A brain disease detection system based on multi-domain fusion features and deep learning

By integrating features from multiple domains and employing deep learning methods, a high-dimensional brain functional connectivity matrix is ​​constructed. This solves the problem of incomplete EEG signal feature extraction in traditional methods, enabling accurate identification and early diagnosis of various brain diseases and improving the diagnostic rate of mental illnesses.

CN117398111BActive Publication Date: 2026-11-10JILIN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311420933.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-30
Publication Date
2026-11-10
Estimated Expiration
2043-10-30

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively extract complex inter-channel communication information from EEG signals, and traditional methods fail to fully reflect the characteristics of EEG signals, resulting in insufficient accuracy in diagnosing brain diseases, especially in areas lacking medical resources where the diagnosis rate of mental illnesses is low.

Method used

By employing multi-domain feature fusion and deep learning methods, a high-dimensional brain functional connectivity matrix is ​​constructed through the combination of synchronous compressed wavelet transform, coherence function, and phase-locked value. A lightweight CNN model is then used for brain disease identification to extract the time-frequency and phase information of EEG signals and reduce noise interference.

Benefits of technology

It has enabled the accurate identification of various brain diseases, improved the early diagnosis rate of mental illnesses, and enhanced the diagnostic capabilities of EEG signals, especially in areas with insufficient medical resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117398111B_ABST
    Figure CN117398111B_ABST
Patent Text Reader

Abstract

This invention belongs to the field of EEG signal analysis technology, specifically involving a brain disease detection system based on multi-domain fusion features and deep learning. First, it combines synchronous compressed wavelet transform (MSWC) and coherence functions, compressing them in the time dimension to obtain time-averaged synchronous compressed wavelet coherence. Then, it combines the phase-locked value obtained from MSWC and wavelet cross-spectrum to propose a new marker, P-MSWC, and uses this marker to construct a high-dimensional functional connectivity matrix of the brain. This marker can comprehensively capture information from the original EEG signal in the time, frequency, and phase domains and has strong anti-interference ability against physiological noise, outperforming traditional functional connectivity markers. Finally, it proposes a lightweight CNN model that effectively utilizes the high-dimensional connectivity matrix of the brain, enabling more accurate and effective detection of various brain-related diseases.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal analysis technology, specifically relating to a brain disease detection system based on multi-domain fusion features and deep learning. Background Technology

[0002] Electroencephalography (EEG) is widely used in the clinical diagnosis and treatment of mental illnesses. It uses electrodes attached to the scalp to record electrical activity in the brain. EEG is inexpensive, painless, and non-invasive, easy to use, and can provide high temporal resolution brain signals.

[0003] Electroencephalography (EEG) can be used to detect and diagnose various brain disorders, including depression, ADHD, Alzheimer's disease, epilepsy, traumatic brain injury, and Parkinson's disease. Current research indicates that some brain disorders directly affect the brain's normal electrophysiological activity. For example, similar to epilepsy, even when not in a seizure phase (i.e., the patient has no symptoms during testing), the brain still generates many abnormal electrical activities. Accurately capturing these activities allows for early diagnosis and treatment of epilepsy. Similarly, in Alzheimer's and Parkinson's diseases, studies show that patients exhibit slower brainwave activity during sleep, with low-frequency waveforms dominating the time, accompanied by numerous abnormal activities. Besides detecting abnormal electrophysiological activity, EEG signals can also be used to detect the levels of electrical activity in different areas. For instance, patients with attention deficit hyperactivity disorder (ADHD) typically have lower levels of activity in the prefrontal cortex, which is responsible for attention and impulse control. EEG can also detect changes in brain activity after traumatic brain injury, helping doctors monitor the recovery of patients and determine the most appropriate treatment.

[0004] Furthermore, with increasing societal pressures, mental illnesses such as depression and anxiety have become common social problems, contributing significantly to the global disease burden. If these illnesses are not diagnosed and treated promptly in their early stages, they can worsen and lead to suicide. Despite this, in some underdeveloped countries and regions, due to a lack of medical resources and professional medical personnel, as well as societal stigma surrounding mental disorders, over 75% of patients do not receive timely and effective diagnosis and treatment. EEG represents neural signals in the cerebral cortex; extracting useful information from EEG signals and exploring the impact of different diseases on the cerebral cortex is a crucial issue. In countries with insufficient medical resources, EEG can be readily used to diagnose mental illnesses. EEG can be used to detect abnormal patterns of brain activity, enabling early detection and treatment of many diseases. In general, EEG is a powerful tool for detecting and diagnosing brain diseases.

[0005] In recent years, various studies have employed machine learning techniques and EEG signals to detect brain-related diseases. This primarily involves extracting time-domain, frequency-domain, and nonlinear features from single-channel EEG signals, combining them with traditional machine learning methods for detection and classification. Traditional single-channel features mainly include statistical characteristics, power spectrum, approximate entropy, and sample entropy. These features are primarily extracted from a single EEG channel and cannot represent the relationships between multiple channels. Traditional research focuses on a single channel, essentially a univariate approach. These methods may not consider the communication between different EEG signals; brain regions are not statically independent, but constantly interact dynamically for normal brain function. Increasing research indicates that brain connectivity can reveal the functions of different brain regions and the complex cortical communication between them, which helps in better studying brain diseases such as depression.

[0006] Currently, several features have emerged for measuring brain connectivity, primarily including those based on information theory (mutual information, transfer entropy, etc.), generalized synchronization (synchronization likelihood index, S-index, etc.), Granger causality (partially oriented coherence, oriented transfer function, etc.), coefficients (Pearson correlation coefficient, Spearman correlation coefficient, etc.), and phase synchronization (phase-locked value, phase lag index, etc.). However, these features can only analyze EEG signals from specific perspectives. As a typical non-stationary physiological signal, EEG signals require not only understanding their time and frequency domain information but also their frequency variation over time, instantaneous frequency, instantaneous amplitude, and phase.

[0007] Electroencephalogram (EEG) signals contain complex brain information, and features extracted solely from the time and frequency domains cannot fully reflect the characteristics of EEG signals. Extracting information from EEG signals from multiple perspectives and combining it with characteristics from different domains will be a new direction for the effective diagnosis of various brain diseases in the future. At the same time, EEG signals themselves carry a large amount of physiological noise, which is similar to the signal itself. Traditional noise removal methods may lose information from the signal itself, affecting the extracted features. Therefore, finding a feature that can measure the multi-channel connectivity of the brain from multiple domains and is robust to noise is crucial.

