Individualized brain network connection matrix construction method based on scalp electroencephalogram and brain network key candidate regulation target selection method

Abstract

The application discloses a kind of individualized brain network connection matrix construction method and brain network key candidate regulatory target selection method based on scalp eeg;Each scalp eeg lead is mapped as a set of oscillators containing multiple phase oscillators, the phase coupling relationship between each oscillator is described using Kuramoto model, and the oscillator connection matrix to be solved is used as an optimizable parameter.By minimizing the reconstruction error between the simulated scalp eeg time series and the real scalp eeg time series, the connection matrix can reflect the dynamic coupling structure corresponding to the real eeg signal.At the same time, non-negative connection constraints, zero diagonal penalties and scale compensation mechanisms are introduced to limit the physical meaning of connection weights, suppress oscillator self-connection, and compensate for amplitude scale changes caused by normalization.The application can obtain individualized brain network connection matrix corresponding to scalp eeg lead sequence and several brain network key candidate regulatory targets.

Description

Technical Field

[0001] This invention relates to the field of scalp EEG signal processing and brain network dynamics modeling technology, and in particular to a method for constructing an individualized brain network connectivity matrix based on scalp EEG and a method for selecting key candidate regulatory targets for brain networks. Background Technology

[0002] Brain network modeling describes the interactions between different neural units and their evolution over time using mathematical models. The Kuramoto model uses coupled phase oscillators to describe the synchronization and desynchronization behavior of the network. It has the advantages of clear model structure, interpretable parameters, and relatively low computational complexity, making it suitable for characterizing the phase coupling relationships between different neural groups in brain networks.

[0003] Existing Kuramoto brain network models typically use structural imaging connectivity, population-averaged connectivity, or artificially pre-defined connectivity as the model's brain network connectivity matrix. These connectivity matrices mainly reflect anatomical pathways or population statistical relationships, and are insufficient to fully characterize the individualized dynamic characteristics of specific subjects under actual EEG recording conditions.

[0004] On the other hand, when multiple phase oscillators are used to jointly characterize a scalp EEG lead, it is necessary to establish a mapping relationship between the microscopic phase oscillator activity and the macroscopic scalp EEG lead signal. Simply applying an equal-weighted average to the outputs of multiple oscillators fails to reflect the differences in the roles of different oscillators within the overall brain network. As the number of oscillators increases, the number of connection parameters to be estimated grows rapidly. Traditional gradient-free search methods are computationally inefficient and struggle to stably obtain high-dimensional brain network connectivity matrices.

[0005] The selection of key candidate regulatory targets for brain networks helps identify candidate nodes with a strong influence on whole-brain functional connectivity, distal network response, and global synchronization from multiple brain network nodes corresponding to multiple scalp EEG leads. This provides a model basis for determining individualized neuromodulation targets, modulation directions, and modulation strengths. Existing methods for selecting regulatory targets often rely on anatomical location, local EEG abnormalities, population average data, or static network topology characteristics, which are insufficient to fully reflect the differences in brain network connectivity patterns and phase dynamics among different subjects. Furthermore, existing methods typically lack simulations of network responses after connection inhibition or enhancement of candidate nodes, making it difficult to quantitatively evaluate the impact of local connectivity changes on the functional connectivity of other nodes and global synchronization.

[0006] Therefore, it is necessary to provide a personalized brain network connectivity matrix construction method that uses individual EEG time series as the training target, can establish a differentiable mapping between oscillators and scalp EEG leads, and ensures connectivity matrix constraints during training. This addresses the technical problems of existing Kuramoto brain network modeling methods, which rely on structural connectivity templates or manually preset connections, making it difficult to directly estimate the connectivity matrix from individual scalp EEG time series, and the difficulty in converting high-dimensional oscillator connectivity matrices into scalp EEG lead-level brain networks. Based on the personalized brain network connectivity matrix, key candidate regulatory targets with significant regulatory effects are screened through virtual connectivity perturbation and network response evaluation for each candidate node. Summary of the Invention

[0007] The purpose of this invention is to provide a method for constructing a personalized brain network connectivity matrix based on scalp EEG and a method for selecting key candidate regulatory targets for brain networks. This method can obtain a personalized brain network connectivity matrix corresponding to the sequence of scalp EEG leads. Furthermore, it can screen out several key candidate regulatory targets that have a strong influence on the patient's whole-brain functional connectivity, remote network propagation, and synchronization state based on the personalized brain network connectivity matrix.

[0008] The present invention adopts the following technical solution:

[0009] A method for constructing a personalized brain network connectivity matrix based on scalp EEG includes the following steps:

[0010] S1: Acquire multi-lead scalp EEG time series and perform preprocessing;

[0011] S2: Establish a Kuramoto brain network model containing N phase oscillators, divide the N phase oscillators into R oscillator sets and establish oscillator-lead mapping, and initialize the initial phase, natural angular frequency and external drive of each phase oscillator;

[0012] S3: Set an N×N dimensional free parameter matrix, apply nonnegative mapping and zero diagonal constraint to the free parameter matrix, and obtain the final connection matrix of the oscillator-level brain network in the current iteration;

[0013] S4: Solve the Kuramoto phase dynamics based on the final connection matrix of the oscillator-level brain network to obtain the phase sequence of each phase oscillator in the target time window; calculate the connection strength of each phase oscillator based on the final connection matrix of the oscillator-level brain network, and normalize the connection strength within each oscillator set to obtain the relative contribution weight of each phase oscillator within the oscillator set.

