CN111419222A

CN111419222A - BNI-based epileptic seizure signal detection method

Info

Publication number: CN111419222A
Application number: CN202010154646.0A
Authority: CN
Inventors: 胡月静; 柏雨露; 汪茜; 高云园; 张启忠
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2020-03-08
Filing date: 2020-03-08
Publication date: 2020-07-17
Anticipated expiration: 2040-03-08
Also published as: CN111419222B

Abstract

The invention discloses a BNI-based epileptic seizure signal detection method, which discusses the effectiveness of a prediction method from the perspective of microscopic neurons: from a microscopic neuron perspective, a nerve quality model (NMM) was used to fit brain deep electrode electroencephalogram (Depth EEG) signals and elucidate the relationship between network structure, kinetic equations, and the generation of epileptic discharges. The concept of brain network epileptogenic index (BNI) was introduced in order to quantify the pathological extent to which a given network can cause epileptic seizures. This is the first time BNI was used as a predictor of seizures. The invention shortens the detection time, reduces the implantation number of the electrodes and can observe good prediction effect.

Description

BNI-based epileptic seizure signal detection method

Technical Field

The invention belongs to the field of Brain Network structure analysis, and relates to a method for detecting epileptic seizure signals by calculating Brain Network Information (BNI) through neural quality modeling based on Brain function Network characteristics.

Background

Epilepsy, a neurological disorder characterized by paroxysmal seizures, has a severe impact on the patient's body, spirit and cognition in long-term frequent seizures. Unfortunately, epilepsy is the fourth most common neurological disorder affecting over 6500 million people worldwide. In response to this situation, there is a clear need for a new method that can predict seizures in these patients and protect them from injury.

With the definition of epilepsy classification, the study shows that the focal epilepsy has certain predictability. For example, a survey conducted in the united states indicates that there may be pre-seizure states from inter-seizure to ictal, which are believed to occur prior to an epileptic seizure, associated with the presence of inter-seizure states most of the time prior to the seizure. If such pre-seizure states can be captured, then seizures can be predicted.

"focal point" (EZ) is an epidemic concept associated with epilepsy and it means "abnormal discharge point of seizure, an indispensable cortical region of epileptic seizure". By implanting brain Depth electrode electroencephalography (Depth EEG) electrodes within the EZ, the neuronal activity of its constituents, as well as the activity observed on other Depth EEG channels, can be assessed and used to build brain Functional networks (FBNs) that describe underlying epileptic mechanisms. Network-based methods are considered to be effective ways of understanding the occurrence of epileptic seizures. Where the term network refers to a collection of "nodes" and "edges" — the FBNs nodes represent the area of the brain with implanted electrodes, while the edges describe the relationship between any two given nodes, usually assessed by a measure of correlation or causal relationship. Mathematical techniques generated by the theory of complexity can then be used to elucidate the link between node dynamics, network connectivity and seizures.

In addition to epilepsy studies, computational models relating to neuronal population levels are of great interest. These models are based on experimental data and a Neural Mass Model (NMM), where the dynamics of each Neural ensemble are represented by second order differential equations.

In addition, NMM employs a mathematical modeling framework that allows for the study of the relationships between the network structure, nodal dynamics, and firing, in which a loop with two different inhibition time scales of interaction between the major neurons, excitatory interneurons, and inhibitory interneurons is embodied.

To address this issue, the focus of the study was on how to quantify the pre-seizure provenance with BNI values derived from the Depth EEG signal and provide a high-level warning of seizures. For each patient, FBNs were calculated based on EZ and the entire brain area (WBA). NMM was further constructed to account for the effects of network topology due to discharges during pre-seizure and seizure periods furthermore, changes in FBNs were compared during these periods. In particular, a new indicator for predicting epilepsy by quantifying seizure potential is employed. To our knowledge, this is the first time BNI is used as a predictor. Finally, the ability of previously reported modeling frameworks to predict epilepsy and optimize prediction time was also tested.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a BNI-based epileptic seizure signal detection method.

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. All signals are sampled by a depth electrode distribution system consisting of 5-18 contacts with the length of 2 mm, the diameter of 0.8 mm and the contact distance of 1.5 mm. The data preprocessing mode is to use a Butterworth filter with a high cut-off frequency of 85Hz for band-pass filtering.

