Method and apparatus for assessing emotional state

Abstract

This application proposes a method and apparatus for assessing emotional states. The method includes: acquiring raw electroencephalogram (EEG) signals from a subject; generating a set of available features from the EEG signals for assessing the subject's emotional state; and using an assessment model to determine the corresponding emotional state category for the subject based on the set of available features and predetermined emotional state categories. The method and apparatus of this application can reduce interference, characterize computer activity from multiple perspectives, and improve assessment performance.

Description

Technical Field

[0001] This application relates to signal processing, and more specifically, to methods, apparatus, and computer storage media for assessing emotional states based on electroencephalogram (EEG) signals. Background Technology

[0002] With societal development, mood-related illnesses have become an increasingly important concern for modern people.

[0003] Routine assessments of emotional states and further diagnosis and treatment of emotional disorders are based on scale assessments and physician experience. Electroencephalography (EEG) records electrical activity generated in the cerebral cortex. It is characterized by high resolution, non-invasiveness, and low cost, and can be used to reflect neural activity in various functional areas (brain regions). Therefore, it has applications in the diagnosis of brain diseases, motor rehabilitation, and as an aid in the assessment of neurological disorders.

[0004] Neurological disorders are brain diseases involving organic lesions and can be detected using equipment and methods that can scan and image brain tissue. Mood disorders, however, are not neurological disorders but rather more psychological, involving unlocalized brain lesions. Therefore, conventional testing methods and data cannot provide satisfactory information. The conventional method of completing questionnaires and then having doctors assess the subject's mood is cumbersome, and the subject's unstable emotional state during questionnaire completion also leads to a certain rate of misdiagnosis. In short, most testing methods cannot pinpoint the specific characteristics of brain activity.

[0005] Current assessment methods based on EEG signals lack universally applicable data and therefore do not offer stable assessment results. In recent years, research on exploring emotions using EEG signals has been on the rise in the fields of psychology and cognitive neuroscience.

[0006] Therefore, there is a need to design and improve assessment schemes for emotional states. Summary of the Invention

[0007] The embodiments of this application aim to solve the problems mentioned above, and propose methods and devices for conducting emotional state assessments.

[0008] According to one aspect of this application, a method for assessing emotional state is proposed, comprising: acquiring raw electroencephalogram (EEG) signals of a subject; generating an available feature set from the EEG signals for assessing the subject's emotional state, wherein the available feature set is selected from a multi-dimensional feature set generated from the EEG signals, the multi-dimensional feature set being associated with at least one of frequency band energy features, brain functional connectivity features, and spectral features of the EEG signals; and using an assessment model to determine an emotional state category corresponding to the subject based on the available feature set and a predetermined emotional state category.

[0009] According to another aspect of this application, a device for assessing emotional states is proposed, comprising: an acquisition unit configured to acquire raw electroencephalogram (EEG) signals of a subject; a feature determination unit configured to generate a set of available features from the EEG signals for assessing the subject's emotional state, wherein the set of available features is selected from a multi-dimensional feature set generated from the EEG signals, the multi-dimensional feature set being associated with at least one of frequency band energy features, brain functional connectivity features, and spectral features of the EEG signals; and an assessment unit configured to determine an emotional state category corresponding to the subject based on the set of available features and a predetermined emotional state category using an assessment model.

[0010] According to another aspect of this application, a computer-readable storage medium is provided that stores a computer program thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0011] According to another aspect of this application, an electronic device is provided, comprising a processor; and a memory for storing executable instructions of the processor, wherein the processor is configured to execute the executable instructions to implement the method as described above.

[0012] By employing the proposed scheme for assessing emotional states, the interference from individual differences among subjects can be minimized through feature construction and screening. This transforms the singular and artifact-prone EEG signals into multi-dimensional features that are relatively stable, allowing for multi-faceted characterization of brain activity. Feature extraction, performed from the perspective of traditional EEG signal analysis, ensures high interpretability and ease of understanding for assessors. The feature screening mechanism in this application yields advantageous features with stable performance across different datasets, thereby enhancing the generalization ability of subsequent assessment tasks. Through this approach, this application provides a relatively stable method and device for assessing and classifying emotional states, presenting the probability and degree of a subject's emotional state category in terms of probability. This scheme can assist physicians in diagnosis to determine if a subject's emotional state involves emotional disorders. Furthermore, for patients with emotional disorders, it can assist in assessing and determining the specific category of depression (MDD) and bipolar disorder (BD). In the assessment scheme, the results of emotional state category assessments can be quantitatively statistically analyzed and visualized, making it more convenient for physicians to refer to during diagnosis. Attached Figure Description

[0013] The above and other features and advantages of this application will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0014] Figure 1This is a schematic flowchart of a method for assessing emotional state according to an embodiment of this application.

[0015] Figure 2 This is a schematic block diagram of a device structure for assessing emotional states according to an embodiment of this application.

[0016] Figure 3 This is a schematic block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0017] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided to make the content of this application comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. In the drawings, the dimensions of some elements may be exaggerated or modified for clarity. The same reference numerals in the drawings denote the same or similar structures, and therefore their detailed descriptions will be omitted.

[0018] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details described, or other methods, elements, etc., can be employed. In other instances, well-known structures, methods, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0019] Traditional assessments and diagnoses of emotional states or illnesses typically employ scale evaluations combined with physician experience. Having subjects actively complete scales (e.g., questionnaires) not only presents challenges due to their emotional state—or even the influence of emotional disorders or illnesses on their state of mind during completion—but the scales themselves may also contain discrepancies between the respondent's subjective interpretation and the objective standards used to create them. While having a physician or specialist interview the subject and complete the scale on their behalf, followed by physician evaluation, can mitigate the issues of subject instability and subjective interpretation to some extent, the involvement of the physician's experience in the evaluation still cannot eliminate subjective factors.

[0020] EEG signal-based emotion assessment helps determine whether a subject's emotional state is normal or likely to have an emotional disorder, or further helps identify the specific emotional disorder (e.g., depression and bipolar disorder). This primarily involves using feature engineering combined with machine learning. The lack of universal data stems from the fact that because subjects of different ages and genders exhibit different signal characteristics in their EEG signals, many EEG-based methods do not correct for age and gender, which can easily lead to biased results.

[0021] This application utilizes EEG signal processing to assess emotions from data sources independent of the scale, without introducing subjective factors. It selects objective indicators from a data-driven perspective that can distinguish different categories of emotional states (e.g., normal mood / emotional disorder (e.g., MDD+BD), depression / bipolar disorder, etc.). Physicians also use EEG signals as an objective standard to assess emotional states and specific disease conditions during the diagnostic process.

[0022] Furthermore, this application, from a feature engineering perspective, covers a broader range of features related to the time domain, frequency domain, and functional connectivity. From these, feature dimensions applicable to more diverse situations (e.g., different genders / ages) are selected, thereby increasing the clinical interpretability of the features. In feature construction and selection, this application aims to minimize the interference of individual differences among subjects to create a relatively stable assessment method. This method can provide auxiliary and supportive data for physicians in diagnosing whether subjects have mood disorders, and whether patients diagnosed with mood disorders belong to a specific category within depression and bipolar disorder. Using machine learning-based classification models and methods, the probability or tendency of a subject's mood state to belong to a specific category of a predetermined mood state / disorder can be obtained. Quantitative statistics and visualization of the assessment results can also improve the convenience for physicians. The assessment results of a subject's mood state provided by the methods and devices provided in this application can assist physicians in making diagnoses based on experience or serve as an important source of analytical data in the physician's diagnostic process.