[0014] S5: Perform a sinusoidal transformation on the phase sequence of each phase oscillator, and then perform a weighted synthesis of the sinusoidal transformation results of the oscillator set according to the relative contribution weight of the phase oscillator in its set, to obtain the simulated EEG time sequence corresponding to each scalp EEG lead;

[0015] S6: Obtain the reconstruction error and connection constraint terms between the simulated EEG time series and the real scalp EEG time series, and construct the loss function;

[0016] S7: Calculate the gradient of the loss function with respect to the free parameter matrix and update the free parameter matrix. Repeat S3 to S7 until the preset convergence condition is met to obtain the final connection matrix of the N×N oscillator-level brain network.

[0017] S8: Divide the final connection matrix of the N×N oscillator-level brain network into R×R connection sub-blocks according to the R oscillator sets, and perform mean pooling on each connection sub-block to output the R×R lead-level individualized brain network connection matrix.

[0018] In step S5, a sinusoidal transformation is performed on the phase sequence of each phase oscillator to obtain the instantaneous output of each phase oscillator at each time point; combined with the relative contribution weight of the phase oscillator in its respective oscillator set, the sinusoidal transformation result of the oscillator set is weighted and synthesized to obtain the simulated EEG time sequence corresponding to each scalp EEG lead; all time points are processed according to the above method to finally obtain the simulated EEG time sequence corresponding to the r-th scalp EEG lead.

[0019] In step S6, a scale compensation factor is introduced to optimize the reconstruction error and obtain the scale-compensated reconstruction error. Then, the scale compensation factor is introduced into the optimized loss function through the scale-compensated reconstruction error. The scale compensation factor is the ratio of the amplitude change ratio of the EEG signal before and after normalization to the amplitude scaling coefficient set during the normalization process. The amplitude change ratio is obtained based on the normalized amplitude statistics and the unnormalized amplitude statistics.

[0020] In step S8, the lead-level connection weights are calculated for the connection sub-blocks corresponding to the source EEG signal leads and the target EEG signal leads. Finally, the lead-level connection weights between all the leads are arranged in the order of the scalp EEG leads to obtain the R×R lead-level individualized brain network connection matrix.

[0021] A method for selecting key candidate regulatory targets in brain networks includes the following steps:

[0022] A: The oscillator sets corresponding to each scalp EEG lead are taken as candidate modulation targets. The input connection weights and output connection weights between the oscillator sets corresponding to each candidate modulation target and the other oscillator sets are adjusted in turn to obtain the modulation oscillator-level brain network connection matrix. The phase evolution process of each phase oscillator is solved according to the modulation oscillator-level brain network connection matrix. According to the relative contribution weight of each phase oscillator in its oscillator set, the phase oscillators belonging to the same scalp EEG lead are weighted and averaged to obtain the weighted average phase time series of each scalp EEG lead.

[0023] B: Calculate the phase lock value based on the weighted average phase time series of each lead to obtain the lead-level functional connectivity matrix before and after modulation;

[0024] C: Based on the whole-network functional reconstruction index, far-end propagation index and global phase sequence parameter change index of the lead-level functional connectivity matrix before and after modulation, a comprehensive modulation score is obtained to evaluate the modulation effect of different candidate modulation objects, and key candidate modulation targets of the brain network are selected according to the comprehensive modulation score.

[0025] The overall functional reconstruction index is used to evaluate the degree of change in the lead-level functional connectivity matrix after modulation relative to the lead-level functional connectivity matrix before modulation; the mean absolute difference between the lead-level functional connectivity matrix after modulation and the lead-level functional connectivity matrix before modulation is used.

[0026] The far-end propagation index is used to evaluate the changes in functional connectivity between leads after excluding the row and column containing the current candidate modulator; the far-end propagation index is obtained by excluding the entire row and column containing the candidate modulator r and retaining only the changes in connectivity between other leads.

[0027] The global phase sequence parameter change index is used to evaluate the degree of change in global synchronization state caused by the control operation; the global phase sequence parameter change index is the difference between the Kuramoto global synchronization sequence parameter before and after control.

[0028] The method for obtaining the global phase sequence parameter change index is as follows: first, obtain the Kuramoto global synchronization sequence parameter; then calculate the average synchronization level before regulation; and finally, analyze the entire analysis time window. The average is taken to obtain the overall synchronization level after regulation, and finally the difference between the overall synchronization level after regulation and the average synchronization level before regulation is used.

[0029] The obtained network-wide functional reconstruction indicators, remote propagation indicators, and global phase sequence parameter change indicators are normalized and then weighted and summed according to preset weights to obtain the comprehensive control score of the r-th candidate control object.

[0030] Beneficial effects:

[0031] This invention maps each scalp EEG lead to a set of oscillators containing multiple phase oscillators, uses the Kuramoto model to describe the phase coupling relationship between the oscillators, and uses the oscillator connection matrix to be solved as an optimizable parameter. By minimizing the reconstruction error between the simulated scalp EEG time series and the real scalp EEG time series, the connection matrix can reflect the dynamic coupling structure corresponding to the real EEG signal. Simultaneously, non-negative connectivity constraints, zero diagonal penalties, and scale compensation mechanisms are introduced to restrict the physical meaning of connection weights, suppress oscillator self-connection, and compensate for amplitude scale changes caused by normalization, respectively. Compared with existing brain network construction methods based on correlation, coherence, or phase synchronization indices, this invention does not directly calculate the statistical relationship between scalp EEG leads, but instead establishes a brain network dynamic model composed of multiple phase oscillators, and performs individualized optimization of the oscillator connection matrix by using the reconstruction error between the simulated EEG signal and the real scalp EEG signal, thereby obtaining a brain network connection matrix with dynamic interpretation ability, individual adaptability, and virtual modulation application value.

