Space-time modeling method for activating network to improve human brain communication process

Abstract

The invention discloses a space-time modeling method for improving a human brain communication process by activating a network. The space-time modeling method comprises the following steps: S1, integrating data; s2, data preprocessing; s3, activity and background of function connection; s4, function connection calculation activity; s5, constructing a high-low activation network; s6, performing graph theory analysis and classification; s7, performing statistical test; and S8, obtaining a result. Experimental results show that the AFC shows high correlation with potential dynamics, in addition, the activation network shows more multi-region interaction in space and has high sensitivity to mental state change, more information of large-scale brain network communication is temporarily revealed with a stable communication effect and low cost, and the method has the advantages of being simple in structure, convenient to operate and high in practicability. According to the method, verification is carried out through an application program, disease classification is achieved, the network activation is an important framework for understanding the brain network, new insights are provided for brain network construction through research, and the potential of using the network activation in network neuroscience research is emphasized.

Description

Technical Field

[0001] This invention relates to the field of human brain communication technology, and in particular to a spatiotemporal modeling method for improving the human brain communication process by activating networks. Background Technology

[0002] The human brain establishes cognitive functions through continuous communication between multiple functional systems. Complex and multifaceted communication is crucial for efficient information processing. Many studies have used structural and functional brain networks to identify the neural mechanisms that allow for efficient processing. However, over time, stable spatial structures have become a major constraint on functional networks that simulate the brain's effective communication processes. Popular functional network methods rely solely on statistical correlations without considering the complexity between dynamic neural fluctuations and basic maintenance. Therefore, capturing precise dynamic neural fluctuations within statistical correlations and establishing the optimal spatiotemporal framework for brain communication remains a significant challenge.

[0003] In the context of brain networks, functional connectivity (FC) has been widely used to describe this process. Static FC has traditionally been calculated using statistical correlations over the entire time series of data acquisition. The exploration of static FC for various physiological signals (EEG, EMG, fMRI, etc.) has greatly expanded our understanding of the functional brain structures behind cognitive processes, diseases, and their neural regulation. However, static FC ignores the detailed changes at every moment in communication and therefore cannot meet the ever-evolving requirements of network neuroscience regarding intrinsic dynamics. Traditional FC calculations rely heavily on indirect measurements, which indirectly encode neural activity information derived from spontaneous neural activity streams in time series. This has been demonstrated in studies using anesthesia, which showed that FC still fluctuates in unconscious states and some characteristics remain similar to those in conscious states. Therefore, a portion of the time-varying FC reflects non-cognitive intrinsic or steady-state processes. This non-cognitive component acts as background during communication, and neural fluctuations must be precisely distinguished from the background.

[0004] In applications on autism spectrum disorder (ASD) and COVID-19 datasets, the proposed activation network extracted richer topological reorganization information that was largely invisible to the DFN. Specifically, the activation network exhibited significant inter-regional connectivity between specific functional subnetworks and was able to achieve topological reorganization in the temporal dimension more effectively. Furthermore, while the DFN could not distinguish between patients and healthy controls, the proposed method revealed a significant decline in the information processing capacity of the patient's brain. Finally, the combination of the two types of networks successfully classified ASD and COVID-19. These findings suggest that the proposed method can serve as a potential analytical framework for elucidating the neural mechanisms of brain dynamics.

[0005] Therefore, the applicant proposed a spatiotemporal modeling method for improving the human brain communication process by activating networks. Summary of the Invention

[0006] In view of the above-mentioned problems in the prior art, the main objective of the present invention is to provide a spatiotemporal modeling method for human brain communication processes that improves upon activation networks.

[0007] The technical solution of this invention is as follows: a spatiotemporal modeling method for improving human brain communication processes by activating networks, comprising the following steps:

[0008] S1. Integrate data;

[0009] S2, Data Preprocessing;

[0010] S3, the activities and background of the functional connections;

[0011] S4, Functional Connection Computing Activity;

[0012] S5, High and Low Activation Network Construction;

[0013] S6. Perform graph theory analysis and classification;

[0014] S7. Statistical tests;

[0015] S8. Obtain the result.

[0016] In a preferred embodiment, step S1 can be further refined into the following steps:

[0017] S11. First, integrate simulated resting-state BOLD-FMRI data;

[0018] S1101. By using simulated resting-state BOLD-fMRI data, the “ground reality” of the time series can be controlled, which can simulate the activity and background properties of functional connectivity to evaluate the dynamic detection capability of the proposed method.

[0019] S1102. The enhancement and reduction levels of background correlation affected by dynamics (ΔFC) are measured to express the activity of functional connectivity. The simulated BOLD-fMRI is generated by a multivariate Gaussian process and a first-order vector autoregression (VAR) model to simulate background and dynamic data, respectively. The background data mainly provides the stability of FC.

[0020] S1103. Generate background data using a pairwise zero-mean multivariate Gaussian process, σ = (σ1, ..., σ2). n The background correlation is represented by the covariance moment ∑ randomly generated between -1 and 1, and then the dynamics of the simulation data are specified.

[0021] S1104. Background correlation is expected to be stochastically influenced by dynamics to reflect the persistent nature of resting-state BOLD-fMRI fluctuations. Dynamic data are estimated in pairs based on a first-order VAR model.

[0022] ε t =Aε t-1 +e t Where A = 0.8 is the autocorrelation coefficient, e t The model residuals, determined through a random Gaussian process, have a mean of 0.2 and a standard deviation (std) of 0.12.

[0023] S1104. Finally, the simulated BOLD-fMRI data is a linear combination of background and dynamic data:

[0024] V=σ+ε t The proposed method was validated by correlating it with ΔFC, generating 5000 data points with a length of 3000 points to ensure the reliability of the output;

[0025] S12, Secondly, integrate autism spectrum disorder data;

[0026] S13, Finally, integrate the COVID-19 data;

[0027] S14. Combining the above three steps, obtain the required data.

[0028] In a preferred embodiment, step S2 can be further subdivided into the following steps:

[0029] S21. Data preprocessing of resting functional volumes involves formal transformation by the Montreal Neuroscience Institute, slice timing, head motion correction, spatial normalization (MNI) space, resampling resolution of 3×3×3 spatial smoothing (full width at half maximum of 6×6×6), linear detrending to reduce low-frequency drift and physiological high-frequency respiratory and cardiac noise, and time bandpass filtering (0.01-0.1Hz). Friston24 parameter model is used to regress head motion effects.

[0030] S22. Motion rotations with a maximum translation of ≥1.5 and / or ≥1.5 degrees in the X, Y, or Z directions of the participants were removed. Linear regression was applied to remove global mean signal, white matter, and cerebrospinal fluid signal. The registered fMRI volumes were partitioned using the Dosenbach 160 region of interest (ROI) template.

[0031] In a preferred embodiment, step S3 can be further subdivided into the following steps:

[0032] S31. The communication dynamic and non-dynamic properties of FC are distinguished by the Activity Functional Connection (AFC) and Background Functional Connection (BFC). Paired time series and sliding window methods are used to identify the time invariance and time specificity of FC in the communication process, where the background represents the time invariance property of FC across the time window.

[0033] S32. AFC is obtained by subtracting the background from the total signal. It measures the degree to which the activity level changes from a general state to a specific time state.