[0023] Below, we will combine the appendix Figure 1 The process shown in the document describes a method for assessing emotional states according to this application.

[0024] The emotional state assessment method according to the embodiments of this application mainly includes the steps of obtaining the raw electroencephalogram (EEG) signal of the subject, S110, performing data preprocessing on the EEG signal, S120, generating a set of available features from the EEG signal for assessing the emotional state of the subject, S130, and using an assessment model to determine the emotional state category corresponding to the subject based on the set of available features and a predetermined emotional state category, etc.

[0025] First, in step S110, the method acquires EEG signals from multiple lead channels of an EEG device. EEG signals can be acquired using an EEG amplifier according to clinical standards, with a sampling rate of 200 Hz or higher and at least 16 lead channels. These lead channels include, for example, FP1, FP2, F3, F4, C3, C4, P4, O1, F7, T3, FZ, CZ, and PZ. The EEG signals from multiple lead channels are synchronously recorded on the same EEG waveform graph along the time axis (in seconds) for comparison.

[0026] In optional step S120, data preprocessing is performed on the acquired raw EEG signal. The purpose of data preprocessing is to reduce the amount of data computation to improve computational efficiency, eliminate the influence of invalid data and interference on detection, and improve the resolution and accuracy of the EEG signal waveform.

[0027] Preprocessing operations include at least one of rereference, downsampling, notch filtering to power frequency, bandpass filtering, and artifact removal.

[0028] The rereference operation uses the average value AV of the original EEG signals from multiple lead channels as the reference signal, and the EEG signal is calibrated by subtracting the average value of the reference signal (AV value) from the signal of each lead channel.

[0029] Downsampling is used to improve computational efficiency. The downsampling frequency should be at least twice the highest frequency of feature analysis on the EEG instrument, generally above 150Hz, for example, 200Hz can be selected to ensure detection accuracy.

[0030] Notch filtering and bandpass filtering are used to remove noise such as power frequency noise, baseline drift noise, and / or high-frequency noise. Filtering is used to avoid the negative impact of various interference signals on detection. Power frequency noise is generally caused by the inherent frequency of the power system. In my country, the operating frequency of AC power is 50Hz (or 60Hz). Power frequency noise can interfere with the normal operation of electronic equipment and may also overwrite the effective information contained in the EEG signal. Notch filters with operating frequencies of 50Hz and / or multiples of 50Hz can be designed to eliminate power frequency noise interference in EEG signals. In this application, a 50Hz notch filter can be designed to remove power frequency noise. High-pass filters with frequencies greater than or equal to a set value are used to remove baseline drift noise from the signal, and low-pass filters with frequencies less than or equal to a set value are used to remove high-frequency noise from the signal. Baseline drift mainly comes from changes in circuit component parameters caused by the physical characteristics of the channel sensor and environmental factors (such as changes in ambient temperature), which leads to the superposition of components unrelated to the measured signal in the signal input. High-frequency noise originates from electromyography interference and environmental noise.

[0031] The purpose of artifact detection is to select high-quality (or clean) signal regions within an EEG signal. Its basic principle is to select clean portions of the EEG signal that meet predetermined signal quality requirements. In contrast to clean EEG signal regions (or time intervals), there are EEG signal regions or time intervals containing artifacts. Artifact removal can be achieved by detecting anomalous signal segments. Specifically, this involves calculating the corresponding signal parameters of the EEG signal within a sliding window of that window width, and comparing these parameters with corresponding threshold values ​​to determine the locations or time intervals where the signal parameters exceed the thresholds (i.e., do not meet signal quality requirements). These are the artifact locations or artifact time intervals. Then, the EEG signal is segmented or sliced, using a time window of predetermined length (also called window width or step size) and predetermined step size to slide and capture a series of EEG signal segments. The artifact percentage represents the proportion of the EEG signal containing artifacts, and it characterizes the quality of the EEG signal. Among these EEG signal segments, those with a satisfactory artifact ratio are selected to obtain a sequence of high-quality (clean) EEG signal segments without artifacts.

[0032] Next, in step S130, a set of available features for assessing the subject's emotional state is generated from (e.g., preprocessed) EEG signals.

[0033] Step S130 may specifically include a sub-step S131 of generating a multi-dimensional feature set from the EEG signal and a sub-step S132 of selecting available feature combinations from the multi-dimensional feature set that can be used as input to the evaluation model. According to embodiments of this application, the multi-dimensional feature set generated in step S131 is associated with at least one of the following: frequency band energy features (involving the frequency domain), brain functional connectivity features, and spectral features (involving the time domain and / or frequency domain) of the EEG signal.

[0034] Sub-step S131 includes sub-step S1311, which determines the brain region EEG signal 102 corresponding to the subject's brain region based on the brain region and signal quality; sub-step S1312, which calculates the frequency band energy characteristics of the brain region EEG signal based on the brain region EEG signal 102 and a predetermined frequency band; sub-step S1313, which calculates the brain functional connectivity characteristics of the brain region EEG signal based on the brain region EEG signal and the predetermined frequency band; sub-step S1314, which transforms the brain region EEG signal 102 into time-frequency plot data to determine the spectral characteristics of the brain region EEG signal; and sub-step S1315, which calculates the median and complexity of the aforementioned frequency band energy characteristics, brain functional connectivity characteristics, and spectral characteristics of the brain region EEG signal 102 on multiple EEG signal segments to form a multi-dimensional feature set 103 based on the median and complexity of these characteristics.

[0035] Sub-step S1311 involves eliminating abnormal lead channels with unsatisfactory signal quality through lead channel fusion, obtaining brain region EEG signals (also known as brain region representative signals) 102 that can be used to represent each brain region of the subject. The purpose of brain region-specific lead fusion is to address the issue that the electrode pads used to acquire the subject's EEG signals may not be positioned according to the system standard, resulting in deviations. To reduce these deviations, the lead channels of the EEG signals are fused according to brain regions.

[0036] Specifically, sub-step S1311 further includes sub-step S1311a for segmenting / slicing the EEG signal into multiple EEG signal segments, sub-step S1311b for removing lead channels with abnormal signal power values ​​for each lead channel, and sub-step S1311c for fusing the EEG signals of each lead channel with normal brain region signal power values ​​to generate brain region EEG signal 102.

[0037] Sub-step S1311a involves sliding a series of EEG signal segments across the preprocessed EEG signal within a predetermined time window (also known as window width or step size) and with predetermined step increments. Since the artifact removal step in data preprocessing S120 has already performed EEG signal slicing or segmentation, multiple generated EEG signal segments can be reused if artifact removal is present. Otherwise, sub-step S1311a performs slicing / segmentation on the EEG signal.

