Depression emotion state grading early warning method and system based on brain-computer interface technology

Abstract

The invention discloses a depressive emotion state grading early warning method and system based on a brain-computer interface technology, and the method comprises the steps: firstly carrying out signal collection and preprocessing, i.e., obtaining EEG signal data through a local storage or user uploading interface, and carrying out the digital filtering and standardization processing of an original EEG signal in sequence; then deep learning model training and analysis are carried out, the preprocessed EEG signals are input into a pre-trained deep learning model, spatial-temporal features are extracted through a multi-modal fusion network architecture, and probability distribution is output; and finally, carrying out graded early warning: judging the current emotional state according to the highest probability category output by the model, and analyzing and generating trend early warning in combination with historical data. Through multi-module collaborative design, deep learning algorithm application and introduction of a brain-computer interface technology, remarkable beneficial effects are achieved in the aspects of depression diagnosis precision, efficiency, safety and personalized intervention, and the method has wide application prospects and social value.

Description

Technical Field

[0001] The present invention relates to the fields of brain-computer interface technology, deep learning algorithm and mental health management technology, and specifically to a depressive emotional state grading early warning method and system based on brain-computer interface technology. Background Art

[0002] Depression is a common mental disorder. Existing methods for diagnosing depression mainly rely on clinical interviews and psychometric tools, such as the Hamilton Depression Rating Scale (HDRS) and the Beck Depression Inventory (BDI). Although these traditional methods are widely used in clinical and research fields, they have obvious limitations, including strong subjectivity, susceptibility to the patient's condition, and time-consuming, which leads to inaccurate diagnostic results.

[0003] In recent years, biomarker-based depression diagnosis methods have gradually attracted attention. These methods usually infer the depressive state by analyzing physiological parameters such as electroencephalogram (EEG) and heart rate variability (HRV). For example, some studies use deep learning models to analyze EEG signals and extract features related to depression. For example, the application of convolutional neural networks (CNN) in the diagnosis of depression based on EEG signal classification. In addition, text analysis based on natural language processing (NLP) technology is also used to identify the language patterns of patients with depression, such as detecting depressive tendencies by analyzing patients' social media or daily conversations. There are also studies that use facial recognition technology to analyze changes in patients' facial expressions and identify non-verbal behaviors unique to patients with depression.

[0004] Although the above methods have made some progress in the automated detection of depression, they still have the following major shortcomings: (1) Limitations: Most studies still focus on identifying basic depressive states rather than specifically targeting the complex characteristics of depression. This results in limited diagnostic generalization and is unable to fully reflect the diversity of depression.

[0005] (2) Individual differences: Existing methods often ignore the diversity and individual differences of depression symptoms, resulting in insufficient accuracy of diagnostic results.

[0006] (3) Lack of stratified diagnosis: Existing biomarker technologies mainly focus on determining the presence of depression, but lack stratified diagnosis of the severity of depression, which limits the possibility of precise treatment and management of patients with depression.

[0007] (4) Data quality and quantity: It is difficult to obtain high-quality EEG data, and the amount of data is limited, which affects the training effect and generalization ability of the model.

[0008] (5) Technology maturity: Although some depression detection products based on these research results have appeared on the market, such as smartphone-based applications, wearable devices, and online psychological assessment tools, most of these products are still in the early stages of development and their actual application effects remain to be verified. Summary of the invention

[0009] In view of the deficiencies in the prior art, the present invention discloses a depressive emotional state grading warning method and system based on brain-computer interface technology to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solution: a depressive emotional state grading early warning method based on brain-computer interface technology, comprising the following steps: S1, signal acquisition and preprocessing: obtain EEG signal data through local storage or user upload interface, and perform digital filtering, standardization and independent component analysis on the original EEG signal in sequence; S2. Deep learning model training and analysis: The preprocessed EEG signal is input into the pre-trained deep learning model, and the spatiotemporal features are extracted through the multimodal fusion network architecture, and the three-category probability distribution including health, potential depression and depression state is output; S3. Gradual warning: Determine the current emotional state based on the highest probability category output by the model, and generate trend warnings based on historical data analysis.

