An ad diagnostic method fusing eeg frequency spectrum normalization deviation and brain network

Abstract

The application discloses an AD diagnosis method fusing brain electric spectrum standardization deviation and brain network, comprising the following steps: S1, carrying out band-pass filtering and segmenting on the multi-channel EEG signals of five brain areas; S2, calculating the brain area-band characteristics of the five different brain areas for each segment; S3, adopting a GAMLSS model to perform standardization modeling, identifying individual deviation information of the disease group about the rPSD of a specific brain area, training a LightGBM classifier by using the individual deviation information, and extracting a structured feature vector by using the internal structure of the classifier; S4, dividing each segment into a plurality of non-overlapping frequency bins according to the frequency resolution, calculating the FCN graph of each frequency bin in the form of a matrix, constructing a three-dimensional image according to the FCN graph, and extracting a visual feature vector by using a Swin transformer network; and S5, constructing a connection vector and inputting the connection vector into the classifier for classification. The method provides a new technical approach for the precise diagnosis and differential diagnosis of neurodegenerative diseases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of modal recognition, and particularly relates to an AD (Alzheimer's disease) diagnosis method. BACKGROUND

[0002] Alzheimer's disease (AD) and frontotemporal dementia (FTD) are highly prevalent neurodegenerative diseases, accounting for 60-80% and 10% of all dementia cases, respectively. The complex clinical and pathological manifestations of these two diseases pose major challenges for diagnosis and treatment.

[0003] Alzheimer's disease is the most common type of dementia, typically characterized by cognitive decline, memory loss, and behavioral changes. Frontotemporal dementia, although less common, tends to affect younger populations and is primarily characterized by personality changes and language impairment. There is currently no cure or treatment that can stop the progression of AD or FTD, which has prompted increasing research interest in age-related brain changes associated with pathological aging.

[0004] Traditional methods of diagnosing AD and FTD rely on clinical assessments, neuroimaging examinations (such as MRI, CT, PET, etc.), and cognitive tests. However, these methods are often costly and time-consuming. Electroencephalography (EEG) has emerged as a promising tool for early diagnosis and monitoring of abnormal brain activity due to its high temporal resolution, non-invasiveness, and low cost.

[0005] In recent years, EEG has become a widely used modality for investigating and clinically supporting neurodegenerative diseases such as AD, FTD, mild cognitive impairment (MCI), and Parkinson's disease (PD). An increasing number of studies have demonstrated that EEG signals can provide valuable biomarkers that can distinguish these diseases from healthy aging. Traditional EEG analysis methods, including spectral power analysis, coherence, phase synchronization, and event-related potentials (ERPs), have been widely used to characterize changes in brain rhythms associated with disease states.

[0006] However, by averaging over population data, population analysis tends to mask individual differences, which can mask important differences in disease progression and individual-level treatment response. From the perspective of precision medicine, this "one-size-fits-all" approach does not take into account unique genetic, environmental, and lifestyle factors that influence each patient's condition.

[0007] Normative modeling has emerged as a powerful approach to enable subject-specific inference by characterizing population-level norms and quantifying individual-level deviations from these norms. While normative modeling is well-established in other fields, such as pediatric medicine using height or weight growth charts, it remains relatively novel in neuroimaging research. In existing research, normative modeling has been used to predict brain age, thereby identifying potential issues in brain development and aging processes.

[0008] Recent MRI-based studies have employed normative modeling to chart age-related trajectories of brain phenotypes throughout the human lifespan. These efforts have not only served to characterize normative age-related brain structural changes but also to describe structural and functional heterogeneity in psychiatric disorders (e.g., schizophrenia and bipolar disorder) and neurodegenerative diseases (e.g., AD).

[0009] Deep learning is popular due to its ability to automatically learn features from raw EEG data. Although there have been a large number of deep learning methods applied to EEG-based disease diagnosis, few methods currently utilize multi-modal models. In the field of neuroscience, many studies have adopted a multi-modal framework. Although the multi-modal framework has achieved significant success in brain disease applications, multi-modal methods for EEG data remain underexplored due to limited sample size and insufficient fusion strategies. SUMMARY

[0010] In view of the technical problems in the prior art that population analysis masks individual differences, multi-modal fusion strategies are insufficient, and diagnostic accuracy needs to be improved, the present application aims to provide an AD diagnosis method fusing normative deviations of brain electrical spectrum and brain network, so as to realize individualized neurophysiological evaluation, effectively fuse heterogeneous data modalities, and significantly improve diagnostic accuracy.

