A dementia recognition method based on electroencephalogram norm network and double-flow attention fusion

Abstract

This invention discloses a dementia identification method based on EEG norm network and dual-stream attention fusion, comprising: collecting resting-state EEG data of subjects and preprocessing the resting-state EEG data; constructing a standardized standard brain network pattern based on age-related changes; calculating the degree of deviation of each brain functional connection of the subject relative to the healthy norm; and extracting image feature vector I from the deviation matrix using a parallel dual-stream architecture. e Sum of numerical eigenvectors V e Based on image feature vector I e Sum of numerical eigenvectors V e Image features ViT were obtained respectively feat Sum of numerical features z feat , fusion of the ViT feat and z feat Obtain the final feature vector F fused The final feature vector F fused The data is fed into a downstream classifier for final classification and the final prediction result is output. This invention's dementia recognition method based on EEG norm networks and dual-stream attention fusion has advantages such as effectively fusing image and numerical features, improving the accuracy of dementia recognition, and enhancing the sensitivity of early diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to a brain-computer interface assisted diagnosis technology for dementia, and in particular to a dementia recognition method based on an electroencephalogram norm network and double-flow attention fusion. BACKGROUND

[0002] Dementia is a group of symptoms with cognitive function degradation (such as memory, thinking, and language ability decline) as the core, often accompanied by brain damage or cognitive abnormality degradation caused by the aging process, and Alzheimer's disease (AD) is the most common type. Notably, neurodegenerative diseases such as AD usually have a long incubation period and a hidden course, and their pathological changes (such as amyloid deposition) often begin years or even decades before the onset of clinical symptoms, making early recognition of the preclinical stage particularly important. In addition, dementia is not a single homogeneous disease, but a brain function impairment caused by multiple causes. Although its macroscopic manifestations are mainly memory loss, judgment decline, and emotional and behavioral abnormalities, different types of dementia have significant heterogeneity in neuropathological mechanisms and brain network damage patterns. This heterogeneity makes it difficult to distinguish disease subtypes simply by observing symptoms, such as distinguishing AD from mild cognitive impairment (MCI). Therefore, there is an urgent need for objective technical means that can simultaneously achieve "early sensitive recognition" and "accurate differential diagnosis" to cope with the increasingly severe aging health challenges.

[0003] The technical development of dementia recognition methods has undergone a paradigm shift from traditional empirical medicine to intelligent diagnosis. (1) Early reliance on clinical scales and single modality imaging, but early sensitivity is low and human dependence is strong; (2) After the 1990s, traditional machine learning realizes preliminary fusion of multiple modalities through manual feature extraction and classic classifiers (such as SVM), but is limited by complex feature engineering and insufficient generalization; (3) After 2010, deep learning promotes automatic feature extraction and time series modeling of images, but still faces the bottleneck of static graph assumption and weak multi-modal integration; (4) In recent years, graph neural networks (GNN) and graph convolution networks (GCN) have become core technologies due to their high efficiency, dynamic graph modeling capabilities, and multi-modal alignment potential. GCN generates time series graphs through sliding windows, captures long-range dynamic evolution of brain networks, and combines automatic preprocessing and multi-modal graph alignment, such as EEG time series and MRI spatial fusion, to significantly improve early recognition sensitivity and enhance clinical interpretability through Grad-CAM and other model visualization technologies.

[0004] The current core technical route for dementia diagnosis can be divided into three categories.

[0005] 1. Based on a static brain functional connectivity analysis scheme, a brain network is constructed using structural MRI (e.g., hippocampal volume measurement) and resting-state fMRI (BOLD signal). Static brain functional connectivity (e.g., Pearson correlation coefficient) between nodes is calculated using brain region activity data at single time points or short time windows (e.g., 5 minutes). Global indicators (clustering coefficient, global efficiency) and local indicators (node ​​degree) are extracted. Finally, binary classification is achieved by combining handcrafted features (e.g., hippocampal volume shrinkage rate) with an SVM / RBF classifier.

[0006] 2. Machine learning scheme based on handcrafted features, which involves manually defining multimodal features (such as hippocampal volume shrinkage rate in structural MRI, Aβ deposition intensity in PET, and EEG data). (Band power ratio), combined with traditional algorithms such as random forest (RF) and logistic regression (LR), learns the mapping relationship between features and disease state, and outputs binary classification results or risk scores.

[0007] 3. Based on a single-modal deep learning approach, the resting-state brain computer functional connectivity matrix or fMRI brain network connectivity matrix is ​​regarded as a two-dimensional image. CNN models (such as ResNet, VGG, etc.) are used to learn its spatial patterns to distinguish different diseases; or the original EEG time series of multiple channels is used as input, and RNN or Long Short-Term Memory Network (LSTM) is used to capture its temporal dynamic characteristics.

[0008] The three methods in the prior art mentioned above all have the following shortcomings.

[0009] 1. Lack of healthy norms for comparison makes it impossible to quantify individual abnormalities. Whether based on static brain connectivity analysis or traditional machine learning methods, most rely on directly comparing the statistical differences between patient groups (e.g., AD / MCI) and healthy controls (HC) in a specific dataset. This method ignores the objective biological laws governing the natural dynamic evolution of brain connectivity with age. Therefore, when evaluating a new individual, even if some of their brain connectivity indicators deviate from the "average level," it is difficult to determine whether this deviation stems from early pathological changes or is merely within the normal physiological fluctuations of their age. This simplistic, one-size-fits-all comparison method relies too heavily on the statistical differences between patient and healthy groups, ignoring the natural physiological evolution of brain function with age, leading to ambiguous diagnostic criteria. Especially in the early stages of disease, weak pathological signals are easily masked by individual differences and age effects, resulting in insufficient sensitivity and specificity in early diagnosis, leading to erroneous diagnostic conclusions.

[0010] 2. A single modality (image or numerical) is insufficient to comprehensively represent pathological patterns. The functional connectivity (FC) matrix is ​​an information-rich data structure containing two different types of features. First, there are numerical topological features, namely the precise connection strength value of each element in the matrix, which can be quantified as network degree, clustering coefficient, and other indicators, reflecting the topological properties of the brain network. Second, there are spatial distribution features, namely the specific patterns of strong and weak connections formed throughout the brain, such as weakened connections in the default mode network (DMN) and abnormal connections in the frontoparietal network. These patterns can be visually represented through images such as heatmaps. Existing solutions often utilize only one modality: machine learning methods typically flatten the matrix into a one-dimensional vector, completely losing spatial distribution information; while deep learning methods, although capable of processing images, may neglect to preserve precise connection strengths during convolution and pooling. This single-dimensional processing approach, relying solely on vectorization (losing spatial patterns) or visualization (ignoring precise numerical values), results in severe information loss. This single-dimensional utilization of features limits the model's comprehensive understanding and characterization of complex pathological states.

