Cognitive feature extraction and classification method and system based on electroencephalogram signals

Abstract

The invention discloses a cognitive feature extraction and classification method and system based on electroencephalogram signals, and the method comprises the steps: constructing a cognitive divergence mapping network, so as to integrate a multi-branch brain region topology module and a neurodynamics physical information network module; the multi-branch brain region topology module divides detection branches of five brain regions (frontal lobe, central lobe, parietal lobe, occipital lobe and temporal lobe) according to 10-20 systems, and through feature extraction of different brain region branches, the feature learning ability of the Alzheimer's disease patient under the condition of cross-brain region signal heterogeneity is remarkably improved; the neurodynamics physical information network module can effectively extract changes of low-frequency and high-frequency components in electroencephalogram signals of the Alzheimer's disease patient by introducing frequency band separation, Fourier transform power spectrum constraint, an adaptive weighting mechanism and neural representation embedding in a frequency spectrum congruence space. The accuracy and generalization of electroencephalogram signal analysis are remarkably improved, and an efficient and non-invasive detection tool is provided for early diagnosis of the Alzheimer's disease.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of medical artificial intelligence and neuroelectrophysiology, and particularly relates to a method and system for extracting and classifying cognitive features based on electroencephalogram (EEG) signals. Background Art

[0002] As a typical representative with the most global disease burden in the spectrum of neurodegenerative diseases, the core pathological mechanism of Alzheimer's disease is manifested as the abnormal deposition of β - amyloid plaques in the extracellular space and the formation of neurofibrillary tangles of microtubule - associated protein Tau in the neurons. These two factors trigger the loss of synaptic plasticity and programmed neuronal death through a cascade amplification effect. Clinically, electroencephalogram (EEG) is commonly used to diagnose Alzheimer's disease. However, due to the non - specificity of early - stage neurophysiological changes, low signal - to - noise ratio, and signal complexity, EEG still faces many challenges in the early diagnosis of Alzheimer's disease.

[0003] In the prior art, the diagnostic methods for Alzheimer's disease based on EEG signals generally adopt a single - branch feature extraction architecture, such as EEGNet, ShallowConvNet, etc. Although these methods show basic diagnostic efficacy in typical cases, they still lack sufficient resolution when facing the EEG signals of complex Alzheimer's disease patients. In addition, traditional time - domain analysis methods are difficult to effectively capture the antagonistic dynamic association between the enhanced low - frequency compensatory power spectrum and the attenuated high - frequency synchrony oscillation caused by neuronal synaptic conduction disorders in the EEG signals of Alzheimer's patients, thus restricting the sensitivity and specificity of early diagnosis. Summary of the Invention

[0004] To solve the deficiencies of the prior art and achieve the purpose of improving the feature learning ability in the case of cross - brain - region signal heterogeneity and enhancing the parsing efficiency of complex EEG signals, the present invention adopts the following technical solutions:

[0005] A cognitive feature extraction and classification system based on EEG signals, including a cognitive divergence mapping network, and the cognitive divergence mapping network includes a multi - branch brain - region topology module and a neurodynamics physical information network;

[0006] The multi - branch brain - region topology module divides the EEG signal source space into multiple isomorphic modeling units according to the characteristics of cerebral cortex functional division, captures regional - specific frequency - domain features respectively, forms the fusion features of each brain region through the combination of cross - channel feature information, and then performs cross - region feature aggregation on the features of each brain region. Compared with the single - branch architecture that is prone to cause coupling interference of cross - brain - region information, feature extraction on different brain - region branches through the multi - branch brain - region topology can improve the feature learning ability in the case of cross - brain - region signal heterogeneity;

[0007] The neural dynamics physical information network module maps EEG signals to the frequency domain and calculates the power spectra of each frequency band by constructing frequency-domain separation modeling. It introduces a power spectral density constraint supervision mechanism and implements an adaptive spectral weight optimization strategy to perform interpretable decoupling and quantitative characterization of abnormal patterns of cross-frequency neural oscillations related to EEG signal data in the spectral isomorphism space. By introducing frequency band separation, Fourier transform power spectrum constraint, adaptive weighting mechanism, and neural representation embedding in the spectral congruence space, the changes in low-frequency and high-frequency components in EEG signals are extracted.

