Severe depression assessment model based on cloud edge collaboration and knowledge guided cross-contrast learning

Abstract

This invention discloses a severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparative learning, belonging to the field of intelligent medical technology. The model employs a cloud-edge collaborative architecture, acquiring fMRI data and demographic information at the edge, combining this with AAL and Harvard brain atlases to extract ROI time series, modeling using Pearson correlation coefficients, conducting inter-group t-tests for significance analysis, and constructing a functional connectivity graph through feature fusion before uploading it to the cloud. On the cloud, based on multi-view brain map hierarchical analysis, it integrates graph attention mechanisms, default mode network medical prior knowledge, and cross-comparative learning to complete the training and optimization of graph attention, subgraph generation, and the prediction network. Then, feature extraction, knowledge-guided pruning, and cross-view fusion are performed on the functional connectivity graph to be tested, and the depression assessment result is obtained through inference by the prediction network. This invention shortens transmission time, reduces bandwidth pressure, avoids privacy leakage risks, and improves assessment accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to a severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning. Background Technology

[0002] Major depressive disorder is a prevalent mental disorder worldwide, characterized by prolonged low mood and cognitive impairment. Currently, clinical assessment relies heavily on physicians' subjective experience and standardized scales (such as the PHQ-9). This approach is limited by the "social masking" mentality of relevant individuals and physicians' subjective biases, making it difficult to achieve accurate early warning.

[0003] Functional magnetic resonance imaging (fMRI), with its non-invasive nature and high spatial resolution, has become an important tool for the objective assessment of depression. Currently, existing techniques typically utilize brain mapping to divide the brain into multiple regions of interest (ROIs) and calculate functional connectivity based on blood oxygen level-dependent (BOLD) signals. Building upon this, researchers often employ deep learning models such as graph convolutional networks to extract topological features of brain networks for automated assessment.

[0004] However, in practical mobile healthcare and remote assessment scenarios, the aforementioned existing technologies still have some significant drawbacks. The challenges and long transmission times of high-dimensional, massive data streams: fMRI data is a high-dimensional spatiotemporal voxel stream with an extremely large data volume. In existing centralized cloud diagnostic architectures, the raw image data needs to be uploaded completely to the cloud, resulting in extremely high bandwidth consumption and transmission latency, severely limiting system response speed and deployment at edge nodes (such as small medical institutions); Patient privacy risks: fMRI data contains highly sensitive personal biometric features. If the raw data is transmitted over an open network, there is a significant risk of privacy breach if the transmission link is intercepted; Semantic corruption and incompleteness caused by video compression: Existing multimedia compression standards (such as H.264 / H.265), when applied to high-precision medical image data, suffer from lossy compression mechanisms that cause subtle, critical pathological features to be filtered out or destroyed as noise before transmission, resulting in data incompleteness and making accurate assessment impossible in the cloud. Summary of the Invention

[0005] The purpose of this invention is to provide a severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning, which solves the problems of high bandwidth load caused by the large amount of fMRI image data, privacy leakage risk caused by raw data transmission, and loss of diagnostic semantics caused by traditional compression in remote assessment scenarios.

[0006] To achieve the above objectives, this invention provides a severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning, comprising the following steps: S1. Acquire functional magnetic resonance imaging (fMRI) data at the edge and extract demographic information to use as sample data. Based on the AAL and Harvard brain maps, segment the blood oxygen level-dependent BOLD signal time series and extract the time series data of each region of interest (ROI). S2. The functional connectivity strength between different ROIs within the two brain maps was calculated at the edge end, and the functional connectivity matrix was constructed using the Pearson correlation coefficient. The significance analysis was performed using the t-test results between the healthy control group and the major depression group. The p-values ​​obtained from the t-test were used to identify and extract key functional connectivity edges with significant differences, and redundant features were removed and data dimensionality was reduced to serve as key functional connectivity features. The key functional connectivity features were fused with demographic information to construct a functional connectivity graph, which was then uploaded to the cloud. S3. Cloud-based multi-view brain map hierarchical analysis of functional connectivity graphs uploaded from the edge based on brain atlas: combining graph attention mechanism, medical prior knowledge and cross-comparison learning to complete the training and optimization of graph attention network, subgraph generation network and prediction network; S4. Based on the functional connectivity matrix of the healthy control group and the severe depression group and the trained graph attention network, the cloud extracts features from the functional connectivity graph of the dataset to be tested uploaded from the edge to obtain the pathological representation map to be tested; based on the trained graph attention network and subgraph generation network, the multi-view brain map in the pathological representation map to be tested is encoded and knowledge-guided pruning is performed to obtain the global semantic relationship features and domain knowledge subgraph features to be tested. S5. The cloud performs cross-perspective fusion of the global semantic relationship features and domain knowledge subgraph features of the AAL and Harvard brain map to be tested, and obtains the high-order embedding representation to be tested. S6. Input the high-order embedding representation to be tested into the trained prediction network, and obtain the depression assessment results of the dataset to be tested through cloud inference.