[0034] S33. Calculate the statistical correlation of paired time series throughout the entire scanning period to obtain the static FC, which is used to extract the time-invariant properties of the FC. To prevent ignoring the detailed communication patterns that occur at each time step, replace the static FC with the time-varying FC. It divides the entire time series into sliding windows and calculates the statistical correlation of each window. With the static FC as the background, we can compare the static FC and the time-varying FC to remove the time-invariant properties of the window FC and emphasize its time-specific properties. Therefore, we can eliminate the continuous anatomical and physiological constraints on the related structures (DFN) by extracting only the dynamic activity of the FC.

[0035] In a preferred embodiment, step S4 can be further subdivided into the following steps:

[0036] S41. An example application for calculating the linear correlation between paired time series using Pearson correlation, and calculating the functional connections between activity and background.

[0037] S42, From the paired discrete non-stationary time series X={x i} i＝1，2，...，n and Y = {y i} i＝1，2，...，n To begin, calculate the static and time-varying FC, as follows:

[0038] Static FC is calculated using the Pearson correlation between the overall X and Y:

[0039]

[0040] Where n is the number of time points. and Let X and Y be the average values ​​of X and Y, respectively. To simplify the calculation, the time series X and Y are normalized using z-scores with a mean of 0 and a standard deviation (std) of 1. Therefore, the equation becomes:

[0041]

[0042] S43. Applying the sliding window method, with a window length of w and a step size of s, since brain activity is dynamic, the window time series does not follow the second-order stationarity assumption, meaning its distribution is not constant over time. Therefore, the time-varying FC of each time window is expressed as:

[0043]

[0044] Where x i，t and y i，t It is the i-th time point within the t-th time window, w is the number of time points within the time window, and x is the number of time points within the time window. t and y t The mean is expressed as x t and y t ;

[0045] S44. The first and second equations are used to analyze different types of correlations. The static FC calculates the overall correlation state over the entire period, ignoring time-related fluctuations. Because it is used to extract the time-invariant properties of statistical correlation, it is considered the background of FC. As a linear correlation, the second equation uses the same segmentation strategy as the time-varying FC for segmentation. The functional connectivity background of the time window t is:

[0046]

[0047] Where x i，t and y i，t It is the i-th time point within the t-th time window, the same as in equation (3). For each time window, the deviation level from the background to the specific time state is calculated, unaffected by the correlation strength:

[0048]

[0049] Where Ac(t) is the AFC value at time t, and when the time-specific correlation is related to its background r win (t)=r back When there is no difference between times t, the value of Ac(t) is 0; otherwise, the value of t is 0. win (t) and r back The difference between Ac(t) is relatively large. The larger the value of Ac(t), the more the functional connection is activated from the background state at time t.

[0050] In a preferred embodiment, step S5 can be further subdivided into the following steps:

[0051] S51. Activation networks use AFC (Activated Flow Coefficient) between various functional systems to simulate brain activity. It is a spatiotemporal structure (node ​​* node * time) represented as a set of nodes and their paired connections, where nodes are ROIs and AFC values ​​measure connections. This method can be used to study the distribution and temporal evolution of connections activated by different levels of brain activity. Activation networks measure the basic activity of functional structures (DFNs), considering high and low activity levels. High activation networks (HAN) and low activation networks (LAN) are constructed using connection sets with the highest and lowest AFC values ​​to represent different patterns in the communication process.

[0052] S52. Using both HAN and LAN, a sliding window method with a window length of 30 TRs and a step size of 3 TRs is used to ensure that AFC has good temporal resolution while maintaining statistical reliability. Then, the activation network is constructed using the obtained AFC values. In order to construct HAN and LAN, sparsity of 10% (ASD) and 25% (COVID-19) is applied to each time window of the activation network to include connections with high or low activation values. DFN is also constructed using time-varying FC for comparison, and the same sparsity is applied to the corresponding dataset.

[0053] In a preferred embodiment, step S6 can be further subdivided into the following steps:

[0054] S61. Graph theory provides a mathematical framework for quantitatively measuring the complex topology of brain networks. The obtained adjacency matrices (activation networks and DFNs) are binarized to represent the presence or absence of connections, and graph theory is used to evaluate the topological reorganization of the network.

[0055] S6101. Specifically, in the above steps, the average clustering coefficient (C), feature path length (L), local efficiency (E1), and global efficiency (Eg) are used as parameters of spatial structure. At the same time, C and E1 measure the information transmission capability within a local range, while L and Eg measure the information transmission capability within a global range. All of these are measurements of the brain's function of local isolation and global integration.

[0056] The S62, ASD, and COVID-19 datasets are classified based on cross-topic graph attributes (C, L, E1, Eg) extracted from HAN, LAN, and DFN. Graph attributes are extracted from each time window and then mixed for each topic to achieve cross-topic classification. The classification performance of the three groups is compared: activation networks (HAN and LAN), DFN, and their combinations. Feature selection is performed on each group before each classification.

[0057] S6201. First, the independent t-test excluded features that did not differ significantly between patients and healthy controls, and a relatively soft significance threshold of p < 0.15 was used to ensure that discriminative features were included.

[0058] S6202. Then, LASSO regression is applied to select the most discriminative features for classification. After feature selection, four mainstream classification strategies are used for 10-fold classification, including Support Vector Machine (SVM), Random Forest (RF), AdaBoost, and Naive Bayes (NB). Linear kernels are applied in the Support Vector Machine, and Sequential Minimum Optimization (SMO) is used as the learning method. The optimal c is determined by... -4 , ..., 10 -1 The parameter space is selected through cross-validation.

[0059] In a preferred embodiment, step S7 can be further subdivided into the following steps:

[0060] S71. Paired t-tests are used to compare the correlation between time-varying FC and background, as well as the differences between AFC and ΔFC. Independent t-tests are used to compare the temporal similarity of network structures between activation networks and DFN, respectively.

[0061] S72. To visualize the different temporal reorganization processes of different network structures, independent t-tests were applied to compare the graph parameters (C and L) between HAN, LAN, and DFN, respectively. Redundancy in network communication efficiency was measured by disrupting the network structure and removing inter- or intra-regional connections. Then, independent t-tests were used to compare the damage levels calculated for C and L on HAN, LAN, and DFN. To select the structure most sensitive to changes in psychological state, the graph parameters (C, L, E1, and Eg) for each structure's HAN, LAN, and DFN were calculated. An independent t-test was used to compare the differences in graph parameters between patient and healthy control datasets, with a significance level of p < 0.05. Multiple comparisons were corrected using the false discovery rate (FDR) at q = 0.05, where all p-values ​​were calculated as two-tailed p-values.

[0062] In a preferred embodiment, after obtaining the results in step S8, the data needs to be compared and analyzed multiple times.

[0063] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0064] This invention proposes the latent active structure (LAS) of DFN to describe the dynamic reorganization of brain networks. The temporal stability of DFN limits current research. We apply AFC to extract time-specific properties of statistical correlation and use computed non-correlation dependencies as connections to compose activation networks. Experimental results show that AFC exhibits a high correlation with latent dynamics. Furthermore, activation networks show more multi-regional interactions spatially and are highly sensitive to changes in mental state. It temporarily reveals more information about large-scale brain network communication with stable communication effects and low cost. The proposed method was validated through an application for disease classification. Activation networks are an important framework for understanding brain networks. This research provides new insights into brain network construction and highlights the potential of using activation networks in network neuroscience research. Attached Figure Description

[0065] Figure 1 This invention provides a verification diagram of the dynamic detection capabilities of FC and AFC in a spatiotemporal modeling method that improves the human brain communication process using an activation network.

