Depression population judgment system based on forehead sparse channel electroencephalogram

Abstract

The invention belongs to the technical field of auxiliary diagnosis of depressive disorder, and particularly provides a depressive population judgment system based on forehead sparse channel electroencephalography, which comprises a data acquisition module for acquiring EEG (electroencephalogram) signals of a to-be-tested person to obtain original EEG signals; the self-supervised characterization learning module is used for enhancing the original electroencephalogram signal to obtain an enhanced electroencephalogram signal, and labeling the original electroencephalogram signal and the enhanced electroencephalogram signal; the feature decoupling module comprises a public encoder and a private encoder, and the public encoder and the private encoder are used for decoupling adaptive features and invariant features from the original electroencephalogram signals and decoupling adaptive features and invariant features from the enhanced electroencephalogram signals; the feature dynamic interaction module is used for realizing dynamic information transmission and optimization of invariant features and adaptive features based on a cross attention mechanism; and the depression crowd judgment module is used for judging depression of the to-be-tested person according to the optimized features.

Description

Technical Field

[0001] The present invention belongs to the technical field of auxiliary diagnosis of depressive disorders, and specifically provides a depressive population determination system based on sparse-channel electroencephalogram of the prefrontal region. Background Art

[0002] Human emotion recognition based on electroencephalogram (EEG) aims to perceive the emotional state of an individual by analyzing EEG. As a non-invasive technology capable of capturing brain activities, EEG can reflect the dynamic changes of various psychological activities such as emotions and cognitions. Different from external emotional cues such as facial expressions, voices, and behaviors, EEG directly comes from the brain and provides insights into emotions from a more objective perspective.

[0003] In recent years, EEG has received extensive attention in affective computing and has been widely applied in fields such as mental health assessment, human-computer affective interaction, and fatigue detection, showing great development potential. For EEG emotion recognition, quite a lot of work analyzes the human emotional state through EEG data of 32, 64 or more channels, which focuses on the neural activities in multiple brain regions. However, too many channels bring several challenges. First, it reduces the operability of the experiment, and the complex design is difficult to achieve cross-scenario generalization. Second, the high-dimensional input significantly increases the computational complexity, reduces the training efficiency, and may affect the stability of the model. In addition, long-term signal acquisition or wearing multiple devices reduces the acceptance of the subjects, especially for patients with emotional disorders, which may affect the data quality and experimental results. Therefore, exploring EEG emotion recognition solutions based on sparse channels is crucial for solving these limitations.

[0004] The prefrontal region plays an important role in emotion processing, regulation, and cognitive control. The EEG activities in this region can better reflect the process of emotion regulation and emotion experience, which provides the possibility of using prefrontal EEG for emotion recognition. Dense-channel EEG can focus on the changes in neural activities in different brain regions from a more comprehensive perspective. In contrast, sparse-channel EEG can only focus on the local activities in a certain region of the brain, which may further reduce the representation acquisition ability of the emotion recognition model.

[0005] Figure 1It shows the performance change of emotion recognition when the same model is applied to dense channels and sparse channels. One way to improve the model's ability to capture key features is to focus on important temporal or spatial positions in EEG signals, usually through attention mechanisms such as channel-spatial attention or self-attention. The role of the attention mechanism is to dynamically adjust the attention weights of the model for different features. The premise of its effective application is that the model extracts valuable features from the input data. However, in sparse-channel EEG data, due to the limitation of spatial coverage, redundant features and noise may be more significant, which in turn affects the feature extraction of the model. Therefore, simply relying on the attention mechanism may not be able to fully compensate for the problems of insufficient spatial information and insufficient data caused by channel sparsity.

[0006] Traditional EEG emotion recognition is usually achieved by collecting or monitoring EEG signals of the whole brain. Although this method can provide relatively comprehensive EEG activity information, it has the following disadvantages:

[0007] (1) Device complexity: Full-brain EEG emotion recognition requires a large number of electrodes for signal collection, usually 32 or more electrodes to cover the entire head. This not only increases the hardware cost but also makes the device design more complex. At the same time, the installation and calibration process of the electrodes is cumbersome and it is difficult to use efficiently in non-laboratory environments, severely restricting the popularization and portable application of the technology.