[0008] According to recent scientific research published in the journal *Science*, the foundation of many brain activities lies not in isolated areas of operation, but in their dynamic connections and interactions. Therefore, the brain is a complex system that integrates many parts to create a coherent biological whole. The autonomous activation of isolated brain regions cannot comprehend complex concepts. Neuroscience research increasingly demonstrates that numerous brain regions are interconnected in different ways, collaboratively constructing complex brain networks to achieve complex cognitive functions. Therefore, utilizing the functional connectivity networks between brain regions for the diagnosis of various diseases will be a new trend in the future.

[0009] Currently, achieving rapid and accurate identification of some mental illnesses using EEG signals remains a challenge. EEG signals, as a typical non-stationary signal, contain a large amount of noise and complex information. Therefore, how to efficiently extract useful information while avoiding noise interference is a worthwhile area of ​​research. Summary of the Invention

[0010] To overcome the above problems, this invention provides a brain disease detection method based on multi-domain fusion features and deep learning, which can accurately identify a variety of brain-related diseases.

[0011] A brain disease detection system based on multi-domain fusion features and deep learning includes a data acquisition and preprocessing module, a feature extraction module, and a brain disease identification module; wherein:

[0012] 1. Data Acquisition and Preprocessing Module: Acquire at least 20 minutes of resting-state EEG signals from participants; perform preliminary preprocessing on the acquired EEG signals; segment the preprocessed EEG signals into segments at fixed 60-second intervals, with each segment overlapping by 25%, and treat each segment as an independent sample. The EEG signal of each independent sample is as follows:

[0013]

[0014] in This represents the time variable of the EEG electroencephalogram (EEG) signal, specifically the time of EEG signal acquisition. This indicates that the time variable Discrete variables obtained by discretization. This represents a specific point in time after time has been discretized. This represents the total number of discrete points in the signal. Indicates the sequence number of the EEG channel. This represents the EEG signals of all channels in this independent sample;

[0015] 2. The feature extraction module uses analytical Morlet wavelets as wavelet basis functions to perform synchronous compressed wavelet transform on all channel signals of each independent sample, obtaining the synchronous compressed wavelet time-frequency matrix of each channel signal. ,in Indicates frequency;

[0016] Select any two channels from each independent sample, where the EEG signals of the two channels are denoted as follows: and Calculate the time-frequency wavelet autospectrum of each channel signal. and and the time-frequency wavelet cross spectrum of the two channel signals. ;

[0017] Based on the definition of coherence function, the two-channel signals are calculated using the obtained time-frequency wavelet autospectrum and time-frequency wavelet crossspectrum. and Synchronous compressed wavelet coherence between , will get Averaging over the time dimension yields the time-averaged synchronous compressed wavelet coherence. Let represent the average coherence of the two channel signals in the frequency dimension, where ;

[0018] Using time-frequency wavelet cross spectrum To calculate the phase difference between two channel signals And calculate the phase-locked value (PLV) between the two channel signals, where ;

[0019] The phase-locked value and the time-averaged synchronous compressed wavelet coherence are fused together, and then fused through a mapping function. Change the range from Mapped to Finally, a new fusion feature, P-MSWC, is obtained, in which... ;

[0020] Iterate through all channels of the signal in each independent sample and calculate the fusion feature between every two channels of the signal in each independent sample. , will integrate features The frequency dimension is expanded to obtain a series of connection matrices constructed from values ​​at each frequency. These connection matrices are combined to form a high-dimensional brain functional connection matrix.

[0021] 3. A lightweight CNN model is constructed in the brain disease identification module. This model integrates the Inception module and six convolutional layers. Some of these convolutional layers incorporate a CBAM attention module. The Inception module reduces the number of parameters while maintaining the same receptive field, replacing the traditional 5×5 convolutional kernel with two 3×3 kernels, and the 3×3 kernels with 1×3 and 3×1 kernels. The CBAM attention module integrates channel attention and spatial attention. This model can learn key information from a high-dimensional brain functional connectivity matrix to identify various brain diseases, achieving accurate detection. Details are as follows:

[0022] 3.1 Construction of a Lightweight CNN Model

[0023] The lightweight CNN model includes an Inception module, followed by six convolutional layers, a fully connected layer, and a Sigmoid layer.

[0024] The Inception module consists of multiple network branches, among which:

[0025] The first branch consists of 5 convolutional layers. The first convolutional layer has 64 kernels with a kernel size of 1×1. The second convolutional layer has 32 kernels with a kernel size of 1×3. The third convolutional layer has 32 kernels with a kernel size of 3×1. The fourth convolutional layer has 32 kernels with a kernel size of 1×3. The fifth convolutional layer has 32 kernels with a kernel size of 3×1.

[0026] The second branch includes 3 convolutional layers. The first convolutional layer has 64 convolutional kernels with a kernel size of 1×1. The second convolutional layer has 32 convolutional kernels with a kernel size of 1×3. The third convolutional layer has 32 convolutional kernels with a kernel size of 3×1.

[0027] The third branch includes one pooling layer and one convolutional layer. The pooling layer uses max pooling with a stride of 1, and the convolutional layer has 32 convolutional kernels, each of which is 1×1 in size.

[0028] The fourth branch includes one convolutional layer with 32 convolutional kernels, each kernel being 1×1 in size;

[0029] The six convolutional layers added after the Inception module are:

[0030] The first convolutional layer group consists of one convolutional layer, one batch normalization layer, and one tanh activation function layer. The convolutional layer has 64 convolutional kernels, and each convolutional kernel is 1×1 in size.

[0031] The second convolutional layer group consists of one convolutional layer, one batch normalization layer, and one Tanh activation function layer. The number of convolutional kernels in the convolutional layer is 16, and the size of each convolutional kernel is 1×1.

[0032] The third convolutional layer group consists of one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 32 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0033] The fourth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 64 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0034] The fifth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 128 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0035] The sixth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, and one dropout layer. The convolutional layer has 256 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0036] Finally, the output is generated through four fully connected layers and one Sigmoid layer. The Sigmoid layer outputs the probability of binary classification, with 0.5 as the dividing line. When the output value is greater than 0.5, the prediction result is defined as diseased, and when it is less than 0.5, the prediction result is defined as healthy.

[0037] The preprocessing content of the data acquisition and preprocessing module is as follows: resample the EEG signal to 256Hz, remove the data from the A1 and A2 electrodes, remove 50Hz power line noise using a notch filter, perform 0.5-70Hz bandpass filtering using an FIR filter, and remove eye movement artifacts using the ASR algorithm.