[0011] Preferably, in step S1, the signal acquisition and preprocessing step specifically includes: S11, remove power supply noise and environmental interference through a 0.5-50Hz bandpass filter; S12. StandardScaler is used to standardize each EEG channel, setting the mean of each channel to 0 and the standard deviation to 1. The standardization formula is:

[0012] in, is each data point in the original EEG signal; is the channel mean of the feature; is the channel standard deviation of the feature; S13. Independent component analysis (ICA) is used to separate and remove artifact components in the EEG signal. Its linear mixed model is expressed as:

[0013]

[0014] Where X∈R N×T is the observation signal matrix, N is the number of channels, T is the number of sampling points; A∈RN×M is the mixing matrix; S∈R M×T is the independent component matrix, M is the number of independent components; The artifact components representing eye movement and electromyography are automatically identified through kurtosis analysis, and the corresponding column vectors are removed from the mixing matrix A to reconstruct the signal; S14, use spherical spline interpolation to repair bad data, and for channels marked as abnormal k , its interpolation voltage V k Through adjacent n Calculation of spatial weights for effective channels:

[0015] in, w i For Channel i The interpolation weights, θ i Bad conductor and effective channel i The spherical angle between c i are the harmonic coefficients solved by Laplace's equation; S15, perform baseline correction to eliminate the DC offset of the signal, and select the [-200,0]ms time window before the stimulus for each trial to calculate the baseline mean μ b , the correction formula is:

[0016] in, X ( t ) is the original potential value;

[0017] μ b is the mean value of the baseline period, T b represents the length of the baseline time window, and t 1 corresponds to the time point -200 ms before stimulation, t 2 corresponds to the time point 0 ms before stimulation.

[0018] Preferably, the preprocessed EEG signal data is trained using a five-fold cross-validation strategy, including: Divide the dataset into training and validation sets; Use the training set to train the model and save the best model weights; The validation set was used to evaluate the model performance and calculate the accuracy, precision, recall, and F1 score.

[0019] Preferably, in step S2, the multimodal fusion network architecture includes: The convolution module consists of three one-dimensional convolutional layers, each with 64, 128, and 256 filters of length 3, followed by a batch normalization layer and a maximum pooling layer. The timing analysis module includes a bidirectional LSTM layer group, the first layer is configured with 128 units and returns the sequence output, and the second layer is configured with 64 units to return the final state output; the timing analysis module adopts a bidirectional LSTM structure, and simultaneously applies 0.3 input dropout and recursive dropout in the first layer LSTM; The classification module consists of two fully connected layers and an output layer, and finally generates a three-dimensional probability vector through the softmax function.

[0020] Preferably, the convolution module is specifically implemented as follows: The first convolutional layer uses L2 regularization coefficient λ=0.01, followed by a batch normalization layer with momentum parameter 0.99 and ε=1e-5; The second convolutional layer uses an L2 regularization coefficient of λ=0.01, followed by a batch normalization layer with a momentum parameter of 0.99 and ε=1e-5; followed by a maximum pooling layer with a window length of 2; The third convolutional layer uses L2 regularization coefficient λ=0.01, followed by a batch normalization layer with momentum parameter 0.99 and ε=1e-5; followed by a maximum pooling layer with a window length of 2; A dropout layer is added after the entire convolution with a dropout rate of 0.5.

[0021] Preferably, the model training process adopts: a) Sparse classification cross entropy loss function combined with L2 regularization term; b) Adam optimizer configures the initial learning rate to 5e-4, β1=0.9, β2=0.999; c) The early stopping mechanism monitors the validation loss and terminates the training if there is no improvement for 15 consecutive rounds; d) Dynamic learning rate: when the validation loss stagnates, the learning rate decays by a factor of 0.5, with a minimum learning rate of 1e-6.

[0022] Preferably, in step S3, the graded warning step includes: 1) Real-time display of three-dimensional probability distribution and classification results corresponding to the maximum probability, including three categories: health, potential depression, and depression; 2) Predict the depression risk level in a specified future period based on time series analysis.

[0023] Preferably, the method implements user interaction through a web page interface, supports standardized formats of bioelectric signals, such as uploading of EEG data files in .npy format, visualization of processing results, and push notification of early warnings.

[0024] The present invention also provides a depressive emotional state grading warning system based on brain-computer interface technology, comprising the following modules: Data acquisition module, used to obtain the user's local or uploaded EEG raw data; Data preprocessing module, used to filter the collected EEG data and analyze and process the independent components; Feature extraction module, which uses deep learning algorithms to extract EEG spatiotemporal features from preprocessed data; The state classification module classifies the collected data into depressive emotional states based on the trained deep learning model; The early warning module is used to display classification results and generate real-time early warning signals.

