Brain signal analysis system of pulse neural network based on global-local coupling and application

Abstract

The embodiment of the invention discloses a brain signal analysis system based on a global-local coupling spiking neural network and application, and relates to the technical field of biological information analysis, and the system comprises a graph embedding layer, a global-local coupling module, a spiking neural network feature extraction module and a jump enhancement output module which are connected in sequence. According to the embodiment of the invention, through the global-local coupling module, the discharge anomaly (microcosmic) of neurons in the functional brain region and the connection anomaly (macroscopic) of the neurons across the functional brain region are captured at the same time, and the two characteristics verify each other, so that the probability of missed judgment and misjudgment is greatly reduced; the brain signal analysis system based on the global-local coupled pulse neural network provided by the embodiment of the invention not only can output a classification result of whether a patient is ill or not, but also can reversely trace key pathological features.

Description

TECHNICAL FIELD

[0001] The embodiment of the application relates to the technical field of biological information analysis, and particularly relates to a brain signal analysis system based on a global-local coupled spiking neural network and application. BACKGROUND

[0002] Brain diseases are complex multi-level disorders, and the main feature is the dual changes of neurons and network levels. Studies have shown that the overactive neurons of the hippocampus and amygdala, combined with the reorganization of functional connections in the prefrontal cortex and default mode network, are likely to cause cognitive dysfunction. Since they mediate the integration of neural circuits supporting cognitive function, the connection patterns in these regions are particularly critical. In addition, stress-related disorders such as major depressive disorder (MDD) and schizophrenia have been proven to be related to amygdala neural activity. Therefore, abnormal activity of specific neuron groups can disrupt neural oscillations, and further lead to cognitive dysfunction. Spiking neural networks (SNNs) simulate the firing behavior of neurons in a biologically reasonable way, providing a unique opportunity to integrate connectivity analysis and population-level neural dynamics.

[0003] With the rapid development of neuroimaging technology, related research has gradually established a systematic brain network research paradigm based on neuroimaging. Functional magnetic resonance imaging (fMRI) has become a key data basis for building brain networks due to its unique whole-brain coverage and high spatial resolution characteristics. The complexity of modeling the topological relationship of the brain has stimulated the interest of researchers, which has given rise to brain network analysis methods based on graph neural networks (GNN). In addition, related research has also proposed methods such as adaptive brain region identification, information theory-driven connectivity selection, and long-range dependency modeling. However, these methods usually focus on modeling functional connection relationships, often ignoring the characterization of neural population dynamics in the brain.

[0004] Spiking neural networks (SNNs) achieve brain network modeling based on biological principles by simulating the dynamic process of biological neurons. In the field of brain cognitive modeling, three-dimensional spiking neural network architecture effectively promotes the analysis of spatio-temporal neural data with the help of biomimetic topological design. Through the spatio-temporal association memory mechanism in SNNs, researchers have achieved brain disease classification based on functional magnetic resonance imaging (fMRI) data. The biological inspired characteristics of SNNs reveal how mindfulness training can restructure brain networks and achieve data analysis of the coupling of brain structure and functional connections. In addition, SNNs based on fMRI topological constraints reveal the auditory coding mechanism from the perspective of information transmission efficiency. Anatomically guided spiking networks use brain anatomical features to guide SNNs to learn brain network characteristics autonomously. Despite these advances, current SNN frameworks still have limitations in brain network modeling, especially the lack of collaborative modeling capabilities to integrate cross-regional network topology and neuron population responses within brain regions. SUMMARY

[0005] This invention provides a brain signal analysis system and application based on a globally-locally coupled spiking neural network, which at least partially solves the above-mentioned problems.

[0006] The first aspect of the present invention provides a brain signal analysis system based on a global-local coupled spiking neural network, the system comprising: a graph embedding layer, a global-local coupling module, a spiking neural network feature extraction module, and a jump enhancement output module connected in sequence; The graph embedding layer is used to transform the feature matrix of functional brain region nodes and the adjacency matrix of functional brain region graph into feature vectors. The feature matrix of functional brain region nodes and the adjacency matrix of functional brain region graph are determined based on the user's resting-state fMRI data. The global-local coupling module is used to integrate parallel local paths and global paths to obtain global-local coupling features. The local paths process the feature matrix of the functional brain region nodes, and the global paths process the feature matrix of the functional brain region nodes and the adjacency matrix of the functional brain region map. The local paths are used to analyze the activity patterns within a single functional brain region and the local connectivity between brain regions, and the global paths are used to analyze the global connectivity between various functional brain regions and global brain region activity. The spiking neural network feature extraction module is used to capture the dependencies between features and strengthen the interactive associations through a dynamic attention weight mechanism, and to process the global-local coupling features output by the global-local coupling module. The jump enhancement output module is used to fuse the pre-pulse layer features and pulse layer binarized features output by the spiking neural network feature extraction module to output a binary classification result of healthy objects and patients with brain diseases. The jump enhancement output module uses a hybrid loss function to optimize the model parameters. The hybrid loss function is a weighted sum of cross-entropy loss and minimum error entropy loss.