[0011] 3. The fusion methods are simplistic and fail to capture dynamic interactions between modalities. Even though some studies have attempted to fuse multimodal features, such as concatenating hippocampal volume and EEG frequency band power, their fusion strategies are merely static feature concatenation or linear weighting, failing to capture deep intermodal relationships. Static fusion methods assume that the information contributions of different modalities are fixed, failing to explore deep, non-linear complementary relationships between modalities. For example, a slight abnormality in the connectivity strength of a certain brain region (numerical modality) may only have decisive diagnostic significance when the whole-brain connectivity presents a specific spatial pattern (image modality). Simple concatenation cannot capture this dynamic, context-dependent interactive information and may even introduce noise redundancy, failing to fully realize the potential of multimodal data fusion and ultimately limiting the improvement of model classification performance. Summary of the Invention

[0012] To avoid the shortcomings of the existing technologies, this invention provides a dementia recognition method based on EEG norm network and dual-stream attention fusion, which effectively integrates image features and numerical features, and improves the recognition accuracy and early diagnosis sensitivity of dementia.

[0013] The present invention adopts the following technical solution to solve the technical problem.

[0014] The present invention provides a dementia identification method based on EEG norm network and dual-stream attention fusion, comprising the following steps:

[0015] Step 1: Data Acquisition and Preprocessing; Acquire resting-state EEG data of the subjects and preprocess the resting-state EEG data;

[0016] Step 2: Modeling Steps; Construct a standardized EEG functional connectivity network model that evolves with age, i.e., a norm network;

[0017] Step 3: Quantification of individual brain function deviation; Calculate the deviation z of each brain functional connection of the subject relative to the corresponding functional connection edge in the healthy norm network. i ; and the degree of deviation of all brain function connections z i Reorganized into whole-brain deviation matrix Z mat This serves as the input data source for subsequent feature extraction in the dual-stream architecture.

[0018] Step 4: Multimodal feature extraction; using a parallel two-stream architecture to extract features from the aforementioned whole-brain deviation matrix Z. mat Extracting image feature vector I e Sum of numerical eigenvectors V e ;

[0019] Step 5: Cross-modal attention fusion and classification diagnosis; based on image feature vector I e Sum of numerical eigenvectors V e Image features ViT mapped to the same high-dimensional semantic space were obtained respectively. feat Sum of numerical features z feat A bidirectional cross-modal attention mechanism is employed to dynamically fuse the two: on the one hand, numerical features z... feat As a query vector, it retrieves image features ViT as key and value vectors. feat This allows for the acquisition of enhanced numerical features that incorporate spatial distribution information; on the other hand, image features ViT feat Used as a query vector, to retrieve the numerical features z that serve as both the key vector and the value vector. feat This process yields enhanced image features incorporating precise numerical intensities. The enhanced features output from the interaction of the two directions are then concatenated to obtain the final feature vector F. fused The final feature vector F fused The data is fed into a downstream classifier for final classification and the final prediction result is output.

[0020] The dementia identification method based on EEG norm network and dual-stream attention fusion of the present invention is also characterized by:

[0021] Furthermore, the subjects in step 1 include patients and healthy control groups.

[0022] Furthermore, the preprocessing in step 1 includes bandpass filtering, removal of physiological artifacts, and signal segmentation.

[0023] Furthermore, in step 2, the preprocessed time segments are processed to obtain the brain functional connectivity feature matrix FC.

[0024] Furthermore, in step 2, the QRF algorithm is used to construct a standardized standard EEG functional network model based on age.

[0025] Furthermore, the QRF model of the QRF algorithm is implemented using the LightGBM framework.

[0026] Furthermore, in step 3, the deviation z of each brain functional connection of the subject relative to the corresponding functional connection edge in the healthy norm network is calculated using a standardized formula. i .

[0027] Furthermore, in step 4, the dual-stream architecture includes an image modal stream and a numerical modal stream.

[0028] Furthermore, the ViT model is used to obtain the image feature vector I in the image modal stream. e The numerical modal flow uses a decision tree model to extract numerical feature vectors V. e .

[0029] Furthermore, in step 5, the image feature vector I e Sum of numerical eigenvectors V e Image features ViT are obtained by mapping their respective encodings to the same high-dimensional semantic space. feat Sum of numerical features z feat .

[0030] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0031] This invention discloses a dementia identification method based on EEG norm network and dual-stream attention fusion, comprising the following steps: data acquisition and preprocessing; acquiring resting-state EEG data of subjects and preprocessing the resting-state EEG data; modeling; constructing a standardized standard EEG functional network model based on age evolution using the QRF algorithm (quantile regression forest algorithm); quantification of individual brain function deviation; calculating the degree of deviation of each brain functional connection of the subject relative to the healthy norm; and multimodal feature extraction; extracting image feature vectors I from the deviation matrix using a parallel dual-stream architecture. e Sum of numerical eigenvectors V e Cross-modal attention fusion and classification diagnosis steps; based on image feature vector I e Sum of numerical eigenvectors V e Image features ViT were obtained respectively feat Sum of numerical features z featBy employing a bidirectional cross-modal attention mechanism, numerical features and image features act as query vectors and key-value pairs to complement each other in retrieving information, thereby fusing them to obtain the final feature vector F. fused The final feature vector F fused The data is fed into a downstream classifier for final classification and the final prediction result is output.

[0032] The dementia identification method based on EEG norm network and dual-stream attention fusion of the present invention can effectively integrate image deviation features and numerical deviation features relative to the norm network, and has the advantages of improving the identification accuracy and early diagnosis sensitivity of dementia. Attached Figure Description

[0033] Figure 1 This is a flowchart of the dementia identification method based on EEG norm network and dual-stream attention fusion of the present invention;

[0034] Figure 2 This is a diagram of the Norm-CrossNet architecture of the present invention;

[0035] Figures 3-6 The graph shows the performance evaluation of the method of this invention from four dimensions: accuracy, F1-score, recall / sensitivity, and precision.