And (2) carrying out FBNs characteristic analysis based on the Depth EEG signal. Introducing the concept of graph theory, establishing an FBNs correlation matrix by applying network analysis, and calculating a corresponding Pearson correlation coefficient and a threshold value. When two channels are changed into a plurality of channels, the Pearson correlation coefficient is changed into a matrix, namely a correlation matrix, from one number, and a proper threshold value is selected to carry out binarization on the correlation matrix. For the matrix after binarization, 0 in the matrix represents a node (i.e. a channel), and 1 represents a degree, so that a brain function network visual map can be drawn.

The specific concept of the graph theory is as follows:

graph theory defines a network as a set of nodes whose pair-wise interactions are defined by edges linking them. In the field of neuroimaging, nodes typically represent brain regions, while edges represent causal or correlative relationships between them.

The specific algorithm of the Pearson phase relation and the threshold is as follows:

the nodes of the FBNs represent brain regions near the electrodes, and the edges between the nodes are determined as Pearson correlation coefficients of the observed nodal activity:

where cov is the covariance of channel i and channel j, and σ_iAnd σ_jRepresenting the respective standard deviations of the activities in nodes i and j.

After determining the node correlation, a statistically significant threshold T is applied to the binarized correlation matrix:

T＝1-(1-α)^1/(L-1)

where L is the average number of disjoint sections, referred to in the process outlined above, α is the required confidence level (α ═ 0.95 herein).

The specific method for binarizing the correlation matrix is as follows:

for any matrix A, the binarization is to select a proper threshold value T to compare with each number in the matrix A, if the threshold value T is larger than the threshold value T, the number of the corresponding position in the matrix A is recorded as 1, and if the threshold value T is smaller than the threshold value T, the number of the corresponding position in the matrix A is recorded as 0. According to the method, each value in the matrix is processed in turn to obtain a binary matrix.

Step (3) fitting the Depth EEG signal with NMM. And (3) substituting the binarization matrix obtained in the step (2) into the NMM, and determining all parameters according to the existing literature to obtain a time-related sequence.

And (4) calculating BNI. In order to study the pathological extent to which a given network can produce epileptic seizures, the concept of BNI was introduced considering the quantitative measurement of FBNs, i.e. the sum of the time at which all nodes in a reference time are discharged divided by the reference time; and (4) calculating BNI on the time series obtained in the step (3), namely using the calculation result in the next step to analyze the epileptic seizure prediction.

The specific algorithm of BNI is as follows:

and (5) carrying out seizure prediction analysis according to BNI. From the experimental results we can assume that there is a risk of morbidity when the BNI exceeds 0.5 during the prediction period. Then in the one hour Depth EEG signal tested, the BNI is shown to trend gradually upward, eventually reaching above the 0.5 threshold at some time, which is the predicted time we want to get.

Compared with the existing algorithms of various electroencephalogram signals, the method has the following characteristics:

on one hand, the method is based on a Depth EEG signal, the greatest advantage is that the data acquisition is closer to the disease source, so that the disease condition can be sensed earlier, and the time of the prediction of the disease is doubled before 2461.74 seconds is detected compared with the time of the NPDC-based method; the invention greatly advances the time of epileptic seizure prediction;

on the other hand, studies have found that the BNI of EZ is more discriminative than the global network and can be used to calculate one of the indices for predicting epileptic seizures. Future surgical protocols attempt to study BNI-like analysis, rather than using a large number of electrodes to cover the entire brain, the electrode area can be distributed directly over the lesion. Therefore, the implantation number of the electrodes can be reduced, the pain and recovery time of a patient can be reduced, and a good prediction effect can be observed.

