Alzheimer's disease classification system based on multi-modal data and brain-like nerve calculation

Abstract

The invention discloses an Alzheimer's disease classification system based on multi-modal data and cranial-like nerve calculation, and the system comprises a first construction module which is used for constructing an electroencephalogram network based on preprocessed EEG data; the second construction module is used for constructing a functional network based on the preprocessed fMRI data; the node embedding module is used for learning node embedding of the electroencephalogram network and the function network through the graph convolutional neural network; the feature extraction module is used for embedding nodes corresponding to the same ROI of a plurality of subjects for extraction and splicing to form a feature vector of each subject; and the model training module is used for training a classification model based on the plurality of tested feature vectors obtained by splicing. According to the Alzheimer's disease classification system based on multi-modal data and cranial nerve-like calculation, EEG and fMRI data are fused, the time-frequency characteristics of electroencephalogram activities and the functional connectivity between brain regions are comprehensively considered, the feature extraction capacity is enhanced, and comprehensive understanding of the disease state is improved.

Description

Technical Field

[0001] The present invention specifically relates to an Alzheimer's disease classification system based on multimodal data and brain-like neural computing. Background Art

[0002] Alzheimer's disease (AD) is a progressive neurodegenerative disease characterized by the gradual loss of neurons in the cerebral cortex, leading to memory loss, cognitive dysfunction and behavioral changes. With the aging of the global population, the incidence of AD has increased year by year, placing a heavy burden on patients' families and social medical systems. Early diagnosis and intervention are crucial to slowing the progression of the disease and improving the quality of life of patients. However, the early symptoms of AD are often hidden and lack specificity, making diagnosis difficult. At present, clinical screening for AD is mainly carried out through cerebrospinal fluid examination, electroencephalogram examination, head CT examination, head magnetic resonance imaging and biomarkers, but these methods have problems such as difficulty in diagnosis, strong subjectivity and lack of early intervention time points. In addition, traditional diagnostic methods rely on clinical evaluation and a series of neuropsychological tests, which are not only time-consuming and labor-intensive, but also have limited accuracy and are difficult to meet the needs of large-scale screening and early diagnosis.

[0003] With the development of medical imaging technology, especially the advancement of functional magnetic resonance imaging (fMRI) and structural magnetic resonance imaging (sMRI), a new perspective has been provided for the early diagnosis of Alzheimer's disease. These imaging technologies can reveal changes in brain structure and function, providing important information for the study of biomarkers of the disease. However, how to extract features that are helpful for diagnosis from complex imaging data and how to build an efficient and accurate diagnostic model remain challenges. Traditional machine learning methods often find it difficult to achieve the ideal accuracy when processing such high-dimensional and complex medical imaging data, and lack a deep understanding of the disease mechanism. Summary of the invention

[0004] The present invention provides an Alzheimer's disease classification system based on multimodal data and brain-like neural computing to solve the above-mentioned technical problems, and specifically adopts the following technical solutions:

[0005] An Alzheimer's disease classification system based on multimodal data and brain-like neural computing, including:

[0006] Data acquisition module, used to collect EEG data and fMRI data of multiple subjects;

[0007] A data processing module is used to pre-process the collected EEG data and fMRI data;

[0008] The first building module is used to build an EEG network based on the preprocessed EEG data;

[0009] The second building module is used to build a functional network based on the preprocessed fMRI data;

[0010] Node embedding module, which learns node embedding of EEG networks and functional networks through graph convolutional neural networks;

[0011] The feature extraction module is used to extract and concatenate the node embeddings corresponding to the same ROI of multiple subjects to form a feature vector for each subject;

[0012] A model training module is used to train a classification model based on the concatenated feature vectors of multiple subjects;

[0013] The classification module is used to output classification results based on the trained classification model and the EEG data and fMRI data of the subject to be tested.

[0014] Furthermore, the specific method for the data processing module to pre-process the collected EEG data is:

[0015] The data processing module performs artifact removal and denoising processing on the EEG data by independent component analysis, and then performs standardization processing.

[0016] Furthermore, the specific method for the data processing module to preprocess the collected fMRI data is:

[0017] The data processing module performs slice timing correction, spatial normalization and Gaussian smoothing on the fMRI data, defines the ROI of the brain region, and extracts the BOLD signal sequence of each brain region from the fMRI data:

[0018] BOLD i (t) for i = 1, 2, ..., N

[0019] Among them, BOLD i (t) represents the BOLD signal of the i-th brain region at time point t, and N is the total number of brain regions.

