Brain-computer interface regulation-oriented personalized brain function network construction and evaluation method

Abstract

The invention belongs to a network construction evaluation method, and aims to solve the technical problems that an existing brain function network construction method is lack of personalized partitions, insufficient in nonlinear neural association description, weak in cross-scene generalization ability and difficult to support precision and large-scale clinical application of brain-computer interface neural regulation. According to the brain-computer interface regulation-oriented personalized brain function network construction and evaluation method provided by the invention, a personalized brain function network is constructed based on neural activity mode characterization, and personalized brain function network construction of heterogeneous fMRI data is realized by fusing neurodynamics knowledge constraint and data-driven modeling; the method can be used as a universal modeling tool for heterogeneous fMRI data, and provides reliable technical support for brain-computer interface nerve regulation and control.

Description

Technical Field

[0001] This application pertains to a network construction and evaluation method, specifically a personalized brain functional network construction and evaluation method for brain-computer interface modulation. Background Technology

[0002] Constructing reliable, personalized brain function networks is a key prerequisite for neuromodulation via brain-computer interfaces. Functional magnetic resonance imaging (fMRI), with its non-invasive advantage of blood oxygen level dependent (BOLD) signals, has become a core technology for characterizing brain region coordination patterns and is widely used in cognitive analysis, disease diagnosis and treatment, and other fields. The existing research has three major limitations: (1) It relies on fixed standard brain region templates, completely ignoring individual anatomical and functional differences, making it difficult to match the functional homogeneity and spatial boundaries of specific data, resulting in insufficient signal consistency within brain regions and directly causing the drift of regulatory targets; (2) It usually uses linear measures such as Pearson correlation and covariance to calculate brain region associations, which cannot capture the nonlinear dependence and dynamic hierarchical relationship of the nervous system, and is prone to causing distortion in functional connectivity calculations; (3) It has a weak ability to adapt to heterogeneous fMRI data, poor stability when facing different populations, acquisition strategies or task paradigms, and most methods rely on population-level statistical priors, which cannot output reliable results in scenarios with a small number of or a single subject data (such as rare disease diagnosis and treatment), which seriously restricts the large-scale clinical application of brain-computer interface neuromodulation.

[0003] To alleviate the aforementioned technical deficiencies, targeted improvement research has been conducted in related fields. Chinese invention patent application CN119741270A discloses a method, apparatus, medium, and device for constructing a functional brain network. It constructs an initial network by extracting time-series features of brain regions using resting-state fMRI, employs an improved Graph Isomorphism Network (GIN) for spatial convolution to fuse information from adjacent brain regions, and uses differentiable graph pooling to cluster brain regions, eliminating disease-irrelevant regions and aggregating functionally similar regions, ultimately forming a coarse-grained brain functional network. This approach effectively solves the problems of complex relationships and redundant information contained in traditional brain functional networks due to overly fine brain region division. However, it still fails to overcome the limitations of standardized partitioning, cannot achieve personalized partitioning construction adapted to heterogeneous datasets, and does not optimize for nonlinear correlations in neural activity, resulting in insufficient cross-scenario generalization ability. Chinese invention patent CN113314216A discloses a method, device, electronic device, and readable storage medium for constructing functional brain networks. It automatically divides functional communities using a community detection algorithm, rather than relying on researchers' prior knowledge. Then, within a sliding time window, it applies a Hilbert transform to the community signals to obtain phase signals. It uses a Kōbun model to fit the phase signals to calculate the coupling matrix between communities, extracts the eigenvectors corresponding to the largest eigenvalues, and obtains a brain functional network connectivity model through K-means clustering. This solves the problem of traditional brain region division relying on prior knowledge potentially missing key areas and generating spurious correlations. Furthermore, it introduces a dynamic model to improve the rationality of correlation calculations. However, it still lacks in personalized adaptability, failing to fully consider the differences in neural activity patterns among different populations and under different acquisition conditions. It also has insufficient adaptability to scenarios with limited subject data, making it difficult to support the personalized, multi-scenario application needs of brain-computer interface modulation. In addition, neither of the above two schemes achieves end-to-end mapping from heterogeneous fMRI input to personalized functional networks, failing to simultaneously solve the three core bottlenecks of region mismatch, distorted correlation measurement, and insufficient cross-domain generalization.

[0004] In summary, although existing technologies have optimized traditional brain functional network construction methods in some aspects, they still fail to fully solve the key problems of lack of personalized partitioning, insufficient characterization of nonlinear associations, and weak cross-scenario generalization ability. These defects directly lead to insufficient physiological rationality, individual specificity, and generalization ability of the constructed brain functional networks, making it difficult to stably support the precise target localization, complex neural activity representation, and multi-scenario personalized modeling needs of brain-computer interface neural modulation. Summary of the Invention

[0005] This application addresses the technical problems of existing brain functional network construction methods, such as lack of personalized partitioning, insufficient characterization of nonlinear neural connections, and weak cross-scenario generalization ability, which make it difficult to support the precise and large-scale clinical application of brain-computer interface neural modulation. It provides a personalized brain functional network construction and evaluation method for brain-computer interface modulation.

[0006] To achieve the above objectives, this application adopts the following technical solution: A method for constructing and evaluating personalized brain functional networks for brain-computer interface modulation includes: Based on a pre-defined neurodynamic equation, combined with time-series data from functional magnetic resonance imaging of specific subjects, prior neurodynamic knowledge is injected into the personalized neural activity pattern representation of specific subjects. The personalized neural activity pattern representation injected with prior knowledge of neurodynamics is generalized and optimized for physiological rationality to obtain the optimized personalized neural activity pattern representation. Brain functional partitioning based on optimized personalized neural activity pattern representation; The brain functional zoning results were interpreted and quantitatively analyzed to obtain brain functional synergy quantitative indicators. Based on the brain functional zoning results, the time dimension is extended to the quantitative indicators of brain functional synergy to obtain a personalized dynamic brain functional network. We constructed a digital twin brain simulation environment corresponding to the brain of a specific subject, and evaluated the performance of personalized dynamic brain functional networks by applying virtual stimuli.

[0007] Furthermore, the method of injecting prior knowledge of neurodynamics into the personalized neural activity pattern representation of a specific subject includes: The optimal set of parameters for the personalized neurodynamic equation is obtained by fitting time-series data from functional magnetic resonance imaging of specific subjects. Substitute the parameter set of the optimal personalized neurodynamic equation into the preset neurodynamic equation to determine the neural activity function corresponding to a specific subject; A physical constraint loss term containing the neural activity function corresponding to a specific subject is constructed, and combined with the data fitting loss term to form a composite loss function; By minimizing the composite loss function, the learning process of personalized neural activity pattern representation is optimized, resulting in a corrected neural activity pattern representation, thus injecting prior neurodynamic knowledge into the personalized neural activity pattern representation of a specific subject.

[0008] Furthermore, the method for generalizing and physiologically rationalizing the personalized neural activity pattern representation infused with prior knowledge of neurodynamics includes: By projecting personalized neural activity pattern representations infused with prior knowledge of neurodynamics onto a low-dimensional latent space, a low-dimensional latent space representation is obtained. A diffusion model is used to capture higher-order statistical regularities and spatiotemporal dependencies in low-dimensional latent space representations, thereby obtaining personalized neural activity pattern representations in low-dimensional latent space. During the training of the diffusion model, neural dynamics equations are used as physical constraints and added to the loss function of the diffusion model training. By using a voxel space decoder, the personalized neural activity pattern representation in the low-dimensional latent space is mapped back to the original voxel space, resulting in an optimized personalized neural activity pattern representation.