[0008] (2) High data calculation cost: Since full-brain EEG signals contain a large amount of redundant information, high data dimensions, signal processing and model training require stronger computing power. This leads to greater difficulty in real-time processing and also places higher requirements on hardware resources, such as high-performance processors and large-capacity storage devices, thus increasing the overall operating cost of the system.

[0009] (3) Poor user acceptance: Traditional EEG devices usually require a head-mounted electrode cap, and conductive gel or electrolyte needs to be applied during use, which brings discomfort to users. In addition, a large number of electrodes and the complex wearing process will make users feel inconvenient, especially for users with emotional disorders such as depression, affecting their enthusiasm and acceptance for daily use. This is particularly obvious in the application of portable, home or wearable emotion monitoring devices. Summary of the Invention

[0010] In view of this, the present application provides a depressive population determination system based on sparse-channel EEG of the forehead, which can accurately and efficiently identify and determine the depressive risk state of users while only relying on a small amount of EEG signals of the forehead channels.

[0011] The technical solution to implement the present invention is as follows:

[0012] A depression population determination system based on sparse channel electroencephalogram of the forehead, comprising: a data acquisition module, a self-supervised representation learning module, a feature decoupling module, a feature dynamic interaction module, and a depression population determination module; wherein,

[0013] The data acquisition module is used to collect the electroencephalogram signal EEG of the person to be tested, and obtain the original electroencephalogram signal X aug ;

[0014] The self-supervised representation learning module is used to perform enhancement processing on the original electroencephalogram signal to obtain an enhanced electroencephalogram signal, and label the original electroencephalogram signal X ori and the enhanced electroencephalogram signal X aug ;

[0015] The feature decoupling module includes a common encoder and a private encoder, and the common encoder and the private encoder are used to decouple the adaptive feature ori and the invariant feature from the original electroencephalogram signal X and decouple the adaptive feature aug and the invariant feature from the enhanced electroencephalogram signal X

[0016] The feature dynamic interaction module, based on the cross-attention mechanism, realizes the dynamic information transfer and optimization between the invariant feature and the adaptive feature ;

[0017] The depression population determination module is used to realize the depression determination of the person to be tested according to the optimized features.

[0018] Optionally, the enhancement processing of the present invention includes at least one of jittering, scaling, inverting, and rearranging.

[0019] Optionally, the jittering in the present invention is: adding Gaussian noise with a set signal-to-noise ratio (SNR) to the original signal; the scaling is: setting a scaling factor to multiply with the original signal; the inverting is: multiplying the amplitude of the original signal by -1; the rearranging is: dividing the original signal into n segments, randomly shuffling them, and then rearranging them.

[0020] Optionally, when the common encoder and the private encoder of the present invention are trained, the loss function is:

[0021]

[0022] wherein, α, β, and γ are hyperparameters used to control the loss intensity, represents the ambiguity loss, represents the similarity loss, represents the Euclidean distance between features;

[0023] Ambiguity loss is: the Huber loss between the recombined features after decoupling and the original features;

[0024] Similarity loss used to make the invariant features in the original EEG signal X ori sufficiently similar to the invariant features in the enhanced EEG signal X while making the adaptive features in the original EEG signal X aug dissimilar to the adaptive features in the enhanced EEG signal X ; where, || represents the L2 norm. ori is used to minimize the distance between invariant features to enhance the similarity between them, optimize the spatial distribution of invariant features and adaptive features, and maintain feature separability. aug

[0025]

[0026] represents the Euclidean distance between features

[0027] Optionally, what is described in the present invention is:

[0027]

[0028] where, || represents the L2 norm.

