A schizophrenia electroencephalogram data recognition method based on pulse neural network

Abstract

A kind of schizophrenia electroencephalogram data recognition method based on pulse neural network, collects the electroencephalogram data of schizophrenia patient and healthy control person;The original data are pretreated;Including signal filtering, artifact removal, signal segmentation and data standardization;Extract multi-scale space-time feature, including time-frequency feature, functional connection feature and spatial feature;Including time-frequency feature extraction, functional connection feature extraction and feature fusion three steps;Three-layer pulse neural network is constructed, including input layer, hidden layer and output layer;The connection model between brain partition based on pulse neural network is constructed;Improved back propagation algorithm is used to train network;The difference of brain partition connection mode of schizophrenia patient and healthy person is analyzed.The present application constructs the brain partition connection model of signal distinction of schizophrenia patient and healthy person by simulating the dynamic activity of brain neuron, realizes the improvement of schizophrenia classification accuracy.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and brain science, specifically to a method for recognizing EEG data of schizophrenia based on spiking neural networks. Background Technology

[0002] Schizophrenia is a serious mental illness with a complex pathogenesis. Its clinical manifestations involve disturbances in perception, thought, emotion, cognition, and behavior, severely impacting patients' daily lives and social interactions, and placing a heavy burden on patients' families and society. Currently, the diagnosis of schizophrenia mainly relies on clinical scale scores and physician experience, which is highly subjective and lacks objective diagnostic indicators, making misdiagnosis and missed diagnosis prone to occur.

[0003] In recent years, with the development of brain science and artificial intelligence technologies, electroencephalography (EEG), as a non-invasive neuroimaging technique, can record the electrical activity of neurons in the cerebral cortex, providing an important tool for studying the neural mechanisms of schizophrenia. Numerous studies have found abnormalities in the electroencephalogram (EEG) signals of schizophrenia patients, such as a significant increase in gamma-band activity in the brain at rest. Simultaneously, alterations in the topology and functional connectivity of brain networks are also observed. These abnormalities can serve as potential biomarkers for the early diagnosis of schizophrenia.

[0004] In the field of EEG data recognition for schizophrenia, there are already many related studies and patents. For example, the application "Method for Identifying Schizophrenia Based on EEG Signals Using an EEGNet Model" (application publication number: CN119632578 A) by Ren Weicong et al. improves the accuracy, timeliness, and objectivity of schizophrenia diagnosis by extracting the gamma band of EEG signals and then using the gamma band to train the EEGNet model. The application "Multi-view Classification Method for Schizophrenia Based on Dynamic Evidence Fusion Neural Network" (authorization announcement number: CN 117542503 B) by Ding Weiping et al. extracts the feature maps of each view of the dynamic brain network connection matrix of schizophrenia patients using edge-to-edge, edge-to-node, and node-to-graph convolutional filters, and improves classification accuracy by fusing multi-view information through evidence theory.

[0005] However, existing methods still have some limitations. Some studies only analyze specific frequency bands or features, failing to fully utilize the information in EEG signals; some models struggle to capture subtle changes and features when processing complex EEG data, resulting in room for improvement in diagnostic accuracy and reliability; furthermore, some methods are inefficient when processing large-scale data, failing to meet the needs of clinical applications.

[0006] Spiking Neural Networks (SNNs), as a third-generation artificial neural network, have neuron models that more closely resemble the characteristics of biological neurons. They can simulate the spiking mechanism and synaptic plasticity of neurons, giving them unique advantages in processing time-series data and simulating biological nervous systems. However, no research has yet applied spiking neural networks to the analysis of EEG data from schizophrenia, particularly in constructing models of brain region connectivity.

[0007] Therefore, developing a method for identifying schizophrenia EEG data based on spiking neural networks, and using spiking neural networks to simulate brain activity and construct a brain region connectivity model to distinguish signals between schizophrenia patients and healthy individuals, has significant theoretical and practical implications. This method is expected to fully extract information from EEG signals, improve the accuracy and reliability of schizophrenia diagnosis, and provide more effective support for clinical treatment. Summary of the Invention

[0008] This invention provides a method for identifying schizophrenia based on electroencephalogram (EEG) data from spiking neural networks. By simulating the dynamic activity of neurons in the brain, a brain region connectivity model is constructed to distinguish signals from those of schizophrenia patients and healthy individuals, thereby improving the accuracy of schizophrenia classification.

[0009] A method for identifying schizophrenia EEG data based on spiking neural networks includes the following steps:

[0010] Step S1: Data Acquisition and Preprocessing: First, EEG data were acquired from schizophrenia patients and healthy controls at a sampling rate of 1000Hz. A 128-lead EEG was used to record 10 minutes of resting state data. The raw data was preprocessed, including signal filtering (0.5-45Hz), artifact removal, signal segmentation, and data standardization, to obtain high-quality data segments.

[0011] Step S2: Feature extraction: Extract multi-scale spatiotemporal features from the preprocessed data, including time-frequency feature extraction, functional connectivity feature extraction, and feature fusion; time-frequency feature extraction uses wavelet transform, functional connectivity features are calculated using phase-locked values, and feature fusion maps electrodes to anatomical brain regions based on standard brain atlases.

[0012] Step S3: Constructing the Spiking Neural Network: Construct a three-layer spiking neural network, including an input layer, hidden layers, and an output layer. The input layer receives preprocessed EEG data, the hidden layer contains multiple spiking neurons to simulate neurons in the cerebral cortex, and the output layer is used for classification decisions.

[0013] Step S4: Brain Partition Connection Model Construction: Based on the spiking neural network, a connection model between brain partitions is constructed, taking into account the constraint of anatomical distance, and the connection weights between neurons are calculated using the improved STDP learning rule.

[0014] Step S5: Model Training and Optimization: Train a three-layer spiking neural network using an improved backpropagation algorithm, optimize the training process with adaptive learning rate and momentum term, and use an early stopping strategy to prevent overfitting.

[0015] Step S6: Classification: Use a trained three-layer spiking neural network to classify EEG data, analyze the differences in brain region connectivity patterns between schizophrenic patients and healthy individuals, and diagnose schizophrenia.

[0016] This invention proposes a method for constructing a brain region connectivity model based on a spiking neural network (SNN), which can effectively simulate the differences in brain connectivity patterns between schizophrenic patients and healthy individuals. The core innovations of this model include:

[0017] 1. Definition and mapping of brain regions

[0018] This invention first establishes a precise mapping relationship between a 128-lead EEG system and 90 AAL brain regions. Each EEG channel is assigned to the nearest AAL brain region based on its spatial location. The specific mapping process includes: obtaining the three-dimensional coordinates of the 128 electrodes using the standard electrode positions of the international 10-5 system; obtaining the centroid coordinates of the 90 brain regions using the standard centroid coordinates of the AAL map; uniformly converting the coordinates to MNI space; calculating the Euclidean distance between each EEG channel and the centroid of the AAL brain region; and assigning each EEG channel to the nearest AAL brain region.

[0019] The number of neurons was allocated to each AAL brain region based on its functional importance and anatomical volume. The prefrontal lobe regions (such as PreCG, SFGdor, MFG, etc.) were assigned a high functional weight of 1.5 due to their important role in schizophrenia; the temporal lobe regions (such as STG, MTG, ITG, etc.) had a weight of 1.3; the parietal lobe regions (such as SPG, IPL, etc.) had a weight of 1.2; the occipital lobe regions (such as SOG, MOG, IOG, etc.) had a weight of 1.0; the limbic system (such as HIP, AMYG, etc.) had a weight of 1.4; and the basal ganglia (such as CAU, PUT, PAL, etc.) had a weight of 1.1. The number of neurons was allocated based on the product of the brain region volume and the functional weight, ensuring that important brain regions received more neuronal resources.

[0020] 2. Spiking neuron model

[0021] This invention uses a modified Leaky Integrate-and-Fire (LIF) neuron model as its foundation and achieves precise capture of EEG signal characteristics from schizophrenia patients by introducing an adaptive threshold mechanism. The model receives the output from a neural pulse coding module as input, converts the input pulse sequence into an input current through a synaptic weight matrix, and then converts it into a change in membrane potential through membrane resistance. When the membrane potential exceeds a threshold, the neuron generates a pulse and resets the membrane potential.

