An adaptive control system and method for abnormal discharge based on neural network phase transition tracking

Abstract

The application discloses an abnormal discharge self-adaptive regulation and control system and method based on neural network phase transition tracking, and belongs to the technical field of neural network regulation and control. The system comprises a host computer, an electroencephalogram acquisition-stimulation integrated device and an FPGA real-time calculation device. The host computer is used for constructing a neural network metastable phase transition direction predictor and an adaptive electroencephalogram signal forward predictor based on a large sample structure magnetic resonance image and a subject's electroencephalogram signal, and transmitting the same to the FPGA real-time calculation device for storage. The electroencephalogram acquisition-stimulation integrated device is used for acquiring an electroencephalogram signal, and generating an electric stimulation signal based on a stimulation instruction issued by the FPGA real-time calculation device. The FPGA real-time calculation device is used for receiving the electroencephalogram signal, calculating metastable state dynamics characteristics based on the electroencephalogram signal, identifying a metastable state phase transition direction based on the metastable state dynamics characteristics, and generating a stimulation instruction when the metastable state phase transition direction is abnormal.

Description

Technical Field

[0001] This invention belongs to the field of neural network control technology, specifically relating to an adaptive control system and method for abnormal discharge based on neural network phase transition tracking. Background Technology

[0002] Abnormal electrical activity in the brain can cause temporary brain dysfunction, affecting normal brain development and function. Severe, recurrent episodes can pose a significant threat to human health and safety. With the development of brain-computer interface (BCI) technology, it has become possible to effectively regulate abnormal brain activity in individuals. Noninvasive brain stimulation (NBS), represented by non-invasive electrical stimulation, has been proven to be an effective treatment for neurological disorders such as Parkinson's disease, depression, and generalized anxiety disorder. Recent research suggests that electrical stimulation should be applied with different parameters at different stages of the disease, leading to the concept of adaptive electrical stimulation. Recently, scientific research on complex abnormal brain activity has shifted from static local areas to a dynamic, whole-brain neural network perspective. Neurodynamic network theory, based on dynamic systems theory, posits that macroscopic brain activity is composed of multiple metastable neural network functional sub-patterns that transform into each other at different times. Before the onset of recurrent neurological disorders such as depression and treatment-resistant anxiety, a sudden shift from a relatively stable pattern to an abnormal activity pattern in the brain's neural network functional sub-patterns has been observed. In physics, such sudden changes of state are called phase transitions.

[0003] Although electrical stimulation neuromodulation technology has achieved some success in intervening in abnormal brain discharges such as epilepsy, the problem of high variability in effectiveness remains. The main reasons are: most existing technologies only focus on the occurrence of abnormal discharge phenomena, failing to identify the abnormal phase transitions of metastable functional sub-modes before the occurrence of abnormal discharges, thus preventing early intervention; simultaneously, they focus too much on the local signals of a single electrode, failing to generate stimulation programs based on the global brain neural network composed of multi-electrode signals; and the stimulation current parameters generated by existing devices are fixed, failing to effectively couple the abnormal brain discharge waveform. In view of this, this invention proposes an adaptive modulation system and method for abnormal discharges based on neural network phase transition tracking to solve the above problems. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing technologies by proposing an adaptive control system and method for abnormal discharges based on neural network phase transition tracking, thereby achieving effective control of abnormal brain discharges.

[0005] To achieve the above objectives, the present invention provides the following solution: an adaptive control system for abnormal discharge based on neural network phase transition tracking, comprising: a host computer, an integrated EEG acquisition-stimulation device, and an FPGA real-time computing device;

[0006] The host computer is used to construct a neural network metastable phase transition direction predictor and an adaptive EEG signal forward predictor based on large sample structural magnetic resonance images and the subject's EEG signals; and transmits the neural network metastable phase transition direction predictor and the adaptive EEG signal forward predictor to the FPGA real-time computing device for storage;

[0007] The integrated EEG acquisition-stimulation device is used to acquire the EEG signals and generate electrical stimulation signals based on the stimulation instructions issued by the FPGA real-time computing device.

[0008] The FPGA real-time computing device is used to receive the EEG signal, calculate metastable dynamic characteristics based on the EEG signal, identify the metastable phase transition direction based on the metastable dynamic characteristics, and generate the stimulation command when the metastable phase transition direction is abnormal.

[0009] More preferably, the host computer includes: a database module, a neural network phase transition predictor generation module, and an adaptive EEG signal predictor generation module;

[0010] The database module is used to store large-sample structural magnetic resonance images and the electroencephalogram (EEG) signals.

[0011] The neural network phase transition predictor generation module is used to construct the neural network metastable phase transition direction predictor;

[0012] The adaptive EEG signal predictor generation module is used to construct the adaptive EEG signal forward predictor.

[0013] More preferably, the method for constructing the neural network metastable phase transition direction predictor includes:

[0014] Using the cortical brain regions corresponding to the scalp electrodes as network nodes and the number of fiber bundles between network nodes as the weights of the connection edges, a structural brain neural network is constructed.