[0029] Optionally, what is described in the present invention is:

[0030]

[0031] where represents aligning the coupled features with the original feature X m

[0032] Optionally, what is described in the present invention is:

[0033]

[0034] where, dist(·) represents the Euclidean distance.

[0035] Optionally, the processing of the adaptive features by the feature dynamic interaction module described in the present invention is:

[0036] First, calculate as query, key, and value values according to and ;

[0037]

[0038] Secondly, information interaction is realized for the adaptive features included in the original features and the enhanced features through the multi-head attention mechanism;

[0039]

[0040] Among them, represents the additional information of represents the additional information of, and d is the dimension of Q and K;

[0041] Finally, the supplementary information is concatenated with the original features to form m≠m′∈{ori,aug}.

[0042] Optionally, the processing of the invariant features by the feature dynamic interaction module of the present invention is as follows:

[0043]

[0044] m∈{ori,aug}

[0045] Among them, W Q and W K and W V represent the projection matrices of query, key, and value.

[0046] Optionally, the depressive population determination module of the present invention is used to receive the optimized features and calculate the probability distribution of the emotional category, and determine the depressive population based on the set determination rules.

[0047] Beneficial effects:

[0048] First, the representation ability of the model is improved through self-supervised tasks

[0049] The present invention introduces a self-supervised learning framework, uses self-supervised tasks to pre-train the model, and enhances the model's ability to extract effective features from sparse-channel electroencephalogram signals. This method can fully mine the potential information in the data, thereby improving the model's representation ability and the recognition accuracy of the depressive state.

[0050] Second, feature decoupling improves the generalization ability and anti-interference ability of the model

[0051] Most of the existing technologies directly extract and classify EEG data, unable to effectively distinguish stable features from changing features in specific scenarios, and are easily interfered by individual differences or environmental noises. By decoupling EEG signal features into invariant features and adaptive features, the present invention enables the model to capture core features related to the depressive state while having stronger environmental adaptability, thereby improving the accuracy and stability of the determination.

[0052] Third, feature optimization based on the dynamic interaction mechanism to enhance the information fusion ability

[0053] Traditional methods mostly use simple feature concatenation or weighted fusion, making it difficult to fully utilize the complementary information between multi-source features. Through the dynamic interaction mechanism, the present invention uses multi-head attention and shiftable attention to dynamically adjust the information flow between invariant features and adaptive features, making feature fusion more accurate, thereby effectively improving the determination performance of the model.

[0054] Fourth, lightweight model design applicable to sparse-channel EEG

[0055] Compared with traditional methods that rely on multi-channel EEG data, the present invention optimizes sparse-channel EEG signals, enabling the system to still obtain high-quality feature representations and accurate determination results for depressive populations even when using only a small amount of channel data. This design not only reduces the complexity of hardware devices but also improves the portability and applicability in practical applications.

[0056] Fifth, by combining feature decoupling, lightweight modeling, and the dynamic interaction mechanism of emotion features, the system significantly reduces the device complexity and computational burden while ensuring the recognition accuracy, improves the real-time performance of the system and the comfort of users wearing it, and is applicable to various portable scenarios such as wearable devices and home monitoring terminals. Brief Description of the Drawings

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0058] Figure 1 is the comparison of EEG emotion recognition results for dense channels and sparse channels;

[0059] Figure 2 is the schematic diagram of the decoupled feature interaction method proposed by the present invention;

[0060] Figure 3 is the self-supervised representation learning module proposed by the present invention;

[0061] Figure 4 The feature decoupling module proposed by the present invention;

[0062] Figure 5 The dynamic interaction module proposed by the present invention. Detailed implementation manners

[0063] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0064] It should be noted that, without conflict, the following embodiments and the features in the embodiments may be combined with each other; and, based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present disclosure.

[0065] It should be noted that the following description relates to various aspects of embodiments within the scope of the appended claims. It will be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is for illustrative purposes only. Based on the present disclosure, those skilled in the art should understand that one aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement a device and / or practice a method. Additionally, this device and / or this method may be implemented using other structures and / or functions in addition to one or more of the aspects described herein.