[0009] Furthermore, by incorporating the neurodynamic equations as physical constraints into the loss function for training the diffusion model, the following results were obtained:

[0010] in, The loss function representing the fine-tuning of physical constraints serves as the total loss of the diffusion model. Represents the physics-driven term. Indicates a data-driven item. The regularization coefficient represents the fit between the balanced data and the physical constraints. Represents a voxel space decoder. This represents the pre-trained base representation model. This represents the initial neural activity representation in the low-dimensional latent space. Representing the equation of neural activity, This represents the raw functional magnetic resonance imaging (fMRI) data input.

[0011] Furthermore, the method for functional brain partitioning based on optimized personalized neural activity pattern representation includes: Determine brain functional partitioning targets based on maximizing homogeneity in the representation of optimized personalized neural activity patterns; Based on the brain functional partitioning objectives, an energy function is constructed to transform the biophysical constraints of brain functional partitioning into mathematical optimization objectives; By combining the energy function and the self-supervised contrastive learning framework, a personalized representation contrastive loss is established. The optimal brain functional partitioning results are obtained by calculating the loss based on personalized representations.

[0012] Furthermore, the personalized representation contrast loss is:

[0013] in, For personalized characterization of contrast loss, Indicates the current voxel Personalized neural activity pattern representation, Indicates the current voxel Personalized neural activity pattern representation, Indicates the relationship with the current voxel Voxel representation of positive samples belonging to the same predictive functional region. This represents a temperature coefficient used to adjust the degree of attention given to difficult samples in contrastive learning. Indicates the relationship with the current voxel The set of negative sample voxels belonging to different prediction functional regions. Represents a set of voxels. express The elements in.

[0014] Furthermore, the method for interpreting and quantifying brain functional characteristics based on the brain functional partitioning results to obtain brain functional synergy quantitative indicators includes: Record the brain functional partitioning results as multiple brain regions; perform the following steps: (1) Based on the brain functional partitioning results, the average value of the personalized neural activity pattern representation of all voxels in each brain region is taken to obtain the regional representation of each brain region. (2) Input the regional representations of all brain regions into the pre-trained collaborative computing model to capture the hierarchical interaction relationships between different brain regions and generate the initial brain functional network. (3) Input the initial brain function network into the downstream task classifier and output the brain function prediction results; wherein, the brain function network is the carrier of brain function synergistic quantitative indicators; (4) Connect steps (1) to (3) into a whole model framework. Through the loss function of the downstream task classifier, update the parameters of the whole model framework synchronously, optimize the brain function network, and obtain the optimal brain function network, that is, obtain the brain function synergistic quantitative index.

[0015] Furthermore, the method for obtaining a personalized dynamic brain function network by extending the time dimension to the quantitative indicators of brain function synergy based on brain functional partitioning results includes: Based on the statistical distribution drift of real-time monitoring signals from functional magnetic resonance imaging (fMRI), the boundaries of local time windows are defined in a personalized manner, and the time series data of fMRI is decomposed into multiple short time segments. For each short segment, a personalized neural activity pattern representation is constructed based on the brain functional partitioning results. Then, nonlinear dependency analysis technology is used to quantify the temporal association between the current short segment and historical short segments in terms of neural activity pattern representation, and a time-varying co-weight matrix is ​​constructed as dynamic prior information. Reuse collaborative computing models to construct transient brain function networks corresponding to short time segments; Based on the time-varying collaborative weight matrix, the transient brain function network sequence, which is composed of transient brain function networks corresponding to each short segment, is subjected to time-weighted fusion to obtain a personalized dynamic brain function network.

[0016] Furthermore, the method for evaluating the performance of personalized dynamic brain functional networks includes: A digital twin brain simulation environment corresponding to the brain of a specific subject is constructed, virtual stimulation is applied, and the rationality analysis of the personalized dynamic brain function network is conducted based on the overlap between the set of regulatory targets identified by the personalized dynamic brain function network and the set of empirical targets. The neuromodulation recovery rate of specific subjects is evaluated based on a pre-set disease diagnosis model. The personalized dynamic brain function network is used as the initial modulation object. The personalized dynamic brain function network is virtually modulated through virtual stimulation. The modulation effect is evaluated and the recovery rate is calculated. The personalized dynamic brain function network after virtual modulation was statistically compared with the brain function network of the healthy control group to verify whether the brain activity pattern after virtual modulation met the statistical regression level.

[0017] Furthermore, the method for calculating the overlap includes:

[0018] in, Indicates the degree of overlap. This represents the set of regulatory targets identified by a personalized dynamic brain function network. This represents the set of empirical targets.

[0019] Compared with the prior art, this application has the following beneficial effects: This application proposes a personalized brain functional network construction and evaluation method for brain-computer interface modulation. Existing brain functional network construction methods generally rely on linear measures such as Pearson correlation, partial correlation, or covariance to calculate functional connectivity between brain regions, which cannot effectively capture the complex hierarchical dependencies widely present in neural activity, resulting in structural biases in the constructed networks when characterizing real brain functional synergies. This application introduces pre-trained co-computation and combines personalized neural activity pattern representation constrained by neurodynamic equations to co-compute brain region interaction associations, significantly improving the personalized expression ability and physiological rationality of functional connectivity for complex neural interactions. In addition, existing technologies usually divide the brain based on fixed maps or simple statistical methods, ignoring the uniqueness of neural activity pattern representations in different scenarios. When facing different populations, different acquisition strategies, or personalized neural modulation scenarios, their output results are often distorted. This application introduces an adaptive partitioning method based on personalized neural activity pattern representations to dynamically calculate brain partitions, significantly improving the consistency of activities within partitions and adaptability to downstream tasks. Furthermore, existing methods heavily rely on specific preprocessing procedures and task paradigms, limiting their generalization ability in constructing brain functional networks when applied to different acquisition protocols and diseases. This application, through a fine-tuning mechanism, can flexibly adapt to new scenarios while preserving large-scale prior knowledge. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a personalized brain functional network construction and evaluation method for brain-computer interface modulation, as described in this application. Figure 2 This is another flowchart illustrating the personalized brain functional network construction and evaluation method for brain-computer interface modulation proposed in this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Existing methods for constructing brain functional networks typically rely on predefined standard brain atlases or statistical methods to segment brain regions. However, due to significant differences in the representation of neural activity patterns among different types of fMRI datasets, such segmentation strategies often fail to individually reflect the functional consistency and spatial boundaries of the data, resulting in large performance fluctuations when applied across datasets and leading to inaccurate localization of brain-computer interface (BCI) modulation targets. Currently, most methods for calculating brain region associations are based on linear metrics such as Pearson correlation coefficient, partial correlation, or covariance, which limits their ability to capture complex nonlinear interactions in the nervous system. These traditional methods struggle to accurately describe the nonlinear relationships and hierarchical dependencies between brain regions, potentially leading to erroneous calculations of functional connectivity, affecting the accuracy of network topology characteristics and their performance in subsequent analyses, and making it difficult to characterize the complex neural activities regulated by BCIs in different scenarios. Furthermore, current methods for constructing brain functional networks have several shortcomings in terms of cross-domain generalization: First, brain-computer interface (BCI) modulation faces various scenarios and fMRI data with different preprocessing strategies, and existing construction processes lack generalization ability; second, most existing frameworks lack effective support for data offset inputs, and most processes rely on population-level statistical priors, which cannot stably output reliable networks when facing application scenarios such as rare diseases or personalized neuromodulation where only a small amount of data or even a single subject can be obtained. This makes it difficult to support personalized modeling of BCI modulation across multiple scenarios.