[0011] To this end, the application provides an AD diagnosis method combining electroencephalogram frequency spectrum normalization deviation and brain network, comprising the following steps: S1, band-pass filtering the multi-channel EEG signals of five brain regions, and segmenting the continuous EEG signals into non-overlapping segments; S2, for each segment in step S1, calculating the relative power spectrum rPSD of all channels in five typical frequency bands, and averaging all channels in each brain region to obtain the brain region-frequency band feature; S3, using the GAMLSS model for normalization modeling, identifying the individual deviation information of the disease group about the rPSD of a specific brain region, and using the individual deviation information to train the LightGBM classifier, and using the internal structure of the classifier to extract a structured feature vector, that is, using the classifier to extract the index of each decision tree reaching the leaf node, and then connecting these leaf node indexes into a feature vector; S4, dividing each segment in step S1 into several non-overlapping frequency bins according to the set frequency resolution, calculating the matrix form of the brain functional connection network FCN of each frequency bin, stacking the FCN of all frequency bins along the frequency dimension to construct a three-dimensional image, and using the Swin transformer network to extract a visual feature vector; S5, connecting the structured feature vector extracted from the LightGBM classifier and the visual feature vector extracted from the Swin transformer network, and inputting them into the classifier for classification to obtain the three-class discrimination between AD, FTD and HC.

[0012] The application proposes a multi-modal integrated normalization diagnosis network (MIND-NET) framework, which combines normalization modeling and multi-modal feature extraction, and is used for neural degenerative disease subtype classification, can not only realize individualized neurophysiological evaluation, but also significantly improve the diagnosis accuracy, and provides a new technical approach for precise diagnosis and differential diagnosis of neural degenerative diseases.

[0013] In addition to the purposes, features and advantages described above, the application has other purposes, features and advantages. The application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0014] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application and their description serve to explain the application, and do not constitute improper limitations on the application. In the drawings:

[0015] Figure 1 is a flowchart of the AD diagnosis method combining electroencephalogram frequency spectrum normalization deviation and brain network of the application;

[0016] Figure 2 shows the distribution of extreme deviations of AD and FTD in five brain regions;

[0017] Figure 3 Differences in bias values for AD, FTD, and HC groups are shown;

[0018] Figure 4 AD and FTD group positive bias and side bias graphs are shown;

[0019] Figure 5 The classification performance of the original features without applying normalized modeling and the features applying normalized modeling in four classical machine learning algorithms is shown. DETAILED DESCRIPTION

[0020] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0021] In view of the technical problems in the prior art that population analysis masks individual differences, multi-modal fusion strategy is insufficient, and diagnostic accuracy needs to be improved, the present application proposes a multi-modal integrated normalized diagnostic network (MIND-NET) framework, which is a novel framework combining normalized modeling and multi-modal feature extraction for neurodegenerative disease subtype classification.

[0022] The method of the present application includes two main stages: first, the present application uses a generalized additive location, scale, and shape model (GAMLSS) to construct normalized trajectories of EEG spectral features based on HC (healthy control group) data. Second, the present application applies individual-level bias derived from this normalized model to assess heterogeneity within AD and FTD cohorts. The MIND-NET framework extracts features from EEG-derived functional connectivity network (FCN) images using a Swin transformer and from structured bias data using LightGBM (LGB), thereby creating a comprehensive multi-modal representation that enhances the classification accuracy of neurodegenerative disease subtypes.

[0023] The overall workflow of the present application is shown in Figure 1 The complete architecture of the MIND-NET model is shown. Starting from EEG preprocessing, (a) (b) show normalized modeling across brain regions and structured data feature extraction using LGB, and (c) is the image feature extraction stage based on a Swin transformer, which uses a conventional Swin transformer to produce a feature vector. Finally, the connected features are input into a classifier to achieve three-class discrimination between AD, FTD, and HC.

[0024] Specifically, the present AD diagnostic method includes the following steps S1 to S5.

[0025] S1, obtaining an EEG dataset

[0026] The present application uses the publicly available EEG dataset OpenNeuro ds004504. In order to conduct more reasonable standardized modeling, the present application uses the data of subjects with age range (61-78 years old) in the dataset, including 25 AD patients, 17 FTD patients and 28 HC individuals. The electrode placement follows the internationally recognized 10-20 system, and 19 channels are located at Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1 and O2. Electrodes A1 and A2 are designated as reference electrodes.

