Multivariable electroencephalogram emotion decoding depression assessment method based on time cluster strategy

Abstract

The invention relates to the technical field of medical diagnosis, in particular to a multivariable electroencephalogram emotion decoding depression assessment method based on a time cluster strategy, and aims at solving the problems that when symptoms of a depression patient are detected at present, false positive results are extremely likely to occur due to the adoption of a multivariable analysis mode, and the depression assessment accuracy and reliability are poor. According to the method, electroencephalogram signals of a depression patient under an emotion recognition task are collected and preprocessed, then an individual decoding mode is calculated based on a time cluster strategy, time dynamic characteristics are extracted, and finally, cross validation is performed by adopting a leave-one-out method. The performance of the model in depressive symptom evaluation is evaluated by calculating the root-mean-square error and significance of the Bayesian model prediction score and the scale score, so that the problem of false positive in multivariate mode analysis is effectively avoided; time dynamic characteristics reflecting possible abnormal nerve states activated in emotion processing are stably extracted by the aid of electroencephalogram signals, depression evaluation is carried out on the basis of the characteristics, and prediction accuracy and reliability of depression evaluation are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical diagnosis, and in particular to a multivariate electroencephalogram emotion decoding depression evaluation method based on a time cluster strategy. BACKGROUND

[0002] Depression, as a mental illness that poses a major challenge to global public health, has long been associated with low clinical recognition accuracy. Current diagnostic methods mainly rely on subjective scale assessment and symptom observation, which are easily influenced by patient expression bias, physician experience difference and environmental factors, leading to high risk of misdiagnosis and missed diagnosis. Research shows that the recognition accuracy of depression relying solely on traditional clinical interviews is low, which seriously hinders the implementation of early intervention and personalized treatment. Therefore, it is an urgent need to develop a quantitative evaluation tool with convenience and objectivity to improve diagnostic accuracy by extracting standardized biomarkers. The neural mechanism of depression is complex. Existing research shows that depression patients generally have emotional processing bias, which is characterized by over-sensitivity to negative stimuli and reduced response to positive stimuli. This feature is considered to be the core cognitive mechanism of the occurrence and maintenance of depression. Traditional behavioral methods, such as reaction time and accuracy, can reflect the overall emotional bias trend, but cannot analyze dynamic neural processes and are easily influenced by the subject's subjective state and environment. Functional magnetic resonance imaging and electroencephalogram research have revealed abnormalities in the amygdala-frontal lobe circuit and the regulation defects of event-related potentials such as P100 and LPP, but the time resolution of functional magnetic resonance imaging is insufficient to capture millisecond-level neural activity. Event-related potential analysis relies on pre-defined time windows and electrode positions, ignoring the high spatiotemporal dynamics of emotional processing and individual neural response variability, resulting in low reproducibility of results. Therefore, the existing technology cannot capture the time dynamic pattern related to emotional bias, and the time characteristics of dynamic neural coding during emotional stimulus processing have not been fully analyzed, which fundamentally restricts the development of high-specificity recognition strategies. In recent years, multivariate pattern analysis has provided a new way to track the dynamic emotional processing of depression individuals by decoding the temporal coding pattern of electroencephalogram signals. However, this method has a significant decoding time sequence false positive risk at the individual level, i.e. random noise or accidental fluctuations in individual neural activity may be misjudged as valid emotional decoding signals, which poses a serious challenge to the screening of standardized depression biomarkers based on multivariate patterns. Therefore, it is urgent to develop a better analysis method to suppress the random fluctuations of individual decoding results and ensure the reliability of multivariate pattern analysis in the analysis of depression neural mechanisms and the improvement of diagnostic accuracy. SUMMARY

[0003] The present application provides a multivariate electroencephalogram emotion decoding depression evaluation method based on a time cluster strategy to solve the problems in the above background.

[0004] In order to achieve the above object, the present application adopts the following technical solutions:

[0005] A multivariate electroencephalogram emotion decoding depression evaluation method based on a time cluster strategy, comprising the following steps:

[0006] Step 1, data collection and preprocessing stage:

[0007] First, the electroencephalogram signals of a plurality of depression patients under an emotion recognition task are collected as samples, then irrelevant electrodes are removed, then band-pass filtering between 0.1Hz-30Hz is performed on the electroencephalogram signals, then independent component analysis method is used to separate and remove electrooculogram and electromyogram and other artifact signals, finally, a time window of 1000ms from 200ms before the stimulus presentation to 800ms after the stimulus presentation is selected, the electroencephalogram signal 200ms before the stimulus is set as a baseline, and abnormal fragments with an amplitude ≥100μV are removed to ensure that subsequent analysis accurately reflects brain activity and realizes electroencephalogram sample division.