[0015] In the aforementioned structural brain neural network, based on excitability values, two node types are defined: abnormal discharge nodes and non-abnormal discharge nodes; each node is assigned a ternary variable at each time point; the ternary variables represent three network activity modes, including: the refractory period activity mode after discharge, the subcritical period activity mode before discharge, and the supercritical period activity mode during discharge.

[0016] Construct node activity transition probabilities weighted by connected edges, and calculate the transition probabilities between three network activity modes based on the node activity transition probabilities.

[0017] More preferably, the control conditions for the node activity transition probability include:

[0018]

[0019] In the formula, P(x i =1 in(t,t+Δ]|x i (t)=-1)) represents the probability that node i changes from subcritical activity mode to supercritical activity mode within a time step (t, t+Δ); I(·) is the step function; z i (t) represents the energy input of node i; E i Represents the excitability value of node i; o(·) represents a small higher-order error; P(x i =0 in(t,t+Δ]|x i (t)=1)) represents the probability that node i changes from supercritical activity mode to refractory activity mode within a time step (t, t+Δ); G(·) is the rate function; t so,i t sf,i τ represents the start and end times of the abnormal discharge at node i, respectively; s,i q si τ represents the variability of the time scale and duration of the anomalous discharge at node i, respectively; r,i q ri Let P(x) represent the variability of the time scale and duration of the refractory period at node i, respectively; i =-1 in(t,t+Δ]x i (t)=0)) represents the probability that node i changes from the refractory period activity mode to the subcritical activity mode within a time step (t, t+Δ).

[0020] More preferably, the method for constructing the adaptive EEG signal forward predictor includes:

[0021] A low-dimensional latent state model is constructed from the latent state matrix, weight matrix, and noise matrix.

[0022] The low-dimensional latent state model is combined with a recursive Kalman filter to construct the adaptive EEG signal forward predictor.

[0023] More preferably, the adaptive EEG signal forward predictor includes:

[0024]

[0025]

[0026] In the formula, This represents the low-dimensional state sequence value from the previous step. This represents the predicted value of the low-dimensional state in the next step. K represents the next predicted value of the signal phase sequence; t Denotes the Kalman gain matrix; A t C t C represents the current model weight matrix. t-1 This represents the model weight matrix from the previous step; y t This represents the phase sequence of electroencephalogram (EEG) signals.

[0027] More preferably, the FPGA real-time computing device includes: a neural network real-time generation module, a neural network phase transition tracking module, and an adaptive stimulus instruction generation module;

[0028] The real-time neural network generation module is used to preprocess the EEG signals and generate a phase synchronization network connection matrix based on the preprocessed EEG signals.

[0029] The neural network phase transition tracking module is used to calculate metastable dynamic characteristics based on the network connection matrix, and input the metastable dynamic characteristics into the neural network metastable phase transition direction predictor to identify abnormal phase transition directions and abnormal nodes.

[0030] When an abnormal phase transition direction is identified, the label of the abnormal node is input to the adaptive instruction generation module. The adaptive instruction generation module obtains the 10-second EEG signal phase sequence of the abnormal node and obtains the predicted EEG signal phase value of the abnormal node for the next 10 seconds based on the adaptive EEG signal forward predictor; and generates the stimulation instruction based on the predicted value.

[0031] This invention also provides an adaptive control method for abnormal discharge based on neural network phase transition tracking. The method, applied to the aforementioned system, includes the following steps:

[0032] S1. Based on large-sample structural magnetic resonance images and subjects' electroencephalogram (EEG) signals, construct a neural network metastable phase transition direction predictor and an adaptive EEG signal forward predictor.

[0033] S2. Collect long-term EEG signals from the subject and calculate metastable dynamic characteristics based on the EEG signals. Identify the metastable phase transition direction based on the metastable dynamic characteristics, the neural network metastable phase transition direction predictor, and the adaptive EEG signal forward predictor. When the metastable phase transition direction is abnormal, generate a stimulus command.

[0034] S3. Based on the stimulation command, generate an electrical stimulation signal and perform electrical stimulation. After the stimulation is completed, continue to collect EEG signals and cycle through S2-S3 to achieve adaptive electrical stimulation regulation of abnormal EEG discharges.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This invention utilizes real-time EEG signals to track the direction of abnormal metastable phase transitions in brain functional activity patterns. It adaptively targets abnormal network nodes before abnormal discharges occur, administering non-invasive electrical interference stimulation for neuromodulation, effectively suppressing the timing and intensity of abnormal brain discharges. By combining FPGA real-time computing devices and a host computer, personalized model training can be performed before modulation. During modulation, based on the acquired and analyzed EEG signals in real-time (with a time error of less than 10 milliseconds), individualized and precise electrical stimulation modulation can be achieved at specific abnormal phase transition moments and in specific abnormal brain regions. This provides a new modulation scheme for people suffering from abnormal brain discharges and has promising application prospects. Attached Figure Description

[0037] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the overall structure of the control system according to an embodiment of the present invention;

[0039] Figure 2 This is a flowchart illustrating the principle of the control system according to an embodiment of the present invention;

[0040] Figure 3 This is a flowchart illustrating the workflow of the host computer neural network phase transition predictor generation module in an embodiment of the present invention.

