Training method, detection method and system of depression detection model

Abstract

The invention discloses a training method, a detection method and a system for a depression detection model, and the training method comprises the steps: obtaining electroencephalogram training data; performing frequency spectrum feature extraction on the electroencephalogram training data to obtain a frequency spectrum feature set; performing time sequence feature extraction on the electroencephalogram training data to obtain a time sequence feature set; performing spectrum proxy graph learning on the spectrum feature set to obtain a spectrum coding representation, and performing time sequence proxy graph learning on the time sequence feature set to obtain a time sequence coding representation; according to the frequency spectrum coding representation, performing double-domain feature fusion on the time sequence coding representation to obtain double-domain fusion features; and according to the double-domain fusion features, performing parameter updating on the initialized depression detection model to obtain a trained depression detection model. According to the training method, a depression detection model can be provided, and the objectivity, the detection efficiency and the detection accuracy of depression detection can be improved. The invention relates to the technical field of artificial intelligence.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a training method, a detection method and a system for a depression detection model. Background Art

[0002] With the continuous development of society, the disease detection of patients with depression has become one of the concerns of people.

[0003] Currently, traditional depression detection methods can generally be divided into two categories. The first category depends on doctor-patient conversations and questionnaire scores, and the diagnosis results depend on the patient's responses and behaviors, as well as the doctor's professional knowledge. This method is easily interfered by the subjective consciousness of doctors and patients, and the objectivity and detection efficiency of depression detection are not good. Also, the second category usually extracts key spatio-temporal features from the optimal channels in each frequency band of electroencephalogram (EEG) signals based on band-pass filtering and channel selection strategies, and realizes depression detection through a classification model. The detection accuracy of this method is not satisfactory.

[0004] Therefore, the problems existing in the prior art still need to be solved and optimized urgently. Summary of the Invention

[0005] The purpose of the present invention is to solve at least to some extent one of the technical problems existing in the related art.

[0006] To this end, an object of an embodiment of the present invention is to provide a training method, a detection method and a system for a depression detection model. Among them, the training method can provide a depression detection model, which is beneficial to improving the objectivity, detection efficiency and detection accuracy of depression detection.

[0007] In order to achieve the above technical purpose, the technical solutions adopted in the embodiments of the present application include:

[0008] In a first aspect, an embodiment of the present application provides a training method for a depression detection model, including:

[0009] Obtain electroencephalogram training data;

[0010] Extract spectral features from the electroencephalogram training data to obtain a spectral feature set, where the spectral feature set includes a plurality of intermediate spectral features, and the spectral frequency bands of each intermediate spectral feature are different;

[0011] Extract temporal features from the electroencephalogram training data to obtain a temporal feature set, where the temporal feature set includes a plurality of cascaded intermediate temporal features, and the temporal receptive field of the latter intermediate temporal feature is larger than that of the previous intermediate temporal feature;

[0012] Perform spectral proxy graph learning on the spectral feature set to obtain a spectral coding representation, and perform temporal proxy graph learning on the temporal feature set to obtain a temporal coding representation;

[0013] According to the spectral coding representation, perform dual-domain feature fusion on the temporal coding representation to obtain a dual-domain fusion feature;

[0014] According to the dual-domain fusion feature, update the parameters of the initialized depression detection model to obtain a trained depression detection model.

[0015] In addition, according to the method of the above embodiments of the present application, the following additional technical features may also be included:

[0016] Further, in an embodiment of the present application, the extracting spectral features from the EEG training data to obtain a spectral feature set includes:

[0017] Perform frequency-domain conversion on the EEG training data to obtain a plurality of EEG frequency band data, and each of the EEG frequency band data has a different frequency band;

[0018] Perform differential entropy feature extraction on all the EEG frequency band data to obtain intermediate spectral features corresponding to each of the EEG frequency band data;

[0019] According to all the intermediate spectral features, obtain the spectral feature set;

[0020] The extracting temporal features from the EEG training data to obtain a temporal feature set includes:

[0021] Input the EEG training data into a plurality of cascaded temporal convolutional structures for feature cascading extraction to obtain intermediate temporal features output by each of the temporal convolutional structures;

[0022] According to all the intermediate temporal features, obtain the temporal feature set.

[0023] Further, in an embodiment of the present application, the performing target proxy graph learning on the target feature set to obtain a first target coding representation includes:

[0024] Perform target proxy graph generation on the target feature set to obtain a target functional proxy graph and a target spatial proxy graph corresponding to the target feature set;

[0025] According to the target functional proxy graph, perform target graph learning on the target spatial proxy graph to obtain the first target coding representation;

[0026] Wherein, the target feature set is a spectral feature set or a temporal feature set.

[0027] Further, in an embodiment of the present application, the generating of the target proxy graph for the target feature set to obtain the target functional proxy graph and the target spatial proxy graph corresponding to the target feature set includes:

[0028] Obtain the preset functional graph construction rule and spatial graph construction rule, as well as the original node data of the target feature set in each brain region;

[0029] Perform adaptive generation processing of proxy nodes on all the original node data to obtain a number of proxy node data, and each proxy node data corresponds to one of the brain regions;

[0030] According to the functional graph construction rule, construct a functional topology graph for the original node data and the proxy node data to obtain the target functional proxy graph;

[0031] According to the spatial graph construction rule, construct a spatial topology graph for the original node data and the proxy node data to obtain the target spatial proxy graph.

[0032] Further, in an embodiment of the present application, the performing of target graph learning on the target spatial proxy graph according to the target functional proxy graph to obtain the first target encoding representation includes:

[0033] Perform first multi-level cascaded graph convolution processing on the target functional proxy graph to obtain a target functional feature representation;

[0034] Perform second multi-level cascaded graph convolution processing on the target spatial proxy graph to obtain a target spatial feature representation;

[0035] Obtain the first target encoding representation according to the target functional feature representation and the target spatial feature representation.

