Training method and system for depression detection model, and detection method of depression detection model
By extracting spectral and temporal features from EEG data and fusing dual-domain features, a depression detection model was constructed, which solved the problem of insufficient objectivity and accuracy in existing depression detection technologies and achieved more efficient depression detection.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2025-09-26
- Publication Date
- 2026-07-30
AI Technical Summary
Existing methods for detecting depression lack objectivity and efficiency, and are not very accurate, making it difficult to accurately capture collaborative connections and long-distance information exchange between brain regions.
By extracting spectral and temporal features from EEG training data, spectral and temporal coding representations are generated, and dual-domain feature fusion is performed to construct a depression detection model. Proxy graph learning and multi-layer cascaded convolutional networks are used for feature interaction and fusion.
It improves the objectivity and accuracy of depression detection, enhances the model's universality across different individuals and situations, and fully captures collaborative connections and long-distance information exchange between brain regions.
Smart Images
Figure CN2025124392_30072026_PF_FP_ABST
Abstract
Description
A training method, detection method, and system for a depression detection model. Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a training method, detection method and system for a depression detection model. Background Technology
[0002] With the continuous development of society, disease detection for patients with depression has become a focus of public attention.
[0003] Currently, traditional methods for detecting depression can generally be divided into two categories. The first type relies on doctor-patient dialogue and questionnaire scores. The diagnostic results depend on the patient's reaction and behavior, as well as the doctor's professional knowledge. This method is easily affected by the subjective consciousness of both the doctor and the patient, resulting in poor objectivity and efficiency in depression detection. The second type is usually based on bandpass filtering and channel selection strategies. It extracts key spatiotemporal features from the optimal channel of each frequency band of the electroencephalogram (EEG) signal and uses a classification model to detect depression. However, the accuracy of this method is not satisfactory.
[0004] Therefore, the problems existing in the current technology still need to be solved and optimized. Summary of the Invention
[0005] The purpose of this invention is to at least partially solve one of the technical problems existing in the related art.
[0006] Therefore, one objective of this invention is to provide a training method, a detection method, and a system for a depression detection model. The training method can provide a depression detection model that improves the objectivity, efficiency, and accuracy of depression detection.
[0007] To achieve the above-mentioned technical objectives, the technical solutions adopted in the embodiments of this application include:
[0008] In a first aspect, embodiments of this application provide a method for training a depression detection model, including:
[0009] Acquire EEG training data;
[0010] The EEG training data is subjected to spectral feature extraction to obtain a spectral feature set, which includes several intermediate spectral features, each of which has a different spectral frequency band.
[0011] Temporal features are extracted from the EEG training data to obtain a temporal feature set, which includes several cascaded intermediate temporal features, wherein the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature.
[0012] Spectrum surrogate graph learning is performed on the spectral feature set to obtain a spectral coded representation, and temporal surrogate graph learning is performed on the temporal feature set to obtain a temporal coded representation;
[0013] Based on the spectral coding representation, the temporal coding representation is subjected to dual-domain feature fusion to obtain dual-domain fused features;
[0014] Based on the dual-domain fusion features, the parameters of the initialized depression detection model are updated to obtain a trained depression detection model.
[0015] In addition, the method according to the above embodiments of this application may also have the following additional technical features:
[0016] Furthermore, in one embodiment of this application, the step of extracting spectral features from the EEG training data to obtain a spectral feature set includes:
[0017] The EEG training data is frequency domain converted to obtain several EEG frequency band data, each of which has a different frequency band.
[0018] Differential entropy feature extraction is performed on all the EEG frequency band data to obtain intermediate spectral features corresponding to each EEG frequency band data;
[0019] The spectral feature set is obtained based on all the intermediate spectral features;
[0020] The step of extracting temporal features from the EEG training data to obtain a temporal feature set includes:
[0021] The EEG training data is input into several cascaded temporal convolutional structures for feature extraction, thereby obtaining the intermediate temporal features output by each temporal convolutional structure.
[0022] The time series feature set is obtained based on all the intermediate time series features.
[0023] Furthermore, in one embodiment of this application, the step of performing target proxy graph learning on the target feature set to obtain a first target encoding representation includes:
[0024] Target proxy graphs are generated from the target feature set to obtain target function proxy graphs and target space proxy graphs corresponding to the target feature set.
[0025] Based on the target function proxy graph, target graph learning is performed on the target space proxy graph to obtain the first target encoding representation;
[0026] The target feature set is either a spectral feature set or a temporal feature set.
[0027] Furthermore, in one embodiment of this application, the step of generating a target proxy graph from the target feature set to obtain a target functional proxy graph and a target spatial proxy graph corresponding to the target feature set includes:
[0028] Obtain preset functional graph construction rules and spatial graph construction rules, as well as the original node data of the target feature set in each brain region;
[0029] All the original node data are subjected to proxy node adaptive generation processing to obtain a number of proxy node data, each of the proxy node data corresponding to a brain region;
[0030] According to the functional graph construction rules, a functional topology graph is constructed on the original node data and the proxy node data to obtain the target functional proxy graph;
[0031] According to the spatial graph construction rules, a spatial topology graph is constructed from the original node data and the proxy node data to obtain the target spatial proxy graph.
