Double-path feature fusion Alzheimer's disease classification method based on windowing EEG (electroencephalogram)

Abstract

The invention provides a double-path feature fusion Alzheimer's disease classification method based on windowing EEG, and relates to the technical field of data processing. The method comprises the following steps: acquiring an original electroencephalogram signal, and performing interference elimination preprocessing on the original electroencephalogram signal to obtain a pure electroencephalogram signal; performing non-overlapping cutting on the pure electroencephalogram signal according to a fixed window length to obtain electroencephalogram signal slices; carrying out feature extraction on the electroencephalogram signal slices by utilizing a two-way feature extraction network to respectively obtain a time domain feature vector and a frequency domain feature vector; splicing the time domain feature vector and the frequency domain feature vector to obtain a time-frequency fusion depth feature vector; performing binary classification operation on the time-frequency fusion depth feature vector to obtain a single-slice classification probability; and carrying out weighted average statistical fusion on the single-slice classification probability to obtain an overall classification probability. According to the method, the combined pain point of single feature, cumbersome algorithm, simple fusion and difficult deployment in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a dual-path feature fusion method for Alzheimer's disease classification based on windowed EEG. Background Technology

[0002] Early risk identification for Alzheimer's disease (AD) is crucial for slowing disease progression. Electroencephalography (EEG) signals, being non-invasive and capable of reflecting abnormalities in brain electrical activity, have become an important technological carrier for AD risk classification, with increasingly urgent application needs in primary healthcare and geriatric checkups. While AD diagnostic technologies and supporting equipment based on EEG signals have developed, several core issues remain: First, feature extraction is often limited to a single time or frequency domain. Manually created features or single-domain deep features cannot fully represent the complex spatiotemporal-frequency dynamics of EEG signals, leading to insufficient classification accuracy and stability. Second, classification models have a large number of parameters, high computational complexity, and high hardware computing power requirements, making efficient deployment on resource-constrained primary healthcare equipment or portable screening terminals difficult. Third, when fusing the classification probabilities of multiple EEG signal slices after windowing, existing methods typically use equal-weighted simple averaging or majority voting, failing to perform statistical fusion based on the specific classification probabilities of each slice. In cases of noise interference or significant individual differences, the robustness of the overall classification still needs improvement. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a dual-path feature fusion Alzheimer's disease classification method based on windowed EEG, which solves the combined pain points of existing technologies such as "single feature + cumbersome algorithm + simple fusion + difficult deployment", and provides a more practical and reliable technical means for early risk identification and disease progression.

[0004] To achieve the above objectives, the present invention provides the following solution: A dual-path feature fusion Alzheimer's disease classification method based on windowed EEG includes: The raw EEG signal is acquired and subjected to interference removal preprocessing to obtain a clean EEG signal. The pure EEG signal is cut into non-overlapping segments according to a fixed window length to obtain several independent EEG signal slices. The EEG signal slices are subjected to feature extraction using a dual-path feature extraction network to obtain time-domain feature vectors and frequency-domain feature vectors, respectively. The time-domain feature vectors are extracted by a one-dimensional convolutional neural network. The frequency-domain feature vectors are obtained by calculating the power spectral density to extract the frequency-domain feature matrix of a preset number of bands, and then performing dimensionality reduction mining using a multilayer perceptron network. The time-domain feature vector and the frequency-domain feature vector are concatenated to obtain a time-frequency fusion deep feature vector; A binary classification operation is performed on the time-frequency fusion deep feature vector to obtain the single-slice classification probability; The overall classification probability is obtained by performing a weighted average statistical fusion of the single-slice classification probabilities of all the EEG signal slices.

