A brain effective connectivity learning and analysis method based on multi-source spatio-temporal collaborative modeling

Abstract

The application discloses a brain effect connection learning and analysis method based on multi-source space-time collaborative modeling, and belongs to the technical field of brain science and medical image processing. The method comprises the following steps: receiving multi-source brain time sequence data, and constructing a unified space-time representation of brain regions; performing space correlation modeling and time collaborative processing on the brain region level time sequence to obtain a unified space-time feature representation; dynamically modeling the unified space-time feature representation through a multi-source collaborative learning mechanism to obtain a brain region fusion feature representation; constructing a connection relationship learning model based on the brain region fusion feature representation, determining the connection relationship between brain regions through a structure generation and optimization process, and obtaining a brain effect connection network. Through unified space-time representation modeling, a multi-source collaborative learning mechanism and a structure feedback optimization process, the application effectively improves the accuracy, stability and interpretability of brain effect connection modeling, and can be widely applied to the fields of brain function analysis, auxiliary diagnosis of neurological diseases and brain state recognition.

Description

Technical Field

[0001] This invention proposes a brain effect connectivity learning and analysis method based on multi-source spatiotemporal collaborative modeling. By constructing a unified spatiotemporal representation, a collaborative learning model, and a structural feedback optimization mechanism, it achieves stable learning and analysis of brain region connectivity relationships. Background Technology

[0002] Brain effector connectivity, used to describe the interactions and information transmission pathways between different brain regions, is an important area of ​​brain science research, with significant applications in analyzing brain functional mechanisms, assisting in the diagnosis of neurological diseases, and recognizing brain states. Existing brain effector connectivity modeling methods mainly rely on statistical modeling or time-series-based analysis, characterizing the dependencies between brain regions over time to infer their interaction structures. However, these methods typically rely on strong prior assumptions, such as linearity or stationarity, making them ill-suited to the nonlinear and dynamic characteristics inherent in complex brain activity.

[0003] With the development of data-driven methods, some studies have introduced machine learning models to model brain region relationships, which has improved modeling capabilities to some extent. However, existing methods mostly analyze time-series brain region data from a single source, failing to fully utilize the complementary information between multi-source data. In addition, data from different sources differ in terms of time scale, dynamic characteristics, and signal expression. Directly applying a unified processing approach can easily lead to inconsistent information expression, thereby affecting the accuracy and stability of connectivity modeling.

[0004] On the other hand, existing methods typically separate feature representation learning from connection structure learning, lacking a unified modeling framework. This makes it difficult for data representation and connection structure to form an effective synergy, easily leading to error accumulation and reducing model reliability. Furthermore, once the connection structure is determined, it usually does not participate in subsequent optimization processes, lacking a feedback mechanism and making it difficult to continuously correct the modeling results.

[0005] Therefore, how to achieve spatiotemporal collaborative modeling of time series data from multiple brain regions within a unified framework, and how to introduce a structural feedback mechanism during the connectivity learning process to improve the stability and accuracy of brain effect connectivity modeling, has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method for learning and analyzing brain effective connectivity based on multisource spatiotemporal collaborative modeling (MSC-EC). The multisource data is preferably time-series data from functional magnetic resonance imaging (fMRI) and electroencephalography (EEG). By modeling a unified spatiotemporal representation of brain time series data from different sources, and combining a collaborative learning mechanism with a structural feedback optimization process, stable learning and analysis of brain region connectivity relationships are achieved. This improves the accuracy of brain effective connectivity modeling and solves problems such as inconsistent multisource data representation and unstable connectivity modeling in existing methods.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A brain effect connectivity learning and analysis method based on multi-source spatiotemporal collaborative modeling includes:

[0009] Receive multi-source brain time-series data;

[0010] Based on the multi-source brain time series data, a unified spatiotemporal representation at the brain region level is constructed;

[0011] Spatial correlation modeling and temporal co-processing are performed on the brain region-level time series to obtain a unified spatiotemporal feature representation;

[0012] The unified spatiotemporal feature representation is dynamically modeled through a multi-source collaborative learning mechanism to obtain brain region fusion feature representation;

[0013] A connectivity learning model is constructed based on the brain region fusion feature representation, and the connectivity between brain regions is determined through a structure generation and optimization process to obtain the brain effect connectivity network.