[0008] Furthermore, in the multi-branch brain region topology module, the EEG signal is abstracted as a superposition of multiple oscillatory components, represented as a non-stationary random process:

[0009]

[0010] where A j represents the time-varying amplitude of the j-th oscillatory component, ω j represents the corresponding angular frequency, which is used to reflect the rhythmic characteristics of a specific frequency band in the EEG signal, t represents time, φ j represents the phase shift, which is used to reflect the phase relationship between different oscillatory components, and ε i (t) represents the noise term, including random interference generated by the experimental environment and physiology;

[0011] To capture the functional characteristics of different brain regions, the acquired EEG signal data is divided into subsets according to neuroanatomical functional regions:

[0012]

[0013] where X represents the input EEG signal data, and C r represents the set of EEG signal channel indices corresponding to the r-th brain region.

[0014] Furthermore, in the EEG signal data N represents the number of EEG signal channels, T represents the length of the time series, and r ∈ {1, 2, 3, 4, 5 +, respectively represent the sets of EEG signal channel indices corresponding to the frontal, central, parietal, occipital, and temporal lobe brain regions.

[0015] Furthermore, in the multi-branch brain region topology module, first, standard convolution is performed on the original EEG signal data of each brain region to capture the local features of the EEG signal data of each brain region:

[0016]

[0017] where, denotes the convolutional kernel parameters, σ(·) denotes the ReLU activation function. This layer simulates the synaptic integration mechanism of local neuron clusters, and through the convolutional kernel weights learn the oscillation patterns of specific frequency bands (such as δ: 1 - 4Hz, θ: 4 - 8Hz). K represents the hyperparameter describing the size of the convolutional kernel, representing the delay / stride number in time or sequence. τ represents the index of time delay, representing the time step at each sliding during the convolution process;

[0018] Subsequently, the features from each brain region undergo depth convolution to extract the region - specific spectral features of each brain region:

[0019]

[0020] Among them, represents the intermediate feature representation of the j - th channel in the r - th brain region at time t, represents the input feature of the r - th brain region and the c - th input channel after time delay, represents the convolutional kernel weight from the c - th input channel to the j - th channel during the depth convolution process. Using depth convolution can reduce the number of parameters, effectively capture local time - frequency domain features, and embody the idea of independent decoupling of channels in each brain region;

[0021] Then, through point - wise convolution (1×1 convolution), effective combination of cross - channel feature information is achieved to form the fused features of each brain region:

[0022]

[0023] Among them, represents the fused feature representation of the j - th channel in the r - th brain region, represents the weight of point - wise convolution, which is used for cross - channel feature fusion. This step can effectively extract the correlation between channels and further enhance the local feature expression ability.

[0024] Furthermore, in the multi - branch brain region topology module, the features of each brain region are further passed through a dense connection module to enhance cross - layer feature fusion:

[0025]

[0026] Among them, represents the feature of the r - th brain region densely connected to the i - th layer, represented after time delay, is the dense connection convolutional kernel weight, representing the convolutional kernel weight of the cross - layer connection between the i - th layer and the l + 1 - th layer. BN(·) represents batch normalization, and L represents the number of layers of dense connection.

[0027] Furthermore, in the multi-branch brain region topology module, the final global average pooling feature of each brain region is expressed as:

[0028]

[0029] where represents the feature vector of the r-th brain region at time point t output from the last layer of the densely connected module, represents the global average pooling operation (Global Average Pooling, GAP) on the feature sequence, compressing the temporal feature vector into a single static feature to enhance the feature expression of the overall region;

[0030] Finally, the features of each brain region are subjected to cross-region feature aggregation to integrate the features independently decoupled and extracted from different brain regions, so as to improve the feature learning ability in the case of cross-brain region signal heterogeneity:

[0031] F (global) = concat(F (r) )

[0032] where concat(·) represents the feature concatenation operation, connecting the feature vectors of different brain regions into a long vector to form a cross-brain region global feature representation.

[0033] Furthermore, in the neurodynamics physical information network module, first, the input electroencephalogram signal x(t) is mapped to the frequency domain through the fast Fourier transform FFT and decomposed into multiple frequency bands of δ (1 - 4 Hz), θ (4 - 8 Hz), α (8 - 13 Hz), β (13 - 30 Hz), γ (30 - 45 Hz), and the power spectrum of each frequency band is calculated: The time-domain signal x(t) is converted to the frequency-domain representation X(f) through FFT:

[0034]

[0035] where P band represents the average power of a specific frequency band, which is used to quantify the energy distribution characteristics of the AD-related pathological frequency band;

[0036] Next, the feature M mid in the middle layer of the network is mapped to the frequency domain through FFT, and its power spectrum P band,mid is calculated, and frequency-domain alignment is performed with the power spectrum P band,orig of the original signal, and the difference between the two is constrained by the relative mean square error MSE j :

[0037]

[0038] Among them, B represents the number of frequency bands, ∈ represents a fractional value. To avoid a zero denominator, the absolute MSE is converted to a relative error to eliminate the difference in energy levels between different frequency bands. This constraint forces the intermediate features of the network to retain the pathological frequency band energy distribution pattern of the original signal.