[0038] Next, in sub-step S1311b, the corresponding EEG signals of leads with abnormal signal power values ​​are removed. For each EEG signal segment, the signal power value is examined for all lead channels (e.g., 19 channels) without distinguishing brain regions. A preset power threshold range is set, which represents the normal overall range of signal power values ​​for lead channels within an EEG signal segment. If the power value of a lead within the time interval of that signal segment is higher or lower than the preset power threshold range, or if the signal power value is outside the preset power threshold range, then the signal power value of that lead channel in that EEG signal segment is considered abnormal, and this is called an abnormal EEG signal interval. If the ratio of the length of the abnormal EEG signal interval to the total interval length of the EEG signal segments examined in that lead channel (also called the abnormal time percentage) is greater than a preset percentage threshold (e.g., 80%), then the EEG signal of that lead channel is considered to have an abnormal signal power value and needs to be removed.

[0039] In sub-step S1311c, for each brain region, the EEG signals (composed of multiple EEG signal segments) of all leads identified as having normal power values ​​are fused to form a brain region EEG signal 102 representing that brain region. The brain region EEG signal 102 has the form of an EEG signal time series composed of EEG signal segments. The fusion operation can include various methods, such as weighted fusion of EEG signals, specifically weighted summation, or taking the mean, median, etc. of the EEG signals. In this way, each brain region has a brain region EEG signal 102 representing the characteristic information of that brain region.

[0040] Returning to sub-step S1311, subsequent sub-steps S1312, S1313, and S1314 involve feature construction based on EEG signals from brain regions 102. First, a predetermined frequency band needs to be defined, i.e., the band of interest to the researcher. For example, the frequency band of interest can be selected based on frequency ranges such as delta (1-4Hz), theta (4-8Hz), alpha (8-12Hz), and beta (12-35Hz), or several narrow bands from 2-20Hz. Some studies have mentioned that certain EEG signal features related to mood disorders may be more pronounced in certain frequency bands; for example, the asymmetry of the alpha band signal in the prefrontal cortex is a marker for distinguishing depression. However, this research has not formed a scientific consensus because, based on the dataset, not all patients with depression exhibit this marker.

[0041] Sub-step S1312 is used to calculate the characteristics of the frequency domain angle. The frequency band energy characteristic is a feature representing the relative frequency band energy of the brain region's EEG signal 102 after calculation relative to a reference frequency band energy. The reference frequency band can be set according to a predetermined reference frequency band of interest to the researcher. The calculation method for the reference frequency band energy is the same as that for the absolute frequency band energy, only the bandwidth is different. For example, the reference frequency band can be set to delta (1-4Hz) and the absolute frequency band to 1-20Hz, or the reference frequency band can be set to theta (4-8Hz) and the absolute frequency band to alpha (8-12Hz).

[0042] The reason for using relative frequency band energy characteristics is that absolute energy is easily affected by the signal amplitude of the subject; for example, the absolute energy of young children is generally higher than that of young adults. Looking at a single data point, the EEG signal amplitude range may differ for each lead channel, resulting in different calculated absolute energies. This can introduce biases from the acquisition environment or equipment. Using relative frequency band energy characteristics aims to reduce the influence of background signals.

[0043] The calculated frequency band energy features include at least one of three dimensions: relative frequency band energy features of brain regions, relative frequency band energy asymmetry of brain regions, and relative frequency band energy complexity of brain regions.

[0044] For multiple EEG signal segments in the brain region's EEG signal, the overall frequency band is divided into multiple frequency bands of interest as described above within the time interval of each signal segment, and the relative frequency band energy of each frequency band is calculated to obtain the relative frequency band energy characteristics of the brain region.

[0045] The relative frequency band energy asymmetry of brain regions is an indicator of asymmetry obtained by calculating the difference in relative frequency band energy between symmetrical left and right brain regions within the time interval of each EEG signal segment. The left and right hemispheres are divided into symmetrical, paired groups of brain regions, such as the left and right frontal lobes, left and right parietal lobes, and left and right occipital lobes. The difference in relative frequency band energy of the EEG signals from the symmetrical left and right brain regions can be directly obtained by subtracting the relative frequency band energy of the left and right symmetrical brain regions, serving as a characteristic of the relative frequency band energy asymmetry of brain regions.

[0046] The relative frequency band energy complexity of brain regions is generated by further processing the characteristic sequences of the relative frequency band energy of brain regions, which correspond to the time series of frequency segments of brain region EEG signal segments, using an entropy-based algorithm.

[0047] Sub-step S1313 is used to calculate the features of the brain functional connectivity angle. This calculation is still performed using the predetermined frequency band of interest from sub-step S1312.

[0048] The calculated brain functional connectivity features include at least one of two dimensions: amplitude connectivity features and phase connectivity features.

[0049] The calculation of amplitude connectivity features includes frequency band filtering based on a predetermined frequency band, then calculating the correlation coefficient between brain region EEG signals 102 at low frequencies to obtain a signal correlation coefficient matrix, and calculating the signal envelope correlation coefficient matrix of the signal envelope of brain region EEG signals 102 at high frequencies.

[0050] In the low-frequency case, the signal correlation coefficient matrix is ​​calculated directly for each EEG signal segment. The signal correlation coefficient matrix has a dimension of M*M, where M is the number of brain regions. Each EEG signal segment corresponds to one signal correlation coefficient matrix. Since the other half of the signal correlation coefficient matrix is ​​symmetrically repeated, its upper triangular matrix can be used. Assuming there are N EEG signal segments, a connectivity feature sequence of length N is obtained, where the number or dimension of features is M*(M-1) / 2 (i.e., the number of unit values ​​in the upper triangular matrix). Finally, the median and standard deviation or entropy over N dimensions are taken. One of the median and standard deviation (entropy) represents the result of stable amplitude connectivity, and the other represents the result of amplitude connectivity variation complexity. Both results have a dimension of M*(M-1) / 2. Assuming the number of predetermined frequency bands of interest is N_freqband, then (N_freqband*(M*M-1) / 2) amplitude connectivity features of number or dimension are obtained.

[0051] In high-frequency cases, the signal correlation coefficient matrix is ​​replaced with the signal envelope correlation coefficient matrix. The reason for calculating the signal envelope correlation coefficient matrix instead of directly calculating the signal correlation coefficient matrix as in the low-frequency case is that this application focuses on the occurrence of high frequencies rather than the details within them. Even slight shifts in the peaks of a high-frequency signal can significantly alter the correlation coefficient, but this correlation information is not very useful in the application scenario of this application. By selecting the signal envelope correlation coefficient matrix, the correlation coefficient is higher when high-frequency signals occur simultaneously in both lead channels, thus reflecting the correlation information of high-frequency occurrence.

[0052] The calculation of phase connectivity features involves, for each signal segment in the EEG signal fragment, dividing it into predetermined frequency bands of interest as described above, extracting the phase values ​​of the EEG signals 102 from each brain region by frequency band, and calculating phase-locked values ​​to obtain a value used to evaluate the frequency consistency between brain regions as the phase connectivity feature. The phase-locked value represents the consistency of the signal in the corresponding frequency band within the time interval of the EEG signal segment in the brain region, and its value range is [0,1], where a value closer to 1 indicates a higher consistency of the rhythm in that frequency band. Similar to the calculation of amplitude connectivity features, based on the obtained phase-locked value sequence between brain regions in each frequency band, the correlation calculation for amplitude is transformed into the calculation of the phase-locked value for phase, thereby obtaining the phase connectivity features in high-frequency and low-frequency cases. According to embodiments of this application, the median and standard deviation / entropy of the data sequence of connectivity features can be further calculated.