[0036] Further, in an embodiment of the present application, the performing of dual-domain feature fusion on the time series encoding representation according to the spectral encoding representation to obtain a dual-domain fusion feature includes:

[0037] Perform intra-domain feature fusion in the frequency domain on the spectral encoding representation to obtain an intra-domain fusion feature in the frequency domain, and perform intra-domain feature fusion in the time domain on the time series encoding representation to obtain an intra-domain fusion feature in the time domain;

[0038] Perform inter-domain feature fusion between the time and frequency domains on the time series encoding representation according to the spectral encoding representation to obtain an inter-domain fusion feature between the time and frequency domains;

[0039] Obtain the dual-domain fusion feature according to the intra-domain fusion feature in the frequency domain, the intra-domain fusion feature in the time domain, and the inter-domain fusion feature between the time and frequency domains.

[0040] Further, in an embodiment of the present application, performing intra-target-domain feature fusion on the second target encoded representation to obtain an intra-target-domain fused feature, including:

[0041] Obtaining a target function feature representation and a target spatial feature representation in the second target encoded representation, where the second target encoded representation is the spectral encoded representation or the temporal encoded representation;

[0042] Performing a first attention analysis on the target function feature representation to obtain a functional attention weight of the target function feature representation, and performing a second attention analysis on the target spatial feature representation to obtain a spatial attention weight of the target spatial feature representation;

[0043] According to the functional attention weight and the spatial attention weight, performing a first weighted fusion on the target function feature representation and the target spatial feature representation to obtain the intra-target-domain fused feature.

[0044] Further, in an embodiment of the present application, the performing inter-time-frequency-domain feature fusion on the temporal encoded representation according to the spectral encoded representation to obtain an inter-time-frequency-domain fused feature, including:

[0045] Performing a third attention analysis on the spectral encoded representation to obtain a frequency-domain attention weight;

[0046] Performing a fourth attention analysis on the temporal encoded representation to obtain a time-domain attention weight;

[0047] According to the frequency-domain attention weight and the time-domain attention weight, performing a second weighted fusion on the spectral encoded representation and the temporal encoded representation to obtain the inter-time-frequency-domain fused feature.

[0048] Second, an embodiment of the present application provides a detection method for a depression detection model, including:

[0049] Obtaining electroencephalogram data to be detected;

[0050] Inputting the electroencephalogram data to be detected into the above-mentioned trained depression detection model to obtain a depression detection result output by the trained depression detection model.

[0051] Third, an embodiment of the present application provides a training system for a depression detection model, including:

[0052] A first processing unit, configured to obtain electroencephalogram training data;

[0053] A second processing unit, configured to extract spectral features from the electroencephalogram training data to obtain a spectral feature set, where the spectral feature set includes a plurality of intermediate spectral features, and the spectral frequency bands of each intermediate spectral feature are different;

[0054] A third processing unit, configured to extract temporal features from the EEG training data to obtain a temporal feature set, where the temporal feature set includes a plurality of cascaded intermediate temporal features, and the temporal receptive field of the latter intermediate temporal feature is larger than that of the previous intermediate temporal feature;

[0055] A fourth processing unit, configured to perform spectral proxy map learning on the spectral feature set to obtain a spectral coding representation, and perform temporal proxy map learning on the temporal feature set to obtain a temporal coding representation;

[0056] A fifth processing unit, configured to perform dual-domain feature fusion on the temporal coding representation according to the spectral coding representation to obtain a dual-domain fusion feature;

[0057] A sixth processing unit, configured to update parameters of an initialized depression detection model according to the dual-domain fusion feature to obtain a trained depression detection model.

[0058] In a fourth aspect, an embodiment of the present application further provides an electronic device, including:

[0059] At least one processor;

[0060] At least one memory, configured to store at least one program;

[0061] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0062] In a fifth aspect, an embodiment of the present application further provides a computer-readable storage medium, in which a program executable by a processor is stored, and the program executable by the processor is used to implement the above method when executed by the processor.

[0063] The advantages and beneficial effects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present application:

[0064] A training method, detection method, and system for a depression detection model disclosed in an embodiment of the present application. In the training method, electroencephalogram (EEG) training data is obtained; spectral feature extraction is performed on the EEG training data to obtain a spectral feature set, where the spectral feature set includes a number of intermediate spectral features, and the spectral frequency bands of each intermediate spectral feature are different; temporal feature extraction is performed on the EEG training data to obtain a temporal feature set, where the temporal feature set includes a number of cascaded intermediate temporal features, and the temporal receptive field of the latter intermediate temporal feature is larger than that of the previous intermediate temporal feature; spectral proxy map learning is performed on the spectral feature set to obtain a spectral encoded representation, and temporal proxy map learning is performed on the temporal feature set to obtain a temporal encoded representation; based on the spectral encoded representation, dual-domain feature fusion is performed on the temporal encoded representation to obtain a dual-domain fusion feature; and based on the dual-domain fusion feature, parameter update is performed on an initialized depression detection model to obtain a trained depression detection model. By performing temporal feature extraction on the EEG training data, the training method obtains a temporal feature set in which the temporal receptive field of the latter intermediate temporal feature is larger than that of the previous intermediate temporal feature, which can effectively improve the comprehensiveness of feature information from the temporal perspective; and by performing spectral feature extraction on the EEG training data, the training method obtains a spectral feature set composed of intermediate spectral features with different spectral frequency bands, which can effectively improve the comprehensiveness of feature information from the spectral perspective; based on the temporal feature set and the spectral feature set, it is beneficial to improve the objectivity and detection accuracy of the depression detection model for depression detection, as well as improve the detection efficiency of depression detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following introduces the relevant technical solution drawings in the embodiments of the present application or the prior art. It should be understood that the drawings below are only for conveniently and clearly expressing some embodiments of the technical solutions in the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative efforts.