[0005] The present invention discloses the following technical effects: This invention provides a dual-path feature fusion method for Alzheimer's disease classification based on windowed EEG. This invention addresses the technical pain points in AD screening, such as single feature extraction, reliance on cloud / specialized hardware, and poor adaptability to primary care settings. It achieves multi-dimensional technical breakthroughs through time-domain-frequency domain dual-path feature mining, lightweight algorithm design, and edge computing hardware and software collaboration. Compared with existing technologies, it has significant advantages in terms of computational efficiency, device adaptability, and clinical applicability. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 A flowchart of an Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG is provided for an embodiment of the present invention. Figure 2 This is a schematic diagram of the technical framework provided for an embodiment of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0010] like Figure 1 As shown, this invention provides a dual-path feature fusion Alzheimer's disease classification method based on windowed EEG, comprising: Step 100: Acquire the raw EEG signal and perform interference removal preprocessing on the raw EEG signal to obtain a clean EEG signal; Step 200: The pure EEG signal is cut into non-overlapping segments according to a fixed window length to obtain several independent EEG signal slices; Step 300: Use a dual-path feature extraction network to extract features from the EEG signal slices to obtain time-domain feature vectors and frequency-domain feature vectors respectively; wherein, the time-domain feature vectors are extracted by a one-dimensional convolutional neural network; the frequency-domain feature vectors are obtained by calculating the power spectral density to extract the frequency-domain feature matrix of a preset number of bands, and then by dimensionality reduction mining through a multilayer perceptron network; Step 400: Concatenate the time-domain feature vector with the frequency-domain feature vector to obtain a time-frequency fusion deep feature vector; Step 500: Perform a binary classification operation on the time-frequency fusion deep feature vector to obtain the single-slice classification probability; Step 600: Perform a weighted average statistical fusion of the single-slice classification probabilities of all the EEG signal slices to obtain the overall classification probability.

[0011] Specifically, such as Figure 2 As shown, the technical framework mainly consists of three main processes: EEG signal acquisition, preprocessing, and standardized windowing: Complete the acquisition, interference removal, and discretization slicing of EEG signals to provide high-quality, standardized input data for feature extraction; A dual-path feature extraction and time-frequency deep fusion network architecture was designed: a dual-path feature extraction module in the time domain and frequency domain was constructed to accurately mine the temporal features and band power spectrum features of EEG signals, and to complete the deep fusion of the dual-path features to maximize the preservation of the differences in Alzheimer's disease features; Slice-level classification and local-to-global statistical fusion: First, the classification operation of single-slice EEG features is realized, and then the results of all slices are integrated through statistical fusion strategy to realize the classification from local EEG features to the whole patient.

[0012] Furthermore, the step of acquiring the raw EEG signal and performing interference removal preprocessing on the raw EEG signal to obtain a clean EEG signal includes: Collect the raw electroencephalogram (EEG) signals from the subject; Based on notch filtering and independent component analysis algorithms, the original EEG signal is preprocessed to remove interference, resulting in the pure EEG signal.

[0013] Specifically, the implementing entities are: EEG acquisition equipment and preprocessing terminal; Signal Acquisition: Raw EEG signals were acquired from subjects in a dark, closed-eye state using an EEG cap. The sampling frequency was set to Fs = 500Hz. The acquisition channels included Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2 (a total of 19 channels). The sampling duration was controlled at 10-20 minutes. The raw EEG signal is denoted as X, and the raw EEG signal is represented as: X∈ 19×N (N is the total number of sampling points); (1) Signal preprocessing: 50Hz power frequency interference was removed by notch filtering, and independent component analysis (ICA) algorithm was used to remove electrooculography artifacts to obtain a pure EEG signal without interference; (2) This ensures the validity of the original signal, eliminates the influence of external interference on subsequent feature extraction, and provides a high-quality data foundation for windowing processing.

[0014] Furthermore, the purified EEG signal is non-overlappingly sliced ​​according to a fixed window length to obtain several independent EEG signal slices, including: Determine the fixed window length used to divide the time series; Based on the fixed window length, the clean EEG signal is subjected to non-overlapping segmentation; The continuous time-domain signal is discretized by the non-overlapping cut to obtain the several independent EEG signal slices.

[0015] Specifically, the executing entity is the edge computing device. Continuous, clean EEG signals are processed according to a fixed window length. Perform non-overlapping segmentation, denoted as w for a certain window, and N as above. The number of sampling points contained in each window is... This results in several independent EEG signal slices (19 channels x 2500 sampling points), where x is the same as above. The dimension of each slice is represented as follows: ; This discretizes continuous time-domain signals, preserving the EEG feature information of a single slice while also achieving data augmentation to improve the generalization ability of subsequent models. At the same time, fixing the slice dimensions provides a unified input standard for feature extraction.