[0032] This invention uses individual EEG time series as the training target, can establish a differentiable mapping between oscillators and scalp EEG leads, and ensures connection matrix constraints during training. It can effectively solve the technical problems of existing Kuramoto brain network modeling methods that rely on structural connection templates or artificially preset connections, make it difficult to directly estimate the connection matrix from individual scalp EEG time series, and make it difficult to convert high-dimensional oscillator connection matrices into scalp EEG lead-level brain networks.

[0033] Furthermore, the present invention can also screen key candidate regulatory targets of brain networks with significant regulatory effects by evaluating virtual connection perturbations and network responses for each candidate node based on an individualized brain network connectivity matrix. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the individualized brain network connectivity matrix construction method of the present invention;

[0035] Figure 2 A comparison of the oscillator-level connectivity matrix of three anonymous subjects and the lead-level individualized brain network connectivity matrix obtained by mean pooling of connector sub-blocks;

[0036] Figure 3 This is a flowchart illustrating the method for selecting key candidate regulatory targets in the brain network according to the present invention.

[0037] Figure 4 This is a schematic diagram of the screening results for candidate regulatory targets based on the patient's individualized brain network connectivity matrix. Detailed Implementation

[0038] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0039] like Figures 1 to 2 As shown, the method for constructing a personalized brain network connectivity matrix based on scalp EEG according to the present invention includes the following steps:

[0040] S1: Acquire multi-lead scalp EEG time series and perform preprocessing;

[0041] In step S1, the R-lead scalp EEG time series of the subject is first obtained. Then, the obtained scalp EEG time series is filtered, rereferenced, resampled and normalized. The processed scalp EEG time series is then divided into at least one target time window.

[0042] Specifically, the R-lead scalp EEG time series of the subjects was first obtained. ;in, express Time of the first R represents the total number of scalp EEG signals in each lead; then, the continuous EEG signal corresponding to the target state is selected and recorded, and the target time window T is extracted; subsequently, bandpass filtering and resampling are performed on the target time window; finally, the signals of each lead are normalized by dividing them by the maximum absolute value within the target time window.

[0043] In this invention, continuous EEG signal recording corresponding to the target state refers to a segment of raw EEG signal that is labeled as a preset state and is continuous in time within a complete scalp EEG recording of the subject. The target state is a preset EEG or physiological state, such as interictal period, preictal period, ictal period, awake state, sleep state, high arousal state, or low arousal state; continuous EEG signal recording means that the EEG signal segment is continuously acquired on the time axis without arbitrarily splicing signals from different time periods or different states.

[0044] In this embodiment, the continuous EEG signal recording corresponding to the target state can be selected as the continuous scalp EEG recording in the open-eye resting state; a target time window with a length of 2 seconds is extracted; the target time window is bandpass filtered from 0.5 to 100 Hz; when the original sampling rate is 1000 Hz, the target time window is downsampled to 512 Hz; and the signals of each lead are normalized by dividing by the maximum absolute value within the target time window.

[0045] S2: Establish a Kuramoto brain network model containing N phase oscillators, divide the N phase oscillators into R oscillator sets and establish an oscillator-lead mapping. Each oscillator set corresponds to a scalp EEG lead, and initialize the initial phase, natural angular frequency and external drive of each phase oscillator.

[0046] In this embodiment, a Kuramoto brain network model containing N phase oscillators is first established, where each phase oscillator represents a local neural population. The phase dynamics of a phase oscillator satisfy:

[0047]

[0048] in, and They represent the first The and the first Each phase oscillator in Phase of time; Indicates the first The natural angular frequency of a phase oscillator; Indicates the global coupling strength; Indicates the first The phase oscillator and the first phase oscillator The connection weights between each phase oscillator Indicates the total number of phase oscillators; This indicates an external drive; the initial phase is initialized within [0, 2π], the natural frequency is initialized according to the bandpass filter frequency range and converted into the natural angular frequency, and the amplitude of the external drive is proportional to the square root of the time step.

[0049] Then the N phase oscillators are divided into R sets of oscillators. Each set of oscillators corresponds to one scalp EEG lead. Each set of oscillators contains N when N is divisible by R. A phase oscillator.

[0050] S3: Set an N×N dimensional free parameter matrix A, apply nonnegative mapping and zero diagonal constraint to the free parameter matrix A, and obtain the final connection matrix K of the oscillator-level brain network in the current iteration;

[0051] In step S3, by defining an N×N dimensional free parameter matrix A that can be updated through gradient optimization, the oscillator-level brain network connectivity matrix K is generated using non-negative mapping and zero diagonal constraints:

[0052] In this embodiment, the free parameter matrix A is first subjected to nonnegative mapping to obtain the initial connection matrix of the oscillator-level brain network. :

[0053]

[0054] in, Represents a linear rectified function;

[0055] Then, the obtained oscillator-level brain network initial connection matrix... By applying zero diagonal constraints, the final connection matrix K of the oscillator-level brain network is obtained;

[0056] In this embodiment, the zero diagonal constraint is achieved by using the initial connection matrix of the oscillator-level brain network. Set the diagonal elements of the matrix to zero, and set the matrix to zero. This can be achieved through at least one of the following methods: element-wise multiplication with the zero-diagonal mask matrix or by adding a zero-diagonal penalty term to the loss function. Free parameter matrix. Initialize to make the matrix A matrix whose elements are located in the range (0, 0.1).

