Cognitive impairment electroencephalogram signal analysis system and method based on multi-dimensional feature fusion

Abstract

This invention relates to the field of electronic digital data processing technology, specifically disclosing a cognitive impairment EEG signal analysis system and method based on multi-dimensional feature fusion. The method includes: 1. sequentially performing interference removal, signal enhancement, and format standardization processing on the raw EEG signal; 2. extracting temporal and frequency domain features from the format-standardized EEG signal using a model to obtain temporal local waveform features and frequency domain rhythm features. The pooling feature map from the temporal local waveform features is then processed by dual attention to obtain attention-enhanced temporal features; 3. fusing the temporal local waveform features, attention-enhanced temporal features, and frequency domain rhythm features; 4. performing adaptive regularization processing on the fused features; 5. mapping the regularized features to different categories of cognitive impairment to achieve predictive classification. This invention effectively enhances the mining of key brain region features, forming a multi-dimensional feature complementary fusion, and reliably improves the accuracy of the analysis results.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically a cognitive impairment EEG signal analysis system and method based on multi-dimensional feature fusion. Background Technology

[0002] Cognitive impairment refers to abnormalities in the brain's higher-level intelligent processing related to learning, memory, and judgment, leading to severe learning and memory disorders, often accompanied by changes such as aphasia, apraxia, agnosia, or apraxia, as seen in Alzheimer's disease. Timely and accurate detection of cognitive impairment is crucial for targeted treatment and slowing disease progression.

[0003] In diagnosing cognitive impairment, the common practice is to classify human cognition into four categories: normal control (NC), mild cognitive impairment (MCI), Alzheimer's disease (AD), and other cognitive impairments (OD). When confirming whether a patient has cognitive impairment and its severity, healthcare professionals typically make a comprehensive judgment based on data obtained from clinical examinations, including the Mini-Mental State Examination (MMSE), Activities of Daily Living (ADL) scales, MRI results, computed tomography (CT) scans, positron emission tomography (PET) electroencephalograms (EEGs), hematological examinations, cerebrospinal fluid analysis, and / or genetic testing.

[0004] In the human body, brain waves are electrical activities originating from brain nerve tissue. Nerve cells generate brain signals (i.e., brain waves) when transmitting signals. Many scholars and neuroscientists have conducted extensive and in-depth research on human brain signals. Research results show that brain signal activity exhibits certain regularities and varies under different cognitive, behavioral, or memory states. It corresponds closely with human consciousness or cognition, thus the diagnosis of cognitive impairment through brain signals has gained widespread attention in the medical community.

[0005] However, in techniques for diagnosing cognitive impairment based on electroencephalogram (EEG) signals, because EEG signals may contain other features or interfering factors unrelated to cognitive impairment, the raw EEG signals containing these interfering factors are usually not used as a direct diagnostic basis. Instead, auxiliary techniques are used to extract and analyze effective features related to cognitive impairment. These extracted and analyzed effective features related to cognitive impairment serve as the technical guidance for diagnosing cognitive impairment. Therefore, as an auxiliary technique, how to reliably and accurately extract and analyze effective features related to cognitive impairment from raw EEG signals is of significant technical guiding importance for ensuring the reliability and accuracy of subsequent EEG-based diagnoses of cognitive impairment.

[0006] With the continuous development of artificial intelligence technology, the extraction and analysis of effective features related to cognitive impairment in EEG signals currently relies on more advanced and reliable neural network computing models. This involves constructing cognitive impairment analysis techniques based on EEG signals using corresponding neural network computing models. Examples include Chinese patent documents such as: "Cognitive Impairment Monitoring System and Method Based on EEG Analysis" (Publication No. CN118873097A, Publication Date November 1, 2024); "EEG Signal Analysis Method and System for Mild Cognitive Impairment Based on Multi-Band PCMI-EBN" (Publication No. CN114631827A, Publication Date June 17, 2022); and "An Early Judgment System for Parkinson's Disease Cognitive Impairment Based on Multidimensional Features of Transcranial Magnetic-Electroencephalography" (Publication No. CN119745323A, Publication Date April 4, 2025). However, most of the publicly available cognitive impairment analysis technologies based on EEG signals, which are built on corresponding neural network computing models, generally suffer from technical problems such as insufficient mining of key brain region features, incomplete representation of single-domain features, and low classification accuracy. This is particularly prominent in the accurate differentiation between MCI and NC classifications, resulting in poor reliability and engineering adaptability. Summary of the Invention

[0007] The technical objective of this invention is to address the unique challenges of extracting and analyzing effective features related to cognitive impairment from raw EEG signals, as well as the shortcomings of existing cognitive impairment analysis techniques based on corresponding neural network computational models. This invention provides a cognitive impairment EEG signal analysis system that effectively enhances the mining of key brain region features, forms a multi-dimensional feature complementarity and fusion based on temporal local waveform features, attention enhancement temporal features, and frequency domain rhythmic features, reliably improves the accuracy of cognitive impairment classification, and also considers engineering practicality. The invention also includes a method for implementing this analysis system.

[0008] The technical objective of this invention is achieved through the following technical solution: a cognitive impairment EEG signal analysis system based on multi-dimensional feature fusion, the cognitive impairment EEG signal analysis system comprising: The raw EEG data input module is used to input the acquired raw multi-channel resting-state EEG signals, which are presented as two-dimensional time-series signals. The bandpass filter module is used to remove signal frequencies with noise interference from the EEG signal and retain signal frequencies that can contain delta rhythm features, theta rhythm features, alpha rhythm features, and beta rhythm features related to cognitive impairment. The artifact removal module is used to remove signal artifacts, including amplitude anomalies and rhythm disorders. The signal cropping and sample augmentation module uses the original EEG signal after artifact removal as a basis and performs non-overlapping random cropping according to a fixed cropping length. The Gaussian noise enhancement module is used to simulate slight noise interference during the actual acquisition of EEG signals, so as to improve the robustness of the deep learning model to the real acquisition scenario of EEG signals and avoid excessive noise from damaging the pathological features of EEG. The timeline scaling module is used to change the temporal rhythm of cropped samples without altering the medical and pathological characteristics of EEG signals, thereby expanding data diversity by simulating the rhythmic differences of EEG signals among different individuals. The amplitude scaling module, without destroying the waveform characteristics and pathological rhythm of the EEG signal, performs a small-amplitude scaling on the amplitude of the cropped sample to simulate the amplitude differences of EEG signals from different individuals and thus expand the diversity of the sample data. The data balancing module performs class balancing on the enhanced training sample set according to the classification of cognitive impairment to ensure the consistency of the learning weights of the deep learning model for each class of samples. The format standardization module is used to perform dimension alignment and remove dimensional interference from the balanced dataset, so as to accelerate the convergence of deep learning model training and improve training stability. The temporal feature extraction module is used to sequentially perform convolution, max pooling, and flattening on the input format-normalized signal to extract temporal local waveform features from the EEG signal, including waveform peaks, waveform valleys, and fluctuation slopes. The dual attention module is used to perform channel attention and temporal attention processing on the pooled feature map output by the temporal feature extraction module, so as to adaptively allocate feature weights, focus on key pathological features, suppress interference from irrelevant features, and output attention-enhanced temporal features. The frequency domain feature extraction module is used to sequentially perform FFT transformation, PSD processing, rhythm energy extraction processing, and fully connected dimensionality reduction processing on the input format-normalized signal in order to extract frequency domain rhythm features, including rhythm energy and power spectral density, from the EEG signal. The feature interaction fusion module is used to stitch and calibrate time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythm features to achieve complementary fusion of multi-dimensional features; The adaptive regularization module is used to adaptively regularize the fused features, suppress overfitting, improve the generalization ability and training stability of the deep learning model, and adaptively adjust the regularization strength to avoid feature loss due to over-regulation. The classification output module is used to map the regularized features to the probability space of four types of cognitive impairment, and to perform classification prediction and output the predicted probability of each category according to the decision rules.

[0009] A cognitive impairment EEG signal analysis method based on multi-dimensional feature fusion, wherein the cognitive impairment EEG signal analysis method is used as the cognitive impairment EEG signal analysis system based on multi-dimensional feature fusion described above; The method for analyzing EEG signals in cognitive impairment includes the following steps: Step 1. The acquired raw EEG signals are sequentially processed through interference removal preprocessing, signal enhancement processing, and format standardization processing; Step 2. The standardized EEG signals are processed by time-domain feature extraction and frequency-domain feature extraction using a deep learning model. The temporal feature extraction process is used to obtain the temporal local waveform features of the EEG signal. The pooled feature map obtained from the temporal feature extraction process is then processed by a dual attention mechanism to obtain attention-enhanced temporal features. The frequency domain rhythmic features of the EEG signal are obtained through the frequency domain feature extraction process. Step 3. Perform feature interaction fusion processing on the time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythm features to achieve multi-dimensional feature complementarity; Step 4. Apply adaptive regularization to the fused features to suppress overfitting; Step 5. Map the regularized features to different categories of cognitive impairment and implement the predictive classification of cognitive impairment according to the decision rules.

[0010] Furthermore, in step 1, the interference removal preprocessing involves sequentially performing bandpass filtering and artifact removal processing on the original EEG signal; The bandpass filtering process involves using an FIR filter to filter the original EEG signal, retaining frequency components within the range of 0.5Hz to 45Hz, to obtain a denoised EEG signal matrix. ; The bandpass filtering process processes single-channel signals in the original EEG signal matrix, satisfying the following relationship: ; In the formula, This represents the signal amplitude after single-channel denoising. Indicates the inverse Fourier transform; Indicates Fourier transform; Indicates the first The first channel, the first The amplitude of the electroencephalogram signal at each sampling point; Let represent the frequency response function of an ideal bandpass filter, satisfying the following relationship: ; The artifact removal process is achieved by combining amplitude threshold denoising with ICA independent component analysis. The amplitude threshold denoising removes amplitude abnormality artifacts, resulting in an initial artifact-removed EEG signal matrix. ; The amplitude threshold denoising process, which processes single-channel signals in the denoised EEG signal matrix, satisfies the following relationship: ; In the formula, This represents the signal amplitude after removing amplitude anomaly artifacts in a single channel; This represents the signal amplitude after single-channel denoising. Indicates the amplitude threshold; The ICA (Independent Component Analysis) method involves visualizing the waveforms of independent components, identifying and removing corresponding independent components, and then reconstructing the remaining effective independent components of the EEG signal using a hybrid matrix to obtain the final artifact-free EEG signal matrix. ; The ICA independent component analysis, for single-channel mixture matrix reconstruction that preserves the effective independent components in the EEG signal matrix, satisfies the following relationship: ; In the formula, This represents the final artifact removal signal amplitude for a single channel; Represents the mixing matrix for a single channel; Represents the independent component matrix of a single channel; This represents the matrix of remaining independent components after removing artifacts from a single channel.