[0016] Furthermore, the temporal feature vector is extracted using a one-dimensional convolutional neural network, including: The EEG signal slices are input into the 1D convolutional neural network, wherein the 1D convolutional neural network comprises; The first convolutional feature is obtained by performing a convolution operation with a kernel size of 7 on the EEG signal slice through the first convolutional layer. After batch normalization of the first convolutional feature, activation processing is performed based on the StarReLU activation function to obtain the first activated feature; Max pooling is used to reduce the dimensionality of the first activation feature to obtain the first pooled feature. The first pooling feature is input into the second convolutional layer to obtain the second convolutional feature; The batch normalization process is sequentially applied to the second convolutional feature, and the second pooling feature is obtained based on the activation process of the StarReLU activation function and the max pooling dimensionality reduction process. The second pooling feature is input into the third convolutional layer to obtain the third convolutional feature; The batch normalization process is sequentially applied to the third convolutional feature, and the third pooling feature is obtained based on the activation process of the StarReLU activation function and the max pooling dimensionality reduction process. In this embodiment, the third pooling feature is flattened into a 1-dimensional vector, and the 1-dimensional vector is regularized. The 1-dimensional vector that has undergone the regularization process is transformed to obtain a temporal feature vector. The first convolutional feature has 64 channels, the second convolutional feature has 128 channels, and the third convolutional feature has 256 channels.

[0017] Furthermore, the frequency domain feature vector is obtained by calculating the power spectral density to extract the frequency domain feature matrix of a preset number of bands, and then performing dimensionality reduction mining using a multilayer perceptron network, including: The frequency domain transformation of the EEG signal slices was performed using the Welch method, and the power spectral density was calculated. The brainwave physiological characteristics are divided into 5 preset bands corresponding to the preset number of bands. The sum of the power spectral density of each preset band is extracted to obtain the frequency domain feature matrix of the preset number of bands. After normalizing the frequency domain feature matrix and flattening it into a one-dimensional feature vector, it is input into the multilayer perceptron network. The frequency domain feature vector is obtained by performing dimensionality reduction mining on the one-dimensional feature vector through the multilayer perceptron network.

[0018] Specifically, this step designs a dual-path approach: a time-domain 1D-CNN feature extraction module (TimeCNN) and a frequency-domain power spectrum + MLP feature extraction module (FreqMLP). These modules respectively mine the core time-domain and frequency-domain features of the EEG signal, as detailed below: The TimeCNN module is used to capture core features such as temporal amplitude changes, temporal correlations, and local textures of EEG slices. It consists of 3 convolutional layers, 3 batch normalization layers, pooling layers, StarReLU activation layers, Dropout layers, and linear layers. First convolution operation: using EEG slices with dimensions 19×2500 For input, use the kernel size. The formula for a 1D convolution with 64 output channels is as follows: ; Batch normalization and activation: After batch normalization of the convolution results, the StarReLU activation function is used to enhance the nonlinear expressive power. The StarReLU activation function is defined as follows: ; The formula for batch normalization + activation is: ; Max pooling: Max pooling with a kernel size of 2 and a stride of 2 is used to reduce the feature dimension. The formula is as follows: ; Second convolution operation: For input, use the kernel size. 1. Output a 1D convolution with 128 output channels, repeating the batch normalization, activation, and pooling steps, with the following formulas in sequence: ; ; ; The third layer of convolution operation: For input, use the kernel size. 1. Output a 1D convolution with 256 output channels, repeating batch normalization, activation, and pooling steps, with the following formulas in sequence: ; ; ; (1) Feature flattening and dimensionality reduction: Flattened into a one-dimensional vector, and after being regularized using Dropout (probability 0.5), the input linear layer is transformed into a 1×512 dimensional temporal feature vector. The formula is: ; The temporal branch design leverages the local feature perception capability of 1D-CNN, improves the comprehensiveness of feature extraction through multi-scale convolutional kernels (7 / 5 / 3), and unifies the feature dimension in the linear layer to prepare for subsequent concatenation.