[0038] The specific contents of the feature extraction module are as follows:

[0039] 1.1 Calculate the synchronous compressed wavelet transform. The specific steps are as follows:

[0040] Step 1: Using the analytical Morlet wavelet as the wavelet basis function, perform a continuous wavelet transform (CWT) on each channel signal in each independent sample to obtain the wavelet coefficients defined in the time domain. :

[0041]

[0042] in: It is the signal of one channel in an independent sample. It is a scaling variable. It is a time-shifted variable. It is the conjugate of the wavelet basis functions. It is Translation in the time dimension The unit is then expanded on a scale to the original size. times, It is a time factor The derivative of represents A quantity that changes infinitely little;

[0043] Step 2: According to Plancherel's theorem, the wavelet coefficients of the single-channel signal defined in the time domain are... Transforming to the frequency domain yields wavelet coefficients defined in the frequency domain. Defined by the following expression:

[0044]

[0045] in It is an angular frequency variable. yes Fourier transform, It is a single-channel signal Fourier transform, It is the imaginary unit;

[0046] For any and If its corresponding wavelet coefficients are defined in the frequency domain The instantaneous angular frequency of a single-channel signal is expressed as:

[0047]

[0048] Step 3: Establish from The mapping transforms the wavelet coefficients from the time-scale domain to the time-angular frequency domain; therefore, the scaling variable... and angular frequency variables Convert to discrete values and ,in Represents a scaled discrete variable. Represents the discrete variable of angular frequency; Represents the interval of the scale variable. Represents the interval of angular frequency variables; their definitions are as follows: and By using arbitrary center frequency bands The intervals on both sides Compression is performed to obtain the synchronous compressed wavelet transform value of the channel signal. :

[0049]

[0050] Step 4: Adjust the time-shift variable Discretization is performed, and a synchronous compressed wavelet transform is established to convert angular frequency to scale and then to frequency to time, thus obtaining the time-frequency coefficient matrix of the channel signal. This is the synchronous compressed wavelet time-frequency matrix:

[0051]

[0052] 1.2 Calculation of Synchronous Compressed Wavelet Coherence

[0053] The amplitude-squared coherence of EEG signals is defined as follows:

[0054]

[0055] in, This represents the signal from two channels in an independent sample. and Related functions, It is a signal and Cross power spectral density, and These are signals and Each has its own power spectral density; combining the amplitude squared coherence function and the synchronous compressed wavelet transform, the synchronous compressed wavelet coherence is constructed, and the specific steps are as follows:

[0056] Step 1: Assume the signal and All lengths Calculate the synchronous compressed wavelet transforms of the two signals respectively to obtain their respective time-frequency coefficient matrices. and Two signals and The time-frequency wavelet cross spectrum is:

[0057]

[0058] Here Indicates complex conjugation;

[0059] Two signals and The time-frequency wavelet autospectral values ​​are as follows:

[0060]

[0061] Step 2: Define two signals and The synchronous compressed wavelet coherence value is:

[0062]

[0063] here It is a smooth function. The smoothing scaling factor in time and frequency;

[0064] 1.3 Calculate the phase-locked loop value

[0065] Through two signals and The phase-locked value (PLV) of two signals is calculated using the time-frequency wavelet cross spectrum. The specific steps are as follows:

[0066] Step 1: From the signal and Time-frequency wavelet cross spectrum To calculate the phase difference between two signals Specifically:

[0067]

[0068] in and They represent The imaginary and real parts, Expressed in radians and with a range of values ​​of ;

[0069] Step 2: Based on the phase difference The phase-locked value (PLV) of the two signals is calculated as follows:

[0070]

[0071] in This indicates that the average value is taken over time, and the range of the phase-locked value (PLV) is as follows: ;

[0072] 1.4 Calculate the fusion features. The specific implementation steps are as follows:

[0073] Step 1: Average the obtained SWC values ​​over the time dimension to obtain the time-averaged synchronous compressed wavelet coherence. Let represent the average coherence of the two signals in the frequency dimension:

[0074]

[0075] The scope of MSWC is as follows: ;

[0076] Step 2: Synchronize the phase-locked value (PLV) and time averaging to compress the wavelet coherence. The fusion process yields new fusion features, which are:

[0077]

[0078] in , Represents PLV, This represents a mapping function that ranges the interval from... Mapped to ; The definition is as follows:

[0079]

[0080] in, ;

[0081] 1.5 Constructing a high-dimensional brain functional connectivity matrix

[0082] Iterate through all channels of the signal in each independent sample and calculate the fusion feature between every two channels of the signal in each independent sample. , will integrate features Expanding the frequency dimension yields a connection matrix constructed from a series of values ​​at each frequency, thus... The frequency dimension is transformed into the depth dimension of the high-dimensional connection matrix; these connection matrices are combined to form a high-dimensional brain functional connection matrix.

[0083] The specific parameters of the four fully connected layers in the lightweight CNN model constructed in the brain disease recognition module are set as follows:

[0084] The first fully connected layer has 64 neurons and uses the Tanh activation function.

[0085] The second fully connected layer has 16 neurons and uses the Tanh activation function;

[0086] The third fully connected layer has 4 neurons and uses the Tanh activation function;

[0087] The fourth fully connected layer has only one neuron and uses the Tanh activation function.

[0088] The training process of the lightweight CNN model constructed in the brain disease recognition module specifically includes the following steps:

[0089] Step 1: Select more than 150 healthy individuals and more than 150 patients with encephalopathy as all participants, collect EEG brain signals from all participants, input the EEG brain signals of all participants into the data preprocessing module for preprocessing, and use all the preprocessed independent samples as the sample dataset. That is, label the sample dataset, set the data label value of healthy individuals to 0, and set the data label value of patients with encephalopathy to 1.

[0090] Step 2: Input each sample in the labeled sample dataset into the feature extraction module, and obtain the high-dimensional brain functional connectivity matrix after processing by the feature extraction module;

[0091] Step 3: Input all the extracted high-dimensional brain function connectivity matrices into the lightweight CNN model for training. The training process ends when the model's loss converges to less than 0.05.

[0092] The beneficial effects of this invention are:

[0093] This invention leverages the noise resistance of synchronous compressed wavelet transform (MSWC) by combining it with a coherence function and constructing an MSWC by averaging over the time dimension. This effectively avoids the influence of physiological noise on EEG signals while simultaneously measuring the connectivity between different brain regions. Furthermore, by utilizing the wavelet cross spectrum during the solution process, the phase difference between signals can be calculated, thereby determining the PLV between signals. By fusing PLV and MSWC into a new feature, a novel marker for measuring the functional connectivity of EEG signals, P-MSWC, is provided. This marker quantifies the connectivity of different EEG channels from three perspectives: time, frequency, and phase. Using P-MSWC to construct a functional connectivity matrix, and extending the frequency dimension, a high-dimensional functional connectivity matrix that accurately represents brain network connectivity is obtained. Finally, a lightweight CNN model is proposed to train this matrix, enabling accurate identification of various brain-related diseases. Attached Figure Description

[0094] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the content of the embodiments of the present invention and these drawings without creative effort.

[0095] Figure 1 This is a flowchart of the present invention.