[0053] Sub-step S1314 is used to calculate the spectral features of the combined time and frequency domains. First, the signal segment time series of the brain region EEG signal 102 is transformed into a time-frequency plot using a Fourier transform (e.g., short-time Fourier transform, STFT), and then the spectral features are calculated based on the time-frequency plot.

[0054] The calculated spectral features include at least one of multiple dimensions such as spectral entropy, spectral centroid, spectral bandwidth, spectral peak value, spectral skewness, spectral kurtosis, spectral stretch, spectral flatness, spectral slope, and spectral attenuation. For multiple EEG signal segments, the median value of each spectral feature on these signal segments and the entropy value, which measures complexity, can be calculated as further feature dimensions.

[0055] Then, in step S1315, the above-mentioned frequency band energy features, brain functional connectivity features, and spectral features of the calculated brain region EEG signal 102 constitute an original feature set. New features are obtained by calculating the median (or mean value) and complexity of the features in these original feature sets over multiple EEG signal segments (i.e., the time domain), thus forming a multi-dimensional feature set 103. The median in the time domain represents the stable representation of the feature across the entire EEG signal, and the complexity represents the variation of the feature across the entire EEG signal. The dimensions of each feature in the obtained multi-dimensional feature set 103 are shown in the table below:

[0056]

[0057] Table 1. Number of feature dimensions in the multidimensional feature set

[0058] Since the median and complexity of each feature are calculated separately, the number or dimension of multi-dimensional features involving multiple aspects such as time domain, frequency domain, and functional connectivity for each subject included in the multi-dimensional feature set 103 is 2*(number of brain regions * number of frequency bands * number of frequency band energy feature subcategories + number of frequency bands * number of connectivity locations * number of functional connectivity feature subcategories + number of brain regions * number of spectral feature subcategories). These multi-dimensional features constitute the feature sample data set.

[0059] According to embodiments of this application, for the acquired EEG signal data or historical EEG signal data, features with the number of dimensions shown in Table 1 can be labeled based on a predetermined emotional state category 101 for subsequent feature selection and model training. The predetermined emotional state category 101 can be a binary classification category, such as {normal mood, emotional disorder}, or, if it is determined that the subject has an emotional disorder (i.e., the subject is a patient with an emotional illness), the predetermined emotional state category 101 becomes {depression (MDD), bipolar disorder (BD)}. If the binary classification categories are represented by 0 and 1 respectively, the label can be abbreviated as {0,1}.

[0060] The process now returns to sub-step S131. After obtaining the multi-dimensional feature set 103, sub-step S132 is executed to select an available feature set from the multi-dimensional feature set.

[0061] Sub-step S132 further includes sub-step S1321, which selects advantageous features from the multi-dimensional feature set to form the original (or first) candidate feature set; sub-step S1322, which transforms the labeled emotional state category of at least one feature in the multi-dimensional feature set to shuffle the emotional state category label; sub-step S1323, which selects advantageous features again based on the shuffled label multi-dimensional feature set to form the pseudo (second) candidate feature set; sub-step S1324, which determines the available feature set 104 based on the features in the original candidate feature set and the pseudo candidate feature set; and sub-step S1325, which corrects the feature distribution of features with non-Gaussian distribution in the determined available feature set 104.

[0062] Sub-step S1321 specifically includes sub-step S1321a, which groups subjects based on their age and gender to facilitate stratified sampling; sub-step S1321b, which extracts feature samples from the feature sample dataset; sub-step S1321c, which determines the ranking of a feature in each feature sample batch; sub-step S1321d, which determines the final ranking of a feature in the multi-dimensional feature set 103; and sub-step S1321e, which selects dominant features. Sub-steps S1321b and S1321c are steps that are executed multiple times, the number of iterations depending on the number of rounds of extraction from the feature sample dataset. Figure 1 As shown in the dashed box on the right side of the middle.

[0063] In sub-step S1321a, the feature samples are grouped (stratified) from the feature sample data set. Considering that features may have different distributions in different age and gender groups, according to the embodiments of this application, the feature samples labeled as two emotional state categories are divided into several age groups, while also distinguishing between gender groups. Age can be divided by elderly, middle-aged, youth, and adolescence, or by age range (e.g., over 80 years old, 70-80 years old, 60-70 years old, and so on, or other age intervals).

[0064] The group / stratified sampling used in this application can be carried out in multiple rounds (e.g., Nr rounds), so sub-steps S1321b and S1321c will be performed Nr times.

[0065] In sub-step S1321b, based on the age / gender grouping, the group with the smallest amount of feature samples (which includes feature samples of brain region EEG signals 102 from m subjects, with each feature sample having the dimension calculated as shown in Table 1 above) can be selected as the number of samples chosen from each group in each round of sampling. In each round of sampling, feature samples of brain region EEG signals 102 from m subjects are extracted from each group. Therefore, the total number of feature samples extracted in one round is m*N_group, where N_group is the number of groups. The feature sample extraction from the groups is carried out with replacement, meaning that in the next round of extraction, the number of feature samples in each group remains the same as when the grouping was completed in sub-step S1321a. The repeated replacement method is applied to each round of sampling. If a group in the previous round has m EEG signal 102 feature samples taken without replacement, then the number of subjects in this group must be greater than the number of sampling rounds Nr*m to ensure that there are feature samples available for extraction in each round of sampling.

[0066] The total number of feature samples in each round of extraction, m*N_group, constitutes a batch of feature sample data. A total of Nr rounds of extraction constitute Nr batches of feature sample data. Nr batches of feature sample data constitute an epoch of feature sample data.

[0067] Next, in sub-step S1321c, the ranking of each feature in the multi-dimensional feature set 103 in the corresponding feature sample batch is determined. That is, for each round of sampling in multiple rounds or for each loop of sampling, the ranking of the feature in the corresponding feature sample batch is determined. Sub-step S1321c further includes sub-step S1321b1, which determines the first ranking of each feature in the feature sample batch based on the importance of each feature in the cross-validation classification experiment; sub-step S1321b2, which determines the second ranking of each feature in the feature sample batch based on the effect size of each feature among the predetermined emotion state categories; and sub-step S1321b3, which determines the third ranking of each feature in the feature sample batch based on the information entropy of each feature among the predetermined emotion state categories.

[0068] Sub-step S1321c1 uses a supervised learning classifier to determine the feature importance ranking, i.e., the first ranking, in a cross-validation-based classification experiment. The supervised learning classifier model used in this step is used to obtain the importance of features for classifying emotional states; therefore, any model that can output features and their corresponding weights or importance results can be used. For example, a random forest classifier can output weights, and a logistic regression model can output the coefficients corresponding to each feature. Ranking based on these weights or coefficient values ​​yields the feature importance ranking value (integer).

[0069] When calculating ranking values, a deduplication operation can be performed to select features from the top 20 based on a predetermined percentage. This predetermined percentage can be a ranking percentage. For example, if several features all have a ranking value of 10, and the top 20 features are required according to the predetermined ranking percentage, then features with a ranking value tied at 10 will be selected. It should be noted that when using a classifier to rank feature importance in sub-step S1321c1, feature samples from each group are grouped into a single batch for classification; therefore, differences between different age and / or gender groups are not considered.