[0036] Figure 7 These are the experimental results (confusion matrix diagram) of the method of the present invention;

[0037] Figure 8 This is the ROC curve analysis diagram of the present invention;

[0038] Figure 9 This is the final fusion feature vector F extracted by the present invention. fused UMAP dimensionality reduction visualization analysis chart;

[0039] Figures 10-12 This is a visualization of the Z-score deviation matrix of the present invention; wherein, Figure 10 A visualization of the Z-score deviation matrix for mild cognitive impairment (MCI); Figure 11 A visualization of the Z-score deviation matrix for Alzheimer's disease (AD); Figure 12 A visualization of the Z-score deviation matrix for the healthy control group HC;

[0040] Figure 13 for Figure 2 A schematic diagram of the age normalization norms in QRF.

[0041] The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings. Detailed Implementation

[0042] The technical terms used in this invention are explained below.

[0043] 1. Constructing a sequence diagram;

[0044] The sliding window technique transforms continuous EEG time series into dynamic network structure sequences. Specifically, within each time window, EEG electrode channels are defined as network nodes, and the calculated functional connectivity strength between channels is defined as network edge weights, thereby generating a series of instantaneous functional connectivity matrices characterizing brain activity states. This process not only preserves the spatial topological information of EEG signals but also captures the dynamic evolution of brain functional connectivity over time through time slicing, providing a standardized spatiotemporal data foundation for subsequent norm comparison and feature extraction.

[0045] 2. Network norm generation;

[0046] Typical network patterns of healthy individuals were extracted through population brain network clustering. A baseline pattern of healthy brain networks varying with age was constructed using the quantile regression forest (QRF) algorithm. The LightGBM framework based on gradient boosting decision trees (GBDT) was utilized, with subject age as a covariate and brain functional connectivity strength as the response variable, targeting different quantiles (e.g., age, brain functional connectivity strength ... The regression model was trained using the formulas 0.025, 0.5, and 0.975. This model learns the conditional probability distribution of brain connectivity strength at different age points by minimizing the pinball loss function, thus obtaining a dynamic confidence interval that covers the physiological fluctuations of most healthy individuals. The norm construction not only quantifies the physiological impact of age on brain function but also provides an age-corrected personalized assessment benchmark for subsequent anomaly detection, supporting accurate pathological identification.

[0047] 3. Bidirectional cross-modal multi-head attention mechanism;

[0048] This invention employs a bidirectional multi-head attention mechanism as its core modality fusion module. Its purpose is to replace the traditional simple concatenation or weighting of feature vectors, achieving deep dynamic interaction and information complementarity between numerical and image modalities. Its working principle is as follows:

[0049] (1) Feature embedding: First, the features from different modalities, i.e., the numerical connection strength vector V, are embedded. e and image global mode vector I e Each encoder projects its own image onto the same high-dimensional semantic space to obtain z. feat and ViT featThis step unifies the feature dimensions, enabling them to interact within the attention mechanism;

[0050] (2) Bidirectional information query: The core of this mechanism is two parallel attention processes:

[0051] Numerical mode focuses on image modes: with z feat For Query, go to ViT feat Find the most relevant key in ViT and evaluate it based on the relevance score. feat The values ​​are weighted and summed.

[0052] Image modalities focus on numerical modalities; conversely, ViT... feat For Query, remove z feat Find relevant keys and weight them accordingly;

[0053] (3) Information fusion: The attention outputs from the two directions are concatenated to form an enhanced feature vector containing bidirectional complementary information. This vector contains both image information weighted from a numerical perspective and numerical information weighted from an image perspective, and is finally fed into the classifier for decision-making.

[0054] See Figures 1-12 The present invention provides a dementia identification method based on EEG norm network and dual-stream attention fusion, comprising the following steps:

[0055] Step 1: Data Acquisition and Preprocessing; Acquire resting-state EEG data of the subjects and preprocess the resting-state EEG data;

[0056] Step 2: Modeling Steps; Constructing an age-based standardized standard EEG functional network model; Specifically, using the Quantile Regression Forest (QRF) algorithm, a healthy network norm that dynamically evolves with age is constructed based on the LightGBM framework;

[0057] The health norms were established based on the publicly available large-scale dataset CUAEEG. This dataset covers a specific age group (70-90 years old) and provides normative data support for the overall average health standards of the aging population. In this invention, multiple electrode channels are paired to form a channel pair, and the brain signal coordination relationship between the two electrodes in any channel pair constitutes a brain functional connection; assuming there are 19 electrode channels, then there are 19 × (19-1) / 2 = 171 channel pairs; that is, when the number of electrodes is 19, there are 171 brain functional connections.

[0058] Step 3: Quantification of individual brain function deviation; Calculate the deviation z of each brain functional connection of the subject relative to the corresponding connection in the healthy norm network. i ;

[0059] Step 4: Multimodal feature extraction step; Image feature vector I is extracted using a parallel two-stream architecture. e Sum of numerical eigenvectors V e ;

[0060] Step 5: Cross-modal attention fusion and classification diagnosis; based on image feature vector I e Sum of numerical eigenvectors V e Image features ViT were obtained respectively feat Sum of numerical features z feat ViT, which fuses image features feat Sum of numerical features z feat Obtain the final feature vector F fused The final feature vector F fused The data is fed into a downstream classifier for final classification and the final prediction result is output.

[0061] like Figure 1 This is a flowchart of a dementia identification method based on EEG norm network and dual-stream attention fusion according to the present invention. Figure 2 The complete Norm-CrossNet architecture of this invention is demonstrated. The complete Norm-CrossNet architecture comprises the following four modules: starting with data acquisition and preprocessing (a), followed by norm establishment and deviation calculation (b), which demonstrates normalization modeling based on the functional connectivity matrix using a QRF model to obtain the deviation matrix; then, a two-stream parallel feature extraction module (c) extracts image feature vectors representing spatial distribution patterns in parallel using a two-stream architecture. e and numerical eigenvectors V representing numerical topological structures e In cross-modal attention fusion (d), the feature vectors processed by the encoder are used to generate the final discriminative feature vectors, which are then input into the MLP classifier to achieve classification among Alzheimer's disease patients (AD), healthy controls (HC), and patients with mild cognitive impairment (MCI). Figure 13 for Figure 2 A schematic diagram of the age normalization norms in QRF.

[0062] In practice, the subjects in step 1 include patients and healthy control groups.

[0063] In practice, the preprocessing in step 1 includes bandpass filtering, removal of physiological artifacts, and signal segmentation.

[0064] In its implementation, this invention utilizes the publicly available dataset CUAEEG. To further standardize the modeling, the dataset is designed for participants aged 70 to 90 years. After screening, the final sample of 193 participants was included in this study, comprising 57 patients with Alzheimer's disease (AD), 62 patients with mild cognitive impairment (MCI), and 74 healthy controls (HC). Electrode placement followed the international 10-20 standard, with 19 channels located at Fp1, F3, C3, P3, O1, Fp2, F4, C4, P4, O2, F7, T3, T5, F8, T4, T6, Fz, Cz, and Pz.