[0032] Further, in one embodiment of this application, the step of performing target graph learning on the target spatial proxy graph based on the target functional proxy graph to obtain the first target encoding representation includes:
[0033] The target function proxy graph is subjected to a first multi-level concatenated graph convolution process to obtain the target function feature representation;
[0034] The target space proxy graph is subjected to a second multi-layer concatenated graph convolution process to obtain the target space feature representation;
[0035] The first target encoding representation is obtained based on the target functional feature representation and the target spatial feature representation.
[0036] Furthermore, in one embodiment of this application, the step of performing dual-domain feature fusion on the temporal coding representation based on the spectral coding representation to obtain dual-domain fused features includes:
[0037] The frequency domain feature fusion is performed on the spectral coding representation to obtain the frequency domain fused feature, and the time domain feature fusion is performed on the temporal coding representation to obtain the time domain fused feature.
[0038] Based on the spectral coding representation, the time-frequency domain feature fusion is performed on the time-series coding representation to obtain the time-frequency domain fused feature;
[0039] The dual-domain fusion feature is obtained based on the frequency domain intra-domain fusion feature, the time domain intra-domain fusion feature, and the time-frequency domain inter-domain fusion feature.
[0040] Furthermore, in one embodiment of this application, target domain intra-domain feature fusion is performed on the second target encoding representation to obtain target domain intra-domain fused features, including:
[0041] Obtain the target functional feature representation and the target spatial feature representation from the second target coding representation, wherein the second target coding representation is the spectral coding representation or the temporal coding representation;
[0042] A first attention analysis is performed on the target functional feature representation to obtain the functional attention weights of the target functional feature representation, and a second attention analysis is performed on the target spatial feature representation to obtain the spatial attention weights of the target spatial feature representation;
[0043] Based on the functional attention weight and the spatial attention weight, the target functional feature representation and the target spatial feature representation are subjected to a first weighted fusion to obtain the target domain intradomain fusion feature.
[0044] Furthermore, in one embodiment of this application, the step of performing time-frequency domain feature fusion on the time-series coding representation based on the spectral coding representation to obtain time-frequency domain fused features includes:
[0045] A third attention analysis is performed on the spectral encoding representation to obtain the frequency domain attention weights;
[0046] A fourth attention analysis is performed on the temporal coding representation to obtain temporal attention weights;
[0047] Based on the frequency domain attention weight and the time domain attention weight, the spectral coding representation and the temporal coding representation are subjected to a second weighted fusion to obtain the time-frequency domain fusion feature.
[0048] Secondly, embodiments of this application provide a detection method for a depression detection model, comprising:
[0049] Acquire EEG data to be tested;
[0050] The EEG data to be detected is input into the trained depression detection model described above to obtain the depression detection result output by the trained depression detection model.
[0051] Thirdly, embodiments of this application provide a training system for a depression detection model, comprising:
[0052] The first processing unit is used to acquire EEG training data;
[0053] The second processing unit is used to extract spectral features from the EEG training data to obtain a spectral feature set, which includes several intermediate spectral features, each of which has a different spectral frequency band.
[0054] The third processing unit is used to extract temporal features from the EEG training data to obtain a temporal feature set. The temporal feature set includes several cascaded intermediate temporal features, and the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature.
[0055] The fourth processing unit is used to perform spectral surrogate graph learning on the spectral feature set to obtain a spectral coded representation, and to perform temporal surrogate graph learning on the temporal feature set to obtain a temporal coded representation;
[0056] The fifth processing unit is used to perform dual-domain feature fusion on the temporal coding representation according to the spectral coding representation to obtain dual-domain fused features;
[0057] The sixth processing unit is used to update the parameters of the initialized depression detection model based on the dual-domain fusion features to obtain a trained depression detection model.
[0058] Fourthly, embodiments of this application also provide an electronic device, including:
[0059] At least one processor;
[0060] At least one memory for storing at least one program;
[0061] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.
[0062] Fifthly, embodiments of this application also provide a computer-readable storage medium storing a processor-executable program, which, when executed by the processor, is used to implement the above-described method.
