Mental disease brain dysfunction detection method based on dynamic causal modeling

Abstract

The invention discloses a mental disease brain dysfunction detection method based on dynamic causal modeling, and belongs to the field of biomedical engineering. The dynamic causal relationship of information transmission among regions of the brain can be accurately captured by adopting dynamic causal modeling based on electroencephalogram, and the method has unique advantages in the aspect of revealing abnormal modes of mental disease information transmission. Different from a traditional functional brain network construction mode, the system pays more attention to capturing the causal relationship and dynamic change of brain intervals when processing and explaining data, the modeling method based on the causal relationship can provide more detailed and personalized information transmission analysis to a certain extent, a more accurate result is provided, and the method is suitable for popularization and application. Furthermore, the functional disorder in the mental disease brain neural network can be sensitively recognized, the pathological mechanisms of bipolar affective disorder, schizophrenia and split affective disorder can be revealed, and support is provided for early diagnosis and personalized treatment.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical engineering, and particularly relates to a method for detecting brain functional disorders of mental diseases based on dynamic causal modeling. Background Art

[0002] Mental diseases, such as bipolar disorder, schizophrenia and schizoaffective disorder, are often accompanied by significant changes in brain neurocognitive functions. Among them, abnormal information transmission is one of the typical manifestations of such diseases, which not only affects the cognitive ability, emotion regulation and social behavior of patients, but also has a significant impact on their daily life. Accurately detecting and identifying the abnormal brain information transmission function of patients with mental diseases is of great significance for the early diagnosis and intervention treatment of the diseases. Compared with the problems such as strong subjectivity existing in the traditional clinical behavior assessment methods, electroencephalogram (EEG) can measure the synchronous sum of postsynaptic potentials of pyramidal cells through electrode devices, and can directly and objectively record and evaluate the brain function abnormalities of patients with mental diseases in real time. At the same time, it also has the advantages of portability, high time resolution, non-invasiveness and low cost. In particular, the dynamic causal modeling (DCM) analysis based on EEG can synchronously reveal the complex dynamic characteristics of brain network circuits in time and space, which is crucial for understanding the brain function abnormalities caused by information transmission disorders in different pathological states of various mental diseases. More importantly, this method can model the causal relationship of information transmission between different brain regions, accurately capture the dynamic transmission process of high-level cognitive information in the brain among different brain regions, and provide a more accurate and comprehensive brain function map than traditional methods. Therefore, through in-depth analysis and understanding of the information transmission mode in the cognitive process, DCM is expected to identify functional disorders in the brain neural network and provide strong support for the early diagnosis and personalized treatment of mental diseases. This paper proposes a system for detecting cognitive function disorders of mental diseases based on dynamic causal modeling. This system mainly performs dynamic causal modeling based on the task-state EEG data of the auditory P300 of patients with mental diseases to capture the disease-specific information transmission abnormal patterns of bipolar disorder, schizophrenia and schizoaffective disorder, so as to realize the detection of corresponding brain functional disorders. In this system, based on the auditory P300 EEG signals of bipolar disorder, schizophrenia and schizoaffective disorder, by integrating various technical means such as dynamic causal modeling, feedback control mechanism analysis and Bayesian model selection, the disease-specific disorders of causal connections between brain regions are accurately characterized and analyzed to reveal the neural mechanisms of abnormal bottom-up and top-down information transmission in mental diseases; then based on the 2×3 factor design unit, this system further detects the main effects of the diagnosis and genetic risk of mental diseases on the intrinsic connectivity and their interaction with the task condition effect, so as to effectively identify the brain functional disorders of bipolar disorder, schizophrenia and schizoaffective disorder, which is of great significance for the accurate and intelligent diagnosis and treatment of mental diseases. Summary of the Invention

[0003] To address the limitations of the current mental illness diagnosis system in objectively quantifying brain function abnormalities mainly characterized by disease information transmission disorders, the present invention proposes a method for detecting brain function disorders of mental illness based on dynamic causal modeling.

[0004] The present invention proposes a method for detecting brain function disorders of mental illness based on dynamic causal modeling, including the following steps:

[0005] Step S1: Preprocess the electroencephalogram (EEG) signals collected from the scalp surface.

[0006] Select the electrodes of interest, re-reference the EEG signals using the reference electrode standardization technique, extract the frequency band of interest and filter out the signal noise components using a 1 - 45 Hz band-pass filter, downsample the signal, segment the signal, perform baseline correction, remove the signal segments containing artifacts, and obtain the event-related potential signals through the superposition and averaging of data segments.

[0007] Step S2: Establish a spatial source prior model:

[0008] The spatial source prior model incorporates three brain regions: the primary auditory cortex, the parietal lobe, and the superior frontal gyrus; the spatial coordinates of the cortical sources in the above regions are as follows: the left (-59, -10, 13) and right (61, -13, 11) primary auditory cortices, the left (-37, -48, 68) and right (28, -56, 63) parietal lobes, and the left (-29, 55, 22) and right (27, 60, 20) superior frontal gyri.

[0009] Step S3: Model the cortical activity sources;

[0010] The electrical activity in each brain region is represented by a single equivalent current dipole, and the dipole is placed at a predefined position on the cerebral cortex, assuming that the electrical activity in this part of the region is caused by the current of this dipole; each dipole has the following two basic properties: (1) Dipole moment: The dipole moment describes the magnitude and direction of the current; (2) Position: The spatial position of the dipole indicates the coordinates of the actual current source in the brain, which is often determined by the brain anatomical coordinate system; the magnitude of the dipole moment is proportional to the current intensity of the nerve activity source, and the direction represents the direction of current flow; its mathematical formula is J = pδ(r - r0), where J is the current density, p is the dipole moment, r0 is the position of the dipole, δ is the Dirac δ function, and r represents the electrode position.