[0041] Figure 4 This is a flowchart of the upper computer adaptive EEG signal predictor generation module in an embodiment of the present invention;

[0042] Figure 5 This is a flowchart illustrating the workflow of the real-time neural network generation module in an FPGA real-time computing device according to an embodiment of the present invention.

[0043] Figure 6 This is a flowchart illustrating the workflow of the neural network phase transition tracking module in an FPGA real-time computing device according to an embodiment of the present invention.

[0044] Figure 7 This is a flowchart illustrating the workflow of the adaptive stimulus instruction generation module for the FPGA real-time computing device in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Example 1:

[0048] like Figure 1 , Figure 2 As shown, this embodiment provides an adaptive control system for abnormal discharges based on neural network phase transition tracking, including: a host computer, an integrated EEG acquisition-stimulation device, and an FPGA real-time computing device. Before control, in the host computer, a neural network metastable phase transition direction predictor and an adaptive EEG signal forward predictor are constructed based on a large sample of structural magnetic resonance imaging (MRI) images and the subject's EEG signals. The host computer is connected to the FPGA real-time computing device via a JTAG cable to transmit the neural network metastable phase transition direction predictor and the adaptive EEG signal forward predictor to the FPGA real-time computing device for storage. During control, the integrated EEG acquisition-stimulation device and the FPGA real-time computing device are connected via a cable or wireless Bluetooth to store the acquired long-term EEG signals to a memory card. After control, the long-term EEG signals are downloaded to the database module of the host computer. The FPGA real-time computing device receives the EEG signals and calculates metastable dynamic characteristics based on the EEG signals. It identifies the metastable phase transition direction based on the metastable dynamic characteristics, and generates stimulation commands when the metastable phase transition direction is abnormal.

[0049] The integrated EEG acquisition and stimulation device features a gated switch, enabling dual operation modes for acquisition and stimulation. In acquisition mode, it collects the subject's EEG signals; in stimulation mode, signal acquisition ceases, and stimulation commands from the FPGA real-time computing device are received to generate electrical stimulation signals. It has 64 signal channels, all capable of both signal acquisition and stimulation signal output. During stimulation, each pair of channels forms a group, consisting of one positive and one negative electrode.

[0050] Specifically, the host computer includes: a database module, a neural network phase transition predictor generation module, and an adaptive EEG signal predictor generation module. The database module stores large samples of structural magnetic resonance imaging (MRI) images and EEG signals. The neural network phase transition predictor generation module constructs a neural network metastable phase transition direction predictor. The adaptive EEG signal predictor generation module constructs an adaptive EEG signal forward predictor.

[0051] like Figure 3 As shown, in the neural network phase transition predictor generation module, a population neural network metastable dynamic model is first constructed based on a large sample of structural magnetic resonance images. Then, the model is fitted based on the long-term EEG signals of the subjects to train an individualized neural network metastable dynamic model and obtain a neural network metastable phase transition direction predictor.

[0052] The specific process includes: using the cortical brain regions corresponding to the scalp electrodes as network nodes, and using the number of fiber bundles between network nodes as the weights of the connection edges, to construct a structural brain neural network; that is, the number of white matter fiber bundles between node i and node j is used as the weight w of the connection edge. ij .

[0053] In the aforementioned brain neural network structure, based on different excitability values ​​E j Two node types are defined: DSCH (Discharge of Abnormal Nodes), which are nodes that exhibit spontaneous discharge behavior; and non-DSCH (Discharge of Non-Abnormal Nodes), which are nodes that cannot spontaneously discharge abnormally but have a probability of entering an abnormal discharge state after interacting with DSCH nodes. The node types are defined as follows: Equations (1) and (2):

[0054] DSCH={j∈{1, 2,...,N}|E j >0}, (1)

[0055] non-DSCH={j∈{1, 2,...,N}|E j <0}。(2)

[0056] Each node is assigned a ternary variable x(t) = {0, -1, 1} at each time point; the ternary variable represents three network activity modes, including: the refractory period activity mode after discharge, the subcritical period activity mode before discharge, and the supercritical period activity mode during discharge.

[0057] Next, a node activity transition probability weighted by connected edges is constructed. Then, the transition probability between the three network activity modes is calculated based on this node activity transition probability. The node activity transition probability is controlled by the following conditions:

[0058]

[0059] In the formula, P(xi = 1 in(t, t + Δ]|x i (t) = -1)) represents the probability that node i changes from the subcritical activity mode to the supercritical activity mode within a time step (t, t + Δ); I(·) is the step function, which is defined as: step function I(y), which is 0 when y < 0, y when 0 < y < 1, and 1 when y > 1; z i (t) represents the energy input of node i; E i represents the excitability value of node i; o(·) represents a small high-order error; P(x i = 0 in(t, t + Δ]|x i (t) = 1)) represents the probability that node i changes from the supercritical activity mode to the refractory activity mode within a time step (t, t + Δ); G(·) is the rate function, which is defined as: velocity function G(T, τ, q), which is 0 when T < τ - q, and 1 / τ + q - T when τ - q < T < τ + q; t so,i 、t sf,i respectively represent the start time and end time of abnormal discharge of node i; τ s,i 、q si respectively represent the variability of the time scale and duration of abnormal discharge of node i; τ r,i ,q ri respectively represent the variability of the time scale and duration of the refractory period of node i; P(x i = -1 in(t,t + Δ]x i (t) = 0)) represents the probability that node i changes from the refractory activity mode to the subcritical activity mode within a time step (t, t + Δ).

