A method for constructing a mood disorder evaluation model based on hierarchical multi-level gating

Abstract

The application provides a kind of construction method of mood disorder evaluation model based on hierarchical multi-level gating, comprising: obtaining the electroencephalogram of a plurality of subjects, labeling label, constructing first training sample set;Using first training sample set for training, obtaining a plurality of electroencephalogram parameter combinations respectively corresponding time sequence dynamic graph network feature extraction model and related time sequence dynamic graph network features;Using first training sample set, time sequence dynamic graph network features to construct second training sample set, train hierarchical multi-level gating model;Based on the time sequence dynamic graph network feature extraction model corresponding to each electroencephalogram parameter combination and hierarchical multi-level gating model, obtain mood disorder evaluation model.The application solves the problem that the mood disorder evaluation model in the prior art does not consider the correlation between different parameter combinations and tasks among brain regions, frequency bands and observation time lengths, does not perform hierarchical screening of different parameter brain networks based on task correlation, resulting in limited evaluation accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a construction method of a mood disorder evaluation model based on hierarchical multi-level gating. BACKGROUND

[0002] Electroencephalogram (EEG) is a signal that is convenient to detect, low in cost and can reflect brain activity, and thus can be used as an objective index for diagnosing mental diseases. Meanwhile, an EEG signal diagram is a multi-channel time sequence signal. The position characteristics of the channels can be used to represent brain functional connectivity, and the time sequence characteristics can represent dynamic changes.

[0003] Mood disorders usually involve abnormal functional connectivity between different regions of the brain, and linkage analysis of EEG signals in different brain regions can reveal the brain function state of patients.

[0004] However, in the traditional method, the connectivity is usually obtained by calculating the correlation or similarity between EEG signal channels in the frequency domain or time domain, and such a calculation method is a linear calculation method, which has limitations in representing complex brain connectivity. Moreover, the mood disorder evaluation model based on EEG signals lacks dynamic modeling of brain networks and decoupling of frequency band, brain region and observation time scale in the modeling process, and the accuracy of the evaluation result needs to be improved. SUMMARY

[0005] In view of the above analysis, the present application aims to provide a construction method of a mood disorder evaluation model based on hierarchical multi-level gating, which is used to solve the problem that the mood disorder evaluation model in the prior art does not consider the correlation between different parameter combinations and tasks among brain regions, frequency bands and observation time lengths, and does not perform hierarchical screening of different parameter brain networks based on task correlation, resulting in limited evaluation accuracy.

[0006] The purpose of the present application is mainly achieved by the following technical solutions:

[0007] The present application provides a construction method of a mood disorder evaluation model based on hierarchical multi-level gating, which comprises the following steps:

[0008] Obtaining EEG signals of a plurality of subjects, extracting corresponding EEG features, and labeling mood disorder related labels to construct a first training sample set;

[0009] Preliminary construction of a time sequence dynamic graph network feature extraction model, training using the first training sample set to obtain a plurality of EEG parameter combinations respectively corresponding to a converged time sequence dynamic graph network feature extraction model and time sequence dynamic graph network features related to each EEG parameter combination;

[0010] The first training sample set, the time series dynamic graph network features related to each combination of electroencephalogram parameters in each sample, and the demographic coding features corresponding to the subjects are used to construct a second training sample set;

[0011] The second training sample set is used to train a hierarchical multi-level gating model; the hierarchical multi-level gating model is used to screen task-related electroencephalogram parameter combinations, and the time series dynamic graph network features related to the screened electroencephalogram parameter combinations are used to predict mood disorder evaluation results;

[0012] Based on the trained time series dynamic graph network features extraction model corresponding to each electroencephalogram parameter combination and the hierarchical multi-level gating model, a mood disorder evaluation model is obtained.

[0013] Further, the time series dynamic graph network feature extraction model is trained by the following method:

[0014] The electroencephalogram features in the first training sample set are grouped according to different electroencephalogram parameter combinations to obtain multiple groups of electroencephalogram features;

[0015] A classification module is connected to the initially constructed time series dynamic graph network feature extraction model, and the classification module is used to output a mood disorder prediction result probability distribution;

[0016] Each group of electroencephalogram features is used to train the initially constructed time series dynamic graph network feature extraction model; the loss function is used for iterative calculation to minimize the loss between the prediction result and the labeled label, and the model parameters optimized by each group of electroencephalogram features are saved to obtain the time series dynamic graph network feature extraction model related to each electroencephalogram parameter combination.

[0017] Further, the second training sample set is constructed by the following method:

[0018] The demographic information corresponding to each subject is coded to obtain demographic coding features of each subject;

[0019] The time series dynamic graph network features related to each electroencephalogram parameter combination extracted are hierarchically deleted for redundant data to obtain an alternative time series dynamic graph network feature set indexed by subject ID;

[0020] The first training sample set, the demographic coding features and the time series dynamic graph network features of each subject are combined to obtain a second training sample set.

[0021] Further, the alternative time series dynamic graph network feature set is obtained by the following method:

[0022] The data in the first training sample set is divided into K folds, and the training data in the K folds is iteratively used to obtain the time series dynamic graph network features related to all electroencephalogram parameter combinations;

[0023] The number of fixed samples is based on the time series dynamic graph network features to calculate the correlation between each parameter combination; and the mean of the correlation is calculated in the sample dimension, and the parameter combination with the absolute value of the correlation greater than the first preset threshold is marked;

[0024] After K iterations, remove the parameter combination with a marked number greater than the second preset threshold to obtain a set of candidate time series dynamic graph network features.

[0025] Further, the demographic coding features are obtained by the following method:

[0026] Read the gender, age, and open and closed eye information of the subject when collecting the electroencephalogram signal, and preprocess them by the following method:

[0027] For age information, use standardization processing to unify the numerical range between different features;

[0028] For gender information, use one-hot encoding method to encode male as (0, 1) and female as (1, 0);

[0029] For open and closed eye information, use one-hot encoding method to encode open eye as (0, 1) and closed eye as (1, 0);

[0030] For each dimension of preprocessed information, input the corresponding information coding module composed of multiple fully connected layers, and use the activation function for nonlinear transformation;

[0031] Encode and merge each information obtained after nonlinear transformation, and map it to a unified representation space to obtain demographic coding features.