[0011] Furthermore, in step 1, the signal enhancement processing involves sequentially performing signal cropping and sample augmentation, Gaussian noise enhancement, time axis scaling, amplitude scaling, and data equalization on the interference-free preprocessed EEG signal. The cropping and sample augmentation process involves performing non-overlapping random cropping of the artifact-removed EEG signals according to a fixed cropping length. The initial augmented sample set obtained after cropping satisfies the following relationship: ; Furthermore, the amplitude of the single-channel signal of the cropped sample satisfies the following relationship: ; In the formula, The signal matrix representing the cropped samples; Indicates a fixed cutting length; Indicates the index of the cropped sample; This represents the total number of samples after cropping. Indicates the first The first cutting sample in the th The first channel, the first The signal amplitude at each sampling point; This represents the final artifact removal signal amplitude for a single channel; Indicates the first A random starting index for each cropped sample, satisfying... , And the interval between any two cropped samples is No overlap; The Gaussian noise enhancement process simulates the slight noise interference in the actual EEG signal acquisition process, so as to improve the robustness of the subsequent deep learning model to the real EEG signal acquisition scenario, while avoiding excessive noise that could damage the pathological features of EEG. After the Gaussian noise enhancement process, an enhanced sample EEG signal matrix is ​​obtained after injecting small-amplitude Gaussian noise. In order to preserve the core pathological rhythm of the original electroencephalogram signal; The Gaussian noise enhancement process, which processes single-channel signals in the cropped sample EEG signal matrix, satisfies the following relationship: ; In the formula, This represents a single-channel cropped sample after injecting small-amplitude Gaussian noise. This represents the signal amplitude of the cropped sample; This indicates small-amplitude Gaussian noise; The time axis scaling process involves changing the time axis length of the cropped sample and then restoring it to the original fixed cropping length through interpolation. To expand sample diversity and simulate the rhythm differences of EEG signals among different individuals; After the time axis scaling process, the enhanced sample EEG signal matrix is ​​obtained after time axis scaling. ; The time axis scaling process, which processes single-channel signals in the cropped EEG signal matrix after Gaussian noise enhancement, satisfies the following relationship: ; In the formula, This represents a single-channel cropped sample after time-axis scaling. This indicates that linear interpolation is used to restore the resampled signal to its original length. ; This indicates that the time axis of the cropped sample is resampled according to the scaling factor; This indicates the scaling factor set. The amplitude scaling process involves slightly scaling the amplitude of the cropped sample to simulate the amplitude differences of EEG signals from different individuals and to expand the sample diversity. After the amplitude scaling process, the amplitude-scaled enhanced sample EEG signal matrix is ​​obtained. ; The amplitude scaling process, which processes single-channel signals in the cropped EEG signal matrix after time-axis scaling, satisfies the following relationship: ; In the formula, This represents a single-channel cropped sample after amplitude scaling. This indicates the set amplitude scaling factor; This represents the amplitude of the c-th channel and the t-th sampling point of the cropped sample; In the initial expanded sample set Based on this, the training sample set is defined as follows after sequentially performing Gaussian noise enhancement processing, time axis scaling processing, and amplitude scaling processing: ; In the formula, This represents the enhanced training sample set; This represents the total number of samples after augmentation; The number of channels representing the brainwave signal; Indicates a fixed cutting length; The data balancing process is performed by weighted random sampling to balance the enhanced training sample set according to the cognitive impairment classification, so as to ensure the consistency of the learning weights of each category of samples by the subsequent deep learning model. The data equalization process involves sampling the enhanced training sample set according to the sampling probability, and the final equalized dataset is defined as follows: ; In the formula, This represents the balanced dataset; 4 indicates the number of categories in the cognitive impairment classification; Indicates the number of samples in each category; The number of channels representing the brainwave signal; Indicates a fixed cutting length; The sampling probability in the data equalization process satisfies the following relationship: ; In the formula, Indicates the sampling probability; The category number indicating the classification of cognitive impairment; Table after sampling, the th The number of samples in a class.

[0012] Furthermore, in step 1, the format standardization process involves dimensional alignment and removal of dimensional interference from the balanced dataset to accelerate the convergence of the subsequent deep learning model training and improve training stability. The format standardization process includes a dimension organization process and a channel-by-channel Z-score standardization process. The dimensionality adjustment process involves merging the category and sample dimensions of the balanced dataset and adjusting it to the input format required by the deep learning model. It satisfies the following relationship: ; In the formula, This represents the dataset after dimension adjustment; This represents the balanced dataset; Indicates batch size; Indicates the number of EEG signal channels; Indicates a fixed cutting length; The channel-by-channel Z-score normalization process normalizes the signal of each channel to a distribution with a mean of 0 and a variance of 1, removes dimensional interference, and ensures that the weights of the features in each channel are consistent, thereby obtaining the format-normalized EEG signal. Its satisfaction , ; The channel-by-channel Z-score normalization process satisfies the following relationship: ; In the formula, This represents the amplitude of a standardized single-channel signal. This represents the amplitude of a single-channel signal after dimension adjustment. Indicates the first The average signal value of each channel; For the first The signal standard deviation of each channel; ; The first The average signal values ​​of each channel satisfy the following relationship: ; In the formula, Indicates a fixed cutting length; Indicates the first Sampling points; c represents the first One channel; The first The standard deviation of the signal in each channel satisfies the following relationship: .

[0013] Furthermore, in step 2, the deep learning model has an input layer, which performs input signal processing and signal de-pathing processing on the received standardized EEG signals. The input signal processing satisfies the following relationship: ; In the formula, Indicates the input signal; Represents the standardized EEG signal matrix; Indicates batch size; Indicates the number of EEG signal channels; Indicates a fixed cutting length; The signal splitting process involves directly dividing the input signal into two paths in the time domain according to the time domain feature extraction architecture and frequency domain feature extraction architecture of the deep learning model, thereby realizing dual-branch parallel feature extraction where the two signals share the same input source.

[0014] Furthermore, in step 2, the deep learning model has a temporal feature extraction architecture, which sequentially performs convolution, max pooling, and flattening processing on the input signal to extract temporal local waveform features from the EEG signal, including waveform peaks, waveform valleys, and fluctuation slopes. The convolution process involves performing a convolution operation on the input signal along the time axis to capture local waveform features in the time domain, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Represents the convolutional feature map; Indicates the activation function; Indicates the number of output channels of the convolution; This indicates the length of the convolution kernel along the time axis; Indicates the first One sampling point; Represents the time-domain input signal; The weight matrix represents the convolution kernel; This represents the bias vector of the convolutional layer; Indicates the temporal length after convolution; The max pooling process reduces the dimensionality of the feature map output by the convolutional layer while preserving the key peak and valley features of the temporal waveform, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Represents the max-pooling feature map; This represents the time-series index after pooling; This indicates the pooling kernel size, and the step size is the same as the pooling kernel size. Indicates the time sequence length after pooling; The flattening process involves flattening the 2D feature map output by the max pooling layer into a 1D feature vector to adapt to the input requirements of subsequent feature fusion and fully connected layers, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, This represents the flattened temporal feature vector; This indicates that the format has been flattened. Represents the max-pooling feature map; Indicates the first The first batch of samples 3D time-domain eigenvalues; This represents the dimension of the temporal feature vector.

[0015] Furthermore, in step 2, the deep learning model has a dual attention architecture, which sequentially performs channel attention processing and temporal attention processing on the pooled feature map output by the temporal feature extraction architecture to adaptively allocate feature weights and focus on key pathological features. The channel attention processing involves multiplying the channel weights by the input feature map channel by channel to obtain a channel-attention-enhanced feature map, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, This represents the feature map after channel attention enhancement; This represents the channel attention weight matrix; Represents the max-pooling feature map; The channel attention weight matrix satisfies the following relationship: ; In the formula, Indicates the first Batch samples, the first The weight values ​​of each channel; Indicates the activation function; This represents the weight matrix of the second fully connected layer; Indicates the activation function; This represents the weight matrix of the first fully connected layer; Intermediate feature map representing channel attention; This represents the bias vector of the first fully connected layer; This represents the bias vector of the second fully connected layer; The intermediate feature maps of the channel attention satisfy the following relationship: ; In the formula, Indicates the first Batch samples, the first The global average pooling value of each convolution output channel; This indicates global average pooling; Represents the max-pooling feature map; Indicates the time sequence length after pooling; This represents the time-series index after pooling; The temporal attention processing involves multiplying the temporal weights by the channel attention-enhanced feature map sequentially to obtain the dual-channel attention-enhanced temporal feature map, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, This represents the temporal feature map after dual-channel attention enhancement; Represents the temporal attention weight matrix; This represents the feature map after channel attention enhancement; The temporal attention weight matrix satisfies the following relationship: ; In the formula, Indicates the first Batch samples, the first The weight values ​​of each time segment; Indicates the activation function; This represents the weight matrix of the second fully connected layer; Indicates the activation function; This represents the weight matrix of the first fully connected layer; Intermediate feature map representing temporal attention; This represents the bias vector of the first fully connected layer; This represents the bias vector of the second fully connected layer; The intermediate feature maps of the temporal attention satisfy the following relationship: ; In the formula, Indicates the first Batch samples, the first Cross-channel global average pooling value for each time segment; This indicates that global average pooling is being performed. This represents the feature map after channel attention enhancement; Indicates the number of output channels of the convolution; The dual-attention architecture, to adapt to subsequent feature fusion, flattens the temporal feature map after dual-channel attention enhancement, resulting in attention-enhanced temporal feature vectors that satisfy the following relationship: ; In the formula, Indicates attention-enhanced temporal features; This indicates that the format has been flattened. This represents the temporal feature map after dual-channel attention enhancement; This represents the dimension of the temporal feature vector.

[0016] Furthermore, in step 2, the deep learning model has a frequency domain feature extraction architecture, which sequentially performs FFT transformation, PSD processing, rhythm energy extraction processing, and fully connected dimensionality reduction processing on the input signal to extract frequency domain rhythm features, including rhythm energy and power spectral density, from the EEG signal. The FFT transformation process converts the time-domain input signal into a frequency-domain signal, capturing the frequency distribution characteristics of the signal, which satisfy the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First The frequency domain complex signal at the frequency point; Represents the Discrete Fourier Transform; Represents the time-domain input signal; The PSD processing converts the frequency domain complex signal into a real-valued form of power spectral density and quantizes the signal energy distribution at different frequencies, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First Power spectral density value at a frequency point; Indicates a fixed cutting length; Represents the square of the modulus of a complex signal in the frequency domain; The rhythm energy extraction process involves screening out frequency bands related to cognitive impairment, extracting the energy characteristics of delta rhythms, theta rhythms, alpha rhythms, and beta rhythms within these bands, eliminating irrelevant frequency components, focusing on pathological features, and satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First The energy value of rhythmic patterns; Indicates the frequency point; Indicates the first A set of frequency indices corresponding to rhythm types; These correspond to four types of rhythms: δ, θ, α, and β, respectively. The rhythm energy extraction process is adapted to subsequent fully connected dimensionality reduction by flattening the obtained rhythm energy feature matrix, transforming the 3D energy feature matrix into a 1D feature vector, satisfying the following relationship: ; In the formula, This represents the flattened frequency domain feature vector; This indicates that the format has been flattened. Represents the rhythmic energy characteristic matrix; express A set of vectors consisting of n real numbers; The fully connected dimensionality reduction process reduces the dimensionality of the flattened rhythmic energy features to adapt them for subsequent feature fusion, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Represents a complex signal in the frequency domain; Indicates the activation function; This represents the weight matrix of the fully connected layer; This represents the bias vector of the fully connected layer; This indicates the output dimension of the frequency domain features.

