Mood disorder assessment system based on multi-level feature fusion

Abstract

The invention provides a mood disorder assessment system based on multi-level feature fusion, and the system comprises a data collection unit which is used for collecting electroencephalogram signals of a plurality of brain regions of a to-be-assessed patient; the electroencephalogram feature extraction unit is used for extracting electroencephalogram features corresponding to the electroencephalogram signals of the brain regions; the multi-level feature extraction unit is constructed on the basis of the symptom features of the multiple testees and the corresponding electroencephalogram features, and is used for performing multi-level electroencephalogram feature latent variable extraction on the electroencephalogram features of the brain regions of the patient to be evaluated to obtain multi-level electroencephalogram feature latent variables; and the feature fusion and classification unit is used for carrying out classification prediction based on the electroencephalogram feature latent variables to obtain a mood disorder assessment result of the patient to be assessed. The method solves the problem that a mood disorder assessment system in the prior art adopts a single feature extraction and learning strategy and has no constraint of symptom information, so that the recognition capability of a model for mood disorders of different functional abnormality types is limited.

Description

Technical Field

[0001] This invention belongs to the field of medical informatics technology, and in particular relates to a multi-level feature fusion mood disorder assessment system. Background Technology

[0002] Electroencephalography (EEG) is a commonly used method for recording electrical activity generated in the brain cortex. It is characterized by high resolution, non-invasiveness, and low cost, and can reflect neural activity in various functional areas of the human brain. Therefore, it has applications in the diagnosis of brain diseases, motor rehabilitation, and assessment of neurological disorders. In recent years, research using EEG to explore mood disorders has also shown an upward trend in the fields of psychology and cognitive neuroscience. However, current EEG-based mood disorder assessment systems typically rely on single feature extraction and learning strategies, failing to fully capture the unique functional patterns of each brain region. This results in limited model recognition capabilities for mood disorders with different functional abnormalities. Furthermore, existing mood disorder assessment systems lack constraints from the clinical symptom dimension when extracting EEG features. This leads the model to focus on learning discriminative features while neglecting disease-related information hidden within electrophysiological indicators, affecting its ability to identify different diseases with similar electrophysiological indicator distributions but different symptoms. Summary of the Invention

[0003] Based on the above analysis, the present invention aims to provide a multi-level feature fusion mood disorder assessment system to solve the problem that the existing technology uses a single feature extraction and learning strategy and lacks symptom information constraints, resulting in the model's limited ability to identify mood disorders of different functional abnormalities.

[0004] The objective of this invention is mainly achieved through the following technical solutions: On one hand, the present invention provides a multi-level feature fusion mood disorder assessment system, comprising: The data acquisition unit is used to acquire electroencephalogram (EEG) signals from multiple brain regions of the patient to be evaluated. The EEG feature extraction unit is used to extract the EEG features corresponding to the EEG signals of each brain region. The multi-level feature extraction unit is constructed based on the symptom characteristics and corresponding EEG characteristics of multiple subjects. It is used to extract multi-level latent variables of EEG characteristics of each brain region of the patient to be evaluated, and obtain multi-level latent variables of EEG characteristics. The feature fusion and classification unit is used to perform classification prediction based on the latent variables of the EEG features to obtain the mood disorder assessment results of the patient to be assessed.

[0005] Furthermore, the EEG features corresponding to the EEG signals include: frequency band energy features, connectivity features, amplitude-phase coupling features, time-spectrum features, microstate features, central rhythm features, and VAEEG features.

[0006] Furthermore, the extraction of the EEG features includes: By using bandpass filtering or power spectral density estimation, the signal power of each frequency band is calculated, and the frequency band energy characteristics of each frequency band are obtained. The connectivity features are obtained by calculating the synchronicity of signals in different brain regions in frequency or phase using coherence and phase-locking methods. The low-frequency phase and high-frequency amplitude signals are extracted using Hilbert transform, and the coupling strength between them is calculated using the modulation index to obtain the amplitude-phase coupling characteristics. By using time-frequency analysis, a dynamic spectrum of signal power changes with time and frequency is obtained, and the time-frequency characteristics are obtained based on the dynamic spectrum. Clustering of multichannel EEG topographic maps at the peak of GFP yields several template microstates. The duration, coverage, and transition patterns of each state are quantified by backfitting to obtain the microstate features. The central rhythm characteristics are obtained by calculating the change in power of a specific frequency band of the sensorimotor cortex relative to the baseline period during a motor task; Using a VAEEG pre-trained model, the potential representations of each 1-second EEG signal within the 1-30Hz frequency band are extracted to obtain the VAEEG features.