[0007] Optionally, the local path includes an input layer, a liquid state machine layer, and a readout layer connected in sequence; The input layer is used to encode the feature matrix of functional brain region nodes into a pulse sequence and transmit it to the liquid state machine layer; The liquid state machine layer is used to process the pulse sequence and capture bio-inspired features. The liquid state machine layer is composed of sparsely randomly connected spiking neurons, which are LIF neurons. The ratio of excitatory to inhibitory neurons in the liquid state machine layer is 4:1. The readout layer is a linear layer used to decode the state of neurons within the liquid state machine layer; the local path is learned unsupervised through impulse temporal dependence plasticity.

[0008] Optionally, the calculation process of the liquid state machine layer includes:

[0009] in, For the input of the liquid state machine layer, It is a nonlinear dynamic system. To read the function, For time step Read out the layer output at any time; The LIF neurons process the pulse sequence based on the following formula:

[0010] in, For time step Presynaptic input, As a leakage factor, For weight parameters, The discharge threshold, It is a pulse output.

[0011] Optionally, the global path employs a pulse graph convolutional network with backpropagation to capture whole-brain dynamics and preserve overall network properties. The calculation process of the global path includes:

[0012] in, Given an adjacency matrix containing self-connected nodes, A This is the original adjacency matrix. I It is the identity matrix. for The degree matrix, To use a new node representation with connection aggregation, For the output of the global path, Represents the pulse layer. Represents the batch normalization layer; X For the input feature matrix, Linear(.) It is a linear transformation function.

[0013] Optionally, the core of the spiking neural network feature extraction module is a spiking self-attention mechanism, the calculation process of which includes:

[0014] in, Q(X), K(X), V(X) These are the query matrix, key matrix, and value matrix, respectively. Linearq(.), Lineark (.), Linearv(.) These are the corresponding linear transformation functions. SSA(X) For pulse self-attention output, MLP(.) denotes a multilayer perceptron; The overall calculation process of the spiking neural network feature extraction module includes:

[0015] in, l This refers to the number of layers in the feature extraction module of the spiking neural network. L This represents the total number of floors. Attn l For the attention output of layer l, O l This is the final output of the l-th layer. X l This is the input for the l-th layer.

[0016] Optionally, the feature fusion process of the jump enhancement output module includes: ; in, O final For the final output of the model, O l The features output by the spiking neural network feature extraction module are the pre-spiking features. SN final (.) represents the final pulse layer. SN final ( O l () represents the binarization feature of the pulse layer.

[0017] Optionally, the calculation process of the hybrid loss function includes:

[0018] Where LTotal is the total loss, LCE is the cross-entropy loss, LMEE is the minimum error entropy loss, and λ is the weight coefficient; The calculation process for the minimum error entropy loss includes: ; in, N For the sample size, G σ (.) Let be a Gaussian kernel function with bandwidth σ. Represents the second-order Rényi entropy Indicates the first i Labels of each sample With the model for the first i Predicted output for each sample The error between; The calculation process for the cross-entropy loss includes: ; in, For the target distribution, To predict the distribution, For network mapping, For category The network output, For input data, These are learnable network parameters.

[0019] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the brain signal analysis system based on a global-local coupled spiking neural network as described in the first aspect of the present invention.

[0020] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the brain signal analysis system based on a global-local coupled spiking neural network as described in the first aspect of the present invention.

[0021] A fourth aspect of the present invention provides a computer program product, including a computer program / instructions, which is implemented by a processor as the brain signal analysis system based on a globally-locally coupled spiking neural network as described in the first aspect of the present invention.

[0022] Traditional methods focus only on changes in connectivity between brain regions (global) or on local neuronal activity (local), easily overlooking crucial pathological information. This invention, through a global-local coupling module, simultaneously captures abnormal neuronal firing within functional brain regions (microscopic) and abnormal connectivity across functional brain regions (macroscopic). These two features corroborate each other, significantly reducing the probability of missed or false diagnoses.

[0023] In clinical data, individual patient differences (such as age and underlying diseases) and equipment differences between different testing centers can generate a large number of "outliers" (such as brain activity data of a patient deviating from the normal range). The proposed solution uses a hybrid design of "cross-entropy loss + minimum error entropy loss" to ensure the classification accuracy of normal samples while weakening the interference of outliers on the model through a Gaussian kernel function. The impact of outliers on gradient updates approaches zero, avoiding the model from being interfered with by extreme data and making the diagnostic results more reliable.