[0066] Figure 2 This invention provides a temporal reconstruction diagram of activation networks in ASD and COVID-19, which improves the spatiotemporal modeling method of human brain communication processes.

[0067] Figure 3 This invention provides an average spatial structure of HAN, LAN, and DFN for an improved spatiotemporal modeling method of human brain communication processes, and their connectivity distribution in ASD and COVID-19.

[0068] Figure 4 This invention provides a spatiotemporal modeling method for improving human brain communication processes by activating networks. The diagram illustrates the damage of (a) HAN, (b) LAN, and (c) DFN under different types of attacks in ASD and COVID-19. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0070] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0071] Example

[0072] An activation network-based spatiotemporal modeling method for human brain communication processes includes the following steps:

[0073] S1. Integrate data;

[0074] S11. First, integrate simulated resting-state BOLD-FMRI data;

[0075] S1101. By using simulated resting-state BOLD-fMRI data, the “ground reality” of the time series can be controlled, which can simulate the activity and background properties of functional connectivity to evaluate the dynamic detection capability of the proposed method.

[0076] S1102. The enhancement and reduction levels of background correlation affected by dynamics (ΔFC) are measured to express the activity of functional connectivity. The simulated BOLD-fMRI is generated by a multivariate Gaussian process and a first-order vector autoregression (VAR) model to simulate background and dynamic data, respectively. The background data mainly provides the stability of FC.

[0077] S1103. Generate background data using a pairwise zero-mean multivariate Gaussian process, σ = (σ1, ..., σ2). n The background correlation is represented by the covariance moment ∑ randomly generated between -1 and 1, and then the dynamics of the simulation data are specified.

[0078] S1104. Background correlation is expected to be stochastically influenced by dynamics to reflect the persistent nature of resting-state BOLD-fMRI fluctuations. Dynamic data are estimated in pairs based on a first-order VAR model.

[0079] ε t =Aε t-1 +e t Where A = 0.8 is the autocorrelation coefficient, e t The model residuals, determined through a random Gaussian process, have a mean of 0.2 and a standard deviation (std) of 0.12.

[0080] S1104. Finally, the simulated BOLD-fMRI data is a linear combination of background and dynamic data:

[0081] V=σ+ε t The proposed method was validated by correlating it with ΔFC, generating 5000 data points with a length of 3000 points to ensure the reliability of the output;

[0082] S12. Secondly, we integrate autism spectrum disorder data, also known as ASD data. The ASD data used in this application comes from the Autism Brain Imaging Data Exchange (ABIDE). To eliminate the influence of multiple recording sites, we selected resting-state fMRI data of children from the University of Michigan site. Detailed information can be found at http: / / fcon_1000.projects.nitrc.org / indi / abide / .

[0083] S13. Finally, consolidate the COVID-19 data. For detailed information on this data, please visit https: / / www.ukbiobank.ac.uk.

[0084] S14. Combining the above three steps, obtain the required data;

[0085] S2, Data Preprocessing;

[0086] S21. The DPABI v3.1 toolbox (reference URL: http: / / rfmri.org / DPABI) can be used for data preprocessing of resting-state functional volumes. This involves formal transformation by the Montreal Neuroscience Institute, slice timing, head motion correction, spatial normalization (MNI), resampling resolution of 3×3×3 spatial smoothing (6×6×6 at half maximum), linear detrending to reduce low-frequency drift and physiological high-frequency respiratory and cardiac noise, and time bandpass filtering (0.01-0.1Hz). The Friston 24 parameter model is used to regress head motion effects.

[0087] S22. Motion rotations with a maximum translation of ≥1.5 and / or ≥1.5 degrees in the X, Y or Z directions of the participants were removed. Linear regression was applied to remove global mean signal, white matter and cerebrospinal fluid signal. The registered fMRI volume was partitioned using the Dosenbach 160 region of interest (ROI) template.

[0088] S3, the activities and background of the functional connections;

[0089] S31. The communication dynamic and non-dynamic properties of FC are distinguished by the Activity Functional Connection (AFC) and Background Functional Connection (BFC). Paired time series and sliding window methods are used to identify the time invariance and time specificity of FC in the communication process, where the background represents the time invariance property of FC across the time window.

[0090] S32. By subtracting the background AFC from the total signal, it measures the degree to which the activity level changes from a general state to a specific time state.

[0091] S33. Calculate the statistical correlation of paired time series within the complete scan period to obtain the static FC, which is used to extract the time-invariant properties of the FC. In order to prevent ignoring the detailed communication patterns that occur at each time point, replace the static FC with the time-varying FC. It divides the entire time series into sliding windows and calculates the statistical correlation of each window. With the static FC as the background, we can compare the static FC and the time-varying FC to remove the time-invariant properties of the window FC and emphasize its time-specific properties. Therefore, we can eliminate the continuous anatomical and physiological constraints on the related structures (DFN) by extracting only the dynamic activity of the FC.

[0092] S4, Functional Connection Computing Activity;

[0093] S41. An example application for calculating the linear correlation between paired time series using Pearson correlation, and calculating the functional connections between activity and background.

[0094] S42, From the paired discrete non-stationary time series X={x i} i＝1，2，...，n and Y = {y i} i＝1，2，...，n To begin, calculate the static and time-varying FC, as follows:

[0095] Static FC is calculated using the Pearson correlation between the overall X and Y:

[0096]

[0097] Where n is the number of time points. and Let X and Y be the average values ​​of X and Y, respectively. To simplify the calculation, the time series X and Y are normalized using z-scores with a mean of 0 and a standard deviation (std) of 1. Therefore, the equation becomes:

[0098]

[0099] S43. Applying the sliding window method, with a window length of w and a step size of s, since brain activity is dynamic, the window time series does not follow the second-order stationarity assumption, meaning its distribution is not constant over time. Therefore, the time-varying FC of each time window is expressed as:

[0100]

[0101] Where x i，t and y i，t It is the i-th time point within the t-th time window, w is the number of time points within the time window, and x is the number of time points within the time window. t and y t The mean is expressed as x t and y t ;

[0102] S44. The first and second equations are used to analyze different types of correlations. Static FC calculates the overall correlation state over the entire period, ignoring time-related fluctuations. Because it is used to extract the time-invariant properties of statistical correlation, it is considered the background of FC. As a linear correlation, the second equation uses the same segmentation strategy as time-varying FC for segmentation. The functional connectivity background of the time window t is:

[0103]

[0104] Where x i，t and y i，t It is the i-th time point within the t-th time window. Similar to equation (3), for each time window, the deviation level from the background to the specific time state is calculated, unaffected by the correlation strength:

[0105]

[0106] Where Ac(t) is the AFC value at time t, and when the time-specific correlation is related to its background r win (t)=r back When there is no difference between times t, the value of Ac(t) is 0; otherwise, the value of t is 0. win (t) and r back The difference between Ac(t) is large. The larger the value of Ac(t), the more the functional connection is activated from the background state at time t.