Drawings

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

FIG. 2 is a diagram of approximate electrode locations of Depth EEG electrodes implanted in WBA (left) and EZ (right) of a patient AR in accordance with an embodiment of the invention;

FIG. 3 is a graph of results of an AR patient fMRI examination according to an embodiment of the present invention;

FIG. 4 is a diagram of a Depth EEG signal according to an embodiment of the present invention;

FIG. 5 is a diagram of the brain network connection-coherence matrix established on EZ (up) and WBA (down) during the period of five minutes of onset (left column) and inter-episode (right column) states in accordance with an embodiment of the present invention;

fig. 6 is a graph of brain network connections-FBNs established on EZ (ascending) and WBA (descending) during inter-episode (left column) and intra-episode (right column) states for a five minute period in accordance with an embodiment of the present invention;

FIG. 7 is a detailed view of the computational components of pyramidal cells (solid line), excitatory interneurons (dotted line), and inhibitory interneurons (dotted line) according to an embodiment of the present invention. The pulse voltage module is marked as 'p- - -v', and the voltage pulse module is marked as 'v- - -p' graph;

fig. 8 is a BNI graph of brain network connections established on EZ (ascending) and WBA (descending) during the period of five minutes of onset (left column) and inter-episode (right column) states in accordance with an embodiment of the present invention;

FIG. 9 is a schematic graph of BNI values within one hour prior to a seizure in accordance with an embodiment of the present invention;

Detailed Description

In this study, we mainly discuss the effectiveness of the prediction method from the microscopic neuron perspective: from a microscopic neuron perspective, NMM was used to fit the Depth EEG signal and elucidate the relationship between the network structure, the kinetic equations and the generation of epileptic discharges. The concept of BNI was introduced in order to quantify the pathological extent to which a given network can cause seizures. This is the first time BNI was used as a predictor of seizures. The results show that the proposed method based on the Depth EEG signal gives good results in only four patient samples. On average, the predicted time to onset was detected twice as early as 2461.74 seconds prior to the NPDC-based method.

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

step (1), collecting electroencephalogram data and preprocessing. The Depth EEG data in this invention were drug-resistant subjects collected from the pediatric Hospital (TCH) who received professional medical team pre-epileptic assessment and detailed surgical sites on the TCH. At the beginning of the experiment, a depth electrode consisting of 5 to 18 contacts was implanted in the patient's brain. Each contact was 2 mm long, 0.8 mm in diameter, and 1.5 mm apart. The Depth EEG signal was recorded at a sampling frequency of 512Hz and the data segment length for each subject was at least four hours longer and contained at least one seizure.

In the present invention, four patients with typical (severe) seizures were analyzed with emphasis, labeled AR, TK, GHG and IM, respectively. This is exemplified by AR patients presenting with Generalized Tonic seizures (GTC). GTC is the most common seizure type associated with epilepsy and also the most common seizure associated with metabolic imbalance, which is the predominant seizure type in only 10% of epileptic patients. This is the approximate electrode position that the surgeon indicates at the time of surgery, as shown in fig. 2. The EZ of the AR is concentrated in the left temporal cavity (fig. 2 right).

The ability to obtain functional MRI (fMRI) is of clinical value in epilepsy, as it can be used to map the spatial location of Inter-Ictal Epileptiform Discharges (IEDs). Epileptic patients use fMRI and SPECT examinations to ensure their safety prior to making a Depth EEG signal. FIG. 3 marks the observed IEDs locations of the AR of the patient. These results were compared to the results of Ictal SPE, CT and PET for confirmation. Local SPECT findings showed perfusion of the posterior lobe and adjacent to the lung lobe resection and the contralateral right apical lobe. PET examination showed a defect in the left temporal parietal wall after resection of the cavity. The position of the EZ is indicated for the present study by diagnosis by a skilled physician.

The test data set for each sample consisted of one hour of pre-episode data, and the present invention used five minute inter-episode and duration data segments as controls. All data fields were determined visually by an experienced epilepsier by the time of the seizure.

Because the original data of the invention is directly collected by the cortical layer, the invention has higher signal-to-noise ratio and high resolution, and the Butterworth filter with high cut-off frequency of 85Hz is needed to carry out band-pass filtering. As can be seen from the Depth EEG signal acquisition map (fig. 4), the signals acquired on the same electrode have high synchronicity, and seizure activity can be clearly observed in the right frontal lobe (RFP) region, but not all the sampling points of RFP can be observed, and the location of the seizure is related to the cortical Depth, which is also the inaccessible region of the EEG signal.

