Negative emotion classification and depression discrimination method based on electroencephalogram signals

Abstract

The invention discloses a negative emotion classification and depression discrimination method based on electroencephalogram signals, and belongs to the technical field of electroencephalogram signal processing, and the method comprises the following steps: 1, obtaining an emotion Deap data set and an HUSM depression data set through network application, and processing the emotion Deap data set based on an emotion VA two-dimensional model to retain negative emotion in the emotion Deap data set; 2, preprocessing the original electroencephalogram signals in the two data sets; 4, carrying out feature parameter screening and dimension reduction on the multi-domain feature fusion matrix; 5, carrying out negative emotion threshold analysis; 6, training a depression SVM classifier; 7, carrying out negative emotion classification on electroencephalogram signals of emotions to be analyzed according to a threshold combination in the step 5, and carrying out negative emotion classification on the electroencephalogram signals of the emotions to be analyzed; and judging whether the electroencephalogram signal of the emotion to be analyzed is depression or not according to the trained depression SVM classifier in the step 6.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram signal processing, in particular to the field of emotion recognition based on electroencephalogram signals, and specifically is a method for classifying negative emotions and distinguishing depression based on electroencephalogram signals. Background Art

[0002] Emotion is a complex psychological state, including multiple dimensions of physiology, behavior, cognition and subjective experience, and occupies an important position in daily life and learning. Psychology divides human emotions into three categories: positive emotions, neutral emotions and negative emotions. Emotion recognition refers to the process of feature extraction and pattern recognition of physiological or non-physiological signals under different emotions based on signal processing. Research on emotion cognition and its brain mechanism, especially the relationship between emotion and cognition and its potential neural basis, has gradually become a frontier hot topic in cognitive neuroscience in recent years. Its research results can not only greatly enrich the content of cognitive neuroscience, but also provide a basis and method for the diagnosis and treatment of emotional disorders and brain diseases.

[0003] Currently, depression identification is often based on subjective scales, questionnaires, etc., but these methods cannot be objective and accurate. With the rise of machine learning and the launch of the China Brain Project, more and more scholars are focusing on brain science research and viewing it as the "ultimate frontier". EEG signals have also been proven to be one of the accurate and efficient physiological bases for studying brain diseases. By preprocessing, extracting and selecting features of EEG signals, and using machine learning classification algorithms, depression can be identified, thereby accurately analyzing the subject's state.

[0004] Emotion recognition based on EEG signals mainly considers two aspects: recognition accuracy and recognition status. Current research is limited to binary classification of positive and negative emotions, which are relatively easy to distinguish, and few studies focus on a specific internal emotion. These emotions can also be divided into different levels and can be expressed through certain features or thresholds of certain features. In addition, the effective information contained in a single feature is not rich enough to achieve a good classification and recognition effect. Therefore, it is necessary to extract multiple electrode channel parameters of multiple EEG features to obtain a higher recognition accuracy. However, too high a feature dimension will lead to a large redundancy in the model, greatly increase the computational complexity and affect the classification and recognition accuracy to a certain extent. It is easy to produce overfitting phenomenon, resulting in poor generalization performance of the model. Summary of the invention

[0005] In order to solve the problems that the existing classification of emotions is too limited and the early diagnosis of depression is not accurate enough, the present invention proposes a method for classifying negative emotions and distinguishing depression based on EEG signals. The present invention starts with negative emotions and further classifies negative emotions through the ANOVA test method into three categories: severe negative emotions, moderate negative emotions and mild negative emotions. The relationship between severe negative emotions and depression brain diseases is then studied to determine whether the subject's condition has reached the standard for illness, thereby achieving efficient recognition effects.

[0006] The present invention proposes a method for classifying negative emotions and distinguishing depression based on EEG signals, comprising the following steps:

[0007] Step 1: Obtain EEG signal data: Obtain the emotion Deap dataset and HUSM depression dataset through the Internet, and process the emotion Deap dataset based on the emotion VA two-dimensional model to retain the negative emotions in it;

[0008] Step 2: Preprocess the EEG signals in the two data sets;

[0009] EEG signals are easily affected by environmental noise and physical magnetic fields. In order to obtain clean and effective EEG signals with high signal-to-noise ratio, the original mixed EEG signals need to be preprocessed. The original mixed EEG signals in the data set are preprocessed using Matlab software and the EEGLAB toolbox that comes with Matlab.