[0057] S4: Based on the final connectivity matrix of the oscillator-level brain network Solving the Kuramoto phase dynamics yields the phase sequence of each phase oscillator within the target time window; and based on the final connectivity matrix of the oscillator-level brain network... Calculate the connection strength of each phase oscillator, and normalize the connection strength within each oscillator set to obtain the relative contribution weight of each phase oscillator within the oscillator set;

[0058] In this embodiment, the first Connection strength of each phase oscillator The final connection matrix of the oscillator-level brain network The Middle The sum of the connection weights of all phase oscillators in the row:

[0059]

[0060] Where N represents the total number of oscillators, Indicates the first The phase oscillator and the first phase oscillator The connection weights of each phase oscillator Represents the final connection matrix of the oscillator-level brain network The Middle The phase oscillator and the first phase oscillator The sum of the connection weights of all phase oscillators (N in total) in the row;

[0061] No. A set of oscillators The first in A phase oscillator in the oscillator set The relative contribution weights within are:

[0062]

[0063] in, Indicates the first Each phase oscillator belongs to its set of oscillators The relative contribution weight within; Indicates the first The connection strength of each phase oscillator; Represents the set of oscillators The Middle The connection strength of each phase oscillator Represents the set of oscillators The sum of the connection strengths of all phase oscillators (N in total); This represents a default positive number used to avoid a denominator of zero.

[0064] S5: Perform a sinusoidal transformation on the phase sequence of each phase oscillator, and then perform a weighted synthesis of the sinusoidal transformation results of the oscillator set according to the relative contribution weight of the phase oscillator in its respective oscillator set to obtain the simulated EEG time sequence corresponding to each scalp EEG lead.

[0065] First, a sinusoidal transformation is performed on the phase sequence of each phase oscillator to obtain the instantaneous output of each phase oscillator at each time point;

[0066] In this embodiment, the r-th scalp EEG lead (i.e., the r-th lead) A set of oscillators ) Each phase oscillator in The instantaneous output at time t is represented as The value range is [-1, 1]; Indicates the first Each phase oscillator in Phase of time;

[0067] Then, combining the relative contribution weights of the phase oscillators within their respective oscillator sets obtained in step S4, the sinusoidal transformation results of the oscillator set, i.e., the instantaneous outputs of each oscillator at each time point, are weighted and synthesized to obtain the simulated EEG time series corresponding to each scalp EEG lead. .

[0068] In this embodiment, the r-th scalp EEG lead is in Analog signal corresponding to time :

[0069]

[0070] in, For the first A set of oscillators The Middle A phase oscillator in the oscillator set The relative contribution weight within; This represents the relative contribution weight of all phase oscillators in the oscillator set and the weight of that oscillator in... The instantaneous outputs at each moment are weighted and synthesized.

[0071] Finally, by processing all time points using the method described above, the simulated EEG time sequence corresponding to the r-th scalp EEG lead is obtained. .

[0072] S6: Obtain the reconstruction error and connection constraint terms between the simulated EEG time series and the real scalp EEG time series, and construct the loss function;

[0073] In this invention, the connection constraint term is used to limit the final connection matrix of the oscillator-level brain network. The structure's regularization term. The connection constraint term does not directly compare simulated EEG with real EEG; instead, it constrains the connection weights in the constraint matrix to meet predetermined conditions.

[0074] In this embodiment, the connection constraint term is a zero diagonal penalty term. :

[0075]

[0076] in, The first term of the brain network connectivity matrix A is represented by the second term. One diagonal element;

[0077] In this embodiment, the reconstruction error and connection constraint terms They respectively satisfy:

[0078]

[0079]

[0080] The loss function is then:

[0081]

[0082] in, Represents the reconstruction error; T represents the number of sampling points within the target time window, i.e., the target time window T; R represents the number of scalp EEG leads; This indicates that the r-th scalp EEG lead is in The analog signal corresponding to the time. This represents the actual signal of the r-th scalp EEG lead at time t; Indicates a zero diagonal penalty term; Represents the loss function; This indicates the weight of the connection constraint term.

[0083] The aforementioned loss function can simultaneously satisfy the requirements of reconstruction consistency between simulated EEG signals and real scalp EEG signals, as well as the structural constraints of the oscillator-level connection matrix, and achieves a trade-off between reconstruction accuracy and connection constraints, providing an optimization basis for obtaining an individualized brain network connection matrix that meets the preset constraints. However, since the EEG signals have been normalized, the numerical scale of the reconstruction error may be compressed, which weakens its driving effect on the update of connection parameters.

[0084] To amplify the reconstruction error on the normalized data and thus avoid insufficient updating of the connection matrix parameters due to an excessively small loss value, this invention also introduces a scale compensation factor S to optimize the reconstruction error. Reconstruction error after scale compensation Therefore, the reconstruction error is compensated by scale. The scale compensation factor S is introduced into the optimized loss function. middle.

[0085] Specifically:

[0086] First, in step S1, the normalized EEG signal (after filtering, rereference, resampling and normalization) and the corresponding unnormalized EEG signal (after filtering, rereference and resampling) are acquired simultaneously.

[0087] Then, amplitude statistics were performed on the normalized EEG signal and the corresponding unnormalized EEG signal to obtain the normalized amplitude statistics. and unnormalized amplitude statistics ;

[0088] Secondly, based on the normalized amplitude statistic and unnormalized amplitude statistics Calculate the percentage change in amplitude of EEG signals before and after normalization. And based on the amplitude scaling factor set during the normalization process. Determine the scale compensation factor ;

[0089] Subsequently, the scale compensation factor Multiply by the reconstruction mean square error Reconstruction error after scale compensation ;

[0090] Finally, based on the reconstruction error after scale compensation Constructing the scale-compensated loss function ;

[0091]

[0092]

[0093] S7: Calculate the loss function (or loss function after scale compensation) The gradient of the free parameter matrix A is calculated relative to the free parameter matrix A, and the free parameter matrix A is updated. S3 to S7 are repeated until the preset convergence condition is met, resulting in the final connection matrix of the N×N oscillator-level brain network. .