[0060] Among them, z i (t) can be expressed as:

[0061]

[0062] In the formula, represents the excitatory input from node j to node i; represents the inhibitory effect of the remaining inactive nodes on the activity of node i.

[0063] The parameter conditions for the occurrence of local abnormal discharge are: [[ID=�7]]

[0064]

[0065] In the formula, E DSCH represents the excitability value of the abnormally discharging node; z DSCH represents the surround inhibition of the abnormal discharging node by the peripheral non-abnormally discharging nodes; when there is no abnormal discharge, z DSCH can be expressed as The structural connection weighting value wW between the surrounding non-abnormal discharge node j and the abnormal discharge node represents the structural connection weighting value wW. DSCH，j Constraints on anomalous discharge nodes; E represents the excitability of non-nominal discharge nodes, w is the global connectivity strength, and b is a parameter greater than 0.

[0066] The parameter conditions for the propagation of abnormal discharge are as follows:

[0067]

[0068] In the formula, E represents the excitability of the non-abnormal discharge node; This indicates that the surrounding anomalous discharge nodes are located on the time scale τ. s Continuous discharge D DCSH The excitatory energy input to non-abnormal discharge nodes; This represents the inhibitory constraint of surrounding non-abnormal discharge nodes on the propagation of abnormal discharge; w is the global connectivity strength, and a and b are parameters greater than 0.

[0069] The parameter range that satisfies formulas (5) and (6) indicates that the network is in the supercritical activity mode. The parameter range that satisfies formula (5) but not formula (6) indicates that the network is in the subcritical activity mode. The parameter range that does not satisfy formulas (5) and (6) indicates that the network is in the refractory period activity mode.

[0070] Subsequently, the long-term EEG signals of the subjects were artificially divided into N non-overlapping segments containing the post-discharge, pre-discharge, and discharge periods. The phase synchronization network of these signals was calculated with a window length of 100ms. The population model was fitted using N-fold cross-validation to train an individualized neural network metastable dynamic model and obtain a neural network metastable phase transition direction predictor.

[0071] like Figure 4 As shown, the specific process for constructing an adaptive EEG signal feedforward predictor includes:

[0072] Given a low-dimensional latent state matrix x of dimension n t A low-dimensional latent state model is constructed using weight matrices A and C, and noise matrices Q, R, and S. This low-dimensional latent state model describes the phase sequence of the EEG signal. The low-dimensional latent state model is as follows:

[0073] x t+1 =A t x t +w t (7)

[0074] y t =C t x t +v t (8)

[0075] In the formula, w t and v t It is modeled as Gaussian white noise, characterized by a covariance matrix containing the noise matrices Q, R, and S:

[0076]

[0077] By combining a low-dimensional latent state model with a recursive Kalman filter, an adaptive EEG signal feedforward predictor is constructed to realize the low-dimensional latent matrix x for the next future step size. t+1 With phase sequence y t+1 Prediction:

[0078]

[0079]

[0080] In the formula, This represents the low-dimensional state value from the previous step. This represents the predicted value of the low-dimensional state in the next step. K represents the next predicted value of the signal phase sequence; t Denotes the Kalman gain matrix; A t C t C represents the current model weight matrix. t-1 This represents the model weight matrix from the previous step; y t This indicates the current step size EEG signal phase value.

[0081] The Kalman gain matrix includes:

[0082]

[0083]

[0084] In the formula, X is a solution to the Riccati equation; R t This represents a current noise covariance matrix.

[0085] To quantify the performance of the adaptive EEG signal forward predictor, the forward prediction error e at time t is:

[0086]

[0087] Therefore, the explained variance of the forward prediction of the i-th channel is used as an evaluation index for prediction performance:

[0088]

[0089] In the formula, T represents the total number of time steps.

[0090] After preprocessing the multi-channel EEG signals, the phase sequence y is obtained by Hilbert transform. t The phase sequence of the EEG signal was artificially divided into T non-overlapping segments of 10 seconds each, and the model parameters were trained using N-fold cross-validation.

[0091] The FPGA real-time computing device utilizes the high-performance Virtex UltraScale+VU19P chip manufactured by Xilinx. This chip includes 9 million system logic units and 224MB of RAM, and possesses comprehensive clock management and frequency synthesis capabilities, meeting the requirements of high-performance systems. Specifically, the FPGA real-time computing device comprises: a neural network real-time generation module, a neural network phase transition tracking module, and an adaptive stimulus command generation module. The neural network real-time generation module embeds a phase synchronization network computation program; the neural network phase transition tracking module contains a microprocessor configured with a pre-trained metastable phase transition direction predictor for the neural network; the adaptive stimulus command generation module contains a microprocessor configured with a pre-trained adaptive EEG signal feedforward predictor for the host computer. The FPGA real-time computing device is controlled by a state machine, operating between a state tracking phase, a stimulus command generation phase, and a standby phase. During the state tracking phase, only the neural network real-time generation module and the neural network phase transition tracking module operate. When an abnormal phase transition direction is detected, the device enters the stimulus command generation phase, and the adaptive stimulus command generation module is enabled. After stimulation begins, the device enters a standby state until the stimulation ends, at which point the next round of signal acquisition begins.