[0066] Figure 1 FIG. is a schematic flowchart of a training method for a depression detection model provided by an embodiment of the present application;

[0067] Figure 2 FIG. is a schematic diagram of a simple structure of a temporal feature extractor provided by an embodiment of the present application;

[0068] Figure 3 FIG. is a schematic diagram of a structural framework of a training system for a depression detection model provided by an embodiment of the present application;

[0069] Figure 4A schematic structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0070] The embodiments of the present application will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and should not be construed as a limitation to the present application. For the step numbers in the following embodiments, they are only set for the convenience of description and explanation, and no limitation is imposed on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0072] Currently, traditional depression detection methods can generally be divided into two categories. The first category depends on doctor-patient conversations and questionnaire scoring, and the diagnosis results depend on the patient's responses and behaviors, as well as the doctor's professional knowledge. This method is easily interfered by the subjective consciousness of doctors and patients, and the objectivity and detection efficiency of depression detection are not good. Also, the second category usually extracts key spatio-temporal features from the optimal channels in each frequency band of electroencephalogram (EEG) signals based on band-pass filtering and channel selection strategies, and realizes depression detection through a classification model. However, the spatio-temporal features extracted by this method are often not comprehensive, making the detection accuracy of the subsequent classification model for detecting depression based on spatio-temporal features unsatisfactory.

[0073] In addition, since depression is a serious mental illness, its abnormal manifestation in the brain presents as an unbalanced connection between brain regions, rather than a simple increase or decrease in the activity of a specific region. There is a part of the prior art that fully simulates and aligns the distribution of EEG electrodes with the neurophysiological structure, encodes all node pairs, and realizes depression detection by learning a dense fully connected adjacency matrix. However, the method of simulation alignment ignores the important cooperative relationships between different brain regions, greatly limiting the performance of depression detection and making it difficult to accurately and efficiently achieve depression detection. Also, since the cooperative connections between specific electrodes learned by the dense fully connected adjacency matrix are general patterns for reasoning, and the cooperative connections between some specific electrodes are redundant and noisy, they inevitably hinder information dissemination.

[0074] In addition, in the brain regions, the axons and dendrites of interneurons are confined to a single brain region, mainly realizing local information exchange; while principal neurons usually extend their axons beyond the brain regions where their cell bodies and dendrites are located, mainly realizing long-distance information exchange. The prior art ignores the targeted modeling of principal neurons, making it difficult to capture the global information interaction between brain regions, resulting in the instability of the overall representation and affecting the detection accuracy of depression detection.

[0075] Furthermore, there is also prior art that uses a two-branch learning method or a progressive graph convolutional network to implement depression detection. The interaction between its global features and local features only occurs after the feature learning process, and it cannot pay attention to the dynamic influence of the two; in addition, the long-distance connections constructed artificially cannot adaptively adjust the connection strength based on signal correlation, making the universality of the depression detection model in different individuals and different scenarios poor.

[0076] Moreover, the prior art ignores the redundancy of time-frequency features in intra-domain and cross-domain features, which easily leads to insufficient extraction of the underlying information in EEG data and low comprehensiveness of the generated feature information, and further results in low detection accuracy of the depression detection model for depression detection.

[0077] In view of this, embodiments of the present invention provide a training method, a detection method and a system for a depression detection model. Among them, the training method extracts temporal features from EEG training data to obtain a temporal feature set in which the temporal receptive field of the latter intermediate temporal feature is larger than that of the previous intermediate temporal feature, which can effectively improve the comprehensiveness of feature information from the temporal perspective; and, the training method extracts spectral features from EEG training data to obtain a spectral feature set composed of intermediate spectral features in different spectral bands, which can effectively improve the comprehensive information of feature information from the spectral perspective; based on the temporal feature set and the spectral feature set, it is beneficial to improve the objectivity and detection accuracy of the depression detection model for depression detection, as well as improve the detection efficiency of depression detection.

[0078] In addition, the training method performs proxy graph learning on the spectral feature set and the temporal feature set respectively, specifically generating corresponding proxy node data for each brain region through an adaptive strategy, and realizing the dynamic interaction of local and global information of brain electrodes by means of the proxy node data, which can enable the depression detection model to fully consider the collaborative connections between different brain regions and have the ability to capture long-distance connections between brain regions; at the same time, a spatial proxy graph is constructed based on the spatial graph construction rules, which can enable the constructed spatial proxy graph to adaptively adjust the connection strength, which is beneficial to improving the stability of the overall representation and the universality of the depression detection model in different individuals and different scenarios.

[0079] In addition, the training method performs graph learning on the proxy graph, specifically by performing multi-level cascaded convolutions on the functional proxy graph and the spatial proxy graph respectively. It can perform interactive learning on the proxy graph across different receptive fields, which is beneficial to fully explore and integrate the key information contained in the EEG data at different scale levels, and is beneficial to further improve the detection accuracy of the depression detection model in detecting depression. Moreover, based on the in-domain feature fusion and inter-domain feature fusion of the encoded representation, the training method can effectively eliminate the redundant parts existing in the EEG data, is beneficial to generating comprehensive and accurate dual-domain fusion features, and is beneficial to further improving the detection accuracy of the depression detection model.