[0010] Step 3, feature extraction and construction of multi-domain feature fusion matrix: For the two data sets in step 2, extract three features from each EEG signal: the PSD ratio of α wave to β wave, the left-right brain asymmetry of α wave PSD, and the nonlinear feature sample entropy; fuse the feature parameters extracted from the two data sets to construct a multi-domain feature fusion matrix.

[0011] Step 4: Perform feature parameter screening and dimensionality reduction on the multi-domain feature fusion matrix: Select feature parameters of the multi-domain feature fusion matrix obtained in step 3, and remove feature parameters that do not have significant differences;

[0012] Step 5: Negative emotion threshold analysis:

[0013] The negative emotion sample set is classified into severe negative emotion, moderate negative emotion and mild negative emotion. The first two feature parameters with the highest significance are analyzed by category, and the optimal threshold combination corresponding to each category of negative emotion is determined for these two feature parameters.

[0014] Step 6: Train the depression SVM classifier:

[0015] The samples belonging to depression in the severe negative emotion samples obtained in step 5 are marked, and then the severe negative emotion samples and the depression sample set obtained in step 4 are imported into the SVM classifier for classification and training;

[0016] Step 7: classify the EEG signal of the emotion to be analyzed as negative emotion according to the threshold combination in step 5, and determine whether the EEG signal of the emotion to be analyzed is depression according to the trained depression SVM classifier in step 6.

[0017] Further, the pretreatment method in step 2 is as follows:

[0018] Electrode positioning was performed using the EEGLAB toolbox. 50 Hz power frequency interference was removed by filtering, high frequency signal interference was removed by 4-45 Hz bandpass filtering, and bad segments were removed. Low frequency interference was removed from the original signal by ICA. After ICA was performed, eye movement noise and artifact components were identified and removed based on the energy distribution of the topographic map and power spectrum analysis. ICA is a linear transformation whose main idea is to separate the signal into a linear combination of statistically independent non-Gaussian sources by minimizing the mutual information between the output components based on the InfoMax principle.

[0019] Furthermore, in step 3, the feature extraction method is as follows:

[0020] Step 3.1, identify the features that can effectively reflect different attention states from the EEG signals. The channels are selected according to the international standard 10-20 system standard electrode placement method. The 6 channels Fp1, Fp2, F3, F4, F7, F8 of the prefrontal lobe with the best emotional EEG and depression EEG performance, plus the channel signals of the two central areas C3 and C4, a total of 8 EEG channels are compared and analyzed. Based on the wavelet packet transform decomposition, 3 features are extracted based on the comparison of frequency domain and time-frequency domain. Each feature contains the parameters of 8 electrode channels, totaling 24 feature parameters.

[0021] The reasons for selecting the three features are as follows: after research, analysis and comparison, alpha waves are related to positive emotions, beta waves are related to negative emotions, when in a negative emotional state, the power of alpha waves will decrease and the power of beta waves will increase accordingly, and the left and right hemispheres of humans have different connections with specific emotions and emotional characteristics. Negative emotions are closely related to depression, and the long-term manifestation of depression is low mood, so the PSD ratio of alpha waves to beta waves and the left and right brain asymmetry of alpha wave PSD are extracted as classification features. The calculation of nonlinear feature sample entropy does not depend on the length of the data and has a small calculation error. It has strong anti-noise and anti-interference capabilities, and can reflect the complexity of brain activities. Therefore, it is suitable for analyzing complex EEG signals, so nonlinear feature sample entropy is extracted.

[0022] The three selected features are introduced as follows:

[0023] (1) PSD ratio of α waves to β waves;

[0024] PSD refers to the amount of signal in the frequency domain, indicating the distribution of power at each frequency point. The power spectrum transforms the EEG signal whose amplitude changes with time in the time domain into a power spectrum graph whose power changes with frequency. Therefore, the distribution and change of EEG rhythm can be directly observed through the power spectrum graph, and the components of each frequency band can be quantitatively analyzed. P(t) is the total energy value in the time domain, and the integrand on the right side of the equation is its energy spectrum density, that is, the energy per unit frequency. Its ratio to the time length of the signal can be called the power spectrum density.

