An electroencephalogram signal processing system

Abstract

This application discloses an electroencephalogram (EEG) signal processing system, relating to the fields of signal processing and deep learning. The system includes: a data acquisition device for acquiring multi-channel EEG signals from a subject; a preprocessing module for preprocessing the multi-channel EEG signals to obtain the frequency domain features of the subject's EEG signals; and a depression recognition module for inputting the frequency domain features of the EEG signals into a trained depression recognition model to obtain a depression recognition prediction result. The depression recognition model includes several sequentially connected ConvKAN Efficient convolutional layers and a classification module. Each ConvKAN Efficient convolutional layer includes a ConvKAN convolutional layer and a KANLinear layer constructed based on a spline interpolation function, which can extract EEG signal features that more fully express depression, thereby improving the accuracy of depression recognition from EEG signals.

Description

Technical Field

[0001] This application relates to the fields of signal processing and deep learning technology, and in particular to an electroencephalogram (EEG) signal processing system. Background Technology

[0002] Depression is a common mental disorder affecting hundreds of millions of people worldwide. Statistics show that depression has become one of the leading causes of global disease burden, profoundly impacting patients' quality of life and social adaptation. Currently, the diagnosis of depression mainly relies on clinician interviews and patient self-reporting, which is subjective and uncertain, making early diagnosis very difficult. To overcome this limitation, increasing research is exploring the application of electroencephalography (EEG) in the diagnosis of depression. EEG can record brain electrical activity in real time, revealing abnormal brain function patterns in patients with depression, providing important evidence for clinical diagnosis. However, due to the complexity of EEG signals and their susceptibility to noise, accurately extracting effective features and achieving efficient classification remains a major challenge in research.

[0003] In recent years, with the rapid development of deep learning technology, many deep learning-based models have been applied to the field of EEG signal processing, especially in depression identification. For example, related techniques have proposed a model based on graph convolutional networks (GCN) combined with secondary subject segmentation and attention mechanisms, an EEG-based depression transformer (EDT) model, and a lightweight depression detection method based on multi-scale dynamic graph convolutional networks and spiking neural networks. However, the features extracted from EEG signals by these methods cannot adequately represent the abnormalities of depression, resulting in low accuracy in EEG-based depression classification. Therefore, there is an urgent need for an EEG signal processing system that can more fully characterize depression, thereby improving the accuracy of EEG-based depression classification. Summary of the Invention

[0004] The purpose of this application is to provide an electroencephalogram (EEG) signal processing system that can extract EEG signal features that more fully express depression, thereby improving the accuracy of depression identification using EEG signals.

[0005] To achieve the above objectives, this application provides the following solution: This application provides a brainwave signal processing system, comprising: Data acquisition equipment used to acquire multi-channel EEG signals from subjects; The preprocessing module is used to: preprocess multi-channel EEG signals to obtain the frequency domain characteristics of the subject's EEG signals; A depression identification module is used to: input the frequency domain features of the EEG signal into a trained depression identification model to predict the depression identification result of the subject; the depression identification model includes a ConvKAN Efficient convolutional module and a classification module; the ConvKAN Efficient convolutional module includes several sequentially connected ConvKAN Efficient convolutional layers, each of the ConvKAN Efficient convolutional layers includes a sequentially connected ConvKAN convolutional layer and a KANLinear layer; the KANLinear layer is constructed based on a spline interpolation function.

[0006] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides an EEG signal processing system, which employs a depression recognition model including a ConvKAN Efficient convolutional module and a classification module. The ConvKAN Efficient convolutional module includes several sequentially connected ConvKAN Efficient convolutional layers, each of which includes a sequentially connected ConvKAN convolutional layer and a KANLinear layer. The KANLinear layer is constructed based on a spline interpolation function. In this depression recognition model, the KANLinear layer replaces the traditional linear convolutional kernel in the convolutional layers. The spline basis function performs nonlinear mapping on the frequency domain features of the input EEG signal, enabling the model to characterize the high-order nonlinear relationship between power features in different frequency bands when recognizing depression. This makes the model more fully representative of depression-related EEG abnormal patterns, thereby significantly improving the accuracy of depression recognition based on EEG signals. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of the structure of an electroencephalogram (EEG) signal processing system provided in an embodiment of this application.