[0080] Referring to Figure 1 , in the embodiment of the present application, a training method for a depression detection model includes:

[0081] Step 110, obtaining EEG training data;

[0082] In the embodiment of the present application, the EEG training data can be the EEG signals collected by the electrodes on the scalp of the subject, or the preprocessed EEG signals. The specific preprocessing steps may include at least one of data filtering, data segmentation, baseline correction, artifact and motion artifact removal, etc. The present application does not limit this.

[0083] Step 120, performing spectral feature extraction on the EEG training data to obtain a spectral feature set, where the spectral feature set includes a number of intermediate spectral features, and the spectral frequency bands of each intermediate spectral feature are different;

[0084] In the embodiment of the present application, step 120 may be to construct a spectral feature extractor, and by inputting the EEG training data into the spectral feature extractor for feature extraction, a spectral feature set corresponding to the EEG training data is obtained. The spectral feature set includes a number of intermediate spectral features in different frequency bands, and each intermediate spectral feature in each frequency band is used to provide the feature information of the corresponding frequency band from the spectral perspective.

[0085] In some embodiments, step 120, performing spectral feature extraction on the EEG training data to obtain a spectral feature set, includes:

[0086] A1. Performing frequency domain conversion on the EEG training data to obtain a number of EEG frequency band data, and the frequency bands of each EEG frequency band data are different;

[0087] A2. Performing differential entropy feature extraction on all the EEG frequency band data to obtain intermediate spectral features corresponding to each EEG frequency band data;

[0088] A3. Obtaining the spectral feature set according to all the intermediate spectral features;

[0089] In the embodiments of the present application, frequency-domain analysis can be performed on electroencephalogram (EEG) training data based on the Fast Fourier Transform (FFT), so as to convert the EEG training data into the frequency domain and obtain EEG frequency-band data divided into several frequency bands. In the embodiments of the present application, taking the example that there are 5 EEG frequency-band data in total, the specific frequency bands of the EEG frequency-band data can be the 1-3 Hz frequency band, the 4-7 Hz frequency band, the 8-13 Hz frequency band, the 14-30 Hz frequency band, and the 31-50 Hz frequency band respectively.

[0090] It can be understood that for a certain EEG frequency-band data, step A2 can be to obtain the differential entropy representation (i.e., the intermediate frequency spectrum feature) of the EEG frequency-band data, and this intermediate frequency spectrum feature is used to characterize the information amount and complexity of the EEG frequency-band data corresponding to the corresponding frequency band. There are already various extraction methods for the specific differential entropy representation, and the same applies to the remaining EEG frequency-band data, which will not be elaborated herein. Additionally, after obtaining the intermediate frequency spectrum features corresponding to each frequency band, all the obtained intermediate frequency spectrum features can be concatenated, and the concatenated frequency spectrum features can be determined as the frequency spectrum feature set, which is used to provide comprehensive feature information from the perspective of the frequency spectrum.

[0091] Step 130: Extract time-series features from the EEG training data to obtain a time-series feature set, where the time-series feature set includes several cascaded intermediate time-series features, and the time-series receptive field of the latter intermediate time-series feature is larger than that of the previous intermediate time-series feature;

[0092] In the embodiments of the present application, step 120 can be to construct a time-series feature extractor. By inputting the EEG training data into the time-series feature extractor for feature extraction, a time-series feature set corresponding to the EEG training data can be obtained. This time-series feature set includes several intermediate time-series features. In terms of scale, the time-series receptive field of the latter intermediate time-series feature is larger than that of the previous intermediate time-series feature.

[0093] In some embodiments, step 130, extracting time-series features from the EEG training data to obtain a time-series feature set, includes:

[0094] B1: Input the EEG training data into several cascaded time-series convolutional structures for feature cascade extraction to obtain the intermediate time-series features output by each time-series convolutional structure;

[0095] B2: Obtain the time-series feature set according to all the intermediate time-series features.

[0096] In the embodiments of the present application, referring to Figure 2, The temporal feature extractor can be improved from a causal convolutional layer, which includes several cascaded temporal convolutional structures. Each temporal convolutional structure is used to capture the potential change trends of EEG training data in each time period. Each temporal convolutional structure can include a one-dimensional convolutional layer (Conv1d) and a batch normalization layer (BatchNorm1d) stacked in sequence.

[0097] It can be understood that step B1 can input the EEG training data into several cascaded temporal convolutional structures, and based on the power-law decay relationship between the cascaded temporal convolutional structures, determine the output of each temporal convolutional structure as the intermediate temporal feature corresponding to this temporal convolutional structure. Specifically, the output length of each temporal convolutional structure is half of the output length of the previous temporal convolutional structure. In terms of scale, the temporal receptive field of the intermediate temporal feature output by each temporal convolutional structure is larger than that of the intermediate temporal feature output by the previous temporal convolutional structure.

[0098] It should be noted that step B2 can be to splice all the obtained intermediate temporal features together and determine the spliced temporal features as the temporal feature set, which is used to provide comprehensive feature information from the time-domain perspective.

[0099] It is worth mentioning that the traditional causal convolutional layer only takes the output of the last layer as the final output, and it does not consider the periodic nature of EEG data in time series. However, the embodiment of the present application is based on a temporal convolutional structure with a power-law decay relationship, which can fully improve the comprehensiveness of feature information from the time-domain perspective and is beneficial to improving the objectivity and detection accuracy of the depression detection model in subsequent depression detection.

[0100] Step 140: Perform spectral proxy graph learning on the spectral feature set to obtain a spectral encoded representation, and perform temporal proxy graph learning on the temporal feature set to obtain a temporal encoded representation;

[0101] In the embodiment of the present application, step 140 can be to perform proxy graph learning on the spectral feature set and the temporal feature set respectively, so as to obtain a spectral encoded representation corresponding to the spectral feature set and a temporal encoded representation corresponding to the temporal feature set.