[0065] Resting-state electroencephalogram (EEG) data were collected from all 193 subjects. A series of standardized preprocessing operations were performed on the raw EEG data, primarily including: bandpass filtering of 0.5–45 Hz to remove low-frequency drift and high-frequency electromyographic noise; identification and removal of physiological artifacts such as electrooculography and electrocardiography using methods such as independent component analysis (ICA); and segmentation of continuous EEG signals into several non-overlapping time segments to provide independent samples for subsequent analysis.

[0066] In specific implementation, step 2 involves processing the preprocessed time segments to obtain the brain functional connectivity feature matrix FC.

[0067] For each time segment, the Welch method is used to estimate the power spectral density (PSD) by calculating the squared amplitude correlation between all electrode channels. Specifically, for N electrode channels (N=19 in this invention), all channel pairs (i, j) are iterated, and the scipy.signal.coherence function is called with the segment length of the Welch method set to calculate the squared amplitude coherence. The scipy.signal.coherence function returns the coherence vector C at different frequency points f. xy The N electrode channels constitute N×(N-1) / 2 brain functional connections. In one embodiment of the present invention, assuming N=19, then there are 19×(19-1) / 2=171 brain functional connections.

[0068] Based on the coherence vector C xy The obtained coherent spectrum is obtained in five preset classical frequency bands: (1) Delta: 1-4 Hz; (2) Theta: 4-8 Hz; (3) Alpha: 8-13 Hz; (4) Beta: 13-30 Hz; (5) Gamma: 30-45 Hz. Then, the different frequency points within each frequency band are averaged. Among them, the frequency band averaging is calculated by calculating the coherence vector C. xy This is achieved by using the average value band_val within a preset frequency band, i.e. , where fmin f max These are the start and end frequencies of each of the five classic frequency bands, respectively, with mean representing the average value operator. Then, the final brain functional connectivity feature matrix FC is obtained for modeling. The parameter in the brain functional connectivity feature matrix FC is the brain functional connectivity strength value band_val. The brain functional connectivity feature matrix FC is an N×N symmetric matrix, and... (i and j are the electrode channel indices, respectively). Since the brain signal coordination relationship between channel pairs (i,j) and (j,i) has bidirectional consistency, the matrix is ​​symmetric. The brain functional connectivity feature matrix FC can be expressed as the following formula (1);

[0069] (1)

[0070] The functional connectivity matrix (FC) comprehensively depicts the topological state of the whole brain functional network of the subject in this frequency band, and serves as the basis for subsequent modeling and analysis.

[0071] In specific implementation, step 2 involves using the QRF algorithm to construct a standardized EEG functional network model based on age.

[0072] In practice, the QRF model of the QRF algorithm is implemented using the LightGBM framework.

[0073] To construct the evolutionary trajectory of EEG functional connectivity in healthy individuals at different ages, this invention employs the Quantile Regression Forest (QRF) algorithm for standardized modeling. This invention constructs an independent QRF model for each individual brain functional connectivity. The QRF algorithm can not only accurately depict the "normative trajectory" of how brain functional connectivity strength dynamically changes with age, but also comprehensively model the nonlinear relationships and heteroscedasticity (i.e., data volatility changing with age) that may exist in neurophysiological data, making it extremely suitable for population modeling and individual anomaly detection of complex EEG biomarkers. In this invention, the main covariate for standardizing the EEG functional network model is age, and the response variable is the brain functional connectivity strength value (FC) between specific channel pairs. ij This refers to the element in the i-th row and j-th column of the functional connectivity feature matrix (FC). This indicator directly reflects the collaborative efficiency of information transmission between brain regions and has clear neurophysiological significance.

[0074] This invention uses a QRF model implemented based on the LightGBM framework and fits it to data from a healthy control group (HC). The fitting process is as follows:

[0075] (1) Use lightgbm.LGBMRegressor and set objective='quantile' to implement quantile regression. For the i-th brain functional connection (i=1-171), use the age variable hc_age of the healthy control group (HC) as the input feature, and use the brain functional connection strength value of this connection in the HC group as the input feature. The response variable (or output) is the target value that the model is trying to predict in regression analysis and machine learning modeling. It is also known as the "output" or "label".

[0076] (2) For the three quantiles ( =0.025 corresponds to the lower bound. =0.5 corresponds to the median. =0.975 corresponds to the upper bound) Three independent LGBM models are trained respectively. The expression of the LGBM model for the i-th brain functional connection can be summarized as the following formula (2);

[0077] (2)

[0078] In formula (2) represents the QRF model corresponding to the i-th brain functional connection (i=1-171); LGBMRegressor represents the regression model function based on the LightGBM framework, used to implement quantile regression; Indicates the quantile parameter. The values ​​are 0.025 (corresponding to the lower bound of the 95% confidence interval), 0.5 (corresponding to the median), or 0.975 (corresponding to the upper bound of the 95% confidence interval);

[0079] objective='quantile' indicates that the model's objective function is set to quantile regression, which is used to learn the conditional probability distribution of the response variable.

[0080] Unlike traditional regression models that only focus on predicting the mean of data, QRF models aim to learn and predict the complete conditional probability distribution of the response variable given covariates. This is achieved by minimizing an asymmetric loss function—the Pinball Loss function—the mathematical formula of which is shown in Equation (3) below.

[0081] (3)

[0082] Formula (3), where y is the actual brain functional connectivity strength value. The QRF model represents the quantile of the brain functional connectivity strength value. The predicted value. By adjusting... The value can model any location in the distribution. In this invention, three independent QRF models were trained for each brain functional connection to accurately estimate its median at any age. =0.5), and the lower bound of the 95% confidence interval ( =0.025) and upper bound ( =0.975).

[0083] After fitting the QRF model, a complete normalized trajectory and confidence interval representing the age-related changes in brain functional connectivity strength in healthy individuals are obtained. The normalized trajectory is a set of LGBM models trained separately for each brain functional connectivity at each quantile, totaling 171 × 3 = 513 LGBM model sets. When a new subject's age is input, the expected healthy trajectory and confidence interval for the i-th brain functional connectivity are calculated using the following formula:

[0084] Median Expectation: ;

[0085] Lower bound of 95% confidence interval: ;

[0086] Upper bound of 95% confidence interval: ;

[0087] The q med,i M represents the expected median of the healthy norm for the i-th brain functional connection at the subject's age; i,0.5 This represents the i-th brain functional connection. The LGBM quantile regression model with q = 0.5; low,i M represents the lower bound of the 95% confidence interval for the healthy norm of the i-th brain functional connectivity; i,0.025 This represents the i-th brain functional connection. The LGBM quantile regression model with q = 0.025; high,i M represents the upper bound of the 95% confidence interval for the healthy norm of the i-th brain functional connectivity; i,0.975 This represents the i-th brain functional connection. The LGBM quantile regression model has a value of 0.975; the predict([age]) function is the model prediction function, with the subject's age as the input and the median predicted value at that age as the output.