[0022] The adaptive threshold mechanism simulates the fatigue and recovery process of biological neurons. When a neuron fires frequently, its threshold gradually increases, requiring stronger input to fire again; when the neuron rests for a period of time, the threshold gradually returns to its initial level. This mechanism prevents neurons from becoming overexcited, making network activity more stable. In the specific implementation, a time window is used to record the firing history of neurons, and a threshold adjustment term is calculated using an exponential decay function. The time window size is set to 100 time steps, and the decay weight of each time step is controlled by a time constant.

[0023] 3. Construction of brain region connectivity model

[0024] Based on the functional connectivity patterns of the AAL brain regions, this invention constructs a neural network with a directed acyclic graph (DAG) structure. The construction process of this network includes: calculating the average functional connectivity strength between each pair of brain regions using functional connectivity features; determining the causal direction between brain regions using the Granger causality index; constructing a directed acyclic graph based on the functional connectivity strength and causal direction; and performing topological sorting on the constructed directed graph to ensure that the network structure is an acyclic graph.

[0025] Neurons within each brain region employ a fully connected structure, with connection weights initialized from a normal distribution with a mean of 1 and a standard deviation of 0.1. Interlayer connection weights are initialized based on functional connectivity strength and spatial distance, using a spatial decay constant of 50 mm to account for the influence of anatomical distance. Based on these rules, a DAG structure containing 90 network layers was constructed, where the input layer consists of sensory-related brain regions (e.g., occipital and temporal lobes), the intermediate layer consists of association cortex regions (e.g., parietal and prefrontal lobes), and the output layer consists of decision-related brain regions (e.g., prefrontal and cingulate gyrus).

[0026] The output layer receives the pulse outputs from all brain regions and integrates them into the final classification result. For each brain region, the average of its output pulse sequence is calculated. The average pulse sequence of all brain regions is then input into a fully connected layer to obtain the probabilities of two categories: the probability of a person with schizophrenia and the probability of a healthy person. This output layer design integrates information from all brain regions, fully utilizes the functional connectivity patterns of the brain, learns the contribution weights of different brain regions to classification through the fully connected layer, and outputs probabilities to facilitate subsequent classification decisions.

[0027] 4. Supervised learning training methods based on gradient descent

[0028] This invention employs a supervised learning method based on gradient descent to train a spiking neural network. This method automatically adjusts network parameters by minimizing the cross-entropy loss function between the predicted output and the true label, enabling the network to accurately distinguish the EEG signal characteristics of schizophrenic patients from those of healthy individuals.

[0029] Trainable parameters during training include neuron parameters and synaptic parameters. Neuron parameters include membrane time constant, threshold, reset potential, and membrane resistance, with initial values ​​set according to the functional characteristics of the brain region. Synaptic parameters include intra-layer connection weights, inter-layer connection weights, fully connected layer weights, and biases, initialized based on functional connectivity strength and spatial distance.

[0030] Because spiking neural networks use a step function as their activation function, and its derivative is zero at non-zero points, gradient descent cannot be directly applied. To address this issue, the sigmoid function is used as an alternative gradient to the step function to calculate the gradient of the loss function with respect to the network parameters. Parameter updates utilize the Adam optimizer, which adaptively adjusts the learning rate to accelerate convergence.

[0031] The training process includes initializing network parameters, forward propagation, backpropagation, and an early stopping strategy. The loss function value is monitored on the validation set. If the validation loss does not improve for 10 consecutive epochs, training is stopped, and the model parameters with the minimum validation loss are saved. Training parameters include optimizer parameters (learning rate 0.001, first moment decay rate 0.9, second moment decay rate 0.999), training control parameters (maximum number of training epochs 1000, early stopping patience value 10, batch size 32, validation set ratio 20%), and parameter constraints (membrane time constant range [5ms, 50ms], threshold range [-70mV, -40mV], reset potential range [-80mV, -60mV], membrane resistance range [5mΩ, 20mΩ], synaptic weight range [-1, 1]).

[0032] This supervised learning method based on gradient descent has the advantages of high computational efficiency, good convergence, strong interpretability and strong generalization ability. It can effectively train spiking neural networks to achieve accurate classification of EEG signals of schizophrenic patients and healthy people. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of the identification method of the present invention;

[0034] Figure 2 This is a flowchart of the data preprocessing module of the present invention;

[0035] Figure 3 This is a flowchart of the feature extraction module of the present invention;

[0036] Figure 4 This is a flowchart of the spiking neural network construction module of the present invention;

[0037] Figure 5 This is a flowchart of the backpropagation algorithm of this invention;

[0038] Figure 6 This is a flowchart of the brain partition connectivity model of the present invention;

[0039] Figure 7 This is a flowchart of the spiking neuron model of the present invention;

[0040] Figure 8 This is a flowchart of the neural pulse coding of the present invention. Detailed Implementation

[0041] A method for identifying schizophrenia EEG data based on spiking neural networks is proposed. This method utilizes spiking neural networks to simulate brain activity and constructs a brain region connectivity model that distinguishes signals between schizophrenia patients and healthy individuals. Figure 1 As shown, specifically:

[0042] Step S1: Data Acquisition and Preprocessing: First, EEG data were acquired from schizophrenia patients and healthy controls at a sampling rate of 1000Hz. A 128-lead EEG was used to record 10 minutes of resting state data. The raw data was preprocessed, including signal filtering (0.5-45Hz), artifact removal, signal segmentation, and data standardization, to obtain high-quality data segments.

[0043] Step S2: Feature extraction: Extract multi-scale spatiotemporal features from the preprocessed data, including time-frequency feature extraction, functional connectivity feature extraction, and feature fusion; time-frequency feature extraction uses wavelet transform, functional connectivity features are calculated using phase-locked values, and feature fusion maps electrodes to anatomical brain regions based on standard brain atlases;

[0044] Step S3: Construction of Spiking Neural Network: Construct a three-layer spiking neural network, including an input layer, a hidden layer, and an output layer; the input layer receives preprocessed EEG data, the hidden layer contains multiple spiking neurons to simulate neurons in the cerebral cortex, and the output layer is used for classification decision-making;

[0045] Step S4: Brain Partition Connection Model Construction: Based on the spiking neural network, a connection model between brain regions is constructed, taking into account the constraint of anatomical distance, and the connection weights between neurons are calculated using the improved STDP learning rule;

[0046] Step S5: Model Training and Optimization: Train a three-layer spiking neural network using an improved backpropagation algorithm, optimize the training process with adaptive learning rate and momentum term, and use an early stopping strategy to prevent overfitting;

[0047] Step S6: Classification: Use a trained three-layer spiking neural network to classify EEG data, analyze the differences in brain region connectivity patterns between schizophrenic patients and healthy individuals, and diagnose schizophrenia.

[0048] To clearly describe the data dimensions in the method of this invention, the following variables are first defined:

[0049] The number of brainwave channels is usually 128. Total number of time points, for example, for 10 minutes of data with a sampling rate of 1000Hz. ; The number of data segments, usually 600; The number of sampling points for each data segment is typically 2000; Number of frequency bands, usually 5 ( , , , , ); : Number of features, usually 2 (energy and entropy); The number of channels is calculated using the following formula: For a 128-channel configuration, the value is 8128. The number of brain regions is typically 90. The number of neurons in the input layer is equal to the feature dimension, which is usually 40920. The number of neurons in the hidden layer is usually 1 / 4 of the number of neurons in the input layer, i.e., 10230. : Number of output layer neurons, usually 2 (for schizophrenia patients and healthy individuals); : Time steps, usually 100; Sample size, for example, 200 (100 patients with schizophrenia and 100 healthy controls).

[0050] 1. Data Preprocessing Module

[0051] like Figure 2 As shown, the data preprocessing module is the foundation of the method of this invention. Its purpose is to remove noise and artifacts from the raw EEG data, improve signal quality, and provide high-quality data for subsequent feature extraction and classification. This module includes four steps: signal filtering, artifact removal, signal segmentation, and data standardization.

[0052] 1.1 Signal Filtering

[0053] Signal filtering is the first step in data preprocessing, and its purpose is to remove high-frequency and low-frequency noise from the raw EEG data while retaining the frequency band (0.5-45Hz) relevant to neural activity. This invention employs a Butterworth filter for bandpass filtering, which has a flat passband response and steep stopband attenuation characteristics.