[0063] The advantages and beneficial effects of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application:
[0064] This application discloses a training method, detection method, and system for a depression detection model. The training method involves: acquiring EEG training data; extracting spectral features from the EEG training data to obtain a spectral feature set, which includes several intermediate spectral features, each with a different spectral frequency band; extracting temporal features from the EEG training data to obtain a temporal feature set, which includes several cascaded intermediate temporal features, where the temporal receptive field of a later intermediate temporal feature is larger than that of a previous intermediate temporal feature; performing spectral surrogate graph learning on the spectral feature set to obtain a spectral encoding representation, and performing temporal surrogate graph learning on the temporal feature set to obtain a temporal encoding representation; performing dual-domain feature fusion on the temporal encoding representation based on the spectral encoding representation to obtain dual-domain fused features; and updating the parameters of an initialized depression detection model based on the dual-domain fused features to obtain a trained depression detection model. This training method extracts temporal features from EEG training data to obtain a temporal feature set where the temporal receptive field of the subsequent intermediate temporal feature is larger than that of the preceding intermediate temporal feature, which can effectively improve the comprehensiveness of feature information from a temporal perspective. Furthermore, this method extracts spectral features from EEG training data to obtain a spectral feature set composed of intermediate spectral features from different spectral bands, which can effectively improve the comprehensiveness of feature information from a spectral perspective. Based on the temporal and spectral feature sets, this method helps improve the objectivity and accuracy of depression detection models, as well as the efficiency of depression detection. Attached Figure Description
[0065] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of this application or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions in this application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0066] Figure 1 is a flowchart illustrating a training method for a depression detection model provided in an embodiment of this application;
[0067] Figure 2 is a simplified structural diagram of a temporal feature extractor provided in an embodiment of this application;
[0068] Figure 3 is a schematic diagram of the structural framework of a training system for a depression detection model provided in an embodiment of this application;
[0069] Figure 4 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0070] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation 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 one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0072] Currently, traditional methods for detecting depression can generally be divided into two categories. The first category relies on doctor-patient dialogue and questionnaire scores. The diagnostic results depend on the patient's reaction and behavior, as well as the doctor's professional knowledge. This method is easily affected by the subjective consciousness of both doctors and patients, resulting in poor objectivity and efficiency in depression detection. The second category is usually based on bandpass filtering and channel selection strategies. It extracts key spatiotemporal features from the optimal channel of each frequency band of the electroencephalogram (EEG) signal and uses a classification model to detect depression. However, the spatiotemporal features extracted by this method are often incomplete, making the accuracy of subsequent classification models in detecting depression based on spatiotemporal features unsatisfactory.
[0073] Furthermore, since depression is a serious mental illness, its abnormal manifestations in the brain are characterized by unbalanced connections between brain regions, rather than a simple increase or decrease in activity in a specific area. Some existing technologies fully simulate and align the distribution of EEG electrodes with neurophysiological structures and encode all node pairs to achieve depression detection by learning a dense, fully connected adjacency matrix. However, this simulated alignment approach ignores important collaborative relationships between different brain regions, significantly limiting the performance of depression detection and making it difficult to achieve accurate and efficient results. Moreover, the collaborative connections between specific electrodes learned by the dense, fully connected adjacency matrix represent a general inference pattern, while the collaborative connections between some specific electrodes are redundant and noisy, inevitably hindering information propagation.
[0074] Furthermore, in brain regions, the axons and dendrites of interneurons are confined to a single brain region, primarily facilitating local information exchange; while principal neurons typically extend their axons beyond the brain region containing their cell bodies and dendrites, primarily facilitating long-distance information exchange. Current technologies neglect targeted modeling of principal neurons, making it difficult to capture global information interactions between brain regions, leading to instability in overall representation and affecting the accuracy of depression detection.
[0075] Furthermore, existing technologies for depression detection use bi-branch learning methods or progressive graph convolutional networks. The interaction between global and local features occurs only after the feature learning process, failing to consider their dynamic impact. In addition, the long-distance connections constructed manually cannot adaptively adjust the connection strength based on signal correlation, resulting in poor universality of depression detection models applied to different individuals and scenarios.
[0076] Furthermore, existing technologies ignore the redundancy of time-frequency features in both intra-domain and cross-domain features, which can easily lead to insufficient extraction of low-level information from EEG data and low comprehensiveness of generated feature information, thereby resulting in low accuracy of depression detection models in detecting depression.
[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. The training method extracts temporal features from EEG training data to obtain a temporal feature set where the temporal receptive field of the subsequent intermediate temporal feature is larger than that of the preceding intermediate temporal feature, effectively improving the comprehensiveness of feature information from a temporal perspective. Furthermore, the training method extracts spectral features from EEG training data to obtain a spectral feature set composed of intermediate spectral features from different spectral bands, effectively improving the comprehensiveness of feature information from a spectral perspective. Based on the temporal feature set and the spectral feature set, it is beneficial to improve the objectivity and accuracy of depression detection in the depression detection model, as well as the detection efficiency of depression detection.
[0078] Furthermore, this training method learns surrogate graphs from spectral and temporal feature sets respectively. Specifically, it generates corresponding surrogate node data for each brain region through an adaptive strategy, and uses the surrogate node data to achieve dynamic interaction between local and global information from brain electrodes. This allows the depression detection model to fully consider the collaborative connections between different brain regions and to capture long-distance connections between brain regions. At the same time, it constructs spatial surrogate graphs based on spatial graph construction rules, which allows the constructed spatial surrogate graphs to adaptively adjust connection strength, which helps to improve the stability of the overall representation and the versatility of the depression detection model in different individuals and scenarios.