[0009] Figure 2 This is an electrode arrangement diagram provided for an embodiment of this application.

[0010] Figure 3 This is a schematic diagram of the ConvKAN model structure provided in an embodiment of this application.

[0011] Figure 4 This is a schematic diagram of the ConvKAN Efficient model structure provided in an embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] In one exemplary embodiment, such as Figure 1 As shown, an electroencephalogram (EEG) signal processing system is provided, including a data acquisition device, a preprocessing module, and a depression recognition module.

[0015] Data acquisition equipment used to acquire multi-channel EEG signals from subjects.

[0016] The preprocessing module is used to preprocess multi-channel EEG signals to obtain the frequency domain characteristics of the subject's EEG signals.

[0017] The depression identification module is used to: input the frequency domain features of the EEG signal into the trained depression identification model to predict the depression identification result of the subject.

[0018] The EEG signal processing system also includes a model training module; the model training module is used to implement steps 1 to 3.

[0019] Step 1: Obtain the EEG dataset; the EEG dataset includes several original EEG signals of samples and the actual depression recognition result corresponding to each original EEG signal of the sample.

[0020] Step 2: Preprocess each sample's original EEG signal to obtain the frequency domain features of the sample EEG signal corresponding to each sample's original EEG signal.

[0021] Step 3: Using the frequency domain features of the sample EEG signals as input and the actual depression recognition results corresponding to the frequency domain features of the sample EEG signals as labels, train the CovnKAN Efficient model to obtain a trained depression recognition model; during the training process: use a dynamic grid update mechanism to adaptively update the grid position of the spline basis functions according to the data distribution of the EEG dataset; perform Kaiming initialization on the weights of the ConvKAN Efficient convolutional layer; and perform L1 regularization on the weight parameters of the spline basis functions.

[0022] The EEG dataset used was the EDRA (Electroencephalographic Data for Depression Recognition) dataset, which includes EEG signals from 50 university students. 26 participants were assigned to a high-risk depression group, and 24 to a low-risk control group. All participants underwent depression assessment using the Patient Reported Outcomes Measurement Information System (PROMIS) to ensure data accuracy and consistency. EEG signal acquisition was performed using 62 electrodes mounted on the participants' heads (electrode arrangement as shown in the image). Figure 2 As shown in the figure, the sampling frequency is 500Hz, and the signal quality is guaranteed by a standardized electrode arrangement.

[0023] During data acquisition, EEG signals not only record brain activity but are also susceptible to environmental noise and physiological interference. Therefore, a series of preprocessing steps are required to extract useful features. The primary task of data preprocessing is filtering to remove low-frequency and high-frequency noise. Bandpass filters are applied to the EEG signals, with a frequency range set from 0.5 to 50 Hz. The bandpass-filtered signals can better capture useful brain activity and remove most physiological and instrument noise.

[0024] Because EEG signals are susceptible to external factors (such as reaction time and fatigue), to increase the number of samples and enhance the robustness of the model, this application first performs time-slicing processing on the 60-second original EEG signal during the sample construction stage, dividing it into multiple non-overlapping short-time sequences of 1 second in length, denoted as sample EEG segments. Each sequence is used as an independent sample in subsequent analysis. In the feature extraction stage, the Welch method is used to estimate the power spectral density for each 1-second sample EEG segment. The Welch method reduces the variance of the spectral estimation and improves the stability of the frequency domain features by segmenting, windowing, and averaging the signal within a single sample.

[0025] Frequency domain analysis of EEG signals plays an important role in the identification of depression. EEG activity in different frequency bands reflects different cognitive states of the brain. EEG frequency bands include... Wave (0.5-4 Hz), Wave (4-8 Hz), Wave (8-13 Hz), Wave (13-30 Hz) and Waves (above 30 Hz), each frequency band corresponds to different physiological and psychological activities. Abnormal changes in each frequency band are commonly observed in the EEG signals of patients with depression, especially in… , , and Frequency domain analysis can effectively capture these changes and provide important features for subsequent classification of depression.