[0020] Furthermore, the specific method of the first building module to build an EEG network based on the preprocessed EEG data is:

[0021] For EEG data, the time domain Pearson correlation coefficient was calculated using the following formula:

[0022]

[0023] Among them, EEG i (t) represents the EEG signal of brain area i at time point t, represents the mean value of the EEG signal of brain area i at all time points, and T is the total number of time points of the signal;

[0024] Construct time-domain EEG network based on time-domain Pearson correlation coefficient;

[0025]

[0026] Among them, θ_time is a preset threshold, which means that the connection will be established only when the time domain Pearson correlation coefficient is greater than the threshold θ_time;

[0027] Fast Fourier transform of EEG data;

[0028] EEG freq (ROI i ,f)=FFT(EEG i (t))

[0029] Among them, EEG freq (ROI i ,f) represents the EEG signal of brain area i at frequency f;

[0030] Based on the transformed EEG data, the frequency domain Pearson correlation coefficient was calculated using the following formula:

[0031]

[0032] Among them, EEG i (t) represents the EEG signal of brain area i at frequency f, represents the mean value of the EEG signal of brain area i at frequency f;

[0033] Construct frequency domain EEG network based on frequency domain Pearson correlation coefficient:

[0034]

[0035] Among them, θ_freq is the threshold of frequency domain correlation, which means that the connection will be established only when the frequency domain Pearson correlation coefficient is greater than θ_freq;

[0036] The total EEG network is obtained by fusing the time domain EEG network and the frequency domain EEG network;

[0037] A total (i,j)=αA time (i,j)+βA freq (i,j)

[0038] Among them, α and β are the weight coefficients controlling the time domain and frequency domain networks, and the fused EEG network is expressed as:

[0039] A total =Fusion(Atime ,A freq ).

[0040] Furthermore, the specific method of the second construction module to construct the functional network based on the preprocessed fMRI data is:

[0041] Construct a set of time series with the same ROI:

[0042] ROI_Collection = {Time_Series k |k=1,2,…,90}

[0043] Among them, Time_Series k represents the time series set of the kth ROI, which is used to construct the nodes of the brain network;

[0044] The functional Pearson correlation coefficient was calculated using the following formula:

[0045]

[0046] Among them, BOLD i t) and BOLD j (t) represents the BOLD signal of brain area i and brain area j at time point t, BOLD i and are the mean BOLD signals of brain area i and brain area j, respectively, and T is the time point number of the BOLD signal;

[0047] Construct a functional network based on the functional Pearson correlation coefficient:

[0048]

[0049] θ_func represents the preset threshold.

[0050] Furthermore, the feature extraction module extracts and concatenates the node embeddings corresponding to the same ROI of multiple subjects to form a feature vector for each subject:

[0051] Z subj_i =[Z subj_i1 , Z subj_i2 , ..., Z subj_ik ]

[0052] Z subj_i represents the feature vector of the subj_i-th subject, Z subj_ik Represents the node embedding of the subj_i-th subject in the k-th ROI, and concatenates the node embeddings of each subject in different ROIs to form a comprehensive feature vector.

[0053] Furthermore, the classification model adopts CNN neural network and the training adopts cross entropy loss function.

[0054] Furthermore, the Alzheimer's disease classification system based on multimodal data and brain-like neural computing also includes a three-dimensional image interpretation module;

[0055] The 3D image interpretation module quantitatively calculates the prediction results of the brain sub-blocks and generates a whole-brain pixel risk value for each subject:

[0056] R subj_i =f(Z subj_i )

[0057] Among them, R subj_i represents the whole brain pixel risk value of the subj_i-th subject, f(Z subj_i ) is a function that calculates the risk value based on the trained model, and generates an individual whole-brain risk visualization map based on the whole-brain pixel risk value of each subject.

[0058] Sort high risk values ​​according to Pareto principle:

[0059] Sorted_Risk_Values=Sort(R subj_i , descending)

[0060] Where Sorted_Risk_Values ​​represents the sorted risk value, and the top 20% of regions are selected;

[0061] Based on the risk value ranking, the high-risk areas of the brain network are combined with clinical data to obtain a comprehensive risk value:

[0062] Combined Risk =α·R subj_i +β·C subj_i

[0063] Among them, α and β are weight coefficients, C i Represents the clinical data of the subj_i-th subject.