[0088] To assess the extent to which Alzheimer's disease (AD) patients deviate from this healthy trajectory, brain functional connectivity data from AD patients were compared with the predicted distribution of age-matched healthy individuals. Specifically, each subject's deviation value (i.e., the Z-score calculated based on the median and confidence interval of their age-matched HC group predictions) was used as input features to train a subsequent multimodal classification model. This workflow enables standardized modeling not only to reveal subject-specific brain functional connectivity biases but also to provide downstream deep learning models with informative, individually calibrated, structured inputs, thereby significantly enhancing the efficiency of integrating and utilizing personalized neurobiomarkers within the framework of neurodegenerative disease analysis.

[0089] In specific implementation, step 3 involves calculating the degree of deviation z of each brain functional connection of the subject relative to the healthy norm using a standardized formula. i .

[0090] For a new subject, the brain functional connectivity matrix is ​​first calculated using the Welch method. Then, the subject's age is input into a pre-trained QRF norm to obtain the expected median and range of fluctuation of brain functional connectivity strength in healthy individuals of the same age.

[0091] We used the predicted median and robust standard deviation (calculated from 95% confidence intervals) output by the QRF model to calculate the deviation of each brain functional connection of the subject from the healthy norm using a standardized formula.

[0092] For the actual value fc_vec[i] of the i-th brain functional connection of the subject, the deviation z i (i.e., Z-score) is calculated through the following steps:

[0093] 1) Based on the subject's age, obtain the median q of their age-matched health norms. med,i and confidence interval .

[0094] 2) Calculate an approximate value for the robust standard deviation. .

[0095] Where SD in the above formula robust,i Let q represent the robust standard deviation of the i-th brain functional connection. high,i q low,i These are the upper and lower bounds of the 95% confidence interval for the healthy norm of this brain functional connection, respectively, and 3.92 is the coefficient (2 × 1.96) corresponding to the 95% confidence interval under the normal distribution.

[0096] 3) Calculate the degree of deviation from standardization. (If SD)robust,i If it is close to 0, then use the minimum value 1e. -6 (replace the denominator). Where z in the above formula... i The standardized deviation (Z-score) of the i-th connection is represented; fc_vec[i] represents the actual strength value of the i-th functional brain connection of the subject; q med,i This represents the expected median of the health norms for this connection;

[0097] 4) Deviation values ​​z of all brain functional connections i Reorganized into a deviation matrix Z mat Deviation value z of a brain functional connection i Construct a one-dimensional vector Z vec (The dimension is the total number of brain functional connections). One-dimensional vector Z vec The following steps are used to reorganize it into an N×N symmetric deviation matrix Z. mat :

[0098] (1) Initialize an N×N zero matrix Z mat ;

[0099] (2) Obtain the index triu_idx of the upper triangle (excluding the diagonal);

[0100] (3) Z vec The value is filled into Z. mat The upper triangular part: Z mat [triu_idx]=Z vec ;

[0101] (4) Through Make it a symmetric matrix; where Z in the left-hand side mat This represents the final constructed N×N symmetric deviation matrix; the right-hand side Z mat Let denote the asymmetric matrix after initially filling the upper triangular portion; let denote the transpose of the initial matrix, which fills the lower triangular portion by the transpose, making the offset matrix satisfy Z. mat [i][j]=Z mat [j][i] conforms to the bidirectional symmetry of brain functional connectivity.

[0102] The deviation matrix Z mat Each element precisely represents whether a certain brain functional connection is abnormally enhanced (z>0) or weakened (z<0), as well as the severity of the abnormality, making it an individualized and standardized pathological feature atlas.

[0103] In specific implementation, step 4, the dual-stream architecture includes an image modal stream and a numerical modal stream.

[0104] In specific implementation, the ViT model is used to obtain the image feature vector I in the image modal stream.e The numerical modal flow uses a decision tree model to extract numerical feature vectors V. e .

[0105] In order to fully extract information from the deviation matrix, this invention designs a parallel dual-stream architecture that treats it as two data modalities (image modality and numerical modality) for feature extraction.

[0106] Image modal stream: The deviation matrix is ​​rendered into a two-dimensional heatmap image. This two-dimensional heatmap image is then fed into a Vision Transformer (ViT) model. The ViT model can effectively capture the macroscopic spatial distribution patterns of anomalous connections formed throughout the brain. In this embodiment, the microsoft / swin-base-patch4-window7-224 model is used; the Swin Transformer is a variant of ViT.

[0107] 1. Preprocessing: ViTFeatureExtractor is used as the preprocessor to load the rendered 2D heatmap image, convert it to RGB format, and perform standard preprocessing (e.g., enlarge to 224*224 pixels, normalize) to adapt to the model input requirements.

[0108] 2. ViT Model Extraction: The ViTModel (specifically, the Swin Transformer architecture) is used for forward propagation of the preprocessed image tensors. In this embodiment of the invention, the microsoft / swin-base-patch4-window7-224 pre-trained model is specifically selected to ensure the effectiveness of feature extraction.

[0109] 3. Final Features: Extract the embedding vectors corresponding to the [CLS] tokens (Class Tokens) from the last hidden state of the ViT model. This vector (768 dimensions in this example) is the image feature vector I representing the spatial distribution pattern of the entire heatmap. e I e This is the extracted final image feature vector.

[0110] ViTFeatureExtractor is a preprocessor used to load rendered 2D heatmap images, convert them to RGB format, and perform standard preprocessing; ViTModel (Swin Transformer) is a visual Transformer model. In this embodiment of the invention, the microsoft / swin-base-patch4-window7-224 model is used to perform forward propagation on the preprocessed image tensor to extract features.

[0111] The 0th token ([CLS] token) of the last hidden state of the ViT model is used to aggregate the global features of the entire image, with an output dimension of 768.

[0112] The [CLS] token (ClassToken) is a special token in the ViT model. The embedding vector corresponding to this token in the last hidden state of the model is used as a feature to represent the spatial distribution pattern of the entire heatmap.