And (2) carrying out FBNs characteristic analysis based on the Depth EEG signal. Introducing the concept of graph theory, establishing an FBNs correlation matrix by applying network analysis, and calculating a corresponding Pearson correlation coefficient and a threshold value. When two channels are changed into a plurality of channels, the Pearson correlation coefficient is changed into a matrix, namely a correlation matrix, from a number, and a proper threshold value is selected to carry out binarization on the correlation matrix. For the matrix after binarization, 0 in the matrix represents a node (i.e. a channel), and 1 represents a degree, so that a brain function network visual map can be drawn.

The specific concept of the graph theory is as follows:

The specific algorithm of Pearson correlation and threshold is as follows:

T＝1-(1-α)^1/(L-1)

The specific method for binarizing the correlation matrix is as follows:

To elucidate the relationship between IEDs and lesions, network analysis was applied to establish the FBNs coherence matrix, and the results are shown in fig. 5. The epilepsy EZ has no great difference between the inter-seizure period and the inter-seizure period, but the global connectivity matrix has a decreasing trend, particularly in the area around the diagonal line, and it can be seen that in the process of the epilepsy seizure, the coherence of the network node and the neighbor nodes is weakened, the cooperative ability is deteriorated, and the ability of information interaction is also weakened.

Although the results of the global and local functional connectivity matrix depict the entire brain's epochs and inter-episodes

The differences between the states are not intuitive. In order to understand the network states of FBNs more clearly, a visual map of brain function network (as shown in fig. 6) is created by using a threshold method. When the epilepsy does not occur, all nodes in the EZ are in dynamic balance and stably work in a coordinated mode, and as the occurrence of the pathological changes, the

network nodes

26 and 27 start to be disconnected from the FBNs, the global FBNs can also see that some areas start to stop exchanging information, so that the FBNs are in a split state, and the transmission of a complete information stream cannot be carried out any more. The global network information interaction center area is also shifted from the inter-seizure period to the seizure period. The pathological changes may occur because the abnormal firing process of the neurons cuts off the interaction process with the neurotransmitters, which further causes the FBNs to fail to connect normally, and the function of some brain areas is degraded, possibly resulting in paralysis.

To study FBNs over time across early epochs, the network dynamics of each FBNs was modeled. These nodes are considered to be the interacting neural masses approximated by the data model described previously. The particular NMM employed in this project embodies an interaction circuit between the major neurons, excitatory interneurons and inhibitory interneurons, with two different time scales of inhibition. The dynamics of each neuron set are expressed by a second-order differential equation, and each neuron population consists of two modules. The first module is to convert the average pulse density of action potentials into an excitatory or inhibitory average Post-synaptic Potential (PSP):

these 12 second order differential equations represent six PSP linear impulse responses, sequential pairs of variables (e.g., y)₁And y₂) Representing an extension of these two-dimensional equations in which the PSP dynamics are given by the first variable (i.e., y)₁，y₃，y₅，y₇，y₉，y₁₁) Specifically, y₁Is a interneuron cell, y₃Is an excitatory interneuron of pyramidal cells, y₅Is a slowly inhibitory interneuron of pyramidal cells, y₇Is a fast inhibitory interneuron of pyramidal cells, y₉Is an inhibition of fast inhibitory interneurons, y₁₁Representing the network propagation delay. R_iInput binarization matrices representing other nodes from the network:

α is a parameter that scales the coupling strength globally, N is the number of nodes in the network S denotes the second module of NMM that can convert the average membrane potential of a group of neurons into the average pulse density (voltage pulse conversion) of the action potentials evoked by these neurons:

wherein e₀Represents half of the maximum discharge rate of NMM, r is v ═ v₀Potential slope of time S-shaped curve, v₀Is the discharge threshold, fig. 7 shows a detailed diagram of the calculation component.

Mathematically, the discharge of this model may be due to different mechanisms, e.g. bistable, minute

Forking, excitatory or intermittent. We have determined the parameters in the formula from the existing literature, which place the model near the saddle nodes on the invariant circle (SNIC) branches, as shown in table 1. In this case, the model node may be discharged due to noise (i.e., it is excitable) and may propagate through the network.