[0007] Preferably, in S1, demographic information includes gender, age, and education level.

[0008] Preferably, S2 specifically includes the following: S21. Based on AAL and Harvard brain maps, the functional connectivity values ​​between different ROI pairs are calculated using the BOLD sequence of each ROI and the Pearson correlation coefficient to construct an initial functional connectivity matrix, where each brain map is defined as a node of the brain functional connectivity map. S22. Perform a t-test between groups on the functional connectivity matrices of the healthy control group and the major depressive disorder group to identify functional connections that are significantly different between the two groups; calculate the t-statistic and p-value for each connection edge in the functional connectivity matrix to assess the significance level of each connection edge in distinguishing between healthy individuals and major depressive disorder. S23. Use the p-value obtained by the intergroup t-test to perform topological screening on the initial functional connectivity matrix. When the p-value between two brain regions is less than the set threshold, it is determined that there is an edge between the two brain regions in the brain map, that is, it is considered that there is a potential abnormal functional connection between the two brain regions. Thus, node edges are obtained, and a sparse brain functional connectivity network reflecting the potential abnormal connection of severe depression is constructed. S24. Encode the demographic information corresponding to the severe depression group data to obtain the pathological information embedding representation; in each brain region, combine its correlation coefficient with the pathological information embedding representation to obtain node features; S25. Based on node features, brain functional connectivity graph nodes and node edges, obtain the functional connectivity graph.

[0009] Preferably, in S3, the specific content is as follows: S31. The multi-head attention mechanism of graph attention network is used to adaptively weight the functional connectivity strength of different brain regions to highlight key brain regions and their interrelationships; based on the obtained attention weighting features, deep graph topology information is extracted to obtain global semantic relationship features. S32. Obtain the functional connectivity map under the brain map transmitted at the edge end, and input it as multi-view graph data into the multi-layer graph attention network to obtain the encoded brain functional map. S33. Obtain key subgraphs using prior knowledge of the default mode network: Based on medical prior knowledge, predetermine the list of core brain regions belonging to the default mode network, then construct the corresponding mask matrix based on the position of the core brain regions in the brain map, multiply the mask matrix with the features encoded by the graph attention network, and use the topology pruning method to filter out connections with low correlation to depression, obtain key subgraph views related to major depression, and encode them through the subgraph generation network to obtain knowledge-guided representation features; S34. Use pooling operations to reduce the dimensionality of the knowledge-guided representation features, transform the graph structure features into a fixed-length feature vector, and obtain the knowledge-guided representation vector representing the perspective of the brain map. S35. Based on the contrastive learning mechanism, minimize the feature differences between different perspectives: Use the projection head to form a positive sample pair with the knowledge-guided representation vector constructed based on the AAL map and the knowledge-guided representation vector constructed based on the Harvard map, minimize the semantic differences in the key sub-map representations under different brain map views, realize the minimization of feature differences under different perspective divisions, and obtain cross-perspective contrastive loss. S36. Concatenate the features of the two knowledge-guided representation vectors in S35 and input them into the prediction network to obtain the sample depression representation vector; calculate the cross-entropy loss between the depression representation vector and the true label as the depression assessment classification loss; combine the depression assessment classification loss with the cross-view comparison loss to obtain the target loss function value; based on the target loss function value, perform collaborative optimization and update of the graph attention network, subgraph generation network, prediction network and projection head; S37. Repeat S31-S36 to obtain the trained graph attention network, subgraph generation network, and prediction network.

[0010] Preferably, the functional connection diagram is represented as follows: , of which Brain regions in a brain map are used as node sets , This represents the number of brain regions in the atlas; node features are represented as... ,in, This is a brain functional connectivity matrix. The information is encoded as a vector; the graph attention network adaptively updates the node representation by the relative position and connection strength between nodes.