[0024] The brain signal analysis system based on a global-local coupled spiking neural network provided in this invention can not only output classification results of "whether or not one is ill," but also trace back key pathological features. Specifically, it can locate the functional brain regions with abnormal discharges through local pathways, and identify abnormal connectivity relationships between functional brain regions through global pathways (such as excessive connectivity in the hippocampus of autistic patients and abnormal association between the thalamus and hippocampus in patients with depression). These findings are consistent with the conclusions of existing clinical research, providing doctors with "visible pathological evidence" rather than the "black box decision-making" of traditional AI models. The identified key case features can assist doctors in further analyzing the association between the abnormal connectivity and the patient's cognitive impairment (such as inattention), providing direction for disease mechanism research and targeted therapy.

[0025] The technical solution provided in this invention uses clinically common resting-state fMRI data as input, without the need for additional special data collection (such as invasive EEG signals), and is compatible with existing hospital testing equipment. Simply converting conventional fMRI data into a "brain region feature matrix + functional connectivity matrix" allows access to this technical solution without modifying existing testing procedures, thus lowering the technical threshold and cost for clinical implementation.

[0026] The technical solutions provided by the embodiments of the present invention can cover a variety of brain diseases and have a wide range of applications. In the verification of autism, depression and schizophrenia, the embodiments of the present invention have shown superior performance compared to traditional methods, indicating that the system architecture design of the embodiments of the present invention has universality. In practical applications, there is no need to adjust the model structure separately for each disease, and it can be quickly extended to the diagnosis of other brain diseases (such as Alzheimer's disease), making it applicable to a wider range of scenarios. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention 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.

[0028] Figure 1 This is a structural block diagram of a brain signal analysis system based on a global-local coupled spiking neural network provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall framework of the brain signal analysis system based on a global-local coupled spiking neural network provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] Recent advances in SNN research have confirmed its ability to characterize static functional connectivity patterns. Building on this, the latest research extends the SNN framework to model dynamic functional and structural connectivity. However, its analytical paradigm is limited to changes in functional connectivity between brain regions, neglecting the representation of neural dynamics within brain regions, thus limiting the completeness of brain modeling. For example, the pathological changes in Alzheimer's disease often begin with microscopic disruption of synaptic plasticity in the hippocampus, while macroscopic functional connectivity abnormalities may not appear for several years.

[0031] It is evident that by simultaneously modeling the dynamic characteristics and functional connectivity differences of neuronal clusters within brain regions, brain changes can be captured from both microscopic and macroscopic perspectives, thereby providing more accurate biomarkers. However, existing SNN frameworks rarely integrate these two perspectives (i.e., network topology and neuronal population response) into a single computational model, which limits their ability to resolve multi-scale biomarkers.

[0032] Based on this, this invention proposes a brain signal analysis system based on a globally-locally coupled spiking neural network. This system simultaneously models cross-brain region connectivity differences and neuronal population firing rate deviations, enabling a dual-perspective analysis of neuropathology. Specifically, it provides two key perspectives: 1) disease sensitivity changes in inter-regional connectivity strength and topology; 2) abnormal firing rate responses of neuronal populations in specific brain regions. Thus, this invention dynamically links macroscopic connectivity patterns with microscopic-scale-inspired spiking population dynamics. The spiking neural network integrates a globally-locally coupled module, a spiking neural network feature extraction module, and a jump output strategy enhanced by a hybrid loss function.

[0033] The global-local coupling module simultaneously captures the spatiotemporal interaction between anomalous firing rates and anomalous brain connectivity. Specifically, the local path is based on a functional magnetic resonance imaging (fMRI)-guided local fluid state machine (LSM). It initializes a specific region's neuronal population through empirical fMRI connectivity, and then updates synaptic weights using pulse time-dependent plasticity (STDP) to simulate local plasticity and construct functional connectivity patterns. The global path optimizes cross-regional connectivity weights through backpropagation of an adaptive pulse graph convolutional network, capturing global topological changes and calculating neural firing dynamics. Furthermore, the jump-enhanced output module with a hybrid loss function constructs a more robust decision boundary by suppressing outlier interference (originating from inter-individual and inter-center differences).

[0034] Specifically, such as Figure 1As shown, it illustrates the structural block diagram of a brain signal analysis system based on a global-local coupling spiking neural network provided in an embodiment of the present invention. The system includes: a graph embedding layer 101, a global-local coupling module 102, a spiking neural network feature extraction module 103, and a jump enhancement output module 104 connected in sequence.

[0035] The graph embedding layer 101 is used to convert the feature matrix of functional brain region nodes and the adjacency matrix of functional brain region graph into feature vectors. The feature matrix of functional brain region nodes and the adjacency matrix of functional brain region graph are determined based on the user's resting-state fMRI data. The global-local coupling module 102 is used to integrate parallel local paths and global paths to obtain global-local coupling features. The local paths process the feature matrix of the functional brain region nodes, and the global paths process the feature matrix of the functional brain region nodes and the adjacency matrix of the functional brain region map. The local paths are used to analyze the activity patterns within a single functional brain region and the local connectivity between brain regions, and the global paths are used to analyze the global connectivity between various functional brain regions and global brain region activity. The spiking neural network feature extraction module 103 is used to capture the dependencies between features and strengthen the interactive associations through a dynamic attention weight mechanism, and to process the global-local coupling features output by the global-local coupling module. The jump enhancement output module 104 is used to fuse the pre-pulse layer features and pulse layer binarized features output by the spiking neural network feature extraction module to output the binary classification results of healthy objects and patients with brain diseases; the jump enhancement output module uses a hybrid loss function to optimize the model parameters, and the hybrid loss function is a weighted sum of cross-entropy loss and minimum error entropy loss.