[0096] Figure 2 This invention relates to EEG data acquisition diagrams and international 10-20 electrode placement standard diagrams.

[0097] Figure 3 It is the graph of the mapping function.

[0098] Figure 4 This is a schematic diagram of the constructed high-dimensional functional connection matrix.

[0099] Figure 5 This is a schematic diagram of the Inception module.

[0100] Figure 6 This is a diagram of the lightweight CNN model structure. Detailed Implementation

[0101] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0102] A brain disease detection system based on multi-domain fusion features and deep learning includes a data acquisition and preprocessing module, a feature extraction module, and a brain disease identification module; such as Figure 1 As shown, where:

[0103] 1. Data Acquisition and Preprocessing Module: Following pre-set experimental requirements and adhering to the international 10-20 electrode standard, at least 20 minutes of resting-state EEG signals were acquired from participants. Preliminary preprocessing of the acquired EEG signals included: resampling the EEG signals to 256Hz, removing data from electrodes A1 and A2, using a notch filter to remove 50Hz power line noise, using an FIR filter for 0.5-70Hz bandpass filtering, and using the ASR algorithm to remove eye movement artifacts. Specific details are as follows:

[0104] 1.1 Develop a data collection plan

[0105] One week prior to EEG signal collection, ensure the participant has abstained from caffeine, tobacco, alcohol, and other abusive substances, and that they have no other mental health conditions. EEG signals are recorded in a quiet room, requiring the participant to relax and sit in a chair. The EEG sensor is placed on the participant's scalp, and data is collected for at least 10 minutes in a resting state with eyes closed. Twenty-one Ag / AgCl electrodes are used, with the average of both earlobes as a reference electrode to measure scalp voltage. The 21 channels are placed according to the international 10-20 electrode placement standard. Nineteen electrodes covering the scalp include those in the frontal lobe (Fp1, Fp2, F3, F4, F7, F8, Fz), temporal lobe (T3, T4, T5, T6), parietal lobe (P3, P4, P7, P8), occipital lobe (O1, O2), and central (C3, C4) regions, plus two electrodes on the left and right earlobes (A1, A2). The collection process and the positions of the 21 electrodes are shown in Figure 2.

[0106] 1.2 Data Preprocessing

[0107] The signals measured by the electrodes covering our scalp are actually the difference between the active electrode and the reference electrode. We used the average potential of electrodes A1 and A2 as our reference potential, thus obtaining the signals from the remaining 19 active electrodes. The sampling frequency of the acquired EEG signals was set to 256Hz, and a notch filter was used to suppress 50Hz power line noise. All EEG signals were subjected to FIR bandpass filtering with cutoff frequencies of 0.5Hz and 70Hz. Artifacts such as heartbeats, eye movements, and muscle activity were removed using the Artifact Space Reconstruction (ASR) algorithm.

[0108] 1.3 Data Segmentation

[0109] Deep learning methods require a large amount of data for training to achieve good generalization results; if the sample size is small, overfitting is likely to occur. Due to privacy concerns regarding medical datasets, there are currently no open-source large-scale EEG datasets available for training. Therefore, to achieve better results, we segmented the acquired 20-minute EEG signal to expand the dataset. The EEG signal was segmented at fixed 60-second intervals, with each segment overlapping by 25%. This strategy is a practical method to overcome the limited availability of data. Each segmented EEG signal is treated as an independent sample, and the EEG signal of each independent sample is as follows:

[0110]

[0111] in This represents the time variable of EEG brainwave signals. This indicates that the time variable Discrete variables obtained by discretization. This represents a specific point in time after time has been discretized. This represents the total number of discrete points in the signal. This indicates the EEG channel number. Excluding reference electrodes A1 and A2, there are a total of 19 available EEG signals. This represents the EEG signals of all channels in this independent sample;

[0112] 2. The feature extraction module uses analytical Morlet wavelets as wavelet basis functions to perform synchronous compressed wavelet transform on all 19 channels of each independent sample, obtaining the synchronous compressed wavelet time-frequency matrix of each channel signal. ,in Indicates frequency, This represents the time variable of the EEG brainwave signal, i.e., the time of EEG signal acquisition;

[0113] Select any two channels from each independent sample, where the EEG signals of the two channels are denoted as follows: and Calculate the time-frequency wavelet autospectrum of each channel signal. and and the time-frequency wavelet cross spectrum of the two channel signals. ;

[0114] Based on the definition of coherence function, the two-channel signals are calculated using the obtained time-frequency wavelet autospectrum and time-frequency wavelet crossspectrum. and Synchronous compressed wavelet coherence between , will get Averaging over the time dimension yields the time-averaged synchronous compressed wavelet coherence. Let represent the average coherence of the two channel signals in the frequency dimension, where ;

[0115] Using time-frequency wavelet cross spectrum To calculate the phase difference between two channel signals And calculate the phase-locked value (PLV) between the two channel signals, where ;

[0116] The phase-locked value and the time-averaged synchronous compressed wavelet coherence are fused together, and then fused through a mapping function. Change the range from Mapped to The purpose of the mapping function is to standardize the fusion features while enhancing the differentiation between the fusion features of patients and control participants; ultimately, a new fusion feature, P-MSWC, is obtained, where... (The P-MSWC obtained from every two channel signals is not a single value, but a series of values ​​with frequency as the variable.)

[0117] Iterate through all channels of the signal in each independent sample and calculate the fusion feature between every two channels of the signal in each independent sample. , will integrate features Expanding the frequency dimension yields a connection matrix constructed from a series of values ​​at each frequency, thus... The frequency dimension is transformed into the depth dimension of a high-dimensional connectivity matrix; these connectivity matrices are combined to form a high-dimensional brain functional connectivity matrix; the fusion feature of the edge between each two channel signals is used as the edge connection strength between these two signals, as detailed below:

[0118] 2.1 Calculate the Synchrosqueezed Wavelet Transform (SWT). The Continuous Wavelet Transform (CWT) is a commonly used time-frequency analysis method. It represents non-stationary signals as two-dimensional functions of time and frequency, allowing for more intuitive analysis and processing. The Synchrosqueezed Wavelet Transform (SWT) combines CWT with time-frequency rearrangement methods. It borrows the style and principles of the EMD algorithm while exhibiting excellent robustness to signal noise, effectively avoiding interference from white noise. It can also obtain relatively clear time-frequency curves when analyzing some noisy medical and physiological signals. The specific steps are as follows:

[0119] Step 1: Using the analytical Morlet wavelet as the wavelet basis function, perform a continuous wavelet transform (CWT) on each channel signal in each independent sample to obtain the wavelet coefficients defined in the time domain. :

[0120]