[0094] In step S7, the AdamW optimizer can be used to update the free parameter matrix A with an initial learning rate of 0.01. When the loss function does not improve for 10 consecutive training epochs, the learning rate is multiplied by 0.5, and the lower limit of the learning rate is 0.001. Each time batch contains 50 consecutive sampling points. Training is stopped when the loss function does not improve for 20 consecutive training epochs and the learning rate reaches the lower limit of the learning rate.

[0095] S8: Based on the R sets of oscillators, finalize the N×N oscillator-level brain network connection matrix. Divide the network into R×R connection sub-blocks, and perform mean pooling on each connection sub-block to output an R×R lead-level individualized brain network connection matrix C.

[0096] In this embodiment, after training convergence, the N×N oscillator-level connection matrix is ​​divided into R×R connection sub-blocks according to the R oscillator sets; for the source EEG signal leads considered to have an impact and output coupling information outwards... and the target EEG signal lead that is considered to be influenced by the source lead. For the corresponding n×n connector sub-blocks, calculate the lead-level connection weights:

[0097]

[0098] in, Represents the set of oscillators Each phase oscillator and oscillator set The lead-level connection weights of each phase oscillator in the middle; Represents the set of oscillators The Middle Each phase oscillator points to the set of oscillators. The Middle The oscillator-level connection weights of each phase oscillator; when each oscillator set contains n phase oscillators. .

[0099] Finally, the lead-level connection weights of all obtained leads are arranged in the order of scalp EEG leads to obtain the R×R lead-level individualized brain network connection matrix C.

[0100] In this invention, the source EEG signal leads and target EEG signal lead All were selected from the R normalized scalp EEG leads obtained in step S1. Among them, the R-th lead, which is considered to have outward output coupling in the directed connectivity relationship, was... Each scalp EEG lead is defined as a source lead, and the first lead considered to receive the coupling effect is... Each scalp EEG lead is defined as a target lead. The ordered lead pairs formed by sequentially traversing the R scalp EEG leads are then... Calculate the lead-level connection weights from each source lead to each target lead.

[0101] In this embodiment, R is 59, n is 40, N is 2360, the dimension of the final connection matrix of the N×N oscillator-level brain network is 2360×2360, and the dimension of the R×R lead-level individualized brain network connection matrix is ​​59×59.

[0102] The present invention will be further described below with reference to specific embodiments:

[0103] The first resting scalp EEG recordings with eyes open were selected from subjects A, B, and C, corresponding to the high-arousal group, low-arousal group, and control group, respectively. Each subject used 59 scalp EEG leads, with the lead order consistent across all three subjects.

[0104] For each subject, a 2-second signal segment from the 10th to the 12th second was recorded. A third-order Butterworth filter was used for bandpass filtering from 0.5 to 100 Hz. When the original sampling rate was 1000 Hz, it was downsampled to 512 Hz using zero-phase decimation to obtain a target signal matrix of 59 × 1000. Then, the maximum absolute value of the signal segment was used for normalization.

[0105] Each scalp EEG lead is configured with 40 phase oscillators, therefore R=59, n=40, N=2360, and the dimension of the oscillator-level connection matrix K to be trained is 2360×2360. The time step dt=1 / 500 s, the global coupling strength F is set to 1; the initial connection weights are randomly initialized in (0,0.1) and the initial diagonal elements are set to zero; the initial phases are randomly initialized in [0,2π]; the external drive is set to a random sequence and follows the sequence. Scaling.

[0106] The 1000 sampling points were divided into 20 time batches, each containing 50 consecutive sampling points. The AdamW optimizer was used with an initial learning rate of 0.01. The ReduceLROnPlateau strategy was employed, multiplying the learning rate by 0.5 when the average loss failed to improve for 10 consecutive rounds, with a minimum learning rate of 0.001. The maximum number of training rounds was 5001. Early stopping was performed when the loss failed to improve for 20 consecutive rounds and the learning rate reached its lower limit. The weights of the reconstruction error and the zero diagonal penalty term in the loss function were set to 0.999 and 0.001, respectively.

[0107] After training, the 2360×2360 oscillator-level connection matrix corresponding to the state with minimum loss is saved. This matrix is ​​then reshaped into a 59×40×59×40 four-dimensional block view, and the mean is calculated along the two oscillator dimensions to obtain a 59×59 lead-level individualized brain network connection matrix.

[0108] Figure 2 The top row shows the 2360×2360 oscillator-level connection matrices for subjects A, B, and C, respectively. Figure 2 The bottom row shows the corresponding 59×59 lead-level connectivity matrix. The figure shows different connectivity weight distributions between different lead pairs, illustrating that the block mean pooling can convert the high-dimensional oscillator connectivity matrix into a brain network matrix corresponding one-to-one with the scalp EEG leads while preserving inter-lead differences. This example is for illustrative purposes only and is not intended to limit statistical differences between different groupings or to establish clinical conclusions.

[0109] In order to be able to obtain individualized brain network connectivity matrices Further analysis of key candidate regulatory targets in brain networks that may trigger brain region lesions, such as... Figures 3 to 4 As shown, this invention also discloses a method for selecting key candidate regulatory targets in brain networks, comprising the following steps:

[0110] A: The oscillator sets corresponding to each scalp EEG lead are taken as candidate modulation targets. The input connection weights and output connection weights between the oscillator sets corresponding to each candidate modulation target and the other oscillator sets are adjusted in turn to obtain the modulation oscillator-level brain network connection matrix. The phase evolution process of each phase oscillator is solved according to the modulation oscillator-level brain network connection matrix. According to the relative contribution weight of each phase oscillator in its oscillator set, the phase oscillators belonging to the same scalp EEG lead are weighted and averaged to obtain the weighted average phase time series of each scalp EEG lead.