[0064] Furthermore, after generating the whole-brain pixel risk value for each subject, the Matthew correlation coefficient at the whole-brain pixel level was calculated using the group data:

[0065]

[0066] in, is the category of the subj_ith sample predicted by the model, is the average of all sample prediction categories, is the actual category of the subj_i-th sample, is the average value of the actual categories of all samples;

[0067] The calculated Matthew correlation coefficients were visualized at the whole-brain pixel level to identify the most correlated brain regions in the group.

[0068] Group_MCC_Visualization=Visualize_MCC(MCC)

[0069] Among them, Group_MCC_Visualization represents the group visualization diagram based on MCC calculation.

[0070] Furthermore, the Alzheimer's disease classification system based on multimodal data and brain-like neural computing also includes a multi-source heterogeneous data fusion module;

[0071] The multi-source heterogeneous data fusion module quantifies the contribution value of each pixel through the trained CNN neural network:

[0072] C subj_i,j =CNN(z subj_i,j )

[0073] Among them, C subj_i,j represents the contribution value of the subj_i-th subject at the j-th pixel, z subj_i,j is the input feature of the subj_i-th subject at the j-th pixel;

[0074] Weighted fusion of EEG data, fMRI data and clinical data,

[0075]

[0076] Among them, γ k is the weight of different data sources, C subj_i,k represents the contribution value of the subj_i-th subject under the k-th data source,

[0077] Use the stacking model to fuse multi-source heterogeneous data to obtain the final risk assessment.

[0078]

[0079] Among them, f stack It is a stacking model. It is the final risk assessment prediction result.

[0080] The Alzheimer's disease classification system based on multimodal data and brain-like neural computing provided by the present invention comprehensively considers the time-frequency characteristics of EEG activity and the functional connectivity between brain regions, which not only enhances the ability of feature extraction, but also improves the comprehensive understanding of the disease state. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0082] Figure 1 It is a schematic diagram of an Alzheimer's disease classification system based on multimodal data and brain-like neural computing of the present invention. DETAILED DESCRIPTION

[0083] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0084] like Figure 1 As shown, an Alzheimer's disease classification system based on multimodal data and brain-like neural computing includes: a data acquisition module, a data processing module, a first construction module, a second construction module, a node embedding module, a feature extraction module, a model training module and a classification module. The data acquisition module is used to collect EEG data and fMRI data of multiple subjects. The data processing module is used to preprocess the collected EEG data and fMRI data. The first construction module is used to build an EEG network based on the preprocessed EEG data. The second construction module is used to build a functional network based on the preprocessed fMRI data. The node embedding module learns the node embedding of the EEG network and the functional network through a graph convolutional neural network. The feature extraction module is used to extract and splice the node embeddings corresponding to the same ROI area of ​​multiple subjects to form a feature vector for each subject. The model training module is used to train a classification model based on the feature vectors of multiple subjects obtained by splicing. The classification module is used to output the classification results based on the trained classification model and the EEG data and fMRI data of the subject to be tested. The above modules are described in detail below.

[0085] The data acquisition module collects EEG data and resting-state functional magnetic resonance imaging (rs-fMRI) data of multiple subjects, hereinafter referred to as fMRI data.

[0086] The specific method for the data processing module to preprocess the collected EEG data is as follows:

[0087] The data processing module removes artifacts and noise from EEG data through independent component analysis (ICA).

[0088] Cleaned_EEG=Denoise(EEG)

[0089] The denoised EEG signal is then normalized to ensure that the amplitudes of all signals are at the same scale and to avoid the amplitude differences between different channels affecting the correlation calculation. Where μ is the mean of the EEG signal and σ is the standard deviation:

[0090]

[0091] In the implementation manner of the present application, the specific method in which the data processing module preprocesses the collected fMRI data is as follows:

[0092] In order to calculate the coordinated changes in blood oxygen level changes between brain regions through fMRI signals and construct a functional connection network, a functional connectivity analysis method based on BOLD (blood oxygen level dependent) signals was used.

[0093] First, slice timing correction is performed to ensure that the data of all brain regions are time-aligned, especially when scanning with different slices.