[0017] Furthermore, in step 3, the feature interaction fusion process first adjusts the dimension of the frequency domain rhythm features to make them consistent with the dimension of the time domain local waveform features, then concatenates the time domain local waveform features, attention-enhanced time domain features, and frequency domain rhythm features, and then performs feature calibration through a fully connected layer. The dimension adjustment of frequency domain rhythm features is achieved using a fully connected layer, satisfying the following relationship: ; In the formula, This represents the frequency domain features after dimension adjustment; Indicates the activation function; This indicates the dimension-adjusted weight matrix of the fully connected layer. This indicates the dimension adjustment of the fully connected layer bias vector; The splicing process of local waveform features in the time domain, attention-enhanced time domain features, and frequency domain rhythmic features satisfies the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the initial fusion features; Represents local waveform characteristics in the time domain; Indicates temporal features after attention enhancement; This represents the frequency domain features after dimension adjustment; It contains complete information encompassing three categories of features: temporal domain, attention-enhanced temporal domain, and frequency domain. The feature calibration process satisfies the following relationship: ; In the formula, This represents the final fused feature vector; Indicates the activation function; This indicates the weight matrix of the fully connected layer for feature fusion calibration; Indicates the final fused feature dimension; It contains complete information encompassing three categories of features: temporal domain, attention-enhanced temporal domain, and frequency domain. This represents the bias vector of the fully connected layer calibrated by fusion features.

[0018] Furthermore, in step 4, the adaptive regularization process satisfies the following relationship: ; And the output dimension matrix is ; In the formula, Features after adaptive regularization; Indicates the activation function; Indicates the weights of the fully connected layer; This represents the final fused feature vector; Indicates the bias of the fully connected layer; This represents the feature dimension after regularization. This represents an additional term in the adaptive regularization loss function.

[0019] Furthermore, in step 5, the decision rule for predicting classification is to select the category with the highest predicted probability as the final classification result of the sample, satisfying the following relationship: ; In the formula, Indicates the first Predicted category labels for batch samples; Indicates the classification category of cognitive impairment; Represents the first in the classification prediction probability matrix Batch of samples belongs to the first The predicted probability of cognitive impairment is represented by the classification prediction probability matrix as follows: ; The first Batch of samples belongs to the first The predicted probability of cognitive impairment satisfies the following relationship: ; In the formula, Softmax Indicates the activation function; This represents the weight matrix of the classification output layer; Features after adaptive regularization; This represents the bias vector of the classification output layer.

[0020] The beneficial technical effects of the present invention are as follows: the above technical measures are designed to address the particularity of extracting and analyzing effective features related to cognitive impairment from raw EEG signals. They are constructed with a neural network computing model that includes a time-domain feature extraction architecture, a dual-attention processing architecture, and a frequency-domain feature extraction architecture as the core. The constructed deep learning model forms an architecture that integrates time-domain and frequency-domain dual branches and introduces a dual-attention mechanism for corresponding channels. Based on the aforementioned deep learning model that integrates time-domain and frequency-domain dual-branch fusion and introduces corresponding channel dual-attention mechanisms, the above technical measures sequentially process the acquired raw EEG signals through interference removal preprocessing, signal enhancement processing, and format standardization. On the one hand, this facilitates the lightweight construction of the deep learning model; on the other hand, the interference removal preprocessing of the acquired raw EEG signals preserves effective pathological features related to cognitive impairment, the signal enhancement processing expands the diversity of samples and enhances the integrity of the pathological features of the samples, and the format standardization processing accelerates the training convergence of the subsequent deep learning model and improves the training stability, thereby contributing to obtaining highly reliable and accurate analysis results. The deep learning model of the aforementioned technical measures combines the physiological characteristics of EEG signals in cognitive impairment. It uses a temporal feature extraction architecture to obtain temporal local waveform fluctuations that reflect the rhythm of neural activity, a frequency domain feature extraction architecture to obtain frequency domain rhythmic features that reflect pathological differences, and a dual-attention processing architecture to adaptively focus on important cognitively relevant channels, such as the prefrontal and temporal lobes, focusing on key pathological features and suppressing irrelevant features. This effectively enhances the extraction of key brain region features and forms a multi-dimensional feature complementarity fusion based on temporal local waveform features, attention-enhanced temporal features, and frequency domain rhythmic features. This reliably improves the accuracy of the analysis results, specifically the accuracy of cognitive impairment classification, especially the feature discrimination between MCI and NC classifications, while also considering engineering practicality. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the system architecture and data processing of the present invention. Detailed Implementation

[0022] This invention relates to the field of electro-digital data processing technology for healthcare information, specifically a cognitive impairment electroencephalogram (EEG) signal analysis system based on multi-dimensional feature fusion, and a method for implementing this analysis system. The following description, in conjunction with the accompanying drawings, illustrates this. Figure 1 The technical solution of this invention will be clearly and thoroughly explained.

[0023] It should be noted that the accompanying drawings of this invention are schematic, and unnecessary details have been simplified to clarify the technical objectives of this invention, so as to avoid obscuring the technical solutions contributed by this invention to the prior art. Furthermore, the terms "approximately" or "basically" used below to refer to quantities or fit relationships mean that reasonable assembly and processing errors are allowed in the industry, and do not literally describe absolute quantities or fit relationships.

[0024] The cognitive impairment EEG signal analysis system of the present invention is implemented based on the following multi-dimensional feature fusion method.

[0025] See Figure 1 As shown, the cognitive impairment EEG signal analysis system of the present invention includes a raw EEG data input module, a bandpass filtering module, an artifact removal module, a signal pruning and sample augmentation module, a Gaussian noise enhancement module, a time axis scaling module, an amplitude scaling module, a data equalization module, a format normalization module, a temporal feature extraction module, a dual attention module, a frequency domain feature extraction module, a feature interaction fusion module, an adaptive regularization module, and a classification output module. Specifically, the bandpass filtering module and the artifact removal module sequentially perform interference-removing preprocessing on the raw EEG signals input from the raw EEG data input module. The signal pruning and sample augmentation module, the Gaussian noise enhancement module, the time axis scaling module, the amplitude scaling module, and the data equalization module sequentially perform enhancement processing on the interference-removed EEG signals to expand sample diversity and completeness. The temporal feature extraction module, the dual attention module, the frequency domain feature extraction module, the feature interaction fusion module, the adaptive regularization module, and the classification output module construct a deep learning model that integrates temporal and frequency domain dual-branch fusion and introduces a corresponding channel dual attention mechanism.

[0026] Specifically, the raw EEG data input module is used to input the acquired raw multi-channel resting-state EEG signals, which are presented as two-dimensional time-series signals; The bandpass filter module is used to remove signal frequencies with noise interference from the EEG signal and retain signal frequencies that can contain delta rhythm features, theta rhythm features, alpha rhythm features, and beta rhythm features related to cognitive impairment. The artifact removal module is used to remove signal artifacts, including amplitude anomalies and rhythm disorders. The signal cropping and sample augmentation module uses the original EEG signal after artifact removal as a basis and performs non-overlapping random cropping according to a fixed cropping length. The Gaussian noise enhancement module is used to simulate slight noise interference during the actual acquisition of EEG signals, so as to improve the robustness of the deep learning model to the real acquisition scenario of EEG signals and avoid excessive noise from damaging the pathological features of EEG. The timeline scaling module is used to change the temporal rhythm of cropped samples without altering the medical and pathological characteristics of EEG signals, thereby expanding data diversity by simulating the rhythmic differences of EEG signals among different individuals. The amplitude scaling module, without destroying the waveform characteristics and pathological rhythm of the EEG signal, performs a small-amplitude scaling on the amplitude of the cropped sample to simulate the amplitude differences of EEG signals from different individuals and thus expand the diversity of the sample data. The data balancing module performs class balancing on the enhanced training sample set according to the classification of cognitive impairment—namely, normal control (NC), mild cognitive impairment (MCI), Alzheimer's disease (AD), and other cognitive impairments (OD)—to ensure that the deep learning model learns consistent weights for each class of samples. The format standardization module is used to perform dimension alignment and remove dimensional interference from the balanced dataset, so as to accelerate the convergence of deep learning model training and improve training stability. The temporal feature extraction module is used to sequentially perform convolution, max pooling, and flattening on the input format-normalized signal to extract temporal local waveform features from the EEG signal, including waveform peaks, waveform valleys, and fluctuation slopes. The dual attention module is used to perform channel attention and temporal attention processing on the pooled feature map output by the temporal feature extraction module, so as to adaptively allocate feature weights, focus on key pathological features, suppress interference from irrelevant features, and output attention-enhanced temporal features. The frequency domain feature extraction module is used to sequentially perform FFT transformation, PSD processing, rhythm energy extraction processing, and fully connected dimensionality reduction processing on the input format-normalized signal in order to extract frequency domain rhythm features, including rhythm energy and power spectral density, from the EEG signal. The feature interaction fusion module is used to stitch and calibrate time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythm features to achieve complementary fusion of multi-dimensional features; The adaptive regularization module is used to adaptively regularize the fused features, suppress overfitting, improve the generalization ability and training stability of the deep learning model, and adaptively adjust the regularization strength to avoid feature loss due to over-regulation. The classification output module is used to map the regularized features to the probability space of four types of cognitive impairment (NC, MCI, AD, and OD). It implements classification prediction and outputs the predicted probabilities of each category according to the decision rule of selecting the category with the highest predicted probability as the final classification result of the sample.

[0027] The aforementioned cognitive impairment EEG signal analysis system is implemented through the following cognitive impairment EEG signal analysis method based on multi-dimensional feature fusion, namely, the following cognitive impairment EEG signal analysis method, which is the specific implementation process of the aforementioned cognitive impairment EEG signal analysis system.

[0028] See Figure 1 As shown, this invention, as a cognitive impairment EEG signal analysis method based on multi-dimensional feature fusion, includes the following five major execution steps, and the specific execution process of each step under each major execution step: Step 1. To ensure the quality of EEG data, solve the problem of uneven data distribution, achieve accurate adaptation with the input format of deep learning models, and improve the training stability and classification performance of deep learning models, it is necessary to perform signal preprocessing, data augmentation, and format standardization on the collected raw EEG data. That is, the collected raw EEG signals are sequentially subjected to interference removal preprocessing, signal augmentation processing, and format standardization processing. Suppose the acquired raw multi-channel resting-state EEG signal (i.e., raw EEG signal) is a two-dimensional time-series signal, and its mathematical definition is: ; In the formula, Represents the original EEG signal matrix; The number of channels represents the EEG signal; 19 channels were used in the experiment, corresponding to different brain regions. This represents the total number of sampling points per channel. In the experiment, the sampling frequency was 100Hz and the acquisition time was 200s. ; In the original EEG signal matrix, the single-channel time-series signal is defined as: ; In the formula, Indicates the first The first channel, the first The amplitude of the EEG signal at each sampling point is expressed in μV. This represents the c-th channel; This represents the t-th sampling point; Step 2. The standardized EEG signal is processed by the temporal feature extraction architecture (i.e., the temporal CNN branch) and frequency feature extraction architecture (i.e., the frequency feature branch) of the deep learning model for corresponding temporal and frequency feature extraction processing. The temporal feature extraction architecture is used to extract temporal features from the input signal to obtain the temporal local waveform features of the EEG signal. To adaptively allocate feature weights and focus on key pathological features, the pooled feature map obtained from temporal feature extraction is processed by the dual attention architecture (i.e., dual attention mechanism) of the deep learning model to obtain attention-enhanced temporal features. The frequency domain feature extraction architecture is used to extract frequency domain features from the input signal to obtain the frequency domain rhythm features of the EEG signal. Step 3. Perform feature interaction fusion processing on the time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythm features to achieve multi-dimensional feature complementarity; Step 4. Apply adaptive regularization to the fused features to suppress overfitting; Step 5. Map the regularized features to different categories of cognitive impairment to achieve predictive classification of cognitive impairment.