[0039] Furthermore, in the neurodynamics physical information network module, the optimization weights are dynamically allocated according to the relative MSE of each frequency band, and the weights are restricted to the interval [0,1] to ensure fair competition among the weights of all frequency bands:

[0040]

[0041] Among them, the weight w j is inversely proportional to the MSE j error, forcing the network to preferentially optimize the non-linear antagonistic relationship between low frequencies (Delta / Theta) and high frequencies (Alpha / Beta / Gamma), thereby accurately modeling the "low-frequency compensation - high-frequency suppression" phenomenon caused by the disruption of neuronal synchrony in AD patients.

[0042] Furthermore, in the neurodynamics physical information network module, a frequency-domain physical information neural network (ND-PINN) constraint is constructed:

[0043]

[0044] Among them, R represents the number of brain regions. The prediction error is normalized by the original power P of the frequency band band,orig to eliminate the energy difference between frequency bands, and the optimization of difficult-to-fit frequency bands is dynamically strengthened through adaptive inverse proportional weighting to ensure that the model captures the antagonistic relationship across frequency bands.

[0045] A method for cognitive feature extraction and classification based on electroencephalogram (EEG) signals includes a cognitive divergence mapping network, and the cognitive divergence mapping network includes a multi-branch brain region topology module and a neurodynamics physical information network;

[0046] Step S1: Organize and partition the EEG signal data, perform fixed validation and four-fold cross-validation respectively, and use data augmentation technology on the EEG signal data;

[0047] Step S2: Construct a cognitive divergence mapping network according to the cognitive feature extraction and classification system based on EEG signals;

[0048] Step S3: Train the cognitive divergence mapping network using the EEG signal data and obtain the weights;

[0049] Step S4: Evaluate the performance of the trained cognitive divergence mapping network weights on the test set, and adjust the model parameters and architecture according to metrics including but not limited to accuracy (Accuracy, Acc), precision (Precision, Pr), F1-score (F-measure, F1), area under the receiver operating characteristic curve (Area Under the Curve, AUC), geometric mean (Geometric Mean, G-Mean), sensitivity (Sensitivity, Se), and specificity (Specificity, Sp) to optimize the model performance.

[0050] The advantages and beneficial effects of the present invention are as follows:

[0051] The cognitive feature extraction and classification method and system based on electroencephalogram signals of the present invention effectively break through the inherent limitations between the coordination of local neural oscillation specificity and distributed functional network coupling (such as prefrontal-parietal compensatory connection reconstruction) in the traditional single-branch feature extraction framework through a multi-branch brain region topology module, and significantly improve the feature learning ability in the case of cross-brain region signal heterogeneity. In addition, the neurodynamics physical information network module effectively decouples the abnormal patterns of cross-frequency neural oscillations in electroencephalogram information of Alzheimer's disease, thereby significantly enhancing the analytical efficiency of complex electroencephalogram signals. Description of the Drawings

[0052] Figure 1 It is a schematic structural diagram of the cognitive divergence mapping network architecture in an embodiment of the present invention.

[0053] Figure 2 It is a schematic structural diagram of the multi-branch brain region topology module in an embodiment of the present invention.

[0054] Figure 3 It is a schematic structural diagram of the neurodynamics physical information network module in an embodiment of the present invention.

[0055] Figure 4 It is a comparison diagram of ROC curves of the AUC performance of the CodimNet model and other methods on the CAUEEG dataset in an embodiment of the present invention.

[0056] Figure 5 It is a time-frequency comparison diagram of the intermediate layer features of the CodimNet model and the RNN model in an embodiment of the present invention.

[0057] Figure 6 It is a PSD diagram of the ND-PINN module in the CodimNet model in an embodiment of the present invention.

[0058] Figure 7 It is a heat map of feature extraction of the CodimNet model in an embodiment of the present invention. Detailed implementation manners

[0059] The following will describe in detail the specific implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the specific implementation manners described herein are only for explaining and understanding the present invention, and are not used to limit the present invention.

[0060] Accurately diagnosing Alzheimer's disease plays a crucial role in achieving early screening, implementing timely intervention, and effectively controlling the risk of disease progression. However, the existing methods for diagnosing Alzheimer's disease show deficiencies when facing complex electroencephalogram (EEG) signals. Relying on single-branch feature extraction, it is difficult to handle the antagonistic dynamic associations between frequencies in complex signals, resulting in limited diagnostic accuracy.