[0094] The original fMRI images are corrected for slice timing, head motion (such as the MCFLIRT tool from FSL), and spatial normalization (matching the images to a standard brain template) to eliminate artifacts caused by head motion and individual brain structure differences and ensure that the signal data at all time points are consistent:

[0095] Normalized_fMRI=Normalize(Motion_Corected_fMRI)

[0096] Gaussian smoothing was performed on the normalized fMRI data to reduce noise and improve signal smoothness:

[0097] Smoothed_fMRI=Gaussian_Smooth(Normalized_fMRI)

[0098] Then define the ROI (Regions of Interest) of the brain region and extract the BOLD signal sequence of each brain region from the fMRI data:

[0099] BOLD i (t) for i = 1, 2, ..., N

[0100] Among them, BOLD i(t) represents the BOLD signal of the i-th brain region at time point t, and N is the total number of brain regions. ROIs can be determined based on anatomical or functional maps (such as AAL or Harvard-Oxford brain region maps).

[0101] In the implementation manner of the present application, the specific method of the first building module to build an EEG network based on the preprocessed EEG data is:

[0102] For EEG data, the time domain Pearson correlation coefficient was calculated using the following formula:

[0103]

[0104] Among them, EEG i (t) represents the EEG signal of brain area i at time point t, represents the mean value of the EEG signal of brain area i at all time points, and T is the total number of time points of the signal.

[0105] Constructing time-domain EEG network based on time-domain Pearson correlation coefficient.

[0106]

[0107] Among them, θ_time is a preset threshold, which means that the connection will be established only when the time domain Pearson correlation coefficient is greater than the threshold θ_time.

[0108] Fast Fourier transform is performed on EEG data to convert time domain signals into frequency domain.

[0109] EEG freq (ROI i ,f)=FFT(EEG i (t))

[0110] Among them, EEG freq (ROI i ,f) represents the EEG signal of brain area i at frequency f.

[0111] Based on the transformed EEG data, the frequency domain Pearson correlation coefficient was calculated using the following formula:

[0112]

[0113] Among them, EEG i (t) represents the EEG signal of brain area i at frequency f, represents the mean value of the EEG signal of brain area i at frequency f.

[0114] Construct frequency domain EEG network based on frequency domain Pearson correlation coefficient:

[0115]

[0116] Among them, θ_freq is the threshold of frequency domain correlation, which means that the connection will be established only when the frequency domain Pearson correlation coefficient is greater than θ_freq.

[0117] The total EEG network is obtained by fusing the time domain EEG network and the frequency domain EEG network.

[0118] A total (i,j)=αA time (i,j)+βA freq (i,j)

[0119] Among them, α and β are the weight coefficients controlling the time domain and frequency domain networks, and the fused EEG network is expressed as:

[0120] A total =Fusion(A time ,A freq ).

[0121] In the implementation manner of the present application, the specific method of the second construction module to construct a functional network based on the preprocessed fMRI data is:

[0122] Construct a set of time series with the same ROI:

[0123] ROI_Collection = {Time_Series k |k=1,2,…,90}

[0124] Among them, Time_Series k Represents the time series set of the kth ROI, which is used to construct the node of the brain network.

[0125] The functional Pearson correlation coefficient was calculated using the following formula:

[0126]

[0127] Among them, BOLD i t) and BOLD j (t) represents the BOLD signal of brain area i and brain area j at time point t, BOLD i and are the mean BOLD signals of brain area i and brain area j, respectively, and T is the time point number of the BOLD signal;

[0128] Construct a functional network based on the functional Pearson correlation coefficient:

[0129]

[0130] θ_func represents a preset threshold, which means that a connection will be established only when the functional Pearson correlation coefficient is greater than the threshold θ_func.

[0131] In an embodiment of the present application, the node embedding module learns the node embedding of the EEG network and the functional network through a graph convolutional neural network. The graph convolutional neural network (GCN) is used to learn the node embedding of the multi-view brain network and further refine the functional connection pattern between brain regions. Through two layers of graph convolution layers, the functional features of each brain region can be mapped to a higher dimensional space.

[0132] The first convolution layer:

[0133]

[0134] Perform convolution operation on the input features, H (0) is the input matrix, W (0) is the weight matrix of the first layer, σ represents a nonlinear activation function (such as ReLU), represents the adjacency matrix after adding self-connection, is a diagonal matrix representing the degree of each row of the adjacency matrix.