[0029] Raw EEG signals are susceptible to noise such as DC drift, power frequency interference, and electrooculography / electromyography artifacts, making direct cropping, expansion, and model training impossible. Therefore, it is necessary to perform interference removal preprocessing on the acquired raw EEG signals. The core objective of interference removal preprocessing is to remove irrelevant noise and artifacts while preserving effective pathological features related to cognitive impairment. Specifically, in step 1 above, the interference removal preprocessing involves sequentially performing bandpass filtering and artifact removal processing on the raw EEG signals.

[0030] Specifically, the raw EEG signal contains DC drift (around 0Hz), power frequency interference (50Hz), and high-frequency noise (>45Hz). This noise is irrelevant to the pathological characteristics of cognitive impairment and needs to be removed by bandpass filtering. Therefore, bandpass filtering uses an FIR filter to filter the raw EEG signal, retaining the signal frequencies related to cognitive impairment, including delta rhythm, theta rhythm, alpha rhythm, and beta rhythm features, i.e., retaining the frequency components in the range of 0.5Hz to 45Hz, resulting in a denoised EEG signal matrix. .

[0031] Bandpass filtering for processing single-channel signals in the original EEG signal matrix satisfies the following relationship: ; In the formula, This represents the signal amplitude after single-channel denoising. This represents the inverse Fourier transform, used to convert a filtered frequency domain signal back to a time domain signal; This represents the Fourier transform, used to convert a time-domain signal into a frequency-domain signal; Indicates the first The first channel, the first The amplitude of the electroencephalogram signal at each sampling point; Let represent the frequency response function of an ideal bandpass filter, satisfying the following relationship: .

[0032] In this way, only frequency components in the range of 0.5Hz to 45Hz are retained, while noise signals outside this range are suppressed.

[0033] Even after filtering, EEG signals may still contain artifacts such as electrooculography (EOG) and electromyography (EMG) artifacts (abnormal amplitude, rhythm disturbances), which need to be further removed to avoid interfering with subsequent cropping operations and the deep learning model's capture of pathological features. Therefore, a combination of "amplitude threshold denoising + ICA independent component analysis" is adopted to balance artifact removal effectiveness and feature preservation. That is, artifact removal is achieved by combining amplitude threshold denoising and ICA independent component analysis.

[0034] Amplitude threshold denoising removes amplitude abnormality artifacts, resulting in an initial artifact-free EEG signal matrix. .

[0035] The normal amplitude range of EEG signals is ±100μV. Signals exceeding this range can be identified as artifacts (such as amplitude abrupt changes caused by blinking or muscle twitching). Therefore, amplitude threshold denoising for single-channel signal processing in the denoised EEG signal matrix satisfies the following relationship: ; In the formula, This represents the signal amplitude after removing amplitude anomaly artifacts in a single channel; This represents the signal amplitude after single-channel denoising. Indicates the amplitude threshold. (Empirical values ​​set based on the physiological characteristics of EEG signals).

[0036] ICA Independent Component Analysis (removal of electrooculography / electromyography artifacts) identifies and removes corresponding independent components by visualizing their waveforms. The remaining effective independent components of the EEG signal are then reconstructed using a hybrid matrix to obtain the final artifact-free EEG signal matrix. This provides a high-quality signal foundation for subsequent random cropping operations.

[0037] Electrooculogram (EOG) and electromyogram (EMG) artifacts are independent sources and can be separated and removed using ICA decomposition. The EEG signal matrix after amplitude thresholding... The ICA decomposition is expressed mathematically as follows: ; In the formula, This represents the initial artifact removal EEG signal matrix; This represents the mixing matrix of a single channel, i.e. , representing the mixing weights of signals from different independent sources; This represents the independent component matrix of a single channel, i.e. Each row corresponds to an independent source signal (including valid EEG signals, electrooculography artifacts, electromyography artifacts, etc.).

[0038] Independent component analysis (ICA) reconstructs a single-channel mixture matrix that preserves the effective independent components of the EEG signal matrix, satisfying the following relationship: ; In the formula, This represents the final artifact removal signal amplitude for a single channel; This represents the mixing matrix of a single channel, i.e. , representing the mixing weights of signals from different independent sources; This represents the independent component matrix of a single channel, i.e. Each row corresponds to an independent source signal (including effective EEG signal, electrooculography artifact, electromyography artifact, etc.). This represents the matrix of remaining independent components after removing artifacts from a single channel.

[0039] Raw long-term EEG signals (general) Since the EEG data of cognitive impairment cannot be directly input into a deep learning model, and the sample size is small, a random cropping method is used to crop the original EEG signal after artifact removal. This achieves both signal length standardization and sample expansion. After random cropping, segmentation and preliminary sample expansion have been achieved, but sample diversity still needs to be improved to further address the small sample size problem and avoid model overfitting. Therefore, in step 1 above, the signal enhancement processing involves sequentially performing signal cropping and sample expansion processing, Gaussian noise enhancement processing, time axis scaling processing, amplitude scaling processing, and data equalization processing on the interference-removed preprocessed EEG signal.

[0040] Specifically, cropping and sample augmentation are performed on the original EEG signals after artifact removal. ( Based solely on this, and according to a fixed cutting length... (Consistent with the input length of the deep learning model, corresponding to a 10.24s EEG signal, which can completely capture the rhythmic features related to cognitive impairment), the artifact-free EEG signal is randomly cropped without overlap. The starting index for each cropping is randomly generated to ensure that the cropping intervals do not overlap and do not exceed the range of the original EEG signal. A single original long-term signal can be cropped to generate multiple standard-length samples, which not only replaces the traditional signal segmentation function, but also achieves the purpose of sample expansion for time-axis cropping.

[0041] The initial augmented sample set obtained after pruning satisfies the following relationship: ; Furthermore, the amplitude of the single-channel signal of the cropped sample satisfies the following relationship: ; In the formula, The signal matrix representing the cropped samples; Indicates a fixed cutting length. To ensure a fixed cropping length (consistent with the input sample length of the deep learning model), the cropped samples can be directly used for model training. Indicates the index of the cropped sample; This represents the total number of samples after cropping, and its value is 3 to 5 times the base sample size. Indicates the first The first cutting sample in the th The first channel, the first The signal amplitude at each sampling point; This represents the final artifact removal signal amplitude for a single channel; Indicates the first A random starting index for each cropped sample, satisfying... , And the interval between any two cropped samples is There is no overlap, ensuring the independence and integrity of each cropped sample. All indices are within the range of the original artifact-free EEG signal (20,000 points).

[0042] The initial augmented sample set is obtained after cropping. This provides a foundation for subsequent enhancement methods, data balancing, and format standardization.

[0043] Gaussian noise enhancement is used to simulate slight noise interference (such as electrode contact noise and environmental interference) in the actual EEG signal acquisition process, in order to improve the robustness of the subsequent deep learning model to the real EEG signal acquisition scenario, while avoiding excessive noise that could damage the pathological features of EEG.

[0044] Let a single cropped sample be (i.e., the 1024-point sample obtained from the above cropping and sample augmentation), after Gaussian noise enhancement processing, an enhanced sample EEG signal matrix with injected small-amplitude Gaussian noise is obtained. In order to preserve the core pathological rhythms of the original electroencephalogram (EEG) signals.

[0045] Gaussian noise enhancement processing, applied to single-channel signals in the cropped sample EEG signal matrix, satisfies the following relationship: ; In the formula, This represents a single-channel cropped sample injected with small-amplitude Gaussian noise, preserving the core pathological rhythms of the original EEG signal. This represents the signal amplitude of the cropped sample; This represents small-amplitude Gaussian noise; among which, Values The mean is 0 and the variance is 0. , , For cropping samples The standard deviation is used to ensure that the noise amplitude is 1% of the original signal, which meets the "small amplitude" requirement and does not damage the pathological characteristics.

[0046] Time axis scaling is achieved by changing the time axis length of the cropped sample (stretching or compressing) and then restoring it to the original fixed cropping length through interpolation. Without altering the medical and pathological characteristics of EEG signals, but only changing the temporal rhythm, it can significantly expand data diversity and simulate the rhythmic differences in EEG signals among different individuals. A scaling factor can be set. Among them, when Corresponding to the original cropped sample timeline (without scaling), when Corresponding to the original cropped sample timeline compression (time acceleration), when The original cropped sample was stretched along the timeline (time sequence slowed down), scaled, and then restored to its original length using linear interpolation. This ensures consistency with the model input dimension. After time-axis scaling, the enhanced sample EEG signal matrix is ​​obtained. The core medical pathological features (such as rhythm amplitude and fluctuation pattern) remain unchanged, and the data is greatly expanded.

[0047] Time-axis scaling processing, and processing of single-channel signals in the cropped EEG signal matrix after Gaussian noise enhancement, satisfy the following relationship: ; In the formula, This represents a single-channel cropped sample after time-axis scaling. This indicates that linear interpolation is used, and the original fixed cutting length is restored through interpolation. This ensures that the dimensions of the augmented samples are consistent with those of the original cropped samples, so as not to affect the input of subsequent deep learning models; This indicates that the time axis of the cropped sample is resampled according to the scaling factor, for example... At that time, the number of time points after resampling was ; This indicates the scaling factor that is set.

[0048] The amplitude of the cropped samples is slightly scaled without destroying the waveform characteristics and pathological rhythms of the signal. This simulates amplitude differences in EEG signals among different individuals (e.g., amplitude differences due to varying acquisition sensitivity or individual physiological variations), further expanding sample diversity. An amplitude scaling factor is set. A point-by-point amplitude scaling method was used to ensure that the scaling amplitude was controlled within ±10%, which is a small adjustment and avoids damaging the waveform characteristics and pathological information of the EEG signal. After amplitude scaling processing, the enhanced sample EEG signal matrix was obtained. .

[0049] Amplitude scaling, which involves processing single-channel signals from the cropped EEG signal matrix after time-axis scaling, satisfies the following relationship: ; In the formula, This represents a single-channel cropped sample after amplitude scaling. This represents the set amplitude scaling factor, which can be randomly selected within the set range during the experiment. Select the corresponding value within to ensure scaling diversity; This represents the amplitude of the c-th channel and the t-th sampling point of the cropped sample.

[0050] The base augmented sample size obtained after random cropping is Based on this, one enhanced sample is generated by sequentially performing the Gaussian noise enhancement process, the time axis scaling process, and the amplitude scaling process described above. Combined with the original cropped sample, the enhancement factor is... (Original cropped samples + 3 augmented samples), the final training sample set is defined as: ; In the formula, This represents the enhanced training sample set; This indicates the total number of samples after enhancement, which can efficiently solve the problem of small samples and ensure the integrity of the pathological features of the enhanced samples; The number of channels representing the brainwave signal; This indicates a fixed cutting length.