[0032] Further, the hierarchical multi-level gating module includes a frequency band gating layer, a brain region combination gating layer, and an observation time unit gating layer, which are respectively used to filter the task-related frequency band, brain region combination, and observation time unit; The hierarchical multi-level gating module is trained by the following method:

[0033] Group the electroencephalogram feature data in the second training sample set by frequency band, and merge it with the corresponding demographic coding features to input the frequency band gating layer to predict the task-related frequency band;

[0034] Group the electroencephalogram feature data corresponding to the predicted task-related frequency band by electroencephalogram parameter combination, and merge it with the corresponding demographic coding features to input the brain region combination gating layer to predict the task-related electroencephalogram combination;

[0035] The EEG features obtained by taking the predicted task-related frequency band and brain region combination as a constraint are grouped according to the observation time unit length, and are spliced with the corresponding demographic coding features of the subjects, and are input into the observation time unit gating layer to screen the task-related observation time unit length;

[0036] Based on the screening results of each gating layer, a plurality of task-related EEG parameter combinations corresponding to each gating layer are obtained, and the time sequence dynamic graph network features obtained based on each task-related EEG parameter combination are used for mood disorder evaluation prediction, and the cross-entropy loss of the prediction results of each gating layer and the labeled label is calculated;

[0037] The cross-entropy losses of each gating layer are weighted and fused to obtain the total loss of the hierarchical multi-level gating module, and the total loss is used for iterative optimization to obtain a converged multi-layer hierarchical gating module.

[0038] Further, the total loss is represented as:

[0039] Loss=W Freq *Freq CE +W brain *Brain CE +W Time *Time CE ;

[0040]

[0041]

[0042] The weights of each gating layer during the weighted fusion satisfy the following constraint conditions:

[0043] W Time >W brain >W Freq ;

[0044] Wherein, Loss is the total loss of the model, W Freq is the frequency band gating layer weight, W brain is the brain region gating layer weight, W Time is the time observation unit gating layer weight, Freq CE is the cross-entropy loss of the frequency band gating layer, Brain CE is the cross-entropy loss of the brain region gating layer, Time CE is the cross-entropy loss of the time observation unit gating layer, y is the labeled label, is the predicted value of the frequency band gating layer, is the predicted value of the brain region gating layer, is the predicted value of the observation time unit gating layer.

[0045] Further, the parameters in the electroencephalogram parameter combination include frequency bands, brain region combinations, and observation time unit lengths.

[0046] The frequency bands include delta, theta, alpha, beta1, and beta2 bands.

[0047] The brain regions include frontal regions, parietal regions, central regions, left temporal regions, right temporal regions, and occipital regions.

[0048] The electroencephalogram parameter combination is constructed by the following method:

[0049] Different brain regions are combined in pairs to obtain multiple brain region combinations.

[0050] Multiple observation time units of different lengths are set.

[0051] Different frequency bands, brain region combinations, and observation time units are combined in a preset level order to obtain multiple electroencephalogram parameter combinations and sub-class parameter combinations corresponding to each parameter.

[0052] Further, the frequency band gating layer includes an input layer, a frequency band sub-level dense gating layer, a frequency band sparse gating layer, and a frequency band classification prediction layer.

[0053] The input layer is used to splice the electroencephalogram features corresponding to each frequency band with the demographic coding features and input them into the frequency band sub-level dense gating layer and the frequency band sparse gating layer, respectively.

[0054] The output quantity of the frequency band sub-level dense gating layer is the same as the number of sub-class parameter combinations corresponding to each frequency band, which is used to extract features of the input data and obtain weights corresponding to each sub-class parameter combination through an activation function; based on the weights, time series dynamic graph network features corresponding to each sub-class parameter combination of each frequency band are weighted and fused to obtain fusion features corresponding to each frequency band.

[0055] The output quantity of the frequency band sparse gating layer is the same as the number of frequency bands, which is used to extract features of the input data and obtain a task relevance index of each frequency band through an activation function; the task relevance index is arranged in descending order, and a preset number of task-related frequency bands are selected.

[0056] The frequency band classification prediction layer is used to receive the fusion features corresponding to the task-related frequency bands to predict a frequency band-related classification result.

[0057] Further, the brain region combination gating layer includes an input layer, a brain region sub-level dense gating layer, a brain region sparse gating layer, and a brain region classification prediction layer.

[0058] The input layer is used to flatten the channel dimension and feature dimension of each brain region combination corresponding to the task-related frequency band, and splice with the demographic coding features, and input into the brain region sub-level dense gated layer and the brain region sparse gated layer respectively;

[0059] The output quantity of the brain region sub-level dense gated layer is the number of the alternative parameter combinations composed of each brain region combination corresponding to the task-related frequency band and each observation time unit; the brain region sub-level dense gated layer is used to extract features from the received data, and obtain the weight of each alternative parameter combination corresponding to each brain region combination through an activation function; and based on the weight, the time sequence dynamic graph network features of each alternative parameter combination corresponding to each brain region combination are weighted and fused to obtain the fusion features corresponding to each brain region combination;

[0060] The output layer quantity of the brain region sparse gated layer is the number of brain region combinations, which is used to extract features from the received data, and obtain the task-related degree index of each brain region combination through an activation function, and arrange the task-related degree index in descending order to obtain a preset number of task-related brain region combinations;

[0061] The brain region classification prediction layer is used to receive the fusion features corresponding to the task-related brain region combination to obtain the brain region related classification result.

[0062] The beneficial effects of the technical solution are:

[0063] 1. The present application uses patient information to fuse electroencephalogram features, and assists in screening the corresponding brain network model based on a hierarchical multi-level gated model, thereby improving the effect of cross-subject mood disorder evaluation;

[0064] 2. The present application constructs multiple parameter combinations based on frequency bands, brain region combinations and observation time lengths, and respectively establishes time sequence brain network evaluation models from different angles, which has more comprehensive feature angles and higher feature interpretability and clinical value.

[0065] 3. The present application uses a pre-trained time sequence dynamic graph network feature extraction model to obtain a set of time sequence dynamic graph network features related to different parameter combinations, freezes most of the parameters of the time sequence dynamic network in the training of the gated network, reduces the complexity of the gated training, and to some extent alleviates the problem of uneven expert load.