[0054] The filter design is based on the following parameters:

[0055] Sampling frequency: Hz; Passband range: Hz to Hz; Filter order: ; Passband ripples: dB; Stopband attenuation: dB.

[0056] The transfer function of the Butterworth filter is:

[0057] in, This represents the transfer function of the Butterworth filter in the complex frequency domain, used to describe the filter's frequency response characteristics. As a complex frequency variable, it is a complex variable in the Laplace transform, used to describe the characteristics of a signal in the frequency domain. The cutoff frequency defines the boundary frequency between the filter's passband and stopband; for the application of this invention, it is set to 0.5Hz and 45Hz. The filter order determines the steepness and filtering performance of the filter; in this invention, it is set to 4th order. The imaginary unit satisfies To convert analog filters into digital filters, this invention uses a bilinear transform:

[0058]

[0059] in, The sampling period is Let be the Z-transform variable.

[0060] The filtering process can be represented as:

[0061] in: This is the original EEG data matrix; This is the filtered data matrix; Let be the transfer function of the digital filter. For example, for 128 channels, 10 minutes, 1000Hz sampling rate data, the input data dimension is... The output data dimension is also .

[0062] 1.2 Artifact Removal

[0063] Artifact removal is the second step in data preprocessing, aiming to remove artifacts generated by non-neural activities such as eye movements and electromyography. We employ Independent Component Analysis (ICA) for artifact removal, specifically using the ICA module from Python's MNE library. The implementation steps of the ICA algorithm are as follows:

[0064] Data preprocessing:

[0065] First, the filtered data... Centralization and whitening are performed using the `preprocessing.ICA` class from the MNE library, with the following parameters set: Number of independent components: (Same as the number of channels, i.e., 128); Maximum number of iterations: 1000; Convergence threshold: 1e-4; Random seed: 42 (to ensure repeatable results).

[0066] Artifact Component Identification: The ICA.find_bads_eog and ICA.find_bads_muscle methods from the MNE library are used to automatically identify eye movement and electromyography artifact components. The identification criteria include:

[0067] Eye movement artifacts: mainly distributed in the forehead area, with the spectrum concentrated in the low frequency (<4Hz).

[0068] Electromyographic artifacts: have high kurtosis values ​​and a wide spectral distribution.

[0069] Signal reconstruction: The ICA.apply method of the MNE library was used to remove the identified artifacts, resulting in a clean EEG signal. .

[0070] The input data dimension for artifact removal is The output data dimension is also For example, for data with 128 channels, 10 minutes, and a 1000Hz sampling rate, both the input and output data dimensions are [missing information]. .

[0071] 1.3 Signal Segmentation

[0072] Signal segmentation is the third step in data preprocessing. Its purpose is to divide continuous EEG signals into fixed-length segments to facilitate subsequent feature extraction and classification. We use overlapping windows for segmentation to maintain signal continuity.

[0073] The segmentation parameters are set as follows: Window length: seconds; overlap rate: Number of sampling points: Number of overlapping points: Total number of segments: .

[0074] The piecewise process can be represented as:

[0075] in: This is the data matrix after artifact removal; This is the segmented data matrix; For segment index; For channel indexing; This is the index for the sampling points. For example, for data with 128 channels, 10 minutes, and a 1000Hz sampling rate, the input data dimension is... The output data dimension is ; This represents the number of segments.

[0076] 1.4 Data Standardization

[0077] Data standardization is the final step in data preprocessing. Its purpose is to eliminate amplitude differences between different channels and ensure that the data has a uniform scale. This invention employs the z-score standardization method:

[0078]

[0079] in: This is the segmented data matrix; The standardized data matrix; For the first Duan Di The mean of the channel; For the first Duan Di The standard deviation of the channel; the input data dimension for data standardization is... The output data dimension is also For example, for data with 600 segments, 128 channels, and 2000 sampling points per segment, both the input and output data dimensions are... .

[0080] 1.5 The input and output data dimensions of the data preprocessing module are shown in Table 1:

[0081] Table 1

[0082]

[0083] 2. Feature Extraction Module

[0084] The feature extraction module is the core of the method of this invention. Its purpose is to extract features from the preprocessed EEG data that can distinguish between patients with schizophrenia and healthy individuals, while maintaining the correspondence with the original EEG channels. For example... Figure 3 As shown, this module includes three steps: time-frequency feature extraction, functional connectivity feature extraction, and feature fusion.

[0085] 2.1 Time-Frequency Feature Extraction

[0086] Time-frequency feature extraction is the first step in feature extraction, aiming to capture the energy distribution and complexity characteristics of each EEG channel across different frequency bands. We employ an improved time-frequency analysis method, combining Short-Time Fourier Transform (STFT) and Hilbert-Huang Transform (HHT), implemented using Python's SciPy and PyHHT libraries.

[0087] The specific steps for time-frequency feature extraction are as follows:

[0088] First, the EEG signals of each channel... A short-time Fourier transform (STFT) is performed using the Hanning window function with a window length of 200 ms and an overlap rate of 50%. The formula for calculating the STFT is:

[0089]

[0090] in, For the first Duan Di Channel in time and frequency The short-time Fourier transform coefficients at a given time and frequency are used to represent the spectral amplitude and phase information at that time and frequency. For the first Duan Di Channel at time point Standardized EEG signal values; The Hanning window function is used to reduce spectral leakage; its expression is: ; The window length is 200 sampling points, corresponding to 200ms. The index of the sampling points within the window, from 0 to... ; This is the starting position of the time window; It is a complex exponential function used to calculate frequency domain components; For complex units ( ).

[0091] Then, empirical mode decomposition (EMD) is performed on the STFT results to obtain the intrinsic mode functions (IMF):

[0092]

[0093] in, For the first Duan Di Channel at time point Standardized EEG signals; For the first Duan Di The first channel Each intrinsic mode function represents the oscillation component of a signal within a specific frequency range; For the first Duan Di The residual term of the channel represents the trend component of the signal or the part that cannot be further decomposed; The total number of intrinsic mode functions (usually 5-8) is automatically determined by the EMD algorithm based on the complexity of the signal; The index of the intrinsic mode function, from 1 to ; Indexed by time point.

[0094] Perform a Hilbert transform on each IMF to obtain the analytic signal:

[0095]

[0096] in, For the first Duan Di Channel 1 The analytic signal of an intrinsic mode function is a complex signal that contains the amplitude and phase information of the original signal; For the first The real part of each intrinsic mode function; for The Hilbert transform of the original signal represents the orthogonal components of the original signal; The Hilbert transform operator is defined as follows: ; For complex units ( ),in Let be the integral variable, representing any point in time on the time axis. Input signal At any moment The function value at point A is calculated by integrating the signal at all times using Cauchy principal value integration.

[0097] Calculate the instantaneous frequency and instantaneous amplitude:

[0098]

[0099] in, For the first Duan Di Channel 1 The intrinsic mode functions at time point The instantaneous frequency, measured in Hz, represents the local oscillation frequency of the signal at that point in time. For the first Duan Di Channel 1 The intrinsic mode functions at time point The instantaneous amplitude represents the local oscillation intensity of the signal at that point in time; For complex numbers The phase angle, in radians; The time derivative operator represents the derivative with respect to time; Modulus operation for complex numbers converts complex numbers into real numbers; This is the conversion factor from angular frequency to frequency.

[0100] The frequency is divided into 5 bands: , , , and Calculate the following characteristics for each frequency band:

[0101] Energy characteristics:

[0102] in, For the first Duan Di Channel in frequency band The energy characteristics represent the average power of the signal within that frequency band; Indicates frequency band index ( , respectively corresponding , , , , frequency band); Indicates belonging to the frequency band All intrinsic mode function indices; The square of the instantaneous amplitude represents the instantaneous power; This represents the number of sampling points for each data segment.

[0103] Entropy characteristics: in, .

[0104] in, For the first Duan Di Channel in frequency band The entropy characteristic represents the complexity and uncertainty of the signal within that frequency band; For time points In frequency band Normalized energy probability density within; It is a logarithmic function with base 2; the larger the entropy value, the higher the complexity of the signal, and the smaller the entropy value, the more regular the signal.

[0105] Complexity characteristics:

[0106] in, For the first Duan Di Channel in frequency band The complexity characteristic represents the combined intensity of the instantaneous frequency and instantaneous amplitude changes of the signal within this frequency band; The derivative of the instantaneous frequency with respect to time represents the rate of change of frequency; The derivative of the instantaneous amplitude with respect to time represents the rate of change of amplitude; the greater the complexity feature, the more drastic the changes in the time-frequency characteristics of the signal.