[0007] Furthermore, the multi-level feature extraction unit includes: The primary symptom-oriented latent variable extraction module is used to extract multiple primary EEG latent variables that are most relevant to global symptom features for each brain region and the whole brain. ; The secondary symptom-guided specific latent variable extraction module is used to extract symptom-specific secondary EEG feature latent variables corresponding to each brain region and the whole brain. ; The third-level subtype fusion latent variable extraction module is used for the extraction of the second-level EEG feature latent variables. Clustering was performed to obtain multiple electroencephalographic subtypes corresponding to each brain region. Based on each EEG subtype, feature extraction and fusion were performed to obtain a three-level EEG feature latent variable of fused symptom subtype features. .

[0008] Furthermore, the first-level symptom-guided latent variable extraction module includes multiple first-level global regression models corresponding to the number of brain regions. The model parameters of the first-level global regression models are obtained through partial least squares iterative training based on the global symptom features corresponding to multiple subjects and the EEG features corresponding to each brain region. The model parameters include the EEG feature loading matrix. and model weight matrix .

[0009] Furthermore, the first-level global regression model extracts the first-level symptom-oriented latent variables for each brain region using the following method: S1: Standardize and initialize various EEG features; ,make , ;in, X represents the EEG feature X updated for the h-th time with the latent variable dimension H as the target; S2: Will and Multiplication yields latent variables , is represented as: m is the number of samples; S3: Update the EEG input features for the next round based on the load matrix, as follows: ; S4: From arrive S2-S3 are executed sequentially to obtain the primary symptom-guided latent variables corresponding to each brain region and the whole brain.

[0010] Furthermore, the secondary symptom-oriented specific latent variable extraction module includes multiple secondary symptom-specific fitting models corresponding to the number of symptom features. The secondary symptom-specific fitting model Based on the aforementioned primary EEG latent variables The corresponding symptom features are obtained through iterative training using partial least squares, and are used to infer the secondary symptom-specific latent variables corresponding to each symptom feature. .

[0011] Furthermore, the secondary symptom-specific fitting model It is constructed using the following method: Based on the EEG feature load matrix corresponding to each brain region and the whole brain Latent variables of primary EEG characteristics of each brain region and the whole brain Perform an inverse transformation to obtain the reconstructed original features. ; Based on the difference between the original EEG features corresponding to each brain region and the whole brain and the reconstructed original features, the first-order residual features corresponding to each brain region and the whole brain are obtained. ; For each symptom feature, the first-order residual feature is used separately. The corresponding second-order symptom-specific fitting model is obtained by training using partial least squares method. .

[0012] Furthermore, the third-level subtype fusion latent variable extraction module includes: The three-level subtype clustering module is used to perform clustering based on the secondary EEG feature latent variables. Clustering was performed to obtain multiple electroencephalographic subtypes corresponding to each brain region and the whole brain. and the posterior probability of each subtype. ; The latent variable extraction module from the perspective of subsymptom subtypes is used to infer subtype latent variables based on the PLS projection model of each pre-trained EEG subtype. ; The subtype fusion module from the perspective of subsymptoms is used to combine the posterior probabilities of each EEG subtype. The values ​​are converted into deterministic weights, and the third-level EEG feature latent variables are obtained based on these deterministic weights. .

[0013] Furthermore, the symptom characteristics are obtained based on mood disorder-related assessment scales, and the mean of each symptom characteristic for each subject is used as the global symptom characteristic.

[0014] The beneficial effects of this technical solution are: 1. The multi-level feature fusion mood disorder assessment system provided by this invention adopts a strategy of multi-level modeling for EEG and symptom-guided fusion, and a strategy of dividing brain regions. An independent EEG feature latent variable extraction model is established for each brain region. Each brain region uses features generated by EEG electrodes corresponding to its spatial location. Different levels of models are used for feature extraction to capture the unique functional patterns of each brain region. Global coordination is achieved through fusion strategy, realizing EEG feature extraction based on symptom information constraints, and improving the model's ability to assess mood disorders of different functional abnormalities. 2. This invention utilizes multi-dimensional clinical symptom indicators and employs a multi-layered modeling strategy. It not only establishes a global symptom fitting projection module but also simultaneously constructs linear fitting models for each of the multi-dimensional clinical features. Using the projection matrix as the feature matrix, it fuses the fitting results from multiple clinical symptom dimensions to assess mood disorders, thus improving the accuracy of the assessment. Furthermore, in practical applications, it can simultaneously output fitting results from multiple clinical symptom dimensions, providing doctors with symptom references to aid in diagnosis. 3. This invention establishes a mapping relationship between EEG features and the symptom space through a fitting projection mechanism. Clinical indicators are only used to train the projection matrix, realizing the conversion from EEG to the symptom semantic space. This ensures that the final classification is mainly based on objective electrophysiological indicators, maintaining the objectivity of the diagnosis. Attached Figure Description