[0066] An embodiment of the present application is a depression population determination system based on frontal sparse channel electroencephalogram, as Figure 2 shown, including: a data acquisition module, a self-supervised representation learning module, a feature decoupling module, a feature dynamic interaction module, and a depression population determination module; wherein,

[0067] The data acquisition module is used to collect the electroencephalogram signal EEG of the person to be tested, and obtain the original electroencephalogram signal X ori ;

[0068] The self-supervised representation learning module is used to perform enhancement processing on the original electroencephalogram signal X aug , obtain the enhanced electroencephalogram signal, and label the original electroencephalogram signal X ori and the enhanced electroencephalogram signal X aug ;

[0069] The feature decoupling module includes a common encoder and a private encoder, and the common encoder and the private encoder are used to decouple the adaptation feature ori and the invariant feature from the original electroencephalogram signal X and from the enhanced electroencephalogram signal Xaug decouple adaptive features and invariant features

[0070] Feature Dynamic Interaction Module, based on the cross-attention mechanism, realizes the dynamic information transfer and optimization of invariant features and adaptive features ;

[0071] Depression Population Determination Module, used to determine the depression of the test subject according to the optimized features.

[0072] In the embodiment of the present application, the collected raw EEG signals in sparse channels are decoupled into invariant features and adaptive features, so that the system can capture the core features related to the depressive state and at the same time have stronger environmental adaptability, thereby improving the accuracy and stability of the determination.

[0073] As Figures 3 - 5 shown, each module in the above system will be described in detail below.

[0074] The data acquisition module of this embodiment collects a small amount of EEG signals from the frontal channels (such as Fp1, Fp2, AF3, AF4), and accurately and efficiently identifies and determines the depression risk state of the user.

[0075] As Figure 3 shown, for the self-supervised representation learning module in this embodiment, a data augmentation strategy is proposed to extract and explore invariant emotional features by generating diverse data representations and designing self-supervised tasks to expand the data space.

[0076] First, for the time features and patterns of the collected EEG signals, four data augmentation methods of jittering, scaling, inversion, and rearrangement are designed. The specific augmentation strategies are as follows:

[0077] Jittering: Add Gaussian noise with a specific signal-to-noise ratio (SNR) to the input data to increase the diversity of the data and the robustness of the model to noise. Scaling: Scale the signal by multiplying the window data by a scaling factor to achieve amplitude diversity of the signal. Inversion: Multiply the original EEG amplitude by -1 to invert the entire window signal to form a spatial inversion of the EEG sequence. Rearrangement: Divide the original EEG into n segments, randomly shuffle them, and then rearrange them.

[0078] Secondly, self-supervised tasks are used to learn the EEG representations under different augmentations, that is, a signal classification network is designed to distinguish between the original data and the augmented data. Let X represent the input EEG signal, which can be divided into two categories: the original signal X ori and the augmented signal X aug , and each input signal is assigned an automatic label L according to the following rules:

[0079]

[0080] Finally, a signal classification network is used to perform a self-supervised representation learning task, with the goal of identifying whether the input signal belongs to the original data or the augmented data.

[0081] This embodiment uses cross-entropy loss to learn the self-supervised model: f s ·[X ori , X aug → L, where f s is a signal classification network for classifying the input signal into the original signal and the augmented signal, and the classification formula is as follows:

[0082]

[0083] Among them, K is the number of categories in the self-supervised task (i.e., the original and the augmented, two types), is the predicted output label of the model f s , and y s is the true label of the self-supervised task.