[0026] Power spectral density (PSD) is an important tool for measuring the distribution of a signal across various frequency bands and reflects the spectral characteristics of the signal. In this application, PSD analysis is used to extract the frequency domain characteristics of the EEG signal. PSD can be calculated using the following formula: (1); in, Indicates the signal at frequency The power spectral density at that point for EEG signal at any given time For the duration of the signal, These are the basis functions for the Fourier transform.

[0027] However, directly using periodogram estimation is susceptible to noise, leading to significant estimation errors. To improve the stability and reliability of power spectral density estimation, this application employs the Welch method to estimate PSD. Based on the periodogram definition, the Welch method divides the signal into multiple overlapping segments, and calculates the periodogram for each segment after windowing. Finally, the power spectral density estimation result is obtained by averaging the periodograms of the segments. This method effectively reduces the impact of noise and improves the statistical stability of frequency domain features. After obtaining the sample EEG segment, the sample EEG segment is divided into several overlapping signal segments. A Fourier transform is performed on each signal segment to obtain the Fourier transform result of the signal segment; the average of the Fourier transform results of all signal segments is taken to obtain the power spectral density estimation result of the sample EEG segment. Specifically, assuming the sample EEG segment... The signal is divided into L segments, each with a length of N, and the segments overlap by a length of M. After windowing each signal segment using a window function, the periodogram (i.e., the Fourier transform result) of each signal segment is calculated. Then, the average of the periodograms of all signal segments is taken to obtain the power spectral density estimate of the entire sample EEG sub-segment.

[0028] In frequency band selection, the main focus is on , , and The power spectral density (PSD) of each frequency band is calculated and integrated to obtain the power value for each band (the power value for each band is the estimated power spectral density for that band). The formula for calculating the power value is: (2); in, , , , They are respectively , , , Power spectral density estimation results for each frequency band. The power spectral density estimation results for each sample EEG sub-segment of each frequency band are integrated to obtain the power spectral density estimation results for each frequency band.

[0029] The obtained frequency band power values ​​are used to compress the power spectrum distribution in the continuous frequency domain into finite-dimensional frequency band-level energy features. These power values, as a quantitative description of the energy intensity of the EEG signal in each typical frequency band, serve as the fundamental feature unit for subsequent feature construction and model input. In subsequent steps, the frequency band power values ​​corresponding to each channel will be uniformly organized into sample-level feature vectors for the training and inference of the depression recognition model.

[0030] In the implementation, the scipy.signal.welch function in Python is used to calculate the power spectral density. This function obtains the power spectral density of the entire EEG segment of the sample based on the Welch method.

[0031] (3); in, It is the first The Fourier transform result of the signal segment. This is the power spectral density estimation result for the signal segment.

[0032] The power spectral density estimation result obtained by formula (3) is used as the direct input for the frequency band integral calculation in formula (2) to ensure the reliability of the frequency band power value in terms of noise suppression and statistical stability, thereby improving the discrimination ability of the final frequency domain features.

[0033] To ensure comparability of features across different channels, the extracted PSD features are standardized. Z-score standardization can be used, with the following formula: (4); in, These are eigenvalues, i.e., the power spectral density estimation results for the frequency band; and These are the mean and standard deviation of the feature, respectively. These are the standardized feature values, i.e., the frequency domain features of the EEG signal. These standardized features will be input into the subsequent depression recognition model. The standardized frequency band power features constitute the final frequency domain feature representation of the EEG signal, serving as the unified output of this step. This output is expressed in "channel..." The features are organized in the form of "frequency bands" and used directly as input features for the depression recognition model in subsequent steps, for model training, validation and testing.