TABLE 1 values of the parameters and their meanings

And (4) calculating BNI. To investigate the pathological extent to which a given network can produce epileptic seizures, the concept of BNI, i.e., the sum of the times at which all nodes are discharged divided by a reference time, was introduced in view of quantitative measurements of FBNs. And (4) calculating BNI on the time series obtained in the step (3), namely using the calculation result in the next step to analyze the epileptic seizure prediction.

The specific algorithm of BNI is as follows:

although most patients spend significantly less than half of their reference time in status epilepticus, the present invention is directed to studying the effect of network changes on the measurability of BNI. Therefore, 0.5 is chosen as the reference duration, which facilitates the calculation to detect such a change. The number of discharges in the model can be defined by the occurrence of large amplitude spikes, as compared to low amplitude dynamics. For ease of handling, if a quiescent period of two seconds occurs after the peak, the return from peak to quiescent is considered. The window period is chosen to account for a visual inspection of the model dynamics of the individual nodes, whether the entire trajectory of the spike is covered. Experiments have shown that a threshold is applied to the mean absolute amplitude of the model output over a sliding window of 0.05s length to extract the peak value for each node. Figure 8 calculates the trend of changes in EZ and BNI of the whole brain before and after a seizure.

As can be seen from fig. 8, even though slight differences in BNI can be observed in WBA, only the BNI values calculated in EZ are increasing during the transition from inter-episode to intra-episode. In both states (inter-and during-episode) there was a significant difference between the BNI values for EZ (P0.0029 <0.05), whereas no significant difference was found in WBA (P0.4972 > 0.05). It can be concluded that measuring BNI in a network established with EZ is more efficient than applying it to WBA networks. Future surgical protocols attempt to study BNI-like analysis, rather than using a large number of electrodes to cover the entire brain, the electrode area can be distributed directly over the lesion. Notably, the frequency of discharges in EZ appeared consistently higher than in WBA, and BNI on EZ was found to be higher during seizures than during interphalangeal periods (P <0.05) with a significant increase in seizure index as the brain approached seizure state.

From the perspective of global FBNs, BNI of each brain functional area during epileptic seizure is below 0.5, and is in a stable discharge state, and due to the influence of disturbance, EZ starts to generate abnormal irregular discharge, BNI index slowly rises until a certain threshold value is exceeded, and finally information transmission states of FBNs are disturbed, so that epileptic seizures are caused. It is difficult to observe a rising trend from the average BNI of global FBNs alone, since BNI values differ only by a fraction, and it is difficult to detect such subtle changes again after averaging of all partitions of the brain. From the above analysis, we propose to detect the BNI value in real time on the EZ, so that the goal of predicting epileptic seizures can be achieved.

And (5) carrying out seizure prediction analysis according to BNI.

Pre-seizure defines a functionally connected network based on Depth EEG recording for pre-seizure and during seizure. On this basis, also the BNI changes within 5 minutes for AR patients and their causes were analyzed. The predicted time of one hour was then used to examine the change in BNI values within one hour prior to the seizure, and the results are shown in FIG. 9. A dramatic increase in BNI was observed in regional brain epilepsy. The BNI remains elevated for epileptic seizures within 5 minutes to one hour after the focal zone around the EZ (typically BNI below 0.5).

In the one hour Depth EEG signal tested, the BNI was shown to be gradually increasing, eventually reaching above the 0.5 threshold at some time, which is the predicted time we want to obtain. The relative activity state (discharge state) of the brain appears to be a process of energy accumulation from the trend of monitored BNI changes, eventually bursting in a state. From fig. 6(a), the brain appears to be in a state of dynamic equilibrium between episodes, the functional areas work in coordination, the discharge state of each brain region is also below a threshold level, and the global brain functional network shows a close association between the entire brain with no unconnected nodes. However, over time, the transition from the initial state breaks this balance: the connections between functional brain regions change, resulting in fluctuations in the discharge frequency, particularly in the EZ. By the above analysis, BNI can be used as an effective predictor for detecting pathogen excretion or diagnosing epilepsy. Due to the inconsistency in the discharge frequency of global FBNs, it is difficult to monitor BNI changes directly from WBA. It is therefore recommended to focus directly on the EZ electrodes during the detection of epilepsy and to avoid implanting as many electrodes in the whole brain as possible. Therefore, the implantation number of the electrodes can be reduced, the pain and recovery time of a patient can be reduced, and a good prediction effect can be observed.