[0025]

[0026] Among them, P(t) is the total energy value in the time domain; P(f) represents the energy spectral density.

[0027] (2) left-right hemisphere asymmetry of alpha wave PSD;

[0028]

[0029] Among them, F right Represents the PSD of the alpha wave right brain, F left represents the left hemisphere PSD of alpha waves; asym is used to represent the left-right hemisphere asymmetry of alpha wave PSD.

[0030] (3) Sample entropy;

[0031] Sample entropy is a nonlinear feature and a new method for measuring time complexity. Compared with other entropy analyses, the calculation of sample entropy does not depend on the length of the data and has a smaller calculation error. It has strong anti-noise and anti-interference capabilities, so it is suitable for analyzing complex EEG signals. Sample entropy is a dimensionless indicator that characterizes the complexity of the EEG signal sequence. The larger the entropy value, the greater the signal complexity and the higher the brain activity. The sample entropy assumes that the time series x(n) = x(1), x(2), ..., x(N) consists of N EEG signal data. The calculation method of sample entropy is as follows: The basic principle algorithm flow is shown in Table 1 below:

[0032] Table 1 Sample entropy algorithm principle

[0033]

[0034]

[0035] As can be seen from the above formula, the sample entropy algorithm mainly includes three parameters, namely the length N of the data, the embedding dimension m and the effective threshold r. The following describes the selection process of the sample entropy parameters:

[0036] m is the embedding dimension, which is the window length set when calculating sample entropy. In general, the value of m is generally set to 1 or 2. When the value of m is greater than 2, the required data length N needs to be very long. Only when N is greater than 5000 will there be a good effect. However, in this case, the required r will also be very large, and the analysis of the signal sequence will cause a large amount of information loss. Therefore, the present invention selects the value of m to be 2.

[0037] r is the effective threshold, indicating the similarity tolerance. When the value of r is large, more information will be lost, and when the value of r is small, the statistical characteristics of the system cannot be estimated ideally. The final conclusion is that when the value of r is usually between 0.1 and 0.25SD(x) (SD(x) is the standard deviation of the sequence), the characteristic parameter is more ideal. The present invention selects 0.2*std as the value of r.

[0038] The parameter N represents the length of the data. There is no restriction on the selection of N. In order to obtain ideal statistical characteristics, the number of data points should be controlled in the range of 100 to 5000.

[0039] Step 3.2: Fuse the 24 feature parameters extracted from each EEG signal to obtain a multi-domain feature fusion vector; fuse the 24 feature parameters extracted from each EEG signal to obtain a 24-dimensional (3 features * 8 channels) feature vector.

[0040] Step 3.3, construct a multi-domain feature fusion matrix;

[0041] According to the operations in steps 3.1 and 3.2, multiple EEG signals in the emotion Deap dataset are processed to obtain a multi-domain feature fusion matrix as a negative emotion sample set;

[0042] According to the operations in steps 3.1 and 3.2, multiple EEG signals in the USM depression dataset are processed, and the obtained multi-domain feature fusion matrix is ​​used as the depression sample set.

[0043] Furthermore, the feature parameter screening and dimensionality reduction method in step 4 is as follows:

[0044] Feature parameter selection is performed on the multi-domain feature fusion matrix obtained in step 3 to reduce redundancy. Significance analysis is performed to determine whether the 24 feature parameters are different. Based on the results, the best feature subset for classifier performance is selected, and feature parameters with no significant differences are removed.

[0045] According to the method of significance analysis, the differences of 24 parameters in the multi-domain feature fusion vectors in the negative emotion sample set and the depression sample set were analyzed, and the p value was calculated to determine the differences of the same type of feature parameters in the two sample sets. p>0.05 indicates that there is no significant difference between the negative emotion sample set and the depression sample set, that is, this type of feature parameter has little effect on the classification result or will interfere with the classification result, and these feature parameters need to be discarded.