[0025] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention improves diagnostic accuracy and intelligence. By introducing deep learning algorithms (such as convolutional neural networks (CNNs) and long short-term memory (LSTM) neural networks (LSTMs) to process EEG signals, the accuracy of depression diagnosis has been significantly improved, and the subjectivity and errors of traditional methods have been reduced. At the same time, the system has achieved full process automation from data collection, preprocessing, feature extraction, model training to early warning output, reducing human intervention and improving diagnostic efficiency and reliability.

[0026] 2. The present invention realizes hierarchical diagnosis and early warning, which can not only identify the existence of depression, but also perform hierarchical diagnosis on people at risk of the disease and normal people. By providing depression warning information to people at risk of the disease, early intervention is achieved to prevent the occurrence and development of depression, providing important support for clinical diagnosis and mental health management.

[0027] 3. The present invention uses non-invasive EEG acquisition equipment and combines brain-computer interface technology to ensure safety and convenience during use, providing a new technical approach for early identification of depression. The system can be applied to a variety of scenarios such as clinical diagnosis and psychological counseling, providing instant diagnostic results and trend predictions, greatly facilitating patients' self-management and doctors' professional guidance.

[0028] 4. The present invention supports data collection and model updating based on the cloud platform, and has the potential for large-scale promotion and application. Its openness and flexibility provide possibilities for future technology upgrades and functional expansion. In addition, the system combines the patient's historical data to not only assess the current emotional state, but also predict the development trend of the disease, and provide customized intervention suggestions based on individual circumstances, which helps to achieve personalized treatment management. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0030] In the attached picture: Figure 1 This is a schematic diagram of a module of a depressive mood state grading warning system based on brain-computer interface technology of the present invention; Figure 2 This is a schematic diagram of a process flow of a depressive mood state classification early warning method based on brain-computer interface technology of the present invention; Numbers in the figure: 01, data acquisition module; 02, data preprocessing module; 03, feature extraction module; 04, state classification module; 05, warning output module. DETAILED DESCRIPTION

[0031] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0032] Implementation example Figure 1 FIG. 1 is a schematic diagram of a depressive mood state grading warning system based on brain-computer interface technology provided by the present invention, which includes the following modules: Data acquisition module, used to obtain the user's local or uploaded EEG raw data; Data preprocessing module, used to filter the collected EEG data and analyze and process the independent components; Feature extraction module, which uses deep learning algorithms to extract EEG spatiotemporal features from preprocessed data; The state classification module classifies the collected data into depressive emotional states based on the trained deep learning model; The early warning module is used to display classification results and generate real-time early warning signals.

[0033] like Figure 2 As shown, the present invention provides a depressive emotional state classification warning method based on brain-computer interface technology, which specifically includes the following steps: Step 1: Signal acquisition and preprocessing (normalization and digital filtering) 1) Signal acquisition: This system supports reading EEG (electroencephalogram) .npy format data files stored locally. The system provides a user upload interface, and users can upload the collected EEG data through the web page, or select the data file to be processed from the system local directory.

[0034] 2) Preprocessing: Digital filtering: Apply digital filters to the raw EEG signal to mainly remove power supply noise and other unnecessary high or low frequency components. Use a bandpass filter with a filtering range of 0.5-50 Hz to retain the core frequency band of the EEG and remove 50Hz power supply noise and environmental noise.

[0035] Standardization: Use StandardScaler to standardize the EEG signal. The standardization operation sets the mean of each feature (each channel of the EEG signal) to 0 and the standard deviation to 1. The converted data is on the same scale, which helps the training of the deep learning model. The standardization formula is: in: is each data point in the original EEG signal; is the mean of the feature channel; is the standard deviation of the feature channel.