[0027] S2, data preprocessing

[0028] The EEG channel signals are subjected to 0.5-45Hz band-pass filtering to remove low-frequency drift and high-frequency electromyographic artifacts, and are uniformly resampled from 500Hz to 200Hz. The continuous EEG signals are segmented into non-overlapping 4-second epochs, and each segment is subjected to zero-mean unit-variance standardization within the channel to eliminate the influence of overall power differences between channels.

[0029] S3, structured data feature extraction

[0030] For each segment, the present application calculates the power spectral density (PSD) from 1 to 45Hz using the Welch method, and calculates the relative power spectrum (rPSD) of all channels in five typical frequency bands: δ (1-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz) and γ (30-45Hz). Subsequently, the present application arranges the EEG channels according to five different brain regions: frontal lobe (Fp1, Fp2, F7, F3, Fz, F4, F8), central (C3, Cz, C4), temporal lobe (T3, T4, T5, T6), parietal lobe (P3, P4, Pz) and occipital lobe (O1, O2). Then the present application averages all channels in each brain region to obtain the final brain region-band features.

[0031] rPSD eliminates the influence of overall EEG power differences between individuals and electrode impedance changes by normalization, enhances the comparability between different subjects and time periods, and reduces the influence of experimental environment and individual physiological state fluctuations. This helps the stability of the subsequent standardized modeling of the present application.

[0032] In order to construct the reference trajectory of EEG signals of healthy individuals at different ages and identify the deviation characteristics of rPSD of specific brain regions of the disease group, the present application adopts GAMLSS for standardized modeling. This method can comprehensively model the nonlinearity, heteroscedasticity, skewness and kurtosis changes in neurophysiological data, making it suitable for population modeling and individual anomaly detection of complex EEG biomarkers. Here, the main covariate is age, and the response variable is the rPSD value of a specific brain region in a specific frequency band. This indicator reflects the intensity of neural activity of the brain region in a specific frequency range and has clear neurophysiological significance.

[0033] The present application uses the gamlss package in R language to fit the GAMLSS model based on HC data. Through the fitting comparison of the candidate distribution family, the present application finally selects the Box-Cox-t distribution (BCT) as the response variable distribution of GAMLSS. The mathematical formula of GAMLSS is as follows:

[0034]

[0035]

[0036] Where F represents the BCT distribution, which is parameterized by the location parameter μ, the scale parameter σ > 0, the power shape parameter ν, and the degree of freedom parameter τ > 0.

[0037] Where is an appropriate link function, X μ and Z μ are known fixed effect and random effect design matrices, and the corresponding coefficient vectors are β μ and γ μ , respectively. represents a smooth basis function used to capture the potential nonlinear effect of the covariate Here, the present application writes a predictor for the location parameter μ; similar formulas apply to other distribution parameters, which are modeled through nonlinear smooth terms.

[0038] After fitting the GAMLSS model, the present application manually calculates the 5%, 50%, and 95% percentiles of each age point based on the four distribution parameters returned by the model to represent the reference range of EEG features in the healthy population. Subsequently, the present application uses linear interpolation to smoothly map these percentile estimates to a regular age sequence, thereby constructing a complete standardized trajectory and confidence interval.

[0039] In order to evaluate the degree of deviation of AD and FTD patients from the standardized reference trajectory, the present application compares their corresponding EEG features with the 5%-95% confidence interval of the healthy population of the same age. The present application calculates the percentage of each category falling within and outside the reference interval to quantify the probability of EEG abnormalities for each category.

[0040] The individual bias information is further used to build classification models. Specifically, the present invention uses each subject's bias value (i.e., bias data computed from the age-matched HC-derived normalized values) as an input feature for training multiple machine learning classifiers to distinguish AD and FTD. The best-performing classifier is then used to extract structured tabular format data. This workflow enables the normalized modeling to not only reveal subject-specific neural biases, but also to provide structured inputs for downstream multi-modal models, thus enhancing the integration of individualized neural signatures into a broader neurodegenerative disease analysis computational framework.

