point, and is the horizontal coordinate of the known data point, and is the vertical coordinate of the known data point.

[0091] Use the polynomial fitting method to perform a fifth-order polynomial fitting on the 50 interpolated points, and the fitted polynomial P(x) formula is:

[0092]

[0093] The fitting coefficients a5, a4, a3, a2, a1, and a0 are determined by the least squares method.

[0094] The fitting curves of each brain region obtained under the full frequency band are as follows Figure 5 shown.

[0095] Step G: Use the dynamic time warping distance (DTW) to measure the similarity between the patient's actual curve and the corresponding standard curve, and obtain the dynamic time warping distance values ​​of the four frequency bands corresponding to the six brain regions. The core idea of ​​the DTW algorithm is to find an optimal way to align the two time series so that the total distance between them is minimized. Specifically, DTW aligns the two sequences through interpolation and delay operations, and then calculates the distance between their corresponding positions, and finally obtains a total warping distance. This method allows a certain amount of time distortion between sequences, so it is suitable for sequences that are offset on the time axis or have different speeds. The smaller the value of the dynamic time warping distance, the more similar the two sequences are in shape. Suppose we have two time series and They are

[0096]

[0097] Calculate the distance matrix D between two sequences using the Euclidean distance:

[0098]

[0099] Construct the cumulative distance matrix C, where C(i,j) represents the minimum cumulative distance from (1,1) to (o,j). The initial conditions and recursive formula are as follows

[0100] C(1,1)=D(1,1)

[0101] C(o,1)=D(o,1)+C(o-1,1),o=2,…,w

[0102] C(1,j)=D(1,j)+C(1,j-1),j=2,…,q

[0103] C(o,j)=D(o,j)+min{C(o-1,j),C(o,j-1),C(o-1,j-1)},o=2,…,w

[0104]

[0105] in is the final DTW distance.

[0106] Table 2 shows the DTW values ​​for each brain region at each frequency band. Overall, with the exception of a few data points with significant discrepancies, the DTW values ​​between the standard curves and the actual curves are generally low, indicating a high degree of similarity between the standard and actual curves. This suggests that the experimental data generally align with the trends and patterns of variation in the actual data.

[0107] Table 2 DTW values ​​of each brain region in each frequency band

[0108]

[0109]

[0110] Step H: According to the data of the 24 adjacency matrices obtained in step D, weights are assigned to different frequency bands and different brain regions, then the dynamic time warping distance value of each brain region and each frequency band is multiplied by the weight and then summed up to integrate the dynamic time warping distance value under each frequency band. The calculation formula is as follows:

[0111]

[0112] Where Total_sum is the sum of all values in the same frequency band in all adjacency matrices, TE k is the data in the adjacency matrix of the kth brain region, Weight k is the weight of each brain region, is the integrated dynamic time warping distance value under each frequency band, is the dynamic time warping distance value of the kth brain region under the frequency band f.

[0113] Table 3 The weight of each brain region

[0114]

[0115]

[0116] Table 4 Weighted DTW of each frequency band

[0117]

[0118] According to the operation process of the above steps, the epilepsy prediction method based on complex brain network and Fiedler value analysis proposed by the present invention can be realized. In order to verify the technical effect of the present invention in epilepsy prediction, the present invention used the public data set CHB-MIT to conduct experiments; the CHB-MIT data set (Children's Hospital Boston-Massachusetts Institute of Technology) is a public epilepsy data set that contains long-term EEG records of 22 epilepsy patients. This data set collects EEG signals of patients in different time periods, including signals during normal state and epileptic seizures. Table 1 shows the Fiedler values ​​under different time windows. It can be seen that the Fiedler value will increase significantly during the attack, indicating that the Fiedler value can be used for complex brain network epilepsy prediction. Table 2 shows the DTW values ​​of each brain region under each frequency band. From the overall table, except for a small amount of data with large differences, the DTW values ​​between most standard curves and actual curves are at a low level, indicating that there is a high similarity between the standard curve and the actual curve. This shows that the experimental data are generally consistent with the trend and change law of the actual data. To integrate the DTW values, each brain region was assigned a weight based on the transfer entropy in the adjacency matrix, as shown in Table 3. The DTW values ​​for the corresponding frequency band and brain region were then multiplied by the weights to obtain the total DTW value for each frequency band, as shown in Table 4. It can be seen that the DTW values ​​for the full, beta, and gamma bands are all relatively small and highly similar. The DTW value for the alpha band is slightly larger. This may be due to the following reasons: First, the actual data is from a single individual, which may include individual differences. For example, this patient may have other diseases or may not have seizures in some brain regions in the alpha band. Second, the base value of the transfer entropy after thresholding in this frequency band is relatively large, resulting in a larger DTW value.

[0119] Overall, the DTW value shows that this method has high reliability in identifying the characteristics of epilepsy patients, which can provide an important reference for subsequent diagnosis and treatment.