[0113] Numerical Modal Flow: The upper triangular portion of the deviation matrix is ​​flattened to form a one-dimensional numerical vector. This one-dimensional numerical vector is used as structured data input into a gradient boosting decision tree model, LightGBM. Features are extracted using the tree structure within the LightGBM model to capture the nonlinear combination relationships between deviation values ​​of different connection strengths.

[0114] The one-dimensional numerical vector Z vec (The dimension represents the total number of brain functional connections) will be used as input to the Gradient Boosting Decision Tree Model (LightGBM). The processing procedure is as follows:

[0115] (1) Training of the decision tree model: First, using Z from the training set vec A LightGBM classifier (LGBMClassifier) ​​is trained using the classifier and its corresponding labels (such as AD, MCI, HC).

[0116] (2) Feature Extraction: After training, the LightGBM classifier is used as a feature extractor. For any one-dimensional numerical vector Z... vec (Whether it's the training set or the test set), input it into the already trained LightGBM model.

[0117] (3) Obtain the leaf node index: Obtain the one-dimensional numerical vector Z by using the predict method of the decision tree model (e.g., setting pred_leaf=True in Python). vec The index of the leaf node that eventually falls on each decision tree in the forest.

[0118] (4) Final features: If the decision tree model has M trees (e.g., M=100), then each one-dimensional numerical vector Z vec All of these will be converted into a new vector containing M leaf node indices. This new vector (or its One-Hot encoding) is the final extracted numerical feature vector V. e It captures the original one-dimensional numerical vector Z. vec Deviation value z of different brain connectivity strengths i The nonlinear combination relationship between them.

[0119] In specific implementation, in step 5, the image feature vector I e Sum of numerical eigenvectors V e Numerical features z are obtained by mapping their respective encodings to the same high-dimensional semantic space. feat and image features ViT feat .

[0120] Simple feature concatenation simply places two feature vectors side-by-side, essentially assuming that information from different modalities is independent and has a fixed contribution. This static fusion method cannot complement, verify, or enhance the deep, dynamic interactions between modalities. However, in clinical diagnosis, a weak numerical anomaly is only significant under specific spatial patterns. Therefore, this invention introduces a bidirectional cross-modal attention mechanism, allowing the model to dynamically and selectively focus on and fuse information from two modalities, much like an expert.

[0121] The image feature vector I extracted in step 4 e Sum of numerical eigenvectors V e Image features ViT are obtained by mapping their respective encodings to the same high-dimensional semantic space. feat Sum of numerical features z feat Among them, z feat and ViT feat It is the original feature z(V) e Numerical characteristics) and ViT (I e The image features are mapped to the same high-dimensional semantic space (e.g., embed_dim=96) by their respective encoders (self.z_encoder and self.ViT_encoder). feat The calculation formula is shown in formula (4) below.

[0122] (4);

[0123] In formula (4), where zval Linear is the original one-dimensional numerical eigenvector output by the numerical modal stream. num A fully connected encoder specifically for numerical modes (parameters are learnable); ReLU is a rectified linear activation function, and Dropout is a random deactivation layer to prevent overfitting; z feat Numerical feature vectors that unify the high-dimensional semantic space after encoding.

[0124] ViT feat The calculation formula is shown in the following formula (5).

[0125] (5);

[0126] In formula (5), where v img The original one-dimensional image feature vector (spatial pattern features extracted by the ViT model) output by the image modal stream, Linear img A fully connected encoder for image modalities (parameters are learnable), ReLU is a rectified linear activation function, and Dropout is a random deactivation layer to prevent overfitting; v feat Image feature vectors that are encoded to unify the high-dimensional semantic space.

[0127] Next is bidirectional information query, the core of which is two parallel cross-query processes based on standard Transformer attention, and its calculation formula is shown in the following formula (6).

[0128] (6);

[0129] In formula (6), Attention is the output of the attention mechanism; Q (Query) is the query vector; K (Key) is the key vector; V (Value) is the value vector; d k is the dimension of the key vector K, used to scale the dot product result and prevent the gradient from becoming too small. Softmax is the normalized exponential function used to convert the dot product result into attention weights between 0 and 1.

[0130] This invention achieves bidirectional querying by exchanging the sources of Q (Query) and K (Key) / V (value):

[0131] (1) Numerical “focus” image: based on numerical feature z feat As a query, remove image features ViT feat Find the most relevant spatial pattern (Key / Value) in the data so that the representation of the numerical feature incorporates the information of its spatial background, as shown in the following formula (7).

[0132] (7);

[0133] In formula (7), query=z feat Numerical feature vector (as the query term); key=ViT feat The image feature vector (used as the key for the query); value=ViT feat The image feature vector (as the value to be weighted); Enhanced numerical features to incorporate spatial pattern information.

[0134] (2) Image “attention” value: Conversely, using the image feature ViTfeat as the query, remove the numerical feature z feat The most critical connection strength values ​​are found in the spatial pattern representation so that the most significant outliers can be highlighted, as shown in the following formula (8).

[0135] (8);

[0136] In formula (8), query=ViTfeat is the image feature vector (as the query term); key=z feat This is a numerical feature vector (serving as the key to be queried); value=z feat The numerical feature vector (as the value to be weighted); Enhanced numerical features to incorporate spatial pattern information.

[0137] Finally, the attention-weighted outputs from the two directions, enhanced with information, are fused to form a final two-dimensional feature vector F that is information-complementary, dynamically enhanced. fused See formula (9) below.

[0138] (9);

[0139] In formula (9) The output of Formula 7: Enhanced numerical features. The output of Equation 8: Enhanced image features, where Concat is the feature concatenation function, F fused This results in the final two-dimensional fused feature vector containing bidirectional complementary information.

[0140] Finally, the vector F fused It is fed into a downstream classifier to make the final classification decision and output the final prediction result, as shown in the following formula (10).

[0141] (10);

[0142] In formula (10), MLP refers to the downstream classification model, the multilayer perceptron classifier. This module receives high-dimensional features F. fusedThrough multi-layer nonlinear transformation, the mapping relationship between features and disease labels is learned.

[0143] In summary, the present invention has the following technical features.