[0034] Furthermore, after feature extraction, this application performed an analysis of variance (ANOVA) to assess the differences in PSD features between the depression group and the control group at different frequency bands. The results of the ANOVA showed that... , , and The frequency domain features showed significant differences between the depression group and the control group, which validates the effectiveness of frequency domain features in depression identification.

[0035] Through the above processing steps, the frequency domain features of the EEG signal were successfully extracted, providing a foundation for subsequent depression classification. Classification methods based on these frequency domain features can improve the accuracy of depression identification and have potential application value.

[0036] The depression recognition model includes a ConvKAN Efficient convolutional module and a classification module. The ConvKAN Efficient convolutional module comprises several sequentially connected ConvKAN Efficient convolutional layers. Each ConvKAN Efficient convolutional layer is a ConvKAN model, and each ConvKAN Efficient convolutional layer includes a sequentially connected ConvKAN convolutional layer and a KANLinear layer. The KANLinear layer is constructed based on a spline interpolation function, specifically a B-spline interpolation function.

[0037] The ConvKAN model, used in convolutional layers, combines CNNs and KANs, introducing KANs into the convolutional layer as a novel convolutional operation. The core idea of ​​this model is to enhance the convolutional layer's ability to represent input signals by using the non-linear feature mapping of KANs, building upon the traditional convolutional kernel. This makes it particularly suitable for processing high-dimensional and complex data, such as electroencephalogram (EEG) signals. The key to the ConvKAN model lies in replacing the traditional KANLinear layer with a KANLinear layer, enabling the convolutional operation to learn more complex feature transformations.

[0038] In the CovnKAN Efficient model, KAN can capture the complex structure and potential patterns of data, demonstrating superior performance in complex signal processing. Applying the CovnKAN Efficient model, which includes the KAN model, to the depression identification task can deeply explore depression-related features in EEG signals, thereby improving classification accuracy and model robustness.

[0039] The ConvKAN model structure is as follows: Figure 3 As shown, it includes an input layer, a hidden layer, and a KANLinear layer, with the hidden layer being a ConvKAN convolutional layer.

[0040] (1) KANLinear layer design: The KANLinear layer is the core module in the ConvKAN model. It uses B-spline interpolation to perform nonlinear transformation on the input features. Traditional convolutional layers extract local features by multiplying the convolutional kernel with the input features, while the KANLinear layer transforms the features through an interpolation mechanism.

[0041] Assume the input features of the KANLinear layer are , If the dimension is , then the output of the KANLinear layer It can be represented as: (5); in, Here is the weight matrix of the KANLinear layer. express OK The set of real matrices of columns; For the bias term of the KANLinear layer, The KANLinear layer uses an interpolation function to process the input features. Perform the transformation to generate a new feature representation. .

[0042] (2) ConvKAN Convolutional Layer Design: The ConvKAN convolutional layer replaces the convolutional kernel with a KANLinear layer, thereby capturing more complex nonlinear relationships in the input data. The workflow of the ConvKAN convolutional layer is as follows: Convolution Unfold: First, the input data... (in For batch size, For the number of channels, and The height and width of the input image are expanded into a matrix. Each column represents the features of a local region during the convolution operation. The convolution unwinding operation is defined as follows: (6); KAN processing: The matrix after convolution expansion The input is fed into a KANLinear layer for feature transformation. This process effectively captures complex nonlinear relationships within the input features. The output of the KANLinear layer... It can be represented as: (7); in, Indicates the number of groups. This refers to the dimension of the output features. In this way, the KANLinear layer can transform the input features... Converted to an output with nonlinear characteristics .

[0043] Output Reshape: Output features after processing by the KANLinear layer It will be reorganized into the standard output format of the convolution operation.

[0044] Specifically, the output features are rearranged according to a predetermined output shape to obtain the final output of the convolution. ,in Number of output channels and This specifies the height and width of the output image.

[0045] (8).