[0066] 4. The present application selects brain regions related to mood disorders, combines the brain regions in pairs, and combines the length of the frequency band and the observation time unit to construct evaluation data, which can identify the correlation between the linkage features between brain regions and mood disorders based on the underlying features of electroencephalogram signals, and accurately extract the features associated with mood disorders by combining the correlation between frequency bands and observation time, for mood disorder evaluation, greatly improving the accuracy of mood disorder evaluation.

[0067] Other features and advantages of the present application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0069] Figure 1 is a construction method flow chart of a mood disorder evaluation model based on hierarchical multi-level gating of an embodiment of the application;

[0070] Figure 2 is a model training flow chart of an embodiment of the application;

[0071] Figure 3 is an electroencephalogram coding model function diagram of an embodiment of the application;

[0072] Figure 4 is a schematic diagram of constructing a timing dynamic graph network feature based on a combination of electroencephalogram multi-parameters of an embodiment of the application;

[0073] Figure 5 is a mood disorder evaluation system schematic diagram constructed based on a mood disorder evaluation model of an embodiment of the application; DETAILED DESCRIPTION

[0074] The preferred embodiments of the present application will be described in detail with reference to the drawings, wherein the drawings form a part of the specification. The drawings are provided to illustrate the embodiments of the present application and to explain the principles of the present application, but are not intended to limit the scope of the present application.

[0075] One embodiment of the present application provides a construction method of a mood disorder evaluation model based on hierarchical multi-level gating, as shown in Figure 1 The method comprises the following steps:

[0076] Step S1: Obtain electroencephalogram signals of a plurality of subjects, extract corresponding electroencephalogram features, and label labels to construct a first training sample set;

[0077] Specifically, the embodiment adopts a resting state multi-channel scalp EEG acquisition device to collect the EEG physiological signals of the subjects, for constructing a first training sample set. In actual application, if the method is applied to depression / non-depression evaluation, the EEG signals corresponding to a plurality of mood disorder patients and healthy subjects are acquired, and depression or non-depression labels are marked; if the method is applied to unidirectional or bidirectional depression evaluation, the EEG signals corresponding to a plurality of unidirectional depression and a plurality of bidirectional depression mood disorder patients are acquired, and unipolar depression or bipolar depression labels are marked, to construct the first training sample set.

[0078] After the EEG signals are collected, the EEG signals are first subjected to data preprocessing, specifically including downsampling, re-referencing, notch filtering to remove power frequency, band-pass filtering, removing eye movement, heart movement and muscle movement using the ICA method, removing abnormal amplitude segments, and cutting the processed segments into EEG segments of a fixed time length; wherein removing the abnormal amplitude segments includes: first time slicing the EEG data after band-pass filtering, then calculating the amplitude in the time slice, and deleting the time slice with amplitude not in the preset range. In the embodiment, the EEG segment time length is set to 120 seconds, and the average value of the evaluation results of all EEG segments is taken as the final evaluation result.

[0079] As shown in Figure 2 , after the EEG signals are subjected to data preprocessing, each EEG segment of the EEG signals is encoded into a latent representation, for training of a time series dynamic graph network feature extraction model and a hierarchical multi-level gating model.

[0080] Specifically, as shown in Figure 3 , each 1-second single-channel EEG signal is encoded into a latent representation by using the pre-trained EEG encoding model, and the latent representation has a corresponding feature domain for different frequency bands of the EEG. According to the electrode position and brain region division, the EEG signals corresponding to different channels are grouped.

[0081] The embodiment uses an existing EEG encoding model VAEEG to realize encoding of 1-second single-channel EEG signals into a latent representation space, wherein each preprocessed signal will be converted into a feature signal under different frequency bands such as delta [1, 4] Hz, theta [4-8 Hz], alpha [8-13] Hz, beta1 [13-25] Hz, beta2 [25-40] Hz, and the dimension of the latent representation corresponding to each 1-second segment and different frequency bands is delta_zim, theta_zim, alpha_zim, beta1_zim, beta2_zim. According to the division of brain function regions, each brain region corresponding latent representation is divided into the corresponding region according to the corresponding electrode position.

[0082] The embodiment extracts stable features of complex electroencephalogram signals through a dimension reduction method of electroencephalogram signals, and plays a certain denoising effect.

[0083] Step S2: initially constructing a time sequence dynamic graph network feature extraction model, training by using the first training sample set, obtaining a plurality of brain electrical parameter combinations respectively corresponding to a converged time sequence dynamic graph network feature extraction model and a time sequence dynamic graph network feature related to each brain electrical parameter combination extracted;

[0084] Specifically, the time sequence dynamic graph network feature extraction model is used to respectively establish time sequence dynamic graph network sequences of the electroencephalogram signals based on electroencephalogram features corresponding to various brain electrical parameter combinations, and to extract features of each time sequence dynamic graph network sequence, to obtain time sequence dynamic graph network features related to each parameter combination.

[0085] The embodiment constructs a plurality of brain electrical parameter combinations in order to extract more comprehensive feature angles through different dimensions, to extract the linkage between different parameters and the correlation features between mood disorders, and to realize more accurate mood disorder evaluation.

[0086] The parameters in the brain electrical parameter combination include frequency bands, brain region combinations and observation time unit lengths; the frequency bands include delta, theta, alpha, beta1 and beta2 frequency bands; and the brain regions include frontal regions, parietal regions, central regions, left temporal regions, right temporal regions and occipital regions.

[0087] The brain electrical parameter combination is obtained by the following method: different brain regions are combined two by two to obtain a plurality of brain region combinations; a plurality of observation time units with different lengths are set; different frequency bands, brain region combinations and observation time units are combined in a preset level order to obtain a plurality of brain electrical parameter combinations and sub-class parameter combinations corresponding to each parameter.