[0015] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a schematic diagram of a multi-level feature fusion mood disorder assessment system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the training process of the multi-level feature extraction unit in an embodiment of the present invention. Detailed Implementation

[0016] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0017] One embodiment of the present invention provides a multi-level feature fusion mood disorder assessment system, such as... Figure 1 As shown, it includes: The data acquisition unit is used to acquire electroencephalogram (EEG) signals from multiple brain regions of the patient to be evaluated. The EEG feature extraction unit is used to extract the EEG features corresponding to the EEG signals of each brain region. The multi-level feature extraction unit is constructed based on the symptom characteristics and corresponding EEG characteristics of multiple subjects. It is used to extract multi-level latent variables of EEG characteristics of each brain region of the patient to be evaluated, and obtain multi-level latent variables of EEG characteristics. The feature fusion and classification unit is used to perform classification prediction based on the latent variables of the EEG features to obtain the mood disorder assessment results of the patient to be assessed.

[0018] Preferably, the data acquisition unit in this embodiment uses a clinically standard EEG amplifier to acquire EEG signals, with a sampling rate range of 200-1000Hz and a channel value range of 16-32 channels.

[0019] This includes multiple brain regions, such as the frontal, parietal, occipital, left temporal, right temporal, and central regions.

[0020] After acquiring the EEG signals of the patient to be evaluated, data preprocessing is performed using a data preprocessing unit, specifically including: Resampling, rereference, notch filtering to remove power frequency, removal of fundamental drift, removal of invalid high-frequency components, and ICA (Inductively Coupled Aspect) method to remove electrooculogram (EOG), electromyogram (EMG), and electrocardiogram (ECG) components, as well as abnormal amplitude segments, are performed. The processed segments are then divided into sample segments of fixed time lengths. The removal of abnormal amplitude segments includes: time-slicing the bandpass-filtered EEG data, calculating the amplitude of the sliced ​​samples, and deleting time slices with amplitudes outside a preset range. In this embodiment, 120 seconds is used as the target slice length.

[0021] Furthermore, the EEG features extracted by the EEG feature extraction unit include: frequency band energy features, connectivity features, amplitude-phase coupling features, time-spectrum features, microstate features, central rhythm features, and VAEEG features.

[0022] The following methods were used to extract EEG features: By using bandpass filtering or power spectral density estimation, the signal power of each frequency band is calculated, and the frequency band energy characteristics of each frequency band are obtained. The connectivity features are obtained by calculating the synchronicity of signals in different brain regions in frequency or phase using coherence and phase-locking methods. The low-frequency phase and high-frequency amplitude signals are extracted using Hilbert transform, and the coupling strength between them is calculated using the modulation index to obtain the amplitude-phase coupling characteristics. By using time-frequency analysis, a dynamic spectrum of signal power changes with time and frequency is obtained, and the time-frequency characteristics are obtained based on the dynamic spectrum. Clustering of multichannel EEG topographic maps at the peak of GFP yields several template microstates. The duration, coverage, and transition patterns of each state are quantified by backfitting to obtain the microstate features. The central rhythm characteristics are obtained by calculating the change in power of a specific frequency band of the sensorimotor cortex relative to the baseline period during a motor task; Using a VAEEG pre-trained model, the potential representations of each 1-second EEG signal within the 1-30Hz frequency band are extracted to obtain the VAEEG features.

[0023] After extracting the EEG features corresponding to each brain region, the multi-angle EEG features corresponding to different brain regions are grouped according to the brain region extents of the frontal, parietal, occipital, left temporal, right temporal, and central regions, thus obtaining the EEG feature set corresponding to each brain region. Furthermore, a global EEG feature set based on spatial location was obtained across the entire brain. Finally, the EEG feature sets of each brain region were obtained. and global EEG feature set The overall original feature set G is represented as: { }, used for subsequent model building and training, where, For frontal region EEG feature set, For the central region EEG feature set, For the parietal region EEG feature set, For the occipital region EEG feature set, The left temporal EEG feature set, The right temporal EEG feature set, This is a global EEG feature set.