[0102] In some embodiments, performing target proxy graph learning on the target feature set to obtain a first target encoded representation includes:

[0103] C1: Generate a target proxy graph for the target feature set to obtain a target functional proxy graph and a target spatial proxy graph corresponding to the target feature set;

[0104] Further, the step C1: Generate a target proxy graph for the target feature set to obtain a target functional proxy graph and a target spatial proxy graph corresponding to the target feature set, includes:

[0105] C11. Obtain the preset function graph construction rule and space graph construction rule, as well as the original node data of the target feature set in each brain region;

[0106] C12. Perform proxy node adaptive generation processing on all the original node data to obtain a number of proxy node data, and each proxy node data corresponds to one of the brain regions;

[0107] C13. According to the function graph construction rule, perform function topology graph construction on the original node data and the proxy node data to obtain the target function proxy graph;

[0108] C14. According to the space graph construction rule, perform space topology graph construction on the original node data and the proxy node data to obtain the target space proxy graph.

[0109] In the embodiment of the present application, the function graph construction rule is used to construct the topological structure of the function proxy graph. Specifically, for the proxy layer hierarchical structure composed of three different connections (edges), the connection can be expressed as:

[0110] E ∈ {e o,o , e o,a , e a,a}

[0111] Wherein, E is the set of connections of the proxy layer hierarchical structure; e o,o is the connection between one original node and another original node; e o,a is the connection between one original node and one proxy node; e a,a is the connection between one proxy node and another proxy node.

[0112] Then the corresponding function graph construction rule can be that the connection between the proxy node and the original node belonging to the same brain region in the adjacency matrix is 1; or, the connection between the proxy node and the original node not belonging to the same brain region in the adjacency matrix is 0. Wherein, the adjacency matrix can be the matrix set corresponding to the connections between all nodes.

[0113] It can be understood that the space graph construction rule is used to construct the topological structure of the space proxy graph. The original node data can be the node positions and node features of the original nodes of the target feature set in each brain region. Specifically, it can be the node positions and node features of the frequency spectrum feature set in each brain region, or the node positions and node features of the time series feature set in each brain region. In the embodiment of the present application, the target feature set is taken as an example of the time series feature set. The content of the frequency spectrum feature set is similar to that of the time series feature set, and can be simply deduced by analogy.

[0114] It should be noted that for brain regions, the brain can be divided into several brain regions based on functional relevance. Specifically, the brain regions can be divided based on the 128 electrodes in the HydroCel electroencephalogram system. This application does not limit the specific brain region division method or the number of divided brain regions here. For example, the number of divided brain regions can be 14. The examples in this application are only for illustration.

[0115] It is worth mentioning that for the original node data of a certain brain region, step C12 can be to adaptively generate proxy node data corresponding to the original node data of the brain region based on the self-attention mechanism. The proxy node data includes the node features and node positions of the proxy nodes. The node features of the proxy nodes can be expressed as:

[0116]

[0117] Among them, is the node feature of the proxy node; is the node feature of the original node; α r is the self-attention weight of the original node data; W v is the value vector of the node feature ; W q is the query vector of the node feature ; W k is the key vector of the node feature ; d k is the dimension of the key vector W k ; softmax(·) is the normalized exponential function; T is the transpose symbol.

[0118] And the node position of the proxy node can be expressed as:

[0119]

[0120] Among them, is the node position of the proxy node; is the node position of the original node.

[0121] It should be added that after obtaining the proxy node data of each brain region, a topological structure corresponding to the original node data and the proxy node data can be constructed based on the functional graph construction rules, and the constructed topological structure can be determined as the target functional proxy graph. Specifically, since the proxy node data is node data corresponding to and adaptively generated from the original node data, the determined target functional proxy graph is a dynamic functional proxy graph; also, if the original node data is the node data of the spectral feature set, the obtained target functional proxy graph is a spectral functional proxy graph; or, if the original node data is the node data of the temporal feature set, the obtained target functional proxy graph is a temporal functional proxy graph.

[0122] It should be noted that step C14 may be to construct a topological structure corresponding to the original node data and the proxy node data based on the spatial graph construction rules, and determine the constructed topological structure as the target functional proxy graph. Specifically, the spatial distances between all nodes in the original node data and the proxy node data can be obtained, and a corresponding spatial topological structure can be generated based on the inverse of the distance. At the same time, a non-negative scaling parameter is used to constrain the connection strength, so as to obtain a static target spatial proxy graph, which may be a time-series spatial proxy graph or a spectral spatial proxy graph.

[0123] C2. Perform target graph learning on the target spatial proxy graph according to the target functional proxy graph to obtain the first target encoded representation;

[0124] Further, step C2, performing target graph learning on the target spatial proxy graph according to the target functional proxy graph to obtain the first target encoded representation, includes:

[0125] C21. Perform first multi-level cascaded graph convolution processing on the target functional proxy graph to obtain a target functional feature representation;

[0126] C22. Perform second multi-level cascaded graph convolution processing on the target spatial proxy graph to obtain a target spatial feature representation;

[0127] C23. Obtain the first target encoded representation according to the target functional feature representation and the target spatial feature representation.

[0128] In the embodiments of the present application, two identical graph learning modules can be constructed, and the target functional proxy graph and the target spatial proxy graph are respectively input into different graph learning modules for graph learning, and then the node feature representations output by the two graph learning modules are aggregated to obtain the first target encoded representation corresponding to the target functional proxy graph and the target spatial proxy graph. The first target encoded representation may specifically be a spectral encoded representation of a spectral feature set or a time-series encoded representation of a time-series feature set.