[0107] S5, High and Low Activation Network Construction;

[0108] S51. Activation networks use AFC (Activated Flow Coefficient) between various functional systems to simulate brain activity. It is a spatiotemporal structure (node ​​* node * time) represented as a set of nodes and their paired connections, where nodes are ROIs and AFC values ​​measure connections. This method can be used to study the distribution and temporal evolution of connections activated by different levels of brain activity. Activation networks measure the basic activity of functional structures (DFNs), considering high and low activity levels. High activation networks (HAN) and low activation networks (LAN) are constructed using connection sets with the highest and lowest AFC values ​​to represent different patterns in the communication process.

[0109] S52. Using both HAN and LAN, a sliding window method with a window length of 30 TRs and a step size of 3 TRs is used to ensure that AFC has good temporal resolution while maintaining statistical reliability. Then, the activation network is constructed using the obtained AFC values. In order to construct HAN and LAN, sparsity of 10% (ASD) and 25% (COVID-19) is applied to each slide of the activation network to include connections with high or low activation values. DFN is also constructed using time-varying FC for comparison, and the same sparsity is applied to the corresponding dataset.

[0110] S6. Perform graph theory analysis and classification;

[0111] S61. Graph theory provides a mathematical framework for quantitatively measuring the complex topology of brain networks. The obtained adjacency matrices (activation networks and DFNs) are binarized to represent the presence or absence of connections, and graph theory is used to evaluate the topological reorganization of the network.

[0112] S6101. Specifically, in the above steps, the average clustering coefficient (C), feature path length (L), local efficiency (E1), and global efficiency (Eg) are used as parameters of spatial structure. At the same time, C and E1 measure the information transmission capability within a local range, while L and Eg measure the information transmission capability within a global range. All of these are measurements of the brain's function of local isolation and global integration.

[0113] The S62, ASD, and COVID-19 datasets are classified based on cross-topic graph attributes (C, L, E1, Eg) extracted from HAN, LAN, and DFN. Graph attributes are extracted from each time window and then mixed for each topic to achieve cross-topic classification. The classification performance of the three groups is compared: activation networks (HAN and LAN), DFN, and their combinations. Feature selection is performed on each group before each classification.

[0114] S6201. First, the independent t-test excluded features that did not differ significantly between patients and healthy controls, and a relatively soft significance threshold of p < 0.15 was used to ensure that discriminative features were included.

[0115] S6202. Then, LASSO regression is applied to select the most discriminative features for classification. After feature selection, four mainstream classification strategies are used for 10-fold classification, including Support Vector Machine (SVM), Random Forest (RF), AdaBoost, and Naive Bayes (NB). Linear kernels are applied in the Support Vector Machine, and Sequential Minimum Optimization (SMO) is used as the learning method. The optimal c is determined by... -4 , ..., 10 -1 The parameter space is selected through cross-validation.

[0116] S7. Statistical tests;

[0117] S71. Paired t-tests are used to compare the correlation between time-varying FC and background, as well as the differences between AFC and ΔFC. Independent t-tests are used to compare the temporal similarity of network structures between activation networks and DFN, respectively.

[0118] S72. To visualize the different time-series reorganization processes of different network structures, independent t-tests were applied to compare the graph parameters (C and L) between HAN, LAN, and DFN, respectively. Redundancy in network communication efficiency was measured by disrupting the network structure and removing inter- or intra-regional connections. Then, independent t-tests were used to compare the damage levels calculated for C and L on HAN, LAN, and DFN. To select the structure most sensitive to changes in mental state, the graph parameters (C, L, E1, and Eg) for each structure's HAN, LAN, and DFN were calculated. An independent t-test was used to compare the differences in graph parameters between patient and healthy control datasets, with a significance level of p < 0.05. Multiple comparisons were corrected using the false discovery rate (FDR) at q = 0.05, where all p-values ​​were calculated as two-tailed p-values.

[0119] S8. To obtain the results, I applied the AFC method to simulate resting-state BOLD-fMRI data to validate its performance using known "foundational facts." Since BOLD-fMRI signal recordings are based on spontaneous neural fluctuations, we used simulated dynamic and background data to model this process, setting the window length and step size to 30 and 30 points, respectively. After combining the dynamic and background data, we measured the dynamics of functional connectivity by the level of change in background correlation (ΔFC), and then validated the performance of the proposed method by correlating it with ΔFC. Figure 1 As shown, Figure 1 a reflects a high correlation between FC and the background of 5000 samples (r=1, p=0.950). Figure 1 b depicts a randomly selected sample for illustration. These results show that FC is more informative of background correlation than dynamic correlation, and the trajectory of time-varying FC is largely influenced by the component providing the dominant dependency. Therefore, FC may not be suitable for directly detecting neurodynamics. We then calculate AFC and ΔFC to validate the communication dynamics detection capability of the proposed method. The results show that AFC and ΔFC exhibit significant differences (t... 4999 =127.674, p<0.001; Figure 1 c), however Figure 1 d also indicates a significant positive correlation between them (r = 0.966, p < 0.001), indicating that the proposed AFC is sensitive to dynamic fluctuations under the influence of background correlation. Furthermore, these findings suggest that AFC is suitable for constructing the dynamic structure of the human brain.

[0120] AFC and time-varying FC were applied to resting-state fMRI data to construct activation networks and DFNs. Communication patterns in the human brain lead to complex structural reorganization processes. Dynamic networks must have structures capable of capturing rich information about reorganization throughout the entire time series. To assess this, the spatial correlation between the temporal mean and windowed fully connected matrix of each participant was compared. Lower correlation indicates greater structural changes over time, suggesting richer information about structural reorganization. The spatial topology of the network within each time window was quantitatively evaluated using graphical parameters such as clustering coefficient (C) and characteristic path length (L). Healthy controls from two datasets were selected for analysis. For each dataset, the distribution of graphical attributes between the selected participants and time windows indicates the unique temporal reorganization process of each network. The temporal evolution of different network topologies was evaluated in terms of local isolation and global integration.

[0121] (a) Temporal similarity of AFC and FC matrices in ASD and (f) COVID-19. To assess the similarity of networks for each participant, we compared the spatial correlation between temporal averages and windowed fully connected matrices. The AFC matrix exhibits greater variability in the temporal domain and is more likely to capture the communicative dynamics of brain activity. (b, c) Probability distributions of graphical parameters of HAN, LAN, and DFN in ASD and (g, h) COVID-19. Graph parameters (C and L) are calculated from each time window to represent the one-off network topology. Mixed graph parameters among participants indicate the temporal reorganization process of the corresponding networks. (d, e) Box plots of graphical parameters of HAN, LAN, and DFN in ASD and (i, j) COVID-19. Combined with their probability distributions, HAN, LAN, and DFN show significant differences in temporal reorganization and network topology. Notably, activation networks and DFN were extracted to describe different aspects of brain activity. Each aspect contains key information; therefore, their graphical parameters cannot be directly compared to suggest which is better, and statistical results indicate significant differences (see reference). Figure 2 );