[0024] Furthermore, the multi-level feature extraction unit includes: The primary symptom-oriented latent variable extraction module is used to extract multiple primary EEG latent variables that are most relevant to global symptom features for each brain region and the whole brain. ; The secondary symptom-guided specific latent variable extraction module is used to extract symptom-specific secondary EEG feature latent variables corresponding to each brain region and the whole brain. ; The third-level subtype fusion latent variable extraction module is used for the extraction of the second-level EEG feature latent variables. Clustering was performed to obtain multiple electroencephalographic subtypes corresponding to each brain region. Based on each EEG subtype, feature extraction and fusion were performed to obtain a three-level EEG feature latent variable of fused symptom subtype features. .

[0025] The first-level symptom-guided latent variable extraction module includes multiple first-level global regression models corresponding to the number of brain regions. The model parameters of the first-level global regression models are obtained through partial least squares iterative training based on the global symptom features corresponding to multiple subjects and the EEG features corresponding to each brain region. The model parameters include the EEG feature loading matrix. and model weight matrix .

[0026] Specifically, the first-level global regression model extracts the first-level symptom-oriented latent variables for each brain region using the following method: S1: Standardize and initialize various EEG features; ,make , ;in, X represents the EEG feature X updated for the h-th time with the latent variable dimension H as the target; S2: Will and Multiplication yields latent variables , is represented as: m is the number of EEG segment samples; S3: Update the EEG input features for the next round based on the load matrix, as follows: ; S4: From arrive S2-S3 are executed sequentially to obtain the primary symptom-guided latent variables corresponding to each brain region and the whole brain.

[0027] Furthermore, the secondary symptom-oriented specific latent variable extraction module includes multiple secondary symptom-specific fitting models corresponding to the number of symptom features. The secondary symptom-specific fitting model Based on the aforementioned primary EEG latent variables The corresponding symptom features are obtained through iterative training using partial least squares, and are used to infer the secondary symptom-specific latent variables corresponding to each symptom feature. .

[0028] Among them, the secondary symptom-specific fitting model It is constructed using the following method: Based on the EEG feature load matrix corresponding to each brain region and the whole brain Latent variables of primary EEG characteristics of each brain region and the whole brain Perform an inverse transformation to obtain the reconstructed original features. ; Based on the difference between the original EEG features corresponding to each brain region and the whole brain and the reconstructed original features, the first-order residual features corresponding to each brain region and the whole brain are obtained. ; For each symptom feature, the first-order residual feature is used separately. The corresponding second-order symptom-specific fitting model is obtained by training using partial least squares method. .

[0029] Furthermore, the third-level subtype fusion latent variable extraction module includes: The three-level subtype clustering module is used to perform clustering based on the secondary EEG feature latent variables. Clustering was performed to obtain multiple electroencephalographic subtypes corresponding to each brain region and the whole brain. and the posterior probability of each subtype. ; The latent variable extraction module from the perspective of subsymptom subtypes is used to infer subtype latent variables based on the PLS projection model of each pre-trained EEG subtype. ; The subtype fusion module from the perspective of subsymptoms is used to combine the posterior probabilities of each EEG subtype. The values ​​are converted into deterministic weights, and the third-level EEG feature latent variables are obtained based on these deterministic weights. .

[0030] Special, such as Figure 2As shown, each module of the multi-level feature extraction unit of the present invention is pre-constructed using the following method: First, multiple mood disorder-related symptom characteristics were obtained from multiple subjects; and EEG characteristics of multiple brain regions and the whole brain of each subject were collected to construct the first training sample set. Specifically, for symptom characteristics, multiple mood disorder-related assessment scales were obtained from the participants. These scales include, but are not limited to, HAMD17 and HCL33. The symptom assessment results for each symptom item were obtained based on the assessment scales, and these results were used as the symptom characteristics for each symptom item. The symptom items used in this embodiment are shown in Table 1. Table 1: Examples of Symptom Items

[0031] The aforementioned method was used to collect EEG signals from multiple brain regions and the entire brain of each subject. Multi-angle EEG features (bandwidth energy features, connectivity features, amplitude-phase coupling features, time-spectrum features, microstate features, central rhythm features, and VAEEG features) were further extracted from the preprocessed EEG signals. Based on the brain region extents of the frontal, parietal, occipital, left temporal, right temporal, and central regions, the multi-angle EEG features corresponding to different brain regions of different subjects were grouped to obtain the EEG feature sets corresponding to each brain region. Furthermore, a global EEG feature set based on spatial location was obtained across the entire brain. .