[0121] in: It is the signal of one channel in an independent sample. It is a scaling variable. It is a time-shifted variable. It is the conjugate of the wavelet basis functions. It is Translation in the time dimension The unit is then expanded on a scale to the original size. times, It is a time factor The derivative of represents A quantity that changes infinitely little;

[0122] Step 2: According to Plancherel's theorem, the wavelet coefficients of the single-channel signal defined in the time domain are... Transforming to the frequency domain yields wavelet coefficients defined in the frequency domain. Defined by the following expression:

[0123]

[0124] in It is an angular frequency variable. yes Fourier transform, It is a single-channel signal Fourier transform, It is the imaginary unit;

[0125] For any and If its corresponding wavelet coefficients are defined in the frequency domain The instantaneous angular frequency of a single-channel signal is expressed as:

[0126]

[0127] Step 3: Use formula (4) to calculate the instantaneous frequency of the single-channel EEG signal, and establish a system from... The mapping transforms wavelet coefficients from the time-scale domain to the time-angular frequency domain; in practical applications, continuous values ​​are converted into discrete values ​​that can be computed by computers. Therefore, the scaling variable in the above process... and angular frequency variables Convert to discrete values and ,in Represents a scaled discrete variable. Represents the discrete variable of angular frequency; Represents the interval of the scale variable. Represents the interval of angular frequency variables; their definitions are as follows: and By using arbitrary center frequency bands The intervals on both sides Compression is performed to obtain the synchronous compressed wavelet transform value of the channel signal. :

[0128]

[0129] Step 4: Adjust the time shift variable in formula (5) Discretization is performed, and a synchronous compressed wavelet transform is established to convert angular frequency to scale and then to frequency to time, thus obtaining the time-frequency coefficient matrix of the channel signal. This is the synchronous compressed wavelet time-frequency matrix:

[0130]

[0131] in, This represents the frequency; thus, it is transformed into a synchronous compressed wavelet transform that can be directly calculated by a computer.

[0132] 2.2 Calculation of Synchronous Compressed Wavelet Coherence

[0133] The coherence function, also known as amplitude-squared coherence, measures the linear correlation between two signals in the frequency domain by calculating the power spectral density. The amplitude-squared coherence of EEG signals is defined as follows:

[0134]

[0135] in, This represents the signal from two channels in an independent sample. and Related functions, It is a signal and Cross power spectral density, and These are signals and Each has its own power spectral density; combining the amplitude squared coherence function and the synchronous compressed wavelet transform, the synchronous compressed wavelet coherence is constructed, and the specific steps are as follows:

[0136] Step 1: Assume the signal and These represent EEG signals from two different channels, each with a length of [missing information]. Calculate the synchronous compressed wavelet transforms of the two signals respectively to obtain their respective time-frequency coefficient matrices. and Two signals and The time-frequency wavelet cross spectrum is:

[0137]

[0138] Here Indicates complex conjugation;

[0139] Two signals and The time-frequency wavelet autospectral values ​​are as follows:

[0140]

[0141] Step 2: Define two signals using formulas (7), (8), and (9). and The Synchrosqueezed Wavelet Coherence (SWC) value is:

[0142]

[0143] here It is a smooth function. This represents the smoothing scaling factor in time and frequency; the smoothing scaling factor in time and frequency is a single value, namely... , here and The difference is that they represent different values; the numerator is used... As a smoothing factor, it is used in the denominator. As a smoothing factor.

[0144] 2.3 Calculate the phase-locked loop value

[0145] Phase-locked value (PLV) is an important method for studying the synchronicity of neural activity in the brain. It can be used to assess the functional connectivity of the brain. PLV first separates phase information from the signal, and then calculates the absolute value of the average phase difference between two signals. When there is strong phase synchronization between two signals, the PLV value will be close to 1, and vice versa. Phase synchronization is widespread in EEG signals. Compared with other synchronization detection methods, the PLV method can easily and quickly detect phase synchronization while maintaining a similar amount of information as other complex measurements.

[0146] Through two signals and The phase-locked value (PLV) of two signals is calculated using the time-frequency wavelet cross spectrum. The specific steps are as follows:

[0147] Step 1: From the signal and Time-frequency wavelet cross spectrum To calculate the phase difference between two signals Specifically:

[0148]

[0149] in and They represent The imaginary and real parts, It is usually expressed in radians and its value ranges from 1 to 2. ;

[0150] Step 2: Calculate the phase difference according to formula (11) The phase-locked value (PLV) of the two signals is calculated as follows:

[0151]

[0152] in This indicates that the average value is taken over time, and the range of the phase-locked value (PLV) is as follows: ;

[0153] 2.4 Computational Fusion Features

[0154] Raw EEG signals are complex and information-dense time series. Assessing the correlation between two signals requires a multidimensional approach to avoid information loss; evaluating the similarity of two signals from a single aspect may lead to incomplete or biased measurements. Therefore, by assessing their similarity from as many angles as possible, it is necessary to conduct a comprehensive and multifaceted analysis of EEG signals to obtain a reliable estimate of their correlation.

[0155] SWC describes the correlation between two signals from three aspects: time domain, frequency domain, and amplitude. PLV extracts the phase information of the signals and describes the difference between the two signals from the perspective of phase. By fusing the two features, SWC and PLV, the correlation between signals can be studied from more directions. The specific steps for fusing features are as follows:

[0156] Step 1: Inspired by the phase-locked value (PLV), the obtained SWC value is averaged over the time dimension to obtain the time-averaged synchronous compressed wavelet coherence. (Mean Synchrosqueezed Wavelet Coherence, MSWC) is used to represent the average coherence of two signals in the frequency dimension.

[0157]

[0158] The scope of MSWC is as follows: ;

[0159] Step 2: Synchronize the phase-locked value (PLV) and time averaging to compress the wavelet coherence. The fusion process yields new fusion features, which are:

[0160]

[0161] in , Represents PLV, This represents a mapping function that ranges the interval from... Mapped to ; The definition is as follows:

[0162]

[0163] in, Represents an unknown variable. The purpose of the mapping function is to standardize fusion features while increasing their discriminative power, enhancing the differentiation between fusion features of patients and control participants, and improving the utility of new features by reducing smaller feature values ​​and increasing larger feature values. The graph of the mapping function is shown below. Figure 3 As shown.

[0164] 2.5 Constructing a high-dimensional brain functional connectivity matrix

[0165] Functional connectivity refers to the temporal correlation of activity between two or more brain regions, indicating that these regions are functionally interconnected and work together to support specific cognitive or behavioral processes. This study used P-MSWC as a biomarker to assess the degree of functional connectivity between different brain regions. All channels in each independent sample were examined, and the P-MSWC values ​​between every two channels in all 19 channels were calculated, totaling [amount missing]. These results are arranged into a 19×19 connectivity matrix, which serves as a high-dimensional brain functional connectivity matrix. Frequency information is also incorporated into the depth information of the connectivity matrix. (The P-MSWC values ​​obtained from every two channel signals are a series of values ​​with frequency as the variable; the connectivity matrix is ​​calculated at each frequency, thus yielding a high-dimensional connectivity matrix.) A schematic diagram of the constructed high-dimensional connectivity matrix is ​​shown below. Figure 4 As shown. The connectivity matrix constructed by this method helps to more comprehensively assess functional connectivity across multiple brain regions and frequency ranges, and enables a more detailed understanding of brain neural communication.