[0084] As Figure 4 shown, the feature decoupling module of this embodiment considers two types of features: the original feature and the augmented feature. Extract low-level EEG signal features from the self-supervised task, that is, the emotion feature is represented as where m ∈ {ori, aug}, that is, X m = [X ori , X aug . This embodiment represents the EEG emotion feature as a combination of invariant features, adaptation features, and emotion-irrelevant factors, which is expressed as follows:

[0085]

[0086] Among them, X m = [X ori , X aug represents the emotion feature, represents the invariant feature, represents the adaptive feature, and E contains various random environmental perturbations, noises, and non-emotion information, that is, the combined features of emotion-irrelevant factors. The goal of the present invention is to make the model focus on and and use them to achieve better emotion recognition performance.

[0087] In order to decouple the emotion feature into the invariant feature and the adaptation feature the common encoder and the private encoder To display the features after predicted decoupling. Formally:

[0088]

[0089] Then, make the invariant features in the original EEG signal X ori be similar enough to the invariant features in the enhanced EEG signal X while making the adaptive features in the original EEG signal X aug be dissimilar to the adaptive features in the enhanced EEG signal X so that the model can focus on the typical features and reduce the interference of redundant features. ori <000018> aug <000022>

[0090] Use cosine similarity to perform the above operations:

[0091]

[0092] In addition, to avoid feature ambiguity, supervise the decoupled features in an autoregressive manner, recombine the decoupled and to obtain the coupled features where align the coupled features with the original feature X m

[0093] Finally, add an ambiguity loss to measure the difference between X m and :

[0094]

[0095] where represents the Huber loss.

[0096] For the finally extracted invariant features and adaptive features, supervise the separability and stability of the enhanced features through the Euclidean distance. Specifically, this module first enhances the distinguishability between the adaptive features to ensure the diversity within the adaptive features. Enhance the similarity between the invariant features by minimizing the distance between them. At the same time, optimize the spatial distinguishability between the invariant features and the adaptive features to maintain their separability. This module improves the model's fine-grained representation ability for emotional features by ensuring effective feature separation, specifically manifested as:

[0097]

[0098] where dist(·) represents the Euclidean distance and can be expressed as:

[0099] ​

[0100] Among them, v i and v j represent different features, and μ is used as a distance normalization factor, defined as the average Euclidean distance of all sample pairs in the batch.

[0101] Finally, the decoupling task loss is defined as follows:

[0102]

[0103] Among them, α, β, and γ are hyperparameters used to control the loss intensity. α and β are set to 1.0, and γ is set to 0.1.

[0104] As Figure 5 shown, the feature dynamic interaction module of this embodiment designs two branches to process the decoupled invariant features and adaptive features respectively, namely the invariant feature branch and the adaptive feature branch;

[0105] For the adaptive features: The representations of the adaptive features may vary in different augmentation types, but they are usually semantically consistent. This means that these features may exhibit different shapes or structures under different augmentations, but the emotional meanings they convey should be the same. For the adaptive features with different representations, cross-attention is used to unify and correlate these feature representations to achieve feature representation alignment. Specifically, and can be replaced by query, key, and value, and they can be expressed as:

[0106]

[0107] Among them, W q 、W k and W v are the weight matrices of the query, key, and value respectively.

[0108] Subsequently, the adaptive features contained in the original features and the augmented features achieve information interaction through the multi-head attention mechanism. In each attention head calculation, first, a form of feature is selected and converted into a set of key-value pairs, and the other form of feature is used as the query. In each attention head calculation space, first, the similarity between the query and the key is calculated, and then the obtained attention weight distribution is applied to the value vector to achieve the interaction between the adaptive features, and its expression form is:

[0109]

[0110] Among them, represents the additional information of denote Additional information, where d is the dimension of Q and K.

[0111] Then, the supplementary information is concatenated with the original features to form m≠m′∈{ori,aug}, and self-attention is used to capture the global dependencies in, further enhancing the effective association between different features.

[0112] In the above process, the model can integrate the feature representations under different enhancement forms from a global perspective to achieve better emotional expression modeling.