[0024] To overcome the aforementioned technical problems, this application proposes a method for constructing and evaluating personalized brain functional networks for brain-computer interface modulation. The following detailed description, in conjunction with embodiments and accompanying drawings, further illustrates this application.

[0025] like Figure 1 The diagram shown is a schematic representation of a personalized brain functional network construction and evaluation method for brain-computer interface modulation, which may include: S101, based on a pre-defined neurodynamic equation and combined with functional magnetic resonance imaging time-series data of specific subjects, injects prior neurodynamic knowledge into the personalized neural activity pattern representation of specific subjects.

[0026] It is important to note that brain neural activity follows specific biophysical laws, and neurodynamic equations are the mathematical descriptions of these laws. Traditional personalized neural activity pattern representations often rely on data-driven approaches, which can easily lead to a disconnect from physiological mechanisms. This application incorporates prior neurodynamic knowledge as a constraint into the neural activity pattern representations of specific subjects, ensuring that the neural activity pattern representations not only possess individual specificity but also closely align with the actual neural activity mechanisms of the brain. This solves the problem of traditional personalized representations being data-driven and detached from physiological reality, giving the neural activity pattern representations neurodynamic physiological rationality. This provides a biologically sound foundation for subsequent optimization of neural activity pattern representations and brain functional partitioning, reduces physiologically meaningless computational results, and preserves the individual specificity of specific subjects.

[0027] S102, generalize and optimize the physiological rationality of the personalized neural activity pattern representation injected with prior knowledge of neurodynamics to obtain the optimized personalized neural activity pattern representation.

[0028] It should be noted that personalized neural activity pattern representations infused with prior neurodynamic knowledge may suffer from insufficient generalization ability or local physiological plausibility biases. This application improves the scenario adaptability of personalized neural activity pattern representations through generalization optimization, and further corrects the deviations between personalized neural activity pattern representations and neurodynamic laws by combining physiological plausibility optimization, ultimately obtaining personalized neural activity pattern representations that are both universally adaptable and physiologically reliable. This enhances the generalization ability of personalized neural activity pattern representations, enabling them to adapt to different neural activity scenarios, further strengthening the neurodynamic and physiological plausibility of personalized neural activity pattern representations, providing a more reliable foundation for subsequent brain function partitioning and synergistic quantification, while reducing the accumulation of errors in subsequent steps and improving stability.

[0029] S103, based on optimized personalized neural activity pattern representation, divides the brain into functional zones.

[0030] The neural activity patterns of different brain regions are specific, and brain regions with similar functions exhibit high homogeneity in their neural activity patterns. This application, based on optimized personalized neural activity pattern representations, divides the brain into multiple functionally independent regions. Each region exhibits high homogeneity in neural activity patterns, while significant functional differences exist between regions. Furthermore, the zoning results closely align with the individual characteristics and neurodynamic patterns of specific subjects. The resulting functional zoning combines individual specificity with physiological rationality, avoiding the problem of traditional general zoning ignoring individual differences. It clarifies the spatial distribution of brain functions, providing clear spatial units for subsequent interpretation and co-quantification of brain functional features. In addition, the zoning results, based on optimized representations, improve the accuracy of functional region division, laying a spatial foundation for the subsequent construction of dynamic brain functional networks.

[0031] S104. The brain functional region results are interpreted and quantitatively analyzed to obtain brain functional synergy quantitative indicators.

[0032] It should be noted that normal brain function depends on the coordinated interaction between different functional areas, and abnormalities in functional coordination are often associated with brain diseases or functional impairments. Based on the functional area division results of step S103, this application interprets the core functions of each functional area and quantitatively analyzes the coordinated interaction relationships between different functional areas, transforming abstract coordinated relationships into calculable and comparable quantitative indicators. Interpreting the characteristics of functional areas makes the quantitative indicators of brain functional coordination more physiologically meaningful, facilitating the identification of subsequent regulatory targets. Furthermore, the quantitative indicators of brain functional coordination possess individual specificity, accurately reflecting the brain functional coordination state of a specific subject, providing support for personalized regulation.

[0033] S105, based on the brain functional partitioning results, extends the time dimension to the quantitative indicators of brain functional synergy, resulting in a personalized dynamic brain functional network.

[0034] Brain neural activity is non-stationary, meaning that the state of functional coordination changes dynamically over time, a dynamic characteristic that traditional static brain functional networks cannot capture. Based on the functional partitioning results of step S103, this paper adds a temporal dimension to the static brain functional coordination quantification indicators of step S104. By analyzing the temporal evolution trajectory of neural activity, a personalized dynamic brain functional network is constructed that reflects changes in the coordinated state at different time stages, achieving a spatiotemporal full-dimensional characterization of brain functional coordination. This overcomes the drawbacks of time averaging in traditional static brain functional networks, accurately capturing the dynamic evolutionary patterns of brain functional coordination. The personalized dynamic brain functional network combines the spatial dimension of functional partitioning with the temporal dimension of dynamic evolution, more closely reflecting the actual state of brain neural activity. This provides a more precise dynamic target basis for subsequent virtual regulation and evaluation, improving the timeliness and accuracy of regulation plans.

[0035] S106: Construct a digital twin brain simulation environment corresponding to the brain of a specific subject, and evaluate the performance of personalized dynamic brain functional networks by applying virtual stimuli.

[0036] It's important to note that a digital twin brain simulation environment is a digital replica of the real brain. It simulates neural modulation processes in a virtual environment, avoiding the risks associated with real-world modulation. By constructing a digital twin brain simulation environment matched to specific subjects and embedding a personalized dynamic brain functional network within it, and applying virtual stimuli to simulate clinical modulation scenarios, the reliability and modulatory potential of the dynamic brain functional network can be evaluated, providing a basis for decision-making in real brain-computer interface modulation. Completing performance evaluation before implementing real neural modulation significantly reduces the risks and costs of clinical modulation. Furthermore, multi-dimensional evaluation ensures the reliability and modulatory potential of the personalized dynamic brain functional network, improving the success rate of subsequent real-world modulation and providing a quantitative basis for precise brain-computer interface modulation, thus propelling the technology from the laboratory to clinical application.

[0037] This application combines prior constraints on neural activity pattern representation driven by neurodynamic knowledge, pre-trained model-driven neural activity pattern constraint-guided representation, and adaptive functional partitioning and functional co-computation guided by neural activity representation to achieve end-to-end mapping from heterogeneous fMRI input to personalized dynamic brain functional networks. This method not only improves the consistency and homogeneity of functional partitioning but also effectively captures complex nonlinear relationships, enhancing the cross-domain generalization ability of the process. It can effectively improve the physiological rationality, personalization, and generalization ability of brain functional network construction, providing important support for large-scale clinical applications such as brain-computer interface neuromodulation.