[0070] For cross-validation classification experiments, a batch of feature sample data is divided into K parts. One part forms the test dataset, and the other K-1 parts serve as the training dataset. The classification experiment is performed K times, and the ranking of the features is obtained each time. The median or average of the K rankings is then taken as the result of the cross-validation experiment. Theoretically, the scores of the feature samples in the test and training datasets can be adjusted. For example, if two feature samples are used as the test dataset, then the training dataset would consist of K-2 feature samples.

[0071] In substep S1321c2, the difference between features of each dimension and the predetermined emotional state categories is determined based on statistical tests, and the effect size of the feature is obtained. The effect sizes of features across all dimensions are ranked to obtain a second ranking. The effect size is represented by Cohen's d and is applicable to continuous data (e.g., continuous EEG signal fragment sequences in this application). It standardizes the effect size to quantify the degree of association between variables. The effect size can be used to compare changes in the data itself before and after, or to compare differences between groups of data. In this application, the function of comparing differences between groups of effect size is used. The effect size in this application is taken as the absolute value of Cohen's d to determine the effect size between two categories in which the feature belongs to the predetermined emotional state category, and is determined by the following formula (1):

[0072]

[0073] In the formula, d is Cohen's d. and These represent the average number of feature samples belonging to two categories within a predetermined emotional state category. Then it is the difference between the averages, S pooled Let n1 be the summative standard deviation of the number of feature samples belonging to the two categories, and n2 be the number of feature samples in each batch belonging to the first and second categories, respectively. and This represents the standard deviation of each feature within each of the two emotion state categories. It should be noted that the parameters... and and S pooled The calculation needs to be performed after all Nr rounds of extraction are completed.

[0074] In substep S1321c3, the third rank of each feature in the feature sample batch is determined based on the information entropy of each feature between predetermined emotional state categories. Similar to substep S1321c2, the information entropy of each feature between two emotional state categories is also calculated in this step. The information entropy is calculated using JS divergence. JS divergence is symmetric and its value is between 0 and 1. The formula (2) for calculating JS divergence is as follows:

[0075]

[0076] Wherein, KL is calculated according to formula (3):

[0077]

[0078] Here, P1 and P2 are the probability distributions of the feature in the two emotion state categories, respectively. KL divergence (relative entropy) measures the difference between two separate probability distributions of feature x (P is the true distribution, Q is the predicted distribution), specifically the difference between P1 and P2. JS divergence is the similarity between two probability distributions P1 and P2, addressing the asymmetry issue of KL divergence. For each feature, the JS divergence of the entropy between the two categories is calculated. For each batch of feature samples, the JS divergence of the entropy is sorted across all features, resulting in the third ranking (integer value) mentioned above.

[0079] After Nr rounds of extraction, in sub-step S1321d, based on the first, second, and third rankings of the corresponding feature sample batches in each round of extraction, the final ranking of each feature in the multi-dimensional feature set 103 is determined.

[0080] Specifically, for the first, second, and third rankings calculated in each round of sampling, the median of each ranking value can be used as the final first, second, and third rankings, respectively. Alternatively, the average of these ranking values ​​can be used as the final first, second, and third rankings, respectively. However, when using the average, the standard deviation statistic also needs to be introduced. For example, if the average and standard deviation are used, sub-steps S1321c2 and S1321c3 need to calculate the average and standard deviation of the corresponding parameters after all Nr rounds of sampling are completed, then calculate the effect size or JS divergence of each feature in each round of sampling, and finally determine the second or third ranking of each feature in each round of sampling.

[0081] Multi-round sampling is used to identify features that are relatively stable and advantageous across several different distributions. Compared to using only the median to determine the final first, second, and third rankings, the mean, a combination of the mean and standard deviation, is susceptible to outliers. For example, if a feature ranks highly in one round of sampling but remains in the middle in most other rounds, the final ranking obtained using the mean cannot represent the feature's stable performance across multiple rounds of sampling. The median, however, does not have this problem. Therefore, if the mean is used to calculate the final rankings for the first through third positions, the standard deviation must be added to reassess the dominant features and eliminate those with large fluctuations, indicating that they do not meet the intended purpose of selecting stable, high-ranking features.

[0082] After obtaining the final values ​​of the first to third rankings, in step S1321e, features that meet the dominance criteria in the final ranking are selected as dominant features for dominant feature integration. Specifically, the average of the final first, second, and third rankings for each feature can be taken to integrate the results of different ranking methods, obtaining the final ranking value for each feature to generate the final ranking list. Further, several top-ranked features can be selected from the final ranking list to form a candidate feature set. For example, features ranked in the top β can be selected, where the value of β is adjustable; for example, it can be set to 0.35, representing the top 35% of features.

[0083] After the initial selection of dominant features in sub-step S1321, an original (first) candidate feature set is obtained. To improve the feature selection effect in the available feature set, in the subsequent sub-step S1322, the emotional state category label of at least one feature in the multi-dimensional feature set is transformed or shuffled, and an adjusted multi-dimensional feature set is obtained through feature type substitution. After adjustment, the new multi-dimensional feature set constitutes a feature sample data set with incorrect classification labels. Obviously, this incorrect feature sample data set will also affect the grouping operation of age and gender. Theoretically, the selection result of dominant features obtained from the incorrect feature sample data set is inaccurate, and those features that rank highly in the incorrect feature ranking of this inaccurate result should not belong to the dominant features in the available feature set used in this application to evaluate emotional state.

[0084] In sub-step S1323, the process described in sub-steps S1321a to S1321e is executed again based on the adjusted multi-dimensional feature set to re-select the advantageous features to form a pseudo (second) candidate feature set. The same settings can be used as those for the selection process of the advantageous features described in sub-step S1321, and will not be repeated here.

[0085] The set of pseudo-candidate features includes advantageous features with inaccurate rankings determined based on incorrect classification labels; these are pseudo-advantageous features. The purpose of sub-step S1323 is to find features that were not originally advantageous but still obtained high rankings during the screening process in sub-step S1321, using the erroneous feature sample data set.

[0086] Then, in sub-step S1324, the features in the candidate (first) feature set selected in sub-step S1321 are compared with the features in the pseudo (second) feature set selected in sub-step S1323 to remove features that appear in both the candidate feature set and the pseudo feature set. After the above removal operation, the remaining features in the candidate feature set are the accurate and usable strength features that are truly associated with the subject's emotional state assessment, constituting the usable feature set 104.

[0087] After obtaining the set of available features 104, distribution correction can be performed on the available features. Generally, the feature distributions of available features include Gaussian and non-Gaussian distributions. For feature dimensions whose distributions do not conform to a Gaussian distribution, in step S1325, non-Gaussian features can be transformed into a normal or approximately normal distribution through operations such as exponential transformation. For each dimension of the non-Gaussian distribution, a Gaussian transformation is performed on all data values ​​corresponding to that dimension feature. In this way, the distributions of all features in the set of available features 104 become Gaussian or normal / approximately normal distributions at the same scale. The purpose of transforming to a Gaussian distribution is to improve the accuracy of the evaluation model in step S140 in determining the emotional state category. When the data value distribution ranges of features in various dimensions differ greatly in the original EEG signal data, directly using data from the original EEG signal for modeling and evaluation may highlight the role of features with larger numerical scales in modeling, while relatively weakening or ignoring the role of features with smaller numerical scales. Therefore, in order to ensure the effectiveness and reliability of the evaluation and classification model, it is necessary to scale the feature values ​​of the features from the original EEG signal data (i.e., the features in the available feature set 104 obtained in step S1324) so ​​that these features have the same or similar weights and influences on the objective function of the evaluation model used to evaluate the emotional state.