[0113] For invariant features: Embed the self-attention mechanism in the feature extraction step, make full use of its advantages in global modeling, and improve the sensitivity and robustness of the model to the emotional core features. Compared with traditional feature extraction methods that only rely on local neighborhood receptive fields, the self-attention mechanism can globally compare and weight the features at all positions in the input feature space, thereby highlighting the key factors most relevant to emotion recognition in the feature representation. Specifically:

[0114] Self-attention measures the correlation between features by calculating the dot product similarity between queries, keys, and values, and dynamically assigns attention weights to features through softmax normalization, which can be defined as:

[0115]

[0116] m∈{ori,aug}

[0117] where, W Q 、W K 、W V represent the projection matrices of query, key, and value.

[0118] This process enables the model to autonomously select and strengthen the most discriminative elements in the emotional signals, while effectively reducing the interference caused by random noise and local deformations brought by data augmentation. Since the essence of invariant features is to maintain a stable representation when processing various input forms, the self-attention mechanism ensures the stability and consistency of these core features under various transformation scenarios by integrating information from different positions and levels.

[0119] Combined with the above constraint loss, the final emotion state decision loss function of the self-supervised representation learning module, feature decoupling module, and feature dynamic interaction module using multi-class cross-entropy during training is:

[0120]

[0121] where, multi-class cross-entropy is the emotion classification loss. λ1 and λ2 are the balance factors for different constraints. In the experiments of this application, λ1 is set to 0.0001 and λ2 is set to 1.0.

[0122] A depression population determination module, configured to implement the depression determination of the to-be-tested person according to the optimized features.

[0123] The present invention obtains an optimized emotional feature representation through a feature dynamic interaction module, and combines multi-class cross-entropy for emotion state determination. In the depression population determination stage, first, the trained classification model is used to classify the fused features for the emotional state to determine whether the tested person has a depression risk. The model calculates the probability distribution of the emotion categories according to the input features, and obtains the final result based on the set determination rules (such as classification threshold or maximum probability category). When the prediction result indicates that the individual's emotional state is in the abnormal or highly depressive range, the individual's historical emotional data can be further combined for comprehensive evaluation to improve the accuracy and stability of depression risk identification.

[0124] An embodiment of this application is a depression population determination system based on frontal sparse-channel electroencephalogram:

[0125] First, the original data is enhanced multiple times, and self-supervised representation learning is performed through an auxiliary task. Then, the emotional features are divided into two parts, invariant features and adaptive features, including the original EEG signal and the enhanced EEG signal. The invariant features have stable emotional attributes, and the adaptive features have different manifestations but have similar or consistent emotional semantic information;

[0126] Secondly, in order to achieve feature decoupling, the original data and the enhanced data are used as sample pairs, and the invariant feature encoder and the adaptive feature encoder are respectively used for decoupling. To reduce the ambiguity problem of the decoupled features, an autoregressive mechanism is introduced to supervise the decoupled features.

[0127] Thirdly, in order to enhance the decoupling ability, the MSE and cosine similarity difference losses are added to strengthen the separability of different features.

[0128] Finally, the decoupled features improve the emotion recognition performance through self-attention and cross-feature interaction.

[0129] The embodiment of this application has the following effects;

[0130] (1) Reduce the device complexity

[0131] By only collecting the EEG signals of the frontal leads (FP1, FP2, FPZ), the number of required electrodes is reduced, the device structure is significantly simplified, and the portability and operation convenience of the system are improved.

[0132] (2) Reduce the data calculation cost

[0133] Through the feature decoupling method, data redundancy is effectively reduced, and only the key information related to emotions is retained, thereby reducing the computational complexity of data processing and model training, and achieving real-time and efficient emotion recognition.

[0134] (3) Improve user acceptance

[0135] This method is applicable to EEG acquisition devices with sparse channels, which are more portable and comfortable, without the need for cumbersome wearing and conductive gel application, greatly improving the user experience and enhancing the usability and acceptance in daily emotion monitoring.