[0088] Specifically, for the length of the observation time unit, three lengths of short, medium and long are preset, respectively corresponding to observation time unit lengths of 5s, 10s and 15s; the embodiment takes the frequency band class as a first-level parameter, the brain region class as a second-level parameter, and the observation time class as a third-level parameter; a fixed level order is combined with different parameters, and the number of brain electrical parameter combinations obtained is: {number of frequency bands * number of brain region combinations * number of time observations};

[0089] The brain region combinations are connection relationships such as frontal region-parietal region and frontal region-central region, and the number of brain region combinations is:

[0090] For each level of parameters, there is a corresponding sub-class parameter combination. For example, for the corresponding delta band, the number of corresponding sub-class parameter combinations is {number of brain region combination class * number of observation time unit class}, such as: delta_frontal-parietal_short observation unit.

[0091] Further, the time sequence dynamic graph network feature extraction model is obtained by training through the following method:

[0092] The electroencephalogram features in the first training sample set are grouped according to different electroencephalogram parameter combinations to obtain multiple groups of electroencephalogram features.

[0093] A classification module is connected to the initially constructed time sequence dynamic graph network feature extraction model, and the classification module is used to output a mood disorder prediction result probability distribution.

[0094] Each group of electroencephalogram features is used to train the initially constructed time sequence dynamic graph network feature extraction model respectively; the loss function is used for iterative calculation to minimize the loss between the prediction result and the labeled label, the model parameters optimized by each group of electroencephalogram features are saved, and a time sequence dynamic graph network feature extraction model related to each electroencephalogram parameter combination is obtained.

[0095] Further, the time sequence dynamic graph network feature extraction model obtains the time sequence dynamic graph network feature related to each parameter combination through a dynamic edge convolution network and a time sequence embedding module.

[0096] The dynamic edge convolution network is used to construct a dynamic graph network sequence related to various electroencephalogram parameter combinations.

[0097] The time sequence embedding module is used to learn the feature relationship of the dynamic graph network sequence through a long short-term memory network, and to obtain a time sequence dynamic graph network feature through a fully connected layer.

[0098] In particular, the dynamic edge convolution network constructs a dynamic graph network sequence related to each electroencephalogram parameter combination based on the similarity of features through the following method:

[0099] the electroencephalogram signal is divided into multiple observation windows according to the length of the observation time unit in the electroencephalogram parameter combination to be processed; for each observation window, the channel dimension of the electroencephalogram feature corresponding to the electroencephalogram parameter combination is taken as a node of a graph; for each node, the distance between the node and other nodes is calculated by using the Euclidean distance, the distances are arranged in ascending order, n nodes with the closest distances are selected as neighbors, and each neighbor node and the node form an edge; the features of the two nodes corresponding to each edge are spliced to obtain the feature of each edge, and the feature of each edge is converted into the feature of the target quantity by using a multilayer perception machine; the features of the target quantity of all edges corresponding to each node are aggregated, and the aggregated features of all nodes are subjected to a global pooling operation to obtain a one-dimensional feature corresponding to the observation window; and the one-dimensional features corresponding to all observation windows are spliced in chronological order to obtain a dynamic graph network sequence related to the electroencephalogram parameter combination.

[0100] Unlike the analysis of a static brain network based on anatomical connections and functional connections, the embodiment adopts a completely data-driven manner to dynamically construct a graph structure according to the similarity of features in each time window. The construction method of a K-neighbor graph can be used to construct a dynamic graph for each observation window. For each observation window, the input dimension of the electroencephalogram feature is (channel number, feature number), wherein the channel number is the total number of channels included in the corresponding brain region combination. The channel dimension of the electroencephalogram feature is taken as a node of a graph. For each node, the distance between the node and other nodes is calculated by using the Euclidean distance, the distances are arranged in ascending order, the top 4 nodes with the closest distances are selected as neighbors, and each node and the node form an edge to construct a graph structure for each observation window.

[0101] Then, the graph structure is subjected to feature aggregation and conversion, including splicing the features of the two nodes of each edge as new features, connecting a nonlinear multilayer perception machine to convert the new features into the feature of the target quantity, reducing the complexity of the model by using one layer of the multilayer perception machine, and aggregating the features of the target quantity corresponding to each edge, i.e., aggregating the features of the target quantity corresponding to all edges for each node, and the aggregation method adopted in the embodiment is mean aggregation.

[0102] After the message aggregation of each node, the features of all nodes of each observation window are subjected to a global pooling operation, so that the graph structure features of each observation window are compressed into one-dimensional features, denoted as EdgeConv_zim. For a fixed frequency band and a brain region combination, only the internal dynamic changes between the specific frequency band and the specific brain region need to be focused on. For each observation window, a graph structure is obtained, which is a dynamic graph that adaptively aggregates edges and nodes. The graph structures of multiple observation windows are spliced to form a dynamic graph network sequence.

[0103] Preferably, the long short-term memory network is formed by connecting multiple time sequence network units in series.

[0104] The time sequence embedding module is used for learning the feature relationship of the dynamic graph network sequence through a long short-term memory network, and performing feature extraction through a full connection layer, as shown in the formula (1) : Figure 4

[0105] Aligning each observation window in the electroencephalogram signal with each time sequence network unit of the long short-term memory network;

[0106] Inputting the one-dimensional feature corresponding to each observation window into the corresponding time sequence network unit; flattening the output of the plurality of time sequence network units to one dimension and inputting the full connection layer to extract the time sequence dynamic graph network feature.

[0107] The present application uses a long short-term memory network to learn the relationship of the dynamic graph network feature. Assuming that a brain electrical segment includes m observation windows, the corresponding long short-term memory network is formed by m time sequence network units (lstm cell) in series, wherein the time sequence information transmission direction is single-phase, the hidden layer of each time sequence network unit is 1 layer, and the output dimension is set as LSTM_zim. The feature of each observation window is aligned with the time sequence network unit (LSTM cell), and the output EdgeConv_zim of the dynamic edge convolution network is used as the input of the corresponding time sequence embedding module. Then, the output of the plurality of time sequence network units is flattened to one dimension to obtain a feature with a length of m*LSTM_zim. The obtained feature is connected to the target output of the full connection layer with EdgeLstm_zim to extract the time sequence dynamic graph network feature.