[0014] Optionally, constructing a unified spatiotemporal representation at the brain region level based on the multi-source brain region time-series data includes:

[0015] By constructing a spatial correlation weight matrix, brain region time series from different sources are mapped to a unified brain region representation space, and data at different time scales are processed uniformly to obtain a spatiotemporally consistent representation at the brain region level.

[0016] Optionally, spatial correlation modeling and temporal co-processing of the brain region-level time series include:

[0017] A spatial weight matrix is ​​constructed based on the spatial correlation between brain regions, and the time series of the brain regions are weighted and transformed.

[0018] Time-scale unification processing of data with different time resolutions is performed based on an adaptive time window;

[0019] By dynamically adjusting the time series using a time response function, a unified spatiotemporal feature representation can be obtained.

[0020] Optionally, dynamically modeling the unified spatiotemporal feature representation through a multi-source collaborative learning mechanism includes:

[0021] By introducing adjustable parameters, the contributions of data from different sources to different brain regions and at different time stages are dynamically weighted.

[0022] By combining and mapping multi-source data using parameterized functions, a fused representation of brain region features is generated.

[0023] Optionally, constructing a connectivity learning model based on the brain region fusion feature representation includes:

[0024] Brain regions are represented as network nodes, and the interactions between brain regions are described by a parameterized connection matrix;

[0025] Modeling of connection direction and connection strength is performed based on the relationship between the historical and current states of brain regions.

[0026] Optionally, determining the connectivity between brain regions through structure generation and optimization processes includes:

[0027] S1: Construct the connection structure parameter space and initialize the connection relationships;

[0028] S2: Evaluate candidate connection structures based on the structural evaluation function and update connection parameters through an optimization process;

[0029] S3: Generate brain effect connectivity structures based on the updated connectivity parameters.

[0030] Optionally, the structural evaluation function is optimized by minimizing a structural loss function, wherein the structural loss function is... for: ;in, For the prediction error term, For structural constraints, These are the weighting coefficients.

[0031] Optionally, a structural feedback optimization mechanism may also be included: the currently learned connectivity structure is fed back as constraint information to the multi-source collaborative learning model to correct the brain region fusion feature representation;

[0032] By minimizing the joint optimization objective function and simultaneously updating the multi-source collaborative learning parameters, connection matrix, and bias terms, collaborative optimization between data representation and connection structure is achieved.

[0033] Optionally, the joint optimization objective function is:

[0034] ;in, This is the structural consistency error term. These are the weighting coefficients for the consistency constraint terms.

[0035] Optionally, the multi-source data includes time-series data from functional magnetic resonance imaging (fMRI) and electroencephalography (EEG).

[0036] The beneficial effects of this invention are as follows:

[0037] (1) Existing methods mostly rely on a single data source for brain effect connectivity modeling, which makes it difficult to take into account the temporal and spatial expressive characteristics of different data. This invention achieves the fusion expression of multi-source data under a unified framework through a multi-source spatiotemporal collaborative modeling mechanism, thereby improving the modeling accuracy.

[0038] (2) Existing methods typically separate feature representation from connection structure learning, lacking a unified optimization mechanism. This invention introduces a structural feedback optimization mechanism, which allows the connection structure to act inversely on the data representation, thereby achieving a closed-loop optimization process for the model and improving the stability of the results.

[0039] (3) This invention optimizes the connection relationship in a continuous space by learning the parameterized connection structure, thereby avoiding the computational complexity problem of traditional discrete search methods and improving the solution efficiency.

[0040] (4) The brain effect connection network constructed in this invention can simultaneously reflect the spatiotemporal dynamic characteristics of multi-source data, providing more reliable support for brain function analysis and related applications. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating a brain effect connectivity learning and analysis method based on multi-source spatiotemporal collaborative modeling, according to an embodiment of the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make various modifications or substitutions to the present invention, all of which should fall within the scope of protection of the present invention.