[0107] The input data dimension for time-frequency feature extraction is The output data dimension is For example, for data with 600 segments, 128 channels, 2000 sampling points per segment, 5 frequency bands, and 3 features, the input data dimension is... The output data dimension is , It refers to the number of frequency bands. It is the number of features.

[0108] 2.2 Functional Connectivity Feature Extraction

[0109] Functional connectivity feature extraction is the second step in feature extraction, and its purpose is to capture the functional coupling relationships between different EEG channels. We employ an improved functional connectivity analysis method that combines phase-locked value (PLV) and Granger causality, implemented using Python's MNE and statsmodels libraries.

[0110] The specific implementation steps for functional connectivity feature extraction are as follows:

[0111] First, the EEG signals of each channel... Perform a Hilbert transform to obtain the analytic signal:

[0112]

[0113] Calculate phase and amplitude:

[0114]

[0115] Calculate the phase lock value (PLV):

[0116]

[0117] in, Indicates the first Channel in segment data and channels The phase lock value is used to quantify the degree of phase synchronization between the two channels; This indicates the number of sampling points for each data segment (usually 2000 sampling points); Indicates the time point index, from 1 to... ; Indicates the first Channel in segment data At the point of time The instantaneous phase; Indicates the first Channel in segment data At the point of time The instantaneous phase; Represents the complex unit ( ); Represents an exponential function; The modulus operation represents the conversion of complex numbers to real numbers. The range of PLV is [0,1], where 0 indicates complete non-synchronization and 1 indicates complete synchronization.

[0118] Then, Granger causality (GC) is calculated. First, autoregressive modeling is performed on the signal:

[0119]

[0120] in, This is the model order (usually 10). and These are the autoregressive coefficients. It represents the residual.

[0121] Calculate the Granger causality index:

[0122]

[0123] in, Indicates channel For the channel Granger causality strength, used to quantify channels For the channel The degree of causal influence; Represents the natural logarithm function; For using only channels Historical information Residual variance when establishing an autoregressive model; To use channels simultaneously and channels Historical information and Residual variance when establishing an autoregressive model; when At that time, it indicates the channel Historical information helps predict channels The future state, that is, the existence of from arrive causal relationship ;on the contrary, This indicates that there is no causal relationship.

[0124] Calculate PLV and GC for each frequency band separately:

[0125]

[0126] in, Indicates the first Channel in segment data and channels In frequency band Phase-locked value characteristics; Indicates the first Channel in segment data For the channel In frequency band Granger causality characteristics; Indicates frequency band The number of sampling points within; Indicates belonging to the frequency band Time point index.

[0127] Finally, the PLV and GC features are merged into a unified functional connectivity feature matrix:

[0128]

[0129] in, Indicates the first Channel in segment data and channels In frequency band The Each functional connection feature Corresponding to PLV characteristics, Corresponding GC characteristics.

[0130] The input data dimension for functional connectivity feature extraction is The output data dimension is For example, for data with 600 segments, 128 channels, 2000 sampling points per segment, 5 frequency bands, and 2 features, the input data dimension is... The output data dimension is , This is the number of features used in the functional connection, which is fixed at 2. Corresponding to PLV characteristics, Corresponding GC characteristics.

[0131] 2.3 Feature Fusion

[0132] Feature fusion is the final step in feature extraction. Its purpose is to combine time-frequency features and functional connectivity features into a unified feature vector while maintaining the correspondence with the original EEG channels. We employ an improved feature fusion method that directly concatenates feature vectors, implemented using Python's NumPy library.

[0133] The specific steps for feature fusion are as follows:

[0134] First, standardize each feature matrix:

[0135]

[0136] in, For the first Duan Di The channel-normalized time-frequency feature vector has a dimension of ; For the first Duan Di The original time-frequency feature vector of the channel; For the first The mean vector of the channel's time-frequency characteristics across all data segments is calculated using the following formula: ; For the first The standard deviation vector of the channel's time-frequency characteristics across all data segments is calculated using the following formula: ; For the first Duan Di Channel and the Standardized functional connectivity feature vectors between channels; For the first Duan Di Channel and the Original functional connectivity feature vectors between channels; For the first Channel and the The mean vector of functional connectivity features between channels across all data segments; For the first Channel and the The standard deviation vector of functional connectivity features between channels across all data segments; subscript This represents all elements in this dimension.

[0137] Then, the functional connectivity features are reconstructed, transforming the features of channel pairs into functional connectivity feature vectors for each channel:

[0138]

[0139] in, For the first Duan Di The reconstructed functional connectivity feature vectors of the channels will replace the original 3D functional connectivity feature matrix. (dimension is) ,Right now Reconstruct it into a one-dimensional vector; This is a matrix reconstruction function that flattens a multidimensional matrix into a one-dimensional vector in row-major order. The total number of brainwave channels (128); The number of frequency bands (5); The number of functional connectivity features is 2 (including PLV and GC features); the dimension of the reconstructed vector is... This vector preserves the first Detailed functional connectivity information of this channel with all other channels across various frequency bands.

[0140] Finally, the standardized time-frequency features and the reconstructed functional connectivity features are directly concatenated:

[0141]

[0142] in, For the first Duan Di The fused feature vector of a channel is the direct concatenation of time-frequency features and functional connectivity features; This represents a vector concatenation operation, which joins two feature vectors sequentially into a longer feature vector. For the first Duan Di The channel-normalized time-frequency eigenvectors are first processed into a two-dimensional feature matrix. Flattened into a one-dimensional vector, with dimension . ; For the first Duan Di The reconstructed functional connectivity feature vector is now a one-dimensional vector with dimension 1. ; the concatenated fused feature vector The total dimension is This feature vector integrates the time-frequency characteristics of this channel and its detailed functional connections with all other channels, providing rich input information for subsequent spiking neural networks.

[0143] The input data for feature fusion includes time-frequency features. and functional connection features The output data dimension is For example, for data with 600 segments and 128 channels, the input data dimensions are respectively... and The output data dimension is , where 1295 = 5×3 (time-frequency characteristics) + 128×5×2 (functional connectivity characteristics).

[0144] 2.4 The input and output data dimensions of the feature extraction module are shown in Table 2:

[0145] Table 2

[0146]

[0147] 3. Spiking Neural Network Construction Module

[0148] The spiking neural network construction module is the core of this invention. Its purpose is to construct a neural network capable of simulating the dynamic activity of brain neurons for processing and classifying electroencephalogram (EEG) features. Figure 4 As shown, this module comprises three main parts: improved neural pulse coding, improved spiking neuron model, and brain region connectivity model based on anatomical constraints.

[0149] 3.1 Improved Neural Pulse Coding

[0150] Neural pulse coding is the first step in processing continuous signals using spiking neural networks. Its purpose is to convert continuous electroencephalogram (EEG) features into discrete pulse sequences. For example... Figure 8 As shown, this invention proposes an improved neural pulse coding method that combines the advantages of time coding and frequency coding to achieve efficient representation of EEG features.

[0151] The improved neural pulse coding method comprises two main parts: adaptive threshold coding based on feature importance and pulse firing rate coding based on a time window. First, the input features... Normalization is performed. Here... Indicates the number of data segments. This indicates the number of brainwave channels (128). This represents the fused feature dimension (1295). The normalization process can be represented as:

[0152]

[0153] in, and The first The first channel The mean and standard deviation of each feature. Indicates the data segment index ( ), Indicates channel index ( ), Indicates feature index ( ).

[0154] Next, the importance weight of each feature is calculated. The importance weight reflects the degree to which a feature contributes to distinguishing between patients with schizophrenia and healthy individuals:

[0155]

[0156] Based on feature importance weights, the 100 most important features are selected for each channel. Specifically, for each channel... We choose importance weights The top 100 largest features:

[0157]

[0158] in, Indicates channel The set of indices for the first 100 selected features. The resulting feature matrix. It can be represented as:

[0159]

[0160] in, Indicates the selected feature index ( ), Indicates channel The The original index of the key features.

[0161] An adaptive threshold is calculated based on the selected feature importance weights. The threshold value is proportional to the feature importance, making important features more likely to generate impulses.