[0019] Furthermore, the FreqMLP module is used to mine the 5-band frequency domain features of EEG signals, consisting of Welch power spectrum calculation, band feature extraction, and MLP dimensionality reduction. (1) Calculation of power spectral density: The Welch method was used to calculate the power spectral density of the EEG slices. Perform a frequency domain transformation to map the time-domain signal to a frequency-power spectral density distribution, as shown in the formula: ; in, For frequency axis, Power spectral density; (2) 5-band feature extraction: The brainwave physiological characteristics are divided into 5 preset bands, and the power spectral density of each band is extracted. The band division and calculation formulas are as follows: ; The final frequency domain feature matrix is ​​19 channels × 5 bands: ; (3) Perform Z-score normalization on the obtained frequency domain feature matrix: Normalize the frequency domain feature matrix P to eliminate the influence of dimensions. The formula is: ; in, The characteristic mean, The characteristic standard deviation; (4) MLP dimensionality reduction: the normalized feature matrix The vector is initially 1×95 dimensional. It is input into a two-layer MLP network to perform high-order feature mining and dimensionality reduction, ultimately resulting in a 1×512 dimensional frequency domain feature vector. The formula is: ; In the frequency domain feature extraction module, the Welch method is used to accurately extract frequency domain power spectrum features, and the 5-band division is specifically tailored to the variation patterns of each band of the EEG in AD patients. Then, MLP is used to achieve high-order mining and dimensionality reduction optimization of frequency domain features, and finally, a linear layer is used to unify the dimension to ensure that the features from both paths can be spliced.

[0020] Furthermore, the step of performing a binary classification operation on the time-frequency fusion deep feature vector to obtain the single-slice classification probability includes: The binary classification operation is performed on the time-frequency fusion deep features; Based on the binary classification operation, dimensionality transformation is performed to obtain log probability features; The logarithmic probability features are converted into Alzheimer's disease classification probabilities using the Softmax function to obtain the single-slice classification probabilities.

[0021] Furthermore, the weighted average statistical fusion of the single-slice classification probabilities of all the EEG signal slices to obtain the overall classification probability includes: Obtain all the single-slice classification probabilities corresponding to all the said EEG signal slices; The weighted average of all the single-slice classification probabilities is calculated to obtain the weighted average statistical fusion. The overall classification probability is obtained by the weighted average statistical fusion.

[0022] Specifically, after time-frequency dual-path feature extraction, the time-domain feature vector output by the TimeCNN module is... (1×512 dimensions) and the frequency domain feature vector output by the FreqMLP module (1×512 dimensions) are concatenated along dimension 1 to obtain the time-frequency fusion deep feature vector. The formula is: ; We obtain a 1×1024 dimensional time-frequency fusion feature vector: ; This integrates the time-domain and five-band frequency-domain dual-depth features of EEG signals, making up for the information loss of single-domain features or single-band features, and maximizing the preservation of Alzheimer's disease feature differences in EEG signals.

[0023] Furthermore, the first linear layer transforms the 1×1024 dimensional fused features into 1×512 dimensional features, and adds StarReLU activation and Dropout (probability 0.5) regularization. The second linear layer: transforms the 1×512 dimensional features into 2-dimensional log-probability logits, using the following formula: ; Classification probability calculation: The log probability is converted into AD classification probability using the Softmax function. The formula is as follows: ; in, The probability of identifying this EEG slice as an AD patient; Weighted average calculation of overall probability: AD classification probability for all slices , The weighted average is calculated using the following formula: .

[0024] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0025] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dual-path feature fusion Alzheimer's disease classification method based on windowed EEG, characterized in that, include: The raw EEG signal is acquired and subjected to interference removal preprocessing to obtain a clean EEG signal. The pure EEG signal is cut into non-overlapping segments according to a fixed window length to obtain several independent EEG signal slices. The EEG signal slices are subjected to feature extraction using a dual-path feature extraction network to obtain time-domain feature vectors and frequency-domain feature vectors, respectively. The time-domain feature vectors are extracted by a one-dimensional convolutional neural network. The frequency-domain feature vectors are obtained by calculating the power spectral density to extract the frequency-domain feature matrix of a preset number of bands, and then performing dimensionality reduction mining using a multilayer perceptron network. The time-domain feature vector and the frequency-domain feature vector are concatenated to obtain a time-frequency fusion deep feature vector; A binary classification operation is performed on the time-frequency fusion deep feature vector to obtain the single-slice classification probability; The overall classification probability is obtained by performing a weighted average statistical fusion of the single-slice classification probabilities of all the EEG signal slices.