[0041] S4, image feature extraction

[0042] After segmenting the raw EEG signal into non-overlapping 4-second windows, considering the impact of segment length on spectral resolution in the Welch method, the present invention uses 4-second segments and a frequency resolution of 0.25 Hz. The frequency resolution Af is inversely proportional to the analysis window time length T. At this resolution, the signal from 0.25 to 25 Hz is divided into 100 non-overlapping frequency bins, covering the critical range from delta to low gamma waves in psychiatric diagnosis. In each frequency band, the present invention computes the coherence function between all pairs of channels to obtain the inter-channel functional connectivity strength. Given the EEG data dimension of C x T (where C is the number of channels and T is the number of time points), each matrix corresponds to a brain functional connectivity network within a frequency bin.

[0043] This procedure computes the in-band coherence according to the following formula:

[0044]

[0045] where P xy (ω) represents the cross-spectral density between x(t) and y(t), while P xx (ω) and P yy (ω) are their respective auto-spectral densities.

[0046] To encode this sequence into a standard image format suitable for input into the Swin transformer, the present invention employs a matrix concatenation method, stacking the 100 frequency bin connectivity maps along the frequency dimension to construct a three-dimensional image. The image is then resampled to a 224 x 224 x 3 dimension to meet the input requirements of the Swin transformer network.

[0047] This representation method not only preserves the brain network structure at different frequencies, but also allows the network to capture cross-frequency functional connectivity patterns from image space. The advantage is to preserve high-frequency resolution microstructure differences rather than roughly dividing frequency bands, thereby enhancing the model's ability to perceive subtle frequency-domain features of neuro-psychiatric states.

[0048] The input image is first processed through the initial L layer of the standard Swin transformer (patch segmentation, linear embedding, and shift window self-attention) to extract multi-scale visual features. Finally, a 1×768 feature vector is obtained in the feature extraction layer, which is fused in parallel with subsequent text features.

[0049] This invention uses rPSD bias based on normalized modeling as input to train an LGB classifier, with the goal of predicting the source label of each sample from the bias features. After training, this invention utilizes the internal structure of the model for feature extraction.

[0050] For any given input sample, this invention does not retrieve its final classification probability, but instead records the indices of the leaf nodes reached in each decision tree. The LGB ensemble of this invention has 512 trees, and these leaf node indices are concatenated into a 1×512 feature vector. This vector is a high-dimensional nonlinear transformation of the original bias scores, capturing deep structural information in the data.

[0051] S5, Feature Fusion Prediction

[0052] To perform the final classification task, this invention fuses information from different modalities. The 1×512-dimensional feature vector extracted by LGB is concatenated with the 1×768-dimensional visual feature vector obtained from the visual model, producing a 1×1280-dimensional joint feature representation. This concatenated vector is then input into the downstream classifier to generate the final prediction.

[0053] Effect Comparison

[0054] After generating normalized trajectories, this invention maps data from AD and FTD subjects onto a model and calculates deviation scores based on the 5th to 95th percentile confidence intervals defined for healthy controls. In the frontal cortex, the deviation rate in the β band was 14.99% for AD and 17.59% for FTD, rising to 18.05% for AD and 25.42% for FTD in the γ band. In the temporal lobe, AD showed a deviation of 19.41% in the β band, while FTD peaked at 20.59% in the α band. The occipital lobe showed the most significant deviation, with an α band AD of 23.16% and an FTD of 34.46%. The parietal lobe showed an α band deviation of 15.85% for AD and 19.65% for FTD, while the central cortex remained relatively moderate across all bands. Overall, extreme deviations in AD and FTD were concentrated in the α, β, and γ bands of the temporal, occipital, and parietal lobes, and FTD also showed significant abnormalities in the β and γ bands of the frontal cortex, such as... Figure 2 .

[0055] The present invention then computes the average deviation of rPSD for AD and FTD patients after normalizing to the GAMLSS trajectory derived for HC. Both AD and FTD cohorts show significantly elevated deviation scores compared to the HC group, indicating significant heterogeneity in their EEG signatures. These group differences are illustrated in Figure 3 Figure 2, where color-coded markers represent individual subjects from AD, FTD, and HC groups, respectively, vividly highlighting the difference patterns in rPSD deviation across cohorts. This figure indicates that the deviation values for AD and FTD groups are significantly higher than the HC group.

[0056] To identify shared deviation patterns and explore heterogeneity between AD and FTD groups, the present invention computes the average power of positive and negative deviations for each group and plots the normalized (0-1) overlap topography.