[0032] After obtaining the first training sample set, a multi-level feature extraction unit is constructed based on the first training sample set, including the sequential construction of a first-level symptom-oriented latent variable extraction module, a second-level symptom-oriented specific latent variable extraction module, and a third-level subtype fusion latent variable extraction module, and the symptom-oriented EEG feature latent variables of each level are obtained. Preferably, the primary symptom-oriented latent variable extraction module is constructed using the following method: The mean of each symptom feature corresponding to each subject is used as the global symptom feature; Based on the global symptom features and the corresponding EEG features of each brain region and the whole brain, a first-level global regression model corresponding to each brain region and the whole brain is obtained by training using partial least squares. A first-level latent variable screening unit is constructed to obtain multiple first-level latent EEG feature variables that are most correlated with the global symptom features for each brain region and the whole brain based on the projected weights W of the first-level global regression model. ; Based on the first-level global regression model and the first-level latent variable screening unit, the first-level symptom-oriented latent variable extraction module is obtained.

[0033] Specifically, this embodiment uses the mean of the symptom set as the global symptom feature, with a dimension of (m, 1), where m represents the sample size. As global symptom features, for each brain region and the whole brain's EEG feature set, as well as the global EEG feature set, a corresponding first-level global regression model was trained using partial least squares (PLS), including: S21: Standardize the various EEG characteristics and symptom characteristics, and represent them as follows: , ;initialization and ,make , ;in, The mean of the EEG characteristic X is... y represents the mean of the symptom characteristic. The standard deviation of EEG characteristics, The standard deviation of the symptom characteristics This represents the EEG input features updated for the hth time with latent variable dimension H as the target. This represents the h-th update of the symptom input features, targeting the latent variable dimension H. S22: with As the target variable, based on the formula By maximizing and linear combination and The covariance is calculated to obtain the h-th weight vector. ;in ; Represented as: ; S23: Based on and Obtain latent variables , is represented as: m is the sample size; S24: Based on and latent variables The load of the h-th EEG feature was obtained. , is represented as: ; among which the load This represents the association between the latent variable and X; S25: Based on and latent variables To obtain the load of symptom characteristics , is represented as: ; among which the load This represents the association between the latent variable and y; S26: Receive the next round of input, whereby the next round of input is represented as: , Iteratively execute S2-S5 to obtain the first-level global regression model corresponding to each brain region and the whole brain. and latent variable matrix EEG feature load matrix and model weight matrix Where W represents the brain region Model parameters.

[0034] For each brain region and global EEG features, the corresponding PLS model is obtained sequentially. And obtain the latent variables of global symptom features corresponding to each brain region and the whole brain EEG feature set. The latent variable is obtained from the projection weights W (dimension of original EEG feature dimension * preset latent variable dimension H) trained by PLS. Each column of W is the optimal linear combination direction of the original EEG features. This direction maximizes the correlation between the projection of the EEG feature in that direction and the global symptoms. Therefore, based on the first-level latent variable screening unit, multiple first-level EEG feature latent variables that are most relevant to the global symptom features can be selected. The latent variables obtained by linearly changing the projection weights W can retain the original feature set and the patterns related to global symptoms, while irrelevant patterns are compressed. The preset latent variable dimension H needs to be obtained through multiple comparisons using the cross-validation method.

[0035] Furthermore, the secondary symptom-guided specific latent variable extraction module is constructed using the following method: Based on the EEG feature load matrix corresponding to each brain region and the whole brain Latent variables of primary EEG characteristics of each brain region and the whole brain Perform an inverse transformation to obtain the reconstructed original features. ; Based on the difference between the original EEG features corresponding to each brain region and the whole brain and the reconstructed original features, the first-order residual features corresponding to each brain region and the whole brain are obtained. ; For each symptom feature, the first-order residual feature is used separately. A corresponding second-order symptom-specific fitting model was established using partial least squares method. ; Utilizing the aforementioned secondary symptom-specific fitting model Inferring the secondary symptom-specific latent variables corresponding to each symptom characteristic ; Construct fusion units to merge secondary symptom-specific latent variables corresponding to each symptom. And simultaneously merge the aforementioned primary EEG latent variables. Secondary EEG latent variables were obtained. ; Based on the secondary symptom-specific fitting model The fusion unit is used to obtain the secondary symptom-guided specific latent variable extraction module.

[0036] It should be noted that this embodiment establishes a secondary symptom-oriented specific latent variable extraction module, which decouples the information of the original features and the global symptom features to obtain symptom-specific features.