[0036] 3) Independent component analysis is used to separate and remove artifact components in the EEG signal. Its linear mixed model is expressed as:

[0037]

[0038] Where X∈R N×T is the observation signal matrix, N is the number of channels, T is the number of sampling points; A∈R N×M is the mixing matrix; S∈R M×T is the independent component matrix, M is the number of independent components; The artifact components representing eye movement and electromyography are automatically identified through kurtosis analysis, and the corresponding column vectors are removed from the mixing matrix A to reconstruct the signal; Spherical spline interpolation is used to repair bad data and the channels marked as abnormal k , its interpolation voltage V k Through adjacent n Calculation of spatial weights for effective channels:

[0039] in, w i For Channel i The interpolation weights, θ i Bad conductor and effective channel i The spherical angle between c i are the harmonic coefficients solved by Laplace's equation; Baseline correction was performed to eliminate the DC offset of the signal. For each trial, the [-200, 0] ms time window before stimulation was selected to calculate the baseline mean. μ b , the correction formula is:

[0040] in, X ( t ) is the original potential value;

[0041] μ b is the mean value of the baseline period, T b represents the length of the baseline time window, and t 1 corresponds to the time point -200 ms before stimulation, t 2 corresponds to the time point 0 ms before stimulation.

[0042] Technical effect: This preprocessing method removes unnecessary noise and ensures data consistency through standardization. The standardized data can improve the convergence speed and accuracy of the deep learning model during training and effectively improve the model performance.

[0043] Step 2: Deep learning model training and analysis 1) Data input: The system inputs the preprocessed EEG signal into the pre-trained deep learning model for classification analysis.

[0044] 2) Training process: Convolutional Neural Network (CNN): This model uses a convolutional neural network (CNN) to extract the spatial features of EEG signals. It has multiple convolutional and pooling layers and can effectively identify spatiotemporal patterns in the signal.

[0045] Long Short-Term Memory Network (BiLSTM): This model also incorporates LSTM layers to process the temporal characteristics of the signal and capture long-term temporal dependencies in the EEG signal. Through bidirectional LSTM, the model is able to analyze information from past and future time steps to improve classification accuracy.

[0046] Output layer: The last layer of the model is a softmax layer, which is used to output the probabilities of the three categories (healthy, potentially depressed, depressed). Finally, the classification is performed based on the category with the largest probability value.

[0047] 3) Multimodal fusion network architecture: a. Input layer: b. Convolutional module (CNN): First convolutional layer: 64 filters of length 3 (kernel_size=3), ReLU activation function, L2 regularization (λ=0.01). Batch Normalization layer: used to accelerate the training process and stabilize the performance of deep networks, momentum parameter 0.99, ε=1e-5. MaxPooling1D layer: uses a pooling window of length 2 (pool_size=2), and the step size is equal to the pooling window size by default, further reducing the amount of data in the time dimension.

[0048] The second convolutional layer: 128 filters of length 3, ReLU activation function, L2 regularization (λ=0.01), Batch Normalization layer (BatchNormalization): used to accelerate the training process and stabilize the performance of deep networks, momentum parameter 0.99, ε=1e-5. The maximum pooling layer (MaxPooling1D): uses a pooling window of length 2 (pool_size=2), and the step size is equal to the pooling window size by default, further reducing the amount of data in the time dimension.

[0049] The third convolutional layer: 256 filters of length 3, ReLU activation function, L2 regularization (λ=0.01), Batch Normalization layer (BatchNormalization): used to accelerate the training process and stabilize the performance of deep networks, momentum parameter 0.99, ε=1e-5. The maximum pooling layer (MaxPooling1D): uses a pooling window of length 2 (pool_size=2), and the step size is equal to the pooling window size by default, further reducing the amount of data in the time dimension; followed by a Dropout layer (0.5) to prevent overfitting.

[0050] c. Timing analysis module (BiLSTM): Bidirectional LSTM layer: contains 128 units, returns sequence output (return_sequences=True), and applies dropout (0.3) and recursive dropout (0.3) at each time step to further avoid overfitting.

[0051] Bidirectional LSTM layer: contains 64 units, and applies dropout (0.3) and recursive dropout (0.3) at each time step to further avoid overfitting; does not return sequence output (return_sequences=False), and only produces output at the last time step.

[0052] d. Classification module: Fully connected layer: has 128 nodes, ReLU activation function, L2 regularization (λ=0.01).

[0053] Dropout layer: set to 0.5 to reduce the risk of overfitting.

[0054] Fully connected layer: has 64 nodes, ReLU activation function, L2 regularization (λ=0.01).

[0055] Output layer: Softmax activation, output three-dimensional probability vector [P(healthy), P(potential depression), P(depression)].

[0056] e. Optimize the training process: Loss function: sparse_categorical_crossentropy combined with L2 regularization.