[0136] In summary, the above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A depression population determination system based on sparse-channel electroencephalogram of the forehead, characterized in that Including: a data acquisition module, a self-supervised representation learning module, a feature decoupling module, a feature dynamic interaction module, and a depression population determination module; wherein A data acquisition module, which is used to collect the electroencephalogram (EEG) signals of the person to be tested, so as to obtain the original EEG signal X aug ; A self-supervised representation learning module for enhancing the original EEG signal to obtain an enhanced EEG signal and labeling the original EEG signal X ori and the enhanced EEG signal X aug with labels; The feature decoupling module includes a common encoder and a private encoder, and the common encoder and the private encoder are used to decouple the adaptation features ori and the invariant features from the original EEG signal X Decouple the adaptation features aug and the invariant features from the enhanced EEG signal X The feature dynamic interaction module, based on the cross-attention mechanism, realizes the dynamic information transfer and optimization of invariant features and adaptable features ; the depression population determination module is configured to implement the depression determination of the to-be-tested person according to the optimized features.

2. The depressive population determination system based on frontal sparse channel electroencephalogram according to claim 1, wherein The enhancement processing includes at least one of jittering, scaling, inverting, and rearranging.

3. The depressive population determination system based on frontal sparse channel electroencephalogram according to claim 2, wherein The jittering is: adding Gaussian noise with a set signal-to-noise ratio (SNR) to the original signal; the scaling is: setting a scaling factor to multiply with the original signal; the inverting is: multiplying the amplitude of the original signal by -1; the rearranging is: dividing the original signal into n segments, randomly shuffling them, and then rearranging them.

4. The depression population determination system based on frontal sparse channel electroencephalogram according to claim 1, characterized in that, When the common encoder and the private encoder are trained, the loss function is: where α, β, and γ are hyperparameters for controlling the loss intensity, represents the ambiguity loss, represents the similarity loss, represents the Euclidean distance between features; Ambiguity loss is: the Huber loss between the recombined features after decoupling and the original features; Similarity loss to make the invariant features in the original EEG signal X ori sufficiently similar to the invariant features in the enhanced EEG signal X while making the adaptive features in the original EEG signal X aug dissimilar to the adaptive features in the enhanced EEG signal X such that the original EEG signal X ori has its adaptive features dissimilar to those in the enhanced EEG signal X aug ; dissimilar Indicates the Euclidean distance between features It is used to minimize the distance between invariant features to enhance the similarity between them, optimize the spatial distribution of invariant features and adaptive features, and maintain the separability of features.

5. The depressive population determination system based on frontal sparse channel electroencephalogram according to claim 4, characterized in that The said is as follows: wherein, ‖‖ represents the L2 norm.

6. The depressive population determination system based on frontal sparse channel electroencephalogram according to claim 4, wherein The said is as follows: Among them, indicates aligning the coupling feature with the original feature X m in alignment.

7. The depressive population determination system based on frontal sparse channel electroencephalogram according to claim 4, characterized in that, The said is as follows: wherein, dist(·) represents the Euclidean distance.

8. The depressive population determination system based on frontal sparse-channel electroencephalogram according to claim 1, wherein The processing of the feature dynamic interaction module for the adaptive features is: First, calculate the query, key, and value values according to and ; Secondly, the adaptive features included in the original features and the enhanced features are subjected to information interaction through the multi-head attention mechanism; Among them, represents additional information of represents additional information of, where d is the dimension of Q and K; Finally, the supplementary information is concatenated with the original features to form m≠m′∈{ori,aug}.

9. The depressive population determination system based on frontal sparse-channel electroencephalogram according to claim 8, characterized in that, The processing of the feature dynamic interaction module for the invariant features is: m ∈ {ori, aug} Among them, W Q , W K , W V represent the projection matrices of query, key, and value.

10. The depressive population determination system based on frontal sparse channel electroencephalogram according to claim 9, characterized in that, The depression population determination module is used to receive the optimized features and calculate the probability distribution of emotional categories, and determine the depression population based on the set determination rules.

Images