[0088] The formula (4) for the transformation operation used for feature distribution correction is as follows:

[0089]

[0090] Where x represents the original distribution of the feature, and y represents the corrected feature distribution. In the embodiments of this application, γ can be selected as 0.2. γ is obtained by observing the QQ plot of the feature. According to the embodiments of this application, the value of γ can also be determined by plotting and observing the data after data transformation using different γ values ​​as parameters. Generally speaking, the closer the feature curve in the QQ plot is to the curve y = x, the more it represents that the γ parameter can make the distribution of the feature closer to a Gaussian distribution (Gaussianization). Alternatively, the optimal γ value can be found by iterating through all possible γ values ​​using the KS test.

[0091] Now, returning to step S140, the evaluation model is used to determine the emotional state category corresponding to the subject based on the distribution-corrected available feature set 104 obtained in sub-step S1325 and the predetermined emotional state category 101.

[0092] According to embodiments of this application, the evaluation model may include a Bayesian model and a regression model. Using the aforementioned available feature set 104 as input and a predetermined emotion state category as output, a Bayesian model and a regression model are established respectively. The parameters of each model are adjusted using a grid search algorithm to obtain the optimal set of evaluation model parameters. Those skilled in the art will understand that other types of supervised classification models capable of predicting classifications and their probabilities based on predetermined emotion state categories for the input feature set can also be used. For example, the Bayesian model can be further adapted to use a Naive Bayes model, and the regression model can be further adapted to use a logistic regression model.

[0093] The prediction and evaluation process of the evaluation model described in step S140 specifically includes a sub-step S141: using a Bayesian model to determine the emotional state category and corresponding classification probability corresponding to the subject, where the classification result is denoted as the first emotional state category and the corresponding first classification probability; a sub-step S142: using a regression model in parallel or sequentially to determine the emotional state category and corresponding classification probability corresponding to the subject, where the classification result is denoted as the second emotional state category and the corresponding second classification probability; and a sub-step S143: weighted fusion of the classification probabilities of the two models to determine the final evaluation result, i.e., the subject's final emotional state category. The weighting weights 105 of the weighted fusion can be determined as follows: the original feature data set from the available feature set 104 is divided multiple times into a training data set and a validation data set; cross-validation is used to obtain the classification effect of the validation data set under each division; and the weighting weights used in the weighted fusion are determined based on the classification effect (i.e., the accuracy of the classification). Generally, the weighting weight corresponding to the most accurate classification result should be selected.

[0094] The training process of the two models and the determination of the weighting weights will be described in detail below.

[0095] Suppose the training data set input to evaluate the model is T = {(X i Y i ), ..., (X n Y n )}. Among them, X i =(X i (1),...,X i (n)) T X i (j) is the feature on the j-th dimension of the i-th feature sample (from the optional feature set 104), where Y∈{0,1} is the predetermined emotional state category. Similar to the above, the emotional state category here is a binary classification, for example, it can be {normal mood, mood disorder}, or it can be {depression (MDD), bipolar disorder (BD)}, with the two categories labeled 0 and 1 respectively.

[0096] During the training of the Naive Bayes model, the conditional probability of each feature sample in the training dataset is first calculated, and then the joint probability of the feature sample between two predetermined emotional state categories is obtained. Since the distribution of these dominant features in the available feature set 104 has been corrected in sub-step S1325 (the corrected distribution is Gaussian or normal / approximately normal), the conditional probability of these feature samples can be obtained using the Gaussian distribution model of the features.

[0097] The formula for calculating the conditional probability of a feature sample is as follows:

[0098]

[0099] in, This indicates that the emotional state category is y. k The mean of the i-th feature in the feature sample. The emotional state category is represented by y. k The variance of the i-th feature in the feature sample.

[0100] Formula (6) is used to calculate the emotional state category c of n feature samples in the training dataset. k Joint probability:

[0101]

[0102] Among them, c k For emotional state categories (c k (∈{0,1}).

[0103] Then, by normalizing the joint probability of each emotional state category according to the following formula (7), we can obtain the classification probability of the above training data set under the binary classification problem as defined by the predetermined emotional state classification, calculated by the Naive Bayes model:

[0104]

[0105] For binary classification problems, N = 2.

[0106] The trained Naive Bayes model can determine the first emotional state category and the first classification probability as determined in sub-step S141.

[0107] In the training process of the logistic regression model, P(Y=1|x) represents the posterior probability estimate that feature sample x is a positive sample in the feature space, and correspondingly, the posterior probability estimate that feature sample x is a negative sample is 1-P(Y=1|x). The posterior probability that feature sample x is a positive sample in the feature space can be calculated using the following formula (8):

[0108]

[0109] Here, w and b are posterior parameters that need to be determined through model training.

[0110] The loss function J(x) of the model is determined using the maximum likelihood estimation method according to formula (9).

[0111]

[0112] Where x is the input feature sample, and n is the number of feature samples in the training dataset.

[0113] Then, the gradient descent method is used to iteratively solve for the parameter combination that minimizes the loss function J(x), i.e., maximizes the data likelihood. During the iteration process, ||J(x)|| is calculated. k+1 )-J(x k When the value of || is less than a set threshold, the loss function converges to its minimum value, and the corresponding parameters are the optimal solution for the parameters of the logistic regression model. According to an embodiment of this application, the threshold can be set to, for example, 0.001. In this way, the parameters of the regression model can be optimized based on the minimum average loss function value of the logistic regression model.

[0114] The trained logistic regression model can determine the second emotional state category and the second classification probability as determined in sub-step S142.

[0115] After the Naive Bayes model and the Logistic Regression model are trained, the fusion weights are calculated to combine the emotion state classification results (first and second emotion state classifications) and their classification probabilities (first and second classification probabilities) of the two models into the final emotion state category.

[0116] Formula (10) gives the method for calculating the fusion weight β:

[0117] P = βP nb +(1-β)P logic (10)

[0118] Among them, P nb P represents the probability that a feature sample obtained based on the Naive Bayes model belongs to the corresponding emotional state category. logic This refers to the probability that feature samples obtained based on a logistic regression model belong to the same emotional state category. In this application, L candidate values ​​{β} for the fusion weights are selected within the interval [0, 1]. i ,...,β L}

[0119] The original feature data set from the available feature set 104 is divided into K parts, and the i-th part is used as the test data set V.i Where i∈{1,2,…,n}, the remaining K-1 parts are used as the training data set T. i The optimal value of the fusion weight β is determined based on a cross-validation experiment similar to that described above.

[0120] Specifically, using the training data set T i After training the Naive Bayes model and the Logistic Regression model, the test dataset V is then used. i The classification probabilities of the two models under the L candidate values ​​of the fusion weights are obtained respectively. The loss function value is calculated based on formula (9) to obtain the classification probability of the model in the test dataset V. i The loss function value Γ on i =(J i (j),…,J i (n)) T , where j∈{β i ,...,β L}

[0121] Test data set V i In this context, i takes values ​​from 1 to K, and the training process for both models is repeated, with the loss function value calculated under the corresponding fusion weight candidate values. This yields the loss function value Γ = {Γ1...Γ} corresponding to the test data set from the K experiments. k}

[0122] Then, the optimal candidate value of the fusion weight β corresponding to the minimum value of the average loss function is determined using the following formula (11), which is used as the final weight 105 for probabilistic fusion.