[0046] Furthermore, the specific operations of step 5 are:

[0047] The negative emotion sample set is divided into three categories, namely severe negative emotion, moderate negative emotion and mild negative emotion. The multi-domain feature fusion matrix of the negative emotion sample set is subjected to ANOVA test according to the category. The first two feature parameters with the highest significance are calculated and analyzed, and a global optimization strategy based on the traversal method is adopted to determine the optimal threshold combination corresponding to each type of negative emotion.

[0048] Furthermore, the specific operations of step 6 are:

[0049] The label of severe negative emotion samples is recorded as 1, among which the label of severe negative emotion samples belonging to depression is recorded as 0. The accuracy is judged by using the five-fold cross-validation method; the severe negative emotion sample data set is randomly divided into 5 subsets of equal size and then five iterations are performed. Each time, four subsets are used as training sets, and the remaining subset is used as a validation set to test the performance of the SVM classifier. After the five iterations, the average accuracy of these five results is calculated to represent the accuracy of the SVM classifier after training.

[0050] Beneficial effects: The method of the present invention not only improves the accuracy of emotion recognition, but also provides a new idea for early detection and intervention of psychological disorders such as depression. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a framework flow chart of the method of the present invention.

[0052] Figure 2 It is a VA two-dimensional model diagram of emotion recognition in the present invention.

[0053] Figure 3 It is an international 10-20 EEG acquisition electrode diagram of the EEG acquisition device of the present invention.

[0054] Figure 4 It is a schematic diagram of the process of preprocessing EEG signals in the present invention.

[0055] Figure 5 It is a partial processing schematic diagram of the EEGLAB toolbox of the present invention.

[0056] Figure 6 This is the signal comparison diagram after ICA denoising of the present invention

[0057] Figure 7 It is a schematic diagram of using wavelet packet transform decomposition before extracting features in the present invention.

[0058] Figure 8 It is a schematic diagram of the combination of multi-domain fusion feature vectors constructed by the present invention.

[0059] Fig. 9 It is a schematic diagram of the feature selection model proposed in the present invention.

[0060] Fig.10 It is the negative emotion grading threshold diagram of the present invention, wherein -1 represents severe negative emotion, 0 represents general negative emotion, and 1 represents mild negative emotion.

[0061] Fig.11 It is a schematic diagram of the present invention using an SVM classifier to classify negative emotions and depression using 5-fold cross validation. DETAILED DESCRIPTION

[0062] The technical solution of the present invention is described in detail below through embodiments, but the protection scope of the present invention is not limited to the embodiments.

[0063] This embodiment proposes a method for classifying negative emotions and distinguishing depression based on EEG signals, and its flow chart is as follows: Figure 1 As shown. This embodiment extracts 3 features of 8 EEG channels to form a feature matrix and performs channel and feature selection analysis on the relationship between depression and negative emotion classification. The specific steps are as follows:

[0064] Step 1) Obtain the emotion Deap dataset and HUSM depression dataset through online application, and process the emotion Deap dataset based on the emotion VA two-dimensional model to retain the negative emotions.

[0065] In this embodiment, the emotion Deap dataset and the HUSM depression dataset are used for analysis. The former is used to analyze the classification of negative emotions, and the latter is used to add the relationship between negative emotions and depression to the former research.

[0066] The classification of negative emotions is based on the emotion VA two-dimensional model. The emotion VA two-dimensional model is as follows: Figure 2As shown in the figure, emotions can be divided into two dimensions: valence and arousal, which can be divided into one coordinate axis: the horizontal axis represents the valence dimension, representing the negativity or positivity of the emotional state, and the coordinate represents the transition from one extreme (happy, joy) to another extreme (sadness, sadness); the vertical axis represents the arousal dimension, representing the state of emotions from strong to weak, reflecting the intensity of emotions. Various complex emotions can be represented by a linear combination of these two dimensions, thus effectively reflecting the continuous change process of emotions.

[0067] Step 2: preprocess the original EEG signals in the two data sets;

[0068] This embodiment combines EEG signals with negative emotion recognition, and preprocesses the EEG signals of the emotion Deap dataset and the HUSM depression dataset. Since the forehead is more sensitive to changes in emotional EEG, the EEG signals of the frontal lobe are selected and the signals of the two central areas C3 and C4 are added for comparison.