[0036] In this embodiment of the invention, brain information medical images of the target object can be acquired to obtain resting-state fMRI data. The graph structure data derived from the resting-state fMRI data is further processed to construct a functional brain region node feature matrix and a functional brain region graph adjacency matrix.

[0037] The functional brain region node feature matrix is ​​used to capture the activity intensity of each functional brain region. The functional brain region map adjacency matrix is ​​used to represent the connectivity relationships between functional brain regions.

[0038] Specifically, such as Figure 2 The diagram illustrates the overall framework of a brain signal analysis system based on a globally-locally coupled spiking neural network, as provided in an embodiment of the present invention. In this embodiment, during model training, it is possible to... fMRI data of the participants and Derive functional brain region structure samples from N participating subjects using brain regions (ROIs). Each sample Indicates the first Functional brain region map structures of each participating subject. (The text abruptly ends here, likely due to an incomplete sentence or a formatting error.) middle, It is a functional brain region map adjacency matrix that represents the connectivity relationships between functional brain regions. This represents the feature matrix of functional brain region nodes that capture the activity intensity of functional brain regions, and its corresponding labels. Used to distinguish between individuals with brain diseases and healthy individuals. In this embodiment of the invention, the model's subsequent training process aims to learn a mapping... And discover biomarkers during the process.

[0039] In this embodiment of the invention, the graph embedding layer is used to transform the feature matrix of functional brain region nodes and the adjacency matrix of functional brain region graph into feature vectors, and to establish a mapping relationship between the basic connectivity patterns of functional brain regions and higher-order pathological features.

[0040] In this embodiment of the invention, the global-local coupling module integrates locally constrained bio-inspired features with globally optimized graph convolutional representations through local and global paths, thereby achieving precise localization of abnormal brain regions and identification of pathology-related biomarkers. Specifically, the local path processes the feature matrix of the functional brain region nodes, initializes a specific region's neuronal population through random connections based on a liquid state machine (LSM), and then updates synaptic weights using pulse temporal dependent plasticity (STDP) to simulate local plasticity and construct functional connectivity patterns. The global path processes the feature matrix of the functional brain region nodes and the graph adjacency matrix of the functional brain region, optimizes cross-regional connection weights through backpropagation based on a globally adaptive pulse graph convolutional network, captures global topological changes, and calculates neural firing dynamics.

[0041] Based on this, the global-local coupling module can output dual-perspective features, achieving a synergistic fusion between microscopic neural dynamics and macroscopic network organization. Specifically, the local pathway focuses on observing details of local brain regions. It simulates the actual workings of neurons in the brain, capturing activity characteristics within small brain regions through "pulse signals" similar to neuronal firing. Functional connectivity patterns are constructed using these activity characteristics, thereby identifying abnormal activity patterns in nerve cells within a specific brain region, such as problems with local connections between brain regions. The global pathway focuses on the overall connectivity of the entire brain, analyzing the connection strength and patterns between different brain regions, such as whether signal transmission between the prefrontal cortex and hippocampus is smooth, and whether cooperation between the left and right hemispheres is normal. Through this global observation, the overall activity intensity of brain regions can be obtained, revealing whether firing frequencies are normal. The global and local pathways work simultaneously, ultimately combining their respective findings to examine both local details and overall connections, thus enabling a more accurate diagnosis of the presence of brain diseases.

[0042] Specifically, in the global-local coupling module 102, the local path includes an input layer, a liquid state machine layer, and a readout layer connected in sequence; The input layer is used to encode the feature matrix of functional brain region nodes into a pulse sequence and transmit it to the liquid state machine layer; The liquid state machine layer is used to process the pulse sequence and capture bio-inspired features. The liquid state machine layer is composed of sparsely randomly connected spiking neurons, which are LIF neurons. The ratio of excitatory to inhibitory neurons in the liquid state machine layer is 4:1. The readout layer is a linear layer used to decode the state of neurons within the liquid state machine layer; the local path is learned unsupervised through impulse temporal dependence plasticity.

[0043] In this embodiment of the invention, the local pathway is learned unsupervised via STDP. The local pathway mainly focuses on the details of neuronal activity in local brain regions (and brain regions with different functions), specifically including: The firing patterns of neuronal populations within a single brain region resemble the electrical signal transmission patterns of nerve cells; the strength and temporal relationships of local connections between neurons in a small brain region, and changes in the activity intensity of local brain regions (such as whether nerve cells in a certain area are overactive or underactive). These characteristics are closer to the actual workings of brain neurons and can reflect microscopic functional abnormalities in local brain regions, such as whether the collaboration of nerve cells in a specific area is disordered, thereby revealing functional connectivity abnormalities between brain regions.