[0057] Figure 4 Subplots (a) and (b) present the positive deviation maps for AD and FTD groups, respectively. In both groups, the beta-band activity in the occipital region shows the most pronounced differences, with additional substantial divergences observed in the theta-band activity in the frontal and parietal regions.

[0058] Figure 4 Subplots (c) and (d) show the negative deviation maps for AD and FTD, respectively, highlighting that the delta and gamma-band activity in the frontal cortex provides the clearest distinction between the two groups.

[0059] These results indicate that the positive deviations distinguishing AD and FTD are primarily characterized by beta-band activity in occipital electrodes and theta-band activity in frontal and parietal electrodes. In contrast, the negative deviations are most prominent in frontal delta and gamma-band activity. This spatial-frequency signature provides an intuitive and informative comparison of shared deviation patterns and group-specific heterogeneity.

[0060] The present invention compares raw features without applying standardized modeling and features with applying standardized modeling, as Figure 5 and evaluates their classification performance using four classical machine learning algorithms: LGB (with default parameters of 512 trees and 63 leaf nodes), multilayer perceptron (MLP), random forest, support vector machine (SVM), and k-nearest neighbors (kNN). Four evaluation metrics are used: accuracy (ACC), F1 score (F1), sensitivity (SENS), and specificity (SPEC).

[0061] It is evident that, in traditional machine learning methods, standardized modeling enhances features yield better performance compared to raw features without applying standardized modeling. In particular, the LGB classifier stands out with an impressive accuracy of 92.05%. This is the main reason for the present invention's choice of LGB for structured data feature extraction.

[0062] The present application uses LGB to perform classification between AD, FTD and HC for each brain region separately. Five-fold cross-validation is adopted to evaluate the performance of the model. As can be seen from Table 1, the experimental results show that the temporal lobe and occipital lobe brain regions achieve classification accuracy of 82.00% and 82.58% respectively, which is significantly better than the other three brain regions. This finding is in strong agreement with the previously identified high heterogeneity characteristics of AD and FTD, thereby confirming the significant advantages and effectiveness of the standardized modeling method proposed by the present application in deep learning model training.

[0063]

[0064] In Table 2, the present application evaluates the influence of different feature extraction schemes on the overall performance of MIND-NET. Model 1 uses Swin transformer architecture Swin-T (tiny) to extract image features, and model 2 uses LGB to process structured data. Specifically, LGB (T:100) uses 100 trees to generate a 1x100 feature vector, and LGB (T:512) uses 512 trees to generate a 1x512 feature vector. The " / " symbol indicates that no structured data features are used in the model. The present application also includes AUC as an additional performance indicator to evaluate classification ability. The experimental results show that the combination of Swin-T and LGB (T:512) achieves the best performance, achieving an accuracy of 97.12%, an AUC of 0.998 and an F1 score of 97.01% under five-fold cross-validation. The use of Swin-T alone produces an accuracy of 93.42%, an AUC of 0.993 and an F1 of 93.03%, while the use of LGB (T:512) alone achieves an accuracy of 91.79%, an AUC of 0.973 and an F1 of 86.97%.

[0065]

[0066] To demonstrate the capabilities of the proposed MIND-NET (Swin-T and LGB (T:512)), in Table 3, the present application not only evaluates the three-class classification performance case 1 (AD vs FTD vs HC), but also evaluates three binary classification tasks: case 2 (AD vs FTD), case 3 (AD vs HC) and case 4 (FTD vs HC). The evaluation indicators include ACC, F1, SEN, SPEC and AUC. To further verify the effectiveness of the proposed model, the present application compares it with several state-of-the-art methods in the benchmark of three-class classification studies on the same dataset.

[0067]

[0068] The experimental results show that the MIND-NET framework of the present application achieves excellent performance in all cases. In the three-class classification task (case 1), the accuracy reaches 97.12% ± 0.31%, the AUC is 0.998 ± 0.002, and the F1 score is 97.01% ± 0.39%. In the binary classification task, case 2 (AD vs FTD) reaches an accuracy of 97.79% ± 0.25%, case 3 (AD vs HC) reaches an accuracy of 98.11% ± 0.34%, and case 4 (FTD vs HC) reaches an accuracy of 98.25% ± 0.24%, all of which show excellent performance.