[0069] The distribution diagram of the electrode channels corresponding to the EEG acquisition equipment is as follows Figure 3 As shown, the present invention selects EEG channels according to the distribution diagram. The 19-lead electrode distribution conforms to the international 10-20 system standard electrode placement method, specifically including forehead electrodes Fp1, Fp2, frontal lobe electrodes F3, F4, F7, F8, Fz, central area electrodes C3, C4, CZ, parietal area electrodes P3, P4, PZ, occipital area electrodes O1, O2, temporal lobe area electrodes T3, T4, T5, T6 and reference electrodes A1, A2.

[0070] The schematic diagram of the EEG signal preprocessing process is as follows Figure 4 As shown, the preprocessing is based on the EEGLAB toolbox and Matlab software, including positioning, filtering, ICA artifact removal and other steps.

[0071] The topographic map after ICA analysis is as follows Figure 5 As shown in the figure, the component attributes of the electrooculogram artifact are concentrated in the frontal eyeball and need to be deleted.

[0072] The comparison of signals before and after ICA denoising is shown in the figure below. Figure 6 As shown in the figure, the original signal and the signal after ICA denoising are compared in the time domain. It can be clearly seen that the denoising and reconstructed signal removes noise while effectively retaining complete information.

[0073] Step 3: Feature extraction and construction of multi-domain feature fusion matrix:

[0074] The preprocessed signal is subjected to feature extraction, and the wavelet packet transform method of time-frequency analysis is used. Wavelet packet analysis can decompose the low-frequency part and the high-frequency part at the same time, making the signal analysis more detailed and the time-frequency more detailed. Based on the wavelet packet decomposition, three features are extracted, namely, the PSD ratio of α wave and β wave, the asymmetry of the left and right brain of α wave PSD, and the sample entropy. The three features of the extracted 8 channels are combined into a multi-domain feature fusion vector.

[0075] Wavelet packet decomposition diagram Figure 7 As shown. A represents low frequency, D represents high frequency, and the number at the end represents the number of layers of wavelet packet decomposition. Since the required brain wave frequency bands are α waves and β waves, the db5 wavelet signal is used to perform a 4-layer decomposition on the 8-channel EEG signals collected in the experiment. After decomposition, the EEG signal α wave and β wave are separated from the original signal.

[0076] The schematic diagram of the combination of feature fusion vectors is shown in Figure 8. The above extracted features are fused to construct a multi-domain feature fusion vector, and a 24-dimensional (8 channels * 3 feature parameters) feature vector F is obtained. The feature parameters represented by F1 to F3 in the figure are the PSD ratio of α wave and β wave, the asymmetry of α wave left and right brain, and the sample entropy.

[0077] 510 EEG signal samples are extracted from the emotion Deap dataset and the USM depression dataset respectively to form a 1020*24 feature matrix.

[0078] The 510 EEG signals in the emotion Deap dataset were processed, and the multi-domain feature fusion vectors obtained formed a multi-domain feature fusion matrix as the negative emotion sample set; the 510 EEG signals in the USM depression dataset were processed, and the multi-domain feature fusion vectors obtained formed a multi-domain feature fusion matrix as the depression sample set.

[0079] Step 4: Perform feature parameter screening and dimensionality reduction on the multi-domain feature fusion matrix:

[0080] The 24 feature parameters in the depression sample set and the negative emotion sample set were subjected to significance analysis and normalization processing to determine the validity and accuracy of the extracted features. Fig. 9 The 24 feature parameters RMS graph is displayed, showing the RMS values ​​of the 24 feature parameters of the 8 channels extracted to show the range of the data. It can be seen from the figure that most of the selected features have certain differences and can distinguish the data well. The α / β PSD ratio feature shows the strongest effectiveness in distinguishing negative emotions and depressive symptoms, with a higher RMS value, but the distinction is not obvious in some channels. The performance of asymmetry features and sample entropy features is consistent and significantly different between different channels.

[0081] According to the method of significance analysis, the differences of 24 parameters in the multi-domain feature fusion vectors in the negative emotion sample set and the depression sample set were analyzed, and the p value was calculated to determine the differences of the same type of feature parameters in the two sample sets. p>0.05 indicates that there is no significant difference between the negative emotion sample set and the depression sample set, that is, this type of feature parameter has little effect on the classification result or will interfere with the classification result, and these feature parameters need to be discarded.