[0038] like Figure 2 The diagram shown is a second flowchart illustrating the personalized brain functional network construction and evaluation method for brain-computer interface modulation proposed in this application, which may include: The method proposed in this application can be implemented based on several main modules: a dynamic prior constraint module for personalized neural activity pattern representation, a personalized neural activity pattern representation module driven by a pre-trained model, an adaptive functional zoning module guided by personalized neural activity pattern representation, a functional co-computation module guided by personalized neural activity pattern representation, and a personalized dynamic brain function network assessment module based on virtual neural modulation.

[0039] S201, Dynamic Prior Constraints for Personalized Neural Activity Pattern Representation (Dynamic Prior Constraints Module for Personalized Neural Activity Pattern Representation).

[0040] A data-driven framework constrained by neurodynamic knowledge is introduced to provide prior neurodynamic knowledge for personalized neural activity pattern representation. Neurodynamic equation parameters are learned from fMRI data, and fine-tuning functions are constructed using neurodynamic knowledge during the neural activity pattern representation learning process. This approach ensures that the represented neural activity patterns not only conform to the data distribution but also to the brain's intrinsic evolutionary mechanisms, thereby addressing the overfitting and uninterpretable issues of purely data-driven models with small sample sizes.

[0041] The constraints of the existing neurodynamic equations are:

[0042] in, This represents the input functional magnetic resonance imaging data. This represents a personalized neural activity pattern representation to be learned. This represents the mapping that reconstructs back to the original data space. This represents the set of parameters in the neurodynamic equations. This represents the pre-defined neurodynamic equation. Represents the physics-driven term. Indicates a data-driven item. This represents the regularization coefficient for balancing data fitting with physical constraints.

[0043] Specifically: (1) Parameter fitting of neurodynamic equations.

[0044] For fMRI time-series data of specific subjects, key free parameters in the neurodynamic equation are fitted. Input data are... ,in, For the number of time points, This is a set of voxels. By solving the neural dynamics equations, a personalized set of physical parameters that can characterize the specific neural activity of this individual is obtained:

[0045] in, This represents the set of parameters for the optimal personalized neurodynamic equation output.

[0046] (2) Constraint construction based on neurodynamic equations.

[0047] A fine-tuning constraint is constructed that integrates data reconstruction error and neurodynamic consistency error. This constraint ensures that the neurodynamic equations conform to both the data distribution and prior neurodynamic knowledge. The physical constraint term is implemented by calculating the error between the personalized neural activity pattern representation and the neurodynamic equations. By minimizing this combined loss, noise that violates neural mechanisms can be effectively eliminated during the representation stage, improving the personalized performance of neural activity pattern representations.

[0048] Taking the wave equation as an example, the following differential equation is solved to obtain the neural activity function. :

[0049] in, Indicates the spatial location of fMRI. Indicates the time points of activity in fMRI. This represents the damping term hyperparameter. Represents spatial scale hyperparameters. Indicates external stimuli, This represents the Laplace operator.

[0050] It should be noted that neural activity functions The optimal parameters serve as the basis for the personalized neural activity pattern representation module driven by the subsequent pre-trained model. .

[0051] The constrained neurodynamic equation is then: The final constraints can be expressed as:

[0052] This application introduces prior knowledge of neurodynamics into a data-driven framework, providing dynamic prior knowledge for personalized neural activity pattern representation. It uses neurodynamic equations to fit fMRI data, and leverages neurodynamic laws to regulate the learning process of neural activity pattern representation, ensuring that the represented neural activity patterns not only conform to the data distribution but also to the brain's intrinsic evolutionary mechanisms. This addresses the overfitting and uninterpretable issues of purely data-driven models with small samples. For fMRI time-series data of specific subjects, key free parameters in the neurodynamic equations are fitted to obtain a set of exclusive physical parameters that accurately characterize the individual's specific neural activity, providing a personalized benchmark for subsequent prior constraints. Furthermore, a composite loss function integrating data reconstruction error and neurodynamic consistency error is constructed. This composite loss function forces the model to find the optimal solution during training that both reconstructs the observed data and satisfies the laws of neurodynamics, effectively eliminating noise that violates neural mechanisms during the representation stage and improving the personalized performance of the representation.

[0053] S202, Pre-trained model-driven personalized neural activity pattern representation (pre-trained model-driven personalized neural activity pattern representation module).

[0054] This embodiment utilizes a pre-trained basic representation model (diffusion model) on a large-scale functional brain imaging dataset to address the problem of insufficient representation of neural activity patterns caused by scene changes. First, high-dimensional fMRI data is compressed into a low-dimensional latent space using a low-dimensional latent space encoder (voxel space encoder). Then, the diffusion model extracts neural activity pattern representations rich in high-order statistical regularities during the reverse denoising process, and the model is reconstructed to restore the voxel space neural activity pattern representations. This basic representation model is then fine-tuned with physical constraints using neurodynamic equations to achieve personalized neural activity pattern representations.

[0055] The overall processing flow described above aims to extract generalized representations from neural activity data and adapt them to individual physical laws. Its core mapping relationship can be summarized as follows:

[0056] in, This represents the spatial neural activity pattern of voxels. Represents a voxel space decoder. This represents a voxel space encoder. This represents the pre-trained base representation model. The parameters represent the latent space characterization process.

[0057] Specifically: (1) Personalized latent space mapping of neural activity.

[0058] To address the high dimensionality and strong spatiotemporal coupling of fMRI data, a pre-trained low-dimensional latent space encoder is constructed. The raw fMRI data is projected into the low-dimensional latent space to overcome the difficulty of direct modeling. The input data format is as follows: , This represents the raw functional magnetic resonance imaging (fMRI) data as input. After processing by a low-dimensional latent space encoder pre-trained on large-scale data, low-dimensional latent space representations that can compactly express brain activity patterns are extracted, achieving dimensionality reduction and preliminary representation of the data.

[0059] The calculation formula for this latent space mapping process is as follows:

[0060] in, This represents the initial neural activity representation in the low-dimensional latent space. This represents a voxel space encoder.

[0061] (2) Personalized neural activity pattern representation in low-dimensional latent space driven by pre-trained models.

[0062] Preliminary neural activity representation in low-dimensional latent space Based on this, utilize the pre-trained basic representation model Further characterization is then performed. By using a pre-trained basic representation model in a low-dimensional latent space, the diffusion model can capture the hidden high-order statistical regularities and spatiotemporal dependencies in the data, constructing a personalized neural activity pattern representation in the low-dimensional latent space. Subsequently, a voxel space decoder is used... (Reconstructing the model) maps these deep features back to the original voxel space, constructing a model to guide subsequent adaptive partitioning and collaborative relationship computation.

[0063] In the low-dimensional latent space, the representation process of the pre-trained basic representation model is as follows: =

[0064] in, This represents the low-dimensional latent space neural activity pattern representation obtained after multi-step reasoning. The following are the low-dimensional latent space neural activity patterns represented in each step of the multi-step reasoning process.

[0065] The calculation formula for the complete neural activity pattern representation process is as follows:

[0066] (3) Fine-tuning of PINN (Physics-Informed Neural Networks) based on the personalized representation of the neural dynamics equation.