[0162]

[0163] in, The baseline threshold is set to 0.5; The threshold adjustment factor is set to 0.2. These two parameters were determined experimentally to ensure that neurons are not over-suppressed or over-activated while maintaining coding efficiency.

[0164] For each time step ( , (Set to 100), calculate the pulse firing probability. The firing probability reflects the degree to which the characteristic value exceeds the threshold:

[0165]

[0166] in, For the sigmoid function, The encoding strength parameter is set to 2.0. This parameter controls the strength of the conversion from eigenvalues ​​to pulse firing probabilities.

[0167] Simultaneously, a time window mechanism is introduced to calculate the average feature value within the time window. This time window mechanism considers the temporal continuity of the features, making the pulse sequence smoother.

[0168]

[0169] in, The time window size is set to 10 time steps. This parameter was determined experimentally, striking a balance between maintaining temporal resolution and considering temporal continuity.

[0170] The pulse firing rate is calculated based on the feature values ​​within the time window. The firing rate reflects the average activity level of the feature within the time window.

[0171]

[0172] Final pulse sequence Generated by the following rules:

[0173]

[0174] This improved neural pulse coding method has the following advantages:

[0175] By adaptively adjusting the threshold based on feature importance weights, important features are more likely to generate pulses. A time window mechanism is introduced to consider the temporal continuity of features, making the pulse sequence smoother. Combining probabilistic coding and firing rate coding improves the information capacity and robustness of the coding. Feature selection reduces feature dimensionality, significantly reducing computational complexity.

[0176] 3.1.1 Improved neural pulse coding input / output data dimensions, as shown in Table 3:

[0177] Table 3

[0178]

[0179] 3.2 Improved spiking neuron model

[0180] like Figure 7As shown, this invention uses an improved Leaky Integrate-and-Fire (LIF) neuron model as its foundation. By introducing an adaptive threshold mechanism and multi-timescale synaptic plasticity, it achieves accurate capture of EEG signal characteristics from patients with schizophrenia. This model receives the output from the neural pulse coding module in section 3.1. As input, where Indicates the number of data segments. This indicates the number of EEG channels (128), where 100 represents the number of important features selected for each channel. Indicates the number of time steps (100).

[0181] The basic LIF model describes the dynamic changes in neuronal membrane potential. First, the input pulse sequence... Through the synaptic weight matrix Converted to input current:

[0182]

[0183] in, The synaptic weight matrix is ​​automatically learned through training, reflecting the importance of different input features; for The input pulse sequence at each moment. Then, the input current. Through membrane resistance This is converted into a change in membrane potential:

[0184]

[0185] in, It is the membrane time constant, which is automatically learned through training and reflects the response speed of the neuron to the input signal; Membrane potential, representing the electrical activity state of a neuron; The resting potential is set to -65mV, representing the steady state of the neuron when there is no input. The membrane resistance is set to 10 MΩ, reflecting the neuron's response characteristics to current.

[0186] When the membrane potential exceeds the threshold At this time, the neuron generates a pulse and resets the membrane potential:

[0187] if then else end

[0188] in, The firing threshold is dynamically adjusted through an adaptive mechanism, reflecting the excitability state of neurons; The reset potential is set to -70mV, representing the recovery state after the neuron fires. The output pulse sequence of a neuron, with dimensions equal to the input. same.

[0189] This basic model achieves precise capture of the EEG signal characteristics of schizophrenia patients through the following two key innovative mechanisms:

[0190] 3.2.1 Adaptive threshold mechanism:

[0191] The adaptive threshold mechanism simulates the fatigue and recovery process of biological neurons. When a neuron fires frequently, its threshold gradually increases, requiring stronger input to fire again; when the neuron rests for a period of time, the threshold gradually returns to its initial level. This mechanism prevents neurons from becoming overexcited, making network activity more stable.

[0192] In the implementation, we used a time window to record the firing history of neurons and calculated the threshold adjustment term using an exponential decay function. The time window size was set to 100 time steps, and the decay weight of each time step was determined by a time constant. Control. The formula for calculating the threshold adjustment is:

[0193]

[0194] in, for The threshold for issuing at any given time; The initial threshold is set to -55mV; The adaptive coefficients are learned automatically through training, with an initial value set to 0.1mV. Set the time window size to 100 time steps; The threshold decay time constant is automatically learned through training, with an initial value set to 50ms. for The pulse issuance status at any given moment (1 indicates issuance, 0 indicates no issuance).

[0195] During the training process, and As trainable parameters, they are automatically adjusted via backpropagation. These parameters can learn different values ​​to adapt to the characteristics of neural activity in different frequency bands. For example, for neurons in the gamma band (30-45Hz), a rapid response to high-frequency stimuli is required. It will learn a relatively large value (close to 0.2mV), while They will learn smaller values ​​(close to 30ms); conversely, for neurons in the delta band (0.5-4Hz), they need to maintain an excitatory state for a longer period of time. It will learn smaller values ​​(close to 0.05mV), while It will learn a relatively large value (close to 70ms).

[0196] 3.2.2 Multi-timescale synaptic plasticity:

[0197] Multi-timescale synaptic plasticity is one of the core features of the spiking neural network in this invention. It achieves precise capture of EEG signal characteristics of schizophrenia patients by simulating the dynamic changes of biological synapses at different time scales. This plasticity mechanism consists of two main parts: short-term plasticity (STP) and long-term plasticity (LTP).

[0198] Short-term plasticity primarily simulates the rapid release and reuptake of presynaptic neurotransmitters. In biological neural systems, when a presynaptic neuron fires a pulse, neurotransmitters are rapidly released into the synaptic cleft, causing a transient change in the postsynaptic potential. This change typically returns to baseline levels within milliseconds to seconds. In our model, short-term plasticity is described by two key variables: using variables and recovery variables Using variables Indicates the probability of presynaptic neurotransmitter release, restoring variables This represents the available neurotransmitter resources. The dynamic changes of these two variables follow these rules:

[0199]

[0200]

[0201] in This is a parameter used to control the probability of neurotransmitter release, initially set to 0.2. The contribution of short-term plasticity to synaptic currents. It can be represented as:

[0202]

[0203] here It is synaptic weight. express The input pulse sequence at time t.

[0204] Long-term plasticity simulates persistent changes in postsynaptic receptors and synaptic structure. When presynaptic and postsynaptic neurons fire impulses sequentially within a short period, persistent changes occur in synaptic strength. These changes can last from minutes to hours, or even longer. In our model, long-term plasticity is expressed through time-dependent changes in synaptic weights. To describe:

[0205]

[0206] in This is the learning rate (initial value set to 0.01). It is the time difference between the firing of neurons before and after the synapse. It is a long-term time constant (initial value set to 50ms). It is a sign function. The contribution of long-term plasticity to synaptic current. It can be represented as:

[0207]

[0208] To integrate the effects of short-term and long-term plasticity, we define the total synaptic current. :

[0209]

[0210] This total synaptic current This information will be fed into the improved LIF neuron model, affecting changes in the neuron's membrane potential. For example... Figure 5 As shown, during training, we automatically adjust the plasticity parameters, including the short-term weight coefficients, using the backpropagation algorithm. Long-term weighting coefficient Short-term time constant and long-term time constant The update rules for these parameters are as follows:

[0211]

[0212]

[0213]

[0214]

[0215] in and These are the learning rates for the weighting coefficients and the time constant, respectively, set to 0.001 and 0.0001. This is the loss function. To ensure the parameters are within a reasonable range, we imposed the following constraints:

[0216]

[0217]

[0218]

[0219]

[0220] This multi-temporal-scale synaptic plasticity mechanism allows networks to adapt to the characteristics of neural activity in different frequency bands. For example, for neurons in the gamma band (30-45Hz), due to the need for rapid response to high-frequency stimuli, the weighting coefficient of short-term plasticity... It will learn larger values ​​(close to 0.9), while the weighting coefficient for long-term plasticity... They will learn smaller values ​​(close to 0.1). Conversely, for neurons in the delta band (0.5-4Hz), the weighting coefficients for long-term plasticity are higher because they need to maintain synaptic strength changes for a longer period. It will learn larger values ​​(close to 0.9), while the weighting coefficient for short-term plasticity... It will learn smaller values ​​(close to 0.1).