[0108] During the training of the time sequence dynamic graph feature extraction model, each electroencephalogram feature in the first training sample set is grouped according to each electroencephalogram parameter combination to obtain a training set corresponding to each electroencephalogram parameter combination; a classifier composed of 2 layers of multilayer perceptron is connected to the time sequence dynamic graph network feature extraction model, and each training set is used to optimize the model parameters using cross-entropy as the loss function. The validation set is used to evaluate its discrimination effect, and when the sensitivity and specificity of the validation set are not less than a certain threshold, the pre-trained model is saved, otherwise, the parameters of the model are adjusted within a preset number of times until the validation set reaches the standard, and if the preset limit is exceeded, the model is ignored for subsequent evaluation model establishment. The model parameters trained by each training set are saved to obtain the dynamic graph network feature extraction model related to each electroencephalogram parameter combination.

[0109] Step S3: constructing a second training sample set using the first training sample set, the time sequence dynamic graph network feature related to each electroencephalogram parameter combination in each sample, and the demographic coding feature corresponding to the subject;

[0110] Specifically, the second training sample set is constructed by the following method:

[0111] ​Encode the demographic information corresponding to each subject to obtain the demographic encoding features of each subject;

[0112] Delete redundant data from the time-series dynamic graph network features related to each combination of extracted electroencephalogram parameters at different levels to obtain an alternative time-series dynamic graph network feature set indexed by subject ID;

[0113] Combine the first training sample set, the demographic encoding features of each subject, and the time-series dynamic graph network features to obtain a second training sample set.

[0114] Further, for each level of sub-class parameter combination, when the features encoded by different time-series dynamic graph networks have a high correlation, the present application introduces a method for reducing model redundancy, deletes redundant data at different levels, and obtains an alternative time-series dynamic graph network feature set for better weight allocation by subsequent gated networks.

[0115] Specifically, the alternative time-series dynamic graph network feature set is obtained by the following method:

[0116] Divide the data in the first training sample set into K folds, and iteratively use the training data in the K folds to obtain time-series dynamic graph network features related to all combinations of electroencephalogram parameters;

[0117] Fix the sample size, calculate the correlation between each parameter combination based on the time-series dynamic graph network features, and calculate the mean of the correlation in the sample dimension. Mark the parameter combinations with an absolute correlation greater than a first preset threshold. In this embodiment, the first preset threshold is set to greater than 0.65.

[0118] After K iterations, remove the parameter combinations with a marked number greater than a second preset threshold to obtain an alternative time-series dynamic graph network feature set. In this embodiment, the second preset threshold is set to 0.8*K.

[0119] Further, the demographic encoding features are obtained by the following method:

[0120] Read the gender, age, and open or closed eye information of the subject when collecting the electroencephalogram signals, and preprocess them by the following method:

[0121] For age information, use standardization processing to unify the numerical range between different features. For gender information, use one-hot encoding to encode males as (0, 1) and females as (1, 0). For open or closed eye information, use one-hot encoding to encode open eyes as (0, 1) and closed eyes as (1, 0).

[0122] For each dimension of the pre-processed information, a corresponding information encoding module composed of multiple fully connected layers is input, and an activation function is used for non-linear transformation; each information obtained after non-linear transformation is encoded and merged, and mapped to a unified representation space to obtain demographic encoding features.

[0123] Step S4: training the hierarchical multi-level gating model using the second training sample set; the hierarchical multi-level gating model is used to screen a task-related electroencephalogram parameter combination, and a mood disorder evaluation result is predicted using a time sequence dynamic graph network feature related to the screened electroencephalogram parameter combination;

[0124] Specifically, the hierarchical multi-level gating model is used to screen a task-related electroencephalogram parameter combination by setting multiple gating layers in sequence based on the electroencephalogram features and demographic encoding features; and a mood disorder evaluation result is predicted based on a time sequence dynamic graph network feature corresponding to a screened electroencephalogram parameter combination.

[0125] The hierarchical multi-level gating model includes a frequency band gating layer, a brain region combination gating layer, an observation time unit gating layer, and a fusion evaluation layer.

[0126] The frequency band gating layer is used to screen a preset number of task-related frequency bands based on electroencephalogram features corresponding to different frequency bands; and a frequency band-related evaluation result is obtained by classifying and predicting a time sequence dynamic graph network feature related to each sub-class parameter combination corresponding to a task-related frequency band.

[0127] The brain region combination gating layer screens a task-related brain region combination using the screened task-related frequency band as a constraint, and classifies and predicts a time sequence dynamic graph network feature related to a sub-class parameter combination corresponding to a combination of a task-related brain region combination and a frequency band to obtain a brain region-related evaluation result.

[0128] The observation time unit gating layer screens a task-related observation time unit using the screened task-related frequency band and brain region combination as a constraint, and classifies and predicts a time sequence dynamic graph network feature related to a parameter combination corresponding to a task-related frequency band, brain region combination, and observation time unit to obtain an observation time-related evaluation result.

[0129] The fusion evaluation layer is used to weight and fuse the frequency band-related evaluation result, the brain region-related evaluation result, and the observation time-related evaluation result to obtain a final mood disorder evaluation result.

[0130] The embodiment is based on the division of each parameter angle into multiple gating layers, the first level is a frequency band gating layer, the second level is a brain region gating layer, and the third level is an observation time gating layer. For the first level gating layer, it is further divided into sub-level layers of multiple frequency band features, from the delta sub-level gating layer to the beta2 sub-level gating layer; for the second level brain region gating layer, any two brain networks are combined to form a brain network combination, from the frontal region-parietal region to the left temporal region-occipital region sub-level gating layer; for the third level observation time gating layer, no sub-gating layer division is required.

[0131] In particular, the frequency band gating layer includes an input layer, a frequency band sub-level dense gating layer, a frequency band sparse gating layer, and a frequency band classification prediction layer.

[0132] The input layer is used to splice the electroencephalogram features corresponding to each frequency band of each sample with the demographic coding features, and input them into the frequency band sub-level dense gating layer and the frequency band sparse gating layer, respectively.

[0133] The frequency band sub-level dense gating layer and the frequency band sparse gating layer are both constructed based on a multi-layer perception mechanism; the output quantity of the frequency band sub-level dense gating layer is the same as the number of sub-class parameter combinations corresponding to each frequency band, which is used to extract features of the input data and obtain weights corresponding to each sub-class parameter combination through an activation function; based on the weights, time sequence dynamic graph network features related to each sub-class parameter combination corresponding to each frequency band are weighted and fused to obtain fusion features corresponding to each frequency band.