[0061] The cognitive feature extraction and classification system based on EEG signals of the present invention realizes the classification and recognition of Alzheimer's disease based on the analysis of 19-channel EEG signals by constructing a cognitive divergence mapping network CodimNet (Cognitive Divergence Mapping Network) based on 19-channel EEG signals. CodimNet is applicable to the early screening of Alzheimer's disease, the classification of mild cognitive impairment, and the mining of biomarkers for neurodegenerative diseases based on 19-channel EEG signals, providing a non-invasive diagnostic framework with both high accuracy and strong generalization ability for clinical use. This network integrates a multi-branch brain region topology module and a neurodynamic physics-informed neural network (ND-PINN) module. The multi-branch brain region topology module divides the detection branches of 5 brain regions (frontal lobe, central, parietal lobe, occipital lobe, temporal lobe) according to the international 10-20 system. By extracting features from different brain region branches, it significantly improves the feature learning ability in the case of cross-brain region signal heterogeneity of Alzheimer's disease patients. The neurodynamic physics-informed neural network module can effectively extract the changes in the low-frequency and high-frequency components of the EEG signals of Alzheimer's disease by introducing frequency band separation, Fourier transform power spectrum constraint, adaptive weighting mechanism, and neural representation embedding in the spectral congruence space. The present invention significantly improves the accuracy and generalization of EEG signal analysis, providing an efficient and non-invasive detection tool for the early diagnosis of Alzheimer's disease.

[0062] As Figure 1As shown in the figure, the cognitive feature extraction and classification method based on electroencephalogram (EEG) signals of the present invention is based on the EEG framework and uses a multi-branch topology with regional stratification for Alzheimer's disease classification. According to the international 10-20 system, the five-branch cortical decomposition divides the 19-channel EEG signals into the frontal lobe, central lobe, parietal lobe, occipital lobe, and temporal lobe. Each branch performs convolutional feature extraction through depthwise convolution (DW-Conv) and pointwise convolution (PW-Conv) transformation, followed by a densely connected convolutional cascade for hierarchical encoding. The neurodynamics physical information network ND-PINN module imposes spectral density constraints to disentangle the integrated frequency-domain prior neurophysiological features. Global average pooling (GAP) aggregates the extracted representations, and these representations are connected and processed through a fully connected classification head to generate the final output.

[0063] As Figure 2 shown, the multi-branch brain region topology module divides the EEG signal source space into five isomorphic modeling units according to the characteristics of the cerebral cortex functional partition: the frontal lobe region, central region, parietal lobe region, occipital lobe region, and temporal lobe region. Each branch uses depthwise separable convolution to construct a local receptive field to capture region-specific frequency-domain features and strengthens the inter-layer gradient flow transmission through a dense connection module (DenseNet). Compared with the single-branch architecture that is prone to coupling interference of cross-brain region information, the multi-branch brain region topology improves the feature learning ability in the case of cross-brain region signal heterogeneity.

[0064] First, the EEG signal can be abstracted as the superposition of multiple oscillatory components, which is mathematically represented as a non-stationary random process:

[0065]

[0066] where, A j represents the time-varying amplitude of the j-th oscillatory component, ω j is the corresponding angular frequency, reflecting the rhythm characteristics of a specific frequency band in the EEG signal, φ j is the phase shift, reflecting the phase relationship between different oscillatory components, and ε i (t) represents the noise term, including random interference generated by the experimental environment and physiology.

[0067] Let represent the input EEG signal, where N = 19 represents the number of EEG signal channels, and T represents the length of the time series. To capture the functional characteristics of different brain regions, we divide the input signal set into subsets according to the neuroanatomical functional partition:

[0068]

[0069] where, C r represents the set of EEG signal channel indices corresponding to the r-th brain region:

[0070] Frontal lobe:

[0071] Central:

[0072] Parietal lobe:

[0073] Occipital lobe:

[0074] Temporal lobe:

[0075] Next, we perform standard convolution on the original EEG input of each brain region to capture local features:

[0076]

[0077] where, are the convolution kernel parameters, and σ(·) is the ReLU activation function. This layer simulates the synaptic integration mechanism of local neuron clusters and learns the oscillation patterns of specific frequency bands (e.g., δ: 1 - 4 Hz, θ: 4 - 8 Hz) through the convolution kernel weights W.

[0078] Subsequently, the features of each brain region are subjected to depth convolution to extract the region - specific spectral features of each brain region:

[0079]

[0080] where, represents the intermediate feature representation of the j - th channel at time t in the r - th brain region. represents the input feature of the r - th brain region and the c - th input channel after time delay. represents the convolution kernel weights from the c - th input channel to the j - th channel during depth convolution, which represents the index of the convolution kernel in the time dimension. Using depth convolution can reduce the number of parameters and effectively capture local time - frequency domain features, embodying the idea of independent decoupling of channels in each brain region.