[0037] For each brain region and the whole brain, the corresponding EEG feature latent variables are used, along with the corresponding EEG feature loading matrix from the first-level global regression model. Perform an inverse transform to reconstruct the original features. This reconstruction of the original features reconstructs the EEG manifestations of global symptom features. In its inverse transformation process, it is based on the corresponding first-level EEG feature latent variables. and EEG feature load matrix Using the reconstruction formula The reconstructed original features are obtained. The difference between these features and the original input EEG features is that the original EEG features include components related to symptom features, as well as individual differences or components related to symptom subtypes. The reconstructed features, however, only include components related to symptom features. Using G-... The first-order residual features are obtained, which contain symptom-specific patterns and individual differences.

[0038] For each symptom item, a second-level symptom-specific fitting model is established using the first-level residual features corresponding to each brain region and the whole brain respectively, through partial least squares method. The potential dimension N of each fitted model also needs to be determined through cross-validation; a pre-trained secondary symptom-specific fitting model is used. Inferring latent variables as symptom-specific latent variables For each brain region and whole-brain feature set, the symptom-specific latent variable and the first-order EEG feature latent variable are combined for each symptom sub-item. Then the input features of the third-level model are obtained. This feature includes both features related to global symptoms and specific features related to corresponding sub-symptoms.

[0039] Furthermore, the three-level subtype fusion latent variable extraction module includes a three-level subtype clustering module, a latent variable extraction module from the perspective of subsymptom subtypes, and a subtype fusion module from the perspective of subsymptoms. The three-level subtype clustering module is constructed using the following method: Based on a preset target cluster size range, secondary EEG latent variables of each brain region and the whole brain are utilized. Train Gaussian mixture models corresponding to different target cluster numbers in sequence; Based on the Bayesian Information Criterion (BIC) value of each Gaussian mixture model, the optimal number of target clusters corresponding to each symptom in each brain region is obtained; The Gaussian mixture model corresponding to the optimal target cluster number is used as a three-level subtype clustering module for inferring clustering results.

[0040] It should be noted that this clustering module uses latent variables of global symptom features and sub-symptom features to establish sub-type features for each sub-symptom. The purpose is to use data to drive the identification of EEG-related subtypes under a given symptom. For each symptom item, corresponding secondary EEG feature latent variables are used in each brain region and in the whole brain subset of EEG features. Features are subjected to Gaussian mixture clustering. The target number of clusters in the Gaussian mixture model is determined using the Bayesian Information Criterion (BIC). Specifically, within a fixed range (from 2 to the maximum candidate value), Gaussian mixture models with different target cluster numbers (number of clusters) are trained sequentially, and the BIC score of each model is recorded. Since a smaller BIC value indicates a better balance between goodness of fit and complexity, the model with the lowest BIC score and the corresponding number of clusters is selected as the optimal solution. After determining the optimal target cluster number, the Gaussian mixture model is trained again using the target cluster number as a parameter. This model is then used to infer the clustering results, including the final EEG physiological subtype (i.e., cluster label). And the posterior probability of each subtype (i.e., the probability of the clustering result). .

[0041] Furthermore, the construction of the latent variable extraction module from the perspective of subsymptom subtypes includes: A secondary symptom-specific fitting model based on each symptom feature in each brain region and the whole brain. The loading matrix parameters for the corresponding second-order symptom-specific latent variables Perform an inverse transformation to obtain the reconstructed specific features. ; Utilizing the first-order residual characteristics Reconstruction-specific features The difference is used as a second-order residual characteristic. Based on the aforementioned electrophysiological subtypes Extract the secondary residual feature samples corresponding to each cluster; Based on the secondary residual feature samples, a corresponding subtype PLS projection model is established for each cluster using partial least squares method. Complete the construction of the latent variable extraction module from the perspective of the subsymptom subtype.

[0042] This embodiment establishes a latent variable extraction module from the perspective of subsymptom subtypes: based on clustering results, a PLS projection model is trained in each subtype to obtain the latent variable projection matrix. The aim is to better integrate subtype information into the feature extraction process.

[0043] For each symptom item, first-order residual features were used in the EEG feature sets for each brain region and the whole brain. Reconstruction-specific features The difference is used as a second-order residual characteristic. Because this feature subtracts components fully related to subsymptoms from the global symptom-related features, it includes specific components for each subsymptom subtype. Based on clustering labels. Extract the corresponding second-order residual features for each cluster. For each sample, a PLS projection model is built for each cluster. Cross-validation was used to determine the target projection dimension. For different subtypes of the same symptom subite in the same brain region, the maximum value of the target projection dimension of all subtypes was used as the target projection dimension of each subtype projection model to ensure the uniformity of feature dimensions of samples across different subtypes. Each subtype PLS projection model was trained and the subtype latent variables were inferred. .