[0079] Furthermore, this training method employs graph learning on surrogate maps, specifically through multi-level cascaded convolutions on both functional and spatial surrogate maps. This allows for interactive learning of the surrogate maps across different receptive fields, facilitating the full extraction and integration of key information contained in EEG data at different scales. This, in turn, improves the detection accuracy of depression detection models. Additionally, this training method, based on intra-domain and inter-domain feature fusion of the encoded representation, effectively eliminates redundant parts in EEG data, generating comprehensive and accurate dual-domain fusion features, further enhancing the detection accuracy of depression detection models.
[0080] Referring to Figure 1, in an embodiment of this application, a method for training a depression detection model includes:
[0081] Step 110: Obtain EEG training data;
[0082] In the embodiments of this application, the EEG training data can be EEG signals collected by electrodes on the subject's scalp, or preprocessed EEG signals. Specific preprocessing steps may include at least one of data filtering, data segmentation, baseline correction, artifact removal, and motion artifact removal, etc., which are not limited in this application.
[0083] Step 120: Extract spectral features from the EEG training data to obtain a spectral feature set, which includes several intermediate spectral features, each of which has a different spectral frequency band.
[0084] In this embodiment of the application, step 120 may be to construct a spectrum feature extractor, and extract features by inputting EEG training data into the spectrum feature extractor to obtain a spectrum feature set corresponding to the EEG training data. The spectrum feature set includes intermediate spectrum features of several different frequency bands, and the intermediate spectrum features of each frequency band are used to provide feature information of the corresponding frequency band from a spectrum perspective.
[0085] In some embodiments, step 120, extracting spectral features from the EEG training data to obtain a spectral feature set, includes:
[0086] A1. The EEG training data is converted to the frequency domain to obtain several EEG frequency band data, each of which has a different frequency band.
[0087] A2. 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;
[0088] A3. Based on all the intermediate spectral features, obtain the spectral feature set;
[0089] In this embodiment, frequency domain analysis can be performed on the EEG training data based on Fast Fourier Transform (FFT) to convert the EEG training data to the frequency domain and obtain EEG frequency band data divided into several frequency bands. In this embodiment, taking a total of 5 EEG frequency bands as an example, the specific frequency bands of the EEG frequency band data can be 1-3 Hz, 4-7 Hz, 8-13 Hz, 14-30 Hz, and 31-50 Hz.
[0090] It is understandable that, for a specific EEG frequency band, step A2 can be to obtain the differential entropy representation of the EEG frequency band data (i.e., intermediate spectral features). These intermediate spectral features are used to characterize the information content and complexity of the corresponding frequency band EEG data. Various methods for extracting the specific differential entropy representation already exist, and the same applies to other EEG frequency band data; therefore, this application will not elaborate further here. Furthermore, after obtaining the intermediate spectral features corresponding to each frequency band, all the obtained intermediate spectral features can be concatenated, and the concatenated spectral features can be determined as a spectral feature set, which is used to provide comprehensive feature information from a spectral perspective.
[0091] Step 130: Extract temporal features from the EEG training data to obtain a temporal feature set. The temporal feature set includes several cascaded intermediate temporal features, wherein the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature.
[0092] In this embodiment of the application, step 120 may be to construct a temporal feature extractor, and extract features by inputting EEG training data into the temporal feature extractor to obtain a temporal feature set corresponding to the EEG training data. The temporal feature set includes several intermediate temporal features. In terms of scale, the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature.
[0093] In some embodiments, step 130, extracting temporal features from the EEG training data to obtain a temporal feature set, includes:
[0094] B1. Input the EEG training data into several cascaded temporal convolutional structures to perform feature cascade extraction, and obtain the intermediate temporal features output by each temporal convolutional structure;
[0095] B2. Based on all the intermediate time series features, the time series feature set is obtained.
[0096] In this embodiment of the application, referring to FIG2, the temporal feature extractor can be obtained by improving the causal convolutional layer, which includes several cascaded temporal convolutional structures. Each temporal convolutional structure is used to capture the potential change trend of EEG training data in various time periods. Each temporal convolutional structure may include sequentially stacked one-dimensional convolutional layers (Conv1d) and batch normalization layers (BatchNorm1d).
[0097] Understandably, step B1 can input EEG training data into several cascaded temporal convolutional structures. Based on the power-law decay relationship between the cascaded temporal convolutional structures, the output of each temporal convolutional structure is determined as the intermediate temporal feature corresponding to that structure. Specifically, the output length of each temporal convolutional structure is half the output length of the previous temporal convolutional structure. Therefore, 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 concatenate all the acquired intermediate temporal features and determine the concatenated temporal features as a temporal feature set, which is used to provide comprehensive feature information from a temporal perspective.
[0099] It is worth mentioning that traditional causal convolutional layers only take the output of the last layer as the final output, and do not consider the periodic nature of EEG data in time. However, the embodiments of this application are based on a temporal convolutional structure with power-law decay, which can fully improve the comprehensiveness of feature information from a temporal perspective, and is conducive to improving the objectivity and detection accuracy of the depression detection model in subsequent depression detection.