[0221] This differentiated plasticity mechanism allows networks to better simulate the dynamic characteristics of brain activity in patients with schizophrenia. For example, in the prefrontal cortex, the weighting coefficients of long-term plasticity are higher due to the need to maintain working memory. Larger values ​​will be learned; however, in the sensory cortex, due to the need for rapid response to external stimuli, the weighting coefficient of short-term plasticity is lower. It learns larger values. This mechanism not only improves the network's ability to capture features of brain electrical signals, but also provides a new perspective for understanding the neural mechanisms of schizophrenia.

[0222] 3.2.3 The input and output data dimensions of the improved spiking neuron model are shown in Table 4:

[0223] Table 4

[0224]

[0225] 3.3 Brain Partition Connectivity Model Based on Anatomical Constraints

[0226] like Figure 6As shown, in this invention, we constructed a brain region connectivity model based on anatomical constraints. This model precisely maps a 128-lead EEG system to 90 AAL brain regions and establishes the corresponding neuronal network structure. The core of this model lies in spatially mapping EEG channels to AAL brain regions and allocating the number of neurons according to the functional importance and anatomical volume of the brain regions.

[0227] 3.3.1 Brain region mapping and neuron allocation

[0228] First, we established a mapping relationship between a 128-lead EEG system and 90 AAL brain regions. Each EEG channel was assigned to the nearest AAL brain region based on its spatial location. The specific mapping process is as follows:

[0229] Coordinate system definition: EEG channel coordinates: Using the international 10⁻⁵ system for standard electrode positions, the three-dimensional coordinates of 128 electrodes are obtained using the `mne.channels.make_standard_montage('standard_1005')` function of the MNE-Python library. These coordinates are based on Talairach space and are in millimeters.

[0230] AAL brain region coordinates: Using the standard brain region centroid coordinates from the AAL atlas, the centroid coordinates of 90 brain regions were obtained through the aal.nii template in the SPM12 software package. These coordinates are also based on Talairach space and are in millimeters.

[0231] Coordinate Alignment: To ensure that the EEG channel and the AAL brain region are in the same spatial coordinate system, we use the following steps for coordinate alignment: transform the EEG channel coordinates from Talairach space to MNI space; transform the AAL brain region coordinates from Talairach space to MNI space using a linear transformation matrix; and perform coordinate transformation using the mne.transforms.apply_trans function of MNE-Python.

[0232] Distance calculation: Calculate the Euclidean distance between each EEG channel and the centroid of the AAL brain region:

[0233] in, Indicates the first The brainwave channel and the first The distance between each AAL brain region and These are the three-dimensional coordinates of the channels and brain regions in MNI space, respectively.

[0234] Brain region allocation: Each EEG channel is assigned to the nearest AAL brain region: in, Indicates the first The AAL brain region number to which each brainwave channel belongs.

[0235] Neuron number allocation: The number of neurons was allocated to each AAL brain region based on its functional importance and anatomical volume. Neuron number The calculation formula is: in, This represents the total number of neurons (128, the same as the number of EEG channels). For the first The volume of each brain region For the first Functional weights of individual brain regions.

[0236] Functional weight Based on the importance of brain regions in schizophrenia: Prefrontal cortex (PreCG, SFGdor, MFG, etc.): Temporal lobe regions (STG, MTG, ITG, etc.): Parietal region (SPG, IPL, etc.): Occipital lobe region (SOG, MOG, IOG, etc.): Edge systems (HIP, AMYG, etc.): Basal ganglia (CAU, PUT, PAL, etc.): .

[0237] 3.3.2 Neuron Network Structure

[0238] Based on the functional connectivity patterns of the AAL brain region, this invention constructs a neural network with a directed acyclic graph (DAG) structure. The construction process of this network is as follows:

[0239] Functional connectivity analysis: using the functional connectivity features extracted in Section 2.2 Calculate the average functional connectivity strength between each pair of brain regions: in, Indicates brain regions and The strength of functional connections between them For the first brain regions in segment data and In the Functional connectivity features of each frequency band.

[0240] Granger causality analysis: using the Granger causality index calculated in Section 2.2 Determine the causal direction between brain regions: in, Indicates from brain region arrive The causal direction, brain region right Granger causality index.

[0241] Directed Acyclic Graph Construction: Based on functional connectivity strength and causal direction, a directed acyclic graph is constructed: Nodes: Each AAL brain region corresponds to a network layer; Edges: When... and At that time, establish from brain regions arrive Directed edges; threshold Set to All The 90th percentile ensures that only the strongest functional connections are retained.

[0242] Topology sorting: Sort the constructed directed graph topology to ensure that the network structure is acyclic: calculate the in-degree of each node; select the node with an in-degree of 0 as the starting layer; arrange the network layers in topological order; remove edges that may form cycles.

[0243] Connection weight initialization:

[0244] Intralayer connectivity: Neurons within each brain region employ a fully connected structure, with connection weights... Initialize to: in, Based on the base weight (0.1), This represents a normal distribution with a mean of 1 and a standard deviation of 0.1.

[0245] Inter-layer connectivity: Initialize inter-layer connectivity weights based on functional connectivity strength and spatial distance. :

[0246] in, brain region and Anatomical distance between them The spatial attenuation constant is 50 mm. For functional connection strength.

[0247] Network structure example: Based on the above rules, this invention constructs a DAG structure containing 90 network layers, wherein: Input layer: sensory-related brain regions (such as occipital lobe and temporal lobe); Intermediate layer: association cortex (such as parietal lobe and prefrontal lobe); Output layer: decision-related brain regions (such as prefrontal lobe and cingulate gyrus).

[0248] 3.3.3 Output Layer Design

[0249] The output layer receives the pulse outputs from all brain regions and integrates them into the final classification result. The specific design is as follows:

[0250] Pulse sequence integration: for each brain region Calculate the average value of its output pulse sequence: in, brain region At time step The output pulse sequence, This represents the total number of time steps (100).

[0251] Fully connected layer: The average impulse sequence of all brain regions is input into a fully connected layer.

[0252] in, This is the weight matrix of the fully connected layer. For bias terms, It is the sigmoid activation function.

[0253] Output probability: The output of the fully connected layer Represents the probability of two categories: The probability of having schizophrenia; The probability of a healthy person.

[0254] This output layer design has the following advantages: it integrates information from all brain regions and makes full use of the brain's functional connectivity patterns; it learns the contribution weights of different brain regions to classification through fully connected layers; and the output probabilities facilitate subsequent classification decisions.

[0255] 3.4 Supervised Learning Training Method Based on Gradient Descent

[0256] 3.4.1 Input and output variable definitions are shown in Table 5:

[0257] Table 5

[0258]

[0259] in: Number of data segments, for example, 600; Number of brainwave channels, for example, 128; The fused feature dimension, for example, 1295; : Number of time steps, for example, 100.

[0260] This invention employs a supervised learning method based on gradient descent to train a spiking neural network. This method automatically adjusts network parameters by minimizing the loss function between the predicted output and the true label, enabling the network to accurately distinguish the EEG signal characteristics of schizophrenic patients from those of healthy individuals.

[0261] 3.4.2 Loss Function Design

[0262] The cross-entropy loss function is used as the training objective:

[0263]

[0264] in, The value of the loss function. Number of data segments For the first The true labels of the data segment (1 represents a schizophrenic patient, 0 represents a healthy person). This is the network's predicted output.

[0265] 3.4.3 Trainable Parameters

[0266] Based on the network structure defined in Section 3.3, the trainable parameters include:

[0267] Neuron parameters: membrane time constant : Set initial values ​​based on the functional characteristics of brain regions (prefrontal lobe 20ms, temporal lobe 15ms, parietal lobe 18ms, occipital lobe 12ms, limbic system 25ms, basal ganglia 10ms).

[0268] threshold The initial value is set to -55mV;

[0269] Reset potential The initial value is set to -70mV;

[0270] membrane resistance The initial value is set to 10MΩ.

[0271] Synaptic parameters: Intralayer connection weights Initial value from Sampling in distribution; inter-layer connection weights Initial values ​​are calculated based on functional connectivity strength and spatial distance; weights of fully connected layers. Initial value from Sampling in distribution; bias of fully connected layers The initial value is set to 0.

[0272] 3.4.4 Gradient Calculation and Parameter Update

[0273] Since neurons in a spiking neural network use a step function as their activation function, and its derivative is 0 at non-zero points, gradient descent cannot be directly applied. To address this issue, we use the sigmoid function as an alternative gradient to the step function:

[0274]

[0275] in, This refers to the membrane potential. The threshold for issuance, This is the sigmoid function.