[0134] The output quantity of the frequency band sparse gating layer is the same as the number of frequency bands, which is used to extract features of the input data and obtain a task relevance index of each frequency band through an activation function; the task relevance index is arranged in descending order, and a preset number of task-related frequency bands are selected.

[0135] The frequency band classification prediction layer is used to receive the fusion features corresponding to the task-related frequency bands, and predict a frequency band-related classification result.

[0136] Specifically, for the frequency band sub-level gating layer, since the preselected frequency bands are fixed, the electroencephalogram features associated in the frequency band dimension are spliced with the demographic coding features, and then connected to the multi-layer perceptron with the same number of output as the number of target super parameter combinations. After the output features pass through the softmax layer, the weight of the features corresponding to each parameter combination is obtained. Based on the obtained weight, the corresponding time sequence dynamic graph network features are weighted and fused to obtain the fusion features corresponding to each frequency band, represented as EdgeLstm_delta, EdgeLstm_theta, EdgeLstm_alpha, EdgeLstm_beta1, and EdgeLstm_beta2. In order to ensure the diversity of features in the early stage of coding and enhance stability, the dense gating mode is used to select and fuse the features corresponding to the sub-level parameters.

[0137] For the frequency band gating layer, the multi-layer perceptron takes the number of frequency bands as the output dimension, and selects the frequency bands ranked top2 after the output features pass through the softmax layer. In this embodiment, the target frequency bands are selected by using the sparse gating mode, aiming to select the frequency bands most relevant to the task purpose.

[0138] The brain region combination gating layer includes an input layer, a brain region sub-level dense gating layer, a brain region sparse gating layer, and a brain region classification prediction layer.

[0139] The input layer is used to flatten the channel dimension and feature dimension of the task-related frequency bands corresponding to each brain region combination, and splice them with the demographic coding features, and input them into the brain region sub-level dense gating layer and the brain region sparse gating layer, respectively.

[0140] The brain region sub-level dense gating layer and the brain region sparse gating layer are both trained based on the multi-layer perceptron.

[0141] The output number of the brain region sub-level dense gating layer is the number of the candidate parameter combinations composed of the task-related frequency bands corresponding to each brain region combination and each observation time unit. The brain region sub-level dense gating layer is used to extract features from the received data, and obtain the weight of each candidate parameter combination corresponding to each brain region combination through an activation function. Based on the weight, the time sequence dynamic graph network features of each candidate parameter combination corresponding to each brain region combination are weighted and fused to obtain the fusion features corresponding to each brain region combination.

[0142] The output number of the brain region sparse gating layer is the number of brain region combinations, which is used to extract features from the received data, and obtain the task relevance index of each brain region combination through an activation function. The task relevance index is arranged in descending order to obtain a preset number of task-related brain region combinations.

[0143] The brain region classification prediction layer is used to receive the fusion features corresponding to the task-related brain region combinations to obtain the brain region related classification result.

[0144] Specifically, for the brain region combination sub-level gating layer, the selected frequency band in the frequency band gating layer is used to constrain the alternative model of the layer gating module. Since the pre-selected brain region selection is fixed, the electrode corresponding to the selected brain region needs to be selected in the channel dimension of each frequency band, the channel part and the feature part of the electroencephalogram feature are flattened, and the demographic coding feature is spliced, input into the multi-layer perception connected with the output of the number of alternative models of the layer, and then the output is input into the softmax layer to obtain the weight of each alternative model. According to the weight, the corresponding time series dynamic graph network feature is weighted and fused to obtain the fusion feature corresponding to each brain region combination. In order to enhance the stability of the model and the diversity of the feature, the embodiment adopts the dense gating mode to fuse the features corresponding to the sub-level parameters.

[0145] For the brain region combination gating layer, the multi-layer perception takes the number of brain regions as the output dimension, and the corresponding frequency band of the top 2 selected features after the softmax layer is sorted; the embodiment adopts the sparse gating mode to select the target brain region, aiming to selectively select the time series dynamic graph network feature extraction model related to the task target.

[0146] The observation time unit gating layer includes an input layer, an observation time sparse gating layer, and an observation time unit classification prediction layer.

[0147] The input layer is used to obtain the alternative electroencephalogram features related to the length of each observation time unit based on the task-related brain region combination and the task-related frequency band, and to splice the flattened alternative electroencephalogram features related to the length of each observation time unit with the demographic coding features respectively, and input the observation time gating layer.

[0148] The output node number of the observation time sparse gating layer is the number of observation time unit length categories, which is used to extract features from the received data and obtain the task-related degree index of each observation time unit length through an activation function, and select the observation time unit length with the maximum task-related degree index as the task-related observation time unit.

[0149] The observation time unit classification prediction layer is used to receive the time series dynamic graph network features corresponding to the task-related observation time unit and the alternative sub-class parameter combination, and obtain the observation time unit related classification result.

[0150] For the observation time unit gating layer, the parameters of the layer are constrained by the task-related frequency band and brain region combination, and the sub-level gating layer is not involved. After flattening the alternative electroencephalogram features related to the length of each observation time unit, they are spliced with the demographic coding features respectively, and then connected to a multi-layer perception machine with the number of output nodes equal to the number of observation time unit parameters. The selected output features are sorted by a softmax layer to obtain the top 1 corresponding observation time unit length. Here, the sparse gating method is used to select the observation time unit length, aiming to selectively select the observation time unit related to the task.

[0151] Further, the hierarchical multi-level gating module is trained by the following method:

[0152] The electroencephalogram feature data in the second training sample set is grouped by frequency band and merged with the corresponding demographic coding features, which are input into the frequency band gating layer to predict the task-related frequency band.

[0153] The electroencephalogram feature data corresponding to the predicted task-related frequency band is grouped by electroencephalogram parameter combination and merged with the corresponding demographic coding features, which are input into the brain region combination gating layer to predict the task-related electroencephalogram combination.

[0154] The electroencephalogram features obtained by taking the predicted task-related frequency band and brain region combination as constraints are grouped by observation time unit length and spliced with the corresponding demographic coding features of the subjects, which are input into the observation time unit gating layer to screen the task-related observation time unit length.