[0135] The second convolution layer:

[0136]

[0137] The deep features are further extracted through the second convolutional layer.

[0138] In the implementation of the present application, the feature extraction module extracts and concatenates the node embeddings corresponding to the same ROI area of ​​multiple subjects to form a feature vector for each subject:

[0139] Z subj_i =[Z subj_i1 , Z subj_i2 , ..., Z subj_ik ]

[0140] Z subj_i represents the feature vector of the subj_i-th subject, Z subj_ik It represents the node embedding of the subj_i-th subject in the k-th ROI, and concatenates the node embeddings of each subject in different ROIs to form a comprehensive feature vector. This feature vector includes multi-dimensional information based on EEG time-frequency features and fMRI functional connectivity features.

[0141] A multi-layer perceptron (MLP) was designed for classification, and the extracted node embeddings were used as input. The classification model was obtained through training to accurately identify Alzheimer's patients and healthy controls.

[0142] MLP model design: MLP forward propagation

[0143] H (1) =σ(W (0) X+b (0) )

[0144] H (2) =σ(W (1) H (1) +b (1) )

[0145] …

[0146]

[0147] Where X is the input feature vector, H (1) is the hidden layer output of layer l, W (1) and b (1) are the weight and bias of the lth layer, It is the result of model prediction, and a multi-layer perceptron (MLP) model is designed to perform the diagnosis task.

[0148] Deep learning algorithms (such as convolutional neural networks and LSTM) are used to model the time series changes of brain area networks to further improve classification performance.

[0149] In an embodiment of the present application, the classification model adopts a CNN neural network and the training adopts a cross entropy loss function.

[0150] Specifically, based on the node embedding of the multi-view brain network, end-to-end training is performed through a deep neural network model. A convolutional neural network (CNN) is used to extract highly recognizable features from EEG and fMRI data to optimize model performance.

[0151] Training: Cross Entropy Loss Function

[0152]

[0153] Where N is the number of samples, C is the number of categories, and y subj_ik is the label that the subj_i-th sample actually belongs to the k-th class, is the probability that the model predicts that the subj_i-th sample belongs to the k-th class. The cross entropy loss function is used to calculate the training error of the model for optimization.

[0154] RMSprop optimizer update rules

[0155]

[0156] w t =w t-1 -ηv t +∈g t

[0157] Among them, v t is the cumulative average of the squared gradient, g t is the gradient of the tth iteration, β is the decay coefficient, η is the learning rate, ∈ is the smoothing term used, and the RMSprop optimizer performs gradient updates to update the model weights.

[0158] Use transfer learning technology to migrate the trained convolutional neural network model to improve training efficiency and retain the learned brain network knowledge.

[0159] In an embodiment of the present application, the Alzheimer's disease classification system based on multimodal data and brain-like neural computing also includes a three-dimensional image interpretation module.

[0160] The 3D image interpretation module quantitatively calculates the prediction results of brain sub-blocks and generates a whole-brain pixel risk value for each subject:

[0161] R subj_i =f(Z subj_i )

[0162] Among them, R subj_i represents the whole brain pixel risk value of the subj_i-th subject, f(Z subj_i ) is a function that calculates the risk value based on the trained model.

[0163] Sort high risk values ​​according to Pareto principle:

[0164] Sorted_Risk_Values=Sort(R subj_i, descending)

[0165] Where Sorted_Risk_Values ​​represents the sorted risk values, and the top 20% regions are selected.

[0166] Based on the risk value ranking, the high-risk areas of the brain network are combined with clinical data (such as cognitive scores) to obtain a comprehensive risk value:

[0167] Combined Risk =α·R subj_i +β·C subj_i

[0168] Among them, α and β are weight coefficients, C i Represents the clinical data of the subj_i-th subject.

[0169] In an embodiment of the present application, after generating the whole-brain pixel risk value of each subject, an individual whole-brain risk visualization map is generated based on the whole-brain pixel risk value of each subject to help clinicians perform visual analysis of lesions.

[0170] After generating the whole-brain pixel risk value for each subject, the Matthew correlation coefficient at the whole-brain pixel level can also be calculated using group data:

[0171]

[0172] in, is the category of the subj_ith sample predicted by the model, is the average of all sample prediction categories, is the actual category of the subj_i-th sample, is the average value of the actual categories of all samples.