[0092] like Figure 5 As shown, the real-time neural network generation module is used for preprocessing the EEG signal. After filtering out the 50Hz power frequency signal using a notch filter, the EEG signal is subjected to real-time artifact removal based on the artifact space reconstruction algorithm. A sliding window with a fixed step size of 100ms is used, followed by a Fast Fourier Transform (FFT) on the EEG signal. A mask function is used for convolution to double the frequency components greater than 0. Then, an inverse FFT is performed to obtain the analytical signal. The coordinate rotating digital computer (CORDIC) method is used to calculate the arctan function of the analytical signal to obtain the instantaneous phase value of the signal. The phase synchronization value between each pair of channels is calculated, and a phase synchronization network connection matrix is ​​generated. The instantaneous phase values ​​of 100 steps and the phase synchronization matrix are stored in RAM.

[0093] like Figure 6As shown, the neural network phase transition tracking module is used to calculate metastable dynamic characteristics based on the network connection matrix. These metastable dynamic characteristics are then input into the neural network metastable phase transition direction predictor to identify the metastable phase transition direction and anomalous nodes. In this embodiment, the metastable dynamic characteristics include global dynamic characteristics and local dynamic characteristics; the global dynamic characteristics include network critical deceleration and network volatility.

[0094] Among them, the network's critical deceleration is characterized by the autocorrelation AC and variance VAR of the first eigenvector:

[0095]

[0096]

[0097] In the formula, p(t) represents the first eigenvector of the network at step t; N represents the number of steps.

[0098] Network volatility is characterized by the global clustering coefficient CC:

[0099]

[0100] in,

[0101]

[0102]

[0103] In the formula, k i w represents the node degree. ij Let M represent the network connection edge value between node i and node j, and M represent the number of nodes that have an edge connection with node i.

[0104] Local dynamic features include: node source index, sink index, and sink connectivity (conn), which are respectively:

[0105] source = 1.41 - |(rr) i ,cr i )-(1 / N,1)|, (21)

[0106] sink i =1.41-|(rr i ,cr i )-(1,1 / N)|, (22)

[0107]

[0108] In the formula, rr i cr represents the rank of the row in the network connectivity matrix where the node is located. iThe rank of the column in the network connection matrix where the node is located is represented by abs(), where abs() represents the absolute value; N is the number of network nodes.

[0109] The network's critical deceleration and volatility characteristics, along with node source, sink, and sink connectivity features, are input into a neural network metastable phase transition direction predictor with 100 steps to identify anomalous phase transition directions. Anomalous nodes are identified by local dynamics features. Anomalous phase transition directions include phase transitions in the metastable neural network from subcritical to supercritical activity modes and from refractory period activity modes.

[0110] like Figure 7 As shown, when an abnormal phase transition direction is detected, the phase sequence of the EEG signal of the abnormal node for 10 seconds (i.e., 100 steps) is obtained, and the phase prediction value of the EEG signal of the abnormal node for the next 10 seconds is obtained based on the adaptive EEG signal forward predictor; the prediction value is reversed to generate a phase region coupling interference stimulation command.

[0111] Example 2:

[0112] This embodiment provides an adaptive control method for abnormal discharge based on neural network phase transition tracking, including the following steps:

[0113] S1. Based on large-sample structural magnetic resonance images and subjects' electroencephalogram (EEG) signals, a neural network metastable phase transition direction predictor and an adaptive EEG signal forward predictor are constructed.

[0114] The method for constructing a neural network metastable phase transition direction predictor includes: firstly, constructing a population neural network metastable dynamic model based on a large sample of structural magnetic resonance images; then, fitting the model based on the subject's long-term electroencephalogram signals, training to obtain an individualized neural network metastable dynamic model, and finally obtaining a neural network metastable phase transition direction predictor.

[0115] The specific process includes: using the cortical brain regions corresponding to the scalp electrodes as network nodes, and using the number of fiber bundles between network nodes as the weights of the connection edges, to construct a structural brain neural network; that is, the number of white matter fiber bundles between node i and node j is used as the weight w of the connection edge. ij .

[0116] In the aforementioned brain neural network structure, based on different excitability values ​​E j Two node types are defined: DSCH (Discharge Anomalous Nodes), which are nodes that exhibit spontaneous discharge behavior; and non-DSCH (Discharge Non-Discharge Nodes), which are nodes that cannot spontaneously discharge anomalously, but have a probability of entering an anomalous discharge state after interacting with DSCH nodes. The node types are defined as follows: Equations (24) and (25) are as follows:

[0117] DSCH={j∈{1,2,...,N}|Ej >0}, (24)

[0118] non - DSCH = {j ∈ {1, 2,..., N}|E j <0}. (25)

[0119] Assign a ternary variable x(t) = {0, - 1, 1} to each node at each time point; the ternary variable represents three network activity patterns, including: the refractory period activity pattern after discharge, the sub - critical period activity pattern before discharge, and the super - critical period activity pattern during discharge.