[0044] In this embodiment of the invention, the input layer encodes external signals into pulse sequences and transmits them to the liquid state machine layer. The liquid state machine layer consists of sparsely randomly connected spiking neurons, maintaining a 4:1 ratio of excitatory to inhibitory neurons. This ratio is biologically plausible and simulates the organizational structure of cortical microcircuits, enabling better learning of the bio-inspired features of various brain regions. The readout layer, typically implemented as a linear layer, is used to decode the states of neurons within the liquid state machine layer.

[0045] Specifically, the calculation process of the liquid state machine layer includes:

[0046] in, For the input of the liquid state machine layer, It is a nonlinear dynamic system (i.e., a liquid state machine layer). To read the function, For time step Read out the layer output at any time.

[0047] In this embodiment of the invention, a nonlinear dynamic system This is achieved through LIF neurons; in this embodiment of the invention, all spiking neurons are LIF neurons.

[0048] The LIF neurons process the pulse sequence based on the following formula:

[0049] in, For time step Presynaptic input, As a leakage factor, For weight parameters, The discharge threshold, It is a pulse output.

[0050] In this embodiment of the invention, considering that local paths may lead to the loss of details in the original graph structure, a global path designed to preserve the overall network properties is introduced to mitigate this information loss. Specifically, in the global-local coupling module 102, the global path employs a pulse graph convolutional network with backpropagation to capture whole-brain dynamics and preserve the overall network properties. The calculation process of the global path includes:

[0051] in, Let A be the adjacency matrix containing self-connections, and I be the original adjacency matrix. for The degree matrix, To use a new node representation with connection aggregation, For the output of the global path, Represents the pulse layer. This represents the batch normalization layer; X is the input feature matrix, and Linear(.) is the linear transformation function.

[0052] In this embodiment of the invention, the global pathway primarily focuses on the global connectivity relationships and global brain activity across different regions of the entire brain. Specifically, this includes: the strength of functional connectivity between different brain regions (e.g., whether signal transmission between the prefrontal and temporal lobes is normal), the network structure patterns across the entire brain (e.g., which brain regions form closely cooperating networks), and the efficiency and coordination of cross-regional information transmission. These characteristics can reflect the macroscopic functional state of the brain as a whole system, such as whether "communication" between different brain regions is smooth, whether there are abnormalities in the overall network structure (e.g., enhancement or weakening of certain key connections), thereby detecting abnormalities in brain activity.

[0053] In simple terms, in this embodiment of the invention, the local path focuses on "the local connectivity between brain regions from the activity within a single brain region", while the global path focuses on "the global connections between brain regions and the global activity of brain regions". The combination of the two can comprehensively capture abnormal features of the brain from the microscopic to the macroscopic level.

[0054] In this embodiment of the invention, a spiking Transformer is used as the main network structure of the spiking neural network feature extraction module. Wherein, The core of the spiking neural network feature extraction module 103 is a spiking self-attention mechanism, the calculation process of which includes:

[0055] in, Q(X), K(X), V(X) These are the query matrix, key matrix, and value matrix, respectively. Linearq(.), Lineark (.), Linearv(.) These are the corresponding linear transformation functions. SSA(X) For pulse self-attention output, MLP(.) denotes a multilayer perceptron; In this embodiment of the invention, after determining the attention level, a short-circuit message aggregation scheme from feature to attention is adopted, followed by processing through another SNN-enhanced MLP module with short-circuiting. The overall calculation process of the spiking neural network feature extraction module includes: ; in, l This refers to the number of layers in the feature extraction module of the spiking neural network. L This represents the total number of floors. Attn l For the attention output of layer l, O l For the first l The final output of the layer,X l For the first l Layer input.

[0056] In the output strategy section, this embodiment of the invention considers that in practical applications, due to the discrete synaptic firing mechanism of synaptic neural networks, the step function has non-differentiable characteristics. Currently, the alternative gradient method is commonly used to address this challenge. However, as the network depth increases, the alternative gradient inevitably leads to approximation errors and gradient vanishing problems. Short-circuit mechanisms can usually be introduced to alleviate gradient vanishing and reduce information loss, but this can cause a conflict between the real-valued features of intermediate layers and the binarization of the neural network. To solve this problem, related research proposes setting a spike activation function at the beginning of each module to maintain local binarization, while forcibly setting a final spike layer at the end of the network to achieve global binarization, thereby ensuring that the network maintains its pure event-driven characteristics. However, binary spike output inevitably leads to information loss. To alleviate the information loss caused by binarization (especially its impact on classification performance), this embodiment of the invention proposes a jump-enhanced output strategy: mixing the features before the spike layer with the subsequent binarized features as the final output.