[0011] Preferably, in S31, the graph attention network includes a multi-head attention layer: Single-head attention coefficient ; This leads to the normalized weights. ; In the formula, Represented as nodes eigenvectors, The weight matrix is ​​a learnable linear transformation. For attention weight vectors, This indicates a splicing operation. Represented as nodes The set of neighbors; the updated features of a node are the weighted sum of the features of all its neighbors; In S33, the subgraph generation network includes a topology pruning layer and a pooling layer: The topology pruning layer utilizes domain knowledge to filter redundant information, and its expression is: ; In the formula, This is the complete brain function map feature matrix encoded by the graph attention network. This is a binary mask matrix constructed based on prior knowledge of the default mode network. This indicates an element-wise multiplication operation.

[0012] Preferably, in S34, the pooling layer uses global average pooling to transform the graph features into vectors: ; In the formula: Embedding of nodes in the pruned subgraph The feature matrix of the subgraph nodes. To guide the representation vector of knowledge representing this perspective; In S35, the cross-view contrast loss function is as follows: Cross-view comparison loss Enhance feature robustness by minimizing the distance between positive sample pairs: ; In the formula, and These represent the feature vectors of the same individual from the perspectives of the Harvard graph and the AAL graph, respectively. The number of samples in a training batch; This represents the temperature coefficient, which is a hyperparameter.

[0013] Preferably, in S36, the expression for the depression assessment classification loss is: ; In the formula: The number of training samples, Indicates the first The true major depressive disorder assessment results of a sample Indicates the first The predicted probability of major depression for each sample; The target loss function value is: ; In the formula, The ratio of the loss function is represented by a hyperparameter.

[0014] Therefore, this invention employs the aforementioned severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning. By introducing the multi-head attention mechanism of graph attention networks, it can adaptively capture complex relational features between brain regions and accurately extract discriminative information. Combined with the default mode network performing knowledge-guided topology pruning, it effectively filters background noise and accurately identifies core biomarkers related to depression. Employing an edge-cloud collaborative architecture, it transforms high-dimensional fMRI raw image streams into lightweight functional connectivity graphs through edge semantic modeling, achieving data compression, reducing network bandwidth pressure in remote assessment, and solving the bottleneck problem of ultra-large-scale multimedia data in network transmission. Following the principle of data minimization, sensitive raw data is retained at the edge, while the cloud only receives abstract graph features without raw pixel details, effectively reducing the temporal attack surface and avoiding the risk of leakage of biometric information at the source.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is an experimental result graph showing the learning rate, batch size, classification loss, and loss ratio hyperparameters in an embodiment of the present invention. Figure 3 This is a diagram showing the experimental results of transmission efficiency in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] Example 1 This invention provides a severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning, the process of which is as follows: Figure 1As shown, it includes the following steps: S1. Acquire functional magnetic resonance imaging (fMRI) data at the edge and extract demographic information (gender, age, education level) as sample data. Based on two brain atlases, AAL and Harvard, segment the time series of blood oxygen level-dependent BOLD signals and extract the time series data of each region of interest (ROI).

[0020] S2. Functional connectivity strengths between different ROIs within the two brain atlases were calculated at the edge points, and a functional connectivity matrix was constructed using the Pearson correlation coefficient. Significance analysis was performed using the t-test results between the healthy control group and the major depressive disorder group. Key functional connectivity edges with significant differences were identified and extracted using the p-values ​​obtained from the t-test, and redundant features were removed and data dimensionality reduced to serve as key functional connectivity features. These key functional connectivity features were then fused with demographic information to construct a functional connectivity graph, which was uploaded to the cloud. Specifically, this includes the following: S21. Based on AAL and Harvard brain atlases, the functional connectivity values ​​between different ROI pairs are calculated using the BOLD sequences of each ROI and the Pearson correlation coefficient to construct an initial functional connectivity matrix, where each brain atlas is defined as a node of the brain functional connectivity map.

[0021] S22. Perform a t-test between groups on the functional connectivity matrices of the healthy control group and the major depressive disorder group to identify functional connections that differ significantly between the two groups; calculate the t-statistic and p-value for each connection edge in the functional connectivity matrix to assess the significance level of each connection edge in distinguishing between healthy individuals and major depressive disorder.