[0067] To ensure that personalized neural activity pattern representations not only conform to statistical data patterns but also to the neurobiological mechanisms of the brain, this embodiment introduces neurodynamic equations as prior knowledge. A loss function incorporating these neurodynamic equations is constructed to measure the representation of neural activity patterns in voxel space. The degree to which the solutions to the neurodynamic equations are satisfied, for example, the neural activity function. By minimizing this loss function, the parameters in the pre-trained base representation model are fine-tuned, ensuring that the final representation of neural activity patterns retains individual specificity while adhering to the dynamic constraints of neural signals. In practical applications, fine-tuning can be performed using the LoRA technique.

[0068] The loss function for fine-tuning the physical constraints is calculated using the following formula:

[0069] in, The loss function representing the fine-tuning of physical constraints serves as the total loss of the diffusion model. Represents the physics-driven term. Indicates a data-driven item. This represents the regularization coefficient for balancing data fitting with physical constraints.

[0070] Taking LoRA as an example, the parameter update process for fine-tuning is as follows:

[0071] in, This represents the low-rank parameter of LoRA. Indicates the learning rate, superscript and They represent the first The result parameters of the first optimization step and the first The optimized parameters are as follows: This represents the Laplace operator.

[0072] This application utilizes a pre-trained basic representation model on a large-scale functional brain imaging dataset to address the problem of insufficient neural activity representation caused by scene changes. It achieves generalized neural activity representation capabilities based on a large amount of functional imaging data, and fine-tunes it using the aforementioned dynamic constraints to realize personalized neural activity pattern representations. A personalized neural activity encoder network projects raw fMRI data into a low-dimensional latent space. Adaptive mapping is achieved through pre-training to address differences in brain activity among individuals. This achieves both preliminary representation of brain activity patterns and efficient dimensionality reduction of the data, further improving the processing speed of the representation. A pre-trained basic representation model is introduced as the representation model for neural activity patterns. Based on the learned cross-group generalized brain activity pattern representation capabilities, information missing in individual data due to low signal-to-noise ratios or acquisition artifacts is supplemented. Personalized neural activity pattern representations that possess both individual uniqueness and statistical robustness are extracted. A Physics-Informed Neural Networks (PINN) architecture is employed to explicitly embed the neurodynamic equations into the fine-tuning process of the pre-trained model. By using the error in the dynamic equations as a penalty term in network training, the weights of the pre-trained model are adjusted accordingly. This process ensures that while the model learns from individual data, the represented neural activity patterns follow the prior knowledge of the neurodynamic equations, achieving a fusion of data-driven and physical-driven approaches.

[0073] S203, Personalized Neural Activity Pattern Representation-Guided Adaptive Functional Partitioning (Personalized Neural Activity Pattern Representation-Guided Adaptive Functional Partitioning Module).

[0074] In this step, functional regions are dynamically constructed based on personalized neural activity pattern representations. Following the principle of maximizing the homogeneity of personalized neural activity pattern representations, a self-supervised learning algorithm is guided by voxel-level personalized neural activity pattern representations to partition functional regions. By defining an energy function that incorporates spatial consistency and homogeneity of personalized representations, the boundaries of functional brain regions are adaptively delineated. This method can identify changes in functional brain regions caused by individual differences.

[0075] The core objective of this step is to find the optimal label configuration vector. The mapping relationship is as follows:

[0076] in, The final result of the personalized functional partitioning 。 This represents the energy function.

[0077] Specifically: (1) Construct the total energy loss of the combined personalized representation, spatial information and gradient information.

[0078] A multidimensional energy function is constructed to transform the biophysical constraints of brain functional partitioning into a mathematical optimization objective. This multidimensional energy function deeply integrates personalized neural activity pattern representation information, the geometric neighborhood structure between voxels, and the rate of change of functional gradients. Specifically, personalized neural activity pattern representation information ensures homogeneity, the geometric neighborhood structure between voxels ensures continuity, and the rate of change of functional gradients ensures boundary accuracy.

[0079] By minimizing this energy function, personalized partitioning of brain function can be achieved.

[0080]

[0081] in, and These represent two voxels at different locations. Defined as , Represents the functional gradient between voxels. Indicates the current voxel Personalized activity pattern representation, Indicates the current voxel Functional area The average functional activity pattern is characterized by This represents the voxel neighborhood relationships in the analyzed brain space. This indicates an indicator function that applies the penalty term when adjacent voxel labels are different. and These represent the hyperparameters that adjust gradient sensitivity and spatial constraint strength, respectively. Indicates the distance between voxels. This indicates energy loss.

[0082] (2) Personalized functional brain region segmentation based on self-supervised learning.

[0083] A self-supervised contrastive learning framework is employed to address the problem of adaptive functional brain region segmentation based on personalized neural activity pattern representations. This is achieved by constructing voxel pairs within the same functional region and voxel pairs across functional regions, and jointly training them using the aforementioned energy function. This process encourages the personalized neural activity pattern representation module, driven by the pre-trained model, to narrow the functional representation distance within the same brain region and widen the functional representation distance between voxels in different brain regions. Ultimately, this results in a personalized whole-brain functional atlas with highly homogeneous activity within each region and significant heterogeneity between different regions.

[0084] This personalized characterization contrast loss The calculation formula is as follows:

[0085] in, This represents the personalized neural activity pattern of the current voxel. Indicates the relationship with the current voxel Voxel representation of positive samples belonging to the same predictive functional region. This represents a temperature coefficient used to adjust the degree of attention given to difficult samples in contrastive learning. Indicates the relationship with the current voxel The set of negative sample voxels belonging to different prediction functional regions. Represents a set of voxels. Indicates the current voxel Personalized neural activity patterns are represented.

[0086] This application dynamically constructs exclusive functional partitions based on personalized neural activity pattern representations. Following the principle of maximizing the homogeneity of functional representations, it adaptively delineates the boundaries of functional brain regions using the personalized neural activity pattern representations constructed by the aforementioned modules. This accurately identifies brain region changes caused by individual differences, resolving the mismatch issues resulting from general atlases and simple statistical divisions. This application constructs a multi-dimensional energy function that integrates personalized neural activity pattern representation information, the geometric neighborhood structure between voxels, and the rate of change of functional gradients. By minimizing this energy function, personalized partitioning of brain functions is achieved. Employing a self-supervised contrastive learning framework, it constructs voxel sample pairs within the same functional region and voxel pairs across functional regions, combining this with the aforementioned energy function to perform functional partitioning, ultimately obtaining a personalized whole-brain functional atlas with homogeneous activity within each partition and heterogeneous activity between sub-partitions.

[0087] S204, Functional Co-computation Guided by Personalized Neural Activity Pattern Representation (Module for Functional Co-computation Guided by Personalized Neural Activity Pattern Representation).