[0100] Step 140: Perform spectral surrogate graph learning on the spectral feature set to obtain a spectral coded representation, and perform temporal surrogate graph learning on the temporal feature set to obtain a temporal coded representation;
[0101] In this embodiment of the application, step 140 may involve performing surrogate graph learning on the spectral feature set and the temporal feature set respectively, so as to obtain the spectral coding representation corresponding to the spectral feature set and the temporal coding representation corresponding to the temporal feature set.
[0102] In some embodiments, target proxy graph learning is performed on the target feature set to obtain a first target encoding representation, including:
[0103] C1. Generate 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;
[0104] Further, step C1, 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, includes:
[0105] C11. Obtain the preset functional graph construction rules and spatial graph construction rules, 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, each of the proxy node data corresponding to a brain region;
[0107] C13. Based on the functional graph construction rules, construct a functional topology graph from the original node data and the proxy node data to obtain the target functional proxy graph;
[0108] C14. Based on the spatial graph construction rules, construct a spatial topology graph from the original node data and the proxy node data to obtain the target spatial proxy graph.
[0109] In this embodiment, the functional graph construction rules are used to construct the topological structure of the functional proxy graph. Specifically, for a proxy layer hierarchical structure composed of three different connections (edges), its connection can be represented as: E∈{e o,o ,e o,a ,e a,a}
[0110] Where E is the set of connections in the proxy layer-level structure; e o,o A connection between one original node and another original node; e o,a A connection between a primary node and a proxy node; e a,a This refers to a connection between one proxy node and another proxy node.
[0111] The corresponding functional graph construction rule can be that, in the adjacency matrix, the connection between the surrogate node belonging to the same brain region and the original node is 1; or, in the adjacency matrix, the connection between the surrogate node not belonging to the same brain region and the original node is 0. The adjacency matrix can be the set of matrices corresponding to all connections between nodes.
[0112] It is understandable that spatial graph construction rules are used to construct the topology of spatial proxy graphs. The original node data can be the node positions and node features of the target feature set in various brain regions. Specifically, it can be the node positions and node features of the spectral feature set in various brain regions, or the node positions and node features of the temporal feature set in various brain regions. In this embodiment, the target feature set is taken as a temporal feature set. The content of the spectral feature set is similar to the content of the temporal feature set and can be easily deduced by analogy.
[0113] It should be noted that, for brain regions, the brain can be divided into several brain regions based on functional relevance. Specifically, brain regions can be divided based on the 128 electrodes in the HydroCel EEG system. This application does not impose any restrictions on the specific brain region division method or the number of brain regions to be divided. For example, the number of brain regions to be divided can be 14. The example in this application is only for illustration.
[0114] It is worth mentioning that, for the original node data of a certain brain region, step C12 can be based on a self-attention mechanism to adaptively generate surrogate node data corresponding to the original node data of that brain region. The surrogate node data includes the node features and node location of the surrogate node. The node features of the surrogate node can be represented as follows:
[0115] in, The node characteristics of the proxy node; The node characteristics of the original node; α r W represents the self-attention weights of the original node data. v Node features The value vector of W; q Node features The query vector; W k Node features The key vector; d k For the key vector W k The dimension; softmax(·) is the normalization exponential function; T is the transpose sign.
[0116] The node position of this proxy node can be represented as:
[0117] in, The node location of the proxy node; This represents the node position of the original node.
[0118] It should be added that after obtaining the surrogate node data for each brain region, a topology corresponding to the original node data and surrogate node data can be constructed based on functional graph construction rules, and the constructed topology is determined as the target functional surrogate graph. Specifically, since the surrogate node data is node data that corresponds to the original node data and is adaptively generated, the determined target functional surrogate graph is a dynamic functional surrogate graph; furthermore, if the original node data is node data of a spectral feature set, the obtained target functional surrogate graph is a spectral functional surrogate graph; or, if the original node data is node data of a temporal feature set, the obtained target functional surrogate graph is a temporal functional surrogate graph.
[0119] It is worth noting that step C14 can be based on spatial graph construction rules to construct a topology corresponding to the original node data and proxy node data, and the constructed topology is determined as the target functional proxy graph. Specifically, the spatial distance between all nodes in the original node data and proxy node data can be obtained, and the corresponding spatial topology can be generated based on the inverse ratio of the distance. At the same time, a non-negative scaling parameter is used to constrain the connection strength, thereby obtaining a static target spatial proxy graph. This target spatial proxy graph can be a temporal spatial proxy graph or a spectral spatial proxy graph.
[0120] C2. Based on the target function proxy graph, perform target graph learning on the target space proxy graph to obtain the first target encoding representation;
[0121] Further, step C2, performing target graph learning on the target spatial proxy graph based on the target functional proxy graph to obtain the first target encoding representation, includes:
[0122] C21. Perform a first multi-level concatenated graph convolution on the target function proxy graph to obtain the target function feature representation;
[0123] C22. Perform a second multi-layer concatenated graph convolution on the target space proxy graph to obtain the target space feature representation;
[0124] C23. Based on the target functional feature representation and the target spatial feature representation, the first target encoding representation is obtained.