[0120] After that, construct the node activity transition probability weighted by the connecting edges and calculate the transition probabilities between the three network activity patterns based on the node activity transition probability. The node activity transition probability is controlled by the following conditions:

[0121]

[0122] In the formula, P(x i = 1 in (t, t + Δ]|x i (t)= - 1)) represents the probability that node i changes from the sub - critical activity pattern to the super - critical activity pattern within a time step (t, t + Δ); I(·) is the step function, which is defined as: for the step function I(y), it is 0 when y < 0, y when 0 < y < 1, and 1 when y > 1; z i (t) represents the energy input of node i; E i represents the excitability value of node i; o(·) represents an infinitesimal of higher order; P(x i = 0 in (t, t + Δ]|x i (t)= 1)) represents the probability that node i changes from the super - critical activity pattern to the refractory period activity pattern within a time step (t, t + Δ); G(·) is the rate function, which is defined as: for the velocity function G(T, τ, q), it is 0 when T < τ - q, and 1 / (τ + q - T) when τ - q < T < τ + q; t so，i 、t sf，i respectively represent the start time and end time of abnormal discharge of node i; τ s，i 、q si respectively represent the variability of the time scale and duration of abnormal discharge of node i; τ r，i ,q ri respectively represent the variability of the time scale and duration of the refractory period of node i; P(x i = - 1 in (t, t + Δ]x i (t)= 0)) represents the probability that node i changes from the refractory period activity pattern to the sub - critical activity pattern within a time step (t, t + Δ).

[0123] Among them, z i (t) can be expressed as:

[0124]

[0125] In the formula, This represents the excitatory input from node j to node i; This indicates the inhibitory effect of the remaining inactive nodes on the activity of node i.

[0126] The parameter conditions for the occurrence of partial abnormal discharge are as follows:

[0127]

[0128] In the formula, E DSCH The excitability value of the abnormal discharge node; z DSCH This indicates the surrounding suppression of abnormal discharge nodes by non-abnormal discharge nodes; when no abnormal discharge occurs, z DSCH It can be represented as The structural connection weighting value wW between the surrounding non-abnormal discharge node j and the abnormal discharge node represents the structural connection weighting value wW. DSCH，j Constraints on anomalous discharge nodes; E represents the excitability of non-nominal discharge nodes, w is the global connectivity strength, and b is a parameter greater than 0.

[0129] The parameter conditions for the propagation of abnormal discharge are as follows:

[0130]

[0131] In the formula, E represents the excitability of the non-abnormal discharge node; This indicates that the surrounding anomalous discharge nodes are located on the time scale τ. s Continuous discharge D DCSH The excitatory energy input to non-abnormal discharge nodes; This represents the inhibitory constraint of surrounding non-abnormal discharge nodes on the propagation of abnormal discharge; w is the global connectivity strength, and a and b are parameters greater than 0.

[0132] The parameter range that satisfies formulas (28) and (29) indicates that the network is in the supercritical activity mode. The parameter range that satisfies formula (28) but not formula (29) indicates that the network is in the subcritical activity mode. The parameter range that does not satisfy formulas (28) and (29) indicates that the network is in the refractory period activity mode.

[0133] Subsequently, the long-term EEG signals of the subjects were artificially divided into N non-overlapping segments containing the post-discharge, pre-discharge, and discharge periods. The phase synchronization network of these signals was calculated with a window length of 100ms. The population model was fitted using N-fold cross-validation to train an individualized neural network metastable dynamic model and obtain a neural network metastable phase transition direction predictor.

[0134] The specific process of constructing an adaptive EEG signal feedforward predictor includes:

[0135] Given a low-dimensional latent state matrix x of dimension n t A low-dimensional latent state model is constructed using weight matrices A and C, and noise matrices Q, R, and S. This low-dimensional latent state model describes the phase sequence of the EEG signal. The low-dimensional latent state model is as follows:

[0136] X t+1 =A t x t +w t (30)

[0137] y t =C t x t +v t (31)

[0138] In the formula, w t and v t It is modeled as Gaussian white noise, characterized by a covariance matrix containing the noise matrices Q, R, and S:

[0139]

[0140] By combining a low-dimensional latent state model with a recursive Kalman filter, an adaptive EEG signal feedforward predictor is constructed to realize the low-dimensional latent matrix x for the next future step size. t+1 With phase sequence y t+1 Prediction:

[0141]

[0142]

[0143] In the formula, This represents the low-dimensional state value from the previous step. This represents the predicted value of the low-dimensional state in the next step. Kt represents the next predicted value of the signal phase sequence; A represents the Kalman gain matrix; t C t C represents the current model weight matrix. t-1 This represents the model weight matrix from the previous step; y tThis indicates the current step size EEG signal phase value.