[0088] This step primarily involves constructing a personalized brain functional network capable of characterizing complex interactions between brain regions. Pre-trained collaborative computing models (such as the Transformer architecture) are used to infer functional collaborations between brain region nodes from personalized neural activity pattern representations. General capabilities for inter-brain region collaborative computing are extracted from large datasets, and feedback signals from downstream tasks guide the optimization of connection weights. This results in a personalized brain functional network that not only reflects the complex hierarchical relationships between brain region activities in different scenarios but also maximizes the performance of downstream tasks. Specifically: (1) Construct a downstream task prediction model (downstream task classifier). The downstream task prediction model first aggregates the voxel-level personalized neural activity pattern representations of each functional brain region after partitioning in step S203 into a region-level representation. Then, it uses collaborative computing to capture the complex hierarchical interaction relationships between regions and generates a personalized brain functional network. Finally, the data is input into the downstream prediction model. The prediction results are then used to provide feedback on task performance, enabling task-oriented personalized optimization of the model.

[0089] The specific calculation process is as follows: (1.1) Aggregation of regional-level personalized neural activity pattern representations.

[0090] It should be noted that the regions mentioned here are the brain regions obtained from the brain functional partitioning results.

[0091] Based on the final personalized functional partitioning results Average the characteristics within the region:

[0092] Obtain whole-brain region-level representation .

[0093] in, This represents the results of the 0th to Kth personalized function partitions. This represents the regional representation of the results of the 0th to Kth personalized functional partitions. This indicates the total number of personalized feature partitioning results. Representation dimension This represents the regional representation of the k-th brain region.

[0094] (1.2) Calculation of pre-trained collaborative relationships.

[0095] As an example, consider a pre-trained collaborative computing model based on the Transformer architecture. The core Attention module can be used to compute complex hierarchical dependencies between activity representations of different brain regions. Stacking Attention modules allows the collaborative computing model to more fully express complex hierarchical dependencies. Therefore, during pre-training on a large amount of neural activity signals using Transformer, a general collaborative relationship computation capability can be learned. This process can be represented as:

[0096] in, This represents the initial brain functional network, which is a personalized brain functional network constructed after the collaborative computational model is pre-trained on a large number of neural activity signals. Represents a collaborative computing model. These represent the parameters of the collaborative computing model.

[0097] (1.3) Prediction and feedback of downstream tasks.

[0098] Downstream task classifier take over As input, output class probability distribution :

[0099] The prediction result The difference from the real label is the key feedback signal for the reverse optimization of the personalized neural activity pattern representation module driven by the pre-trained model, the adaptive functional partitioning module guided by the personalized neural activity pattern representation, and the functional co-computation module guided by the personalized neural activity pattern representation.

[0100] (2) End-to-end joint training strategy for personalized pipeline construction.

[0101] This embodiment designs an end-to-end joint training strategy, chaining the models involved in each stage from steps (1.1) to (1.3) into a differentiable overall model framework. Key parameters are updated synchronously by predicting the loss function through downstream tasks. This global optimization strategy eliminates the error accumulation caused by step-by-step processing, ensuring personalized performance of the constructed brain functional network. Specifically: (2.1) Loss function of downstream task classifier.

[0102] Using standard cross-entropy loss Measuring the distribution of downstream task predictions With real labels Differences between them:

[0103] Where C represents the total number of iterations, and c represents the sequence number of the iteration. This represents the class probability distribution in the c-th iteration of optimization. This represents the regional representation of the personalized functional partitioning result corresponding to the c-th iteration optimization.

[0104] (2.2) Parameter synchronization update strategy.

[0105] In a single forward-backward propagation, the optimizer (such as Adam, in conjunction with an early stopping strategy) aims to minimize... To achieve this, the following two parameters will be updated simultaneously: Parameters in the pre-trained model-driven personalized neural activity pattern representation module (such as the low-rank increment parameters of LoRA) are used to fine-tune the generation of personalized neural activity pattern representations based on upstream injected neurodynamic prior knowledge, and parameters in the collaborative computation model are used to optimize the modeling of collaborative relationships between brain regions.

[0106] This application constructs a personalized functional network capable of characterizing complex interactions between brain regions. It infers the collaborative relationships between brain region nodes from personalized neural pattern representations using a pre-trained model. It extracts the general capabilities of brain region collaborative computation from a large dataset and optimizes connection weights using feedback signals from a downstream task classifier. This constructs a personalized brain functional network that reflects the complex hierarchical relationships between brain region activities in different scenarios and maximizes downstream task performance. Specifically, a downstream task classifier is constructed, receiving the computed personalized brain functional network as input and outputting prediction results. The prediction results provide feedback on task performance, enabling task-oriented personalized model optimization. An end-to-end joint training strategy is also designed, chaining the models involved in each stage into a differentiable overall model framework. The parameters of all models are updated synchronously using the prediction loss function of the downstream task classifier. This global optimization strategy eliminates the error accumulation caused by step-by-step processing, ensuring the personalized performance of the constructed brain functional network.

[0107] S205, Construction of Personalized Dynamic Brain Functional Networks Based on Time-Varying Synergistic Relationships.

[0108] Addressing the non-stationary nature of brain neural signals, a personalized dynamic brain function network was constructed capable of accurately capturing time-varying synergistic relationships. This network aims to overcome the averaging limitations of traditional static networks in the time dimension by deeply analyzing the dynamic evolution trajectory of neural signals over time, accurately identifying time windows and their functional synergistic patterns. It not only reveals the dynamic evolution mechanism of brain functional topology but also provides personalized dynamic brain function network support for neural modulation based on real-time neural activity interactions. Specifically: (1) Personalized temporal decomposition of neural activity.

[0109] Long-term fMRI data A refined temporal decomposition is performed. By monitoring the statistical distribution drift of signals in real time and defining the boundaries of local time windows in a personalized manner, the continuous and non-stationary time series is decomposed into a series of short-term segments with relatively stable internal states, thereby achieving the natural capture of the dynamic changes in neural activity.

[0110] This personalized time series decomposition can be described as:

[0111] in, Indicates the first identified A set of statistically distributed abrupt change points Indicates the first identified -1 set of statistically distributed abrupt change points, Indicates the first A short segment, express The elements in.

[0112] (2) Construction of personalized dynamic brain function network based on time-varying synergistic relationship.

[0113] For each short-term segment of neural activity obtained from the above decomposition, a personalized neural activity pattern representation is first constructed. Then, using nonlinear dependency analysis, the temporal correlation between the current short-term segment and historical short-term segments in terms of neural activity pattern representation is quantified, thereby constructing a time-varying collaborative weight matrix as dynamic prior information. Based on this, the aforementioned collaborative computation model is reused to construct the transient brain function network corresponding to each short-term segment. Finally, the time-varying collaborative weight matrix is ​​used to perform temporal weighted fusion of these transient brain function network sequences, ultimately outputting a personalized dynamic brain function network.

[0114] The construction process of this personalized dynamic brain function network can be described as follows:

[0115] in, Indicates time segment The corresponding transient brain function network, Indicates time segment The calculated time-varying collaborative weights, Indicates time segment The final personalized dynamic brain functional network state, Indicates time segment The corresponding personalized dynamic brain function network state.