[0011] Step S4: Model the neural model:

[0012] The CMC model is adopted as the neural model, and the CMC model is the microcircuit neural mass model; the CMC model assumes that each neural source contains four cell populations, which have different functions and interaction modes; the electrical activity of each population is described by current changes, and these currents interact in the cerebral cortex activity to generate the overall neural activity signal;

[0013] Step S5: Construct the extrinsic connections:

[0014] The presence or absence of forward connections, backward connections, and intrinsic connections is defined, and 64 different transmission models are constructed, including 63 candidate models, covering forward, backward, and / or intrinsic connection effects, and a complete model without external connections;

[0015] Step S6: Electromagnetic model modeling:

[0016] In the boundary element method modeling process, a spherical approximation model is first selected, which contains three concentric layers, and different head regions are simulated through a fine node distribution; then, in these models, 4000 test dipoles are randomly placed, their potential distributions are calculated, and the data is fitted to a single dipole to evaluate the localization error; a refined triangular potential approximation method and an isolation problem method are adopted; during this process, the setting of the electrode position relative to the outer layer nodes has an important impact on the error evaluation, and the model accuracy is further improved through virtual refinement processing;

[0017] Step S7: Model inversion of the dynamic causal model:

[0018] Through the above modeling steps, the preprocessed EEG data in step S1 is input into the defined dynamic causal model for model inversion;

[0019] Step S8: Bayesian model selection:

[0020] A model with the maximum log evidence among all test models is found through the Bayesian model, and assuming that the prior probabilities of all models are equal, to determine which model can better explain the data of each research effect and interaction effect;

[0021] Step S9: Brain dysfunction detection:

[0022] After determining the optimal model, by statistically comparing and visualizing the posterior estimates of the model's intrinsic connectivity, the differences among patients, relatives, and the control group can be identified; then, the connectivity differences exceeding the threshold are defined as significant effects, so as to achieve the purpose of identifying the brain dysfunction connection edges and presenting the results of abnormal regions.

[0023] Furthermore, step S1 specifically includes:

[0024] Step S11: Select 32 channels from the collected EEG signal channels for subsequent analysis, including the following channels: Fp1, AF3, F3, F7, FC5, FC1, C3, T7, CP5, CP1, P3, P7, PO3, O1, Oz, Pz, Fp2, AF4, F2, F4, F8, FC6, FC2, Cz, C4, T8, CP6, CP2, P4, P8, PO4, O2; These channels are defined according to the international 10-20 standard lead system. Perform re-referencing processing on the EEG signals, use reference electrode standardization technology to remove the influence of the reference electrode, and enhance the data quality; Use a 1-45 Hz band-pass filter to extract the frequency band of interest and filter out the signal noise components, and downsample the signals; Segment each segment of EEG data, and select the time interval [-200, 800] ms, where 0 ms represents the start time of the stimulus.

[0025] Step S12: Baseline correction and artifact rejection: Perform baseline correction on the EEG signals, select the time interval [-200, 0] ms as the baseline, and perform artifact removal processing; Use a threshold of ±75 μV for artifact rejection to ensure the removal of artifact signals caused by factors such as eye movement and muscle activity; Further clean the processed EEG signals to ensure that the data quality meets the requirements of subsequent analysis;

[0026] Step S13: Obtain P300 data: By averaging the trials of the target stimulus and the standard stimulus respectively, calculate the group-averaged ERP of each participant, and then estimate the P300 waveform of all participants. The amplitude of P300 is obtained by extracting the average amplitude in the time window of ±20 ms near the maximum positive peak within the time window of 300 to 600 milliseconds after the target stimulus.

[0027] Furthermore, the neural model in step S4 is constructed as follows:

[0028] Step S41: Model the neural intracellular populations as four cell populations, including: Superficial pyramidal cells: Superficial pyramidal cells are important excitatory neurons in the cortex, responsible for transmitting information from the upper cortical layers; Deep pyramidal cells: Deep pyramidal cells are located in the deeper layers of the cortex and participate in long-distance connections between cortices; Spiny stellate cells: Spiny stellate cells are another type of excitatory neurons in the cortex, located in the middle layer of the cortex; Inhibitory interneurons: Inhibitory interneurons play an inhibitory role in the neural network. They generate inhibitory signals through connections with other neurons;

[0029] Step S42: The activity of the neural source is described by the change in membrane potential; The change in membrane potential is affected by the input current and the interconnections between neurons; Assume there is a neuron population, and its change in membrane potential is expressed as: Among them, V(t) is the membrane potential, τ is the time constant of the membrane potential, which reflects the response speed of neurons to stimuli, and I ext (t) is the external input current, and I syn (t) is the synaptic input current from other neuron populations;

[0030] Step S43: The synaptic input current comes from the connections between different neural populations, and these connections are excitatory or inhibitory; the synaptic current is expressed by the following formula: I syn (t) = ∑ i w i ·g i (V(t) - V rest )·θ(t), where w i is the connection strength or synaptic weight, representing the connection strength between neural populations, g i is the conductance of each neuron population, reflecting their response strength to the input, V rest is the resting potential, and θ(t) is the step function, indicating whether the neuron is activated;

[0031] Step S44: In the CMC model, the synaptic input includes both excitatory input and inhibitory input; the excitatory input increases the membrane potential of the neuron, while the inhibitory input decreases its membrane potential; the excitatory synaptic input is expressed as: I exc (t) = w exc ·(V(t) - V rest ), and the inhibitory synaptic input can be expressed as: I inh (t) = w inh ·(V(t) - V rest ), where w exc and w inh respectively represent the connection weights of excitatory and inhibitory synapses;

[0032] Step S45: In the neuron population, the synaptic current of the feedback connection is expressed as: I rev (t) = w rev ·(V deep (t) - V shallow (t)), where V deep (t) and V shallow (t) respectively represent the membrane potentials of deep and shallow pyramidal cells, and w rev represents the synaptic weight of the feedback connection; the calculation of the self-inhibition current is expressed as: I self (t) = -w self ·g(V(t) - V rest ), and w selfIt represents the self-inhibitory synaptic weight, and g(.) represents the self-inhibitory synaptic transmission function, which describes the decay or delay effect of the synaptic current over time;

[0033] Step S46: The output of the neuron population is determined by the activities of the neurons, and the output current is calculated from the membrane potential or other electrophysiological signals; the output of the population is expressed by the following formula: where τ0 is the time constant of the output current.