[0123]

[0124] At this point, the weighting weight 105 used in sub-step S143 is determined, enabling the evaluation model to determine the subject's emotional state category and its final classification probability by weighted fusion of the classification results of the two models.

[0125] It should be noted that the method for assessing emotional state described in this application can be executed based on a binary classification system, setting {normal mood, mood disorder} as a predetermined emotional state category 101, to obtain an assessment result indicating whether the subject's mood is normal or that a mood disorder exists. Then, using a binary classification system, setting {depression (MDD), bipolar disorder (BD)} as a new predetermined emotional state category 101 in the method, the method for assessing emotional state described in this application is executed again for the EEG signal data of subjects (e.g., patients) who were identified as having a mood disorder in the first assessment, to obtain a further assessment result indicating whether the mood disorder of these subjects belongs to depression (MDD) or bipolar disorder (BD).

[0126] Figure 2 A device 200 for assessing emotional states according to an embodiment of this application is shown.

[0127] The device 200 may include an acquisition unit 210 for acquiring raw electroencephalogram (EEG) signals of a subject, a feature determination unit 220 for generating a set of available features from the EEG signals for assessing the subject's emotional state, and an assessment unit 230 for determining the emotional state category corresponding to the subject based on the set of available features and a predetermined emotional state category using an assessment model.

[0128] In addition to completing Figure 1 The sub-step S110 shown in the diagram can also be completed by the data preprocessing sub-step S120 by the acquisition unit 210. Optionally, the data preprocessing sub-step S120 can also be completed in the feature determination unit 220.

[0129] Feature determination unit 220 mainly completes Figure 1 The sub-step S130 shown generates a set of available features, and the functions implemented by the more specific sub-steps (e.g., sub-steps S131 to S132, S1311 to S1315, S1311a to S1311c, S1321 to S1325, S1321c1 to S1321c3, etc.) further included in this sub-step.

[0130] Evaluation Unit 230 completed Figure 1 The functions of sub-step S140 shown, and the functions implemented by more specific sub-steps in this step (e.g., sub-steps S141 to S143).

[0131] Those skilled in the art will understand that the details of these steps and sub-steps have been incorporated above. Figure 1 The flowchart of the method described in this application has been presented and will not be repeated here.

[0132] The device 200 may also include an output unit (not shown) that presents the assessment results to the user in a manner such as visualization. Furthermore, the device 200 may also include an input unit (not shown) that receives user input in various interactive ways. For example, the user can input a predetermined emotional state category 101 to set a binary classification for the device 200 to assess the subject's emotional state category, and the device 200 can also automatically complete assessment tasks under multiple predetermined emotional state categories according to a pre-defined process.

[0133] By employing the proposed scheme for assessing emotional states, the interference from individual differences among subjects can be minimized through feature construction and screening. This transforms the singular and artifact-prone EEG signals into multi-dimensional features that are relatively stable, allowing for multi-faceted characterization of brain activity. Feature extraction, performed from the perspective of traditional EEG signal analysis, ensures high interpretability and ease of understanding for some features. The feature screening mechanism in this application yields advantageous features with stable performance across different datasets, thereby enhancing the generalization ability of subsequent assessment tasks. Through this approach, this application provides a relatively stable method and device for assessing and classifying emotional states, and presents the probability and degree of a subject's emotional state belonging to different categories in terms of probability. This scheme can assist physicians in diagnosis to determine whether a subject's emotional state involves emotional disorders, and for patients with emotional disorders, it can further assist in assessing and determining whether the disorder falls under specific categories of depression (MDD) and bipolar disorder (BD). In the assessment scheme, the assessment results of emotional state categories can be quantitatively statistically analyzed and visualized, making it more convenient for doctors to refer to during diagnosis and treatment. For example, doctors can further assess the degree of related emotional disorders or diseases based on the probability values ​​between emotional state categories given by the scheme in this application.

[0134] It should be noted that although several modules or units of the system for assessing emotional states are mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units. Components shown as modules or units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0135] In exemplary embodiments of this application, a computer-readable storage medium is also provided, on which a computer program is stored, the program including executable instructions that, when executed by, for example, a processor, can implement the steps of the method for assessing emotional states described in any of the above embodiments. In some possible implementations, various aspects of this application can also be implemented as a program product including program code that, when run on a terminal device, causes the terminal device to perform the steps described in this specification for the method of assessing emotional states according to various exemplary embodiments of this application.

[0136] The program product for implementing the above-described method according to embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this application is not limited thereto. In this document, a readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0137] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0138] The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0139] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0140] In an exemplary embodiment of this application, an electronic device is also provided, which may include a processor and a memory for storing executable instructions of the processor. The processor is configured to perform the steps of the method for assessing emotional states in any of the above embodiments by executing the executable instructions.

[0141] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0142] The following reference Figure 3 To describe an electronic device 300 according to this embodiment of the present application. Figure 3 The electronic device 300 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0143] like Figure 3 As shown, the electronic device 300 is presented in the form of a general-purpose computing device. The components of the electronic device 300 may include, but are not limited to: at least one processing unit 310, at least one storage unit 320, a bus 330 connecting different system components (including storage unit 320 and processing unit 310), a display unit 340, etc.

[0144] The storage unit stores program code that can be executed by the processing unit 310, causing the processing unit 310 to perform the steps described in the method for assessing emotional states according to various exemplary embodiments of this application. For example, the processing unit 310 can perform actions such as... Figure 1The steps are shown in the figure.

[0145] The storage unit 320 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 3201 and / or a cache storage unit 3202, and may further include a read-only memory unit (ROM) 3203.

[0146] The storage unit 320 may also include a program / utility 3204 having a set (at least one) program module 3205, such program module 3205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0147] Bus 330 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0148] Electronic device 300 can also communicate with one or more external devices 400 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 300, and / or with any device that enables electronic device 300 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 350. Furthermore, electronic device 300 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 360. Network adapter 360 can communicate with other modules of electronic device 300 via bus 330. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 300, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0149] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the method for assessing emotional states according to the embodiments of this application.

[0150] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the appended claims.

Claims

1. A method for assessing emotional states, comprising: Obtain raw electroencephalogram (EEG) signals from the subjects; A set of available features for assessing the subject's emotional state is generated from the EEG signal, including: Generate a multi-dimensional feature set from the EEG signal; and The available feature set is selected from the multi-dimensional feature set, wherein dominant features are selected from the multi-dimensional feature set, and the selected dominant features constitute a first candidate feature set; the emotional state category labeled by at least one feature in the multi-dimensional feature set is modified; dominant features are selected again from the multi-dimensional feature set, and the selected dominant features constitute a second candidate feature set; the available feature set is determined by removing dominant features from the second candidate feature set from the first candidate feature set; the feature distribution of features with non-Gaussian distribution in the available feature set is transformed into a normal distribution or an approximately normal distribution; wherein the multi-dimensional feature set is associated with at least one of the frequency band energy features, brain functional connectivity features, and spectral features of the EEG signal; The evaluation model is used to determine the emotional state category corresponding to the subject based on the available feature set and the predetermined emotional state category.