[0116] This application addresses the non-stationary nature of brain neural activity by establishing a dynamic brain functional network construction framework capable of capturing time-varying synergistic relationships. It aims to overcome the limitations of static networks' time averaging by analyzing the dynamic evolution of neural signals over time and identifying functional synergistic patterns within different time intervals. This reveals the dynamic changes in the brain's functional topology, providing a personalized dynamic brain functional network for neural modulation through real-time interactive changes in neural activity. Specifically, long-term fMRI data is temporally decomposed based on changes in the statistical characteristics of neural activity signals. By detecting statistical distribution drift, the boundaries of local time intervals are individually determined, thus decomposing continuous non-stationary time series into several relatively stable short segments, achieving natural capture of dynamic changes in neural activity. For each decomposed neural activity time segment, a personalized neural activity pattern representation is constructed. Nonlinear dependency analysis is used to calculate the temporal correlation between the neural activity pattern representations of the current segment and the preceding segment, obtaining a time-varying synergistic weight matrix, which serves as a dynamic prior. Building upon this, the aforementioned collaborative computation module for brain inter-region activities is reused to construct transient brain functional networks for each segment. The time-varying collaborative weight matrix is ​​then used to perform temporal weighted fusion of the transient network sequences, resulting in the final personalized dynamic brain functional network. This process captures the brain's interaction patterns across different time segments, revealing the dynamic evolution of brain functional networks over time.

[0117] S206, Personalized Brain Functional Network Assessment Based on Virtual Neural Modulation (Personalized Dynamic Brain Functional Network Assessment Module Based on Virtual Neural Modulation).

[0118] A digital twin brain simulation environment for virtual neuromodulation was constructed to assess the reliability and modulatory potential of personalized brain functional networks before implementing real neuromodulation. Virtual stimuli were applied to the constructed digital twin brain model to simulate changes in brain state under personalized modulatory targets based on the personalized brain functional network. The rationality of the modulatory targets was further evaluated, the recovery rate was calculated based on a virtual doctor, and the correlation between the modulated network and the healthy group was analyzed. This not only validated the physiological realism of the personalized dynamic brain functional network but also predicted the modulatory effect, providing crucial decision-making basis for brain-computer interface modulation.

[0119] (1) Rationality assessment of personalized regulatory targets.

[0120] Based on prior knowledge in neuroscience, a rationality analysis is conducted on regulatory targets identified using personalized dynamic brain functional networks. The anatomical and functional rationality of the selected targets is assessed by quantitatively comparing and visually contrasting the calculated personalized targets with clinically recognized empirical targets through spatial mapping.

[0121] The specific method for calculating overlap is as follows: Define a personalized target set as... The empirical target set is By calculating the overlap between personalized target sets and empirical target sets in brain space. (Such as Dice coefficient or Jaccard index) to quantify reasonableness:

[0122] A higher overlap score indicates that personalized targets, while retaining individual specificity, are more in line with general neuromodulation anatomy.

[0123] (2) Assessment of rehabilitation rate based on virtual doctor neuromodulation.

[0124] A high-precision disease diagnosis model is introduced as a "virtual doctor" to monitor and evaluate brain state after virtual intervention. By simulating the changes in the connection weights of brain functional networks caused by stimulation of specific targets, the study observes whether the patient's brain network changes towards a healthy state.

[0125] Let the initial pathological brain network be Apply virtual stimulation The subsequent network status is .Will Input pre-trained disease diagnosis model :

[0126] in, The diagnostic results output by the pre-trained disease diagnosis model.

[0127] Recovery rate calculation: Statistics are based on the pre-trained disease diagnosis model on the test set after virtual modulation. The proportion of samples classified as healthy represents the virtual neuromodulation recovery rate. This indicator directly predicts the potential clinical efficacy of the modulation program, providing prospective guidance for brain-computer interface modulation strategies.

[0128] (3) Correlation analysis between virtual regulation and the healthy group network.

[0129] The personalized dynamic brain function network after virtual modulation was statistically compared with the brain network of the healthy control group to verify whether the brain activity pattern after virtual modulation has statistically returned to the normal level.

[0130] Calculate similarity metrics: Calculate the applied virtual stimulus The subsequent network status Average network of the healthy group Pearson correlation coefficient between them:

[0131] in, For similarity measurement results, This is the Pearson correlation coefficient.

[0132] Then determine the effectiveness: if virtual stimuli are applied The subsequent network status With health group If the similarity is significantly higher than before regulation, then the regulation scheme based on this personalized dynamic brain function network is deemed effective.

[0133] This application constructs a digital twin brain simulation environment for virtual neuromodulation, used to evaluate the reliability and modulatory potential of personalized brain networks before implementing real neuromodulation. Virtual stimulation is applied to the constructed digital twin brain model to simulate changes in brain state under personalized modulatory targets based on personalized brain functional networks. The rationality of the modulatory targets is further evaluated, the recovery rate is assessed based on a virtual doctor, and the correlation with the healthy group network is analyzed. This module not only verifies the physiological authenticity of the network but also predicts the modulatory effect, providing a decision-making basis for brain-computer interface modulation. The rationality of the modulatory targets based on personalized brain networks is analyzed based on prior knowledge. The rationality of the selected targets is evaluated by visualizing and calculating the overlap between personalized modulatory targets and empirical targets. Furthermore, a disease diagnosis model is introduced as a virtual doctor. The recovery rate of virtual neuromodulation based on personalized targets is calculated by monitoring whether the patient's brain network is diagnosed as healthy after virtual modulation. The personalized brain functional network after virtual modulation is statistically compared with the brain network of the healthy control group to calculate the similarity between the modulated state and the healthy group. If the characteristics of the modulated network are significantly similar to the healthy group, the personalized network is deemed to be accurately constructed and the modulation scheme effective.

[0134] It should be noted that the validity of this application has been systematically verified through the construction of personalized brain functional networks in multiple scenarios. Based on publicly available fMRI brain imaging data, disease diagnosis tasks, brain decoding tasks, and brain fingerprinting tasks covering autism, Parkinson's disease, Alzheimer's disease, depression, and ADHD were constructed, involving 680 subjects. In the experiment, all subjects achieved personalized brain functional network construction by representing brain activity patterns. Experimental results show that this application demonstrates excellent personalized brain functional network construction capabilities in various scenarios, and can stably support various downstream applications and analysis needs.

[0135] Furthermore, the method proposed in this application possesses high generalization and scalability. Although disease diagnosis, brain fingerprint recognition, and motor imagery decoding are used as verification tasks in specific embodiments, the technical solution of this application is not limited to these specific tasks. Since this application is based on pre-trained neural activity pattern representation, it can theoretically support the application to cover any task without reconstructing the overall architecture for specific task data. In addition, for task scenarios with unsatisfactory performance, this application can adopt a modular plug-and-play strategy for each step in the above embodiments, requiring only separate training of the LoRA module to achieve the extension to that task. Moreover, the personalized brain functional network construction framework based on pre-trained neural activity pattern representation proposed in this application can be further extended to brain network construction tasks with other data modalities, such as electroencephalography (EEG). Its basic principles and methodological framework do not depend on specific neural signal types. Personalized brain functional network construction can be performed under neural activity pattern representation without substantial modifications to the core technical route, greatly expanding the application boundaries and innovative potential of brain functional networks in basic neuroscience and clinical translation.