2. The Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG as described in claim 1, characterized in that, The process of acquiring raw EEG signals and performing interference removal preprocessing on the raw EEG signals to obtain clean EEG signals includes: Collect the raw electroencephalogram (EEG) signals from the subject; Based on notch filtering and independent component analysis algorithms, the original EEG signal is preprocessed to remove interference, resulting in the pure EEG signal.

3. The Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG as described in claim 1, characterized in that, The pure EEG signal is non-overlappingly sliced ​​according to a fixed window length to obtain several independent EEG signal slices, including: Determine the fixed window length used to divide the time series; Based on the fixed window length, the clean EEG signal is subjected to non-overlapping segmentation; The continuous time-domain signal is discretized by the non-overlapping cut to obtain the several independent EEG signal slices.

4. The Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG as described in claim 1, characterized in that, The temporal feature vector is extracted using a one-dimensional convolutional neural network, including: The EEG signal slices are input into the 1D convolutional neural network, wherein the 1D convolutional neural network comprises; The first convolutional feature is obtained by performing a convolution operation with a kernel size of 7 on the EEG signal slice through the first convolutional layer. After batch normalization of the first convolutional feature, activation processing is performed based on the StarReLU activation function to obtain the first activated feature; Max pooling is used to reduce the dimensionality of the first activation feature to obtain the first pooled feature. The first pooling feature is input into the second convolutional layer to obtain the second convolutional feature; The batch normalization process is sequentially applied to the second convolutional feature, and the second pooling feature is obtained based on the activation process of the StarReLU activation function and the max pooling dimensionality reduction process. The second pooling feature is input into the third convolutional layer to obtain the third convolutional feature; The batch normalization process is sequentially applied to the third convolutional feature, and the third pooling feature is obtained based on the activation process of the StarReLU activation function and the max pooling dimensionality reduction process. In this embodiment, the third pooling feature is flattened into a 1-dimensional vector, and the 1-dimensional vector is regularized. The 1-dimensional vector that has undergone the regularization process is transformed to obtain a temporal feature vector. The first convolutional feature has 64 channels, the second convolutional feature has 128 channels, and the third convolutional feature has 256 channels.

5. The Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG according to claim 1, characterized in that, The frequency domain feature vector is obtained by calculating the power spectral density to extract the frequency domain feature matrix of a preset number of bands, and then performing dimensionality reduction mining using a multilayer perceptron network, including: The frequency domain transformation of the EEG signal slices was performed using the Welch method, and the power spectral density was calculated. The brainwave physiological characteristics are divided into 5 preset bands corresponding to the preset number of bands. The sum of the power spectral density of each preset band is extracted to obtain the frequency domain feature matrix of the preset number of bands. After normalizing the frequency domain feature matrix and flattening it into a one-dimensional feature vector, it is input into the multilayer perceptron network. The frequency domain feature vector is obtained by performing dimensionality reduction mining on the one-dimensional feature vector through the multilayer perceptron network.

6. The Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG according to claim 1, characterized in that, The step of performing a binary classification operation on the time-frequency fusion deep feature vector to obtain the single-slice classification probability includes: The binary classification operation is performed on the time-frequency fusion deep features; Based on the binary classification operation, dimensionality transformation is performed to obtain log probability features; The logarithmic probability features are converted into Alzheimer's disease classification probabilities using the Softmax function to obtain the single-slice classification probabilities.

7. The Alzheimer's disease classification method based on dual-path feature fusion using windowed EEG according to claim 1, characterized in that, The weighted average statistical fusion of the classification probabilities of each individual slice of all the EEG signal slices to obtain the overall classification probability includes: Obtain all the single-slice classification probabilities corresponding to all the said EEG signal slices; The weighted average of all the single-slice classification probabilities is calculated to obtain the weighted average statistical fusion. The overall classification probability is obtained by the weighted average statistical fusion.

Images