[0155] Based on the screening results of each gating layer, a plurality of task-related electroencephalogram parameter combinations corresponding to each gating layer are obtained. The time series dynamic graph network features obtained based on each task-related electroencephalogram parameter combination as an index are used for mood disorder assessment prediction, and the cross-entropy loss of the prediction results of each gating layer and the labeled label is calculated.

[0156] The cross-entropy losses of each gating layer are weighted and fused to obtain the total loss of the hierarchical multi-level gating module. The total loss is used for iterative optimization to obtain a converged multi-level gating module.

[0157] The total loss is represented as:

[0158] Loss=W Freq *Freq CE +W brain *Brain CE +W Time *Time CE ;

[0159]

[0160] The weights of each gating layer in the weighted fusion satisfy the following constraint condition:

[0161] W Time >W brain >W Freq ;

[0162] Wherein, Loss is the total loss of the model, W Freq is the frequency band gating layer weight, W brain is the brain region gating layer weight, W Time is the time observation unit gating layer weight, Freq CE is the cross-entropy loss of the frequency band gating layer, Brain CE is the cross-entropy loss of the brain region gating layer, Time CE is the cross-entropy loss of the time observation unit gating layer, y is the labeled label, is the predicted value of the frequency band gating layer, is the predicted value of the brain region gating layer, is the predicted value of the observation time unit gating layer.

[0163] In the training process, the embodiment respectively connects the dynamic graph network features corresponding to the frequency band gating layer, the brain network combination gating layer and the observation time unit gating layer to the multi-layer perception machine output classification result, and calculates the cross-entropy of the result output by each level. Since the model design is to design the gating mechanism and realize the feature selection of the dynamic graph network from the frequency band layer to the brain network combination layer to the observation time unit layer, the principle is to realize better mood disorder evaluation through the feature angle of layer-by-layer progression and the constantly refined feature background, so when setting the cross-entropy weight, the weight of each level should be kept W Time >W brain >W Freq .

[0164] Step S5: Based on the time sequence dynamic graph network feature extraction model corresponding to each combination of the trained brain electrical parameter, and the hierarchical multi-level gating model, a mood disorder evaluation model is obtained.

[0165] The mood disorder evaluation model obtained based on the model construction method of the embodiment can realize the evaluation of normal / depressive mood disorders, and unipolar / bipolar mood disorders. The stability and generalization of the model are high, and it can better assist the clinical psychiatrists in improving the mood disorder identification. Based on the mood disorder evaluation model constructed by the embodiment, a mood disorder evaluation system can be composed of a data acquisition unit, a brain electrical coding unit, etc., as shown in Figure 5 , to realize a portable and stable end-to-end mood disorder evaluation.

[0166] To sum up, the embodiment of the present application provides a construction method of a mood disorder evaluation model based on a hierarchical multi-level gating, which fuses electroencephalogram features by using patient information, assists in screening of a corresponding brain network model based on a hierarchical multi-level gating model, and improves the effect of cross-subject mood disorder evaluation; and the present application constructs multiple parameter combinations based on frequency bands, brain region combinations, and observation time lengths, and respectively establishes mood disorder evaluation models of different angles of time sequence brain networks, can recognize the relevance between the linkage features between brain regions and mood disorders based on the underlying features of electroencephalogram signals, accurately extracts features associated with mood disorders by combining the relevance between frequency bands and observation time, is used for mood disorder evaluation, has more comprehensive feature angles, greatly improves the accuracy of mood disorder evaluation, and has higher feature interpretation and clinical value.

[0167] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing relevant hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium is a disk, an optical disk, a read-only memory, a random access memory, etc.

[0168] The above is only a preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for constructing a mood disorder assessment model based on hierarchical multi-level gating, characterized in that, The method comprises the following steps: obtaining brain electrical signals of multiple subjects, extracting corresponding brain electrical features, and labeling mood disorder related labels to construct a first training sample set; a time series dynamic graph network feature extraction model is initially constructed, and the first training sample set is used for training to obtain a plurality of brain electrical parameter combinations respectively corresponding to a converged time series dynamic graph network feature extraction model and extracted time series dynamic graph network features related to each brain electrical parameter combination; a second training sample set is constructed using the first training sample set, the time series dynamic graph network features related to each brain electrical parameter combination in each sample, and the demographic coding features corresponding to the subjects; a hierarchical multi-level gating model is trained using the second training sample set; the hierarchical multi-level gating model is used to screen task-related brain electrical parameter combinations, and the time series dynamic graph network features related to the screened brain electrical parameter combinations are used to predict mood disorder evaluation results; the hierarchical multi-level gating model comprises a frequency band gating layer, a brain region combination gating layer, and an observation time unit gating layer, which are respectively used to screen task-related frequency bands, brain region combinations, and observation time units; the hierarchical multi-level gating model is trained by the following method: group the brain electrical feature data in the second training sample set by frequency band, and merge with the corresponding demographic coding features, input the frequency band gating layer, and predict the task-related frequency band; group the brain electrical feature data corresponding to the predicted task-related frequency band by brain electrical parameter combination, and merge with the corresponding demographic coding features, input the brain region combination gating layer, and predict the task-related brain electrical combination; group the brain electrical features obtained by taking the predicted task-related frequency band and brain region combination as constraints by observation time unit length, and respectively splice with the corresponding demographic coding features of the subjects, input the observation time unit gating layer, and screen the task-related observation time unit length; based on the screening results of each gating layer, obtain a plurality of task-related brain electrical parameter combinations corresponding to each gating layer, perform mood disorder evaluation prediction based on the time series dynamic graph network features obtained based on each task-related brain electrical parameter combination as an index, and calculate the cross-entropy loss of the prediction results of each gating layer and the labeled label respectively; weight and fuse the cross-entropy losses of each gating layer to obtain the total loss of the hierarchical multi-level gating model, and iteratively optimize the total loss to obtain a converged multi-layer gating model; based on the trained time series dynamic graph network feature extraction model corresponding to each brain electrical parameter combination and the hierarchical multi-level gating model, a mood disorder evaluation model is obtained. 2.The method of constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 1, wherein, The time series dynamic graph network feature extraction model is trained by the following method: group the brain electrical features in the first training sample set by different brain electrical parameter combinations to obtain multiple groups of brain electrical features; connect a classification module to the initially constructed time series dynamic graph network feature extraction model, and the classification module is used to output a mood disorder prediction probability distribution; The time series dynamic graph network feature extraction model preliminarily constructed is trained by using each set of electroencephalogram features; iterative calculation is performed through a loss function to minimize the loss between the prediction result and the labeled label, and the model parameters optimized by using each set of electroencephalogram features are saved to obtain a time series dynamic graph network feature extraction model related to each combination of electroencephalogram parameters. 3.The method of constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 1, wherein, The second training sample set is obtained by the following method: The demographic information corresponding to each subject is encoded to obtain demographic encoding features of each subject; The extracted time series dynamic graph network features related to each combination of electroencephalogram parameters are hierarchically deleted to obtain a candidate time series dynamic graph network feature set indexed by subject ID; The first training sample set, the demographic encoding features of each subject, and the time series dynamic graph network features are combined to obtain the second training sample set.