[0173] The calculated Matthew correlation coefficients were visualized at the whole-brain pixel level to identify the most correlated brain regions in the group.

[0174] Group_MCC_Visualization=Visualize_MCC(MCC)

[0175] Among them, Group_MCC_Visualization represents the group visualization diagram based on MCC calculation.

[0176] In an embodiment of the present application, the Alzheimer's disease classification system based on multimodal data and brain-like neural computing also includes a multi-source heterogeneous data fusion module.

[0177] The multi-source heterogeneous data fusion module quantifies the contribution value of each pixel through the trained CNN neural network:

[0178] C subj_i,j =CNN(z subj_i,j )

[0179] Among them, C subj_i,j represents the contribution value of the subj_i-th subject at the j-th pixel, z subj_i,j It is the input feature of the subj_i-th subject at the j-th pixel.

[0180] Weighted fusion of EEG data, fMRI data and clinical data,

[0181]

[0182] Among them, γ k is the weight of different data sources, C subj_i,k represents the contribution value of the subj_i-th subject under the k-th data source,

[0183] Use the stacking model to fuse multi-source heterogeneous data to obtain the final risk assessment.

[0184]

[0185] Among them, f stack It is a stacking model. It is the final risk assessment prediction result.

[0186] The Alzheimer's disease classification system based on multimodal data and brain-like neural computing in this application also includes intelligent diagnosis and decision support modules. A deep learning model is used to perform real-time intelligent diagnosis of Alzheimer's disease and provide decision support for doctors. Through personalized diagnostic result reports, the patient's disease risk is prompted and further diagnosis and treatment recommendations are given.

[0187] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form, and any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. An Alzheimer's disease classification system based on multimodal data and brain-like neural computing, characterized in that: Include: Data acquisition module, used to collect EEG data and fMRI data of multiple subjects; A data processing module is used to pre-process the collected EEG data and fMRI data; The first building module is used to build an EEG network based on the preprocessed EEG data; The second building module is used to build a functional network based on the preprocessed fMRI data; Node embedding module, which learns node embedding of EEG networks and functional networks through graph convolutional neural networks; The feature extraction module is used to extract and concatenate the node embeddings corresponding to the same ROI of multiple subjects to form a feature vector for each subject; A model training module is used to train a classification model based on the concatenated feature vectors of multiple subjects; The classification module is used to output classification results based on the trained classification model and the EEG data and fMRI data of the subject to be tested.

2. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 1, characterized in that: The specific method of preprocessing the collected EEG data by the data processing module is: The data processing module performs artifact removal and denoising processing on the EEG data by independent component analysis, and then performs standardization processing.

3. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 1, characterized in that: The specific method of preprocessing the collected fMRI data by the data processing module is: The data processing module performs slice timing correction, spatial normalization and Gaussian smoothing on the fMRI data, defines the ROI of the brain region, and extracts the BOLD signal sequence of each brain region from the fMRI data: BOLD i (t)fori=1,2,...,N Among them, BOLD i (t) represents the BOLD signal of the i-th brain region at time point t, and N is the total number of brain regions.

4. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 3, characterized in that: The specific method of the first building module to build an EEG network based on the preprocessed EEG data is: For EEG data, the time domain Pearson correlation coefficient was calculated using the following formula: Among them, EEG i (t) represents the EEG signal of brain area i at time point t, represents the mean value of the EEG signal of brain area i at all time points, and T is the total number of time points of the signal; Construct time-domain EEG network based on time-domain Pearson correlation coefficient; Among them, θ_time is a preset threshold, which means that the connection will be established only when the time domain Pearson correlation coefficient is greater than the threshold θ_time; Fast Fourier transform of EEG data; EEG freq (ROI i ,f)=FFT(EEG i (t)) Among them, EEG freq (ROI i ,f) represents the EEG signal of brain area i at frequency f; Based on the transformed EEG data, the frequency domain Pearson correlation coefficient was calculated using the following formula: Among them, EEG i (t) represents the EEG signal of brain area i at frequency f, represents the mean value of the EEG signal of brain area i at frequency f; Construct frequency domain EEG network based on frequency domain Pearson correlation coefficient: Among them, θ_freq is the threshold of frequency domain correlation, which means that the connection will be established only when the frequency domain Pearson correlation coefficient is greater than θ_freq; The total EEG network is obtained by fusing the time domain EEG network and the frequency domain EEG network; A total (i,j)=αA time (i,j)+βA freq (i,j) Among them, α and β are the weight coefficients controlling the time domain and frequency domain networks, and the fused EEG network is expressed as: A total =Fusion(A time ,A freq )。 5. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 4, characterized in that: The specific method of the second construction module to construct a functional network based on the preprocessed fMRI data is: Construct a set of time series with the same ROI: ROI_Collection={Time_Series k |k=1,2,…,90} Among them, Time_Series k represents the time series set of the kth ROI, which is used to construct the nodes of the brain network; The functional Pearson correlation coefficient was calculated using the following formula: Among them, BOLD i t) and BOLD j (t) represents the BOLD signal of brain area i and brain area j at time point t, BOLD i and are the mean BOLD signals of brain area i and brain area j, respectively, and T is the time point number of the BOLD signal; Construct a functional network based on the functional Pearson correlation coefficient: θ_func represents the preset threshold.

6. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 5, characterized in that: The feature extraction module extracts and concatenates the node embeddings corresponding to the same ROI of multiple subjects to form a feature vector for each subject: WITH subj_i =[From subj_i1 ,WITH subj_i2 ,...,WITH subj_ik ] Z subj_i represents the feature vector of the subj_i-th subject, Z subj_ik Represents the node embedding of the subj_i-th subject in the k-th ROI, and concatenates the node embeddings of each subject in different ROIs to form a comprehensive feature vector.

7. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 6, characterized in that: The classification model uses CNN neural network and the training adopts cross entropy loss function.

8. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 7, characterized in that: The Alzheimer's disease classification system based on multimodal data and brain-like neural computing also includes a three-dimensional image interpretation module; The 3D image interpretation module quantitatively calculates the prediction results of the brain sub-blocks and generates a whole-brain pixel risk value for each subject: R subj_i =f(Z subj_i ) Among them, R subj_i represents the whole brain pixel risk value of the subj_i-th subject, f(Z subj_i ) is a function that calculates the risk value based on the trained model, and generates an individual whole-brain risk visualization map based on the whole-brain pixel risk value of each subject. Sort high risk values ​​according to Pareto principle: Sorted_Risk_Values=Sort(R subj_i ,descending) Where Sorted_Risk_Values ​​represents the sorted risk value, and the top 20% of regions are selected; Based on the risk value ranking, the high-risk areas of the brain network are combined with clinical data to obtain a comprehensive risk value: Combined Risk =α·R subj_i +β·C subj_i Among them, α and β are weight coefficients, C i Represents the clinical data of the subj_i-th subject.

9. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 8, characterized in that: After generating the whole-brain pixel risk value for each subject, the Matthew correlation coefficient at the whole-brain pixel level was calculated using the group data: in, is the category of the subj_ith sample predicted by the model, is the average of all sample prediction categories, is the actual category of the subj_i-th sample, is the average value of the actual categories of all samples; The calculated Matthew correlation coefficients were visualized at the whole-brain pixel level to identify the most correlated brain regions in the group. Group_MCC_Visualization=Visualize_MCC(MCC) Among them, Group_MCC_Visualization represents the group visualization diagram based on MCC calculation.

10. The Alzheimer's disease classification system based on multimodal data and brain-like neural computing according to claim 7, characterized in that: The Alzheimer's disease classification system based on multimodal data and brain-like neural computing also includes a multi-source heterogeneous data fusion module; The multi-source heterogeneous data fusion module quantifies the contribution value of each pixel through the trained CNN neural network: C subj_i,j =CNN(z subj_i,j ) Among them, C subj_i,j represents the contribution value of the subj_i-th subject at the j-th pixel, z subj_i,j is the input feature of the subj_i-th subject at the j-th pixel; Weighted fusion of EEG data, fMRI data and clinical data, Among them, γ k is the weight of different data sources, C subj_i,k represents the contribution value of the subj_i-th subject under the k-th data source, Use the stacking model to fuse multi-source heterogeneous data to obtain the final risk assessment. Among them, f stack It is a stacking model. It is the final risk assessment prediction result.

Images