[0044] In this embodiment, the input is time-series data from multiple brain regions. Assume there are a total of... There are 10 data sources, each represented as a brain region-level time series. Each data source is represented as:

[0045]

[0046] in, ; Indicates the number of brain regions; Indicates the first The duration of each data source; Indicates the first The first data source The brain region in the first The values ​​are taken at each time point. Since different data sources differ in terms of time length, temporal resolution, and dynamic response characteristics of brain regions, they need to be processed within a unified modeling framework.

[0047] First, spatial correlation modeling is performed on the time series data of multiple brain regions. For the first... Using multiple data sources, construct a spatial association matrix between brain regions:

[0048]

[0049] in, Indicates the first brain regions of each data source brain regions Spatial association weights are determined based on the spatial coordinate alignment of brain regions. Specifically, brain regions are matched according to their positional relationships in a unified spatial coordinate system across different data sources, and association weights between these regions are constructed accordingly. The original brain region time series is transformed using a spatial association matrix to obtain a spatially enhanced representation.

[0050]

[0051] Right now

[0052]

[0053] in, Indicates the th after spatial correlation transformation The first data source The brain region in the first The representation values ​​at each time point. To ensure the stability of the weight allocation for each brain region, a normalization constraint can be applied to the spatial correlation matrix:

[0054]

[0055] Through the above steps, data from different sources are mapped into a unified brain region association expression framework in the spatial dimension, thereby reducing the inconsistency between multi-source data caused by differences in brain region expression.

[0056] In terms of time dimension, to address the issue of inconsistent time resolution across different data sources, this embodiment introduces a time-coordinated processing mechanism. Let the length of the unified time axis be... For each data source, a mapping relationship is constructed from the original timeline to the unified timeline. For the unified time point... , define the first The set of time windows corresponding to each data source is:

[0057]

[0058] in, Indicates the first Data sources and a unified time point The corresponding center position; This represents the radius of the time window; This represents the set of time indices used for aggregation. Based on this time window, temporal aggregation is performed on the spatially augmented representation to obtain the time-aligned representation:

[0059]

[0060] in, Represents a set The number of elements in; Indicates the first The first data source individual brain regions at the same time point The time-coordinated representation is used. To further characterize the dynamic response differences between data from different sources, this embodiment also introduces a time response function. Furthermore, the time-aligned representation is dynamically corrected using convolution.

[0061]

[0062] in, Indicates the first The response length of each data source; Indicates the first The data source in the first Response weights at each delay location; This represents the unified time representation after dynamic response correction. Through formulas (7) and (8), data from different sources are presented in a unified time frame for time-series collaborative representation.

[0063] After obtaining a unified spatiotemporal representation, a multi-source collaborative learning model is constructed. Let the fused brain region feature representation be... ,in Indicates the first individual brain regions at the same time point The fusion feature is composed of a weighted combination of data from different sources:

[0064]

[0065] in, This represents the co-weight of the m-th data source at brain region i and time point t. This weight is not a fixed value but is dynamically determined by learnable parameters. In one embodiment, the weight can be defined using a normalized exponential function:

[0066]

[0067] in, For the first The data source is in the brain region. Time point The learnable scoring parameters at the given location. Through formula (10), it can be guaranteed that:

[0068]

[0069] That is, the sum of the contributions of data from different sources to the same brain region at the same time is 1, thus giving the multi-source collaborative learning process a clear physical meaning and interpretability.

[0070] Based on this, a brain effect connectivity learning model is constructed. Let the brain effect connectivity matrix be:

[0071]

[0072] in, Indicates brain regions Pointing to brain regions The connection strength; if This indicates the brain region. brain regions There is a relationship of interaction. Regarding brain regions... At the point of time The state of the brain region is represented by a dynamic modeling approach based on the state of the brain region at the previous moment:

[0073]

[0074] in, This indicates brain regions predicted based on connectivity structures. At the point of time The state; Indicates brain regions The bias term. Formula (13) gives the connectivity matrix. With brain region dynamic representation The way these connections are linked transforms the problem of learning brain region connectivity into a parameter matrix. The problem to be solved.