[0008] Step 2, decoding mode calculation stage:

[0009] A sample linear support vector machine classifier classifies two types of emotions in step 1 at each sample point, then principal component analysis is performed on the voltage values of the electroencephalogram electrode channels, the principal components explaining 99.99% of the variance of the data are retained, the electroencephalogram response mode is constructed, then the two types of emotion samples are divided into 10 subsets and the test set is rotated each time, then the training data and the test data are input into the classifier after z-score standardization, then the classification results of each sample point are repeatedly calculated to obtain an individual emotion decoding dynamic accuracy curve changing with time, the emotion bias processing mode of the depression patient is quantified, then statistical analysis is performed on the individual emotion decoding dynamic curve and the opportunity level to determine the authenticity of the decoding accuracy at each sample point, then Fisher algorithm is used to replace the labels to obtain the following accuracy matrix: s is the number of sample points, and each row of the matrix represents the accuracy of 999 times of classifier iteration at the current sample point, the accuracy in the empty distribution is sorted in ascending order, then the decoding of the depression patient at the current sample point is calculated as the proportion of the zero distribution greater than or equal to the accuracy obtained by using the correct label, and the formula is as follows: index is the index of the correct label, when the decoding accuracy obtained by using the correct label is greater than or equal to 95% of the zero distribution, i.e. s I ≥0.95, it is determined that the decoding of the current sample point is significant, and the preliminary significant decoding time sequence can be obtained by repeating the process for each sample point: The significant time sequence is then corrected, and the time cluster is composed of adjacent sample points I s ≥0.95, and the size is calculated as the accumulation of significant points in the time cluster, i.e. Thus, the time cluster definition and screening are completed, and then the most significant cluster under each permutation label is defined, and the most significant time cluster c is extracted is formed, that is, MaxTimeClusterSequence=(c max , c max，1 , c max，2 ,..., c max，999 ), and each significant time cluster under the correct label is formed into a null distribution of multiple time clusters with the maximum significant time cluster distribution, and the time clusters are sorted in ascending order, and the condition of the individual participant in the time cluster of the current correct label is calculated as the proportion of the zero distribution of the time cluster obtained with the correct label, that is, the formula

[0010] is used to determine whether the time cluster of the current true label is true, and the time cluster of each depressive patient is verified for authenticity, and finally the significant decoding time sequence of the depressive individual is screened out.

[0011] Step 3, feature extraction stage:

[0012] The time dynamic feature is extracted in the individual decoding mode in the step 2, that is, the individual level area under the curve, which is defined as the area under the curve of the individual in the significant region and indicates the possible abnormal neural state activated in emotional processing.

[0013] Step 4, depression evaluation stage:

[0014] First, the individual level area under the curve of each depressive individual in step 3 is extracted, and the label of each individual is the score obtained by the clinician based on the Hamilton Depression Scale-17 evaluation, the feature integration is completed, then the Bayesian model is used as the regression model, and the leave-one-out cross-validation strategy is used, and the single subject data is used as the test set and the remaining subject data is used as the training set for iterative training and verification, and finally the performance of the model in the depression symptom evaluation can be evaluated by calculating the root mean square error and the prediction significance of the prediction score of the leave-one-out cross-validation and the clinician evaluation.

[0015] The present application has the following beneficial effects:

[0016] The present application firstly processes the high time resolution electroencephalogram signals of depression patients in an emotion recognition task, then obtains the individual emotion decoding dynamic curve through multivariate pattern analysis, subsequently calculates the individual significant decoding time sequence through the individual decoding method based on time cluster, combines with the individual emotion decoding dynamic curve to form the individual decoding mode, and finally extracts the time dynamic characteristics of the individual decoding dynamic mode to construct the index system for measuring the abnormal emotion processing of depression, which not only optimizes and expands the calculation method of the individual significant decoding time sequence, but also realizes the accurate quantification of the neural markers of abnormal emotion processing of depression patients, and more accurately realizes the depression evaluation based on electroencephalogram signals. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 It is the overall technical route schematic diagram of the present application.

[0018] Figure 2 It is the time cluster definition and screening stage method schematic diagram of the present application.