[0057] Specifically, in this embodiment of the invention, the feature fusion process of the jump enhancement output module 104 includes: ; in, O final For the final output of the model, O l The features output by the spiking neural network feature extraction module are the pre-spiking features. SN final (.) represents the final pulse layer. SN final ( O l () represents the binarization feature of the pulse layer.

[0058] The above feature fusion process can intuitively demonstrate the increase in the amount of information output by the network, and significantly alleviates information loss compared to the original binary pulse output.

[0059] In the loss function design section, this embodiment of the invention considers the significant individual differences between patients and healthy control groups, and the cross-institutional variability caused by differences in equipment and operating procedures further complicates the data. Traditional loss functions based on cross-entropy (CE) may lack robustness when handling such complex data distributions. To address this limitation, this embodiment of the invention introduces Minimum Error Entropy (MEE) as an additional optimization objective for the model. By minimizing the information entropy of the classification error, MEE effectively suppresses outlier interference caused by individual and site differences while reducing reliance on assumptions about a specific data distribution.

[0060] In cross-entropy loss, the optimization objective is essentially to measure the difference between two probability distributions, and the optimization process is performed independently, point-to-point. In contrast, MEE considers the statistical dependencies between error samples. Error information Quantification can be achieved through Rényi entropy:

[0061] in, Indicate target With model output The error between them It is the entropy order (usually taken as 2 to simplify calculation). The information potential energy is given by the following formula:

[0062]

[0063] in The bandwidth is Gaussian kernel function, It is an error The probability density function. Clearly, the criterion of minimizing error entropy... Equivalent to maximizing information potential energy Accordingly, the MEE loss function is defined in this embodiment of the invention as follows:

[0064] Therefore, in this embodiment of the invention, the calculation process of the hybrid loss function includes:

[0065] in, L Total For the total loss, L CE For cross-entropy loss, L MEE The minimum error entropy loss is represented by λ, where λ is the weighting coefficient. The calculation process for the minimum error entropy loss includes: ; in, N For the sample size, G σ (.) Let be a Gaussian kernel function with bandwidth σ. Represents the second-order Rényi entropy. Indicates the first i Labels of each sample With the model for the first i Predicted output for each sample The error between; The calculation process for the cross-entropy loss includes: ; in, For the target distribution, To predict the distribution, For network mapping, For category The network output, For input data, These are learnable network parameters.

[0066] To verify L MEE To ensure robustness, this invention provides a theoretical analysis of gradient perturbations caused by outliers. In determining... L CE Next, the gradient is calculated:

[0067] Similarly, L MEE The gradient is given by the following equation:

[0068] If there are outliers in the input, the corresponding network will produce abnormal output. L CE gradient and It exhibits a linear relationship, that is, when hour As outliers increase, their impact on parameter updates grows linearly, directly affecting gradient updates. In contrast, L MEE Gaussian kernel mapping causes the kernel function contribution to decrease as outliers increase: when hour This causes outliers to contribute close to zero to the gradient.

[0069] In this embodiment of the invention, verification embodiments are also provided to verify the above-mentioned brain signal analysis system based on global-local coupling spiking neural networks. Specifically, in this embodiment of the invention, three sets of brain network analysis-related fMRI datasets are used to evaluate the above-mentioned system: (1) the ABIDE dataset, which includes 528 patients with autism (ASD) and 571 healthy controls (HC); (2) the REST-meta-MDD dataset, which includes 848 patients with depression (MDD) and 794 healthy controls; and (3) the SRPBS dataset, which includes 92 patients with schizophrenia and 92 healthy controls.

[0070] In this embodiment of the invention, the control system is divided into two categories. The first category is artificial neural networks (ANNs), including generalized graph networks and dedicated brain networks. Specifically, these include: GCN, GAT, GIN, SIB, DIR-GNN, ProtGNN, BrainGNN, IBGNN, CI-GNN, BrainIB, ContrastPooL, and ALTER. The second category is spiking neural networks, including SpikeGCN and SpikeGT.

[0071] Experimental results show that the system provided by this invention significantly outperforms the two comparison systems on all three datasets. Specifically, compared with Artificial Neural Networks (ANNs), in general graph networks, the system of this invention achieves significant improvements in accuracy: 6.2% improvement on the REST-meta-MDD dataset, 5.1% improvement on the ABIDE dataset, and 9.2% improvement on the SRPBS dataset. Compared with specialized brain information neural networks, the system provided by this invention also performs excellently: accuracy improvements of 0.9%, 2.5%, and 3.2% on the REST-meta-MDD, ABIDE, and SRPBS datasets, respectively. Compared with Self-Organizing Neural Networks (SNNs), the system provided by this invention achieves improvements of 4.9%, 4.3%, and 4.8% on the REST-meta-MDD, ABIDE, and SRPBS datasets, respectively. Experimental results demonstrate that the system provided by this invention consistently outperforms existing methods on all evaluation datasets.

[0072] The brain signal analysis system based on a globally-locally coupled spiking neural network provided in this invention can detect the abnormal firing rate response of a specific group of neurons and identify brain regions in patients that exhibit abnormal neural activity compared to healthy control groups.