[0051] The enhanced training sample set still suffers from class imbalance; for example, the number of samples in MCI (mild cognitive impairment) is the smallest, while the number of samples in NC (normal controls) is the largest. If directly used for deep learning model training, this would cause the deep learning model to favor the majority class, resulting in a low recognition rate for the minority class and failing to meet the needs of clinical auxiliary diagnosis. Data balancing involves class-balancing the enhanced training sample set according to the four categories of cognitive impairment to ensure consistent learning weights for each category in subsequent deep learning models. Let the number of samples in the enhanced training sample set for AD (Alzheimer's disease), MCI (mild cognitive impairment), NC (normal controls), and OD (other cognitive impairments) be as follows: ; Usually present The imbalanced distribution is defined as follows: the maximum number of samples in each class is... The goal of data balancing is to make the number of samples in each category approximately equal. .

[0052] Data balancing involves a combination of oversampling and undersampling, employing weighted random sampling. The core of weighted random sampling is assigning different sampling probabilities to each class: oversampling with replacement (increasing the number of samples) is performed on the minority class, while undersampling without replacement (reducing the number of samples) is performed on the majority class, ensuring a balanced sample size across classes after sampling. Therefore, data balancing involves sampling the augmented training set according to the sampling probabilities, ultimately resulting in a balanced dataset defined as follows: ; In the formula, This represents the balanced dataset; 4 indicates the number of categories in the cognitive impairment classification; Indicates the number of samples in each category. And satisfy , For the first time after sampling The number of samples in each class is adjusted to achieve a balanced distribution of samples across all classes, preventing the model from being biased towards the majority class. The number of channels representing the brainwave signal; This indicates a fixed cutting length.

[0053] In data equalization, the sampling probability is inversely proportional to the number of samples in each category, satisfying the following relationship: ; In the formula, Indicates the sampling probability; The category numbering indicates the classification of cognitive impairment. ; Indicates the number of samples taken. The number of samples in each class; Minority class sample size The smaller the value, the higher the sampling probability. The larger the number of samples, the more samples are taken (oversampling); the number of majority class samples The larger the sampling probability, the higher the sampling probability. The smaller the value, the fewer the number of samples (undersampling).

[0054] The balanced dataset needs to undergo format standardization to align dimensions and remove dimensional interference, thereby accelerating the training and convergence of the subsequent deep learning model and improving training stability. Therefore, in step 1 above, the format standardization process involves aligning dimensions and removing dimensional interference from the balanced dataset to accelerate the training and convergence of the subsequent deep learning model and improve training stability.

[0055] Specifically, the format standardization process includes dimensional organization and channel-by-channel Z-score standardization.

[0056] The dimensionality adjustment process involves merging the category and sample dimensions of the balanced dataset and adjusting it to the input format required by the deep learning model. It satisfies the following relationship: ; In the formula, This represents the dataset after dimension adjustment; This represents the balanced dataset; This indicates the batch size, which can be adjusted according to training needs to ensure that the sample dimensions input to the model perfectly match the model input layer during each training session. This indicates a fixed cut length, which is the number of sampling points per channel.

[0057] Because different channels have varying acquisition sensitivities, the signal amplitudes of different EEG channels differ in dimension. Directly inputting these signals into a deep learning model can lead to the model overemphasizing channels with larger amplitudes, negatively impacting training performance. Channel-wise Z-score normalization normalizes the signal of each channel to a distribution with a mean of 0 and a variance of 1, removing dimensional interference and ensuring consistent feature weights across channels, thus yielding a format-normalized EEG signal. Its satisfaction , .

[0058] The channel-by-channel Z-score normalization process satisfies the following relationship: ; In the formula, This represents the amplitude of a standardized single-channel signal. This represents the amplitude of a single-channel signal after dimension adjustment. Indicates the first The average signal value of each channel; For the first The signal standard deviation of each channel; Indicates a minute value. This is used to avoid cases where the denominator is 0 (when the channel signal is constant, the standard deviation is 0, so adding...). This ensures that standardized operations are performed correctly.

[0059] No. The average signal values ​​of each channel satisfy the following relationship: ; In the formula, Indicates a fixed cutting length. =1024 sampling points; Represents the t-th sampling point; c represents the first One channel.

[0060] No. The standard deviation of the signal in each channel satisfies the following relationship: .

[0061] The preprocessed and standardized EEG data requires feature extraction and analysis using a neural network computational model. Combining the physiological characteristics of EEG signals in cognitive impairment, a deep learning model was constructed that integrates a temporal feature extraction architecture, a dual-attention processing architecture, and a frequency domain feature extraction architecture, incorporating a channel attention mechanism. This model uses temporal waveform fluctuations to reflect neural activity rhythms, frequency domain features to represent pathological differences, and different channels to correspond to pathological associations in different brain regions, thereby improving the accuracy of the final analysis and classification, especially in the feature differentiation between MCI and NC subtypes.

[0062] In step 2 above, the deep learning model has an input layer. The core function of this input layer is to process the received EEG signals after they have been standardized into input signals and perform signal splitting processing, so as to provide adaptive input for the subsequent dual-branch feature extraction of the time-domain CNN branch and the frequency-domain feature branch.

[0063] The input layer processes the input signal in accordance with the following relationship: ; In the formula, Indicates the input signal; Represents the standardized EEG signal matrix; This indicates the batch size, which should be consistent with the batch size in the format standardization to ensure consistent input dimensions. This indicates the number of EEG signal channels, corresponding to the number of brain regions channels in the original acquisition. This indicates a fixed cut length, which is the number of sampling points per channel.

[0064] The signal splitting process of this input layer is to split the input signal... Using a time-domain input approach, the feature extraction is directly divided into two paths according to the time-domain and frequency-domain feature extraction architectures of the deep learning model. These paths are respectively input into the subsequent time-domain convolutional neural network (CNN) branch (i.e., the time-domain CNN branch) and the frequency-domain feature branch. This achieves parallel feature extraction with two branches sharing the same input source, ensuring consistency and correlation in feature extraction. Path 1 (Time-domain branch input): The temporal CNN branch is used to extract temporal local waveform features of EEG signals; Path 2 (Frequency Domain Branch Input): The frequency domain feature branch is used to extract the frequency domain rhythm features of EEG signals.

[0065] In step 2 above, the deep learning model has a temporal feature extraction architecture, which is used for the input signal. The convolution, max pooling, and flattening processes are performed sequentially. The core function of this temporal CNN branch is to extract temporal local features (such as waveform peaks, waveform valleys, and fluctuation slopes) from EEG signals. These features directly reflect the abnormal neural activity (such as rhythmic fluctuation disorders) in patients with cognitive impairment. The branch adopts a lightweight "convolution-pooling-flattening" structure, which is suitable for small sample training scenarios and avoids parameter redundancy.

[0066] Specifically, convolution is the initial extraction process of temporal features, which modulates the input signal along the time axis. Convolution operations are performed to capture local waveform features in the temporal domain, introducing nonlinearity to enhance feature representation capabilities, satisfying the following relationship (1D convolution, along the time axis). (convolution) ; And the output dimension matrix is ; In the formula, Represents the convolutional feature map; This represents the activation function. By introducing nonlinearity, gradient vanishing is suppressed, positive features are preserved, and the amplitude characteristics of EEG signals are consistent. Indicates the number of output channels of the convolution. This is used to expand the feature dimensions and enhance the feature expressive power; This indicates the length of the convolution kernel along the time axis. It is compatible with a timing length of 1024 points and can effectively capture the fluctuation characteristics of local waveforms. Indicates the first One sampling point; Represents the time-domain input signal; The weight matrix represents the convolution kernel and can be trained to capture local features from different channels and time windows. This represents the bias vector of the convolutional layer, a trainable parameter used to adjust the baseline of the feature output; Indicates the temporal length after convolution. .

[0067] Max pooling is used to reduce the dimensionality of temporal features and preserve key features. Its core purpose is to reduce the dimensionality of the feature map output by the convolutional layer to reduce model parameters, suppress overfitting, and preserve key peak and valley features of the temporal waveform (these features are highly correlated with pathological abnormalities in cognitive impairment), satisfying the following relationship (max pooling, stride = 2, pooling kernel size). ): ; And the output dimension matrix is ; In the formula, Represents the max-pooling feature map; This represents the time series index after pooling, corresponding to the time series segment after pooling; This indicates the pooling kernel size, and the stride is consistent with the pooling kernel size to ensure non-overlapping pooling and avoid feature redundancy; This indicates the time series length after pooling, with the dimensionality halved, effectively reducing dimensionality. .

[0068] The core advantage of max pooling is that it can preserve the maximum value within each pooling window, that is, the key peak features of the time domain waveform, which matches the abnormal waveform characteristics of EEG signals in cognitive impairment (such as abnormally high or low peak values).

[0069] In addition, the feature map of max pooling Simultaneously, it serves as the input feature map for the subsequent dual-channel attention architecture, providing a temporal feature foundation for attention weight learning.

[0070] Flattening is a temporal feature vector transformation that flattens the 2D feature map (channel × temporal) output of the max pooling layer into a 1D feature vector to adapt to the input requirements of subsequent feature fusion and fully connected layers, satisfying the following relationship (flattening operation): ; And the output dimension matrix is ; In the formula, This represents the flattened temporal feature vector; This indicates that the format has been flattened. Represents the max-pooling feature map; Indicates the first The first batch of samples The time-domain feature values ​​provide time-domain feature input for subsequent feature fusion; Represents the dimension of the time-domain feature vector. .

[0071] In step 2 above, the deep learning model employs a dual-attention architecture, which is one of the core architectures of the deep learning model. Located after the max-pooling layer of the temporal feature extraction architecture and before the feature interaction fusion layer, it serves as a crucial transitional architecture between the temporal feature extraction architecture and multi-feature fusion. It only affects the max-pooling feature map output by the temporal feature extraction architecture and does not directly interfere with the feature extraction process of the frequency domain feature extraction architecture. In other words, the dual-attention architecture sequentially performs channel attention and temporal attention processing on the pooled feature map output by the temporal feature extraction architecture to adaptively allocate feature weights, focus on key pathological features (pathologically relevant channels and time-series segments), and suppress irrelevant feature interference. The dual-attention architecture executes in a hierarchical order of "channel attention → temporal attention," first optimizing channel weights and then optimizing temporal weights. The input is the max-pooling feature map of the temporal CNN branch. The output is the attention-enhanced temporal signature.

[0072] Specifically, the core objective of channel attention processing is to adaptively allocate channel weights. Given the multi-channel nature of EEG signals, the weights of each channel are adaptively allocated, focusing on EEG signal channels related to pathological features of cognitive impairment (such as channels with high incidence of pathological abnormalities in the frontal and temporal lobes), suppressing the influence of irrelevant channels (such as interfering channels), and improving the targeting of feature extraction.

[0073] Channel attention processing involves multiplying the channel weights by the input feature map channel by channel to obtain a channel-attention-enhanced feature map, satisfying the following relationship: ; And the output dimension matrix is , and input feature map Consistent dimensions ensure input adaptation for subsequent temporal attention layers; In the formula, This represents the feature map after channel attention enhancement; This represents the channel attention weight matrix; This represents the max-pooling feature map.