[0144] 1. Utilizing a quantile regression forest (QRF) model based on large-scale healthy control (HC) data, this invention accurately constructs a nonlinear "normative trajectory" of brain functional connectivity strength changing with age by minimizing the pinball loss function. The QRF model is an EEG norm network. Therefore, this invention can convert the brain functional connectivity data of any new subject (whether AD, MCI, or HC) into a standardized Z-score deviation matrix relative to their age-matched healthy population. This aims to achieve a paradigm shift from "group comparison" to "individual quantification," thereby enabling accurate and objective assessment of the severity of an individual's deviation from the healthy trajectory, resolving diagnostic ambiguity caused by individual differences and age effects, and significantly improving the sensitivity to capturing early, weak pathological signals such as MCI. Based on this, it solves the problem of existing technologies relying on statistical differences between patient and healthy groups, ignoring the natural physiological evolution of brain function with age, leading to ambiguous diagnostic criteria.

[0145] 2. A parallel two-stream architecture is used to fully characterize these two types of information, achieving two-stream attention fusion;

[0146] Image modal stream: The goal is to render the Z-score deviation matrix as a heatmap and use the Visual Transformer (ViT) to capture its macroscopic, global spatial distribution patterns (e.g., weakened connectivity in the default mode network or aberrant clusters of connectivity in specific brain regions).

[0147] Numerical modal flow aims to flatten the deviation matrix into a one-dimensional numerical vector and extract its precise, topological numerical combination features (e.g., nonlinear relationships between different connectivity strengths) using models such as LightGBM. This invention aims to completely preserve and extract the two core information elements of pathological patterns through this dual-flow design, solving the problem of incomplete single-modal representation.

[0148] The functional connectivity (FC) matrix inherently contains two core pieces of information: precise numerical topological information and macroscopic spatial distribution information. This invention employs a parallel, dual-stream architecture, addressing the problem that existing technologies either vectorize (losing spatial patterns) or visualize (ignoring precise numerical values), resulting in significant information loss.

[0149] 3. Employ advanced dynamic fusion mechanisms. For example... Figure 2This invention introduces a bidirectional cross-modal attention mechanism, aiming to move beyond static addition between image and numerical modalities. Instead, these modalities act as dynamic contexts for querying, verifying, and enhancing each other. For example, a weak numerical anomaly (from the numerical stream) only has diagnostic significance within a specific global spatial mode (from the image stream). The attention mechanism of this invention aims to capture this context-dependent complementary relationship. Through a bidirectional process of numerical values ​​"focusing" on the image and the image "focusing" on the numerical values, it fuses a final feature vector with higher information density, stronger discriminative power, and less noise. Based on this, it solves the problem that simple feature concatenation in existing technologies cannot capture deep intermodal relationships.

[0150] The dementia identification method based on EEG norm network and dual-stream attention fusion of the present invention has the following advantages.

[0151] 1. By using QRF-based standardized modeling, the discriminative power of pathological features is significantly improved.

[0152] Traditional methods directly use the original functional connectivity (FC) matrix, whose values ​​are greatly affected by physiological factors such as age, resulting in weak pathological signals, making it difficult to distinguish pathological changes from normal physiological fluctuations, especially in early diagnosis. This invention employs the quantile regression forest (QRF) described in S2-S3, establishing a "norm" of age-related changes in functional connectivity based on healthy control group (HC) data, and converting the FC matrix of all subjects into a standardized Z-score deviation matrix. To verify the effectiveness of this method, we designed a comparative experiment, using (1) the original FC feature vector and (2) the QRF Z-score feature vector of this invention, respectively, and feeding them into different baseline classifiers (LightGBM, SVM, Random Forest) for a three-class classification task. Performance was evaluated from four dimensions: accuracy, F1-score, recall / sensitivity, and precision. Figures 3 to 6 As shown.

[0153] Figure 7 The comparison charts strongly demonstrate that the normalization norm construction mechanism in steps 2-3 of this invention is not a simple feature processing step, but a crucial signal normalization step with significant statistical importance. It filters out noise from physiological factors such as age, extracting the mixed and weak pathological signals from the original data into standardized Z-score features with high discriminative power. This results in a qualitative improvement in the model's accuracy, F1 score, recall, and precision, providing a fundamental basis for achieving high performance in subsequent dual-stream models.

[0154] 2. Superiority of Dual-Stream Parallel Architecture and Cross-Modal Attention Fusion (Ablation Experiment). Existing methods either use only a single modality (e.g., image only or numerical only), resulting in information loss; or use simple feature concatenation, failing to achieve dynamic interaction and information enhancement between modalities. This invention adopts the dual-stream architecture in step 4 (ViT extracts image spatial patterns + vectorization extracts numerical topology), and uses the bidirectional cross-modal attention mechanism in step 5 for dynamic fusion. To fully verify the advantages provided by this invention, comparative experiments were designed for the following three types of baseline models:

[0155] (1) Single-modal baseline model (numerical features): This model only sends the numerical feature vectors into the classifier for classification.

[0156] (2) Single-modal baseline model (image features): This model only sends the image feature vectors into the classifier for classification.

[0157] (3) Simple fusion baseline model: The numerical feature vector and the image feature vector are simply concatenated and then fed into the classifier.

[0158] Table 1: Performance Comparison of Each Model in the Three-Class Classification Task

[0159]

[0160] As shown in Table 1 above, compared with traditional single-modal methods (models A and B) and simple multimodal fusion methods (model C), the present invention (Norm-CrossNet) has achieved significant improvements in all performance metrics through innovative bimodal feature representation and cross-modal attention fusion mechanism.

[0161] To further verify the contribution of each key technical module (QRF normalized modeling, two-stream architecture, and bidirectional cross-modal attention) in this invention, an ablation experiment was also designed in this embodiment, and the results are shown in Table 2. The experimental results show that the absence of any one module will lead to a decrease in classification performance, fully demonstrating the synergistic enhancement effect of the technical solution of this invention.

[0162] Table 2: Ablation Experiment Results on the Impact of Each Key Technical Module of the Invention on Recognition Performance

[0163]

[0164] As shown in Table 2 above, the effectiveness of each core module of this invention has been verified through fine-grained ablation experiments:

[0165] (1) Necessity of QRF normalization modeling: Comparing Experiment 1 (original FC baseline, 53.80%) and Experiment 3 (QRF numerical features, 62.15%), it can be seen that the accuracy improved by about 9 percentage points after introducing QRF to construct the norm deviation. This proves that filtering out individual age differences and converting the original connection strength into a normalized Z-score are the prerequisites for extracting effective pathological signals.

[0166] (2) Complementarity of dual-stream features: The single-modal performance of Experiment 2 (image only, 71.46%) and Experiment 3 (numerical only, 62.15%) was limited, while Experiment 4 (simple stitching, 80.42%) combined the two and the accuracy was significantly improved. This shows that there is a significant complementary relationship between spatial distribution pattern (image) and topological connectivity strength (numerical), and the dual-stream architecture effectively makes up for the shortcomings of single viewpoint.