[0276] Based on the alternative gradient, we can calculate the gradient of the loss function with respect to the network parameters:

[0277] Gradient of neuron parameters:

[0278] Gradient of synaptic parameters:

[0279] 3.4.5 Training Process

[0280] The training process includes the following steps:

[0281] Initialize network parameters: Neuron parameters: set according to the initial values ​​defined in Section 3.3; Synapse parameters: set according to the initialization method defined in Section 3.3.

[0282] Forward propagation: transferring input features Converted into a pulse sequence via pulse coding ;Will The input is fed into a spiking neural network to calculate the membrane potential and output pulse at each time step; the average pulse sequence for each brain region is also calculated. ;Will The input is fed into a fully connected layer to obtain the predicted output. .

[0283] Backpropagation: Calculating the loss function Gradients to network parameters; update parameters using the Adam optimizer.

[0284] Early stopping strategy: Monitor the loss function value on the validation set; if the validation loss does not improve after 10 consecutive epochs, stop training; save the model parameters with the minimum validation loss.

[0285] 3.4.6 Training Parameter Settings

[0286] The key parameter settings during training are as follows:

[0287] Optimizer parameters: learning rate : 0.001; First-order moment decay rate : 0.9; Second-order moment decay rate : 0.999; Numerical stability constant :1e-8.

[0288] Training control parameters: Maximum number of training rounds: 1000; Early stopping patience value: 10; Batch size: 32; Validation set ratio: 20%.

[0289] Parameter constraints: Membrane time constant range: [5ms, 50ms]; Threshold range: [-70mV, -40mV]; Reset potential range: [-80mV, -60mV]; Membrane resistance range: [5mΩ, 20mΩ]; Synaptic weight range: [-1, 1].

[0290] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing EEG data of schizophrenia based on spiking neural networks, characterized in that... Includes the following steps: Step S1: Data Acquisition and Preprocessing: First, EEG data were acquired from schizophrenia patients and healthy controls at a sampling rate of 1000Hz. A 128-lead EEG was used to record 10 minutes of resting state data. The raw data was preprocessed, including signal filtering, artifact removal, signal segmentation, and data standardization, to obtain high-quality data segments. Step S2: Feature extraction: Extract multi-scale spatiotemporal features from the preprocessed data, including time-frequency feature extraction, functional connectivity feature extraction, and feature fusion; Time-frequency features were extracted using wavelet transform, functional connectivity features were calculated using phase-locked values, and feature fusion was based on standard brain atlases to map electrodes to anatomical brain regions. Step S3: Construction of Spiking Neural Network: Construct a three-layer spiking neural network, including an input layer, a hidden layer, and an output layer; the input layer receives preprocessed EEG data, the hidden layer contains multiple spiking neurons to simulate neurons in the cerebral cortex, and the output layer is used for classification decision-making; Step S4: Brain Partition Connection Model Construction: Based on the spiking neural network, a connection model between brain regions is constructed, taking into account the constraint of anatomical distance, and the connection weights between neurons are calculated using the improved STDP learning rule; Step S5: Model Training and Optimization: Train a three-layer spiking neural network using an improved backpropagation algorithm, optimize the training process with adaptive learning rate and momentum term, and use an early stopping strategy to prevent overfitting; Step S6: Classification: Use a trained three-layer spiking neural network to classify the EEG data and analyze the differences in brain region connectivity patterns between schizophrenic patients and healthy individuals.

2. The method for identifying schizophrenia EEG data based on spiking neural networks according to claim 1, characterized in that... The specific process of signal filtering is as follows: The transfer function of the Butterworth filter is: ; in, This represents the transfer function of the Butterworth filter in the complex frequency domain, used to describe the frequency response characteristics of the filter; The complex frequency variable is a complex variable in the Laplace transform, used to describe the characteristics of a signal in the frequency domain; The cutoff frequency defines the boundary frequency between the filter's passband and stopband. The filter order determines the steepness of the filter and its filtering performance. The imaginary unit satisfies Use bilinear transform to convert analog filters into digital filters: ； in, The sampling period is Let Z be the variable for transformation; The filtering process is represented as follows: in: This is the original EEG data matrix; This is the filtered data matrix; The transfer function of the digital filter; Indicates the number of brainwave channels; This is the sampled data.

3. The method for identifying schizophrenia EEG data based on spiking neural networks according to claim 2, characterized in that... The specific process for artifact removal is as follows: Data preprocessing: First, the filtered data... To perform centralization and whitening, initialize the process using the `preprocessing.ICA` class from the MNE library, setting the following parameters: Number of independent components: Same as the number of channels; Maximum number of iterations: 1000; Convergence threshold: 1e-4; Random seed: 42; Artifact component identification: The ICA.find_bads_eog and ICA.find_bads_muscle methods from the MNE library are used to automatically identify eye movement and electromyography artifacts. Identification criteria include: Eye movement artifacts: mainly distributed in the forehead region, with the spectrum concentrated in the low frequency; Electromyographic artifacts: have high kurtosis and a wide spectral distribution; Signal reconstruction: The ICA.apply method from the MNE library was used to remove identified artifacts, resulting in a clean EEG signal. ; The input data dimension for artifact removal is The output data dimension is also .

4. The method for identifying schizophrenia EEG data based on spiking neural networks according to claim 3, characterized in that... The specific process of signal segmentation is as follows: segmentation is performed using an overlapping window method to maintain signal continuity; the segmentation parameters are set as follows: The piecewise process is represented as follows: in: This is the data matrix after artifact removal; This is the segmented data matrix; For segment index; For channel indexing; For sampling point index; This represents the number of sampling points; The number of overlapping points; This represents the number of segments.

5. The method for recognizing EEG data of schizophrenia based on spiking neural networks according to claim 4, characterized in that... The specific process of data standardization is as follows: The z-score standardization method is used. ； in: This is the segmented data matrix; The standardized data matrix; For the first Duan Di The mean of the channel; For the first Duan Di The standard deviation of the channel; the input data dimension for data standardization is... The output data dimension is also .

6. The method for identifying schizophrenia EEG data based on spiking neural networks according to claim 5, characterized in that... The specific process of time-frequency feature extraction is as follows: First, the EEG signals of each channel... A short-time Fourier transform is performed using the Hanning window function with a window length of 200 ms and an overlap rate of 50%. The formula for calculating the short-time Fourier transform is as follows: ； in, For the first Duan Di Channel in time and frequency The short-time Fourier transform coefficients at a given time and frequency are used to represent the spectral amplitude and phase information at that time and frequency. For the first Duan Di Channel at time point Standardized EEG signal values; The Hanning window function is used to reduce spectral leakage; its expression is: ; The length of the window; The index of the sampling points within the window, from 0 to... ; This is the starting position of the time window; It is a complex exponential function used to calculate frequency domain components; For complex units; Then, empirical mode decomposition is performed on the short-time Fourier transform results to obtain the intrinsic mode functions: ； in, For the first Duan Di Channel at time point Standardized EEG signals; For the first Duan Di The first channel Each intrinsic mode function represents the oscillation component of a signal within a specific frequency range; For the first Duan Di The residual term of the channel represents the trend component of the signal or the part that cannot be further decomposed; The total number of intrinsic mode functions is automatically determined by the empirical mode decomposition algorithm based on the complexity of the signal; The index of the intrinsic mode function, from 1 to ; Indexed by time point; Performing a Hilbert transform on each intrinsic mode function yields the analytic signal: ； in, For the first Duan Di Channel 1 The analytic signal of an intrinsic mode function is a complex signal that contains the amplitude and phase information of the original signal; For the first The real part of each intrinsic mode function; for The Hilbert transform of the original signal represents the orthogonal components of the original signal; The Hilbert transform operator is defined as follows: ; For the complex unit; where Let be the integral variable, representing any point in time on the time axis. For input signal At any moment The function value at that location; Calculate the instantaneous frequency and instantaneous amplitude: ； in, For the first Duan Di Channel 1 The intrinsic mode functions at time point The instantaneous frequency, measured in Hz, represents the local oscillation frequency of the signal at that point in time. For the first Duan Di Channel 1 The intrinsic mode functions at time point The instantaneous amplitude represents the local oscillation intensity of the signal at that point in time; For complex numbers The phase angle, in radians; The time derivative operator represents the derivative with respect to time; Modulus operation for complex numbers converts complex numbers into real numbers; This is the conversion factor from angular frequency to frequency; The frequency is divided into 5 bands: , , , and For each frequency band, calculate the following characteristics: Energy characteristics: ; in, For the first Duan Di Channel in frequency band The energy characteristics represent the average power of the signal within that frequency band; Indicates frequency band index; Indicates belonging to the frequency band All intrinsic mode function indices; The square of the instantaneous amplitude represents the instantaneous power; The number of sampling points for each data segment; Entropy characteristics: in, ; in, For the first Duan Di Channel in frequency band The entropy characteristic represents the complexity and uncertainty of the signal within that frequency band; For time points In frequency band Normalized energy probability density within; It is a logarithmic function with base 2; the larger the entropy value, the higher the complexity of the signal, and the smaller the entropy value, the more regular the signal. Complexity characteristics: ; in, For the first Duan Di Channel in frequency band The complexity characteristic represents the combined intensity of the instantaneous frequency and instantaneous amplitude changes of the signal within this frequency band; The derivative of the instantaneous frequency with respect to time represents the rate of change of frequency; The derivative of the instantaneous amplitude with respect to time represents the rate of change of amplitude; the greater the complexity feature, the more drastic the changes in the time-frequency characteristics of the signal. The input data dimension for time-frequency feature extraction is The output data dimension is , It refers to the number of frequency bands. It is the number of features.