[0081] Then, through point - wise convolution (1×1 convolution), effective combination of cross - channel feature information is achieved to form the fused features of each brain region:

[0082]

[0083] where, represents the fused feature representation of the j - th channel in the r - th brain region, the weights of point - wise convolution, which are used for cross - channel feature fusion. This step can effectively extract the correlation between channels and further enhance the local feature expression ability.

[0084] Subsequently, the features of each brain region are further enhanced by a DenseNet module for cross-layer feature fusion:

[0085]

[0086] Among them, represents the feature of the r-th brain region at the i-th layer of dense connection, which is the representation after a time delay. represents the weights of the dense connection convolutional kernel, and represents the weights of the convolutional kernel for cross-layer connection between the i-th layer and the l+1-th layer. BN(·) represents batch normalization.

[0087] Then, the final global average pooling feature of each brain region is output as:

[0088]

[0089] Among them, represents the feature vector of the r-th brain region at time point t after passing through the last layer of the dense connection module. Performing a global average pooling operation (Global Average Pooling, GAP) on the feature sequence compresses the temporal feature vector into a single static feature, strengthening the feature expression of the overall region.

[0090] Finally, the features of the frontal lobe region, central region, parietal lobe region, occipital lobe region, and temporal lobe region are aggregated across regions, integrating the features independently decoupled and extracted from different brain regions to improve the feature learning ability in the case of cross-brain region signal heterogeneity:

[0091] F (global) = concat(F (1) , F (2) , F (3) , F (4) , F (5) )

[0092] Among them, concat(·) represents the feature concatenation operation, which connects the feature vectors of different brain regions into a long vector to form a global feature representation across brain regions.

[0093] As Figure 3 shown, the neurodynamics physical information network module realizes the interpretable decoupling and quantitative characterization of abnormal cross-frequency neural oscillation patterns related to Alzheimer's electroencephalogram in the spectral isomorphic space by constructing frequency-domain separated modeling, introducing a power spectral density constraint supervision mechanism, and implementing an adaptive spectral weight optimization strategy.

[0094] First, the input EEG signal x(t) is mapped to the frequency domain through the Fast Fourier Transform (FFT), decomposed into five frequency bands: δ (1 - 4 Hz), θ (4 - 8 Hz), α (8 - 13 Hz), β (13 - 30 Hz), and γ (30 - 45 Hz), and the power spectrum of each frequency band is calculated: The time-domain signal x(t) is converted to the frequency-domain representation X(f) through FFT:

[0095]

[0096] where P band represents the average power of a specific frequency band and is used to quantify the energy distribution characteristics of the AD-related pathological frequency bands.

[0097] Next, the intermediate layer features M mid of the network are mapped to the frequency domain through FFT, and its power spectrum P band,mid is calculated and frequency-domain aligned with the power spectrum P band,orig of the original signal. The difference between the two is constrained by the relative mean square error MSE j :

[0098]

[0099] where B = 5 is the number of frequency bands, ∈ = 10 -6 To avoid a zero denominator, the absolute MSE is converted to a relative error to eliminate the difference in energy magnitudes between different frequency bands. This constraint forces the intermediate features of the network to retain the energy distribution pattern of the pathological frequency bands of the original signal.

[0100] Then, the optimization weights are dynamically allocated according to the relative MSE of each frequency band, and the weights are restricted to the interval [0, 1] to ensure fair competition among the weights of all frequency bands:

[0101]

[0102] where the weight w j is inversely proportional to the MSE j error, forcing the network to preferentially optimize the non-linear antagonistic relationship between low frequencies (Delta / Theta) and high frequencies (Alpha / Beta / Gamma), thereby accurately modeling the "low-frequency compensation - high-frequency suppression" phenomenon caused by the disruption of neuronal synchrony in AD patients.

[0103] The final frequency-domain physics-informed neural network (ND-PINN) is constrained as:

[0104]

[0105] which passes through the original power P band,origNormalize the prediction error, eliminate the energy difference between frequency bands, and dynamically strengthen the optimization of difficult-to-fit frequency bands through adaptive inverse proportional weighting to ensure that the model captures the antagonistic relationship across frequency bands.

[0106] Apply the weights obtained from training to the test set for evaluating the classification effect. We take the main and key evaluation metrics for model evaluation, including Accuracy (Acc), Precision (Pr), F-measure (F1), Area Under the Curve (AUC), Geometric Mean (G-Mean), Sensitivity (Se), and Specificity (Sp). Their definitions are as follows:

[0107]

[0108]

[0109] Among them, TP i represents the number of true positives for each class C i (i = 1, 2, 3), FP i represents the number of false positives mispredicted as C i , FN i represents the number of false negatives that belong to C i but are mispredicted as other classes, TN i represents the number of true negatives that do not belong to C i and are correctly predicted as non-C i , and AUC i represents the area under the receiver operating characteristic curve for each class.