After significant success of patient AR, the same procedure was used to test in three other subjects. As can be seen from table 2, in EZ, the BNI in the pre-seizure state is significantly higher than the BNI in the seizure state. However, this change was not observed in WBA, which is consistent with AR results. This also illustrates the consistency of the experimental results. The average predicted time for all subjects was 2467.74 seconds, outperforming many currently available methods.

TABLE 2 BNI in patients pre-and peri-ict

In this study, a new index BNI was proposed for epilepsy prediction. FBNs were first created from Depth EEG data of four epileptic patients. Neural mass modeling methods were then used to study the association between the structure of FBNs and their propensity to firing. BNI values were calculated to quantify the inter-seizure and pre-seizure epileptogenesis. According to the method, the time of the epileptic seizure prediction is greatly advanced, so that doctors and patients have more sufficient time to take countermeasures, thereby realizing more effective epileptic seizure prediction.

Claims

1. A BNI based seizure signal detection method, comprising the steps of:

step (1), collecting electroencephalogram data and preprocessing;

step (2), brain function network characteristic analysis based on brain depth electrode electroencephalogram signals;

introducing a concept of graph theory, establishing a brain function network correlation matrix by applying network analysis, and calculating a corresponding Pearson correlation coefficient and a threshold value; when two channels are changed into a plurality of channels, the Pearson correlation coefficient is changed into a matrix, namely a correlation matrix, from one number, and a threshold value is selected to carry out binarization on the correlation matrix; for the matrix after binarization, 0 in the matrix represents a node, namely a channel, and 1 represents a degree, and a brain function network visual graph is drawn;

step (3) fitting Depth EEG signals with neural quality models

Substituting the binarization matrix obtained in the step (2) into a neural quality model to obtain a sequence related to time;

step (4) calculating BNI value

In order to study the pathological extent to which a given network can produce epileptic seizures, the concept of BNI was introduced considering quantitative measurements of brain functional networks, i.e. the sum of the time at which all nodes in a reference time are discharged divided by the reference time; calculating BNI for the time sequence obtained in the step (3), namely using the calculation result in the next step to analyze the epileptic seizure prediction;

the specific algorithm of BNI is as follows:

and (5) analyzing according to the value of the BNI.

2. The BNI-based seizure signal detection method of claim 1, wherein said BNI-based seizure signal detection method comprises: the electroencephalogram data acquisition and preprocessing are carried out; the method specifically comprises the following steps:

the method comprises the steps that collected electroencephalogram data are obtained by sampling through a depth electrode distribution system consisting of 5-18 contacts with the length of 2 mm, the diameter of 0.8 mm and the contact distance of 1.5 mm; the data preprocessing mode is to use a Butterworth filter with a high cut-off frequency of 85Hz for band-pass filtering.

3. The BNI-based seizure signal detection method of claim 1, wherein said BNI-based seizure signal detection method comprises: introducing a graph theory, specifically:

graph theory defines a network as a set of nodes, with node pair interactions defined by edges linking them; where nodes represent brain regions and edges represent causal or correlative relationships between nodes.

4. The BNI-based seizure signal detection method of claim 1, wherein said BNI-based seizure signal detection method comprises: the specific algorithm of the Pearson correlation coefficient and the threshold is as follows:

the nodes of the brain functional network represent brain regions near the electrodes, and the edges between the nodes are determined as Pearson correlation coefficients of the observed node activity:

where cov (i, j) is the covariance of channel i and channel j, and σ_iAnd σ_jRepresents the respective standard deviations of the activities in nodes i and j;

T＝1-(1-α)^1/(L-1)

where L is the average of the disjoint cross sections, α is the confidence required, and α is 0.95.

5. The BNI-based seizure signal detection method of claim 1, wherein said BNI-based seizure signal detection method comprises: the specific method for binarizing the correlation matrix is as follows:

for any matrix A, binarization is to select a proper threshold value T to compare with each number in the matrix A, if the threshold value T is larger than the threshold value T, the number of the corresponding position in the matrix A is recorded as 1, and if the threshold value T is smaller than the threshold value T, the number of the corresponding position in the matrix A is recorded as 0.