[0073] Specifically, to analyze neuronal population dynamics, this invention conducted a rigorous statistical comparison of brain region neural activity between patients and healthy controls. The results showed that the differences in synaptic activity between patients with brain diseases and healthy controls were statistically significant.

[0074] For patients with schizophrenia, the brain signal analysis system based on a globally-locally coupled spiking neural network provided in this invention has revealed abnormal activity in multiple brain regions, including the left prefrontal cortex (PreCG.L), left frontal cortex (MFG.L), left posterior frontal cortex (PCL.L), right cingulate gyrus (CAU.R), and left pons (MOG.L). These findings are consistent with previous studies, suggesting that abnormal activity in these areas may be a potential cause of schizophrenia. These regions are primarily involved in working memory maintenance, conflict monitoring, and goal-oriented behavior regulation.

[0075] For patients with autism, significant activity differences were observed in the left olfactory bulb, right insula, left cingulate gyrus, left cingulate nucleus, and right temporoparietal lobe. Further literature indicates that abnormalities in these areas are closely related to core symptoms of ASD. These areas are responsible for detecting internal sensory signals and coordinating the allocation of attentional resources.

[0076] For patients with major depressive disorder: significant activity changes were observed in the PCG.L, CAL.L, PoCG.L, PCUN.L, and CAU.L regions. Previous studies have confirmed that abnormal activity in these brain regions may contribute to the development of depression, and their abnormal processing mechanisms may lead to negative information bias.

[0077] Comparative analysis of three types of brain diseases revealed significantly different neural activity patterns in abnormal brain regions compared to healthy controls. These results demonstrate that the brain signal analysis system based on a globally-locally coupled spiking neural network provided in this invention can effectively simulate brain activity and identify brain regions exhibiting abnormal neural activity in patients compared to healthy controls, thus allowing these abnormally active brain regions to serve as biomarkers. Clinical and neuroimaging results also validate the rationale behind the biomarkers discovered in this invention.

[0078] To analyze functional patterns under different brain diseases, this embodiment of the invention also analyzed the connectivity of various functional brain regions in the patient group and the healthy control group. The results showed significant differences between the patient group and the healthy control group.

[0079] Experimental results show that the brain signal analysis system based on a global-local coupled spiking neural network provided in this embodiment of the invention successfully learned the unique connection patterns between schizophrenia patients and healthy control groups, and obtained functional connections that can distinguish between patients and control groups.

[0080] Specifically, regarding schizophrenia: functional connectivity between the right side of the frontoparietal cortex and the right side of the triangular prefrontal cortex (frontoparietal network) was significantly enhanced. This finding is consistent with previous studies that patients with schizophrenia have enhanced connectivity between the sensorimotor network and the frontoparietal network.

[0081] Regarding autism: Enhanced connectivity in the HIP.L region is consistent with previous studies revealing hippocampal dysfunction in autism.

[0082] For major depressive disorder: Increased connectivity between the THA.R and HIP.L regions is consistent with previous research findings – altered connectivity in this region is one of the pathophysiological features of major depressive disorder.

[0083] Experimental results show that the brain signal analysis system based on a globally-locally coupled spiking neural network provided in this invention can accurately reproduce clinically observed abnormal functional connectivity patterns, effectively simulate the physiological mechanisms of brain functional connectivity, and reveal connectivity patterns specific to brain diseases. Clinical and neuroimaging results also validate the rationality of the discovered biomarkers.

[0084] Therefore, the brain signal analysis system based on a globally-locally coupled spiking neural network provided in this invention achieves a dual-perspective analysis of neuropathology by jointly modeling differences in brain region connectivity and deviations in the population firing rate of neurons in brain regions. This establishes a robust decision-making method for the diagnosis of brain diseases and can identify dual-perspective biomarkers from abnormalities in functional brain region connectivity and neuronal firing rates, providing novel insights for clinical research and treatment strategies.

[0085] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing, implements the brain signal analysis system based on a global-local coupled spiking neural network as described in any of the above embodiments.

[0086] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the brain signal analysis system based on a global-local coupled spiking neural network as described in any of the above embodiments.

[0087] Based on the same inventive concept, embodiments of the present invention provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the brain signal analysis system based on a globally-locally coupled spiking neural network as described in any of the above embodiments.