[0111] In this embodiment, the set of oscillators corresponding to the r-th scalp EEG lead Modulated connection matrix The corresponding connection weights for:

[0112] or

[0113] in, For connection control coefficients; when When, it indicates a reduction in the correlation strength of the candidate oscillator set, i.e., the candidate control object; when This indicates an increase in the connection strength related to the candidate regulatory target. For example, the connection inhibition coefficient can be set to 0.25 and the connection enhancement coefficient to 1.75, respectively. Represents the final connection matrix of the pre-oscillator-level brain network. The row points to the first Inter-phase oscillator connection weights;

[0114] Following existing research on virtual brain network damage and network coupling perturbation that simulates changes in node network interactions by weakening, strengthening, or removing adjacent connections of candidate nodes, this application simulates the reduction or enhancement of the input and output information transmission capabilities of the candidate oscillator set by altering the coupling connection weights between the candidate oscillator set and other oscillators. Furthermore, this application simultaneously adjusts the input and output connections corresponding to the candidate oscillator set to examine the propagation effect of these connection changes on the entire brain network. Specifically, this application employs a method of virtually perturbing the connection matrix by adjusting all input and output edges of each candidate node.

[0115] In this invention, when comparing different candidate oscillator sets, the natural frequency, initial phase, external drive, calculation time window, and other dynamic parameters are kept consistent. Only the connection weights related to the current candidate modulation object are changed to reduce the influence of the initial state and random perturbations on the comparison results. The phase sequence of each phase oscillator within the corresponding time window is solved based on the unmodified connection matrix and the modulated connection matrix, respectively. For the oscillator set corresponding to the r-th scalp EEG lead... Using the relative contribution weights of the oscillators within the set, a weighted circular average is calculated on the unit phase vectors of each phase oscillator within the set, and the argument of the weighted average vector is taken to obtain the first... Weighted average phase time series of scalp EEG leads.

[0116] B: Calculate the phase lock-in value based on the weighted average phase time series of each lead to obtain the lead-level functional connectivity matrix before and after modulation. ;

[0117] Specifically, the The first lead and the first The phase lock value for each lead is;

[0118]

[0119] in, Indicates the first The first lead and the first The phase lock value for each lead, where T represents the total number of sampling points within the calculation time window. Indicates an imaginary singular number. and They represent the first The first lead and the first Each lead at time The average phase.

[0120] The phase lock values ​​between all leads were obtained using the method described above. Then, the obtained phase lock values ​​between all leads were arranged in the order of scalp EEG leads to obtain the final lead-level functional connectivity matrix before and after R×R modulation. .

[0121] C: Lead-level functional connectivity matrix before and after modulation The system uses the whole-network functional reconstruction index, remote propagation index, and global phase sequence parameter change index to obtain a comprehensive control score for evaluating the control effect of different candidate control objects. And based on the comprehensive regulation score We sorted and screened key candidate regulatory targets of brain networks that have a strong influence on the whole brain functional connectivity, remote network propagation and synchronization status of patients.

[0122] In this invention, the whole-network functional reconstruction index is used to evaluate the post-control lead-level functional connectivity matrix. Compared to the lead-level functional connectivity matrix before modulation The overall degree of change; the remote propagation index, used to evaluate the changes in functional connectivity between the remaining leads after excluding the row and column where the current candidate control target is located; the global phase sequence parameter change index, used to evaluate the degree of change in global synchronization state caused by the control operation.

[0123] In this embodiment, the whole network function reconstruction indicator is the post-control lead-level functional connection matrix. Functional connection matrix with pre-regulatory lead level The mean absolute difference between :

[0124]

[0125] in, This represents the overall network functional reconstruction index obtained after regulating the r-th candidate oscillator set using the m-th regulation mode; r represents the candidate regulation object number; m represents the connection inhibition mode or connection enhancement mode; R represents the total number of scalp EEG leads; u and v represent the scalp EEG lead numbers; This represents the phase lock value between the u-th scalp EEG lead and the v-th scalp EEG lead after the m-th modulation mode is used to modulate the r-th candidate oscillator set; This represents the baseline phase lock value between the u-th and v-th scalp EEG leads when no connectivity modulation is performed. This represents the diagonal elements of the exclusion function connection matrix; the absolute value represents the magnitude of the simultaneous statistical changes in phase lock value enhancement and reduction. In this invention, the mean absolute difference... The larger the value, the more significant the network functional reconfiguration caused by regulation.

[0126] In this invention, the remote propagation index excludes the entire row and column containing the candidate control object r, retaining only the others. Changes in connection between leads Obtain;

[0127]

[0128] in, This represents the distal propagation index obtained after regulating the r-th candidate oscillator set using the m-th regulation mode; r represents the candidate regulation object number; m represents the connectivity inhibition mode or connectivity enhancement mode; R represents the total number of scalp EEG leads; u and v represent the scalp EEG lead numbers other than the r-th candidate lead. This represents the phase lock value between the u-th scalp EEG lead and the v-th scalp EEG lead after the m-th modulation mode is used to modulate the r-th candidate oscillator set; This represents the baseline phase lock value between the u-th and v-th scalp EEG leads when no connectivity modulation is performed. and This indicates the entire row and column containing the candidate lead to be excluded; This represents the diagonal elements of the exclusion function connection matrix.