[0144] The Kalman gain matrix includes:

[0145]

[0146]

[0147] In the formula, X is a solution to the Riccati equation; R t This represents the current noise covariance matrix.

[0148] To quantify the performance of the adaptive EEG signal forward predictor, the forward prediction error e at time t is:

[0149]

[0150] Therefore, the explained variance of the forward prediction of the i-th channel is used as an evaluation index for prediction performance:

[0151]

[0152] In the formula, T represents the total number of time steps.

[0153] After preprocessing the multi-channel EEG signals, the phase sequence y is obtained by Hilbert transform. t The phase sequence of the EEG signal was artificially divided into T non-overlapping segments of 10 seconds each, and the model parameters were trained using N-fold cross-validation.

[0154] S2. Collect long-term EEG signals from the subject and calculate metastable dynamic characteristics based on the EEG signals. Identify the metastable phase transition direction based on the metastable dynamic characteristics, the neural network metastable phase transition direction predictor, and the adaptive EEG signal forward predictor. When the metastable phase transition direction is abnormal, generate a stimulation command.

[0155] First, the EEG signal is preprocessed. After filtering out the 50Hz power frequency signal using a notch filter, the EEG signal is subjected to real-time artifact removal based on the artifact space reconstruction algorithm. A sliding window with a fixed step size of 100ms is used. Then, a Fast Fourier Transform (FFT) is performed on the EEG signal, and a mask function is used for convolution to double the frequency components greater than 0. Next, an inverse FFT is performed to obtain the analytical signal. The coordinate rotating digital computer (CORDIC) method is used to calculate the arctan function of the analytical signal to obtain the instantaneous phase value of the signal. The phase synchronization value between each pair of channels is calculated to generate the phase synchronization network connection matrix. The instantaneous phase values ​​of 100 steps and the phase synchronization matrix are stored in RAM.

[0156] Subsequently, metastable dynamic characteristics are calculated based on the network connection matrix. These characteristics are then input into a neural network metastable phase transition direction predictor to identify the metastable phase transition direction and anomalous nodes. In this embodiment, the metastable dynamic characteristics include global dynamic characteristics and local dynamic characteristics; the global dynamic characteristics include network critical deceleration and network volatility.

[0157] Among them, the network's critical deceleration is characterized by the autocorrelation AC and variance VAR of the first eigenvector:

[0158]

[0159]

[0160] In the formula, p(t) represents the first eigenvector of the network at step t; N represents the number of steps.

[0161] Network volatility is characterized by the global clustering coefficient CC:

[0162]

[0163] in,

[0164]

[0165]

[0166] In the formula, k i w represents the node degree. ij Let M represent the network connection edge value between node i and node j, and M represent the number of nodes that have an edge connection with node i.

[0167] Local dynamic features include: node source index, sink index, and sink connectivity (conn), which are respectively:

[0168] source = 1.41 - |(rr) i ,cr i )-(1 / N,1)|, (44)

[0169] sink i =1.41-|(rr i ,cr i )-(1,1 / N)|, (45)

[0170]

[0171] In the formula, rr i cr represents the rank of the row in the network connectivity matrix where the node is located. iThe rank of the column in the network connection matrix where the node is located is represented, abs() represents the absolute value, and N is the number of network nodes.

[0172] The network's critical deceleration and volatility characteristics, along with node source, sink, and sink connectivity features, are input into a neural network metastable phase transition direction predictor with 100 steps to identify anomalous phase transition directions. Anomalous nodes are identified by local dynamics features. Anomalous phase transition directions include phase transitions in the metastable neural network from subcritical to supercritical activity modes and from refractory period activity modes.

[0173] When an abnormal phase transition direction is detected, the phase sequence of the EEG signal of the abnormal node for 10 seconds (i.e., 100 steps) is obtained, and the phase prediction value of the EEG signal of the abnormal node for the next 10 seconds is obtained based on the adaptive EEG signal forward predictor; the prediction value is reversed to generate a phase region coupling interference stimulation command.

[0174] S3. Based on the stimulation command, generate an electrical stimulation signal and perform electrical stimulation. After the stimulation is completed, continue to collect EEG signals and cycle through S2-S3 to achieve adaptive electrical stimulation regulation of abnormal EEG discharges.