[0022] S23. The p-value obtained by the inter-group t-test is used to perform topological screening on the initial functional connectivity matrix. When the p-value is less than the set threshold, which is set to 0.05 in this embodiment, it is determined that there is an edge between two brain regions in this brain map. That is, it is considered that there is a potential abnormal functional connection between these two brain regions in major depression. Thus, node edges are obtained, and a sparse brain functional connectivity network reflecting the potential abnormal connection of major depression is constructed.

[0023] S24. Encode the demographic information corresponding to the severe depression group data to obtain the pathological information embedding representation; in each brain region, combine its correlation coefficient with the pathological information embedding representation to obtain node features.

[0024] S25. Based on node features, brain functional connectivity graph nodes and node edges, obtain the functional connectivity graph.

[0025] The functional connection diagram is represented as follows , of which Brain regions in a brain map are used as node sets , This represents the number of brain regions in the atlas; node features are represented as... ,in, This is a brain functional connectivity matrix. The information is encoded as a vector; the graph attention network adaptively updates the node representation by the relative position and connection strength between nodes.

[0026] S3. Cloud-based multi-view hierarchical analysis of functional connectivity graphs uploaded from the edge terminal, constructed using a brain atlas: Combining graph attention mechanisms, medical prior knowledge, and cross-comparison learning, the training and optimization of the graph attention network, subgraph generation network, and prediction network are completed. Specific details are as follows: S31. The multi-head attention mechanism of graph attention network is used to adaptively weight the functional connectivity strength of different brain regions to highlight key brain regions and their interrelationships; based on the obtained attention weighted features, deep graph topology information is extracted to obtain global semantic relationship features.

[0027] Graph attention networks include multi-head attention layers: Single-head attention coefficient ; This leads to the normalized weights. ; In the formula, Represented as nodes eigenvectors, The weight matrix is ​​a learnable linear transformation. For attention weight vectors, This indicates a splicing operation. Represented as nodes The set of neighbors; the updated features of a node are the weighted sum of the features of all its neighbors; S32. Obtain the functional connectivity map under the brain map transmitted at the edge end, and input it as multi-view graph data into the multi-layer graph attention network to obtain the encoded brain functional map.

[0028] S33. Obtain key subgraphs using prior knowledge of the default mode network: Based on medical prior knowledge, predetermine the list of core brain regions belonging to the default mode network, and then construct the corresponding mask matrix according to the position of the core brain regions in the brain map. Multiply the mask matrix with the features encoded by the graph attention network, and use the topology pruning method to filter out connections with low correlation to depression, obtain key subgraph views related to major depression, and encode them through the subgraph generation network to obtain knowledge-guided representation features.

[0029] Subgraph generation networks include topology pruning layers and pooling layers: The topology pruning layer utilizes domain knowledge to filter redundant information, and its expression is: ; In the formula, This is the complete brain function map feature matrix encoded by the graph attention network. This is a binary mask matrix constructed based on prior knowledge of the default mode network. This indicates an element-wise multiplication operation.

[0030] S34. Use pooling operations to reduce the dimensionality of the knowledge-guided representation features, transform the graph structure features into a fixed-length feature vector, and obtain the knowledge-guided representation vector representing the perspective of the brain map.

[0031] The pooling layer uses global average pooling to transform graph features into vectors: ; In the formula: Embedding of nodes in the pruned subgraph The feature matrix of the subgraph nodes. The knowledge-guided representation vector represents this perspective.

[0032] S35. Based on the contrastive learning mechanism, minimize the feature differences between different perspectives: Use the projection head to form a positive sample pair with the knowledge-guided representation vector constructed based on the AAL map and the knowledge-guided representation vector constructed based on the Harvard map, minimize the semantic differences in the representation of key subgraphs under different brain map views, achieve the minimization of feature differences under different perspective divisions, and obtain cross-view contrastive loss.

[0033] The cross-view comparison loss function is as follows: Cross-view comparison loss Enhance feature robustness by minimizing the distance between positive sample pairs: ; In the formula, and These represent the feature vectors of the same individual from the perspectives of the Harvard graph and the AAL graph, respectively. The number of samples in a training batch; This represents the temperature coefficient, which is a hyperparameter.

[0034] S36. Concatenate the features of the two knowledge-guided representation vectors in S35 and input them into the prediction network to obtain the sample depression representation vector; calculate the cross-entropy loss between the depression representation vector and the true label as the depression assessment classification loss; combine the depression assessment classification loss with the cross-view comparison loss to obtain the target loss function value; based on the target loss function value, perform collaborative optimization and update of the graph attention network, subgraph generation network, prediction network and projection head.