[0019] Figure 3 It is the feature extraction stage individual level area under the curve feature difference graph of refractory depression patients and remitted refractory depression patients after intervention of the present application. DETAILED DESCRIPTION

[0020] The present application will be further described below in combination with the drawings and specific embodiments.

[0021] The present embodiment will be clearly and completely described in combination with the drawings on the depression task-state electroencephalogram data set.

[0022] Figure 1 It is the technical route graph of the depression evaluation method based on the time cluster strategy multivariate electroencephalogram emotion decoding.

[0023] All samples of the data set come from refractory depression patients in the psychiatry department of an affiliated hospital of a certain university, a total of 20 subjects are enrolled, and the intervention treatment is carried out by inhaling 50% laughing gas / 50% oxygen through a mask, the inhalation time is 1h, and the electroencephalogram task-state data of the subjects in the emotion valence recognition task are collected before and after the intervention, and table 1 is the explanation of the subject information of the depression task-state electroencephalogram data.

[0024] Table 1 demographic information of the data set

[0025]

[0026] A multivariate electroencephalogram emotion decoding depression evaluation method based on a time cluster strategy, comprising a data collection and preprocessing stage, a decoding mode calculation stage, a feature extraction stage and a depression evaluation stage, specifically as follows:

[0027] 1. Preprocessing stage: The collected EEG task-state data of the subjects performing the emotional valence recognition task is preprocessed, specifically:

[0028] (1) Remove irrelevant electrodes: In the EEG signal collection, not all electrodes have directly related brain activity information. In this example, we removed the Cz electrode which does not record brain activity information.

[0029] (2) EEG signal filtering: Apply a band-pass filter between 0.1 Hz and 30 Hz.

[0030] (3) Independent component analysis to remove artifacts: Use independent component analysis to separate and remove electrooculogram (horizontal, vertical) and electromyogram artifact signals.

[0031] (4) EEG sample division: Select a time window of 1000 ms from 200 ms before stimulus presentation to 800 ms after stimulus presentation, and set the 200 ms before stimulus as the baseline. For abnormal segments with amplitude exceeding 100 μV, they are removed. In this example, subjects who provide excessive EEG artifacts are deleted, leaving 18 subjects.

[0032] 2. Decoding mode calculation stage: The purpose of the decoding mode calculation stage is to extract the decoding dynamic mode of the depressed individuals to represent the neural representation of abnormal emotional processing. The specific implementation steps are as follows:

[0033] (1) Sample linear support vector machine classifier to classify two types of emotions at each sampling point.

[0034] (2) Use the voltage value of the EEG electrode channel as the input feature of the classifier. In order to reduce noise and optimize computational efficiency, principal component analysis is performed on the voltage data, retaining the principal components that explain 99.99% of the variance of the data, thereby constructing a stable EEG response pattern.

[0035] (3) To prevent bias of the classifier and ensure equal number of samples of the two types of emotions, in data processing, the larger sample of one type of emotional data is randomly sampled to match the number of the other type of sample, in order to improve the generalization ability of the classifier.

[0036] (4) Use 10-fold cross-validation to divide the two types of emotional samples into 10 subsets, and rotate the test set each time to ensure the stability and generalization of the classification results. In this example, each subset includes 7-9 trials. In addition, to further improve the signal-to-noise ratio, the test set only contains two samples, i.e. the mean of the EEG response pattern of the two types of emotions in the current fold.

[0037] (5) Before inputting into the classifier, z-score standardization is performed on the training data and test data to optimize the classification performance.

[0038] (6), for each sampling point, the process is repeated 100 times, and the final decoding accuracy result at the current sampling point is the average result of the classification accuracy of 1000 cross-validations. By repeatedly calculating the classification results of each sampling point, the individual emotion decoding dynamic accuracy curve changing with time, i.e. the individual emotion decoding dynamic curve, is obtained.

[0039] (7), then the calculation of the significant decoding time sequence of the depression individual is carried out. Firstly, statistical analysis and calculation are carried out at the individual level and the chance level to obtain the preliminary significant decoding time sequence. The present application uses permutation labels (happiness and sadness) to generate another 999 times of classifier iterations for the individual to create an empty distribution for each sampling point. The labels are replaced using the Fisher algorithm, and the decoding accuracy obtained by replacing the labels is as follows:

[0040]

[0041] s represents the number of sampling points, and each row of the matrix represents the accuracy of 999 times of classifier iterations at the current sampling point. An empty distribution of decoding accuracy (1000 permutations) is created for each sampling point. The accuracy in the empty distribution is sorted in ascending order, and then the decoding of the depression patient at the current sampling point is calculated as the proportion of the zero distribution greater than or equal to the accuracy obtained by using the correct label. The formula is as follows:

[0042]

[0043] index is the index of the correct label. When the decoding accuracy obtained by using the correct label is greater than or equal to 95% of the zero distribution (I s ≥0.95), it is determined that the decoding of the current sampling point is significant. Repeating this process for each sampling point can obtain the preliminary significant decoding time sequence:

[0044]

[0045] (8), definition and screening of time clusters. A large number of point-by-point tests may lead to many accidental significant results (false positives), and random noise or accidental fluctuations of individual neural activity may be misjudged as effective significant decoding signals. Therefore, the significant time sequence needs to be corrected. Here, time clusters are defined to capture the time continuity of significant decoding, as shown in Figure 2 , the time cluster is composed of adjacent sampling points I s ≥0.95, and the size is calculated as the cumulative of the significant points in the time cluster:

[0046]

[0047] The most significant cluster under each permutation label is defined as follows:

[0048]

[0049] For each permutation, the most significant time cluster c max is extracted, and the maximum significant time cluster distribution is formed:

[0050] MaxTimeCluserSequence=(c max，1 , c max，2 ,..., c max，999 )

[0051] (9) The significant time cluster under the correct label is evaluated for authenticity, each significant time cluster under the correct label is compared with the maximum significant time cluster distribution to form a null distribution (1000 in total) of time clusters, the time clusters in the null distribution are sorted in ascending order, the proportion of the time clusters of the individual participant under the current correct label that is greater than or equal to the zero distribution of the time clusters obtained with the correct label is calculated, and whether the time cluster of the current true label is true is judged by the following formula:

[0052]

[0053] Through the above method, the authenticity of the time cluster of each refractory depression patient is tested, and finally the significant decoding time sequence of the individual is screened out, and the individual significant decoding time sequence obtained is combined with the individual emotion decoding dynamic curve to form an individualized decoding mode.

[0054] 3. Feature extraction stage: The purpose of the feature extraction stage is to calculate the degree of abnormal emotion processing bias of refractory depression patients. In order to effectively capture the time representation changes of depression patients in abnormal emotion processing, the individual level area under the curve is extracted in the individual decoding mode, which is defined as the area under the curve of the individual in the significant region. This feature may indicate the possible neural state activated in emotion processing, Figure 3 The figure of the difference in the individual level area under the curve feature of the refractory depression patients and the refractory depression patients relieved after intervention in the feature extraction stage of the present application.

[0055] 4、 Depression assessment stage: the present application adopts a Bayesian regression model (Naive Bayesian Regression) to comprehensively quantitatively evaluate the symptom severity of patients with refractory depression, and constructs a Bayesian probability regression framework based on the emotional processing time dynamic characteristics obtained in the feature extraction stage. In order to verify the model efficiency, a leave-one-out cross-validation (LOOCV) strategy is adopted: the data of a single subject is sequentially taken as the test set, and the data of the remaining subjects is taken as the training set for iteration, and finally the root mean square error of the predicted score and the clinical evaluation is calculated. In order to further improve the evaluation accuracy of the prediction performance, random guessing is introduced as a control, and the pairing relationship between the individual time dynamic characteristics and the corresponding scale scores is randomly disturbed, and 999 arrangements are performed to generate the random distribution of the root mean square error of each time dynamic characteristic (a total of 1000 times of root mean square error iteration), including the results of the true pairing relationship. This method can compare the prediction results with the zero distribution, so as to calculate the prediction significance of the time dynamic characteristics. This evaluation system can realize comprehensive and objective classification of depression symptoms from mild to severe, and provide data-driven decision support for individualized treatment strategies.

[0056] The specific evaluation steps are as follows:

[0057] (1) Feature integration: in the feature extraction stage, the time dynamic characteristics of each refractory depression patient can be obtained, that is, each individual obtains the individual level of area under the curve, and the label of each individual is the score obtained by the clinician based on the Hamilton Depression Scale-17 evaluation. In this way, the feature value and label of each individual are obtained.

[0058] (2) Model selection: a Bayesian model is used as the regression model.

[0059] (3) Leave-one-out cross-validation: in order to verify the evaluation accuracy and stability of the model, a leave-one-out cross-validation strategy is adopted, and the data of a single subject is sequentially taken as the test set, and the data of the remaining subjects is taken as the training set for iterative training and verification, in order to evaluate the stability and prediction effect of the model on different data subsets.