[0129] The remote propagation metric can measure whether the functional connectivity between other nodes has changed significantly after excluding the candidate control object r.

[0130] In this invention, the global phase sequence parameter change index is the Kuramoto global synchronization sequence parameter. The difference before and after adjustment; the specific method for obtaining it is as follows:

[0131] First, obtain the Kuramoto global synchronization order parameter. for

[0132]

[0133] in, This represents the global phase order parameter of the entire oscillator network at time t when controlling the r-th candidate oscillator set and using the m-th control mode. For the total number of oscillators, Indicates an imaginary singular number. This represents the set of oscillators corresponding to the r-th scalp EEG lead. The Middle Each phase oscillator in Phase of time; This represents mapping the phase of the oscillator to a vector on the unit circle;

[0134] Then, calculate the average synchronization level before regulation. :

[0135] ;

[0136] in, The instantaneous synchronization level at time t in the uncontrolled state can be expressed as described above. The calculation formula is used to obtain the result.

[0137] Subsequently, the entire analysis time window was analyzed. The overall synchronization level after regulation is obtained by taking the average. :

[0138] ;

[0139] Finally, the Kuramoto global synchronization sequence parameters are calculated. The difference before and after regulation, i.e., the synchronous change caused by regulation, is: ;

[0140] The acquired network-wide functional reconstruction indicators, remote propagation indicators, and global phase sequence parameter change indicators are normalized and then weighted and summed according to preset weights to obtain the comprehensive control score of the r-th candidate control object. :

[0141]

[0142] in, , and These are the normalized full-network functional reconstruction indicators, remote propagation indicators, and global phase sequence parameter change indicators, respectively. , , The weights of each indicator; + + In this embodiment, the following can be used: , , .

[0143] Each candidate node will be ranked according to its comprehensive control score. Arranged from highest to lowest importance, several key candidate regulatory targets of brain networks with strong influence on the patient's whole-brain functional connectivity, distant network propagation, and synchronization status were screened. These key candidate regulatory targets of brain networks are the regulatory targets that may cause lesions.

[0144] like Figure 4 As shown in the figure, (a) represents the patient's individualized brain network connectivity matrix; (b) represents the ranking of the comprehensive modulation scores of candidate lead nodes; (c) and (d) represent the PLV changes relative to the baseline state after connection inhibition and connection enhancement of the top-ranked candidate nodes, respectively; and (e) represents the response of the candidate node in three aspects: whole-network functional reconstruction, remote propagation, and global synchronization sequence parameters.

[0145] This experiment used an individualized brain network connectivity matrix obtained through training from a single patient as the input matrix. First, the oscillator-level connectivity matrix was divided into blocks and averaged according to the oscillator sets corresponding to the scalp EEG leads to obtain the lead-level brain network connectivity matrix, as shown in Figure (a). Then, 59 scalp EEG lead nodes were sequentially selected as candidate modulation nodes. Connectivity inhibition and connectivity enhancement were applied to all input and output connections of the current candidate node, respectively, to obtain brain network connectivity matrices under different modulation conditions. Connectivity inhibition was achieved by multiplying the relevant connection weights of the candidate nodes by a preset modulation coefficient less than 1, while connectivity enhancement was achieved by multiplying the connection weights by a preset modulation coefficient greater than 1. Except for the connections related to the current candidate node, the remaining connection weights, initial phase, intrinsic frequency, and computation time parameters remained unchanged to allow for pairing and comparison of different candidate nodes.

[0146] Under each modulation condition, the phase change of each lead node is calculated based on the modulation connectivity matrix, and the phase lock value between lead nodes is further calculated to obtain the modulation functional connectivity matrix. The modulation functional connectivity matrix is ​​compared with the baseline functional connectivity matrix to calculate the degree of functional connectivity reconstruction across the entire network, the degree of far-end propagation after excluding candidate nodes, and the degree of change in global synchronization state. The above indicators are normalized and weighted to obtain the comprehensive modulation score of each candidate lead node, and then sorted from high to low according to the comprehensive score, as shown in Figure (b).

[0147] Figure (c) shows the change in phase lock value between leads relative to baseline after implementing connectivity inhibition on top-ranked candidate nodes; Figure (d) shows the change in phase lock value after implementing connectivity enhancement, where different colors indicate enhanced or weakened functional connectivity; Figure (e) further compares the changes in whole-network functional reconstruction, distal propagation, and global phase order parameters of the candidate node under both connectivity inhibition and connectivity enhancement modulation directions. By comparing the changes in indicators under different candidate nodes and different modulation directions, candidate modulation targets with significant modulating effects on the functional connectivity and global synchronization state of the patient's brain network can be screened, providing a basis for subsequently determining the direction and intensity of modulation.