[0167] (3) The key role of cross-modal attention mechanism: Comparing Experiment 4 (simple splicing, 80.42%) and Experiment 7 (bidirectional attention of this invention, 93.57%), the performance achieved a qualitative leap after introducing the bidirectional cross-modal attention mechanism, improving by 13.15 percentage points. This fully demonstrates that only through the bidirectional dynamic interaction of numerical "attention" to the image and the image "attention" to the numerical value can the deep correlation between modalities be maximized, which is the core technical support for achieving accurate identification of dementia.

[0168] 3. High recognition accuracy and early diagnostic sensitivity verified by experiments; the complete technical solution (S1-S5) proposed in this invention has demonstrated excellent performance in the tri-classification task of distinguishing between Alzheimer's disease (AD), mild cognitive impairment (MCI) and healthy controls (HC).

[0169] To objectively evaluate the model's generalization ability and eliminate interference from individual-specific characteristics, this embodiment abandons the traditional random shuffling and instead employs a rigorous subject-independent 5-fold cross-validation strategy. Specifically, the 193 subjects were randomly divided into five mutually exclusive subsets based on their individual IDs, strictly ensuring that all data from the same subject exists only in either the training or test set, thus preventing data leakage. The final performance metric is the average of the five validation results.

[0170] Under the aforementioned strict experimental conditions, the experimental results (see...) Figure 7(Confusion matrix diagram), the overall recognition accuracy of this model is as high as 93.57%. The model is robust; for example, the recognition recall rate for the healthy control group (HC) is 98.6% (73 / 74), and the recognition recall rate for MCI is also as high as 95.2% (59 / 62). This demonstrates that this method has high sensitivity in capturing early and subtle pathological features.

[0171] like Figure 8 ROC curve analysis further confirmed the model's superior discriminative ability. When distinguishing between AD, HC, and MCI classes, the AUC (area under the curve) values ​​reached 0.941, 0.987, and 0.979, respectively. The AUC values ​​for all classes were very close to 1, indicating that the model maintains high robustness across various thresholds.

[0172] like Figure 9 The final fusion feature vector F extracted by this invention is obtained through the UMAP algorithm. fused Nonlinear dimensionality reduction visualization was performed. It was then visually observed that the three groups of subjects (HC, MCI, and AD) exhibited clearly defined clustered distributions in the feature space. The healthy control group (HC) formed highly independent, tightly packed clusters; while the pathological groups (AD and MCI), although biologically continuous, still showed clear class boundaries within the feature space of this invention. This result, from the perspective of high-dimensional manifold representation, validates the superior performance of the two-stream attention fusion mechanism in integrating cross-modal complementary features.

[0173] 4. It possesses clinical interpretability and application value.

[0174] This invention not only boasts high diagnostic accuracy, but the Z-score deviation matrix generated in step 3 is itself a highly interpretable biomarker. By visualizing this Z-score deviation matrix (e.g., Figures 10-12 As shown in the figure, clinicians can visually see which specific brain functional connections (such as frontal-parietal connections) in a particular patient (such as MCI patients) show abnormal enhancement (Z>0, usually indicated in red) or weakening (Z<0, usually indicated in blue) compared to the age-matched healthy norm.

[0175] This provides clinical practice with individualized and standardized pathological feature maps, which effectively solves the problems that traditional "black box" models cannot explain, and provides an objective basis for personalized treatment and efficacy evaluation. It has certain clinical application value and translational potential.

[0176] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0177] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A data analysis method based on EEG norm network and dual-stream attention fusion, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing; Acquire resting-state EEG data of the subjects and preprocess the resting-state EEG data; Step 2: Modeling Steps; Construct a standardized standard EEG functional network model that varies with age, i.e., a norm network; The main covariate of the standardized standard EEG functional network model is age, and the response variable is the brain functional connectivity strength value between specific channel pairs; The brain functional connectivity matrix contains two different types of features: one is numerical topological features, i.e., the precise connectivity strength value of each element in the matrix; the other is spatial distribution features, i.e., the specific patterns of strong and weak connections formed throughout the brain; Step 3: Quantification of individual brain function deviation; Calculate the deviation z of each brain functional connection of the subject relative to the corresponding functional connection in the healthy norm network matrix. i ; The degree of deviation z in connecting all brain functions i Reorganized into whole-brain deviation matrix Z mat This serves as the input data source for subsequent feature extraction in a dual-stream architecture. Step 4: Multimodal feature extraction step; using a parallel two-stream architecture, image feature vectors I representing the spatial distribution pattern are extracted from the deviation matrix z. e and numerical eigenvectors V representing numerical topological structures e The dual-stream architecture includes an image modal stream and a numerical modal stream. Step 5: Cross-modal attention fusion and classification diagnosis; based on the numerical feature vector V e and image feature vector I e Image features ViT were obtained respectively feat Sum of numerical features z feat A bidirectional cross-modal attention mechanism is used to apply ViT to image features. feat Sum of numerical features z feat Perform dynamic fusion, fusing image features ViT feat Sum of numerical features z feat Obtain the final feature vector F fused The final feature vector F fused The data is fed into a downstream classifier for final classification and judgment, and the final prediction result is output.

2. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, The subjects in step 1 include patients and healthy control groups.

3. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, The preprocessing in step 1 includes bandpass filtering, removal of physiological artifacts, and signal segmentation.

4. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, In step 2, the time segments obtained after preprocessing are processed to obtain the brain functional connectivity matrix FC.

5. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, In step 2, the quantile regression forest (QRF) algorithm is used to construct a standardized standard EEG functional connectivity network model based on age.

6. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 5, characterized in that, The QRF model of the QRF algorithm is implemented using the LightGBM framework.

7. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, In step 3, the deviation z of each brain functional connection of the subject relative to the healthy norm network is calculated using a standardized formula. i .

8. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, In step 4, the image modal stream renders the deviation matrix z into a two-dimensional heatmap image; the numerical modal stream flattens out the upper triangular part of the deviation matrix z to form a one-dimensional numerical vector.

9. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 8, characterized in that, The image feature vector I is obtained using the ViT model in the image modal stream. e The numerical modal flow employs a gradient boosting decision tree model to extract numerical feature vectors V. e .

10. The data analysis method based on EEG norm network and dual-stream attention fusion according to claim 1, characterized in that, In step 5, the image feature vector I e Sum of numerical eigenvectors V e Image features ViT are obtained by mapping their respective encodings to the same high-dimensional semantic space. feat Sum of numerical features z feat .

Images