4. The method for constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 3, wherein, The candidate time series dynamic graph network feature set is obtained by the following method: The data in the first training sample set is divided into K folds, and the training data in the K folds is iteratively used to obtain time series dynamic graph network features related to all combinations of electroencephalogram parameters; The correlation between each parameter combination is calculated based on the time series dynamic graph network features while the number of samples is fixed; The mean of the correlation is calculated in the sample dimension, and the parameter combinations with an absolute value of correlation greater than a first preset threshold are marked; After K iterations, the parameter combinations with a number of marks greater than a second preset threshold are removed to obtain the candidate time series dynamic graph network feature set.

5. The method for constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 3, wherein, The demographic encoding features are obtained by the following method: The gender, age, and open or closed eye information of the subject when collecting the electroencephalogram signal are read and preprocessed by the following method: For age information, standardization processing is used to unify the numerical range between different features; For gender information, a one-hot encoding method is used to encode males as (0, 1) and females as (1, 0); For open or closed eye information, a one-hot encoding method is used to encode open eyes as (0, 1) and closed eyes as (1, 0); For each dimension of preprocessed information, the corresponding information encoding module composed of multiple fully connected layers is input, and an activation function is used for non-linear transformation; Each information obtained after non-linear transformation is encoded and merged, and mapped to a unified representation space to obtain demographic encoding features.

6. The method for constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 1, wherein, The total loss is represented as: ； ； ； ； The weights of each gating layer during weighted fusion satisfy the following constraint condition: ； wherein, is the total loss for the model, is the band gating layer weight, is the brain region gating layer weight, is the time observation unit gating layer weight, is the cross-entropy loss for the band gating layer, is the cross-entropy loss for the brain region gating layer, is the cross-entropy loss for the time observation unit gating layer, is the annotated label, is the predicted value for the band gating layer, is the predicted value for the brain region gating layer, is the predicted value for the observation time unit gating layer.

7. The method for constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 1, wherein, The parameters in the electroencephalogram parameter combination include frequency bands, brain region combinations, and observation time unit lengths; The frequency bands include delta, theta, alpha, beta1, and beta2 bands; The brain regions include frontal region, parietal region, central region, left temporal region, right temporal region, and occipital region; The electroencephalogram parameter combination is obtained by the following method: Different brain regions are combined in pairs to obtain multiple brain region combinations; Multiple observation time units of different lengths are set; Different frequency bands, brain region combinations, and observation time units are combined in a preset level order to obtain multiple electroencephalogram parameter combinations and corresponding sub-class parameter combinations of each parameter. 8.The method of constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 1, wherein, The frequency band gating layer comprises an input layer, a frequency band sub-level dense gating layer, a frequency band sparse gating layer, and a frequency band classification prediction layer; The input layer is configured to splice the electroencephalogram features corresponding to each frequency band and the demographic coding features, and input the features into the frequency band sub-level dense gating layer and the frequency band sparse gating layer respectively; The output quantity of the frequency band sub-level dense gating layer is the same as the quantity of the sub-class parameter combinations corresponding to each frequency band, and the frequency band sub-level dense gating layer is configured to extract features from the input data and obtain the weight of each sub-class parameter combination through an activation function; Based on the weight, the time series dynamic graph network features of each sub-class parameter combination corresponding to each frequency band are fused to obtain the fusion features of each frequency band; The output quantity of the frequency band sparse gating layer is the same as the quantity of the frequency bands, and the frequency band sparse gating layer is configured to extract features from the input data and obtain the task relevance index of each frequency band through an activation function, and the task relevance index is arranged in descending order to obtain a preset quantity of task-related frequency bands; The frequency band classification prediction layer is configured to receive the fusion features of the task-related frequency bands and predict the frequency band-related classification result. 9.The method of constructing a mood disorder assessment model based on hierarchical multi-level gating according to claim 1, wherein, The brain region combination gating layer comprises an input layer, a brain region sub-level dense gating layer, a brain region sparse gating layer, and a brain region classification prediction layer; The input layer is configured to flatten the channel dimension and the feature dimension of the task-related frequency bands corresponding to each brain region combination, splice the flattened dimension with the demographic coding features, and input the features into the brain region sub-level dense gating layer and the brain region sparse gating layer respectively; The output quantity of the brain region sub-level dense gating layer is the quantity of the candidate parameter combinations composed of the task-related frequency bands corresponding to each brain region combination and each observation time unit; the brain region sub-level dense gating layer is configured to extract features from the received data and obtain the weight of each candidate parameter combination corresponding to each brain region combination through an activation function; and based on the weight, the time series dynamic graph network features of each candidate parameter combination corresponding to each brain region combination are fused to obtain the fusion features of each brain region combination; The output quantity of the brain region sparse gating layer is the quantity of the brain region combinations, and the brain region sparse gating layer is configured to extract features from the received data and obtain the task relevance index of each brain region combination through an activation function; the task relevance index is arranged in descending order to obtain a preset quantity of task-related brain region combinations; The brain region classification prediction layer is configured to receive the fusion features of the task-related brain region combinations and obtain the brain region-related classification result.

Images