[0044] Furthermore, the subtype fusion module from the subsymptom perspective obtains the three-level symptom subtype fusion features through the following method: The posterior probabilities of each subtype The entropy transformed into a subtype probability distribution is expressed as: ; Calculate maximum entropy Where N represents the number of target subtypes; use This normalizes the entropy into a deterministic fraction. Based on the subtype latent variables The deterministic scores yielded the third-level EEG latent variables. .

[0045] This embodiment establishes a subtype fusion module from the perspective of subsymptoms, uses deterministic computation methods to enhance the feature effects of subtypes with high determinism, and weakens the influence of features of subtypes with low determinism.

[0046] Furthermore, the feature fusion and classification unit in this embodiment is also obtained through pre-training and is used to perform classification prediction based on the latent variables of the EEG features to obtain the mood disorder assessment results.

[0047] The feature fusion and classification unit includes a feature fusion module and a classification module. During training, based on the latent variables of EEG features at various levels of multiple subjects, corresponding mood disorder diagnostic labels are labeled to construct a second training sample set. The feature fusion module and classification module are iteratively trained using the second training sample set. Specifically, the feature fusion module merges the first to third level EEG feature latent variables corresponding to each brain region of each sample using a weighted fusion method, and merges the weighted fused EEG feature latent variables at the brain region level to obtain multi-level symptom-oriented fusion features. A classification module, such as a machine learning or deep learning classifier, is constructed. Utilizing the multi-level symptom-oriented fusion features and corresponding mood disorder diagnostic labels, the feature fusion module and classification module are iteratively trained to obtain the optimal weights of the feature fusion module and a converged classification module. Based on the classification module, the mood disorder assessment result is output.

[0048] During training, the cross-entropy loss function can be used for iterative optimization, but this embodiment does not limit it.

[0049] The mood disorder assessment system constructed based on this embodiment can achieve multi-level EEG feature extraction and fusion based on symptom information constraints without collecting additional assessment scales and symptom feature information from the patients to be assessed in practical applications, thereby improving the model's ability to assess mood disorders.

[0050] In summary, the multi-level feature fusion mood disorder assessment system of this invention employs a multi-level modeling strategy, combining EEG and symptom-guided fusion with brain region segmentation. It establishes independent EEG feature latent variable extraction models for each brain region, using features generated by EEG electrodes corresponding to their spatial location. Different levels of models are used for feature extraction, capturing the unique functional patterns of each brain region. A fusion strategy enables global collaboration, achieving EEG feature extraction based on symptom information constraints, thus improving the model's assessment ability for mood disorders with different functional abnormalities. Furthermore, this invention uses multi-dimensional clinical symptom indicators. Through a multi-level modeling strategy, it not only establishes a global symptom fitting projection module but also constructs linear fitting models for multi-dimensional clinical features, using the projection matrix as the feature matrix and outputting fitting results for multiple clinical symptom dimensions, providing doctors with symptom references for auxiliary diagnosis. By establishing a mapping relationship between EEG features and the symptom space through the fitting projection mechanism, clinical indicators are only used to train the projection matrix, realizing the conversion from EEG to the symptom semantic space. This ensures that the final classification is primarily based on objective electrophysiological indicators, maintaining the objectivity of the diagnosis.

[0051] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-level feature fusion mood disorder assessment system, characterized in that, include: The data acquisition unit is used to acquire electroencephalogram (EEG) signals from multiple brain regions of the patient to be evaluated. The EEG feature extraction unit is used to extract the EEG features corresponding to the EEG signals of each brain region. The multi-level feature extraction unit is constructed based on the symptom characteristics and corresponding EEG characteristics of multiple subjects. It is used to extract multi-level latent variables of EEG characteristics of each brain region of the patient to be evaluated, and obtain multi-level latent variables of EEG characteristics. The feature fusion and classification unit is used to perform classification prediction based on the latent variables of the EEG features to obtain the mood disorder assessment results of the patient to be assessed.

2. The mood disorder assessment system based on multi-level feature fusion according to claim 1, characterized in that, The EEG features corresponding to the EEG signals include: frequency band energy features, connectivity features, amplitude-phase coupling features, time-spectrum features, microstate features, central rhythm features, and VAEEG features.