7. The method for identifying schizophrenia EEG data based on a spiking neural network according to claim 6, characterized in that... The specific process for extracting the functional connectivity features is as follows: First, the EEG signals of each channel... Perform a Hilbert transform to obtain the analytic signal: ； Calculate phase and amplitude: ； Calculate the phase lock value (PLV): ； in, Indicates the first Channel in segment data and channels The phase lock value is used to quantify the degree of phase synchronization between the two channels; This indicates the number of sampling points in each data segment; Indicates the time point index, from 1 to... ; Indicates the first Channel in segment data At the point of time The instantaneous phase; Indicates the first Channel in segment data At the point of time The instantaneous phase; Indicates the complex unit; Represents an exponential function; Modulus operation for complex numbers converts complex numbers into real numbers. The range of PLV is [0,1], where 0 indicates complete non-synchronization and 1 indicates complete synchronization. Then, Granger causality is calculated; first, autoregressive modeling is performed on the signal: ； in, The model order is... and These are the autoregressive coefficients. For residuals; Calculate the Granger causality index: ； in, Indicates channel For the channel Granger causality strength, used to quantify channels For the channel The degree of causal influence; Represents the natural logarithm function; For using only channels Historical information Residual variance when establishing an autoregressive model; To use channels simultaneously and channels Historical information and Residual variance when establishing an autoregressive model; when At that time, it indicates the channel Historical information helps predict channels The future state, that is, the existence of from arrive causal relationship ;on the contrary, This indicates that there is no causal relationship. Calculate PLV and GC for each frequency band separately: ； ； in, Indicates the first Channel in segment data and channels In frequency band Phase-locked value characteristics; Indicates the first Channel in segment data For the channel In frequency band Granger causality characteristics; Indicates frequency band The number of sampling points within; Indicates belonging to the frequency band Time point index; Finally, the PLV and GC features are merged into a unified functional connectivity feature matrix: ； in, Indicates the first Channel in segment data and channels In frequency band The Each functional connection feature Corresponding to PLV characteristics, Corresponding GC characteristics; The input data dimension for functional connectivity feature extraction is The output data dimension is , This is the number of features used in the functional connection, which is fixed at 2. Corresponding to PLV characteristics, Corresponding GC characteristics.

8. The method for identifying schizophrenia EEG data based on spiking neural networks according to claim 7, characterized in that... The specific process of feature fusion is as follows: First, standardize each feature matrix: ； in, For the first Duan Di The channel-normalized time-frequency feature vector has a dimension of ; For the first Duan Di The original time-frequency feature vector of the channel; For the first The mean vector of the channel's time-frequency characteristics across all data segments is calculated using the following formula: ; For the first The standard deviation vector of the channel's time-frequency characteristics across all data segments is calculated using the following formula: ; For the first Duan Di Channel and the Standardized functional connectivity feature vectors between channels; For the first Duan Di Channel and the Original functional connectivity feature vectors between channels; For the first Channel and the The mean vector of functional connectivity features between channels across all data segments; For the first Channel and the The standard deviation vector of functional connectivity features between channels across all data segments; subscript This represents all elements of that dimension; Then, the functional connectivity features are reconstructed, transforming the features of channel pairs into functional connectivity feature vectors for each channel: ； in, For the first Duan Di The reconstructed functional connectivity feature vectors of the channels will replace the original 3D functional connectivity feature matrix. Reconstruct it into a one-dimensional vector; This is a matrix reconstruction function that flattens a multidimensional matrix into a one-dimensional vector in row-major order. This represents the total number of brainwave channels. Number of frequency bands; The number of functional connection features; the dimension of the reconstructed vector is This vector preserves the first Detailed functional connectivity information of this channel with all other channels across various frequency bands; Finally, the standardized time-frequency features and the reconstructed functional connectivity features are directly concatenated: ； in, For the first Duan Di The fused feature vector of a channel is the direct concatenation of time-frequency features and functional connectivity features; This represents a vector concatenation operation, which joins two feature vectors sequentially into a longer feature vector. For the first Duan Di The channel-normalized time-frequency eigenvectors are first processed into a two-dimensional feature matrix. Flattened into a one-dimensional vector, with dimension . ; For the first Duan Di The reconstructed functional connectivity feature vector is now a one-dimensional vector with dimension 1. ; the concatenated fused feature vector The total dimension is ; The input data for feature fusion includes time-frequency features. and functional connection features The output data dimension is .

9. The method for identifying schizophrenia EEG data based on spiking neural networks according to claim 8, characterized in that... The specific process of neural pulse coding is as follows: Adaptive threshold coding based on feature importance and pulse firing rate coding based on time window; first, for the input features Normalization was performed; among which Indicates the number of data segments. Indicates the number of brainwave channels. The fused feature dimensions are represented as follows: The normalization process is represented as follows: ； in, and The first The first channel The mean and standard deviation of each feature. Indicates the data segment index. , Indicates the channel index. , Indicates feature index, ; Then, the importance weight of each feature is calculated: ； Based on feature importance weights, the 100 most important features are selected for each channel; specifically, for each channel... Choose importance weights The top 100 largest features: ； in, Indicates channel The set of indices for the first 100 selected features; the feature matrix after selection. Represented as: ； in, This indicates the selected feature index. , Indicates channel The The original index of the key features; An adaptive threshold is calculated based on the selected feature importance weights; the threshold value is proportional to the importance of the feature, making important features more likely to generate impulses. ； in, The baseline threshold; These are threshold adjustment coefficients; these two parameters were determined experimentally to ensure that neurons are not over-suppressed or over-activated while maintaining coding efficiency. For each time step , , Set to 100 to calculate the pulse firing probability; the firing probability reflects the degree to which the characteristic value exceeds the threshold: ； in, For the sigmoid function, This is the coding strength parameter; it controls the intensity of the conversion from feature values ​​to pulse firing probabilities. Simultaneously, a time window mechanism is introduced to calculate the average feature value within the time window. This time window mechanism takes into account the temporal continuity of the features, making the pulse sequence smoother. ； in, The time window size is set to 10 time steps to strike a balance between maintaining time resolution and considering time continuity. Based on the feature values ​​within the time window, the pulse firing rate is calculated. The firing rate reflects the average activity level of the feature within the time window. ； Final pulse sequence Generated by the following rules: 。 10. The method for identifying schizophrenia EEG data based on a spiking neural network according to claim 9, characterized in that... The specific design of the output layer is as follows: Pulse sequence integration: for each brain region Calculate the average value of its output pulse sequence: ; in, brain region At time step The output pulse sequence, This represents the total number of time steps. Fully connected layer: The average impulse sequence of all brain regions is input into a fully connected layer. in, This is the weight matrix of the fully connected layer. For bias terms, It is the sigmoid activation function; Output probability: The output of the fully connected layer Represents the probability of two categories: The probability of having schizophrenia; The probability of a healthy person.

Images