[0122] First, the AFC matrix provides more information on network changes in human brain communication than the DFN, and the windowed AFC matrix shows a significantly lower mean correlation with its time-averaged structure (ASD:

[0123] t 108 =-66.106, p<0.001, COVID-19: t 626=-84.604, p<0.001), in contrast, the structure of the DFN is relatively stable over time, consistent with previous findings. Therefore, the activation network extracted from the AFC exhibits richer network structure changes over time and has greater potential to extract the communication dynamics of brain activity. Secondly, as Figure 2 As shown in b, c, g, and h, HAN, LAN, and DFN exhibit different graph parameter distributions, indicating the unique evolutionary process of each network structure during the same period. Specifically... Figure 2 d, e, i, j show that the topologies of HAN, LAN, and DFN differ in terms of local isolation and global integration functions, with C and L showing significantly different graph parameters (p < 0.001, FDR-corrected). These results imply different topological organization patterns of activating networks and DFN during communication. In summary, activating networks describe richer structural changes in brain networks and exhibit different reorganization patterns in the temporal domain compared to DFN. However, to better understand the spatiotemporal characteristics of activating networks, we need to analyze their details and verify the reliability of the information provided by activating networks.

[0124] The spatial structure of the activation network and DFN was calculated using time-averaged networks. Healthy controls were selected from two datasets. The HAN, LAN, and DFN were averaged according to time windows and participants. To better illustrate the network structure, we summarized the connectivity proportions located in six specific subnetworks: the cingulate operculum (CON), sensorimotor network (SMN), occipital lobe network (ON), frontoparietal lobe network (FPN), default mode network (DMN), and cerebellar network (CN).

[0125] The network is the average of all participants from each dataset's health control, (ac, e.g.) the connection distribution of HAN, LAN, and DFN, i.e., the percentage of connections located within or between subnetworks out of the total connections. HAN and LAN are associated with specific functional subnetworks. HAN connections are mainly associated with SMN, and LAN connections are mainly associated with DMN. However, DFN is more about the predefined structure of the human brain, showing the balanced connection distribution of each subnetwork. In particular, HAN and LAN tend to communicate with other regions. (d, h) the proportion of inter-regional and intra-regional connections, which indicates that the activation network and DFN are different structures, focusing on opposite communication patterns (inter-regional or intra-regional communication). CON: occipital network, SMN: sensorimotor network, ON: occipital network, FPN: frontoparietal network, DMN: default mode network, CN: cerebellar network;

[0126] like Figure 3As shown, the spatial distributions of HAN, LAN, and DFN differ. HAN and LAN are associated with specific functional subnets: HAN connections are primarily associated with SMN (ASD: 11.635%; COVID-19: 6.761%), while LAN connections are primarily associated with DMN (ASD: 16.981%; COVID-19: 12.107%), indicating a high proportion of connections located in these regions. In contrast, DFN better conforms to the predefined structure of the human brain, showing a balanced connection distribution in each subnet. Figure 3 c, g), In addition, important subnets of HAN and LAN tend to communicate with other areas, with a high proportion of connections between SMN in HAN (ASD: 24.175%; COVID-19: 30.236%) and DMN in LAN (ASD: 25.590%; COVID-19: 30.236%) and other areas (COVID-19: 15.739%). Figure 3 As shown in d and h, HAN (ASD: 79.874%; COVID-19: 92.233%) and LAN (ASD: 74.057%; COVID-19: 64.780%) contain significant inter-regional connectivity, while DFN does not contain sufficient inter-regional connectivity (ASD: 29.088%; COVID-19: 51.258%). This suggests that the activation network and DFN are different structures, focusing on opposite communication patterns (inter-regional or intra-regional communication) of brain networks.

[0127] The spatial structure of HAN, LAN and DFN is evaluated based on natural selection criteria: stable communication efficiency and low cabling cost. To evaluate communication efficiency, we simulated the disruption to the network spatial structure by removing inter-area and intra-area connections respectively to determine the impact on communication efficiency.

[0128] Figure 4 The diagram shows the damage levels of HAN, LAN, and DFN under their respective damage types (e.g., removal of inter-area or intra-area connections). The subplots below show the correlation between the damage level and the proportion of damage to the most severely damaged network, with connections removed. (a, b, e, f) damage caused by the removal of inter-area connections on C and L, and (c, d, g, h) damage caused by the removal of inter-area connections on C and L. These results indicate that the active network structure has high redundancy in the face of damage, and the DFN structure is relatively vulnerable to both damage types. The active network demonstrates the importance of inter-area connections, as evidenced by the positive correlation between the degree of damage on C and the proportion of inter-area connections removed. The DFN demonstrates the importance of intra-area connections through the negative correlation between the degree of damage on L and the proportion of inter-area connections removed, and the positive correlation between the degree of damage on C and L and the proportion of inter-area connections removed.

[0129] like Figure 4 As shown, removing only the inter-area connectivity will affect both the HAN and LAN.

[0130] C (p < 0.001, FDR-corrected) caused more significant damage, indicating that local isolation has a significant impact on brain function. However, the DFN structure has lower damage redundancy because removing interregional connections significantly affects L (p < 0.001, FDR-corrected), while removing intraregional connections affects both C and L (p < 0.001, FDR-corrected). This suggests that the activated network structure has high redundancy and can maintain relatively stable communication efficiency during structural reorganization. Furthermore, it shows the importance of interregional connections in the activated network, as removing more interregional connections causes greater damage to C (ASD: r = 0.556, p < 0.001; COVID-19: r = 0.643, p < 0.001), although... Thus, the results of DFN highlight the importance of intra-regional connections. On the one hand, the impairment of L is negatively correlated with the proportion of inter-regional connections removed (ASD: r = -0.795, p < 0.001; COVID-19: r = -0.527, p < 0.001). On the other hand, the impairment of C and L is positively correlated with the proportion of intra-regional connections removed (C: ASD: r = 0.775, p < 0.001; COVID-19: r = 0.758, p < 0.001; L: ASD: r = 0.447, p < 0.001; COVID-19: r = 0.494, p < 0.001). This indicates that Activation Network and DFN focus on different aspects of the communication process by focusing on inter-regional and intra-regional connections. Simultaneously, the simulation results demonstrate that the fluctuation of FC is significantly constrained by background relevance. Figure 2 (a, b) Changing the DNF structure requires overcoming the influence of background relevance, which is costly. Experimental results show that its network structure exhibits high temporal relevance under real-world conditions. Figure 3 a, f), and prove that its structure is stable while maintaining stable communication efficiency. Therefore, the structure described by the activation network has relatively low wiring cost and stable communication efficiency to realize its time evolution.

[0131] The table below shows the graphical characteristics of HAN, LAN, and DFN between patients and healthy controls, highlighting statistically significant differences (p < 0.05). The data are presented as mean values:

[0132]

[0133] For practical applications, graph theory analysis is used to compare network performance to reveal changes in mental states. The HAN, LAN, and DFN of each participant are averaged along a time window and then used to calculate graph properties (clustering coefficient (C), characteristic path length (L), local efficiency (E1), and global efficiency (Eg)). Table 1 compares the differences in graph properties between healthy controls and ASD and COVID-19 patients. For ASD, HAN showed a significant difference.