[0057] Optimizer: Adaptive Moment Estimation (Adam), initial learning rate 5e-4, β1=0.9, β2=0.999.

[0058] Regularization strategy: Dropout (0.5) + weight decay (decay = 1e-4).

[0059] Early stopping mechanism: monitor the validation loss and terminate the training if there is no improvement for 15 consecutive rounds.

[0060] Dynamic learning rate: When validation loss stagnates, the learning rate decays by a factor of 0.5 (minimum 1e-6).

[0061] 4) K-fold cross validation: K-fold cross validation (K=5) was used to train and evaluate the model.

[0062] Divide the dataset into training and validation sets.

[0063] Train the model using the training set and save the best model weights.

[0064] The validation set was used to evaluate the model performance and calculate the accuracy, precision, recall, and F1 score.

[0065] 5) Training technology effect: The deep learning model can accurately classify different types of EEG signals and accurately identify potential depressive symptoms based on EEG features. The trained model can achieve a diagnostic accuracy of more than 97% in clinical applications.

[0066] Step 3: Graded warning and intervention 1) Web interface: The system provides users with a friendly web interface, where users can upload their own EEG signal data files and view the analysis results on the interface. The system will display classification information based on the results predicted by the model, including three categories: "healthy", "potential depression" and "depression".

[0067] The interface displays the probability values ​​of the classification, helping users understand the model's confidence in each category.

[0068] 2) Early warning mechanism: Real-time warning: The system classifies the patient's current status based on the model and predicts the trend of the disease through historical data analysis. Combined with the patient's historical data, the system can warn of the risk of depression in the next few weeks or months.

[0069] Personalized intervention: Based on the classification results and predicted disease development trends, the system provides personalized intervention plans, such as recommending regular psychological counseling, drug treatment, or lifestyle adjustments.

[0070] 3) Technical effects: Through real-time classification and early warning mechanisms, the system can help users identify depressive symptoms early and provide personalized intervention measures to slow down the deterioration of the disease and improve the patient's quality of life.

[0071] The specific workflow includes: first, the user logs in to the system and wears a brain-computer interface device to collect the user's brain wave data and transmit it to the system. The system preprocesses the received data and performs filtering and independent component analysis. The preprocessed data is sent to the feature extraction module to extract key features related to depression. The feature data is input into the pre-trained model to classify the depressive mood state. The classification results are processed by the early warning output module. If it is found that the user's depressive mood state reaches the early warning standard, an early warning message is sent to the user. After receiving the early warning message, the user can seek professional psychological counseling or treatment.

[0072] Through multi-module collaborative design, application of deep learning algorithms and introduction of brain-computer interface technology, the present invention has achieved remarkable beneficial effects in the accuracy, efficiency, safety and personalized intervention of depression diagnosis, and has broad application prospects and social value.

[0073] Finally, it should be noted that the above description is only a preferred example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A depressive mood state grading warning method based on brain-computer interface technology, characterized in that: The following steps are involved: S1, signal acquisition and preprocessing: obtain EEG signal data through local storage or user upload interface, and perform digital filtering, standardization and independent component analysis on the original EEG signal in sequence; S2. Deep learning model training and analysis: The preprocessed EEG signal is input into the pre-trained deep learning model, and the spatiotemporal features are extracted through the multimodal fusion network architecture, and the three-category probability distribution including health, potential depression and depression state is output; S3. Gradual warning: Determine the current emotional state based on the highest probability category output by the model, and generate trend warnings based on historical data analysis.