[0082] Step 5: Negative emotion threshold analysis:

[0083] The negative emotion sample set is divided into three categories, namely severe negative emotion, moderate negative emotion and mild negative emotion. The multi-domain feature fusion matrix of the negative emotion sample set is subjected to ANOVA test according to the category. The first two feature parameters with the highest significance (the first two feature parameters with the best classification performance) are calculated and analyzed, and a global optimization strategy based on the traversal method is adopted to determine the optimal threshold combination corresponding to each type of negative emotion.

[0084] Fig.10 is the threshold for negative emotion classification. According to the VA two-dimensional model, the negative emotion data of the negative emotion sample set are divided into three groups: the negative emotion sample set is divided into severe negative emotion, moderate negative emotion and mild negative emotion. After analysis, it is found that the PSD ratio of α / β of feature parameter 4 (Fp2) corresponds to Fig.10 A) and characteristic parameter 10 (the PSD ratio of α / β of F4, corresponding to Fig.10 The analysis of B) in the above results in the best results; these feature parameters have the largest differences under different negative emotional states, indicating that they play an important role in emotion classification. Then, the thresholds of these features under different emotional states are further analyzed to optimize the accuracy of the emotion classification model. For severe negative emotions, moderate negative emotions, and mild negative emotions, all possible threshold combinations of these two feature parameters are systematically traversed, and the classification accuracy corresponding to each pair of thresholds is evaluated, and finally the threshold pair that maximizes the model performance is selected. This strategy ensures that the optimal segmentation point is found in the feature space, thereby giving the negative emotion sample set a higher ability to distinguish, enabling it to more accurately and reliably identify different levels of negative emotion, further improving the classification effect.

[0085] Step 6: Train the depression SVM classifier:

[0086] The samples belonging to depression in the severe negative emotion samples (multi-domain feature fusion matrix) obtained in step 5 are marked, and combined with the depression sample set (multi-domain feature fusion matrix), they are imported into the SVM classifier for classification and training.

[0087] The label of severe negative emotion samples is recorded as 1, and the label of severe negative emotion samples belonging to depression is recorded as 0. The accuracy is determined by the five-fold cross-validation method. The severe negative emotion sample data set is randomly divided into five subsets of equal size and then iterated five times. Each time, four subsets are used as training sets, and the remaining subset is used as a validation set to test the performance of the SVM classifier. After the five iterations, the average accuracy of the five results is calculated to represent the accuracy of the trained SVM classifier.

[0088] like Fig.11 As shown: This figure is a diagram showing the accuracy of depression recognition using the SVM classifier. It can be clearly seen from the results in the figure that after the recognition method proposed by the present invention, efficient emotion and depression classification is achieved, providing multiple possible directions for future EEG medical research.

[0089] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and details may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for classifying negative emotions and distinguishing depression based on EEG signals, characterized in that: The steps include: Step 1: Obtain the emotion Deap dataset and HUSM depression dataset through online application, and process the emotion Deap dataset based on the emotion VA two-dimensional model to retain the negative emotions; Step 2: preprocess the original EEG signals in the two data sets; Step 3: Feature extraction and construction of multi-domain feature fusion matrix: For the EEG signals in the two data sets, three features are extracted from the EEG signals: the PSD ratio of α waves to β waves, the left-right brain asymmetry of α wave PSD, and the nonlinear feature sample entropy. Among them, PSD refers to the quantity of the signal in the frequency domain. The feature parameters extracted from each EEG signal are fused to obtain a multi-domain feature fusion vector, and the multi-domain feature fusion vectors in the two data sets are combined to construct a multi-domain feature fusion matrix. Step 4: Perform feature parameter screening and dimensionality reduction on the multi-domain feature fusion matrix: Select feature parameters for the multi-domain feature fusion matrix obtained in step 3, and remove feature parameters that do not have significant differences; Step 5: Negative emotion threshold analysis: The negative emotion sample set is classified into severe negative emotion, moderate negative emotion and mild negative emotion. The first two feature parameters with the highest significance are analyzed by category, and the optimal threshold combination corresponding to each category of negative emotion is determined for these two feature parameters. Step 6: Train the depression SVM classifier: The samples belonging to depression in the severe negative emotion samples obtained in step 5 are marked, and then the severe negative emotion samples and the depression sample set obtained in step 4 are imported into the SVM classifier for classification and training; Step 7: Negative emotion classification and depression discrimination of the emotions to be analyzed: The EEG signal of the emotion to be analyzed is classified as negative emotion according to the threshold combination in step 5, and the trained depression SVM classifier in step 6 is used to determine whether the EEG signal of the emotion to be analyzed is depression.