[0069] The comparison with other state-of-the-art methods shows that the MIND-NET framework of the present application is significantly superior to the prior art, as shown in Table 4. The EEGConvNeXt method of Acharya et al. achieves an accuracy of 95.70% in three-class classification, while the method of the present application achieves 97.12%, an improvement of 1.42%. The custom CNN combined with visual transformer method of Chen et al. only achieves an accuracy of 79.12%, while the method of the present application has been significantly improved. The LightGBM method combined with wavelet transform of Hachamnia et al. achieves an accuracy of 93.21% in three-class classification, which is still lower than the 97.12% of the present application. These comparison results clearly demonstrate the technical advantages of the MIND-NET framework proposed by the present application.

[0070]

[0071] The above description is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A data analysis method integrating EEG spectrum normalization bias and brain network, characterized in that, Includes the following steps: S1. Bandpass filtering is performed on the multi-channel EEG signals from five brain regions to segment the continuous EEG signals into non-overlapping segments. S2. For each segment in step S1, calculate the relative power spectrum rPSD of all channels in five typical frequency bands, and average all channels in each brain region to obtain brain region-frequency band characteristics. S3. The GAMLSS model is used for normalized modeling, and individual bias information about rPSD in specific brain regions of the disease group is identified. This individual bias information is used to train the LightGBM classifier, and the internal structure of the classifier is used to extract structured feature vectors. S4. Divide each segment in step S1 into several non-overlapping frequency boxes according to the set frequency resolution, calculate the matrix form of the functional connectivity network (FCN) of each frequency box, stack the FCNs of all frequency boxes along the frequency dimension to construct a three-dimensional image, and use the Swing transformer network to extract visual feature vectors. S5. The structured feature vector extracted from the LightGBM classifier and the visual feature vector extracted from the Swin transformer network are concatenated and input into the classifier for classification to obtain the three-class discrimination between AD, FTD and HC.

2. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, The multichannel EEG signals from five brain regions include 19 channels: frontal lobe (Fp1, Fp2, F7, F3, Fz, F4, F8), central lobe (C3, Cz, C4), temporal lobe (T3, T4, T5, T6), parietal lobe (P3, P4, Pz), and occipital lobe (O1, O2).

3. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, In step S1, the EEG signal is bandpass filtered from 0.5 to 45 Hz and then uniformly resampled from 500 Hz to 200 Hz. Each segment after segmentation is standardized with zero mean and unit variance within the channel.

4. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, In step S2, five typical frequency bands are used: δ (1-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz), and γ (30-45Hz).

5. The data analysis method for integrating EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, In the GAMLSS model, the covariate is age, and the response variable is the rPSD value of a specific brain region and frequency band. The Box-Cox-t distribution is chosen as the distribution of the response variable in GAMLSS.

6. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 5, characterized in that, Step S3 also includes: S31. Calculate the 5%, 50%, and 95th percentiles for each age point based on the distribution parameters returned by the GAMLSS model to represent the reference range of EEG characteristics in healthy individuals. Then, use linear interpolation to smoothly map these percentile estimates to regular age sequences. S32. Compare the EEG characteristics of patients with AD and FTD to be evaluated with the 5%-95% confidence interval of healthy individuals of the same age, calculate the percentage of each category falling within and outside the reference interval, to quantify the probability of EEG abnormality in each category, and thus assess the degree to which AD and FTD patients deviate from the standardized reference trajectory.

7. The data analysis method for fusing EEG spectrum normalization bias and brain network as described in claim 1, characterized in that, Individual bias information regarding rPSD in specific brain regions of disease groups is identified as bias data calculated from normalized values ​​derived from age-matched HC.

8. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, In step S4, the frequency resolution is set to 0.25 Hz. At this resolution, the signal from 0.25 to 25 Hz is divided into 100 non-overlapping frequency boxes, covering the key range from delta waves to low gamma waves in the diagnosis of mental illness.

9. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, In step S4, the 3D image is first resampled to 224×224×3 dimensions, and then input into the Swintransformer model, outputting a 1×768 visual feature vector.

10. The data analysis method for fusing EEG spectrum normalization deviation and brain network as described in claim 1, characterized in that, The 1×512-dimensional feature vector extracted from the LightGBM classifier is concatenated with the 1×768-dimensional visual feature vector obtained from the Swin transformer model to produce a 1×1280-dimensional concatenation vector, which is used as input to the downstream classifier to produce a predicted classification.

Images