[0074] The channel attention weight matrix is ​​learned through a two-layer fully connected network (FC+ReLU+FC+Sigmoid, where FC is a fully connected network, ReLU is a function, and Sigmoid is a function), achieving adaptive allocation of channel weights and satisfying the following relationship: ; In the formula, Indicates the first Batch samples, the first The weight value of each channel is determined by its weight; the larger the weight, the greater the contribution of that channel to the pathological features. This represents the activation function, which normalizes the channel weights to the [0,1] interval, representing the importance of each channel; ; This represents the weight matrix of the second fully connected layer, which are trainable parameters that map the compressed features back to... The number of channels is the same as the original number of channels; Indicates the activation function; This represents the weight matrix of the first fully connected layer, which are trainable parameters. 3D channel features compressed to Dimensionality reduces parameter redundancy; Intermediate feature map representing channel attention; This represents the bias vector of the first fully connected layer, which are trainable parameters. This represents the bias vector of the second fully connected layer, which is a trainable parameter.

[0075] The intermediate feature maps of channel attention satisfy the following relationship: ; In the formula, Indicates the first Batch samples, the first The global average pooling value of each convolutional output channel is used to aggregate the temporal global features of that channel and capture the overall information at the channel level. This indicates global average pooling; Represents the max-pooling feature map; Indicates the time sequence length after pooling; This represents the time-series index after pooling.

[0076] The core objective of temporal attention processing is to adaptively allocate the weights of temporal segments. Building upon channel attention optimization, it further focuses on pathological feature segments related to time (such as temporal segments with abnormal rhythms in the EEG signals of patients with cognitive impairment), suppresses interference from irrelevant temporal segments, enhances the specificity of temporal features, and complements channel attention.

[0077] Temporal attention processing involves multiplying the temporal weights by the channel-attention-enhanced feature map sequentially to obtain the dual-channel attention-enhanced temporal feature map, satisfying the following relationship: ; And the output dimension matrix is The dimension is consistent with the input feature map; In the formula, This represents the temporal feature map after dual-channel attention enhancement; Represents the temporal attention weight matrix; This represents the feature map after channel attention enhancement.

[0078] The temporal attention weight matrix employs a two-layer fully connected network, consistent with channel attention, to learn temporal weights and achieve adaptive weight allocation for temporal segments, satisfying the following relationship: ; In the formula, Indicates the first Batch samples, the first The weight value of each time series segment is determined by the weight of the segment. The larger the weight, the richer the pathological features contained in the time series segment. This represents the activation function, which normalizes the temporal weights to the [0,1] interval to ensure the interpretability of the weights; This represents the weight matrix of the second fully connected layer, which are trainable parameters that map the compressed features back to... The dimension is consistent with the original time series length; Indicates the activation function; This represents the weight matrix of the first fully connected layer, which are trainable parameters. Time-series features compressed to Dimensionality reduces parameter redundancy; Intermediate feature map representing temporal attention; This represents the bias vector of the first fully connected layer, which are trainable parameters. This represents the bias vector of the second fully connected layer, which is a trainable parameter.

[0079] Among them, the intermediate feature map of temporal attention is the feature map enhanced by channel attention. Cross-channel global average pooling is performed to aggregate the cross-channel features of each time segment, satisfying the following relationship: ; In the formula, Indicates the first Batch samples, the first The cross-channel global average pooling value of each time series segment is used to aggregate the cross-channel information of that time series segment; This indicates that global average pooling is being performed. This represents the feature map after channel attention enhancement; This indicates the number of output channels for the convolution.

[0080] The dual-attention architecture enhances the temporal feature map with dual-channel attention to adapt to subsequent feature fusion. After flattening, the attention-enhanced temporal feature vectors are obtained, satisfying the following relationship: ; In the formula, Indicates attention-enhanced temporal features; This indicates that the format has been flattened. This represents the temporal feature map after dual-channel attention enhancement; Represents the dimension of the time-domain feature vector. The feature dimensions are completely consistent with those of the flattened branches of the temporal CNN, ensuring dimensionality adaptation for subsequent feature fusion.

[0081] In step 2 above, the deep learning model has a frequency domain feature extraction architecture. The core function of this architecture is to extract frequency domain rhythmic features (i.e., delta rhythm energy, theta rhythm energy, alpha rhythm energy, beta rhythm energy, power spectral density, etc.) from the EEG signal. These features are core pathological indicators of cognitive impairment (e.g., decreased alpha rhythm energy and increased delta rhythm energy in AD patients). The frequency domain feature extraction architecture adopts a hierarchical structure of "frequency domain transformation - power spectral density calculation - rhythm extraction - dimensionality reduction". That is, the frequency domain feature extraction architecture sequentially performs FFT transformation, PSD processing, rhythm energy extraction processing, and fully connected dimensionality reduction processing on the input signal to extract frequency domain rhythmic features from the EEG signal, including rhythm energy and power spectral density.

[0082] Specifically, the core purpose of FFT transform is to convert the time domain to the frequency domain. FFT transform transforms the input signal in the time domain... The signal is converted to the frequency domain to capture its frequency distribution characteristics, laying the foundation for subsequent power spectral density calculations, and satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First The frequency domain complex signal at the frequency point; It represents the Discrete Fourier Transform, consistent with the definition of the Fourier Transform in bandpass filtering, ensuring mathematical and logical coherence. Represents the time-domain input signal; Represents the field of complex numbers; Represents the frequency point, with a value range of 100. Hz (matching the sampling frequency of 100Hz, Nyquist frequency is 50Hz), number of frequency points .

[0083] The core purpose of PSD processing is to convert a frequency domain complex signal into a real-valued form of power spectral density (PSD), quantifying the signal energy distribution at different frequencies. PSD values ​​are directly related to rhythmic abnormalities in cognitive impairment (e.g., abnormal increases or decreases in the energy of pathological rhythms), satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First Power spectral density value at a frequency point; This indicates a fixed clipping length, which is also the number of sampling points per channel. This is used to normalize the power spectral density, ensuring that the PSDs of signals of different lengths are comparable; It represents the square of the modulus of a complex signal in the frequency domain, converts the complex number into a non-negative real number, and characterizes the energy intensity of the signal at that frequency point.

[0084] The core purpose of rhythm energy extraction processing is to screen pathology-related rhythms, identify frequency bands (0.5Hz–45Hz, consistent with the bands processed by the bandpass filtering above) related to cognitive impairment, and extract the energy features of four core rhythms within this band—namely, delta rhythm energy features, theta rhythm energy features, alpha rhythm energy features, and beta rhythm energy features. Irrelevant frequency components are eliminated, focusing on pathological features, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First The energy value of rhythmic patterns; Indicates the frequency point; Indicates the first The frequency index set corresponding to the rhythm type only contains frequency points in the range of 0.5Hz to 45Hz; These correspond to four types of rhythms: δ, θ, α, and β.

[0085] Among them, the four core rhythm energy characteristics (closely related to cognitive impairment) and their frequency ranges are as follows: The frequency range of the delta rhythm energy characteristics is 0.5 Hz to 4 Hz (elevated energy is often observed in patients with cognitive impairment). The frequency range of the theta rhythm energy characteristics is 4Hz to 8Hz (related to memory dysfunction); The frequency range of alpha rhythm energy characteristics is 8 Hz to 13 Hz (energy reduction is often observed in AD patients); The frequency range of β-rhythm energy characteristics is 13Hz to 45Hz (related to cognitive arousal state).

[0086] To adapt to subsequent fully connected dimensionality reduction, the rhythm energy extraction process flattens the obtained rhythm energy feature matrix. Flattening transforms the 3D energy feature matrix into a 1D feature vector, satisfying the following relationship: ; In the formula, This represents the flattened frequency domain feature vector; This indicates that the format has been flattened. Represents the rhythmic energy characteristic matrix; express A set of vectors consisting of n real numbers.

[0087] The core purpose of fully connected dimensionality reduction is frequency domain feature optimization. It reduces the dimensionality of the flattened rhythmic energy features, removes redundant information, and introduces nonlinearity to enhance the expressive power of frequency domain features, adapting them to subsequent feature fusion, and satisfying the following relationship: ; And the output dimension matrix is This provides frequency domain feature input for subsequent feature fusion; In the formula, Represents a complex signal in the frequency domain; This represents the activation function, introducing nonlinearity to maintain consistency with the activation logic of the temporal feature extraction architecture; represents the weight matrix of the fully connected layer, which are trainable parameters used to map M-dimensional energy features to the target dimension; This represents the bias vector of the fully connected layer, which are trainable parameters; Indicates the frequency domain feature output dimension. This allows for integration and adaptation with subsequent time-domain features.

[0088] In step 3 above, the feature interaction fusion processing involves processing the temporal local waveform features output by the temporal feature extraction architecture. Dual-attention enhanced temporal features output by dual-channel attention architecture and the frequency domain features output by the frequency domain feature extraction architecture Interactive fusion is performed to achieve complementarity of multi-dimensional features, enhance the expressive power of features, and provide more comprehensive pathological feature support for subsequent classification. This feature interactive fusion processing adopts a "splicing + feature calibration" fusion strategy to ensure feature complementarity and no redundancy. Specifically, the frequency domain rhythm features are first adjusted in dimension to match the dimensions of the time domain local waveform features. Then, the time domain local waveform features, attention-enhanced time domain features, and frequency domain rhythm features are spliced ​​together, and finally, feature calibration is performed through a fully connected layer.

[0089] Specifically, for the dimensionality adjustment of frequency domain rhythm features, a fully connected layer is used to implement dimensionality mapping, satisfying the following relationship: ; In the formula, This indicates the frequency domain features after dimension adjustment, which have the same dimensions as the time domain features. Indicates the activation function; The dimension-adjustable weight matrix of the fully connected layer represents the trainable parameters that map 1024-dimensional frequency domain features to 31744 dimensions. This represents the dimension-adjustable bias vector of the fully connected layer, which is a trainable parameter.

[0090] The initial fused features are obtained by concatenating the time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythmic features, satisfying the following relationship: ; And the output dimension matrix is ; In the formula, Indicates the initial fusion features; Represents local waveform characteristics in the time domain; Indicates temporal features after attention enhancement; This represents the frequency domain features after dimension adjustment; It contains complete information including three types of features: time domain, attention-enhanced time domain, and frequency domain. .

[0091] Feature calibration is performed through fully connected layers to remove redundant information and optimize the expressive power of the fused features, resulting in the final fused features that satisfy the following relationship: ; In the formula, This represents the final fused feature vector, which contains multi-dimensional pathological features and provides input for subsequent regularization and classification outputs. Indicates the activation function; This represents the weight matrix of the fully connected layer for feature fusion calibration, which are trainable parameters. Indicates the final fused feature dimension. It balances feature representation capability with model training efficiency; It contains complete information including three types of features: time domain, attention-enhanced time domain, and frequency domain. ; This represents the bias vector of the fully connected layer for feature fusion calibration, which is a trainable parameter.

[0092] For small-sample training scenarios, an adaptive regularization mechanism is introduced to suppress overfitting of the deep learning model, improve its generalization ability and training stability, and adaptively adjust the regularization strength to avoid feature loss caused by over-regularization. Therefore, in step 4 above, the adaptive regularization process satisfies the following relationship: ; And the output dimension matrix is The feature dimensions have been further optimized, taking into account both generalization ability and feature representation, and providing adaptive input for the classification output layer; In the formula, Features after adaptive regularization; Indicates the activation function; Indicates the weights of the fully connected layer; This represents the final fused feature vector; Indicates the bias of the fully connected layer; This represents the feature dimension after regularization. ; This represents an additional term in the adaptive regularization loss function.