[0129] It can be understood that step C21 may be to use the node features and node positions of the original nodes in the target functional proxy graph, as well as the node features and node positions of the proxy nodes as the input of a multi-level cascaded graph convolution network, and perform multi-scale graph learning through the multi-level cascaded graph convolution network to obtain a target functional feature representation.

[0130] It should be noted that the multi-level cascaded graph convolution processing of the target space proxy graph is similar to the content of the multi-level cascaded graph convolution processing of the aforementioned target function proxy graph, and can be simply deduced by analogy. Therefore, this application will not elaborate here. Step C23 can perform an aggregation operation on the target function feature representation and the target space feature representation to obtain the first target encoding representation.

[0131] Step 150: Perform dual-domain feature fusion on the time-series encoding representation according to the spectrum encoding representation to obtain a dual-domain fusion feature.

[0132] In the embodiments of the present application, feature fusion can be performed on the spectrum encoding representation and the time-series encoding representation from the time-domain and frequency-domain perspectives to obtain a dual-domain fusion feature.

[0133] In some embodiments, step 150: Perform dual-domain feature fusion on the time-series encoding representation according to the spectrum encoding representation to obtain a dual-domain fusion feature, includes:

[0134] D1: Perform intra-frequency-domain feature fusion on the spectrum encoding representation to obtain an intra-frequency-domain fusion feature, and perform intra-time-domain feature fusion on the time-series encoding representation to obtain an intra-time-domain fusion feature.

[0135] Furthermore, performing intra-target-domain feature fusion on the second target encoding representation to obtain an intra-target-domain fusion feature includes:

[0136] D11: Obtain the target function feature representation and the target space feature representation in the second target encoding representation, where the second target encoding representation is the spectrum encoding representation or the time-series encoding representation;

[0137] D12: Perform first attention analysis on the target function feature representation to obtain the functional attention weight of the target function feature representation, and perform second attention analysis on the target space feature representation to obtain the spatial attention weight of the target space feature representation;

[0138] D13: Perform first weighted fusion on the target function feature representation and the target space feature representation according to the functional attention weight and the spatial attention weight to obtain the intra-target-domain fusion feature.

[0139] In the embodiments of the present application, the second target encoding representation can be the spectrum encoding representation or the time-series encoding representation in step D1. Taking the second target encoding representation as the spectrum encoding representation in step D1 as an example in the embodiments of the present application, the case where the second target encoding representation is the time-series encoding representation in step D1 can be simply deduced by analogy.

[0140] It can be understood that step D11 may be to obtain the spectral functional feature representation and the spectral spatial feature representation in the spectral coding representation; then, based on the cross-attention network, obtain the functional attention weights corresponding to the spectral functional feature representation and the spatial attention weights corresponding to the spectral spatial feature representation respectively. Step D13 may perform weighted fusion on the spectral functional feature representation and the spectral spatial feature representation according to the functional attention weights and the spatial attention weights, so as to obtain the in-frequency-domain fusion feature, which can effectively eliminate the redundant part existing in the EEG training data in the frequency domain.

[0141] It should be noted that in another embodiment, step D13 may also introduce a difference loss with orthogonal constraints, and perform weighted fusion on the spectral functional feature representation and the spectral spatial feature representation based on the difference loss between the spectral functional feature representation and the spectral spatial feature representation, the functional attention weights, and the spatial attention weights, so as to obtain the in-frequency-domain fusion feature, which can more effectively eliminate the redundant part existing in the EEG training data in the frequency domain.

[0142] D2. Perform time-frequency domain cross-feature fusion on the time series coding representation according to the spectral coding representation to obtain a time-frequency domain cross-fusion feature;

[0143] Further, the step D2, performing time-frequency domain cross-feature fusion on the time series coding representation according to the spectral coding representation to obtain a time-frequency domain cross-fusion feature, includes:

[0144] D21. Perform a third attention analysis on the spectral coding representation to obtain frequency domain attention weights;

[0145] D22. Perform a fourth attention analysis on the time series coding representation to obtain time domain attention weights;

[0146] D23. Perform a second weighted fusion on the spectral coding representation and the time series coding representation according to the frequency domain attention weights and the time domain attention weights to obtain the time-frequency domain cross-fusion feature.

[0147] In the embodiments of the present application, the contents of steps D21 to D23 are similar to the contents of the foregoing steps D11 to D13, and can be simply deduced by analogy. Among them, steps D11 to D13 achieve feature fusion within the frequency domain and within the time domain, while steps D21 to D23 achieve feature fusion between the frequency domain and the time domain. The steps D21 to D23 can effectively eliminate the redundant part existing between the frequency domain and the time domain of the EEG training data, and the present application will not elaborate here.

[0148] D3. Obtain the dual-domain fusion feature according to the in-frequency-domain fusion feature, the in-time-domain fusion feature, and the inter-time-frequency-domain fusion feature.

[0149] In the embodiment of the present application, the in-frequency-domain fusion feature, the in-time-domain fusion feature, and the inter-time-frequency-domain fusion feature can be integrated to obtain the dual-domain fusion feature corresponding to the EEG training data.

[0150] Step 160. Update the parameters of the initialized depression detection model according to the dual-domain fusion feature to obtain a trained depression detection model.

[0151] In the embodiment of the present application, step 160 may first be to construct a classifier composed of three serially connected residual blocks and a multi-layer perceptron (MLP), and then input the dual-domain fusion feature into the classifier for prediction and classification to obtain the depression detection result corresponding to the EEG training data. Then, based on the depression detection result and the training label corresponding to the EEG training data, the accuracy of the depression detection model is evaluated, so as to update the parameters of the depression detection model.