[0136] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A personalized brain function network construction and evaluation method for brain-computer interface regulation, characterized in that, The method comprises the following steps: Based on the preset neural dynamics equation, the functional magnetic resonance imaging time series data of a specific subject is combined to inject neural dynamics prior knowledge into the personalized neural activity pattern representation of the specific subject; Generalization and physiological rationality optimization are performed on the personalized neural activity pattern representation injected with neural dynamics prior knowledge to obtain an optimized personalized neural activity pattern representation; Based on the optimized personalized neural activity pattern representation, brain function partitioning is performed; The brain function partitioning result is interpreted and quantitatively analyzed to obtain a brain function coordination quantization index; Based on the brain function partitioning result, the time dimension is expanded on the brain function coordination quantization index to obtain a personalized dynamic brain function network; A digital twin brain simulation environment corresponding to the brain of the specific subject is built, and the performance of the personalized dynamic brain function network is evaluated by applying a virtual stimulus.

2. The method of claim 1, wherein the method is a method of constructing and evaluating a personalized brain functional network for brain-computer interface regulation. The method for injecting neural dynamics prior knowledge into the personalized neural activity pattern representation of the specific subject comprises the following steps: According to the functional magnetic resonance imaging time series data of the specific subject, the parameter set of the optimal personalized neural dynamics equation is fitted; The parameter set of the optimal personalized neural dynamics equation is substituted into the preset neural dynamics equation to determine the neural activity function corresponding to the specific subject; A physical constraint loss term containing the neural activity function corresponding to the specific subject is constructed, and a data fitting loss term is combined to form a compound loss function; By minimizing the compound loss function, the learning process of the personalized neural activity pattern representation is optimized to obtain a corrected neural activity pattern representation, thereby completing the injection of neural dynamics prior knowledge into the personalized neural activity pattern representation of the specific subject. 3.The method of claim 1, wherein, The method for generalizing and optimizing the physiological rationality of the personalized neural activity pattern representation injected with neural dynamics prior knowledge comprises the following steps: The personalized neural activity pattern representation injected with neural dynamics prior knowledge is projected into a low-dimensional latent space to obtain a low-dimensional latent space representation; A diffusion model is used to capture the high-order statistical rules and spatiotemporal dependence relationships in the low-dimensional latent space representation to obtain a personalized neural activity pattern representation in the low-dimensional latent space; the diffusion model is trained with the neural dynamics equation as a physical constraint and added to the loss function for training the diffusion model; The personalized neural activity pattern representation in the low-dimensional latent space is mapped back to the original voxel space using a voxel space decoder to obtain an optimized personalized neural activity pattern representation.

4. The method of claim 3, wherein the method is used for constructing and evaluating personalized brain function network for brain-computer interface regulation. The result obtained by adding the neural dynamics equation as a physical constraint to the loss function for training the diffusion model is: wherein, represents a loss function of physical constraint fine-tuning as the total loss of the diffusion model, represents a physical driving term, represents a data-driven term, represents a regularization coefficient balancing data fitting and physical constraint, represents a voxel space decoder, represents a pre-trained base representation model, represents a low-dimensional latent space preliminary neural activity representation, represents a neural activity equation, represents input raw functional magnetic resonance imaging data.

5. The method of claim 1, wherein the method is a method of constructing and evaluating a personalized brain functional network for brain-computer interface regulation. The method for performing brain function partitioning based on the optimized personalized neural activity pattern representation comprises the following steps: A brain function partitioning target is determined based on the principle of maximizing the homogeneity of the optimized personalized neural activity pattern representation; According to the brain function partitioning target, an energy function is constructed to convert the biophysical constraints of brain function partitioning into a mathematical optimization target; An individualized representation contrast loss is established by combining the energy function and a self-supervised contrast learning framework; The optimal brain function partitioning result is calculated based on the individualized representation contrast loss.

6. The method of claim 5, wherein the method is a method of personalized brain function network construction and evaluation for brain-computer interface regulation. The individualized representation contrast loss is: in, For personalized characterization of contrast loss, Indicates the current voxel Personalized neural activity pattern representation, Indicates the current voxel Personalized neural activity pattern representation, Indicates the relationship with the current voxel Voxel representation of positive samples belonging to the same predictive functional region. This represents a temperature coefficient used to adjust the degree of attention given to difficult samples in contrastive learning. Indicates the relationship with the current voxel The set of negative sample voxels belonging to different prediction functional regions. Represents a set of voxels. express The elements in.

7. The method of claim 1, wherein the method is a method of constructing and evaluating a personalized brain functional network for brain-computer interface regulation. The method for interpreting and quantitatively analyzing the brain function partitioning result to obtain a brain function coordination quantification index comprises the following steps: The brain function partitioning result is recorded as a plurality of brain regions; the following steps are performed: (1) Based on the brain function partitioning result, the individualized neural activity pattern representation of all voxels in each brain region is averaged to obtain a regional-level representation of each brain region; (2) The regional-level representations of all brain regions are input into a pre-trained coordination calculation model to capture the hierarchical interaction relationship between different brain regions and generate an initial brain function network; (3) The initial brain function network is input into a downstream task classifier to output a brain function prediction result; wherein the brain function network is a carrier of the brain function coordination quantification index; (4) Steps (1) to (3) are connected as a whole model framework, the parameters of the whole model framework are updated synchronously through the loss function of the downstream task classifier, the brain function network is optimized, and the optimal brain function network is obtained, i.e., the brain function coordination quantification index is obtained.

8. The method of claim 7, wherein the method is a method of constructing and evaluating a personalized brain functional network for brain-computer interface regulation. The method for expanding the time dimension on the basis of the brain function partitioning result and the brain function coordination quantification index to obtain a personalized dynamic brain function network comprises the following steps: Based on the statistical distribution drift of the real-time monitoring signal of functional magnetic resonance imaging, the boundary of the local time window is personalized defined, and the functional magnetic resonance imaging time series data is disassembled into a plurality of short-time segments; For each short-time segment, a personalized neural activity pattern representation is constructed based on the brain function partitioning result, and a time-varying coordination weight matrix is constructed as dynamic prior information by using a nonlinear dependence analysis technique to quantize the time sequence correlation between the current short-time segment and the historical short-time segment in the neural activity pattern representation; The short-time segment corresponding instantaneous brain function network is constructed by reusing the coordination calculation model; The instantaneous brain function network sequence composed of the instantaneous brain function networks corresponding to each short-time segment is time sequence weighted fused according to the time-varying coordination weight matrix to obtain a personalized dynamic brain function network. 9.The method of claim 1, wherein, The method for evaluating the performance of the personalized dynamic brain function network comprises the following steps: A digital twin brain simulation environment corresponding to the brain of a specific subject is built, a virtual stimulus is applied, and a rationality analysis is performed on the personalized dynamic brain function network according to the overlap degree between the identified regulatory target set and the experience target set of the personalized dynamic brain function network; The neural regulation rehabilitation rate of the specific subject is evaluated based on a preset disease diagnosis model, the personalized dynamic brain function network is taken as an initial regulation object, the personalized dynamic brain function network is virtually regulated through the virtual stimulus, the regulation effect is evaluated, and the rehabilitation rate is calculated; The personalized dynamic brain function network after virtual regulation is statistically compared with the brain function network of a healthy control group to verify whether the brain activity pattern after virtual regulation meets the statistical regression normal level.

10. The method of claim 9, wherein the method is a method of constructing and evaluating a personalized brain functional network for brain-computer interface regulation. The method for calculating the overlap degree comprises the following steps: wherein, represents the degree of overlap, represents the set of regulatory targets identified by the personalized dynamic brain function network, represents the set of empirical targets.

Images