[0046] In the ConvKAN convolutional layer, the spatial dimension of the output (height) and width The value is calculated based on the input dimensions and the parameters of the convolution kernel. Assume the input height and width are respectively... and The size of the convolution kernel is Step size is Fill with The output height and width can be calculated using the following formula: (9); (10); in, This indicates a floor operation. In this way, the ConvKAN convolutional layer can guarantee that the output dimension of the convolution operation meets expectations.

[0047] The ConvKAN model, based on traditional convolutional neural networks, effectively enhances the model's ability to model complex data by introducing the KANLinear layer. In particular, when processing time-series data such as electroencephalograms, it can capture high-dimensional features that traditional convolution cannot express.

[0048] ConvKAN Efficient is an optimized and improved version of the ConvKAN model. Through dynamic grid update mechanism, optimized weight initialization and regularization mechanism, it improves the computational efficiency and performance of the model, especially when processing EEG data, it shows significant advantages.

[0049] The ConvKAN Efficient convolutional module performs layer-by-layer nonlinear mapping on the frequency domain features of the standardized EEG signal, outputting a global feature aggregation. This global feature aggregation is then fed into the classification module, which outputs a depression prediction result, including both depressed and non-depressed symptoms. In a specific example, the ConvKAN Efficient convolutional module consists of three sequentially connected ConvKAN Efficient convolutional layers. The ConvKAN Efficient model structure is as follows: Figure 4 As shown.

[0050] Dynamic mesh update mechanism: In the ConvKAN model, convolution operations are implemented using KAN. The core mechanism of KAN is to utilize... Spline basis functions are used to generate convolutional kernels, which are continuously adjusted during training to adapt to the input data. The ConvKAN Efficient model builds upon this by introducing a dynamic grid update mechanism, enabling the network to dynamically adjust the grid shape based on the distribution of the input data. This mechanism helps the network adapt to different data distributions and improves its ability to model complex nonlinear features.

[0051] Assuming input data If the feature dimension is d, the grid size is N, and the order of the spline basis function is m, then spline basis functions The calculation can be expressed as: (11); in, These are the weights to be learned, i.e., the grid positions; It is a spline basis function. These are the input features. In the ConvKAN Efficient model, as training progresses, the grid positions { The convolutional kernels are adaptively updated based on the data distribution, thereby optimizing their feature extraction capabilities. This is achieved through a data-driven reparameterization / statistical redistribution strategy. The grid is essentially the knot positions of spline basis functions. These knots determine the "resolution distribution" of the nonlinear mapping in the input space and which input intervals are modeled with greater precision. The spline basis functions obtained through this mechanism are not directly used as output but rather as the core computational primitives of the KANLinear layers within each ConvKAN Efficient convolutional layer of the ConvKAN Efficient model. During forward propagation, the spline basis functions participate in constructing the nonlinear convolutional kernel, mapping the local EEG features of the EEG signal frequency domain. During backward propagation, the grid positions of the spline basis functions are dynamically adjusted according to the data distribution of the input data of the CovKAN Efficient model, enabling the KANLinear layers to adapt to the EEG feature distributions of different subjects and different brain regions.

[0052] Therefore, the dynamic grid update mechanism indirectly adjusts the feature extraction method of each ConvKAN Efficient convolutional layer of the ConvKANEfficient model by influencing the shape and distribution of spline basis functions, enabling the model to have stronger adaptive capabilities in the input-to-output mapping process.

[0053] Optimized weight initialization: Kaiming initialization (also known as He initialization) is introduced for weight initialization. This initialization method considers the input and output dimensions of each layer and is initialized using the following formula: The weight initialization formula for the ConvKANEfficient convolutional layer is: (12); in, These are the weights of the ConvKAN Efficient convolutional layer; Indicates a Gaussian distribution; This represents the number of input neurons in the current ConvKAN Efficient convolutional layer. Through Kaiming initialization, the ConvKAN Efficient model can achieve better gradient propagation in the early stages of training, thereby improving training stability, especially in deep networks. The weights obtained in this step serve as parameters for the initial mapping functions of each layer, directly determining the propagation method of the frequency domain features of the input EEG signal in the initial stage of the network. Kaiming initialization ensures that the input features maintain a reasonable variance range when propagating between layers in multi-layer ConvKAN Efficient convolutional modules, thus avoiding gradient vanishing or exploding problems and enabling the spline basis function-driven nonlinear convolution to stably participate in model training.