[0166] 3. A lightweight CNN model is constructed in the brain disease recognition module. This model integrates the Inception module and six convolutional layers. Some of these convolutional layers incorporate a CBAM (Convolutional Block Attention Module). The Inception module reduces the number of parameters while maintaining the same receptive field, replacing the traditional 5×5 convolutional kernel with two 3×3 kernels, and replacing the 3×3 kernels with 1×3 and 3×1 kernels, thus reducing the number of training parameters. The CBAM attention module integrates channel attention and spatial attention, enabling the model to focus more on important features in the input, thereby improving model performance. This model can learn key information from the high-dimensional brain functional connectivity matrix to identify various brain diseases, achieving accurate detection. Details are as follows:

[0167] 3.1 Construction of a Lightweight CNN Model

[0168] Convolutional Neural Networks (CNNs) can efficiently process high-dimensional data by sharing convolutional kernels. Furthermore, they offer advantages over traditional machine learning algorithms by automatically performing feature extraction and implementing end-to-end models. In particular, CNNs can effectively identify complex connections presented in the brain's connectivity matrix, and can very effectively distinguish the differences between the brain networks of patients and healthy individuals.

[0169] In this invention, we propose a lightweight CNN model based on a high-dimensional functional connectivity matrix to identify patients with brain diseases and healthy individuals. The lightweight CNN model includes an Inception module, followed by six convolutional layers, a fully connected layer, and a Sigmoid layer. Each of the six convolutional layers contains a CBAM attention module, which incorporates channel attention and spatial attention mechanisms.

[0170] The Inception module structure is as follows: Figure 5 As shown, it consists of multiple cascaded network branches, where:

[0171] The first branch consists of 5 convolutional layers. The first convolutional layer has 64 kernels with a kernel size of 1×1. The second convolutional layer has 32 kernels with a kernel size of 1×3. The third convolutional layer has 32 kernels with a kernel size of 3×1. The fourth convolutional layer has 32 kernels with a kernel size of 1×3. The fifth convolutional layer has 32 kernels with a kernel size of 3×1.

[0172] The second branch includes 3 convolutional layers. The first convolutional layer has 64 convolutional kernels with a kernel size of 1×1. The second convolutional layer has 32 convolutional kernels with a kernel size of 1×3. The third convolutional layer has 32 convolutional kernels with a kernel size of 3×1.

[0173] The third branch includes one pooling layer and one convolutional layer. The pooling layer uses max pooling with a stride of 1, and the convolutional layer has 32 convolutional kernels, each of which is 1×1 in size.

[0174] The fourth branch includes one convolutional layer with 32 convolutional kernels, each kernel being 1×1 in size;

[0175] Different kernel sizes in the Inception module correspond to different receptive fields. While larger kernels produce better global features, they also increase computational cost. To extract features more effectively, we used kernels of 1×1, 3×3, and 5×5 sizes. To reduce the number of parameters while maintaining the same receptive field, we adopted a local module from Inception V3, where the 5×5 kernel was replaced by two 3×3 kernels, and the 3×3 kernels were replaced by 1×3 and 3×1 kernels.

[0176] To better select more key information from the high-dimensional brain connectivity matrix for identifying various brain diseases, a CBAM attention module was added to three of the six convolutional layers following the Inception module. The CBAM attention module includes channel attention and spatial attention mechanisms. The overall network architecture is as follows: Figure 6 As shown.

[0177] The six convolutional layers added after the Inception module are:

[0178] The first convolutional layer group consists of one convolutional layer, one batch normalization layer, and one tanh activation function layer. The convolutional layer has 64 convolutional kernels, and each convolutional kernel is 1×1 in size.

[0179] The second convolutional layer group consists of one convolutional layer, one batch normalization layer, and one Tanh activation function layer. The number of convolutional kernels in the convolutional layer is 16, and the size of each convolutional kernel is 1×1.

[0180] The third convolutional layer group consists of one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 32 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0181] The fourth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 64 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0182] The fifth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 128 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0183] The sixth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, and one dropout layer. The convolutional layer has 256 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%.

[0184] In the proposed network model, two 1×1 convolutional kernels are used after the Inception module to perform feature dimensionality reduction. This step helps remove redundant features and enables cross-channel information exchange, while increasing the non-linearity of the features. All convolutional modules in the network structure consist of convolutional layers, batch normalization layers, and Tanh activation functions. Batch normalization layers accelerate network training and convergence while preventing the risks of gradient explosion and overfitting. The kernel size and stride of the max-pooling layer are both set to 2 to reduce the size of the feature maps by half. Furthermore, to prevent model overfitting, a Dropout layer is added with a dropout ratio of 20%.

[0185] Finally, the output is generated through four fully connected layers and one Sigmoid layer, with the specific model parameters set as follows:

[0186] The first fully connected layer has 64 neurons and uses the Tanh activation function.

[0187] The second fully connected layer has 16 neurons and uses the Tanh activation function;

[0188] The third fully connected layer has 4 neurons and uses the Tanh activation function;

[0189] The fourth fully connected layer has one neuron and uses the Tanh activation function.

[0190] Finally, the Sigmoid layer is used to output the binary classification probability, which will output a value between 0 and 1. With 0.5 as the dividing line, when the output value is greater than 0.5, the prediction result is defined as diseased, and when it is less than 0.5, the prediction result is defined as healthy.

[0191] The lightweight CNN model is used to identify patients with brain diseases and healthy individuals. The training process of the model specifically includes the following steps:

[0192] Step 1: Select more than 150 healthy individuals and more than 150 patients with encephalopathy as all participants, collect EEG brain signals from all participants, input the EEG brain signals of all participants into the data preprocessing module for preprocessing, use all the preprocessed independent samples as the sample dataset, label the sample dataset, set the data label value of healthy individuals to 0, and set the data label value of patients with encephalopathy to 1.

[0193] Step 2: Input each sample in the labeled sample dataset into the feature extraction module, and obtain the high-dimensional brain functional connectivity matrix after processing by the feature extraction module;

[0194] The feature extraction module calculates frequency values ​​using MATLAB R2021a. The frequency values ​​were calculated to one decimal place and duplicates were eliminated. A total of 225 frequency values ​​between 0.5 and 70 Hz were selected. Finally, a 19×19×225 three-dimensional matrix was generated for each sample. The 225 represents the depth dimension of the matrix, which is the high-dimensional feature matrix extracted from the sample. The 19×19 matrix represents the functional connectivity matrix of 19 channels. Each element is between 0 and 1. The larger the value, the greater the correlation between the two channel pairs.