[0110] Based on these metric results, an individual-based independent evaluation method is adopted. Diagnosis is performed based on the complete individual EEG signals. By implementing a fixed validation and four-fold cross-validation scheme, the differences in classification performance between the proposed cognitive disagreement mapping network and existing methods are systematically compared, and the effectiveness of the proposed module is evaluated to verify its applicability and advantages in related tasks.

[0111] Table 1 Differences in classification performance between the fixed validation of the cognitive disagreement mapping network and existing methods on the CAUEEG dataset

[0112]

[0113] Table 1 Experimental data shows that this model has achieved optimal performance in key indicators such as accuracy (76.27%), area under the receiver operating characteristic curve (85.27%), and specificity (88.28%). Compared with the sub-optimal model EEGNet, this model has a 0.29 percentage point increase in sensitivity, a 2.07 percentage point increase in F1 score, and a 1.89 percentage point increase in geometric mean, indicating that it is more robust in dealing with class imbalance problems. This model adopts a multi-branch separation modeling strategy, establishing independent feature extraction channels in Alzheimer's disease-related brain regions such as the parietal lobe and temporal lobe, effectively capturing the unique brain wave changes in different brain regions, reflected in the increase in sensitivity (76.19%). The ND-PINN module constructs a spectral congruence space through Fourier transform, constrains the power spectral density, and effectively retains Alzheimer's disease-specific rhythm features such as delta waves and theta waves, enabling the model to still maintain an area under the receiver operating characteristic curve value of 85.27% in a complex noise environment, a 19.04% increase compared to the pure time-domain method.

[0114] Table 2 Differences in classification performance between the cognitive divergence mapping network and existing methods on the CAUEEG dataset in four-fold cross-validation

[0115]

[0116] Table 2 We used four-fold cross-validation to compare and evaluate the performance of different deep learning models in the three-class classification task of Alzheimer's disease based on electroencephalogram (EEG) signals. In four-fold cross-validation, all indicators decreased slightly overall, but our model was still the highest in all evaluation indicators. When using the dataset preprocessed by independent component analysis (ICA), the model performance was slightly lower than that of the dataset without ICA preprocessing (accuracy decreased from 66.13% to 64.26%), indicating that the model can autonomously learn artifact correction and effectively utilize signals containing artifacts, thereby enhancing its robustness and generalization ability.

[0117] As Figure 4 shown, CodimNet has a higher area under the receiver operating characteristic curve value in all four folds than other models, and the average area under the receiver operating characteristic curve value is significantly higher than other methods, indicating that this model has strong stability and generalization ability in the classification task. Especially in fold-3 (77.97%) and fold-4 (80.66%), it significantly outperforms other models.

[0118] Through the four-fold cross-validation scheme, we conducted ablation experiments. Systematically evaluated the effectiveness of the proposed module to verify its applicability and advantages in related tasks.

[0119] Table 3 Evaluating the contribution of regions to Alzheimer's disease diagnosis through cortical branch-specific ablation analysis

[0120]

[0121] Table 3 Experimental data shows that our model is significantly superior to the ablation variants of each branch in core indicators such as accuracy (66.13±1.38%), sensitivity (65.70±1.43%), and G-Mean (73.44±1.12%), verifying the advantage of the multi-branch architecture in processing spatially non-stationary signals, that is, by separately modeling the spatio-temporal dynamics of different brain regions, effectively coping with the cross-brain region heterogeneous frequency-domain distortion caused by Aβ deposition, so as to achieve multi-dimensional accurate capture of the pathological characteristics of Alzheimer's disease.

[0122] Table 4 Quantitative evaluation of the classification effect of the Alzheimer's disease EEG three-classification model under different numbers of ND-PINNs

[0123]

[0124] Table 4 shows that the neurodynamics physical information network module has a significant impact on the three-classification performance of Alzheimer's disease EEG. When the number of ND-PINN modules increases from 1 to 5, the comprehensive performance of the model shows a gradient improvement trend except for a marginal deviation when using 4 ND-PINN modules, proving that ND-PINN effectively captures the global inhibition and local compensation mechanisms of relevant neural oscillations in Alzheimer's disease EEG signals through frequency-domain separation.