[0175] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An abnormal discharge adaptive regulation system based on neural network phase transition tracking, characterized in that, The application relates to a brain electrical signal acquisition-stimulation integrated device and an FPGA real-time computing device. The host computer is used for constructing a neural network metastable phase transition direction predictor and an adaptive brain electrical signal forward predictor based on large-sample structural magnetic resonance images and brain electrical signals of a subject, and transmitting the neural network metastable phase transition direction predictor and the adaptive brain electrical signal forward predictor to the FPGA real-time computing device for storage; wherein the method for constructing the neural network metastable phase transition direction predictor comprises the following steps: A structural brain neural network is constructed by taking a scalp electrode corresponding cortex brain region as a network node and taking the number of fiber bundles between network nodes as a connection edge weight; In the structural brain neural network, two types of nodes are defined based on excitability values, including an abnormal discharge node and a non-abnormal discharge node; a ternary variable is assigned to each node at each time point; the ternary variable represents three network activity modes, including a post-discharge refractory period activity mode, a pre-discharge subcritical period activity mode and a discharge supercritical period activity mode; A node activity transition probability weighted by a connection edge is constructed, and a transition probability between the three network activity modes is calculated based on the node activity transition probability; The method for constructing the adaptive brain electrical signal forward predictor comprises the following steps: constructing a low-dimensional latent state model from a latent state matrix, a weight matrix and a noise matrix; and combining the low-dimensional latent state model with a recursive Kalman filter to construct the adaptive brain electrical signal forward predictor. The phase transition refers to a sudden transition of a neural network functional submode of the brain from a relatively stable mode to an abnormal activity mode. The brain electrical signal acquisition-stimulation integrated device is used for acquiring the brain electrical signals and generating an electrical stimulation signal based on a stimulation instruction issued by the FPGA real-time computing device. The FPGA real-time computing device is used for receiving the brain electrical signals, calculating metastable state dynamics characteristics based on the brain electrical signals, identifying a metastable state phase transition direction based on the metastable state dynamics characteristics, and generating the stimulation instruction when the metastable state phase transition direction is abnormal. The host computer comprises a database module, a neural network phase transition predictor generation module and an adaptive brain electrical signal predictor generation module.

2. The neural network-based phase tracking abnormal discharge adaptive regulation system according to claim 1, wherein, The database module is used for storing large-sample structural magnetic resonance images and the brain electrical signals. The neural network phase transition predictor generation module is used for constructing the neural network metastable phase transition direction predictor. The adaptive brain electrical signal predictor generation module is used for constructing the adaptive brain electrical signal forward predictor. The control condition of the node activity transition probability comprises:

3. The neural network-based phase tracking abnormal discharge adaptive regulation system according to claim 1, wherein, The adaptive brain electrical signal forward predictor comprises: ， where, denotes the probability of node i transitioning from subcritical active mode to supercritical active mode within a time step (t, t+∆); I(∙) is the step function; z i (t) denotes the energy input to node i; E i denotes the excitability value of node i; o(∙) denotes a higher-order small error; denotes the probability of node i transitioning from supercritical active mode to refractory active mode within a time step (t, t+∆); G(∙) is the rate function; t so,i , t sf,i denote the start time and end time of an abnormal discharge of node i, respectively; τ s,i , q si denote the time scale and variability of the duration of an abnormal discharge of node i, respectively; τ r,i , q ri denote the time scale and variability of the duration of a refractory period of node i, respectively; denotes the probability of node i transitioning from refractory active mode to subcritical active mode within a time step (t, t+∆).

4. The neural network-based phase tracking abnormal discharge adaptive regulation system according to claim 1, wherein, The FPGA real-time computing device comprises a neural network real-time generation module, a neural network phase transition tracking module and an adaptive stimulation instruction generation module. ， ， wherein, represents the low-dimensional state sequence value of the previous step, represents the predicted value of the low-dimensional state at the next step; represents the next step prediction value of the signal phase sequence; K t represents the Kalman gain matrix; A t、 C t represents the current model weight matrix, C t-1 represents the model weight matrix of the previous step; y t represents the electroencephalogram signal phase sequence.

5. The neural network based phase tracking abnormal discharge adaptive regulation system according to claim 1, wherein, The neural network real-time generation module is used for pre-processing brain electrical signals and generating a phase synchrony network connection matrix based on the pre-processed brain electrical signals. ​ The neural network phase transition tracking module is configured to calculate metastable state dynamics features based on the network connection matrix, and input the metastable state dynamics features into the neural network metastable phase transition direction predictor to identify abnormal phase transition directions and abnormal nodes; When an abnormal phase transition direction is identified, the label of the abnormal node is input into the adaptive stimulation instruction generation module, the adaptive stimulation instruction generation module acquires a 10s electroencephalogram phase sequence of the abnormal node, obtains a 10s electroencephalogram phase prediction value of the abnormal node in the future based on the adaptive electroencephalogram forward predictor, and generates the stimulation instruction based on the prediction value.

6. The method of adaptive regulation of abnormal discharge based on neural network phase transition tracking, the method is applied to the system of any one of claims 1-5, characterized in that, The method comprises the following steps: S1, constructing a neural network metastable phase transition direction predictor and an adaptive electroencephalogram forward predictor based on a large sample structural magnetic resonance image and electroencephalogram signals of a subject; S2, acquiring long-time electroencephalogram signals of the subject, calculating metastable state dynamics features based on the electroencephalogram signals, identifying metastable phase transition directions based on the metastable state dynamics features, the neural network metastable phase transition direction predictor and the adaptive electroencephalogram forward predictor, and generating a stimulation instruction when the metastable phase transition direction is abnormal; S3, generating an electric stimulation signal based on the stimulation instruction, the electric stimulation signal being used for non-invasive electric stimulation, and continuing to acquire electroencephalogram signals after the non-invasive electric stimulation is completed, and circulating S2-S3 to realize adaptive electric stimulation regulation of electroencephalogram abnormal discharge.

Images