[0148] The above description is merely a preferred embodiment of the present invention. For those skilled in the art, any equivalent substitutions made to the preprocessing method, integration algorithm, non-negative mapping function, connection strength definition, loss function, or gradient optimizer without departing from the principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a personalized brain network connectivity matrix based on scalp electroencephalography (EEG), characterized in that: Includes the following steps: S1: Acquire multi-lead scalp EEG time series and perform preprocessing; S2: Establish a Kuramoto brain network model containing N phase oscillators, divide the N phase oscillators into R oscillator sets and establish oscillator-lead mapping, and initialize the initial phase, natural angular frequency and external drive of each phase oscillator; S3: Set an N×N dimensional free parameter matrix, apply nonnegative mapping and zero diagonal constraint to the free parameter matrix, and obtain the final connection matrix of the oscillator-level brain network in the current iteration; S4: Solve the Kuramoto phase dynamics based on the final connection matrix of the oscillator-level brain network to obtain the phase sequence of each phase oscillator in the target time window; calculate the connection strength of each phase oscillator based on the final connection matrix of the oscillator-level brain network, and normalize the connection strength within each oscillator set to obtain the relative contribution weight of each phase oscillator within the oscillator set. S5: Perform a sinusoidal transformation on the phase sequence of each phase oscillator, and then perform a weighted synthesis of the sinusoidal transformation results of the oscillator set according to the relative contribution weight of the phase oscillator in its set, to obtain the simulated EEG time sequence corresponding to each scalp EEG lead; S6: Obtain the reconstruction error and connection constraint terms between the simulated EEG time series and the real scalp EEG time series, and construct the loss function; S7: Calculate the gradient of the loss function with respect to the free parameter matrix and update the free parameter matrix. Repeat S3 to S7 until the preset convergence condition is met to obtain the final connection matrix of the N×N oscillator-level brain network. S8: Divide the final connection matrix of the N×N oscillator-level brain network into R×R connection sub-blocks according to the R oscillator sets, and perform mean pooling on each connection sub-block to output the R×R lead-level individualized brain network connection matrix.

2. The method for constructing a personalized brain network connectivity matrix according to claim 1, characterized in that: In step S5, a sinusoidal transformation is performed on the phase sequence of each phase oscillator to obtain the instantaneous output of each phase oscillator at each time point; combined with the relative contribution weight of the phase oscillator in its respective oscillator set, the sinusoidal transformation result of the oscillator set is weighted and synthesized to obtain the simulated EEG time sequence corresponding to each scalp EEG lead; all time points are processed according to the above method to finally obtain the simulated EEG time sequence corresponding to the r-th scalp EEG lead.

3. The method for constructing a personalized brain network connectivity matrix according to claim 1, characterized in that: In step S6, a scale compensation factor is introduced to optimize the reconstruction error and obtain the scale-compensated reconstruction error. Then, the scale compensation factor is introduced into the optimized loss function through the scale-compensated reconstruction error. The scale compensation factor is the ratio of the amplitude change ratio of the EEG signal before and after normalization to the amplitude scaling coefficient set during the normalization process. The amplitude change ratio is obtained based on the normalized amplitude statistics and the unnormalized amplitude statistics.

4. The method for constructing a personalized brain network connectivity matrix according to claim 1, characterized in that: In step S8, the lead-level connection weights are calculated for the connection sub-blocks corresponding to the source EEG signal leads and the target EEG signal leads. Finally, the lead-level connection weights between all the leads are arranged in the order of the scalp EEG leads to obtain the R×R lead-level individualized brain network connection matrix.

5. A method for selecting key candidate regulatory targets in brain networks based on any one of the individualized brain network connectivity matrix construction methods according to claims 1 to 4, characterized in that, Includes the following steps: A: The oscillator sets corresponding to each scalp EEG lead are taken as candidate modulation targets. The input connection weights and output connection weights between the oscillator sets corresponding to each candidate modulation target and the other oscillator sets are adjusted in turn to obtain the modulation oscillator-level brain network connection matrix. The phase evolution process of each phase oscillator is solved according to the modulation oscillator-level brain network connection matrix. According to the relative contribution weight of each phase oscillator in its oscillator set, the phase oscillators belonging to the same scalp EEG lead are weighted and averaged to obtain the weighted average phase time series of each scalp EEG lead. B: Calculate the phase lock value based on the weighted average phase time series of each lead to obtain the lead-level functional connectivity matrix before and after modulation; C: Based on the whole-network functional reconstruction index, far-end propagation index and global phase sequence parameter change index of the lead-level functional connectivity matrix before and after modulation, a comprehensive modulation score is obtained to evaluate the modulation effect of different candidate modulation objects, and key candidate modulation targets of the brain network are selected according to the comprehensive modulation score.

6. The method for selecting key candidate regulatory targets in brain networks according to claim 5, characterized in that: The overall functional reconstruction index is used to evaluate the degree of change in the lead-level functional connectivity matrix after modulation relative to the lead-level functional connectivity matrix before modulation; the mean absolute difference between the lead-level functional connectivity matrix after modulation and the lead-level functional connectivity matrix before modulation is used.

7. The method for selecting key candidate regulatory targets in brain networks according to claim 5, characterized in that: The far-end propagation index is used to evaluate the changes in functional connectivity between leads after excluding the row and column containing the current candidate control target. The far-end propagation index is obtained by excluding the entire row and column containing the current candidate control target and retaining only the changes in connectivity between other leads.

8. The method for selecting key candidate regulatory targets in brain networks according to claim 5, characterized in that: The global phase sequence parameter change index is used to evaluate the degree of change in global synchronization state caused by the control operation; the global phase sequence parameter change index is the difference between the Kuramoto global synchronization sequence parameter before and after control.

9. The method for selecting key candidate regulatory targets in brain networks according to claim 8, characterized in that: The method for obtaining the global phase sequence parameter change index is as follows: first, obtain the Kuramoto global synchronization sequence parameter; then calculate the average synchronization level before regulation; then take the average over the entire analysis time window to obtain the overall synchronization level after regulation; finally, calculate the difference between the overall synchronization level after regulation and the average synchronization level before regulation.

10. The method for selecting key candidate regulatory targets in brain networks according to claim 5, characterized in that: The acquired network-wide functional reconstruction indicators, remote propagation indicators, and global phase sequence parameter change indicators are normalized and then weighted and summed according to preset weights to obtain the comprehensive control score of the current candidate control object.

Images