[0035] The expression for the depression assessment classification loss is: ; In the formula: The number of training samples, Indicates the first The true major depressive disorder assessment results of a sample Indicates the first The predicted probability of major depression for each sample; The target loss function value is: ; In the formula, The ratio of the loss function is represented by a hyperparameter.

[0036] S37. Repeat S31-S36 to obtain the trained graph attention network, subgraph generation network, and prediction network.

[0037] S4. Based on the functional connectivity matrix of the healthy control group and the severe depression group, and the trained graph attention network, the cloud extracts features from the functional connectivity graph of the dataset to be tested uploaded from the edge to obtain the pathological representation map to be tested; based on the trained graph attention network and subgraph generation network, the multi-view brain map in the pathological representation map to be tested is encoded and knowledge-guided pruning is performed to obtain the global semantic relationship features and domain knowledge subgraph features to be tested.

[0038] S5. The cloud performs cross-perspective fusion of the global semantic relationship features and domain knowledge subgraph features of the AAL and Harvard brain map to be tested, and obtains the high-order embedding representation to be tested.

[0039] S6. Input the high-order embedding representation to be tested into the trained prediction network, and obtain the depression assessment results of the dataset to be tested through cloud inference.

[0040] This embodiment compares the proposed method (Co-Brain) with existing methods. Five-fold cross-validation was employed, and performance was evaluated using accuracy (Acc), sensitivity (Sen), specificity (Spec), F1 score (F1), and transmission time. Accuracy (Acc) represents the proportion of correctly predicted samples out of the total sample count; a higher value is better. Sensitivity (Sensitivity, also known as recall), represents the proportion of samples that are actually positive but are correctly predicted as positive, measuring the model's ability to identify positive samples; a higher value is better. Specificity represents the proportion of samples that are actually negative but are correctly predicted as negative, reflecting the model's ability to distinguish negative samples; a higher value is better. The F1 score is the harmonic mean of accuracy and recall, used to comprehensively evaluate the model's balanced performance on both positive and negative classes; a higher value is better. Transmission time refers to the time consumed for the raw data to be transmitted from the initiator to the cloud processing unit to obtain the results. The results are shown in Table 1.

[0041] Table 1. Experimental Results of Different Methods

[0042] Table 1 presents the experimental results of the proposed method and other comparative methods across multiple performance metrics. As shown in Table 1, the proposed method demonstrates the most outstanding performance in sensitivity and transmission time, both surpassing existing methods. Regarding transmission efficiency, this metric measures the communication time from the initiator to the cloud processing unit to obtain the results: due to the cloud-edge collaborative framework employed in this method, effective feature compression of the data at the edge reduces the transmission time to between 21.47 ms and 25.90 ms, significantly shorter than other models, effectively optimizing the communication burden. Although it did not surpass all comparative methods in some metrics such as specificity, the results for each metric are satisfactory. Ablation experiments show that Co-Brain-GAT outperforms Co-Brain-GCN (with attention mechanism removed) in multiple metrics. Although the specificity of Co-Brain-GAT is lower than that of the ablation model Co-Brain-GCN, its significant gain in sensitivity greatly reduces the risk of omission, making it more valuable for clinical primary screening. The comprehensive experimental results demonstrate that this method significantly optimizes data transmission efficiency in remote scenarios while addressing the challenges of limited samples and improving diagnostic sensitivity.

[0043] Figure 2 These are the results of experiments conducted in this embodiment with different learning rates, batch sizes, and loss ratios. For example... Figure 2 As shown in Figure a, when the batch size increases from 16 to 32, both the accuracy and F1 score of the model show an upward trend, reaching a peak at 32. When the batch size is further increased to 64 and 128, the model performance decreases significantly. Figure 2 The horizontal axis of b represents the denominator (numerator is 1) of the cross-view contrast loss ratio. The line shows that as the value increases, the model performance first increases and then decreases, with the best performance at 7. Figure 2 In the experiment with the hyperparameter c representing the learning rate, the model performance rapidly increased after starting with a very small learning rate (0.00001), reaching its peak at 0.0003. Afterward, as the learning rate continued to increase, the model performance declined sharply. The results from the three experiments show that the optimal batch size, the optimal loss ratio, and the optimal learning rate for this method are 32 and 1 / 7, respectively.