2. According to claim 1, a depressive emotional state classification warning method based on brain-computer interface technology is characterized in that: In step S1, the signal acquisition and preprocessing steps specifically include: S11, remove power supply noise and environmental interference through a 0.5-50Hz bandpass filter; S12. StandardScaler is used to standardize each EEG channel, setting the mean of each channel to 0 and the standard deviation to 1. The standardization formula is: ； in, is each data point in the original EEG signal; is the channel mean of the feature; is the channel standard deviation of the feature; S13. Independent component analysis is used to separate and remove artifact components in the EEG signal. Its linear mixed model is expressed as: ； Where X∈R N×T is the observation signal matrix, N is the number of channels, T is the number of sampling points; A∈R N×M is the mixing matrix; S∈R M×T is the independent component matrix, M is the number of independent components; The artifact components representing eye movement and electromyography are automatically identified through kurtosis analysis, and the corresponding column vectors are removed from the mixing matrix A to reconstruct the signal; S14, use spherical spline interpolation to repair bad data, and for channels marked as abnormal k , its interpolation voltage V k Through adjacent n Calculation of spatial weights for effective channels: ； in, w i For Channel i The interpolation weights, θ i Bad conductor and effective channel i The spherical angle between c i are the harmonic coefficients solved by Laplace's equation; S15, perform baseline correction to eliminate the DC offset of the signal, and select the [-200,0]ms time window before the stimulus for each trial to calculate the baseline mean μ b , the correction formula is: ； in, X ( t ) is the original potential value; ； μ b is the mean value of the baseline period, T b represents the length of the baseline time window, and t 1 corresponds to the time point -200 ms before stimulation, t 2 corresponds to the time point 0 ms before stimulation.

3. According to claim 1, a depressive emotional state classification warning method based on brain-computer interface technology is characterized in that: The preprocessed EEG signal data is trained using a five-fold cross-validation strategy, including: Divide the dataset into training and validation sets; Use the training set to train the model and save the best model weights; The validation set was used to evaluate the model performance and calculate the accuracy, precision, recall, and F1 score.

4. According to claim 1, a depressive emotional state classification warning method based on brain-computer interface technology is characterized in that: In step S2, the multimodal fusion network architecture includes: The convolution module consists of three one-dimensional convolutional layers, each with 64, 128, and 256 filters of length 3, followed by a batch normalization layer and a maximum pooling layer. The timing analysis module includes a bidirectional LSTM layer group, the first layer is configured with 128 units and returns the sequence output, and the second layer is configured with 64 units to return the final state output; the timing analysis module adopts a bidirectional LSTM structure, and simultaneously applies 0.3 input dropout and recursive dropout in the first layer LSTM; The classification module consists of two fully connected layers and an output layer, and finally generates a three-dimensional probability vector through the softmax function.

5. The method for early warning of depression mood state based on brain-computer interface technology according to claim 4, characterized in that: The convolution module is specifically implemented as follows: The first convolutional layer uses L2 regularization coefficient λ=0.01, followed by a batch normalization layer with momentum parameter 0.99 and ε=1e-5; The second convolutional layer uses an L2 regularization coefficient of λ=0.01, followed by a batch normalization layer with a momentum parameter of 0.99 and ε=1e-5; followed by a maximum pooling layer with a window length of 2; The third convolutional layer uses L2 regularization coefficient λ=0.01, followed by a batch normalization layer with momentum parameter 0.99 and ε=1e-5; followed by a maximum pooling layer with a window length of 2; A dropout layer is added after the entire convolution with a dropout rate of 0.

5.

6. The method for early warning of depression mood status based on brain-computer interface technology according to claim 1, characterized in that: The model training process uses: a) Sparse classification cross entropy loss function combined with L2 regularization term; b) Adam optimizer configures the initial learning rate to 5e-4, β1=0.9, β2=0.999; c) The early stopping mechanism monitors the validation loss and terminates the training if there is no improvement for 15 consecutive rounds; d) Dynamic learning rate: when the validation loss stagnates, the learning rate decays by a factor of 0.5, with a minimum learning rate of 1e-6.

7. The method for early warning of depression mood status based on brain-computer interface technology according to claim 1, characterized in that: In step S3, the hierarchical warning step includes: 1) Real-time display of three-dimensional probability distribution and classification results corresponding to the maximum probability, including three categories: health, potential depression, and depression; 2) Predict the depression risk level in a specified future period based on time series analysis.

8. The method for early warning of depression mood status based on brain-computer interface technology according to claim 1, characterized in that: The method realizes user interaction through a web page interface, and supports uploading of EEG data files in a standardized format of bioelectric signals, visual display of processing results, and push notification of early warnings.

9. A depressive mood state grading warning system based on brain-computer interface technology, characterized in that: Includes the following modules: Data acquisition module, used to obtain the user's local or uploaded EEG raw data; Data preprocessing module, used to filter the collected EEG data and analyze and process the independent components; Feature extraction module, which uses deep learning algorithms to extract EEG spatiotemporal features from preprocessed data; The state classification module classifies the collected data into depressive emotional states based on the trained deep learning model; The early warning module is used to display classification results and generate real-time early warning signals.

Images