[0075] To learn the optimal connection structure, this embodiment constructs a prediction error term:

[0076]

[0077] in, This represents the fitting error of the connectivity structure to the dynamic representation of brain regions. To avoid an overly dense connectivity matrix, a structural constraint term is also introduced:

[0078]

[0079] in, The complexity of the connection matrix is ​​constrained, leading to a sparser network structure. The resulting structure learning objective function is:

[0080]

[0081] in, These are the weighting coefficients for the structural constraint terms, used to balance fitting accuracy and structural complexity.

[0082] The key to this invention lies in the fact that the connection structure learning is not completed independently, but rather a structural feedback optimization mechanism is further introduced. Specifically, after obtaining the current connection matrix... Then, the connection matrix is ​​fed back as constraint information to the multi-source collaborative learning model to correct the fused feature representation. Therefore, the structural consistency error is defined as:

[0083]

[0084] in, Indicates a point in time Brain region fusion feature vector; Represents the L2 norm; This reflects the degree of consistency between the current fused representation and the connection structure. If If the value is large, it indicates that the current feature representation has not yet fully matched the learned connectivity relationships, and the parameters in the multi-source collaborative learning model need to be updated.

[0085] Therefore, the joint optimization objective of the entire method can be written as:

[0086]

[0087] in, The weight coefficients of the consistency constraint term are denoted by . By minimizing formula (18), the multi-source collaborative learning parameters can be updated simultaneously. Connection matrix and bias terms This enables collaborative optimization between data representation and connection structure.

[0088] After iterative optimization, the final brain effect connectivity matrix is ​​obtained:

[0089]

[0090] Based on this, a brain effect connectivity network can be constructed:

[0091]

[0092] in, Represents a set of brain region nodes; Indicates by A defined set of directed edges; This represents the final connection strength matrix. Furthermore, brain effect connectivity analysis can be performed based on this connection matrix, such as calculating node in-degree, out-degree, critical path strength, and local subnetwork connectivity density. Taking node out-degree as an example, it can be expressed as:

[0093]

[0094] in, Indicates brain regions The degree of departure; This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise. This analysis process can identify key brain regions that play a major role in network propagation.

[0095] To verify the effectiveness of the proposed method in the brain effect connectivity learning task, this embodiment conducts experimental evaluation on a simulated dataset. The simulated dataset is a 6-node network with a known real connectivity structure. The performance of each method is evaluated by comparing the connectivity structures recovered by different methods with the real structures.

[0096] This experiment selected several representative brain effect connectivity modeling methods as comparison methods, including statistical analysis-based methods Patel, lsGC, and pwLiNGAM, and deep learning-based methods DiffAN, MetaRLEC, and FSTA-EC. Simultaneously, the proposed method MSC-EC was compared and analyzed with the aforementioned methods.

[0097] To comprehensively evaluate model performance, accuracy, precision, recall, F1 score, and structural Hamming distance (SHD) were used as evaluation metrics. SHD measures the difference between the predicted connectivity structure and the true structure; a lower SHD value indicates better structure recovery.

[0098] Patel 0.65 0.60 0.13 0.21 5.20 lsGC 0.52 0.40 0.40 0.40 7.20 pwLiNGAM 0.67 0.55 0.50 0.50 5.00 DiffAN 0.70 0.81 0.36 0.49 4.40 MetaRLEC 0.67 1.00 0.17 0.29 5.00 FSTA-EC 0.41 0.41 1.00 0.58 8.80 MSC-EC 0.73 0.58 0.77 0.65 4.20

[0099] Overall, the proposed method demonstrates superior performance across multiple evaluation metrics, particularly in structural reconstruction. Firstly, in terms of accuracy, the method achieves 0.73, the highest among all compared methods, indicating high accuracy in the overall connectivity prediction task. Secondly, in terms of recall, the method achieves the highest value of 0.77, demonstrating its ability to effectively identify real-world connectivity relationships. In contrast, some traditional and deep learning methods exhibit lower recall, indicating significant missed detections during connectivity reconstruction. Regarding precision, the method achieves 0.58, slightly lower than some conservative methods, but combined with recall, it shows that the proposed method maintains good prediction accuracy while avoiding missed detections, achieving a better overall balance. Finally, in terms of the comprehensive F1-score, the method achieves 0.65, the highest among all methods, further demonstrating that it achieves the best balance between precision and recall, exhibiting stronger overall performance. The proposed method achieves the lowest SHD (Structured Hierarchy Value) of 4.20, indicating that its predicted connection structure is closest to the actual structure. In contrast, while the FSTA-EC method achieves a recall of 1.00, its SHD is as high as 8.80, indicating that this method suffers from over-connection, leading to a large number of erroneous edges.