[0152] Specifically, for a machine learning model, the accuracy of the model prediction result can be measured by a loss function. The loss function is defined on a single training data and is used to measure the prediction error of a training data. Specifically, the loss value of the training data is determined by the label of the single training data and the prediction result of the model for the training data. During actual training, a training data set has many training data, so generally a cost function is used to measure the overall error of the training data set. The cost function is defined on the entire training data set and is used to calculate the average value of the prediction errors of all training data, which can better measure the prediction effect of the model. For a general machine learning model, based on the foregoing cost function, plus a regularization term that measures the model complexity, it can be used as the training objective function. Based on this objective function, the loss value of the entire training data set can be obtained. There are many types of commonly used loss functions. For example, the 0-1 loss function, the square loss function, the absolute loss function, the logarithmic loss function, the cross-entropy loss function, etc. can all be used as the loss function of the machine learning model, which will not be elaborated one by one here. In the embodiment of the present application, any one of the loss functions can be selected to determine the training loss value. For example, the cross-entropy loss function. Based on the training loss value, the parameters of the model are updated using the backpropagation algorithm, and after several iterations, a trained depression detection model can be obtained. The specific number of iterations can be preset in advance, or it is considered that the training is completed when the accuracy requirement is met in the test set.

[0153] The embodiment of the present application also provides a detection method for a depression detection model, including

[0154] Step 210: Obtain the EEG data to be detected;

[0155] Step 220: Input the EEG data to be detected into the above-trained depression detection model, and obtain the depression detection result output by the trained depression detection model.

[0156] In the embodiment of the present application, the EEG data to be detected can be obtained, and the EEG data to be detected is input into the trained depression detection model. The trained depression detection model is used to detect depression in the EEG data to be detected, so as to obtain the depression detection result.

[0157] Next, a training system for a depression detection model proposed according to an embodiment of the present application will be described in detail with reference to the accompanying drawings.

[0158] Refer to Figure 3 , a training system for a depression detection model proposed in the embodiment of the present application includes:

[0159] The first processing unit 101 is configured to obtain EEG training data;

[0160] The second processing unit 102 is configured to extract spectral features from the EEG training data to obtain a spectral feature set. The spectral feature set includes a plurality of intermediate spectral features, and the spectral frequency bands of each intermediate spectral feature are different;

[0161] The third processing unit 103 is configured to extract temporal features from the EEG training data to obtain a temporal feature set. The temporal feature set includes a plurality of cascaded intermediate temporal features, and the temporal receptive field of the latter intermediate temporal feature is larger than that of the previous intermediate temporal feature;

[0162] The fourth processing unit 104 is configured to perform spectral proxy graph learning on the spectral feature set to obtain a spectral coding representation, and perform temporal proxy graph learning on the temporal feature set to obtain a temporal coding representation;

[0163] The fifth processing unit 105 is configured to perform dual-domain feature fusion on the temporal coding representation according to the spectral coding representation to obtain a dual-domain fusion feature;

[0164] The sixth processing unit 106 is configured to update the parameters of the initialized depression detection model according to the dual-domain fusion feature to obtain a trained depression detection model.

[0165] It can be understood that the content in the above method embodiments is applicable to the system embodiments of the present application. The functions specifically implemented by the system embodiments of the present application are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those in the above method embodiments.

[0166] Reference Figure 4 , the embodiments of the present application also provide an electronic device, including:

[0167] At least one processor 201;

[0168] At least one memory 202, configured to store at least one program;

[0169] When the at least one program is executed by the at least one processor 201, the at least one processor 201 implements the above method embodiments.

[0170] Similarly, it can be understood that the content in the above method embodiments is applicable to the device embodiments of the present application. The functions specifically implemented by the device embodiments of the present application are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those in the above method embodiments.

[0171] The embodiments of the present application also provide a computer-readable storage medium, in which a program executable by the processor 201 is stored, and the program executable by the processor 201 is used to implement the above method embodiments when executed by the processor 201.

[0172] Similarly, the content in the above method embodiments is applicable to the computer-readable storage medium embodiments of the present application. The functions specifically implemented by the computer-readable storage medium embodiments of the present application are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those in the above method embodiments.

[0173] In some alternative embodiments, the functions / operations mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two consecutive blocks shown may actually be executed substantially simultaneously or the blocks can sometimes be executed in the reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example for the purpose of providing a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logical flows presented herein. Alternative embodiments are foreseeable, in which the order of various operations is changed and the sub-operations described as part of a larger operation are executed independently.

[0174] In addition, although the present application has been described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated in a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It should also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present application. Rather, considering the attributes, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the modules will be understood within the ordinary skills of an engineer. Thus, those skilled in the art can implement the present application as set forth in the claims without undue experimentation. It should also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0175] If a function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method according to the embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0176] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in connection with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0177] More specific examples (nonexhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which a program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0178] It should be understood that various parts of the present application can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0179] In the foregoing description of the present specification, the descriptions referring to the terms "one embodiment / example", "another embodiment / example", or "certain embodiments / examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0180] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the claims and their equivalents.

[0181] The above has specifically described the preferred embodiments of the present application, but the present application is not limited to the embodiments. Those skilled in the art can make various equivalent deformations or substitutions without departing from the spirit of the present application, and these equivalent deformations or substitutions are all included within the scope defined by the claims of the present application.

Claims

1. A method for training a depression detection model, characterized in that: include: Obtain EEG training data; Extracting spectrum features from the EEG training data to obtain a spectrum feature set, wherein the spectrum feature set includes a plurality of intermediate spectrum features, and each of the intermediate spectrum features has a different spectrum band; Extracting time series features from the EEG training data to obtain a time series feature set, wherein the time series feature set includes a plurality of cascaded intermediate time series features, and a time series receptive field of a later intermediate time series feature is larger than a time series receptive field of a previous intermediate time series feature; Performing spectrum proxy graph learning on the spectrum feature set to obtain a spectrum coding representation, and performing time series proxy graph learning on the time series feature set to obtain a time series coding representation; According to the spectrum coding representation, dual-domain feature fusion is performed on the time series coding representation to obtain dual-domain fusion features; According to the dual-domain fusion features, the parameters of the initialized depression detection model are updated to obtain a trained depression detection model.