[0054] Improved Regularization Mechanism: To avoid overfitting and enhance the model's generalization ability, the ConvKAN Efficient model improves its regularization mechanism. In the ConvKAN Efficient model, L1 regularization is used to constrain model complexity for the spline basis function weights W. The L1 regularization formula for the spline basis function weight parameters is as follows: (13); in, The weight parameter regularization function is used for the spline basis functions. It is a regularization hyperparameter; It is the first i The weight parameters of each spline basis function. L1 regularization forces the network to learn sparse convolutional kernels, thus maintaining efficiency while controlling complexity. This regularization term suppresses unimportant spline weights, enabling the ConvKAN Efficient model to form a sparse and discriminative nonlinear mapping structure inside the ConvKAN Efficient convolutional layers.

[0055] Furthermore, the ConvKAN Efficient model incorporates entropy regularization, enhancing its stability and robustness. The model training module also uses entropy regularization to constrain the distribution of the weight parameters of the spline basis functions, as shown in the following formula: (14); in, It is the entropy regularization function; It is the first iThe normalized probability distribution of the weight parameters of each spline basis function is used. Entropy regularization is used to constrain the smoothness of the spline weight distribution, preventing the model from becoming overly dependent on individual spline nodes during training. Through the above regularization mechanism, the ConvKAN Efficient model improves its overall generalization performance while maintaining its expressive power.

[0056] The above process yields a trained depression recognition model. Subsequently, a data acquisition device collects multi-channel EEG signals from the subject. This device includes multiple electrodes positioned in different brain regions of the subject's head, including the frontal lobe, central lobe, parietal lobe, temporal lobe, and occipital lobe. In one embodiment, the electrodes are arranged according to the international 10–20 EEG electrode system to obtain multi-channel EEG signals covering the entire brain region.

[0057] The following describes the preprocessing of the acquired multi-channel EEG signals. Regarding the preprocessing of the multi-channel EEG signals, the preprocessing module is used to: The multi-channel EEG signals include EEG signals in multiple frequency bands, wherein the frequency bands include... , , and The EEG signal is divided into bands; time-sliced ​​to obtain several EEG sub-segments; the power spectral density is estimated for each EEG sub-segment using the Welch method, resulting in a power spectral density estimation result for each EEG sub-segment in each band; the power spectral density estimation result for each EEG sub-segment in each band is integrated to obtain a power spectral density estimation result for each band; the power spectral density estimation results for all bands are standardized to obtain the frequency domain characteristics of the EEG signal.

[0058] In terms of using the Welch method to estimate the power spectral density of all the EEG sub-segments, the preprocessing module is used to: divide each EEG sub-segment of each frequency band into several overlapping signal segments; perform Fourier transform on each signal segment to obtain the Fourier transform result of the signal segment; and take the average of the Fourier transform results of all signal segments based on the above formula (3) to obtain the power spectral density estimation result of the EEG sub-segment.

[0059] Then, the power spectral density estimation results of the EEG sub-segments are substituted into the integral formula (2) of the corresponding frequency band to obtain the power spectral density estimation results of each frequency band. Then, based on formula (4), the power spectral density estimation results of each frequency band are Z-score standardized to obtain the final frequency domain characteristics of the EEG signal.

[0060] The frequency domain features of the EEG signal are input into a trained depression recognition model to predict the subject's depression status. This prediction result can be displayed on a screen for doctors to review and use, assisting in the diagnosis of depression.

[0061] Compared with existing depression recognition methods based on traditional convolutional neural networks or single frequency domain features, the depression recognition method proposed in this application based on EEG signal frequency domain features and the ConvKAN Efficient model has significant advantages in terms of recognition accuracy, modeling ability, training stability, and generalization performance. Its technical effects are mainly reflected in the following aspects: (a) Improve the accuracy of depression identification and enhance the ability to express complex EEG patterns: The method proposed in this application can effectively improve the accuracy of depression identification, especially when processing EEG signals with strong nonlinear and non-stationary characteristics, it shows better discrimination performance.