[0134] C:t 108 =-2.572, p=0.011; L:t 108 =2.359, p=0.020; E1:t 108 =-2.799, p=0.006; Eg: t 108 =-2.513, p=0.013) Patients (C: 0.536±0.190; L: 2.020±0.081; E1: 0.677±0.165; Eg: 0.533±0.013, mean±std) and healthy controls (C: 0.609±0.092; L: 1.989±0.055; E1: 0.744±0.069; Eg: 0.538±0.009); DFN showed a significant decrease only between patients (0.652±0.034) and healthy controls (0.666±0.029) (t 108 =-2.450, p=0.016), for COVID-19, HAN still showed a significant difference (C:t). 611 =-2.033, p=0.043; L:t 611 =2.312, p=0.021; E1:t 611 =-2.090, p=0.037; Eg: t 611 =-1.996, p=0.046) in patients (C: 0.559±0.059; L: 1.754±0.006; E1: 0.778±0.031; Eg: 0.624±0.001) and healthy controls (C: 0.569±0.065; L: 1.753±0.004; E1: 0.783±0.034; Eg: 0.625±0.001), however, there was no significant difference between patients using DFN and healthy controls (p>0.05).

[0135] The classification accuracy for ASD and COVID-19 can be found in the table below:

[0136]

[0137] Where AN stands for Activation Network; DFN for Dynamic Functional Network; MIXED for Mixed Features of Activation Network and DFN; SVM for Support Vector Machine; RF for Random Forest; Note: Naive Bayes, the highest classification accuracy in each dataset is highlighted, and the data is expressed as mean (standard deviation);

[0138] A data-driven approach was used to evaluate the practical application of activation networks and DFNs. Within each time window, graph attributes (C, L, E1, Eg) were extracted as features for feature selection and classification. For better comparison, the features were divided into three groups based on network type: activation networks, DFNs, and combinations of both. Data-driven feature selection methods (independent t-tests and LASSO regression) were applied to extract discriminative features for each group. Finally, these features were validated using various machine learning classifiers (SVM, RF, Adaboost, NB). The SVM with mixed features achieved the best classification accuracy in both datasets (Table 2) (ASD: 88.182% ± 7.484%; COVID-19: 76.066% ± 4.247%). The selected features effectively distinguished different psychological states between patients and healthy controls. Notably, when using mixed features without adding new, non-redundant information, the classification accuracy did not significantly improve. In fact, compared to other features, the mixed features showed a significant improvement in classification accuracy (ASD: t-tests). 18 =-2.724, p=0.042; COVID-19: t 18 = -6.574, p < 0.001, FDR-corrected) DFN in both datasets. This improvement shows that activation networks contain key information representing other aspects of brain communication processes, and combining these two networks can provide a more comprehensive understanding of brain networks;

[0139] S81. The results were obtained by comparing and analyzing the data multiple times. The specific conclusions obtained in this application are as follows:

[0140] This study proposes a novel framework for separating active and contextual components in functional connectivity to distinguish the communication dynamics and non-dynamics of functional connections (FCs). To capture the topological reorganization of brain networks, we propose activation networks as a spatiotemporal structural framework. We first verify the dynamic detection capability of AFCs through simulation, and then apply activation networks to empirical datasets to reveal their advantages in extracting interregional connections between specific functional systems and establishing effective economic structures, thus providing richer information on structural reorganization. In addition, both HAN and LAN are crucial and play different roles in activation networks. HAN is particularly sensitive to changes in mental state, and when combined with LAN and DFN, they describe a more comprehensive picture of communication processes, which will be discussed in further detail below.

[0141] BOLD-fMRI data were simulated using a combination of predefined temporal correlation and first-order VAR models. We applied them to construct the background and dynamic components of the correlation, respectively. Various models were used to generate the dynamics of functional connectivity, such as large-scale dynamics models, neural field models, and neural network models. Since the underlying dynamics are unknown, the model selection is an arbitrary choice based on certain assumptions about physiological characteristics. Furthermore, it varies depending on the current problem. The current study is interested in the background and active components of functional connectivity. Experimental results show that the predefined temporal correlation and first-order VAR models are suitable for simulating the corresponding data in this study. Therefore, the selected simulation model is suitable for the current problem.

[0142] Simulation results show that the calculated time-varying functional connectivity (FC) mainly reflects the background correlation trajectory, and the estimated dynamic data has the least impact on the background correlation. Figure 1 (a) and (b) A key issue in practical applications is whether non-dynamic dependence dominates functional connectivity. If the detected BOLD-fMRI primarily reflects neural activity, then time-varying FCs can effectively capture communication processes, and any weak noise will not affect their dynamic detection capability. However, BOLD-fMRI signals are an indirect measurement of brain activity, and neural dynamics are not the only attribute involved. Therefore, the calculated statistical correlations are limited by the non-dynamic dependence of the time series, and their fluctuations can only be partly attributed to neural dynamics. Studies have shown that during slow-wave sleep and anesthesia, cognitive processes are absent or decreased, yet the topological structure of BOLD-fMRI correlations still exists, which supports this point. Furthermore, the network structure composed of FCs is very similar across task rest states and time, indicating that the predefined degree of dependence within FCs is insensitive to potential neural fluctuations and therefore cannot describe the complex communication processes during brain activity. Thus, FCs are dominated by their non-dynamic dependence. These findings support using dynamic attributes to assess connectivity rather than correlation strength. Figure 2 The simulation results of c and d demonstrate the ability of AFC to extract the true value of FC, indicating its reliability in dynamic modeling of neural activity;

[0143] Correlational activity measurements were applied to ASD and COVID-19 datasets, respectively. Brain dynamics lead to complex topological reorganization of brain networks. In the experiment, due to the limitation of temporal network topological changes, DFN struggled to describe the reorganization process. This observation is consistent with previous literature, highlighting the stable topology of brain activity under various task and rest conditions. Gratton et al. suggested that common organizational principles and personal characteristics dominate functional networks, while rest tasks and daily variations are not the main factors in functional networks. However, extracting underlying temporal information is fundamental to spatiotemporal structure and must be sensitive enough to describe the topological reorganization of communication processes. Experimental results show that, compared to DFN, activation networks exhibit richer variations and different reorganization patterns. Figure 2 This indicates that the activation network has the potential to extract dynamic reorganization of brain networks, and further exploration and verification of the reliability of the proposed method in terms of the spatial and temporal characteristics of the activation network;

[0144] Network architecture helps us understand, predict, and optimize the behavior of dynamic systems. We constructed the average spatial structures of HAN, LAN, and DFN. The results show that they focus on different communication patterns, such as... Figure 4 As shown, the DFN almost describes the outline of each functional system, and its connections are mainly located within the functional system, showing significant intra-regional connections. For the activation network, it eliminates the constraints of network structure and highlights the importance of inter-regional connections. Cole et al. believe that cognitive tasks lead to complex changes in the resting state functional network, increasing multi-regional interactions and reducing intra-network connectivity. In addition, HAN and LAN are associated with some specific functional subnetworks. HAN is more involved in inter-regional communication of SMN, while LAN is more involved in inter-regional communication of DMN. SMN is relatively more active in communication with others and is very important in coordinating with the sensory and behavioral functions of CON in the resting state. It is also related to top-down control of sensory areas. In contrast, the activity of DMN in the resting state seems to be significantly lower. This finding links the previous result of DMN inactivity to the functional needs of the human brain.