[0125] In this embodiment of the application, two identical graph learning modules can be constructed, and the target functional proxy graph and the target spatial proxy graph can be input into different graph learning modules for graph learning. Then, by aggregating the node feature representations output by the two graph learning modules, a first target encoding representation corresponding to the target functional proxy graph and the target spatial proxy graph can be obtained. Specifically, the first target encoding representation can be a spectral encoding representation of a spectral feature set or a temporal encoding representation of a temporal feature set.
[0126] It is understandable that step C21 can be to take the node features and node positions of the original nodes in the target function proxy graph, as well as the node features and node positions of the proxy nodes, as input to a multi-layer cascaded graph convolutional network, and perform multi-scale graph learning through the multi-layer cascaded graph convolutional network to obtain the target function feature representation.
[0127] It should be noted that the multi-level convolutional processing of the target space proxy graph is similar to that of the aforementioned multi-level convolutional processing of the target function proxy graph, and can be easily deduced by analogy. Therefore, this application will not elaborate further. 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.
[0128] Step 150: Based on the spectral coding representation, perform dual-domain feature fusion on the temporal coding representation to obtain dual-domain fused features;
[0129] In the embodiments of this application, feature fusion can be performed on the spectral coding representation and the temporal coding representation from both the time domain and frequency domain perspectives to obtain dual-domain fused features.
[0130] In some embodiments, step 150, fusing dual-domain features on the temporal coding representation according to the spectral coding representation to obtain dual-domain fused features, includes:
[0131] D1. Perform frequency domain feature fusion on the spectral coding representation to obtain frequency domain fused features, and perform time domain feature fusion on the temporal coding representation to obtain time domain fused features;
[0132] Furthermore, the second target encoding representation is subjected to target domain intra-domain feature fusion to obtain target domain intra-domain fused features, including:
[0133] D11. Obtain the target functional feature representation and the target spatial feature representation from the second target coding representation, wherein the second target coding representation is the spectrum coding representation or the time-series coding representation;
[0134] D12. Perform a first attention analysis on the target functional feature representation to obtain the functional attention weight of the target functional feature representation, and perform a second attention analysis on the target spatial feature representation to obtain the spatial attention weight of the target spatial feature representation.
[0135] D13. Based on the functional attention weight and the spatial attention weight, perform a first weighted fusion on the target functional feature representation and the target spatial feature representation to obtain the target domain intradomain fusion feature.
[0136] In this embodiment of the application, the second target encoding representation can be either the spectral encoding representation or the temporal encoding representation in step D1. In this embodiment of the application, the second target encoding representation is taken as the spectral encoding representation in step D1. The second target encoding representation can be derived by analogy as the temporal encoding representation in step D1.
[0137] Understandably, step D11 can involve obtaining the spectral functional feature representation and the spectral spatial feature representation from the spectral coding representation; then, based on the cross-attention network, obtaining the functional attention weights corresponding to the spectral functional feature representation and the spatial attention weights corresponding to the spectral spatial feature representation. Step D13 can perform weighted fusion of the spectral functional feature representation and the spectral spatial feature representation according to the functional attention weights and spatial attention weights, thereby obtaining the intra-frequency domain fusion feature. This intra-frequency domain fusion feature can effectively eliminate the redundant parts of the EEG training data in the frequency domain.
[0138] It should be noted that, in another implementation, step D13 can also introduce a difference loss with orthogonal constraints, and perform weighted fusion of 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 weight, and the spatial attention weight, so as to obtain the frequency domain intradomain fusion feature. This frequency domain intradomain fusion feature can more effectively eliminate the redundant parts of the EEG training data in the frequency domain.
[0139] D2. Based on the spectral coding representation, perform time-frequency domain feature fusion on the time-series coding representation to obtain time-frequency domain fused features;
[0140] Further, step D2, which involves fusing time-frequency domain features on the time-series coding representation based on the spectral coding representation to obtain time-frequency domain fused features, includes:
[0141] D21. Perform a third attention analysis on the spectral coding representation to obtain the frequency domain attention weights;
[0142] D22. Perform a fourth attention analysis on the temporal coding representation to obtain the temporal attention weights;
[0143] D23. Based on the frequency domain attention weight and the time domain attention weight, perform a second weighted fusion on the spectral coding representation and the temporal coding representation to obtain the time-frequency domain fusion feature.
[0144] In this embodiment, steps D21 to D23 are similar to those of steps D11 to D13 described above, and can be easily deduced by analogy. Steps D11 to D13 perform feature fusion within the frequency domain and feature fusion within the time domain, while steps D21 to D23 perform feature fusion between the frequency and time domains. These steps D21 to D23 can effectively eliminate redundant parts of the EEG training data between the frequency and time domains, which will not be elaborated further in this application.
[0145] D3. Based on the frequency domain intra-domain fusion feature, the time domain intra-domain fusion feature, and the time-frequency domain inter-domain fusion feature, the dual-domain fusion feature is obtained.