2. The method of claim 1, wherein, Generating a multi-dimensional feature set from the EEG signal further includes: The brain region EEG signal corresponding to the subject's brain region is determined based on the brain region and signal quality, wherein the brain region EEG signal includes multiple EEG signal segments; The multidimensional feature set is generated from the EEG signals of the brain region.

3. The method of claim 2, wherein, Determining the EEG signal corresponding to the subject's brain region based on brain region and signal quality further includes: For each lead channel of the EEG signal, remove the lead channel with abnormal signal power value; For each brain region of the subject, EEG signals from lead channels with normal signal power values ​​were fused to generate the EEG signal for that brain region.

4. The method of claim 3, wherein, Generating the multi-dimensional feature set from the EEG signals of the brain region further includes: The frequency band energy characteristics of the brain region EEG signal are calculated based on the brain region EEG signal and a predetermined frequency band. The frequency band energy characteristics include at least one of the following: relative frequency band energy characteristics of the brain region, relative frequency band energy asymmetry of the brain region, and relative frequency band energy complexity of the brain region. Brain functional connectivity features of the brain region EEG signals are calculated based on the brain region EEG signals and the predetermined frequency band, wherein the brain functional connectivity features include at least one of amplitude connectivity features and phase connectivity features; The brain region EEG signal is transformed into time-frequency plot data to determine the spectral characteristics of the brain region EEG signal. These spectral characteristics include at least one of spectral entropy, spectral centroid, spectral bandwidth, spectral peak value, spectral skewness, spectral kurtosis, spectral stretching, spectral flatness, spectral slope, and spectral attenuation. The median and complexity of the frequency band energy features, brain functional connectivity features, and spectral features of the EEG signals of the brain regions on the multiple EEG signal segments are calculated to form the multidimensional feature set, wherein the median and the complexity are labeled based on the predetermined emotional state category.

5. The method of claim 1, wherein, Selecting dominant features from the multi-dimensional feature set further includes: The multidimensional feature set is grouped based on the age and gender of the subjects; Multiple rounds of feature sample extraction are performed from the grouped multi-dimensional feature set, wherein the feature samples obtained in each round of feature sample extraction constitute a feature sample batch. For each batch of feature samples, determine the ranking of each feature in the multi-dimensional feature set within the corresponding batch of feature samples; The final ranking of each feature in the multi-dimensional feature set is determined based on the ranking of each feature in all feature sample batches. The features that satisfy the dominance criteria in the final ranking are selected as dominant features.

6. The method of claim 5, wherein, The multi-round feature sample extraction from the grouped multi-dimensional feature set further includes: In each round of feature sample extraction, feature samples with the minimum number of features are extracted from all groups based on the number of features in the group with the minimum number of features; and Before the next round of feature sample extraction, the extracted feature samples from each group are replaced.

7. The method of claim 5, wherein, Determining the rank of each feature in the multi-dimensional feature set within the corresponding feature sample batch further includes: For each batch of feature samples, Each feature is ranked first in its sample batch based on its importance in the cross-validation classification experiment. Each feature is ranked second in the sample batch of that feature based on the effect size of each feature among the predetermined emotional state categories; The third ranking of a feature in a batch of features is determined based on the information entropy of each feature among the predetermined emotional state categories.

8. The method of claim 7, wherein, The classification experiments were conducted using a supervised learning classifier model, including a random forest classifier.

9. The method of claim 7, wherein, The effect size includes Cohen's d.

10. The method of claim 7, wherein, The information entropy is calculated based on JS divergence.

11. The method of claim 7, wherein, Determining the final ranking of each feature in the multi-dimensional feature set based on the ranking of each feature across all feature sample batches further includes: The final ranking of each feature in the multi-dimensional feature set is determined based on the median or average of the first, second, and third rankings of each feature across all feature sample batches.

12. The method of claim 1, wherein, Determining the available feature set based on the first candidate feature set and the second candidate feature set further includes: The available feature set is formed by removing the features included in the second candidate feature set from the first candidate feature set.

13. The method of claim 1, wherein, The evaluation model includes a Bayesian model and a regression model. Determining the emotional state category corresponding to the subject based on the available feature set and predetermined emotional state categories using the evaluation model further includes: The Bayesian model is used to determine the first emotional state category and the first classification probability corresponding to the subject based on the available feature set and the predetermined emotional state category; The regression model is used to determine the second emotional state category and the second classification probability corresponding to the subject based on the available feature set and the predetermined emotional state category; The first classification probability and the second classification probability are weighted and fused to determine the final emotional state category corresponding to the subject.

14. The method of claim 13, wherein, During the training phase of the Bayesian model: The conditional probability of a feature sample is determined based on the feature distribution of each feature sample in the training dataset. Calculate the joint probability of the feature sample on the training dataset based on the conditional probability; The joint probability is normalized to determine the classification probability of the training data set under the predetermined emotional state category.

15. The method of claim 13, wherein, During the training phase of the regression model, the parameters of the regression model are optimized based on the minimum average loss function value of the regression model.

16. The method according to claim 13, characterized in that, The weights for weighted fusion are determined using the training dataset through cross-validation.

17. The method of claim 1, wherein, Before generating a set of available features from the EEG signal for assessing the subject's emotional state, the method further includes: The original EEG signal is subjected to data preprocessing, which includes at least one of downsampling, rereference, notch filtering to remove power frequency, bandpass filtering, and artifact removal.

18. The method of any one of claims 1 to 17, wherein, The predetermined emotional state categories include: Normal mood, mood disorder; or Depression, bipolar disorder.

19. The method of claim 18, wherein, The method further includes: If the predetermined emotional state categories include normal emotion and emotional disorder, determine the emotional state category corresponding to the subject; and / or For subjects whose emotional state category is determined to be emotional disorder, the predetermined emotional state category is set to include depression and bipolar disorder, and the emotional state category corresponding to the subject is determined again.

20. A device for assessing emotional states, comprising: The acquisition unit is configured to acquire the subject's raw electroencephalogram (EEG) signals; A feature determination unit is configured to generate a set of available features from the EEG signal for assessing the subject's emotional state, comprising: generating a multi-dimensional feature set from the EEG signal; selecting the available feature set from the multi-dimensional feature set, wherein dominant features are selected from the multi-dimensional feature set, and the selected dominant features constitute a first candidate feature set; modifying the labeled emotional state category of at least one feature in the multi-dimensional feature set; again selecting dominant features from the multi-dimensional feature set, and the selected dominant features constitute a second candidate feature set; determining the available feature set by removing dominant features from the second candidate feature set from the first candidate feature set; transforming the feature distribution of features with non-Gaussian distributions in the available feature set to a normal or approximately normal distribution, wherein the multi-dimensional feature set is associated with at least one of the frequency band energy features, brain functional connectivity features, and spectral features of the EEG signal; An assessment unit is configured to use an assessment model to determine the emotional state category corresponding to the subject based on the available feature set and a predetermined emotional state category.

21. The apparatus of claim 20, wherein, The device is configured to implement the method according to any one of claims 2 to 19.

22. A computer-readable storage medium having a computer program stored thereon, the computer program including executable instructions that, when executed by a processor, implement the method according to any one of claims 1 to 19.

23. An electronic device, comprising: include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the executable instructions to implement the method according to any one of claims 1 to 19.

Images