[0093] The additional term in the aforementioned adaptive regularization loss function is assumed to be a fusion feature. The output after the fully connected layer is Then the following relationship is satisfied: ; In the formula, This represents the adaptive regularization strength, and its value is dynamically adjusted with the number of training iterations. The dimension index representing the fused feature, from 1 to ; This represents the total number of dimensions of the fused features; The square of the fusion feature of the b-th batch and the d-th dimension is the core expression of the L2 regular expression.

[0094] The adaptive regularization strength satisfies the following relationship: ; In the formula, Indicates the initial regularization strength. Based on small sample training experience; This represents the natural constant, approximately equal to 2.71828; Indicates the attenuation coefficient. This controls the decay rate of the regularization strength; This indicates the number of training iterations. As the number of iterations increases, the regularization strength gradually decreases to avoid over-regularization in later training stages.

[0095] The core purpose of predictive classification is to map the adaptively regularized features into the probability space of four types of cognitive impairment (AD, MCI, NC, OD), and to perform classification prediction according to decision rules, outputting the predicted probabilities of each category to meet the needs of clinical auxiliary diagnosis. Therefore, in step 5 above, the decision rule for predictive classification is to select the category with the highest predicted probability as the final classification result of the sample, satisfying the following relationship: ; In the formula, Indicates the first Predicted category labels for batch samples; Indicates the classification category of cognitive impairment; Represents the first in the classification prediction probability matrix Batch of samples belongs to the first The predicted probability of cognitive impairment is represented by the classification prediction probability matrix as follows: .

[0096] Among them, the Batch of samples belongs to the first The predicted probability of cognitive impairment, based on a fully connected layer and a softmax activation function, satisfies the following relationship: ; In the formula, Softmax This represents an activation function that maps features to the [0,1] interval, ensuring that the sum of the predicted probabilities of each category is 1, thus providing probabilistic interpretability. ; Represents the classification output layer weight matrix, which contains trainable parameters, where Number of classification categories (AD, MCI, NC, OD); Features after adaptive regularization; This represents the bias vector of the classification output layer, which can be trained parameters.

[0097] The above specific technical solutions are only used to illustrate the present invention, and are not intended to limit it.

[0098] Although the present invention has been described in detail with reference to the specific technical solutions described above, those skilled in the art should understand that modifications can still be made to the specific technical solutions described above, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the present invention.

Claims

1. A cognitive impairment electroencephalogram signal analysis system based on multi-dimensional feature fusion, characterized in that, The cognitive impairment EEG signal analysis system includes: The raw EEG data input module is used to input the acquired raw multi-channel resting-state EEG signals, which are presented as two-dimensional time-series signals. The bandpass filter module is used to remove signal frequencies with noise interference from the EEG signal and retain signal frequencies that can contain delta rhythm features, theta rhythm features, alpha rhythm features, and beta rhythm features related to cognitive impairment. The artifact removal module is used to remove signal artifacts, including amplitude anomalies and rhythm disorders. The signal cropping and sample augmentation module uses the original EEG signal after artifact removal as a basis and performs non-overlapping random cropping according to a fixed cropping length. The Gaussian noise enhancement module is used to simulate slight noise interference during the actual acquisition of EEG signals, thereby improving the robustness of deep learning models to real EEG signal acquisition scenarios. The time axis scaling module is used to expand data diversity by simulating the rhythm differences of EEG signals from different individuals; The amplitude scaling module is used to simulate the amplitude differences of EEG signals from different individuals to increase the diversity of sample data. The data balancing module is used to perform class balancing on the augmented training sample set according to the classification of cognitive impairment. The format standardization module is used to perform dimension alignment and remove dimensional interference from the balanced dataset. The temporal feature extraction module is used to sequentially perform convolution, max pooling, and flattening on the input format-normalized signal to extract temporal local waveform features from the EEG signal, including waveform peaks, waveform valleys, and fluctuation slopes. The dual attention module is used to perform channel attention and temporal attention processing on the pooled feature map output by the temporal feature extraction module, so as to adaptively allocate feature weights, focus on key pathological features, and output attention-enhanced temporal features. The frequency domain feature extraction module is used to sequentially perform FFT transformation, PSD processing, rhythm energy extraction processing, and fully connected dimensionality reduction processing on the input format-normalized signal in order to extract frequency domain rhythm features, including rhythm energy and power spectral density, from the EEG signal. The feature interaction fusion module is used to stitch and calibrate time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythm features to achieve complementary fusion of multi-dimensional features; The adaptive regularization module is used to adaptively regularize the fused features to suppress overfitting. The classification output module is used to map the regularized features to the probability space of four types of cognitive impairment, and to perform classification prediction and output the predicted probability of each category according to the decision rules.

2. A cognitive impairment electroencephalogram signal analysis method based on multi-dimensional feature fusion, characterized in that, The cognitive impairment EEG signal analysis method is used to implement the cognitive impairment EEG signal analysis system based on multi-dimensional feature fusion as described in claim 1. The method for analyzing EEG signals in cognitive impairment includes the following steps: Step 1. The acquired raw EEG signals are sequentially processed through interference removal preprocessing, signal enhancement processing, and format standardization processing; Step 2. The standardized EEG signals are processed by time-domain feature extraction and frequency-domain feature extraction using a deep learning model. The temporal feature extraction process is used to obtain the temporal local waveform features of the EEG signal. The pooled feature map obtained from the temporal feature extraction process is then processed by a dual attention mechanism to obtain attention-enhanced temporal features. The frequency domain rhythmic features of the EEG signal are obtained through the frequency domain feature extraction process. Step 3. Perform feature interaction fusion processing on the time-domain local waveform features, attention-enhanced time-domain features, and frequency-domain rhythm features to achieve multi-dimensional feature complementarity; Step 4. Apply adaptive regularization to the fused features to suppress overfitting; Step 5. Map the regularized features to different categories of cognitive impairment and implement the predictive classification of cognitive impairment according to the decision rules. 3.The cognitive impairment electroencephalogram signal analysis method based on multi-dimensional feature fusion of claim 2, characterized in that, In step 1, the interference removal preprocessing involves sequentially performing bandpass filtering and artifact removal processing on the original EEG signal; The band-pass filtering processing is to filter the original electroencephalogram signal by using an FIR filter, to retain frequency components in the range of 0.5Hz-45Hz, and to obtain an electroencephalogram signal matrix after denoising ; The bandpass filtering process processes single-channel signals in the original EEG signal matrix, satisfying the following relationship: ； In the formula, denotes the signal amplitude after single-channel denoising; denotes an inverse Fourier transform; Indicates Fourier transform; Indicates the first The first channel, the first The amplitude of the electroencephalogram signal at each sampling point; Let represent the frequency response function of an ideal bandpass filter, satisfying the following relationship: ； The artifact removal process is achieved by combining amplitude threshold denoising with ICA independent component analysis. The amplitude threshold denoising removes amplitude abnormality artifacts, resulting in an initial artifact-removed EEG signal matrix. ; The amplitude threshold denoising process, which processes single-channel signals in the denoised EEG signal matrix, satisfies the following relationship: ； In the formula, This represents the signal amplitude after removing amplitude anomaly artifacts in a single channel; This represents the signal amplitude after single-channel denoising. Indicates the amplitude threshold; The ICA (Independent Component Analysis) method involves visualizing the waveforms of independent components, identifying and removing corresponding independent components, and then reconstructing the remaining effective independent components of the EEG signal using a hybrid matrix to obtain the final artifact-free EEG signal matrix. ; The ICA independent component analysis, for single-channel mixture matrix reconstruction that preserves the effective independent components in the EEG signal matrix, satisfies the following relationship: ； In the formula, This represents the final artifact removal signal amplitude for a single channel; Represents the mixing matrix for a single channel; Represents the independent component matrix of a single channel; This represents the matrix of remaining independent components after removing artifacts from a single channel.

4. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2, characterized in that, In step 1, the signal enhancement processing involves sequentially performing signal cropping and sample augmentation, Gaussian noise enhancement, time axis scaling, amplitude scaling, and data equalization on the interference-free preprocessed EEG signal. The cropping and sample augmentation process involves performing non-overlapping random cropping of the artifact-removed EEG signals according to a fixed cropping length. The initial augmented sample set obtained after cropping satisfies the following relationship: ； Furthermore, the amplitude of the single-channel signal of the cropped sample satisfies the following relationship: ； In the formula, The signal matrix representing the cropped samples; Indicates a fixed cutting length; Indicates the index of the cropped sample; This represents the total number of samples after cropping. Indicates the first The first cutting sample in the th The first channel, the first The signal amplitude at each sampling point; This represents the final artifact removal signal amplitude for a single channel; Indicates the first A random starting index for each cropped sample, satisfying... , And the interval between any two cropped samples is No overlap; The Gaussian noise enhancement process simulates the slight noise interference in the actual EEG signal acquisition process, so as to improve the robustness of the subsequent deep learning model to the real EEG signal acquisition scenario, while avoiding excessive noise that could damage the pathological features of EEG. After the Gaussian noise enhancement process, an enhanced sample EEG signal matrix is ​​obtained after injecting small-amplitude Gaussian noise. In order to preserve the core pathological rhythm of the original electroencephalogram signal; The Gaussian noise enhancement process, which processes single-channel signals in the cropped sample EEG signal matrix, satisfies the following relationship: ； In the formula, This represents a single-channel cropped sample after injecting small-amplitude Gaussian noise. This represents the signal amplitude of the cropped sample; This indicates small-amplitude Gaussian noise; The time axis scaling process involves changing the time axis length of the cropped sample and then restoring it to the original fixed cropping length through interpolation. To expand sample diversity and simulate the rhythm differences of EEG signals among different individuals; After the time axis scaling process, the enhanced sample EEG signal matrix is ​​obtained after time axis scaling. ; The time axis scaling process, which processes single-channel signals in the cropped EEG signal matrix after Gaussian noise enhancement, satisfies the following relationship: ； In the formula, This represents a single-channel cropped sample after time-axis scaling. This indicates that linear interpolation is used to restore the resampled signal to its original length. ; This indicates that the time axis of the cropped sample is resampled according to the scaling factor; This indicates the scaling factor set. The amplitude scaling process involves slightly scaling the amplitude of the cropped sample to simulate the amplitude differences of EEG signals from different individuals and to expand the sample diversity. After the amplitude scaling process, the amplitude-scaled enhanced sample EEG signal matrix is ​​obtained. ; The amplitude scaling process, which processes single-channel signals in the cropped EEG signal matrix after time-axis scaling, satisfies the following relationship: ； In the formula, This represents a single-channel cropped sample after amplitude scaling. This indicates the set amplitude scaling factor; This represents the amplitude of the c-th channel and the t-th sampling point of the cropped sample; In the initial expanded sample set Based on this, the training sample set is defined as follows after sequentially performing Gaussian noise enhancement processing, time axis scaling processing, and amplitude scaling processing: ； In the formula, This represents the enhanced training sample set; This represents the total number of samples after augmentation; The number of channels representing the brainwave signal; Indicates a fixed cutting length; The data balancing process is performed by weighted random sampling to balance the enhanced training sample set according to the cognitive impairment classification, so as to ensure the consistency of the learning weights of each category of samples by the subsequent deep learning model. The data equalization process involves sampling the enhanced training sample set according to the sampling probability, and the final equalized dataset is defined as follows: ； In the formula, This represents the balanced dataset; 4 indicates the number of categories in the cognitive impairment classification; Indicates the number of samples in each category; The number of channels representing the brainwave signal; Indicates a fixed cutting length; The sampling probability in the data equalization process satisfies the following relationship: ； In the formula, Indicates the sampling probability; The category number indicating the classification of cognitive impairment; Indicates the number of samples taken. The number of samples in a class.

5. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2, characterized in that, In step 1, the format standardization process involves dimensional alignment and removal of dimensional interference from the balanced dataset to accelerate the convergence of the subsequent deep learning model training and improve training stability. The format standardization process includes a dimension organization process and a channel-by-channel Z-score standardization process. The dimensionality adjustment process involves merging the category and sample dimensions of the balanced dataset and adjusting it to the input format required by the deep learning model. It satisfies the following relationship: ； In the formula, This represents the dataset after dimension adjustment; This represents the balanced dataset; Indicates batch size; Indicates the number of EEG signal channels; Indicates a fixed cutting length; The channel-by-channel Z-score normalization process normalizes the signal of each channel to a distribution with a mean of 0 and a variance of 1, removes dimensional interference, and ensures that the weights of the features in each channel are consistent, thereby obtaining the format-normalized EEG signal. Its satisfaction , ; The channel-by-channel Z-score normalization process satisfies the following relationship: ； In the formula, represents The standardized single-channel signal amplitude is represented. This represents the amplitude of a single-channel signal after dimension adjustment. Indicates the first The average signal value of each channel; For the first The signal standard deviation of each channel; ； The first The average signal values ​​of each channel satisfy the following relationship: ； In the formula, Indicates a fixed cutting length; Indicates the first Sampling points; c represents the first One channel; The first The standard deviation of the signal in each channel satisfies the following relationship: 。 6. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2, characterized in that, In step 2, the deep learning model has an input layer, which performs input signal processing and signal de-segmentation processing on the received standardized EEG signals. The input signal processing satisfies the following relationship: ； In the formula, Indicates the input signal; Represents the standardized EEG signal matrix; Indicates batch size; Indicates the number of EEG signal channels; Indicates a fixed cutting length; The signal splitting process involves directly dividing the input signal into two paths in the time domain according to the time domain feature extraction architecture and frequency domain feature extraction architecture of the deep learning model, thereby realizing dual-branch parallel feature extraction where the two signals share the same input source.

7. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2 or 6, characterized in that, In step 2, the deep learning model has a temporal feature extraction architecture, which sequentially performs convolution, max pooling, and flattening on the input signal to extract temporal local waveform features from the EEG signal, including waveform peaks, waveform valleys, and fluctuation slopes. The convolution process involves performing a convolution operation on the input signal along the time axis to capture local waveform features in the time domain, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, Represents the convolutional feature map; Indicates the activation function; Indicates the number of output channels of the convolution; This indicates the length of the convolution kernel along the time axis; This represents the t-th sampling point; Represents the time-domain input signal; The weight matrix represents the convolution kernel; This represents the bias vector of the convolutional layer; Indicates the temporal length after convolution; The max pooling process reduces the dimensionality of the feature map output by the convolutional layer while preserving the key peak and valley features of the temporal waveform, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, Represents the max-pooling feature map; This represents the time-series index after pooling; This indicates the pooling kernel size, and the step size is the same as the pooling kernel size. Indicates the time sequence length after pooling; The flattening process involves flattening the 2D feature map output by the max pooling layer into a 1D feature vector to adapt to the input requirements of subsequent feature fusion and fully connected layers, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, This represents the flattened temporal feature vector; This indicates that the format has been flattened. Represents the max-pooling feature map; Indicates the first The first batch of samples 3D time-domain eigenvalues; This represents the dimension of the temporal feature vector.

8. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2 or 6, characterized in that, In step 2, the deep learning model has a dual attention architecture, which sequentially performs channel attention processing and temporal attention processing on the pooled feature map output by the temporal feature extraction architecture to adaptively allocate feature weights and focus on key pathological features. The channel attention processing involves multiplying the channel weights by the input feature map channel by channel to obtain a channel-attention-enhanced feature map, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, This represents the feature map after channel attention enhancement; This represents the channel attention weight matrix; Represents the max-pooling feature map; The channel attention weight matrix satisfies the following relationship: ； In the formula, Indicates the first Batch samples, the first The weight values ​​of each channel; Indicates the activation function; This represents the weight matrix of the second fully connected layer; Indicates the activation function; This represents the weight matrix of the first fully connected layer; Intermediate feature map representing channel attention; This represents the bias vector of the first fully connected layer; This represents the bias vector of the second fully connected layer; The intermediate feature maps of the channel attention satisfy the following relationship: ； In the formula, Indicates the first Batch samples, the first The global average pooling value of each convolution output channel; This indicates global average pooling; Represents the max-pooling feature map; Indicates the time sequence length after pooling; This represents the time-series index after pooling; The temporal attention processing involves multiplying the temporal weights by the channel attention-enhanced feature map sequentially to obtain the dual-channel attention-enhanced temporal feature map, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, This represents the temporal feature map after dual-channel attention enhancement; Represents the temporal attention weight matrix; This represents the feature map after channel attention enhancement; The temporal attention weight matrix satisfies the following relationship: ； In the formula, Indicates the first Batch samples, the first The weight values ​​of each time segment; Indicates the activation function; This represents the weight matrix of the second fully connected layer; Indicates the activation function; This represents the weight matrix of the first fully connected layer; Intermediate feature map representing temporal attention; This represents the bias vector of the first fully connected layer; This represents the bias vector of the second fully connected layer; The intermediate feature maps of the temporal attention satisfy the following relationship: ； In the formula, Indicates the first Batch samples, the first Cross-channel global average pooling value for each time segment; This indicates that global average pooling is being performed. This represents the feature map after channel attention enhancement; Indicates the number of output channels of the convolution; The dual-attention architecture, to adapt to subsequent feature fusion, flattens the temporal feature map after dual-channel attention enhancement, resulting in attention-enhanced temporal feature vectors that satisfy the following relationship: ； In the formula, Indicates attention-enhanced temporal features; This indicates that the format has been flattened. This represents the temporal feature map after dual-channel attention enhancement; This represents the dimension of the temporal feature vector.

9. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2 or 6, characterized in that, In step 2, the deep learning model has a frequency domain feature extraction architecture, which sequentially performs FFT transformation, PSD processing, rhythm energy extraction processing, and fully connected dimensionality reduction processing on the input signal to extract frequency domain rhythm features, including rhythm energy and power spectral density, from the EEG signal. The FFT transformation process converts the time-domain input signal into a frequency-domain signal, capturing the frequency distribution characteristics of the signal, which satisfy the following relationship: ； And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First The frequency domain complex signal at the frequency point; Represents the Discrete Fourier Transform; Represents the time-domain input signal; The PSD processing converts the frequency domain complex signal into a real-valued form of power spectral density and quantizes the signal energy distribution at different frequencies, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First Power spectral density value at a frequency point; Indicates a fixed cutting length; Represents the square of the modulus of a complex signal in the frequency domain; The rhythm energy extraction process involves screening out frequency bands related to cognitive impairment, extracting the energy characteristics of delta rhythms, theta rhythms, alpha rhythms, and beta rhythms within these bands, eliminating irrelevant frequency components, focusing on pathological features, and satisfying the following relationship: ； And the output dimension matrix is ; In the formula, Indicates the first Batch, No. Channel, First The energy value of rhythmic patterns; Indicates the frequency point; Indicates the first A set of frequency indices corresponding to rhythm types; These correspond to four types of rhythms: δ, θ, α, and β, respectively. The rhythm energy extraction process is adapted to subsequent fully connected dimensionality reduction by flattening the obtained rhythm energy feature matrix, transforming the 3D energy feature matrix into a 1D feature vector, satisfying the following relationship: ； In the formula, This represents the flattened frequency domain feature vector; This indicates that the format has been flattened. Represents the rhythmic energy characteristic matrix; express A set of vectors consisting of n real numbers; The fully connected dimensionality reduction process reduces the dimensionality of the flattened rhythmic energy features to adapt them for subsequent feature fusion, satisfying the following relationship: ； And the output dimension matrix is ; In the formula, Represents a complex signal in the frequency domain; Indicates the activation function; This represents the weight matrix of the fully connected layer; This represents the bias vector of the fully connected layer; This indicates the output dimension of the frequency domain features.

10. The method for analyzing EEG signals of cognitive impairment based on multi-dimensional feature fusion according to claim 2, characterized in that, In step 3, the feature interaction fusion process first adjusts the dimension of the frequency domain rhythm features to make them consistent with the dimension of the time domain local waveform features, then concatenates the time domain local waveform features, attention-enhanced time domain features, and frequency domain rhythm features, and then performs feature calibration through a fully connected layer. The dimension adjustment of frequency domain rhythm features is achieved using a fully connected layer, satisfying the following relationship: ； In the formula, This represents the frequency domain features after dimension adjustment; Indicates the activation function; This indicates the dimension-adjusted weight matrix of the fully connected layer. This indicates the dimension adjustment of the fully connected layer bias vector; The splicing process of local waveform features in the time domain, attention-enhanced time domain features, and frequency domain rhythmic features satisfies the following relationship: ； And the output dimension matrix is ; In the formula, Indicates the initial fusion features; Represents local waveform characteristics in the time domain; Indicates temporal features after attention enhancement; This represents the frequency domain features after dimension adjustment; It contains complete information encompassing three categories of features: temporal domain, attention-enhanced temporal domain, and frequency domain. The feature calibration process satisfies the following relationship: ； In the formula, This represents the final fused feature vector; Indicates the activation function; This indicates the weight matrix of the fully connected layer for feature fusion calibration; Indicates the final fused feature dimension; It contains complete information encompassing three categories of features: temporal domain, attention-enhanced temporal domain, and frequency domain. This indicates the bias vector of the fully connected layer calibrated by fusion features; In step 4, the adaptive regularization process satisfies the following relationship: ； And the output dimension matrix is ; In the formula, Features after adaptive regularization; Indicates the activation function; Indicates the weights of the fully connected layer; This represents the final fused feature vector; Indicates the bias of the fully connected layer; This represents the feature dimension after regularization. This represents an additional term in the adaptive regularization loss function; In step 5, the decision rule for predicting classification is to select the category with the highest predicted probability as the final classification result of the sample, satisfying the following relationship: ； In the formula, Indicates the first Predicted category labels for batch samples; Indicates the classification category of cognitive impairment; Represents the first in the classification prediction probability matrix Batch of samples belongs to the first The predicted probability of cognitive impairment is represented by the classification prediction probability matrix as follows: ; The first Batch of samples belongs to the first The predicted probability of cognitive impairment satisfies the following relationship: ； In the formula, Softmax Indicates the activation function; This represents the weight matrix of the classification output layer; Features after adaptive regularization; This represents the bias vector of the classification output layer.

Images