[0146] In this embodiment of the application, the frequency domain intra-domain fusion features, the time domain intra-domain fusion features, and the time-frequency domain inter-fusion features can be integrated to obtain the dual-domain fusion features corresponding to the EEG training data.
[0147] Step 160: Update the parameters of the initialized depression detection model based on the dual-domain fusion features to obtain a trained depression detection model.
[0148] In this embodiment, step 160 may first involve constructing a classifier consisting of three cascaded residual blocks and a multilayer perceptron (MLP). Then, the dual-domain fusion feature is input into the classifier for prediction and classification, thereby obtaining a depression detection result corresponding to the EEG training data. Next, the accuracy of the depression detection model is evaluated based on the depression detection result and the training labels corresponding to the EEG training data, thereby updating the parameters of the depression detection model.
[0149] Specifically, for machine learning models, the accuracy of model predictions can be measured by a loss function. The loss function is defined on a single training data point and measures the prediction error of that data point. Specifically, the loss value is determined by the label of the individual training data point and the model's prediction result for that data point. However, in actual training, a training dataset contains many data points. Therefore, a cost function is generally used to measure the overall error of the training dataset. The cost function is defined on the entire training dataset and calculates the average prediction error of all training data points, providing a better measure of the model's prediction performance. For general machine learning models, the aforementioned cost function, plus a regularization term to measure model complexity, serves as the training objective function. Based on this objective function, the loss value of the entire training dataset can be calculated. Many types of loss functions are commonly used, such as 0-1 loss, squared loss, absolute loss, logarithmic loss, and cross-entropy loss, which will not be elaborated upon here. In this embodiment, any one of these loss functions can be selected to determine the training loss value, such as the cross-entropy loss function. Based on the training loss value, the backpropagation algorithm is used to update the model parameters. After several iterations, a well-trained depression detection model can be obtained. The specific number of iterations can be preset, or training can be considered complete when the accuracy requirement on the test set is met.
[0150] This application also provides a detection method for a depression detection model, including...
[0151] Step 210: Acquire EEG data to be tested;
[0152] Step 220: Input the EEG data to be detected into the trained depression detection model to obtain the depression detection result output by the trained depression detection model.
[0153] In this embodiment of the application, EEG data to be tested can be acquired and input into a trained depression detection model. The trained depression detection model can then be used to detect depression in the EEG data to obtain a depression detection result.
[0154] The following describes in detail, with reference to the accompanying drawings, a training system for a depression detection model proposed according to an embodiment of this application.
[0155] Referring to Figure 3, a training system for a depression detection model proposed in this embodiment includes:
[0156] The first processing unit 101 is used to acquire EEG training data;
[0157] The second processing unit 102 is used to extract spectral features from the EEG training data to obtain a spectral feature set, wherein the spectral feature set includes several intermediate spectral features, and each intermediate spectral feature has a different spectral frequency band.
[0158] The third processing unit 103 is used to extract temporal features from the EEG training data to obtain a temporal feature set. The temporal feature set includes several cascaded intermediate temporal features, and the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature.
[0159] The fourth processing unit 104 is used to perform spectral surrogate graph learning on the spectral feature set to obtain a spectral coded representation, and to perform temporal surrogate graph learning on the temporal feature set to obtain a temporal coded representation;
[0160] The fifth processing unit 105 is used to perform dual-domain feature fusion on the temporal coding representation according to the spectral coding representation to obtain dual-domain fused features;
[0161] The sixth processing unit 106 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.
[0162] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0163] Referring to Figure 4, this application embodiment also provides an electronic device, including:
[0164] At least one processor 201;
[0165] At least one memory 202 is used to store at least one program;
[0166] When the at least one program is executed by the at least one processor 201, the at least one processor 201 implements the method embodiment described above.
[0167] Similarly, it can be understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0168] This application also provides a computer-readable storage medium storing a program executable by a processor 201, which, when executed by the processor 201, is used to implement the above-described method embodiments.
[0169] Similarly, the content of the above method embodiments is applicable to the present computer-readable storage medium embodiments. The specific functions implemented by the present computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0170] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.
[0171] Furthermore, although this application is 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 into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.
[0172] If a function is implemented as 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 this application, in essence, or the part that contributes to the prior art, or a part of the 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0173] 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 embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0174] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0175] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0176] In the foregoing description of this specification, the references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0177] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.
[0178] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.