[0034] Furthermore, the specific method of step 5 is as follows:

[0035] By defining the presence or absence of forward connections, backward connections, and intrinsic connections, three forward models are constructed:

[0036] The f1 model is defined as the forward connections from l-SP to l-SF and from r-SP to r-SF;

[0037] The f2 model is defined as the forward connections from l-A1 to l-SP and from r-A1 to r-SP;

[0038] The f3 model is defined as the forward connections from l-A1 flowing through l-SP to l-SF and from r-A1 flowing through r-SP to r-SF;

[0039] Three backward models:

[0040] The b1 model is defined as the backward connections from l-SF to l-SP and from r-SF to r-SP;

[0041] The b2 model is defined as the backward connections from l-SP to l-A1 and from r-SP to r-A1;

[0042] The b3 model is defined as the backward connections from l-SF flowing through l-SP to l-A1 and from r-SF flowing through r-SP to r-A1;

[0043] Seven intrinsic connection models:

[0044] The i1 model is defined as the intrinsic connection between the two brain regions of l-SF and r-SF;

[0045] The i2 model is defined as the intrinsic connection between the two brain regions of l-SP and r-SP;

[0046] The i3 model is defined as the intrinsic connection between the two brain regions of l-A1 and r-A1;

[0047] The i4 model is defined as the intrinsic connection among the four brain regions of l-SF, r-SF, l-SP, and r-SP;

[0048] The i5 model is defined as the intrinsic connection among the four brain regions of l-SP, r-SP, l-A1, and r-A1;

[0049] The i6 model is defined as the intrinsic connectivity of four brain regions: l-SF, r-SF, l-A1, and r-A1;

[0050] The i7 model is defined as the intrinsic connectivity of six brain regions: l-SF, r-SF, l-SP, r-SP, l-A1, and r-A1;

[0051] One unconnected model n is defined as all brain regions being independent of each other without connections;

[0052] By combining different combinations of forward, backward, and intrinsic connectivity models, 3×3×7 = 63 different transmission models were constructed; in addition, a complete model n without external connections was set up, and a total of 64 different models were constructed;

[0053] Among them, l represents the left brain region, r represents the right brain region, SP represents the parietal lobe, SF represents the superior frontal gyrus, and A1 represents the primary auditory cortex.

[0054] Furthermore, the model inversion process based on dynamic causal modeling in step S7 is as follows:

[0055] Step S71: The core calculation of Bayesian model inversion depends on Bayes' formula. During the model inversion process, the posterior probability distribution p(θ|D) and the Bayes factor are calculated. Given two models M1 and M2, the Bayes factor β 12 is used to measure the advantage of M1 over M2: where p(D|M) is the marginal likelihood of the data given model M;

[0056] Step S72: Perform dynamic causal modeling at the group level by creating a 2×3 factorial design unit, which includes two levels of "task condition" and three levels of "group". The two levels of "task condition" are standard stimulus and target stimulus, and the three levels of "group" are patients, relatives, and control groups; the group effects studied include: (1) "diagnostic effect": comparison of patients with the combined group of relatives and controls; (2) "genetic risk effect": comparison of relatives with the control group; test the main effects of diagnosis and genetic risk on intrinsic connectivity, as well as their interaction effects with the task condition effect.

[0057] Furthermore, the Bayesian model selection process in step S8 is as follows:

[0058] Step S81: Given multiple models M1, M2,..., M k , each model has a corresponding parameter set θ1, θ2,..., θ k ; according to Bayes' theorem, when given data D, the posterior probability P(M i |D) of model M i is calculated by the following formula: Among them, P(M i |D) is the posterior probability of model M i . P(D|M i ) is the marginal likelihood of model M i , also known as the model evidence; P(M i ) is the prior probability of model M i . P(D) is the total likelihood of the data, which is used to normalize the calculation of the posterior probability; in Bayesian model selection, the goal is to select the model with the maximum marginal likelihood P(D|M i ) by calculating the marginal likelihood of all candidate models and comparing them to determine which model best explains the data;

[0059] Step S82: Calculate the marginal likelihood P(D|M i ). The marginal likelihood represents the fitness of model M i given the data D;

[0060] The marginal likelihood is calculated by integrating over all parameters of the model: P(D|M i ) = ∫P(D|θ i , M i )P(θ i , M i )dθ i , where P(D|θ i , M i ) is the likelihood of the data given the parameter θ i and model M i . P(θ i , M i ) is the prior distribution of the parameters of model M i ;

[0061] Step S83: Calculate the free energy approximation;

[0062] Introduce a variational distribution q(θ), and approximate the marginal likelihood with the free energy: logP(D|M i ) = E q [logP(D|θ i , M i )] - KL(q(θ i ) ∥ P(θ i , M i ), where E q [logP(D|θ i , M i )] is the expectation of the variational distribution q(θ) of the parameter θ i . KL(q(θ i ) ∥ P(θ i , M i)) is the variational distribution q(θ i ) and the Kullback-Leibler divergence between the true parameter prior P(θ i , M i );

[0063] Step S84: Model selection;

[0064] Select the model with the maximum log evidence, and calculate the log marginal likelihood logP(D|M i ): logP(D|M i ) = log(∫P(D|θ i , M i )P(θ i , M i )dθ i ); Then, compare the log evidence of different models and select the model with the maximum log evidence;

[0065] Step S85: Establish the criteria for model comparison;

[0066] A difference in log evidence greater than 3 is strong evidence indicating that the model with greater evidence is significantly better than other models;

[0067] Furthermore, a connectivity difference of 20% or more in Step S9 is considered a significant effect to achieve the purpose of identifying the obstacle edge and presenting the results.