[0125] Table 5 Quantitative evaluation of ND-PINN across different intermediate layer constraints

[0126]

[0127] Table 5 shows that there is a significant correlation between the position depth of the intermediate layer constraint and the model performance. CodimNet (constraining the 10th layer) achieves the best performance in multiple indicators, and its deep layer constraint mechanism more effectively captures the frequency-domain feature variations related to the pathology of Alzheimer's disease. From the perspective of neurodynamics, the deep intermediate layer constraint can strengthen the feature analysis ability in the Beta / Gamma high-frequency band, thus maintaining a high specificity of 82.77±0.81%, revealing the advantageous mechanism of the deep intermediate layer constraint in maintaining the fidelity of frequency-domain supervision.

[0128] Table 6 Quantitative evaluation of the impact of removing ND-PINNs in different brain region branches on the classification performance

[0129]

[0130] Table VI The experimental results show that when removing a single PINN branch, all performance indicators decline, verifying the effectiveness of multi-brain region collaborative modeling. The unique neurodegenerative dynamics characteristics of each brain region still need to be spatiotemporally decoupled by independent modeling branches. The absence of any single branch in ND-PINN will lead to the dimensional collapse of the global feature representation space.

[0131] Table VII Classification performance of different neural networks instantiated in the multi-branch paradigm

[0132]

[0133]

[0134] Based on the analysis of the ablation experiment results in Table VII, the performance of traditional recurrent neural network architectures (RNN / LSTM / GRU / BiLSTM) in the multi-branch framework shows significant limitations, lacking the ability to capture the spectral heterogeneity characteristics of EEG signals. Our model compensates for the frequency-domain information loss caused by time-domain aliasing in traditional models through a neurodynamics-guided frequency-domain supervision mechanism.

[0135] Table VIII Evaluating the impact of different brain rhythm constraints on EEG classification of Alzheimer's disease by ND-PINN

[0136]

[0137] The experimental data in Table VIII show that when removing specific frequency band constraints, the model performance presents differential attenuation, and the indicators are all better than the ablation model, confirming the importance of multi-band collaborative supervision for neurodynamics representation.

[0138] To verify the effectiveness of the CodimNet model in extracting EEG signal features of Alzheimer's disease patients, this study constructs a comparative analysis graph and a heat map containing time-frequency graphs, original PSD graphs, and PSD graphs after ND-PINN processing based on the forward propagation layer feature visualization technology of the hook function.

[0139] The specific analysis is as follows:

[0140] As Figure 5 shown, the time-frequency graph (Figure a) of the CodimNet model reveals significant differences in low-frequency and high-frequency features among Alzheimer's disease patients, patients with mild cognitive impairment, and normal individuals. In contrast, the time-frequency features of the RNN model (Figure b) have lower discrimination, and the feature differences among different categories (normal, mild cognitive impairment, Alzheimer's disease) are not as obvious as those of the CodimNet.

[0141] As Figure 6As shown, after ND-PINN processing (the third column), the low-frequency power is significantly improved, especially in the Delta (0-4Hz) and Theta (4-8Hz) bands, while the power in the high-frequency band (>10Hz) decreases more significantly, thus more clearly reflecting the characteristic EEG changes in Alzheimer's disease patients.

[0142] As Figure 7 shown, in the heatmaps generated by the CodimNet model, there are significant differences in the characteristic activation intensities of different category samples, reflecting the ability of the model to distinguish Alzheimer's disease, mild cognitive impairment, and normal samples.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cognitive feature extraction and classification system based on electroencephalogram signals, including a cognitive divergence mapping network, characterized in that: The cognitive disagreement mapping network includes a multi-branch brain region topology module and a neurodynamics physical information network; In the multi-branch brain region topology module, according to the characteristics of cerebral cortex functional division, the source space of electroencephalogram (EEG) signals is divided into multiple isomorphic modeling units to capture regional-specific frequency domain features respectively. Through the combination of cross-channel feature information, the fusion features of each brain region are formed, and then the features of each brain region are aggregated across regions; In the neurodynamics physical information network module, the EEG signals are mapped to the frequency domain and the power spectra of each frequency band are calculated. A power spectral density constraint supervision mechanism is introduced and an adaptive spectral weight optimization strategy is implemented to perform interpretable decoupling and quantitative characterization of abnormal patterns of cross-frequency band neural oscillations related to EEG signal data in the spectral isomorphic space.

2. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 1, characterized in that: In the multi-branch brain region topology module, the EEG signals are abstracted as the superposition of multiple oscillatory components, expressed as a non-stationary random process: Among them, A j represents the time-varying amplitude of the j-th oscillatory component, ω j represents the corresponding angular frequency, which is used to reflect the rhythm characteristics of a specific frequency band in the EEG signal, t represents time, and φ j represents the phase offset, which is used to reflect the phase relationship between different oscillatory components, and ε i (t) represents the noise term; The obtained EEG signal data are divided into subsets according to neuroanatomical functional regions; Among them, X represents the input electroencephalogram signal data, and C r represents the set of electroencephalogram signal channel indices corresponding to the r-th brain region.

3. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 2, wherein: The electroencephalogram signal data where N represents the number of electroencephalogram signal channels, T represents the length of the time series, and r ∈ {1, 2, 3, 4, 5}, respectively represent the electroencephalogram signal channel index sets corresponding to the frontal lobe, central lobe, parietal lobe, occipital lobe, and temporal lobe brain regions.

4. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 2, characterized in that: In the multi-branch brain region topology module, first, the local features of the EEG signal data of each brain region are captured: Among them, represents the convolution kernel parameters, σ(·) represents the activation function, simulating the synaptic integration mechanism of local neuron clusters, and learning the oscillation patterns of specific frequency bands through the convolution kernel weights K represents the hyperparameter describing the convolution kernel size, and τ represents the index of time delay; Subsequently, the regional-specific spectral features of each brain region are extracted: Among them, represents the intermediate feature representation of the j-th channel in the r-th brain region at time t, represents the input feature of the r-th brain region and the c-th input channel after time delay, represents the convolutional kernel weight from the c-th input channel to the j-th channel during the depth convolution process; Then, through the combination of cross-channel feature information, the fusion features of each brain region are formed: Among them, represents the fused feature representation of the j-th channel in the r-th brain region, represents the weight of the pointwise convolution, which is used to fuse features across channels.

5. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 4, wherein: In the multi-branch brain region topology module, the features of each brain region are further passed through a dense connection module to enhance cross-layer feature fusion: Among them, represents the feature that the r-th brain region is densely connected to the i-th layer, is the weight of the densely connected convolutional kernel, representing the convolutional kernel weight of the cross-layer connection between the i-th layer and the (l + 1)-th layer. BN(·) represents batch normalization, and L represents the number of layers of dense connection.

6. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 5, wherein: In the multi-branch brain region topology module, the final global average pooling feature of each brain region is output as: Among them, represents the feature vector of the output of the last layer of the dense connection module for the r-th brain region at time point t, represents the global average pooling operation on the feature sequence; Finally, the features of each brain region are aggregated across regions: F (global) = concat(F (r) ) Among them, concat(·) represents the feature splicing operation, which connects the feature vectors of different brain regions into a long vector to form a global feature representation across brain regions.

7. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 1, wherein: In the neurodynamics physical information network module, first, the input EEG signal x(t) is mapped to the frequency domain through the fast Fourier transform (FFT) into multiple frequency bands, and the power spectra of each frequency band are calculated: the time-domain signal x(t) is converted to the frequency-domain representation X(f) through FFT; Among them, P band represents the average power of a specific frequency band and is used to quantify the energy distribution characteristics of the relevant frequency band; Next, the intermediate network layer feature M mid is mapped to the frequency domain by FFT to calculate its power spectrum P band,mid , and is frequency-domain aligned with the power spectrum P band,orig of the original signal. The difference between the two is constrained by the relative mean square error MSE j : Among them, B represents the number of frequency bands, ∈ represents a small value, and the absolute mean squared error (MSE) is converted to a relative error to eliminate the difference in energy levels of different frequency bands.

8. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 7, characterized in that: In the neurodynamics physical information network module, the optimization weights are dynamically allocated according to the relative MSE of each frequency band; Among them, the weight w j is inversely proportional to the MSE j error, forcing the network to preferentially optimize the non-linear antagonistic relationship between low and high frequencies.

9. The cognitive feature extraction and classification system based on electroencephalogram signals according to claim 7, wherein: In the neurodynamics physical information network module, a frequency-domain physical information neural network constraint is constructed; where R represents the number of brain regions, and through the original power P of the frequency band band,orig Normalize the prediction error to eliminate the energy difference between frequency bands, and dynamically strengthen the optimization of difficult-to-fit frequency bands through adaptive inverse proportional weighting.

10. A cognitive feature extraction and classification method based on electroencephalogram signals, including a cognitive divergence mapping network, characterized in that: The cognitive disagreement mapping network includes a multi-branch brain region topology module and a neurodynamics physical information network; Step S1: Divide the EEG signal data; Step S2: Construct a cognitive disagreement mapping network according to the EEG signal-based cognitive feature extraction and classification system according to any one of claims 1 to 9; Step S3: Use the EEG signal data to train the cognitive disagreement mapping network; Step S4: Perform performance evaluation on the trained cognitive disagreement mapping network to optimize the performance.

Images