Claims
1. A training method for a depression detection model, characterized in that, include: Acquire EEG training data; The EEG training data is subjected to spectral feature extraction to obtain a spectral feature set, which includes several intermediate spectral features, each of which has a different spectral frequency band. Temporal features are extracted from the EEG training data to obtain a temporal feature set, which includes several cascaded intermediate temporal features, wherein the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature. Spectrum surrogate graph learning is performed on the spectral feature set to obtain a spectral coded representation, and temporal surrogate graph learning is performed on the temporal feature set to obtain a temporal coded representation; Based on the spectral coding representation, the temporal coding representation is subjected to dual-domain feature fusion to obtain dual-domain fused features; Based on 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 step of extracting spectral features from the EEG training data to obtain a spectral feature set includes: The EEG training data is frequency domain converted to obtain several EEG frequency band data, each of which has a different frequency band. Differential entropy feature extraction is performed on all the EEG frequency band data to obtain intermediate spectral features corresponding to each EEG frequency band data; The spectral feature set is obtained based on all the intermediate spectral features; The step of extracting temporal features from the EEG training data to obtain a temporal feature set includes: The EEG training data is input into several cascaded temporal convolutional structures for feature extraction, thereby obtaining the intermediate temporal features output by each temporal convolutional structure. The time series feature set is obtained based on all the intermediate time series features.
3. The method according to claim 1, characterized in that, Target proxy graph learning is performed on the target feature set to obtain the first target encoding representation, including: Target proxy graphs are generated from the target feature set to obtain target function proxy graphs and target space proxy graphs corresponding to the target feature set. Based on the target function proxy graph, target graph learning is performed on the target space proxy graph to obtain the first target encoding representation; The target feature set is either a spectral feature set or a temporal feature set.
4. The method according to claim 3, characterized in that, The step of generating a target proxy map from the target feature set to obtain a target function proxy map and a target space proxy map corresponding to the target feature set includes: Obtain preset functional graph construction rules and spatial graph construction rules, as well as the original node data of the target feature set in each brain region; All the original node data are subjected to proxy node adaptive generation processing to obtain a number of proxy node data, each of the proxy node data corresponding to a brain region; According to the functional graph construction rules, a functional topology graph is constructed on the original node data and the proxy node data to obtain the target functional proxy graph; According to the spatial graph construction rules, a spatial topology graph is constructed from 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 spatial proxy graph based on the target functional proxy graph to obtain the first target encoding representation includes: The target function proxy graph is subjected to a first multi-level concatenated graph convolution process to obtain the target function feature representation; The target space proxy graph is subjected to a second multi-layer concatenated graph convolution process to obtain the target space feature representation; The first target encoding representation is obtained based on the target functional feature representation and the target spatial feature representation.
6. The method according to claim 1, characterized in that, The step of fusing dual-domain features on the temporal coding representation based on the spectral coding representation to obtain dual-domain fused features includes: The frequency domain feature fusion is performed on the spectral coding representation to obtain the frequency domain fused feature, and the time domain feature fusion is performed on the temporal coding representation to obtain the time domain fused feature. Based on the spectral coding representation, the time-frequency domain feature fusion is performed on the time-series coding representation to obtain the time-frequency domain fused feature; The dual-domain fusion feature is obtained based on the frequency domain intra-domain fusion feature, the time domain intra-domain fusion feature, and the time-frequency domain inter-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 fused features, including: Obtain the target functional feature representation and the target spatial feature representation from the second target coding representation, wherein the second target coding representation is the spectral coding representation or the temporal coding representation; A first attention analysis is performed on the target functional feature representation to obtain the functional attention weights of the target functional feature representation, and a second attention analysis is performed on the target spatial feature representation to obtain the spatial attention weights of the target spatial feature representation; Based on the functional attention weight and the spatial attention weight, the target functional feature representation and the target spatial feature representation are subjected to a first weighted fusion to obtain the target domain intradomain fusion feature.
8. The method according to claim 6, characterized in that, The step of fusing time-frequency domain features on the time-series coding representation according to the spectral coding representation to obtain time-frequency domain fused features includes: A third attention analysis is performed on the spectral encoding representation to obtain the frequency domain attention weights; A fourth attention analysis is performed on the temporal coding representation to obtain temporal attention weights; Based on the frequency domain attention weight and the time domain attention weight, the spectral coding representation and the temporal coding representation are subjected to a second weighted fusion to obtain the time-frequency domain fusion feature.
9. A detection method for a depression detection model, characterized in that, include: Acquire 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 1-8 to obtain the depression detection result output by the trained depression detection model.
10. A training system for a depression detection model, characterized in that, include: The first processing unit is used to acquire EEG training data; The second processing unit is used to extract spectral features from the EEG training data to obtain a spectral feature set, which includes several intermediate spectral features, each of which has a different spectral frequency band. The third processing unit is used to extract temporal features from the EEG training data to obtain a temporal feature set. The temporal feature set includes several cascaded intermediate temporal features, and the temporal receptive field of the later intermediate temporal feature is larger than that of the earlier intermediate temporal feature. The fourth processing unit is used to perform spectral surrogate graph learning on the spectral feature set to obtain a spectral coded representation, and to perform temporal surrogate graph learning on the temporal feature set to obtain a temporal coded representation; The fifth processing unit is used to perform dual-domain feature fusion on the temporal coding representation according to the spectral coding representation to obtain dual-domain fused features; The sixth processing unit is used to update the parameters of the initialized depression detection model based on the dual-domain fusion features to obtain a trained depression detection model.