[0068] The advantages of the present invention are as follows:

[0069] Adopting electroencephalogram-based dynamic causal modeling can accurately capture the dynamic causal relationships of information transmission between various regions of the brain, and has unique advantages in revealing abnormal patterns of information transmission in mental diseases. Different from the traditional method of constructing functional brain networks, this system pays more attention to capturing the causal relationships and dynamic changes between brain regions when processing and interpreting data. This causal relationship-based modeling method can, to a certain extent, provide a more detailed and personalized analysis of information transmission, provide more accurate results, and further enable sensitive identification of functional disorders in the brain neural networks of mental diseases, helping to reveal the pathological mechanisms of bipolar disorder, schizophrenia, and schizoaffective disorder, and providing support for early diagnosis and personalized treatment. Brief description of the drawings

[0070] Figure 1 It is a step schematic diagram of a mental disease brain dysfunction detection system based on dynamic causal modeling provided by the present invention.

[0071] Figure 2 It is a step schematic diagram of the establishment of a mental disease brain dysfunction detection system based on dynamic causal modeling provided by the present invention.

[0072] Figure 3 This provides an external connection definition diagram for model construction of the present invention.

[0073] Figure 4 This provides a relative log evidence and posterior probability diagram for each model based on dynamic causal modeling of the present invention.

[0074] Figure 5 This provides a posterior estimate diagram of the intrinsic connection of the optimal model under the optimal model of each source and experimental effect of the present invention. Detailed implementation manners

[0075] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0076] In a first aspect, according to an embodiment of the present invention, a mental disorder brain dysfunction detection system based on dynamic causal modeling is proposed. Please refer to Figure 1 , which includes the following steps:

[0077] Step S1: Perform the following preprocessing process on the EEG signals collected from the head surface:

[0078] Select the electrodes of interest, re-reference the EEG signals using the reference electrode standardization technique, extract the frequency band of interest and filter out the signal noise components using a 1 - 45 Hz band-pass filter, downsample the signal, segment the signal, perform baseline correction, remove the signal segments containing artifacts, and obtain the event-related potential signals through superposition averaging of data segments;

[0079] Step S2: Establish a spatial source prior model:

[0080] According to literature research, the definition of the prior model in this system incorporates three brain regions that play a crucial role in P300 cognitive information processing: the primary auditory cortex, the parietal lobe, and the superior frontal gyrus. The spatial coordinates of the cortical sources in the above regions are as follows: left (-59, -10, 13) and right (61, -13, 11) primary auditory cortex (l- / r-A1); left (-37, -48, 68) and right (28, -56, 63) parietal lobe (l- / r-SP); left (-29, 55, 22) and right (27, 60, 20) superior frontal gyrus (l- / r-SF);

[0081] Step S3: Cortical activity source modeling:

[0082] To simplify the model, it is assumed that the electrical activity in each brain region can be represented by a single equivalent current dipole, which can significantly reduce the computational complexity. By placing the dipole at predefined positions in the cerebral cortex (such as the primary auditory cortex, parietal lobe, frontal lobe, etc.), it can be assumed that the electrical activity in this part of the region is mainly caused by the current of this dipole. Each dipole has the following two basic properties: (1) Dipole moment: The dipole moment describes the magnitude and direction of the current, and the unit is usually ampere-meter (A·m). (2) Position: The spatial position of the dipole indicates the coordinates of the actual current source in the brain, which is often determined by the brain anatomical coordinate system. The magnitude of the dipole moment is proportional to the current intensity of the neural activity source, and the direction represents the direction of current flow. Its mathematical formula is J = pδ(r - r0), where J is the current density, p is the dipole moment, r0 is the position of the dipole, δ is the Dirac δ function, and r represents the electrode position;

[0083] Step S4: Neural model modeling:

[0084] The CMC model is used for neural model modeling. The CMC model assumes that each neural source contains four cell populations, and these populations have different functions and interaction methods. The electrical activity of each population is described by current changes, and these currents interact in the cerebral cortex activity to generate the overall neural activity signal;

[0085] Step S5: Construct external connections:

[0086] Three forward models (f) were constructed by defining the presence or absence of forward connections, backward connections, and intrinsic connections: the f1 model is defined as the forward connections from l-SP to l-SF and from r-SP to r-SF; the f2 model is defined as the forward connections from l-A1 to l-SP and from r-A1 to r-SP; the f3 model is defined as the forward connections where l-A1 flows through l-SP to l-SF and r-A1 flows through r-SP to r-SF. Three backward models (b): the b1 model is defined as the backward connections from l-SF to l-SP and from r-SF to r-SP; the b2 model is defined as the backward connections from l-SP to l-A1 and from r-SP to r-A1; the b3 model is defined as the backward connections where l-SF flows through l-SP to l-A1 and r-SF flows through r-SP to r-A1. And seven intrinsic connection models (i): the i1 model is defined as the intrinsic connection between the two brain regions of l-SF and r-SF; the i2 model is defined as the intrinsic connection between the two brain regions of l-SP and r-SP; the i3 model is defined as the intrinsic connection between the two brain regions of l-A1 and r-A1; the i4 model is defined as the intrinsic connection among the four brain regions of l-SF, r-SF, l-SP, and r-SP; the i5 model is defined as the intrinsic connection among the four brain regions of l-SP, r-SP, l-A1, and r-A1; the i6 model is defined as the intrinsic connection among the four brain regions of l-SF, r-SF, l-A1, and r-A1; the i7 model is defined as the intrinsic connection among the six brain regions of l-SF, r-SF, l-SP, r-SP, l-A1, and r-A1. One no-connection model n is defined as all brain regions being independent of each other without connections, see Figure 3 . By combining different combinations of forward, backward, and intrinsic connection models, a total of 3(f) × 3(b) × 7(i) = 63 different transmission models were constructed. In addition, there is also a complete no-external-connection model n. Therefore, a total of 64 different models were constructed;

[0087] Step S6: Electromagnetic model modeling:

[0088] In the boundary element method modeling process, a spherical approximation model was first selected, which consists of three concentric layers, and different head regions were simulated through a fine node distribution. Then, in these models, 4000 test dipoles were randomly placed, their potential distributions were calculated, and the data were fitted to a single dipole to evaluate the localization error. To improve the calculation accuracy, a refined triangular potential approximation method and an isolation problem method were adopted. During this process, the setting of the electrode position relative to the outer layer nodes had an important impact on the error evaluation, and the accuracy of the model was further improved through virtual refinement processing;