[0195] Step 3: Input all the extracted high-dimensional brain functional connectivity matrices into the lightweight CNN model for training. In order to train and evaluate the proposed network, observe the loss value at each epoch stage. The training process ends when the model loss converges to less than 0.05.

[0196] Using 100 epochs, each training iteration outputs a value between 0 and 1, with 0.5 as the dividing line. When the output value is greater than 0.5, the prediction result is defined as diseased, and when it is less than 0.5, the prediction result is defined as healthy.

[0197] Step 4: During training, set the batch size to 32 and use the Adam optimizer with an initial learning rate of 1e-4; use the binary cross-entropy loss function, defined as:

[0198]

[0199] in Indicates the actual value. The values ​​represent predicted values. The network was trained and tested on a machine equipped with an Nvidia RTX 3060 GPU. The entire network model was deployed using Python 3.8.8 and the PyTorch framework. Table 1 shows the comprehensive hyperparameter information for the entire CNN model.

[0200] Table 1. Detailed hyperparameter information for each module of the CNN model.

[0201]

[0202] The key technologies of this invention are twofold:

[0203] (1) A novel marker integrating the three domains of time, frequency and phase was developed. This marker can be used to construct the brain functional connectivity matrix in multi-channel EEG signals.

[0204] (2) A lightweight convolutional neural network model was built, which can effectively utilize the brain functional connectivity matrix constructed by novel biomarkers to achieve more accurate and efficient detection of various brain diseases.

[0205] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the scope of protection of the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, any person skilled in the art can make equivalent substitutions or changes based on the technical solution and inventive concept of the present invention within the scope of the technology disclosed in the present invention. These simple modifications are all within the scope of protection of the present invention.

[0206] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0207] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A brain disease detection system based on multi-domain fusion features and deep learning, characterized in that, It includes a data acquisition and preprocessing module, a feature extraction module, and a brain disease identification module; among which: (1) The data acquisition and preprocessing module acquires at least 20 minutes of resting-state EEG signals from participants; and performs preliminary preprocessing on the acquired EEG signals; the preprocessed EEG signals are segmented at fixed time intervals of 60 seconds, with each segment overlapping by 25%, and each segment is treated as an independent sample. The EEG signal of each independent sample is as follows: in This represents the time variable of the EEG electroencephalogram (EEG) signal, specifically the time of EEG signal acquisition. This indicates that the time variable Discrete variables obtained by discretization. This represents a specific point in time after time has been discretized. This represents the total number of discrete points in the signal. Indicates the sequence number of the EEG channel. This represents the EEG signals of all channels in this independent sample; (2) The feature extraction module uses the analytical Morlet wavelet as the wavelet basis function to perform synchronous compressed wavelet transform on all channel signals of each independent sample to obtain the synchronous compressed wavelet time-frequency matrix of each channel signal. ,in Indicates frequency; Select any two channels from each independent sample, where the EEG signals of the two channels are denoted as follows: and Calculate the time-frequency wavelet autospectrum of each channel signal. and and the time-frequency wavelet cross spectrum of the two channel signals. ; Based on the definition of coherence function, the two-channel signals are calculated using the obtained time-frequency wavelet autospectrum and time-frequency wavelet crossspectrum. and Synchronous compressed wavelet coherence between , will get Averaging over the time dimension yields the time-averaged synchronous compressed wavelet coherence. Let represent the average coherence of the two channel signals in the frequency dimension, where ; Using time-frequency wavelet cross spectrum To calculate the phase difference between two channel signals And calculate the phase-locked value (PLV) between the two channel signals, where ; The phase-locked value and the time-averaged synchronous compressed wavelet coherence are fused together, and then fused through a mapping function. Change the range from Mapped to Finally, a new fusion feature, P-MSWC, is obtained, in which... ; Iterate through all channels of the signal in each independent sample and calculate the fusion feature between every two channels of the signal in each independent sample. , will integrate features The frequency dimension is expanded to obtain a series of connection matrices constructed from values ​​at each frequency. These connection matrices are combined to form a high-dimensional brain functional connection matrix. (3) A lightweight CNN model is constructed in the brain disease identification module. This model integrates the Inception module and 6 convolutional layer groups. Among them, some convolutional layer groups also incorporate the CBAM attention module. The Inception module reduces the number of parameters while maintaining the same convolutional receptive field. The traditional 5×5 convolutional kernel is replaced by two 3×3 convolutional kernels, and the 3×3 convolutional kernels are replaced by 1×3 and 3×1 convolutional kernels. The CBAM attention module integrates channel attention and spatial attention. This model can learn key information from the high-dimensional brain functional connectivity matrix to identify various brain diseases, so as to achieve accurate detection of brain diseases. The specific contents are as follows: The lightweight CNN model includes an Inception module, followed by six convolutional layers, a fully connected layer, and a Sigmoid layer. The Inception module consists of multiple network branches, among which: The first branch consists of 5 convolutional layers. The first convolutional layer has 64 kernels with a kernel size of 1×1. The second convolutional layer has 32 kernels with a kernel size of 1×3. The third convolutional layer has 32 kernels with a kernel size of 3×1. The fourth convolutional layer has 32 kernels with a kernel size of 1×3. The fifth convolutional layer has 32 kernels with a kernel size of 3×1. The second branch includes 3 convolutional layers. The first convolutional layer has 64 convolutional kernels with a kernel size of 1×1. The second convolutional layer has 32 convolutional kernels with a kernel size of 1×3. The third convolutional layer has 32 convolutional kernels with a kernel size of 3×1. The third branch includes one pooling layer and one convolutional layer. The pooling layer uses max pooling with a stride of 1, and the convolutional layer has 32 convolutional kernels, each of which is 1×1 in size. The fourth branch includes one convolutional layer with 32 convolutional kernels, each kernel being 1×1 in size; The six convolutional layers added after the Inception module are: The first convolutional layer group consists of one convolutional layer, one batch normalization layer, and one tanh activation function layer. The convolutional layer has 64 convolutional kernels, and each convolutional kernel is 1×1 in size. The second convolutional layer group consists of one convolutional layer, one batch normalization layer, and one Tanh activation function layer. The number of convolutional kernels in the convolutional layer is 16, and the size of each convolutional kernel is 1×1. The third convolutional layer group consists of one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 32 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%. The fourth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 64 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%. The fifth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, one dropout layer, and one CBAM attention module. The convolutional layer has 128 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20%. The sixth convolutional layer group includes one convolutional layer, one batch normalization layer, one Tanh activation function layer, one pooling layer, and one dropout layer. The convolutional layer has 256 convolutional kernels, each with a size of 3×3. The pooling layer uses max pooling with a stride of 2, and the dropout layer accounts for 20% of the total layer size. Finally, the output is generated through four fully connected layers and one Sigmoid layer. The Sigmoid layer outputs the probability of binary classification, with 0.5 as the dividing line. When the output value is greater than 0.5, the prediction result is defined as diseased, and when it is less than 0.5, the prediction result is defined as healthy.