[0062] The aforementioned technical advantages primarily stem from the ConvKAN Efficient model structure. Specifically, this model replaces traditional linear convolutional kernels with KANLinear layers in its convolutional layers, using spline basis functions to nonlinearly map the frequency domain features of the input EEG signal. This allows the model to characterize higher-order nonlinear relationships between power features across different frequency bands. Compared to related techniques that rely solely on linear convolutions or fixed activation functions, this model provides a more comprehensive representation of depression-related EEG abnormalities, thereby significantly improving classification accuracy.

[0063] (ii) Enhance the robustness of the model to differences in participants and channels, and improve generalization performance: The method described in this application maintains stable recognition performance under different subjects, different electrode channels, and different data distributions, significantly reducing the model's sensitivity to individual differences and noise interference.

[0064] The effectiveness of this technique stems from the synergistic effect of the dynamic grid update mechanism and the frequency domain feature modeling method. On the one hand, by extracting power spectral density and frequency band power, the original time-domain EEG signal is converted into frequency domain features of EEG signals with clear physiological significance, effectively reducing the influence of time-domain noise and random fluctuations. On the other hand, the dynamic grid update mechanism in the ConvKAN Efficient model can adaptively adjust the grid position of the spline basis function according to the input feature distribution, enabling the model to maintain high-resolution modeling ability of key feature regions under different subjects and different data distributions, thereby significantly improving the model's generalization ability.

[0065] (III) Improve model training stability and convergence efficiency, and reduce training difficulty: Compared with existing deep learning models that are prone to gradient instability or slow convergence in the early stages of training, the method in this application exhibits faster convergence speed and higher training stability during training.

[0066] The effectiveness of this technique primarily stems from the introduced optimized weight initialization strategy and regularization mechanism. By employing the Kaiming initialization method to initialize the spline basis function weights, the variance of each layer's output remains stable during forward propagation, effectively avoiding the vanishing or exploding gradient problems. Simultaneously, by combining L1 regularization and entropy regularization constraints on the spline basis function weight parameters, redundant parameters are suppressed while maintaining the model's expressive power, thus making the training process more stable and efficient.

[0067] (iv) Reduce model complexity and improve computational efficiency while ensuring high recognition performance: This application improves recognition performance without significantly increasing model computational complexity, making it suitable for deployment and application under limited computing power conditions.

[0068] The effectiveness of this technique stems from the structural optimization design of the ConvKAN model. By using sparsity constraints on spline basis functions, the representation of the convolutional kernel in the feature dimension becomes more compact, reducing the proportion of invalid features involved in the calculation. At the same time, in the frequency domain feature extraction stage, the original EEG signal is reduced in dimensionality through power spectral density and frequency band integration, which reduces the size of the subsequent model input and improves the overall system efficiency.

[0069] (v) Improve the interpretability and medical application value of depression identification results: Compared to black-box methods that rely solely on deep features, the method in this application improves the interpretability of the model's decision-making process to some extent.

[0070] The effectiveness of this technique stems from a feature-based modeling approach that utilizes power spectral density and frequency band power. This is achieved through explicit modeling... , , , By using EEG frequency band energy features that are highly correlated with depression, the model input has a clear neurophysiological meaning. Furthermore, by using spline basis functions to nonlinearly weight the features of different frequency bands, it is helpful to analyze the contribution of different frequency bands in depression identification, thereby enhancing the application value of the method in medical auxiliary diagnosis scenarios.

[0071] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, 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 this application.