[0100] The above experiments show that the proposed method MSC-EC in this embodiment has better performance and higher accuracy in learning brain effect connectivity networks on simulated datasets compared with other methods, and therefore has great application prospects in computer-aided diagnosis of brain diseases.

[0101] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A brain effect connectivity learning and analysis method based on multi-source spatiotemporal collaborative modeling, characterized in that, Includes the following steps: Receive multi-source brain time-series data; Based on the multi-source brain time series data, a unified spatiotemporal representation at the brain region level is constructed; Spatial correlation modeling and temporal co-processing are performed on the brain region-level time series to obtain a unified spatiotemporal feature representation; The unified spatiotemporal feature representation is dynamically modeled through a multi-source collaborative learning mechanism to obtain brain region fusion feature representation; A connectivity learning model is constructed based on the brain region fusion feature representation, and the connectivity between brain regions is determined through a structure generation and optimization process to obtain the brain effect connectivity network.

2. The method according to claim 1, characterized in that, The construction of a unified spatiotemporal representation of brain regions based on multi-source brain time-series data includes: By constructing a spatial correlation weight matrix, brain region time series from different sources are mapped to a unified brain region representation space, and data at different time scales are processed uniformly to obtain a spatiotemporally consistent representation at the brain region level.

3. The method according to claim 1, characterized in that, Spatial correlation modeling and temporal co-processing of the aforementioned brain region-level time series include: A spatial weight matrix is ​​constructed based on the spatial correlation between brain regions, and the time series of the brain regions are weighted and transformed. Time-scale unification processing of data with different time resolutions is performed based on an adaptive time window; By dynamically adjusting the time series using a time response function, a unified spatiotemporal feature representation can be obtained.

4. The method according to claim 1, characterized in that, Dynamic modeling of the unified spatiotemporal feature representation through a multi-source collaborative learning mechanism includes: By introducing adjustable parameters, the contributions of data from different sources to different brain regions and at different time stages are dynamically weighted. By combining and mapping multi-source data using parameterized functions, a fused representation of brain region features is generated.

5. The method according to claim 1, characterized in that, Constructing a connectivity learning model based on the aforementioned brain region fusion feature representation includes: Brain regions are represented as network nodes, and the interactions between brain regions are described by a parameterized connection matrix; Modeling of connection direction and connection strength is performed based on the relationship between the historical and current states of brain regions.

6. The method according to claim 1, characterized in that, Determining the connectivity relationships between brain regions through structure generation and optimization processes includes: S1: Construct the connection structure parameter space and initialize the connection relationships; S2: Evaluate candidate connection structures based on the structural evaluation function and update connection parameters through an optimization process; S3: Generate brain effect connectivity structures based on the updated connectivity parameters.

7. The method according to claim 6, characterized in that, The structural evaluation function is optimized by minimizing a structural loss function. for: ;in, For the prediction error term, For structural constraints, These are the weighting coefficients.

8. The method according to claim 1, characterized in that, It also includes a structural feedback optimization mechanism: The learned connectivity structure is fed back as constraint information to the multi-source collaborative learning model to correct the brain region fusion feature representation; By minimizing the joint optimization objective function and simultaneously updating the multi-source collaborative learning parameters, connection matrix, and bias terms, collaborative optimization between data representation and connection structure is achieved.

9. The method according to claim 8, characterized in that, The joint optimization objective function is: ;in, This is the structural consistency error term. These are the weighting coefficients for the consistency constraint terms.

10. The method according to claim 1, characterized in that, The multi-source data includes time-series data from functional magnetic resonance imaging (fMRI) and electroencephalography (EEG).

Images