[0089] Step S7: Model inversion of the dynamic causal model:

[0090] Through the above modeling steps, the preprocessed EEG data in step S1 is input into the defined dynamic causal model for model inversion;

[0091] Step S8: Bayesian model selection:

[0092] Use the Bayesian model to find the model with the maximum log evidence (free energy approximation) among all tested models, and assume that the prior probabilities of all models are equal to determine which model can better explain the data of each research effect and interaction effect. This method balances the accuracy and complexity of the model, thus selecting the most general model;

[0093] Step S9: Brain dysfunction detection:

[0094] After determining the optimal model, by statistically comparing and visualizing the posterior estimates of the model's intrinsic connectivity, the differences among patients, relatives, and the control group can be identified. Subsequently, based on the literature, a connectivity difference of 20% or more is defined as a significant effect, so as to achieve the purpose of identifying the connected edges of brain dysfunction and presenting the results of abnormal regions.

[0095] Furthermore, step S1 specifically includes:

[0096] Step S11: Select 32 channels from the 64 EEG signal channels collected for subsequent analysis, including the following channels: Fp1, AF3, F3, F7, FC5, FC1, C3, T7, CP5, CP1, P3, P7, PO3, O1, Oz, Pz, Fp2, AF4, F2, F4, F8, FC6, FC2, Cz, C4, T8, CP6, CP2, P4, P8, PO4, O2; These channels are defined according to the international 10-20 standard lead system. Perform re-referencing processing on the EEG signals, use the reference electrode standardization technique to remove the influence of the reference electrode, and enhance the data quality. Use a 1-45Hz band-pass filter to extract the frequency band of interest and filter out the signal noise components, and downsample the signals. Segment each segment of EEG data, and select the time interval [-200, 800] ms, where 0 ms represents the start time of the stimulus.

[0097] Step S12: Baseline correction and artifact rejection: Perform baseline correction on the EEG signals, select the time interval [-200, 0] ms as the baseline, and perform artifact rejection. Use a threshold of ±75 μV for artifact rejection to ensure the removal of artifact signals caused by factors such as eye movement and muscle activity. Further clean the processed EEG signals to ensure that the data quality meets the requirements of subsequent analysis.

[0098] Step S13: Obtain P300 data: By averaging the trials of the target stimulus and the standard stimulus respectively, calculate the group-averaged ERP of each participant, and then estimate the P300 waveform of all participants. The amplitude of P300 is obtained by extracting the average amplitude within a ±20 ms time window near the maximum positive peak within the 300 - 600 ms time window after the target stimulus.

[0099] Further, the neural model of the mental disorder brain dysfunction detection system based on dynamic causal modeling in step S4 is constructed as follows:

[0100] Step S41: Model the neural intracellular populations as four cell populations, including: Superficial Pyramidal Cells (SPC): Superficial Pyramidal Cells are important excitatory neurons in the cortex, responsible for transmitting information from the upper layer of the cortex. They send excitatory signals to other cortical regions or deep regions through external connections. The activities of these cells play a key role in regulating the excitability of local neural circuits. Deep Pyramidal Cells (DPC): Deep Pyramidal Cells are located in the deeper layers of the cortex and are usually involved in long-distance connections between cortices. These cells can transmit information back to the superficial cortex through feedback connections and contribute to information exchange between brain regions. They also undertake the function of feedback regulation. Spiny Stellate Cells (SSC): Spiny Stellate Cells are another type of excitatory neurons in the cortex, mainly located in the middle layer of the cortex, especially layer IV. They are activated by incoming external inputs to generate excitatory currents. The interaction between Spiny Stellate Cells and Pyramidal Cells helps with information processing and transmission, especially during the processing of sensory inputs. Inhibitory Interneurons (IIN): Inhibitory Interneurons play an inhibitory role in the neural network. They generate inhibitory signals through connections with other neurons. By regulating the activities of excitatory neurons, Inhibitory Interneurons play a crucial role in maintaining the stability of the neural network and avoiding over-excitation. Their activities contribute to the fine control of neuronal synchrony and response time.

[0101] Step S42: The activity of the neural source is described by the change in membrane potential. The change in membrane potential is affected by the input current and the mutual connection between neurons. Suppose there is a neuron population, and its change in membrane potential can be expressed as: where V(t) is the membrane potential, τ is the time constant of the membrane potential, reflecting the response speed of the neuron to the stimulus, I ext (t) is the external input current, I syn(t) is the synaptic input current from other neuron populations.

[0102] Step S43: The synaptic input current comes from the connections between different neural populations, which can be excitatory or inhibitory. The synaptic current can be expressed by the following formula: I syn (t) = ∑ i w i ·g i (V(t) - V rest )·θ(t), where w i is the connection strength or synaptic weight, representing the connection strength between neural populations, g i is the conductance of each neuron population, reflecting their response strength to the input, V rest is the resting potential, and θ(t) is the step function, indicating whether the neuron is activated (i.e., whether the membrane potential exceeds the threshold).

[0103] Step S44: In the CMC model, the synaptic input includes both excitatory (positive) input and inhibitory (negative) input. Excitatory input increases the membrane potential of the neuron, while inhibitory input decreases its membrane potential. The excitatory synaptic input can be expressed as: I exc (t) = w exc ·(V(t) - V rest ), and the inhibitory synaptic input can be expressed as: I inh (t) = w inh ·(V(t) - V rest ), where w exc and w inh represent the connection weights of excitatory and inhibitory synapses respectively.

[0104] Step S45: In the neuron population, feedback connections and self-inhibition are very important mechanisms for regulating neuron activity. Feedback connections come from the connections of deep pyramidal cells to superficial pyramidal cells, and can regulate the information transmission between brain regions. The self-inhibition mechanism is achieved through inhibitory interneurons. The synaptic current of the feedback connection can be expressed as: I rev (t) = w rev ·(V deep (t) - V shallow (t)), where V deep (t) and V shallow (t) represent the membrane potentials of deep and superficial pyramidal cells respectively, and w rev represents the synaptic weight of the feedback connection. The current calculation of self-inhibition can be expressed as: I self (t) = -w self ·g(V(t) - V rest ), where wself denotes the self-inhibitory synaptic weight, and g(.) denotes the self-inhibitory synaptic transmission function, which describes the decay or delay effect of the synaptic current over time.