[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable terminal device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions can also be loaded onto a computer or other programmable terminal device to cause a series of operational steps to be performed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0094] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0095] The above provides a detailed description of a brain signal analysis system and its application based on a global-local coupled spiking neural network provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A brain signal analysis system based on a globally-locally coupled spiking neural network, characterized by, The system comprises a graph embedding layer, a global-local coupling module, a spiking neural network feature extraction module and a jump enhancement output module connected in sequence. The graph embedding layer is configured to convert a functional brain region node feature matrix and a functional brain region graph adjacency matrix into a feature vector, the functional brain region node feature matrix and the functional brain region graph adjacency matrix being determined based on resting state fMRI data of a user. The global-local coupling module is configured to integrate a local path and a global path arranged in parallel to obtain a global-local coupling feature, the local path processing the functional brain region node feature matrix, and the global path processing the functional brain region node feature matrix and the functional brain region graph adjacency matrix, the local path being configured to analyze activity patterns within a single functional brain region and local connection relationships between brain regions, and the global path being configured to analyze global connection relationships between the functional brain regions and global brain region activities. The spiking neural network feature extraction module is configured to capture inter-feature dependencies and strengthen interaction correlations through a dynamic attention weight mechanism, and process the global-local coupling feature output by the global-local coupling module. The jump enhancement output module is configured to fuse a pre-spike layer feature and a spike layer binary feature output by the spiking neural network feature extraction module, and output a result of binary classification of a healthy subject and a brain disease patient, the jump enhancement output module adopting a hybrid loss function to optimize model parameters, the hybrid loss function being a weighted sum of a cross-entropy loss and a minimum error entropy loss.

2. The brain signal analysis system based on a globally-locally coupled spiking neural network according to claim 1, characterized in that, The local path comprises an input layer, a liquid state machine layer and a readout layer connected in sequence. The input layer is configured to encode the functional brain region node feature matrix into a pulse sequence and transmit the pulse sequence to the liquid state machine layer. The liquid state machine layer is configured to process the pulse sequence and capture biologically inspired features, the liquid state machine layer being composed of pulse neurons with sparse random connections, the pulse neurons being LIF neurons, and a ratio of excitatory neurons to inhibitory neurons in the liquid state machine layer being 4:

1. The readout layer adopts a linear layer to decode states of neurons in the liquid state machine layer, and the local path performs unsupervised learning through pulse temporal dependency plasticity.

3. The brain signal analysis system based on a globally-locally coupled spiking neural network according to claim 2, characterized in that, The calculation process of the liquid state machine layer comprises: wherein, is an input to the liquid state machine layer, is a nonlinear dynamical system, is a readout function, is a time step is a readout layer output; The LIF neuron processes the pulse sequence based on the following formula: ； wherein, is the presynaptic input at time step , is the leak factor, is the weight parameter, is the firing threshold, is the pulse output.

4. The global-local coupling based spiking neural network system for brain signal analysis according to claim 1, wherein, The global path adopts a pulse graph convolution network with back propagation to capture whole brain dynamics and preserve overall network properties, and a calculation process of the global path comprises: ； wherein, is a self-connected adjacency matrix, A is an original adjacency matrix, I is an identity matrix, is is a degree matrix, is a new node representation with aggregated connections, is an output of a global path, represents a pulse layer, represents a batch normalization layer; X is an input feature matrix, Linear(.) is a linear transformation function.

5. The global-local coupling based spiking neural network system for brain signal analysis according to claim 1, wherein, A core of the spiking neural network feature extraction module is a pulse self-attention mechanism, and a calculation process of the pulse self-attention mechanism comprises: ； wherein, Q(X), K(X), V(X) are a query matrix, a key matrix, a value matrix, respectively, Linearq(.), Lineark(.), Linearv(.) are corresponding linear transformation functions, respectively, SSA(X) is a pulse self-attention output, and MLP(.) denotes a multi-layer perceptron. An overall calculation process of the spiking neural network feature extraction module comprises: ； wherein, l is the number of layers of the pulse neural network feature extraction module, L is the total number of layers, Attn l is the l-th layer attention output, O l is the l-th layer final output, X l is the l-th layer input.

6. The global-local coupling based spiking neural network system for brain signal analysis according to claim 1, wherein, A feature fusion process of the jump enhancement output module comprises: ； wherein, O final is the final output of the model, O l is the pre-spike feature output by the spiking neural network feature extraction module, SN final (.) is the final spike layer, SN final ( O l ) is the binary spike layer.

7. The global-local coupling based spiking neural network system for brain signal analysis according to claim 1, wherein, The calculation process of the mixed loss function includes: wherein, L Total is the total loss, L CE is the cross-entropy loss, L MEE is the minimum error entropy loss, and λ is a weight coefficient. A calculation process of the minimum error entropy loss comprises: ； in, N For the sample size, G σ (.) Let be a Gaussian kernel function with bandwidth σ. Represents the second-order Rényi entropy. Indicates the first i Labels of each sample With the model for the first i Predicted output for each sample The error between; The calculation process of the cross-entropy loss comprises: ； wherein, is the target distribution, is the predicted distribution, is the network mapping, is the class of the network output, is the input data, is the learnable network parameter.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the brain signal analysis system based on the global-local coupled spiking neural network according to any one of claims 1-7 when executing the computer program.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the brain signal analysis system based on the global-local coupled spiking neural network according to any one of claims 1-7.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to implement the brain signal analysis system based on the global-local coupled spiking neural network according to any one of claims 1-7.

Images