[0060] (4) Result evaluation: the performance of the model in depression symptom evaluation is evaluated by calculating the root mean square error of the predicted score and the clinical evaluation and the prediction significance of the leave-one-out cross-validation. The results show that significant prediction effect is observed in the data set, the root mean square error is 3.73, and the significance p value is 0.003.

Claims

1. A multivariate electroencephalogram emotion decoding depression assessment method based on a time cluster strategy, characterized in that, The method comprises the following steps: Step 1, data collection and preprocessing stage: First, the EEG signals of multiple depression patients in the emotional recognition task are collected as samples, then irrelevant electrodes are removed, then band-pass filtering between 0.1 Hz and 30 Hz is performed on the EEG signals, then independent component analysis method is used to separate and remove EOG and EMG artifacts, finally, a time window of 1000 ms from 200 ms before the stimulus presentation to 800 ms after the stimulus presentation is selected, the EEG signal 200 ms before the stimulus is set as the baseline, and the abnormal fragments with amplitude ≥ 100 μV are removed to ensure that the subsequent analysis accurately reflects the brain activity and realizes the division of EEG samples; Step 2, decoding mode calculation stage: The sampling linear support vector machine classifier classifies the two types of emotions in step 1 at each sampling point, and then performs principal component analysis on the voltage values of the electroencephalogram electrode channels to retain the principal components that explain 99.99% of the variance of the data, construct an electroencephalogram response pattern, and then divide the two types of emotional samples into 10 subsets and test the set in each rotation, and then input the training data and test data into the classifier after z-score standardization, and then calculate the classification result of each sampling point repeatedly to obtain the individual emotion decoding dynamic accuracy curve changing with time, quantify the emotional bias processing pattern of the depressed patients, and then statistically analyze the individual emotion decoding dynamic curve and the opportunity level to determine the authenticity of the decoding accuracy at each sampling point, and then use the Fisher algorithm to replace the label to obtain the following accuracy matrix: s is the number of sampling points, and each row of the matrix represents the accuracy of 999 classifier iterations at the current sampling point. The accuracy in the empty distribution is sorted in ascending order, and then the decoding of the depressed patients at the current sampling point is calculated as the proportion of the zero distribution obtained by the correct label that is greater than or equal to the accuracy, and the formula is as follows: index is the index of the correct label, and when the decoding accuracy obtained by the correct label is greater than or equal to 95% of the zero distribution, i.e. I s ≥ 0.95, it is determined that the decoding at the current sampling point is significant, and the process is repeated for each sampling point to obtain the preliminary significant decoding time sequence: The significant time sequence is then corrected, and the time cluster is composed of adjacent sampling points I s ≥ 0.95, and the size is calculated as the cumulative significant points within the time cluster, i.e. Thus, the time cluster definition and screening are completed, and then by defining the most significant cluster under each permutation label, the most significant time cluster c max is extracted to form the maximum significant time cluster distribution, i.e. MaxTimeClusterSequence = (c max,1 , c max,2 ,..., c max,999 ), and then each significant time cluster under the correct label is compared with the maximum significant time cluster distribution to form multiple time cluster empty distributions, and the time clusters are sorted in ascending order. The proportion of the zero distribution obtained by the correct label that is greater than or equal to the time cluster is calculated as the proportion of the zero distribution obtained by the correct label that is greater than or equal to the time cluster, and the formula is as follows: The time cluster of the current real label is judged whether it is real or not, and the authenticity of each time cluster of the depression patient is tested, and finally the significant decoding time sequence of the depression individual is screened out. Step 3, feature extraction stage: The time dynamic feature is extracted in the individual decoding mode in the step 2, that is, the individual level area under the curve, which is defined as the area under the curve of the individual in the significant region and indicates the possible abnormal neural state activated in the emotional processing; Step 4, depression evaluation stage: First, the area under the curve of each depression individual in step 3 is extracted, the label of each individual is the score obtained by the clinician based on the Hamilton Depression Scale-17 evaluation, the feature integration is completed, then the Bayesian model is used as the regression model, the leave-one-out cross-validation strategy is used, the data of a single subject is used as the test set and the data of the remaining subjects is used as the training set for iterative training and verification, finally, the performance of the model in the depression symptom evaluation can be evaluated by calculating the root mean square error and the prediction significance of the prediction score of the leave-one-out cross-validation and the clinician evaluation.

Images