[0105] Step S46: The output of the neuron population is usually determined by the activities of neurons, and the output current can be calculated from the membrane potential or other electrophysiological signals. The output of the population can be expressed by the following formula: where τ0 is the time constant of the output current.

[0106] Furthermore, the model inversion process of the mental disorder brain dysfunction detection system based on dynamic causal modeling in step S7 is as follows:

[0107] Step S71: The core calculation of Bayesian model inversion depends on the Bayes formula. During the model inversion process, the posterior probability distribution p(θ|D) and the Bayes factor are usually calculated. The Bayes factor is a measure of the goodness of fit of a model to the data and can be used for model comparison. Suppose there are two models M1 and M2, then the Bayes factor β 12 is used to measure the advantage of M1 over M2: where p(D|M) is the marginal likelihood of the data given model M.

[0108] Step S72: Perform dynamic causal modeling at the group level by creating a 2×3 factorial design unit, which includes two levels of "task condition" (standard stimulus and target stimulus) and three levels of "group" (patients, relatives, and control group). The group effects studied include: (1) "diagnostic effect" (comparison of patients combined with relatives and control group); (2) "genetic risk effect" (comparison of relatives with control group). The main effects of diagnosis and genetic risk on intrinsic connectivity, as well as their interaction effects with the task condition effect, are tested.

[0109] Furthermore, the Bayesian model selection process of the mental disorder brain dysfunction detection system based on dynamic causal modeling in step S8 is as follows:

[0110] Step S81: The basic idea of Bayesian model selection. Bayesian model selection is based on the Bayes theorem and selects the most appropriate model by calculating the posterior probability of the model given the data. The posterior probability of each candidate model is evaluated by comparing its marginal likelihood. The marginal likelihood represents the probability of a certain model given the observed data. Suppose there are multiple models M1, M2,..., M k , each with a corresponding parameter set θ1, θ2,..., θ k . According to the Bayes theorem, when given the data D, the posterior probability P(M i |D) of model M i can be calculated by the following formula: Among them, P(M i |D) is the posterior probability of model M i , P(D|M i ) is the marginal likelihood of model M i , also known as the model evidence. P(M i ) is the prior probability of model M i , and P(D) is the total likelihood of the data, which is used to normalize the calculation of the posterior probability. In Bayesian model selection, the goal is to select the model with the maximum marginal likelihood P(D|M i ) by calculating the marginal likelihood of all candidate models and comparing them to determine which model best explains the data;

[0111] Step S82: Marginal likelihood. The marginal likelihood P(D|M i ) is a key metric that represents the fitness of model M i given the data D. The marginal likelihood can be calculated by integrating over all the parameters of the model:

[0112] P(D|M i ) = ∫P(D|θ i , M i )P(θ i , M i )dθ i , where P(D|θ i , M i ) is the likelihood of the data given the parameter θ i and model M i , and P(θ i , M i ) is the prior distribution of the parameters of model M i . The calculation of the marginal likelihood is usually quite complex, especially in high-dimensional parameter spaces. To simplify the calculation, a common approach is to estimate the marginal likelihood using free energy approximation.

[0113] Step S83: Free energy approximation. Free energy approximation is an approximate method for calculating the marginal likelihood. By introducing a variational distribution q(θ), the marginal likelihood can be approximated by the free energy: logP(D|M i ) ≈ E q [logP(D|θ i , M i )] - KL(q(θ i ) ∥ P(θ i , M i ), where E q [logP(D|θ i , M i )] is the expectation with respect to the parameter θ iThe expectation of the variational distribution q(θ). The KL(q(θ i )) ∥ P(θ i , M i )) is the Kullback-Leibler divergence between the variational distribution q(θ i ) and the true parameter prior P(θ i , M i ). In this way, BMS can provide a more efficient approximation when calculating the model marginal likelihood.

[0114] Step S84: Model selection: Maximum log evidence. The core of Bayesian model selection lies in selecting the model with the maximum log evidence. To compare different models, calculate the log marginal likelihood logP(D|M i ): logP(D|M i ) =

[0115] log(∫P(D|θ i , M i )P(θ i , M i )dθ i ). Then, compare the log evidence of different models and select the model with the maximum log evidence. Generally, the greater the difference in evidence between models, the better the model can explain the data.

[0116] Step S85: Criteria for model comparison. In Bayesian model selection, the difference in log evidence is usually used to judge which model is better. It is generally considered that a difference in log evidence greater than 3 is strong evidence indicating that the model with greater evidence is significantly better than other models. For example: If △logP(D|M1)-logP(D|M2)>3, then model M1 is significantly better than model M2 in data interpretation. This result can be shown as a log selection process, see Figure 4 .

[0117] Step S9: After determining the optimal model, visualize the posterior estimate of its internal connectivity to identify differences between patients, relatives, and the control group. According to the literature, a connectivity difference of 20% or more is considered a significant effect, so as to achieve the purpose of identifying the obstacle edge and presenting the results.