[0145] We also need to verify from the perspective of natural selection whether the structure of the activation network meets the basic requirements of most natural systems. According to this theory, dynamic large-scale brain communication is expected to exhibit stable communication efficiency and low wiring cost. In terms of wiring cost, as discussed in the simulation section, FC is not easy to modify due to the strong influence of non-dynamic dependencies. However, AFC breaks this limitation, isolating dynamic dependencies from FC and allowing relatively free fluctuations. Cajal's law describes the optimization principle of saving space, cytoplasm and conduction time in neural circuits. Based on the constrained structure, the operation optimization on DFN cannot rebuild its structure from scratch to obtain the optimum. Activation networks are not subject to this constraint, and the diversity of their temporal structure also proves their low wiring cost. Furthermore, stable communication efficiency is another aspect of natural selection and is used to analyze the impact of lesions on global and local communication capabilities. Activated networks exhibit higher redundant communication efficiency and highlight the importance of inter-regional connectivity. Their structure is highly adaptable and can achieve network reorganization during communication. In contrast, DFNs exhibit the opposite behavior. These findings are consistent with previous experimental results, indicating that DFN structures have a high degree of spatial similarity. The small-world characteristics of brain network topology are a balance between local isolation and global integration, which is crucial for maintaining optimal brain function. Redundant interactions during communication improve efficiency, robustness, and the ability to recover from brain damage. Considering the lower wiring cost of activated networks, natural selection in network construction highlights the importance of activated networks in the temporal domain.

[0146] After exploring the spatial and temporal activation characteristics, it is necessary to discuss the relationship between HAN and LAN. Although both describe the activity of brain networks, their graphical parameter distributions differ, indicating that each network has a unique structural reorganization process. Figure 2 Furthermore, activation networks exhibit stronger correlations with subnetworks of specific functions. HAN and LAN each have prominent functional systems, and the information they extract is unique. Therefore, the combination of HAN and LAN constitutes a complete view of activation networks. The application of HAN and LAN depends on the aspects we want to analyze: dynamic communication between highly active functional systems or neural activity-independent structures during communication.

[0147] To apply the activation network to real-world scenarios, we assessed its reliability by comparing network topologies between healthy controls and patients using graph properties. Experimental results showed that HAN revealed significant differences in network topologies between the two groups. In both ASD and COVID-19, HAN best described the decline in brain processing capacity (Table 1). The activation network was sensitive to changes in mental state. Furthermore, the classification results demonstrated the good performance of the activation network. Feature combinations across HAN, LAN, and DFN achieved the best classification performance in both ASD and COVID-19. In particular, the features of the activation network significantly improved the classification performance of DFN. This not only demonstrates the reliability of the activation network but also shows that both the activation network and DFN contain key and discriminative information about changes in mental state. In previous literature, the extensive contributions of DFN have demonstrated its ability to extract cognitive content. The activation network expands the content of brain networks, making their combination more suitable for revealing the characteristics of brain networks.

[0148] In explaining our proposed activation network, certain aspects of the research require special consideration. First, choosing the statistical method is a primary issue in activation network computation. Numerous model-based and model-free computational methods are available for calculating directional or non-directional connectivity, such as mutual information, coherence analysis, and transfer entropy. In BOLD-fMRI studies, Pearson correlation is the most commonly used method, and linear correlation is the simplest method for calculating time-invariant and time-specific properties. Furthermore, the activity and context of functional connectivity is a general idea to which various statistical correlation methods can be applied. Integrating their time-invariant and time-specific features is a key challenge in using activation networks, especially for nonlinear and directional correlation techniques, which is an area requiring further exploration. Second, activation networks do not describe functional networks but rather their activity in the time domain. Current research on dynamic brain networks mainly focuses on the temporal changes of DFNs based on statistical correlation. However, previous studies and our experimental results indicate that DFNs have limitations in extracting the dynamic reorganization of brain networks. We have expanded the perspective for constructing brain networks by including non-correlation dependencies in addition to statistical correlations.

[0149] This invention proposes the latent active structure (LAS) of DFN to describe the dynamic reorganization of brain networks. The temporal stability of DFN limits current research. We apply AFC to extract time-specific properties of statistical correlation and use computed non-correlation dependencies as connections to compose activation networks. Experimental results show that AFC exhibits a high correlation with latent dynamics. Furthermore, activation networks show more multi-regional interactions spatially and are highly sensitive to changes in mental state. It temporarily reveals more information about large-scale brain network communication with stable communication effects and low cost. The proposed method was validated through an application for disease classification. Activation networks are an important framework for understanding brain networks. This research provides new insights into brain network construction and highlights the potential of using activation networks in network neuroscience research.

[0150] Finally, it should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spatiotemporal modeling method for human brain communication processes improved by activating networks, characterized in that: Includes the following steps: S1. Integrate data; S2, Data Preprocessing; S3, the activities and background of the functional connections; S4, Functional Connection Computing Activity; S5, High and Low Activation Network Construction; S6. Perform graph theory analysis and classification; S7. Statistical tests; S8. Obtain the result.

2. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S1 can be further broken down into the following steps: S11. First, integrate simulated resting-state BOLD-FMRI data; S1101. By using simulated resting-state BOLD-fMRI data, the "ground reality" of the time series can be controlled, which can simulate the activity and background properties of functional connectivity to evaluate the dynamic detection capability of the proposed method. S1102. The enhancement and reduction levels of background correlation affected by dynamics (ΔFC) are measured to express the activity of functional connectivity. The simulated BOLD-fMRI is generated by a multivariate Gaussian process and a first-order vector autoregression (VAR) model to simulate background and dynamic data, respectively. The background data mainly provides the stability of FC. S1103. Generate background data using a pairwise zero-mean multivariate Gaussian process, σ = (σ1, ..., σ2). n The background correlation is represented by the covariance moment Σ, which is randomly generated between -1 and 1, and then the dynamics of the simulation data are specified. S1104. Background correlation is expected to be stochastically influenced by dynamics to reflect the persistent nature of resting-state BOLD-fMRI fluctuations. Dynamic data are estimated in pairs based on a first-order VAR model. ε t =Aε t-1 +e t Where A = 0.8 is the autocorrelation coefficient, e t The model residuals, determined through a random Gaussian process, have a mean of 0.2 and a standard deviation (std) of 0.

12. S1104. Finally, the simulated BOLD-fMRI data is a linear combination of background and dynamic data: V=σ+ε t The proposed method was validated by correlating it with ΔFC, generating 5000 data points with a length of 3000 points to ensure the reliability of the output; S12, Secondly, integrate autism spectrum disorder data; S13, Finally, integrate the COVID-19 data; S14. Combining the above three steps, obtain the required data.

3. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S2 can be further divided into the following steps: S21. Data preprocessing of resting functional volumes involves formal transformation by the Montreal Neuroscience Institute, slice timing, head motion correction, spatial normalization (MNI) space, resampling resolution of 3×3×3 spatial smoothing (full width at half maximum of 6×6×6), linear detrending to reduce low-frequency drift and physiological high-frequency respiratory and cardiac noise, and time bandpass filtering (0.01–0.1Hz). Friston24 parameter model is used to regress head motion effects. S22. Motion rotations with a maximum translation of ≥1.5 and / or ≥1.5 degrees in the X, Y, or Z directions of the participants were removed. Linear regression was applied to remove global mean signal, white matter, and cerebrospinal fluid signal. The registered fMRI volumes were partitioned using the Dosenbach 160 region of interest (ROI) template.

4. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S3 can be further divided into the following steps: S31. The communication dynamic and non-dynamic properties of FC are distinguished by the Activity Functional Connection (AFC) and Background Functional Connection (BFC). Paired time series and sliding window methods are used to identify the time invariance and time specificity of FC in the communication process, where the background represents the time invariance property of FC across the time window. S32. AFC is obtained by subtracting the background from the total signal. It measures the degree to which the activity level changes from a general state to a specific time state. S33. Calculate the statistical correlation of paired time series throughout the entire scanning period to obtain the static FC, which is used to extract the time-invariant properties of the FC. To prevent ignoring the detailed communication patterns that occur at each time step, replace the static FC with the time-varying FC. It divides the entire time series into sliding windows and calculates the statistical correlation of each window. With the static FC as the background, we can compare the static FC and the time-varying FC to remove the time-invariant properties of the window FC and emphasize its time-specific properties. Therefore, we can eliminate the continuous anatomical and physiological constraints on the related structures (DFN) by extracting only the dynamic activity of the FC.

5. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S4 can be further divided into the following steps: S41. An example application for calculating the linear correlation between paired time series using Pearson correlation, and calculating the functional connections between activity and background. S42, From the paired discrete non-stationary time series X={x i } i＝1，2，...，n and Y = {y i } i＝1，2，…，n To begin, calculate the static and time-varying FC, as follows: Static FC is calculated using the Pearson correlation between the overall X and Y: Where n is the number of time points. and Let X and Y be the average values ​​of X and Y, respectively. To simplify the calculation, the time series X and Y are normalized using z-scores with a mean of 0 and a standard deviation (std) of 1. Therefore, the equation becomes: S43. Applying the sliding window method, with a window length of w and a step size of s, since brain activity is dynamic, the window time series does not follow the second-order stationarity assumption, meaning its distribution is not constant over time. Therefore, the time-varying FC of each time window is expressed as: Where x i，t and y i，t It is the i-th time point within the t-th time window, w is the number of time points within the time window, and x is the number of time points within the time window. t and y t The mean is expressed as x t and y t ; S44. The first and second equations are used to analyze different types of correlations. The static FC calculates the overall correlation state over the entire period, ignoring time-related fluctuations. Because it is used to extract the time-invariant properties of statistical correlation, it is considered the background of FC. As for linear correlation, the second equation uses the same segmentation strategy as the time-varying FC for segmentation. The functional connectivity background of the time window t is: Where x i，t and y i，t It is the i-th time point within the t-th time window. Similar to equation (3), for each time window, the deviation level from the background to the specific time state is calculated, unaffected by the correlation strength: Where Ac(t) is the AFC value at time t, and when the time-specific correlation is related to its background r win (t)=r back When there is no difference between times t, the value of Ac(t) is 0; otherwise, the value of t is 0. win (t) and r back The difference between Ac(t) is relatively large. The larger the value of Ac(t), the more the functional connection is activated from the background state at time t.

6. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S5 can be further divided into the following steps: S51. Activation networks use AFC (Activated Functional Components) between various functional systems to simulate brain activity. It is a spatiotemporal structure (node ​​* node * time) represented as a set of nodes and their paired connections, where nodes are ROIs. Connections are measured by AFC values. This method can be used to study the distribution and temporal evolution of connections activated by different levels of brain activity. Activation networks measure the basic activity of functional structures (DFNs), considering high and low activity levels. High activation networks (HAN) and low activation networks (LAN) are constructed using connection sets with the highest and lowest AFC values ​​to represent different patterns in the communication process. S52. Using both HAN and LAN, a sliding window method with a window length of 30 TRs and a step size of 3 TRs is used to ensure that AFC has good temporal resolution while maintaining statistical reliability. Then, the activation network is constructed using the obtained AFC values. In order to construct HAN and LAN, sparsity of 10% (ASD) and 25% (COVID-19) is applied to each slide of the activation network to include connections with the top or bottom activation values. DFN is also constructed using time-varying FC for comparison, and the same sparsity is applied to the corresponding dataset.

7. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S6 can be further divided into the following steps: S61. Graph theory provides a mathematical framework for quantitatively measuring the complex topology of brain networks. The obtained adjacency matrices (activation networks and DFNs) are binarized to represent the presence or absence of connections, and graph theory is used to evaluate the topological reorganization of the network. S6101. Specifically, in the above steps, the average clustering coefficient (C), feature path length (L), local efficiency (E1), and global efficiency (Eg) are used as parameters of spatial structure. At the same time, C and E1 measure the information transmission capability within a local range, while L and Eg measure the information transmission capability within a global range. All of these are measurements of the brain's function of local isolation and global integration. The S62, ASD, and COVID-19 datasets are classified based on cross-subject attributes (C, L, E1, Eg) extracted from HAN, LAN, and DFN. Graph attributes are extracted from each time window and then blended for each topic to achieve cross-topic classification. The classification performance of the three groups is compared: activation networks (HAN and LAN), DFN, and their combinations. Feature selection is performed on each group before each classification. S6201. First, the independent t-test excluded features that did not differ significantly between patients and healthy controls, and a relatively lenient significance threshold of p < 0.15 was used to ensure that discriminative features were included. S6202. Then, LASSO regression is applied to select the most discriminative features for classification. After feature selection, four mainstream classification strategies are used for 10-fold classification, including Support Vector Machine (SVM), Random Forest (RF), AdaBoost, and Naive Bayes (NB). Linear kernels are applied in the Support Vector Machine, and Sequential Minimum Optimization (SMO) is used as the learning method. The optimal c is determined by... -4 , ..., 10 -1 The parameter space is selected through cross-validation.

8. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: Step S7 can be further divided into the following steps: S71. Paired t-tests are used to compare the correlation between time-varying FC and background, as well as the differences between AFC and ΔFC. Independent t-tests are used to compare the temporal similarity of network structures between activation networks and DFN, respectively. S72. To visualize the different temporal reorganization processes of different network structures, independent t-tests were applied to compare the graph parameters (C and L) between HAN, LAN, and DFN, respectively. Redundancy in network communication efficiency was measured by disrupting the network structure and removing inter- or intra-regional connections. Then, independent t-tests were used to compare the damage levels calculated for C and L on HAN, LAN, and DFN. To select the structure most sensitive to changes in psychological state, the graph parameters (C, L, E1, and Eg) for each structure's HAN, LAN, and DFN were calculated. An independent t-test was used to compare the differences in graph parameters between patient and healthy control datasets, with a significance level of p < 0.

05. Multiple comparisons were corrected using the false discovery rate (FDR) at q = 0.05, where all p-values ​​were calculated as two-tailed p-values.

9. The spatiotemporal modeling method for human brain communication processes improved by activation networks according to claim 1, characterized in that: After obtaining the results in step S8, the data needs to be compared and analyzed multiple times.

Images