2. The method for classifying negative emotions and distinguishing depression based on EEG signals according to claim 1, characterized in that: The pretreatment method in step 2 is: The EEGLAB toolbox was used for electrode positioning. The 50 Hz power frequency interference was removed by a filter, and the high-frequency signal interference was removed by a 4-45 Hz band-pass filter. Bad segments were removed, and low-frequency interference was removed from the original signal by ICA. After performing ICA, eye movement noise and artifact components were identified and removed based on the energy distribution of the topographic map and power spectrum analysis.

3. The method for classifying negative emotions and distinguishing depression based on EEG signals according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1, feature extraction; Identify the features that can effectively reflect different attention states from the EEG signals. The channels are selected according to the international standard 10-20 system standard electrode placement method. A total of 8 EEG channels are compared and analyzed based on the channel signals of the 6 channels Fp1, Fp2, F3, F4, F7, F8 of the prefrontal lobe and the two central areas C3 and C4. Based on the wavelet packet transform decomposition, the PSD ratio of α wave to β wave, the left-right brain asymmetry of α wave PSD, and the nonlinear characteristic sample entropy are extracted based on the frequency domain and time-frequency domain comparison. Each feature contains the parameters of 8 electrode channels, totaling 24 feature parameters. Step 3.2: Fuse the 24 feature parameters extracted from each EEG signal to obtain a multi-domain feature fusion vector: Step 3.3, construct a multi-domain feature fusion matrix; According to the operations in step 3.1 and step 3.2, multiple EEG signals in the emotion Deap dataset are processed, and the obtained multi-domain feature fusion vectors form a multi-domain feature fusion matrix as a negative emotion sample set; According to the operations in step 3.1 and step 3.2, multiple EEG signals in the USM depression dataset are processed, and the obtained multi-domain feature fusion vectors form a multi-domain feature fusion matrix as the depression sample set.

4. The method for classifying negative emotions and distinguishing depression based on EEG signals according to claim 1, characterized in that: The feature parameter screening and dimensionality reduction method in step 4 is as follows: According to the method of significance analysis, the differences of 24 feature parameters in the multi-domain feature fusion vectors in the negative emotion sample set and the depression sample set were analyzed, and the p value was calculated to determine the differences of the same type of feature parameters in the two sample sets. p>0.05 indicates that there is no significant difference between the negative emotion sample set and the depression sample set, that is, this type of feature parameter has little effect on the classification result or will interfere with the classification result, and these feature parameters need to be discarded.

5. The method for classifying negative emotions and distinguishing depression based on EEG signals according to claim 1, characterized in that: The specific operations of step 5 are: The negative emotion sample set is divided into three categories, namely severe negative emotion, moderate negative emotion and mild negative emotion. The multi-domain feature fusion matrix of the negative emotion sample set is subjected to ANOVA test according to the category. The first two feature parameters with the highest significance are calculated and analyzed, and a global optimization strategy based on the traversal method is adopted to determine the optimal threshold combination corresponding to each type of negative emotion.

6. The method for classifying negative emotions and distinguishing depression based on EEG signals according to claim 1, characterized in that: The specific operations of step 6 are: The label of severe negative emotion samples is recorded as 1, among which the label of severe negative emotion samples belonging to depression is recorded as 0. The accuracy is judged by using the five-fold cross-validation method; the severe negative emotion sample data set is randomly divided into 5 subsets of equal size and then five iterations are performed. Each time, four subsets are used as training sets, and the remaining subset is used as a validation set to test the performance of the SVM classifier. After the five iterations, the average accuracy of these five results is calculated to represent the accuracy of the SVM classifier after training.

Images