Claims

1. A method for predicting epileptic seizures based on complex brain networks and Fiedler value analysis, characterized in that: Here are the steps: Step A: Filter the EEG data and select the data by stages; Butterworth filters and zero-phase filters were used to filter signals of different frequency ranges and time structures. Four frequency bands of EEG signals were obtained after filtering: full-band 0.5-60Hz, alpha band 8-13Hz, beta band 14-29Hz, and gamma band 30-60Hz. The overall data set for each frequency band was then divided into stages according to the temporal sequence of epileptic seizures, resulting in staged data segments at the time of onset, five minutes before onset, 5-10 minutes before onset, 10-15 minutes before onset, 15-20 minutes before onset, and 20-25 minutes before onset. A total of 24 staged data segments were obtained for different frequency bands and stages. Step B: Divide brain regions; The dataset used a bipolar lead method to record EEG, with each subject using 23 bipolar channels for recording. These 23 bipolar channels were used as nodes of the brain network and divided into six brain regions: the left region: FP1-F7, F7-T7, T7-P7, P7-O1; the left middle region: FP1-F3, F3-C3, C3-P3, P3-O1; the right middle region: FP2-F4, F4-C4, C4-P4, P4-O2; the right region: FP2-F8, F8-T8, T8-P8, P8-O2; the central region: FZ-CZ, CZ-PZ; the peripheral region: P7-T7, T7-FT9, FT9-FT10, FT10-T8, T8-P8; Step C: Calculate the transfer entropy between nodes in each brain region corresponding to each subject and each epileptic seizure stage data segment to obtain the corresponding transfer entropy matrix; Step D: Perform threshold processing on the transfer entropy matrix, generate the adjacency matrix and construct the network; First, the transfer entropy matrices obtained from each stage data segment of all subjects and all epileptic seizures were processed with corresponding means, resulting in a total of 24 mean matrices. Then, the data in each mean matrix was sorted according to numerical size, and the percentile was selected as the threshold as needed. The data in the mean matrix was compared with the threshold. Data below the threshold was set to 0, and other data remained unchanged to obtain an adjacency matrix. Next, the adjacency matrix was converted into a graph to form a brain network. The edges between nodes in the brain network graph are directional, with arrows indicating the direction of the connection. The thickness of the edge reflects the strength of the connection between nodes, and the thicker the edge, the stronger the connection relationship. Step E: Calculate the Fiedler value in each stage data segment corresponding to the six brain regions; Step F: For each frequency band and each brain region, a polynomial fitting method is used to draw a fitting curve of the Fiedler value changes over time. For each frequency band, a standard curve for each of the six brain regions is obtained. All frequency bands are processed in the same way to obtain a total of 24 standard curves. Then, the data of a single patient are tested to obtain the corresponding 24 curves for each patient as the actual curves. Step G: Use dynamic time warping (DTW) to measure the similarity between the patient's actual curve and the corresponding standard curve; Step H: Determine weights and integrate dynamic time warping distance values; According to the data in the 24 adjacency matrices obtained in step D, weights are assigned to different frequency bands and different brain regions. Then, the dynamic time warping distance value of each brain region and each frequency band is multiplied by the weight and then added up to obtain the dynamic time warping distance value under each frequency band. The calculation formula is as follows: ； in, is the sum of all values ​​in the same frequency band in all adjacency matrices, It is The data in the adjacency matrix of each brain region, is the weight of each brain region, is the integrated dynamic time warping distance value in each frequency band, It is brain regions in the frequency band The dynamic time warping distance value under .

2. The epileptic seizure prediction method based on complex brain network and Fiedler value analysis according to claim 1, characterized in that: The calculation method of transfer entropy in step C is: For two EEG signal time series and Approximated by a Markov process, use express, use If , then a causality measure is used to calculate the deviation of the probability expressed by the following formula from the generalized Markov condition; ； in and They are and Markov process Order and order; the right side of the formula is given by its front The probability of its value is calculated after the history of the step, while the left side is considering and This probability is estimated when the history of The transition probability and The past has nothing to do with arrive When there is no causal relationship between the above equations, they are completely satisfied. However, in order to calculate and The causal relationship is defined using the Kullback-Leibler divergence between two probability distributions. arrive The transfer entropy of: ； This formula calculates the arrive Directed information flow; Based on the above definition, the time series is calculated by the following formula arrive The transfer entropy of: ； in is the time index of the discrete value in the sequence, Represents the prediction time, which is a discrete value time interval; and They are and dimensional delay embedding vector, details are as follows: ； in, Represents the delay time.

3. The epileptic seizure prediction method based on complex brain network and Fiedler value analysis according to claim 1, characterized in that: The calculation process of the Fiedler value in step E is as follows: ； in is the Laplace matrix, is the degree matrix, is the adjacency matrix, is the eigenvalue, is with The identity matrix has the same size; the eigenvalues ​​of the Laplacian matrix are arranged in ascending order, the smallest eigenvalue is always 0, corresponding to the eigenvector with the same elements; the Fiedler value is the second smallest eigenvalue, that is: .

4. The epileptic seizure prediction method based on complex brain network and Fiedler value analysis according to claim 1, characterized in that: The process of obtaining the fitting curve in step F: Before fitting the curve, it is necessary to use linear interpolation to expand the Fiedler values ​​in the six phase data segments of each frequency band and each brain region into N points. The linear interpolation formula is: ； in is the interpolation point, and is the horizontal coordinate of the known data point, and is the ordinate of the known data point; Use the polynomial fitting method to fit the N points after interpolation with a fifth-order polynomial to obtain the fitted polynomial The formula is: ； Among them, the coefficients of the fitting Determined by the least squares method.

5. The epileptic seizure prediction method based on complex brain network and Fiedler value analysis according to claim 1, characterized in that: The calculation method of DTW in step G is: Suppose we have two time series and , respectively ； Calculate the distance matrix between two sequences , using Euclidean distance: ； Constructing a cumulative distance matrix ,in From (1,1) to The minimum cumulative distance; the initial conditions and recursive formula are as follows ； in is the final DTW distance.

Images