Claims

1. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling, comprising the following steps: Step S1: preprocessing the EEG signals collected by the head meter; Select the electrode of interest, use the reference electrode standardization technology to re-reference the EEG signal, use 1-45Hz bandpass filtering to extract the frequency band of interest and filter out the signal noise component, downsample the signal, segment the signal, perform baseline correction, remove the signal segments containing artifacts, and obtain the event-related potential signal by superimposing and averaging the data segments; Step S2: Establish spatial source prior model: The spatial source prior model incorporates three brain regions: primary auditory cortex, parietal lobe, and superior frontal gyrus; the spatial coordinates of the cortical sources in the above regions are as follows: left (-59, -10, 13) and right (61, -13, 11) primary auditory cortex, left (-37, -48, 68) and right (28, -56, 63) parietal lobe, left (-29, 55, 22) and right (27, 60, 20) superior frontal gyrus; Step S3: cortical activity source modeling; The electrical activity in each brain region is represented by a single equivalent current dipole, which is placed at a predefined position in the cerebral cortex. It is assumed that the electrical activity in this area is caused by the current of the dipole. Each dipole has the following two basic properties: (1) Dipole moment: The dipole moment describes the magnitude and direction of the current. (2) Position: The spatial position of the dipole indicates the coordinates of the actual current source in the brain, which is often determined by the brain anatomical coordinate system. The magnitude of the dipole moment is proportional to the current intensity of the neural activity source, and the direction indicates the direction of current flow. Its mathematical formula is J = pδ(r-r0), where J is the current density, p is the dipole moment, r0 is the position of the dipole, δ is the Dirac delta function, and r represents the electrode position. Step S4: Neural model building: The CMC model is used as a neural model. The CMC model is a microcircuit neural quality model. The CMC model assumes that each neural source contains four cell populations, which have different functions and interaction modes. The electrical activity of each population is described by the change of current, and these currents interact in the cerebral cortex activity to produce an overall neural activity signal. Step S5: Building external connections: The presence or absence of forward, reverse, and intrinsic connections was defined, and 64 different transmission models were constructed, including 63 candidate models covering forward, reverse, and / or intrinsic connection effects, as well as a complete model without external connections; Step S6: electromagnetic model building: The boundary element method modeling process first selected a spherical approximation model with three concentric layers and simulated different head regions through a refined node distribution. Then, 4,000 test dipoles were randomly placed in these models, their potential distribution was calculated, and the data was fitted to a single dipole to evaluate the positioning error. A refined triangular potential approximation method and an isolation problem method were used. In this process, the setting of the electrode position relative to the outer layer nodes had an important influence on the error evaluation, and the accuracy of the model was further improved through virtual refinement. Step S7: Model inversion of dynamic causal model: Through the above modeling steps, the EEG data preprocessed in step S1 is input into the defined dynamic causal model to perform model inversion; Step S8: Bayesian model selection: We used a Bayesian approach to find the model with the largest log-evidence among all models tested, assuming equal prior probabilities for all models, to determine which model better explained the data for each study effect and interaction effect; Step S9: Brain dysfunction detection: After determining the optimal model, the differences between patients, relatives and control groups can be identified by statistical comparison and visualization of the posterior estimate of the model's intrinsic connectivity; then, the connectivity difference exceeding the threshold is defined as a significant effect, thereby achieving the purpose of identifying the connection edges of brain dysfunction and presenting the results of abnormal regions.

2. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling as claimed in claim 1, characterized in that: The step S1 specifically includes: Step S11: 32 channels are selected from the EEG signal acquisition channels for subsequent analysis, including the following channels: Fp1, AF3, F3, F7, FC5, FC1, C3, T7, CP5, CP1, P3, P7, PO3, O1, Oz, Pz, Fp2, AF4, F2, F4, F8, FC6, FC2, Cz, C4, T8, CP6, CP2, P4, P8, PO4, O2; this channel is defined according to the international 10-20 standard lead system. The EEG signal is re-referenced, and the reference electrode standardization technology is used to remove the influence of the reference electrode and enhance the data quality; a 1-45Hz bandpass filter is used to extract the frequency band of interest and filter out the signal noise component, and the signal is downsampled; each segment of EEG data is segmented and the time interval [-200, 800] ms is selected, where 0ms represents the start time of the stimulus. Step S12: Baseline correction and artifact removal: perform baseline correction on the EEG signal, select the time interval of [-200,0]ms as the baseline, and perform artifact removal processing; use a threshold of ±75μV for artifact removal to ensure that artifact signals caused by factors such as eye movements and muscle activities are removed; further clean up the processed EEG signal to ensure that the data quality meets the requirements of subsequent analysis; Step S13: Obtain P300 data: Calculate the group average ERP for each participant by averaging the trials of the target stimulus and the standard stimulus, and then estimate the P300 waveform of all participants. The amplitude of P300 is obtained by extracting the average amplitude of the ±20 millisecond time window around the maximum positive peak in the time window of 300 to 600 milliseconds after the target stimulus.

3. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling as claimed in claim 1, characterized in that: The neural model of step S4 is constructed as follows: Step S41: Modeling the neural cell population into four cell populations including: superficial pyramidal cells: superficial pyramidal cells are important excitatory neurons in the cortex, responsible for transmitting information from the upper cortex; deep pyramidal cells: deep pyramidal cells are located in the deeper layers of the cortex and participate in long-distance connections between cortices; spinous astrocytes: spinous astrocytes are another type of excitatory neurons in the cortex, located in the middle layer of the cortex; inhibitory interneurons: inhibitory interneurons play an inhibitory role in the neural network, and they generate inhibitory signals through connections with other neurons; Step S42: The activity of the neuron is described by the change of membrane potential; the change of membrane potential is affected by the input current and the interconnection between neurons; assuming there is a neuron population, the change of its membrane potential is expressed as: Among them, V(t) is the membrane potential, τ is the time constant of the membrane potential, which reflects the response speed of neurons to stimulation, and I ext (t) is the external input current, I syn (t) is the synaptic input current from other neuronal populations; Step S43: Synaptic input current comes from the connections between different neural populations, which are excitatory or inhibitory; the synaptic current is expressed by the following formula: I syn (t)=∑ i w i ·g i (V(t)-V rest )·θ(t), where w i is the connection strength or synaptic weight, which represents the connection strength between neural groups, g i is the conductance of each neuron population, reflecting the strength of their response to input, V rest is the resting potential, θ(t) is a step function indicating whether the neuron is activated; Step S44: In the CMC model, synaptic inputs include both excitatory inputs and inhibitory inputs; excitatory inputs increase the membrane potential of the neuron, while inhibitory inputs decrease its membrane potential; excitatory synaptic inputs are represented by: I exc (t) = w exc ·(V(t)-V rest ), the inhibitory synaptic input can be expressed as: I inh (t) = w inh ·(V(t)-V rest ), where w exc and w inh denote the connection weights of excitatory and inhibitory synapses, respectively; Step S45: In the neuron population, the synaptic current of the reverse connection is expressed as: I rev (t) = w rev ·(V deep (t)-V shallow (t)), where V deep (t) and V shallow (t) represents the membrane potential of deep and superficial pyramidal cells, respectively, w rev It represents the synaptic weight of the reverse connection; the self-inhibitory current calculation is expressed as: I self (t) = -w self g(V(t)-V rest ), w self represents the autoinhibitory synaptic weight, g(.) represents the autoinhibitory synaptic transfer function, which describes the decay or delay effect of the synaptic current over time; Step S46: The output of the neuron population is determined by the activity of the neurons, and the output current is calculated by the membrane potential or other electrophysiological signals; the output of the population is expressed by the following formula: Where τ0 is the time constant of the output current.

4. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling as claimed in claim 1, characterized in that: The specific method of step 5 is: By defining the presence or absence of forward connection, reverse connection and intrinsic connection, three forward models are constructed: The f1 model is defined as l-SP to l-SF and r-SP to r-SF forward connections; The f2 model is defined as l-A1 to l-SP and r-A1 to r-SP forward connections; The f3 model is defined as the forward connection of l-A1 flowing through l-SP to l-SF and r-A1 flowing through r-SP to r-SF; 3 backward models: The b1 model is defined as l-SF to l-SP and r-SF to r-SP backward connections; The b2 model was defined as l-SP to l-A1 and r-SP to r-A1 backward connections; The b3 model is defined as a backward connection with l-SF flowing through l-SP to l-A1 and r-SF flowing through r-SP to r-A1; 7 Intrinsic Connection Models: The i1 model is defined as the intrinsic connection between the l-SF and r-SF brain regions; The i2 model is defined as the intrinsic connection between the two brain regions of l-SP and r-SP; The i3 model is defined as the intrinsic connection between the two brain regions l-A1 and r-A1; The i4 model is defined as the intrinsic connectivity of the four brain regions: l-SF, r-SF, l-SP, and r-SP; The i5 model is defined as the intrinsic connections of four brain regions: l-SP, r-SP, l-A1, and r-A1; The i6 model is defined as the intrinsic connections of four brain regions: l-SF, r-SF, l-A1, and r-A1; The i7 model is defined as the intrinsic connections of six brain regions: l-SF, r-SF, l-SP, r-SP, l-A1, and r-A1; 1. Disconnected model n is defined as all brain regions are independent and unconnected; By combining different forward, backward and internal connection models, 3×3×7=63 different transmission models were constructed. In addition, a complete model without external connection n was set up, and a total of 64 different models were constructed. Among them, l represents the left brain area, r represents the right brain area, SP represents the parietal lobe, SF represents the superior frontal gyrus, and A1 represents the primary auditory cortex.

5. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling as claimed in claim 1, characterized in that: The model inversion process based on dynamic causal modeling in step S7 is as follows: Step S71: The core calculation of the Bayesian model inversion relies on the Bayesian formula. During the model inversion process, the posterior probability distribution p(θ|D) and the Bayesian factor are calculated. Suppose there are two models M1 and M2, then the Bayesian factor β 12 Used to measure the advantages of M1 over M2: Where p(D|M) is the marginal likelihood of the data given the model M; Step S72: Dynamic causal modeling was performed at the group level by creating a 2×3 factorial design unit, which included two levels of "task condition" and three levels of "group". The two levels of "task condition" were standard stimulus and target stimulus, and the three levels of "group" were patients, relatives, and control groups. The group effects studied included: (1) "diagnosis effect": patients were compared with relatives and controls; (2) "genetic risk effect": relatives were compared with controls; the main effects of diagnosis and genetic risk on intrinsic connectivity were tested, as well as their interactions with task condition effects.

6. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling as claimed in claim 1, characterized in that: The Bayesian model selection process of step S8 is as follows: Step S81: Suppose there are multiple models M1, M2, ..., M k , each model has a corresponding parameter set γ1,γ2,...,γ k ; According to Bayes' theorem, given data D, model M i The posterior probability P(M i |D) is calculated by the following formula: Among them, P(M i |D) is model M i The posterior probability, P(D|M i ) is the model M i The marginal likelihood of P(M i )Model M i The prior probability of the data, P(D) is the total likelihood of the data, which is used to normalize the calculation of the posterior probability; in Bayesian model selection, the goal is to select the model with the maximum marginal likelihood P(D|M i ) model, by calculating the marginal likelihood of all candidate models and comparing them to determine which model best explains the data; Step S82: Calculate the marginal likelihood P(D|M i ), the marginal likelihood represents the model M i Fitness under given data D; The marginal likelihood is calculated by integrating over all parameters of the model: P(D|M i )=∫P(D|θ i ,M i )P(θ i ,M i )dθ i , where P(D|θ i ,M i ) data under given parameter θ i and Model M i The likelihood under i ,M i )Model M i The parameter prior distribution of ; Step S83: Calculate free energy approximation; Introduce a variational distribution q(θ) and use free energy to approximate the marginal likelihood: logP(D|M i )=E q [logP(D|θ i ,M i )]-KL(q(θ i )∥P(θ i ,M i )), where E q [logP(D|θ i ,M i )] is the parameter θ i The expectation of the variational distribution q(θ) of KL(q(θ i )∥P(θ i ,M i )) is the variational distribution q(θ i ) and the true parameter prior P(θ i ,M i ) Step S84: model selection; The model with the largest log-evidence is selected and the log marginal likelihood logP(D|M i ):logP(D|M i )=log(∫P(D|θ i ,M i )P(θ i ,M i )dθ i ); then, compare the log-evidence of different models and select the model with the largest log-evidence; Step S85: establishing the criteria for model comparison; A difference in log-evidence greater than 3 is strong evidence that the model with greater evidence is significantly better than the other models.

7. A method for detecting brain dysfunction in mental illness based on dynamic causal modeling as claimed in claim 1, characterized in that: In step S9, a connectivity difference of 20% or more is considered a significant effect, thereby achieving the purpose of identifying the barrier edge and presenting the results.

Images