2. The method according to claim 1, characterized in that The extracting spectrum features of the EEG training data to obtain a spectrum feature set includes: Performing frequency domain conversion on the EEG training data to obtain a plurality of EEG frequency band data, each of which has a different frequency band; Performing differential entropy feature extraction on all the EEG frequency band data to obtain intermediate frequency spectrum features corresponding to each of the EEG frequency band data; According to all the intermediate spectrum features, obtaining the spectrum feature set; The step of extracting time series features from the EEG training data to obtain a time series feature set includes: Inputting the EEG training data into a plurality of cascaded temporal convolution structures for feature cascade extraction, and obtaining intermediate temporal features output by each of the temporal convolution structures; The timing feature set is obtained according to all the intermediate timing features.

3. The method according to claim 1, characterized in that The target proxy graph is learned for the target feature set to obtain the first target encoding representation, including: Generating a target proxy graph for the target feature set to obtain a target function proxy graph and a target space proxy graph corresponding to the target feature set; According to the target function proxy graph, performing target graph learning on the target space proxy graph to obtain the first target encoding representation; Wherein, the target feature set is a spectrum feature set or a time series feature set.

4. The method according to claim 3, characterized in that The generating of a target proxy graph for the target feature set to obtain a target function proxy graph and a target space proxy graph corresponding to the target feature set includes: Obtaining preset functional map construction rules and spatial map construction rules, as well as original node data of the target feature set in each brain region; Performing proxy node adaptive generation processing on all the original node data to obtain a plurality of proxy node data, each of the proxy node data corresponds to one of the brain regions; According to the functional graph construction rule, construct a functional topology graph for the original node data and the proxy node data to obtain the target functional proxy graph; According to the spatial graph construction rule, a spatial topology graph is constructed for the original node data and the proxy node data to obtain the target spatial proxy graph.

5. The method according to claim 3, characterized in that: The step of performing target graph learning on the target space proxy graph according to the target function proxy graph to obtain the first target encoding representation includes: Performing a first multi-layer cascade graph convolution process on the target function proxy graph to obtain a target function feature representation; Performing a second multi-layer cascade graph convolution process on the target space proxy graph to obtain a target space feature representation; The first target coding representation is obtained according to the target function feature representation and the target space feature representation.

6. The method according to claim 1, characterized in that The step of fusing dual-domain features on the temporal coding representation according to the spectral coding representation to obtain dual-domain fused features includes: Performing frequency domain feature fusion on the spectrum coding representation to obtain frequency domain fusion features, and performing time domain feature fusion on the time series coding representation to obtain time domain fusion features; According to the spectrum coding representation, the time-frequency domain feature fusion is performed on the time series coding representation to obtain the time-frequency domain fusion feature; The dual-domain fusion feature is obtained according to the frequency domain intra-domain fusion feature, the time domain intra-domain fusion feature and the time-frequency domain fusion feature.

7. The method according to claim 6, characterized in that The second target encoding representation is subjected to target domain intra-domain feature fusion to obtain target domain intra-domain fusion features, including: Acquire a target function feature representation and a target spatial feature representation in the second target coding representation, wherein the second target coding representation is the spectrum coding representation or the temporal coding representation; Performing a first attention analysis on the target functional feature representation to obtain a functional attention weight of the target functional feature representation, and performing a second attention analysis on the target spatial feature representation to obtain a spatial attention weight of the target spatial feature representation; According to the functional attention weight and the spatial attention weight, a first weighted fusion is performed on the target functional feature representation and the target spatial feature representation to obtain the target domain intra-domain fusion feature.

8. The method according to claim 6, characterized in that The step of fusing the time-frequency domain features of the time series coding representation according to the spectrum coding representation to obtain the time-frequency domain fusion features includes: Performing a third attention analysis on the spectral coded representation to obtain a frequency domain attention weight; Performing a fourth attention analysis on the temporal coding representation to obtain a temporal attention weight; According to the frequency domain attention weight and the time domain attention weight, a second weighted fusion is performed on the spectrum coding representation and the time series coding representation to obtain the time-frequency domain fusion feature.

9. A method for detecting a depression detection model, characterized in that: include: Obtaining EEG data to be tested; The EEG data to be detected is input into the trained depression detection model as described in any one of claims 1 to 8 to obtain the depression detection result output by the trained depression detection model.

10. A depression detection model training system, characterized in that: include: A first processing unit, used for acquiring EEG training data; A second processing unit is used to extract spectrum features from the EEG training data to obtain a spectrum feature set, wherein the spectrum feature set includes a plurality of intermediate spectrum features, and each of the intermediate spectrum features has a different spectrum band; A third processing unit is used to extract time series features from the EEG training data to obtain a time series feature set, wherein the time series feature set includes a plurality of cascaded intermediate time series features, and a time series receptive field of a later intermediate time series feature is larger than a time series receptive field of a previous intermediate time series feature; a fourth processing unit, configured to perform spectrum proxy graph learning on the spectrum feature set to obtain a spectrum coding representation, and to perform time series proxy graph learning on the time series feature set to obtain a time series coding representation; a fifth processing unit, configured to perform dual-domain feature fusion on the temporal coding representation according to the spectrum coding representation to obtain a dual-domain fusion feature; The sixth processing unit is used to update the parameters of the initialized depression detection model according to the dual-domain fusion features to obtain a trained depression detection model.

Images