Claims

1. An electroencephalographic signal processing system, characterized by, The electroencephalogram (EEG) signal processing system includes: Data acquisition equipment used to acquire multi-channel EEG signals from subjects; The preprocessing module is used to: preprocess multi-channel EEG signals to obtain the frequency domain characteristics of the subject's EEG signals; A depression identification module is used to: input the frequency domain features of the EEG signal into a trained depression identification model to predict the depression identification result of the subject; the depression identification model includes a ConvKAN Efficient convolutional module and a classification module; the ConvKAN Efficient convolutional module includes several sequentially connected ConvKAN Efficient convolutional layers, each of the ConvKAN Efficient convolutional layers includes a sequentially connected ConvKAN convolutional layer and a KANLinear layer; the KANLinear layer is constructed based on a spline interpolation function.

2. The electroencephalographic signal processing system of claim 1, wherein, In terms of preprocessing multi-channel EEG signals, the preprocessing module is used for: Multi-channel EEG signals include EEG signals from multiple frequency bands; time-slicing is performed on the EEG signals of each frequency band to obtain several EEG sub-segments; Using the Welch method, the power spectral density of all the EEG sub-segments is estimated to obtain the power spectral density estimation results for each EEG sub-segment in each frequency band. The power spectral density estimation results for each EEG sub-segment of each frequency band are integrated to obtain the power spectral density estimation results for each frequency band. The power spectral density estimation results for all frequency bands are standardized to obtain the frequency domain characteristics of the EEG signal.

3. The EEG signal processing system according to claim 2, characterized in that, Frequency bands include , , and wave bands.

4. The electroencephalographic signal processing system of claim 2, wherein, In estimating the power spectral density of all the EEG segments using the Welch method, the preprocessing module is used for: Each frequency band and each EEG sub-segment is divided into several overlapping signal segments. Perform a Fourier transform on each signal segment to obtain the Fourier transform result of the signal segment; The power spectral density estimation result of the EEG sub-segment is obtained by averaging the Fourier transform results of all signal segments.

5. The electroencephalographic signal processing system of claim 2, wherein, The standardization process is Z-score standardization.

6. The electroencephalographic signal processing system of claim 1, wherein, The EEG signal processing system further includes a model training module; the model training module is used for: Acquire an EEG dataset; the EEG dataset includes several raw EEG signals of samples and the actual depression recognition result corresponding to each raw EEG signal of the sample; Each of the original EEG signals of the sample is preprocessed to obtain the frequency domain features of the sample EEG signal corresponding to each original EEG signal of the sample. Using the frequency domain features of the sample EEG signals as input and the actual depression recognition results corresponding to the frequency domain features of the sample EEG signals as labels, the CovnKAN Efficient model is trained to obtain a trained depression recognition model. During the training process: a dynamic grid update mechanism is used to adaptively update the grid position of the spline basis functions according to the data distribution of the EEG dataset; the weights of the ConvKAN Efficient convolutional layer are initialized using Kaiming; and the weight parameters of the spline basis functions are L1 regularized.

7. The electroencephalographic signal processing system of claim 6, wherein, The model training module is used to: during forward propagation, use spline basis functions to construct nonlinear convolution kernels to map the frequency domain features of EEG signals; and during backward propagation, dynamically adjust the grid positions of the spline basis functions according to the data distribution of the input data of the CovnKAN Efficient model.

8. The electroencephalographic signal processing system of claim 6, wherein, The weight initialization formula for the ConvKAN Efficient convolutional layer is: ； wherein, is the weight of the ConvKAN Efficient convolutional layer; denotes a Gaussian distribution; is the number of input neurons of the ConvKAN Efficient convolutional layer.

9. The electroencephalographic signal processing system of claim 6, wherein, The L1 regularization formula for the weight parameters of the spline basis functions is as follows: ； wherein, is a weight parameter of the k-th spline basis function; is a regularization hyperparameter; is the k-th weight parameter of the spline basis functions; i is the k-th weight parameter of the spline basis functions.

10. The electroencephalographic signal processing system of claim 6, wherein, The model training module is also used to: constrain the distribution of the weight parameters of the spline basis functions using an entropy regularization algorithm, as shown in the following formula: ； wherein, is an entropy regularizing function; is the normalized probability distribution of the i th spline basis function weight parameter.

Images