2. The brain disease detection system based on multi-domain fusion features and deep learning according to claim 1, characterized in that, The preprocessing content of the data acquisition and preprocessing module is as follows: resample the EEG signal to 256Hz, remove the data from the A1 and A2 electrodes, remove 50Hz power line noise using a notch filter, perform 0.5-70Hz bandpass filtering using an FIR filter, and remove eye movement artifacts using the ASR algorithm.

3. The brain disease detection system based on multi-domain fusion features and deep learning according to claim 1, characterized in that, The specific contents of the feature extraction module are as follows: S1.1, Calculate the synchronous compressed wavelet transform, the specific steps are as follows: Step 1: Using the analytical Morlet wavelet as the wavelet basis function, perform a continuous wavelet transform (CWT) on each channel signal in each independent sample to obtain the wavelet coefficients defined in the time domain. : in: It is the signal of one channel in an independent sample. It is a scaling variable. It is a time-shifted variable. It is the conjugate of the wavelet basis functions. It is Translation in the time dimension The unit is then expanded on a scale to the original size. times, It is a time factor The derivative of represents A quantity that changes infinitely little; Step 2: According to Plancherel's theorem, the wavelet coefficients of the single-channel signal defined in the time domain are... Transforming to the frequency domain yields wavelet coefficients defined in the frequency domain. Defined by the following expression: in It is an angular frequency variable. yes Fourier transform, It is a single-channel signal Fourier transform, It is the imaginary unit; For any and If its corresponding wavelet coefficients are defined in the frequency domain The instantaneous angular frequency of a single-channel signal is expressed as: Step 3: Establish from The mapping transforms the wavelet coefficients from the time-scale domain to the time-angular frequency domain; therefore, the scaling variable... and angular frequency variables Convert to discrete values and ,in Represents a scaled discrete variable. Represents the discrete variable of angular frequency; Represents the interval of the scale variable. Represents the interval of angular frequency variables; their definitions are as follows: and By using arbitrary center frequency bands The intervals on both sides Compression is performed to obtain the synchronous compressed wavelet transform value of the channel signal. : Step 4: Adjust the time-shift variable Discretization is performed, and a synchronous compressed wavelet transform is established to convert angular frequency to scale and then to frequency to time, thus obtaining the time-frequency coefficient matrix of the channel signal. This is the synchronous compressed wavelet time-frequency matrix: S1.2 Calculation of Synchronous Compressed Wavelet Coherence The amplitude-squared coherence of EEG signals is defined as follows: in, This represents the signal from two channels in an independent sample. and Related functions, It is a signal and Cross power spectral density, and These are signals and Each has its own power spectral density; combining the amplitude squared coherence function and the synchronous compressed wavelet transform, the synchronous compressed wavelet coherence is constructed, and the specific steps are as follows: Step 1: Assume the signal and All lengths Calculate the synchronous compressed wavelet transforms of the two signals respectively to obtain their respective time-frequency coefficient matrices. and Two signals and The time-frequency wavelet cross spectrum is: Here Indicates complex conjugation; Two signals and The time-frequency wavelet autospectral values ​​are as follows: Step 2: Define two signals and The synchronous compressed wavelet coherence value is: here It is a smooth function. The smoothing scaling factor in time and frequency; S1.3 Calculate the phase-locked loop value Through two signals and The phase-locked value (PLV) of two signals is calculated using the time-frequency wavelet cross spectrum. The specific steps are as follows: Step 1: From the signal and Time-frequency wavelet cross spectrum To calculate the phase difference between two signals Specifically: in and They represent The imaginary and real parts, Expressed in radians and with a range of values ​​of ; Step 2: Based on the phase difference The phase-locked value (PLV) of the two signals is calculated as follows: in This indicates that the average value is taken over time, and the range of the phase-locked value (PLV) is as follows: ; S1.4 Calculate the fusion features. The specific implementation steps are as follows: Step 1: Average the obtained SWC values ​​over the time dimension to obtain the time-averaged synchronous compressed wavelet coherence. Let represent the average coherence of the two signals in the frequency dimension: The scope of MSWC is as follows: ; Step 2: Synchronize the phase-locked value (PLV) and time averaging to compress the wavelet coherence. The fusion process yields new fusion features, which are: in , Represents PLV, This represents a mapping function that ranges the interval from... Mapped to ; The definition is as follows: in, ; S1.5 Constructing a High-Dimensional Brain Functional Connectivity Matrix Iterate through all channels of the signal in each independent sample and calculate the fusion feature between every two channels of the signal in each independent sample. , will integrate features Expanding the frequency dimension yields a connection matrix constructed from a series of values ​​at each frequency, thus... The frequency dimension is transformed into the depth dimension of the high-dimensional connection matrix; these connection matrices are combined to form a high-dimensional brain functional connection matrix.

4. The brain disease detection system based on multi-domain fusion features and deep learning according to claim 1, characterized in that, The specific parameters of the four fully connected layers in the lightweight CNN model constructed in the brain disease recognition module are set as follows: The first fully connected layer has 64 neurons and uses the Tanh activation function. The second fully connected layer has 16 neurons and uses the Tanh activation function; The third fully connected layer has 4 neurons and uses the Tanh activation function; The fourth fully connected layer has only one neuron and uses the Tanh activation function.

5. The brain disease detection system based on multi-domain fusion features and deep learning according to claim 1, characterized in that, The training process of the lightweight CNN model constructed in the brain disease recognition module specifically includes the following steps: Step 1: Select more than 150 healthy individuals and more than 150 patients with encephalopathy as all participants, collect EEG brain signals from all participants, input the EEG brain signals of all participants into the data preprocessing module for preprocessing, use all the preprocessed independent samples as the sample dataset, and label the sample dataset, that is, set the data label value of healthy individuals to 0 and the data label value of patients with encephalopathy to 1. Step 2: Input each sample in the labeled sample dataset into the feature extraction module, and the high-dimensional brain functional connectivity matrix is ​​obtained after processing by the feature extraction module; Step 3: Input all the extracted high-dimensional brain functional connectivity matrices into the lightweight CNN model for training. The training process ends when the model's loss converges to less than 0.05.