3. The multi-level feature fusion mood disorder assessment system according to claim 2, characterized in that, The extracted EEG features corresponding to the EEG signals of each brain region include: By using bandpass filtering or power spectral density estimation, the signal power of each frequency band is calculated, and the frequency band energy characteristics of each frequency band are obtained. The connectivity features are obtained by calculating the synchronicity of signals in different brain regions in frequency or phase using coherence and phase-locking methods. The low-frequency phase and high-frequency amplitude signals are extracted using Hilbert transform, and the coupling strength between them is calculated using the modulation index to obtain the amplitude-phase coupling characteristics. By using time-frequency analysis, a dynamic spectrum of signal power changes with time and frequency is obtained, and the time-frequency characteristics are obtained based on the dynamic spectrum. Clustering of multichannel EEG topographic maps at the peak of GFP yields several template microstates. The duration, coverage, and transition patterns of each state are quantified by backfitting to obtain the microstate features. The central rhythm characteristics are obtained by calculating the change in power of a specific frequency band of the sensorimotor cortex relative to the baseline period during a motor task; Using a VAEEG pre-trained model, the potential representations of each 1-second EEG signal within the 1-30Hz frequency band are extracted to obtain the VAEEG features.

4. The mood disorder assessment system based on multi-level feature fusion according to claim 1, characterized in that, The multi-level feature extraction unit includes: The primary symptom-oriented latent variable extraction module is used to extract multiple primary EEG latent variables that are most relevant to global symptom features for each brain region and the whole brain. ; The secondary symptom-guided specific latent variable extraction module is used to extract symptom-specific secondary EEG feature latent variables corresponding to each brain region and the whole brain. ; The third-level subtype fusion latent variable extraction module is used for the extraction of the second-level EEG feature latent variables. Clustering was performed to obtain multiple electroencephalographic subtypes corresponding to each brain region. Based on each EEG subtype, feature extraction and fusion were performed to obtain a three-level EEG feature latent variable of fused symptom subtype features. .

5. The mood disorder assessment system based on multi-level feature fusion according to claim 4, characterized in that, The primary symptom-guided latent variable extraction module includes multiple primary global regression models corresponding to the number of brain regions. The model parameters of the primary global regression models are obtained through partial least squares iterative training based on the global symptom features corresponding to multiple subjects and the EEG features corresponding to each brain region. The model parameters include the EEG feature loading matrix. and model weight matrix .

6. The mood disorder assessment system with multi-level feature fusion according to claim 5, characterized in that, The first-level global regression model extracts the first-level symptom-oriented latent variables for each brain region using the following method: S1: Standardize and initialize various EEG features; ,make , ;in, X represents the EEG feature X updated for the h-th time with the latent variable dimension H as the target; S2: Will and Multiplication yields latent variables , is represented as: m is the number of samples; S3: Update the EEG input features for the next round based on the load matrix, as follows: ; S4: From arrive S2-S3 are executed sequentially to obtain the primary symptom-guided latent variables corresponding to each brain region and the whole brain.

7. The mood disorder assessment system based on multi-level feature fusion according to claim 5, characterized in that, The secondary symptom-oriented specific latent variable extraction module includes multiple secondary symptom-specific fitting models corresponding to the number of symptom features. The secondary symptom-specific fitting model Based on the aforementioned primary EEG latent variables The corresponding symptom features are obtained through iterative training using partial least squares, and are used to infer the secondary symptom-specific latent variables corresponding to each symptom feature. .

8. The mood disorder assessment system based on multi-level feature fusion according to claim 7, characterized in that, The secondary symptom-specific fitting model It is constructed using the following method: Based on the EEG feature load matrix corresponding to each brain region and the whole brain Latent variables of primary EEG characteristics of each brain region and the whole brain Perform an inverse transformation to obtain the reconstructed original features. ; Based on the difference between the original EEG features corresponding to each brain region and the whole brain and the reconstructed original features, the first-order residual features corresponding to each brain region and the whole brain are obtained. ; For each symptom feature, the first-order residual feature is used separately. The corresponding second-order symptom-specific fitting model is obtained by training using partial least squares method. .

9. The mood disorder assessment system based on multi-level feature fusion according to claim 5, characterized in that, The three-level subtype fusion latent variable extraction module includes: The three-level subtype clustering module is used to perform clustering based on the secondary EEG feature latent variables. Clustering was performed to obtain multiple electroencephalographic subtypes corresponding to each brain region and the whole brain. and the posterior probability of each subtype. ; The latent variable extraction module from the perspective of subsymptom subtypes is used to infer subtype latent variables based on the PLS projection model of each pre-trained EEG subtype. ; The subtype fusion module from the perspective of subsymptoms is used to combine the posterior probabilities of each EEG subtype. The values ​​are converted into deterministic weights, and the third-level EEG feature latent variables are obtained based on these deterministic weights. .

10. The mood disorder assessment system based on multi-level feature fusion according to claim 4, characterized in that, The symptom features were obtained based on mood disorder-related assessment scales, and the mean of each symptom feature for each subject was used as the global symptom feature.

Images