[0044] Figure 3 This is the compression ratio achieved in this embodiment. By comparing the original data size and the graph data size transmitted to the cloud, this method achieves a compression ratio of 10.9 times, effectively solving the existing problems of difficult and long transmission times.

[0045] Therefore, this invention employs the aforementioned severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning. By introducing the multi-head attention mechanism of graph attention networks, it can adaptively capture complex relational features between brain regions and accurately extract discriminative information. Combined with the default mode network performing knowledge-guided topology pruning, it effectively filters background noise and accurately identifies core biomarkers related to depression. Employing an edge-cloud collaborative architecture, it transforms high-dimensional fMRI raw image streams into lightweight functional connectivity graphs through edge semantic modeling, achieving data compression, reducing network bandwidth pressure in remote assessment, and solving the bottleneck problem of ultra-large-scale multimedia data in network transmission. Following the principle of data minimization, sensitive raw data is retained at the edge, while the cloud only receives abstract graph features without raw pixel details, effectively reducing the temporal attack surface and avoiding the risk of leakage of biometric information at the source.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparative learning, characterized in that, Includes the following steps: S1. Acquire functional magnetic resonance imaging (fMRI) data at the edge and extract demographic information to use as sample data. Based on the AAL and Harvard brain maps, segment the blood oxygen level-dependent BOLD signal time series and extract the time series data of each region of interest (ROI). S2. The functional connectivity strength between different ROIs within the two brain maps was calculated at the edge end, and the functional connectivity matrix was constructed using the Pearson correlation coefficient. The significance analysis was performed using the t-test results between the healthy control group and the major depression group. The p-values ​​obtained from the t-test were used to identify and extract key functional connectivity edges with significant differences, and redundant features were removed and data dimensionality was reduced to serve as key functional connectivity features. The key functional connectivity features were fused with demographic information to construct a functional connectivity graph, which was then uploaded to the cloud. S3. Cloud-based multi-view brain map hierarchical analysis of functional connectivity graphs uploaded from the edge based on brain atlas: combining graph attention mechanism, medical prior knowledge and cross-comparison learning to complete the training and optimization of graph attention network, subgraph generation network and prediction network; S4. Based on the functional connectivity matrix of the healthy control group and the severe depression group and the trained graph attention network, the cloud extracts features from the functional connectivity graph of the dataset to be tested uploaded from the edge to obtain the pathological representation map to be tested; based on the trained graph attention network and subgraph generation network, the multi-view brain map in the pathological representation map to be tested is encoded and knowledge-guided pruning is performed to obtain the global semantic relationship features and domain knowledge subgraph features to be tested. S5. The cloud performs cross-perspective fusion of the global semantic relationship features and domain knowledge subgraph features of the AAL and Harvard brain map to be tested, and obtains the high-order embedding representation to be tested. S6. Input the high-order embedding representation to be tested into the trained prediction network, and obtain the depression assessment results of the dataset to be tested through cloud inference.

2. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning as described in claim 1, characterized in that, In S1, demographic information includes gender, age, and education level.

3. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning according to claim 2, characterized in that, S2 specifically includes the following: S21. Based on AAL and Harvard brain maps, the functional connectivity values ​​between different ROI pairs are calculated using the BOLD sequence of each ROI and the Pearson correlation coefficient to construct an initial functional connectivity matrix, where each brain map is defined as a node of the brain functional connectivity map. S22. Perform an intergroup t-test on the functional connectivity matrices of the healthy control group and the major depressive disorder group to identify functional connectivity that differs significantly between the two groups. For each connection edge in the functional connectivity matrix, the t-statistic and p-value of the two groups were calculated to evaluate the significance level of each connection edge in distinguishing between health and major depressive disorder. S23. Use the p-value obtained by the intergroup t-test to perform topological screening on the initial functional connectivity matrix. When the p-value between two brain regions is less than the set threshold, it is determined that there is an edge between the two brain regions in the brain map, that is, it is considered that there is a potential abnormal functional connection between the two brain regions. Thus, node edges are obtained, and a sparse brain functional connectivity network reflecting the potential abnormal connection of severe depression is constructed. S24. Encode the demographic information corresponding to the severe depression group data to obtain the pathological information embedding representation; in each brain region, combine its correlation coefficient with the pathological information embedding representation to obtain node features; S25. Based on node features, brain functional connectivity graph nodes and node edges, obtain the functional connectivity graph.

4. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning as described in claim 3, characterized in that, In S3, the specific content is as follows: S31. The multi-head attention mechanism of graph attention network is used to adaptively weight the functional connectivity strength of different brain regions to highlight key brain regions and their interrelationships; based on the obtained attention weighting features, deep graph topology information is extracted to obtain global semantic relationship features. S32. Obtain the functional connectivity map under the brain map transmitted at the edge end, and input it as multi-view graph data into the multi-layer graph attention network to obtain the encoded brain functional map. S33. Obtain key subgraphs using prior knowledge of the default mode network: Based on medical prior knowledge, predetermine the list of core brain regions belonging to the default mode network, then construct the corresponding mask matrix based on the position of the core brain regions in the brain map, multiply the mask matrix with the features encoded by the graph attention network, and use the topology pruning method to filter out connections with low correlation to depression, obtain key subgraph views related to major depression, and encode them through the subgraph generation network to obtain knowledge-guided representation features; S34. Use pooling operations to reduce the dimensionality of the knowledge-guided representation features, transform the graph structure features into a fixed-length feature vector, and obtain the knowledge-guided representation vector representing the perspective of the brain map. S35. Based on the contrastive learning mechanism, minimize the feature differences between different perspectives: Use the projection head to form a positive sample pair with the knowledge-guided representation vector constructed based on the AAL map and the knowledge-guided representation vector constructed based on the Harvard map, minimize the semantic differences in the key sub-map representations under different brain map views, realize the minimization of feature differences under different perspective divisions, and obtain cross-perspective contrastive loss. S36. Concatenate the features of the two knowledge-guided representation vectors in S35 and input them into the prediction network to obtain the sample depression representation vector; calculate the cross-entropy loss between the depression representation vector and the true label as the depression assessment classification loss; combine the depression assessment classification loss with the cross-view comparison loss to obtain the target loss function value; based on the target loss function value, perform collaborative optimization and update of the graph attention network, subgraph generation network, prediction network and projection head; S37. Repeat S31-S36 to obtain the trained graph attention network, subgraph generation network, and prediction network.

5. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning according to claim 4, characterized in that, The functional connection diagram is represented as follows: , of which Brain regions in a brain map are used as node sets , This represents the number of brain regions in the atlas; node features are represented as... ,in, This is a brain functional connectivity matrix. The information is encoded as a vector; the graph attention network adaptively updates the node representation by the relative position and connection strength between nodes.

6. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning according to claim 5, characterized in that, In S31, the graph attention network includes a multi-head attention layer: Single-head attention coefficient ; This leads to the normalized weights. ; In the formula, Represented as nodes eigenvectors, The weight matrix is ​​a learnable linear transformation. For attention weight vectors, This indicates a splicing operation. Represented as nodes The set of neighbors; the updated features of a node are the weighted sum of the features of all its neighbors; In S33, the subgraph generation network includes a topology pruning layer and a pooling layer: The topology pruning layer utilizes domain knowledge to filter redundant information, and its expression is: ； In the formula, This is the complete brain function map feature matrix encoded by the graph attention network. This is a binary mask matrix constructed based on prior knowledge of the default mode network. This indicates an element-wise multiplication operation.

7. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning according to claim 6, characterized in that, In S34, the pooling layer uses global average pooling to transform graph features into vectors: ； In the formula: Embedding of nodes in the pruned subgraph The feature matrix of the subgraph nodes. To guide the representation vector of knowledge representing this perspective; In S35, the cross-view contrast loss function is as follows: Cross-view comparison loss Enhance feature robustness by minimizing the distance between positive sample pairs: ； In the formula, and These represent the feature vectors of the same individual from the perspectives of the Harvard graph and the AAL graph, respectively. The number of samples in a training batch; This represents the temperature coefficient, which is a hyperparameter.

8. The severe depression assessment model based on cloud-edge collaboration and knowledge-guided cross-comparison learning according to claim 7, characterized in that, In S36, the expression for the depression assessment classification loss is: ； In the formula: The number of training samples, Indicates the first The true major depressive disorder assessment results of a sample Indicates the first The predicted probability of major depression for each sample; The target loss function value is: ； In the formula, The ratio of the loss function is represented by a hyperparameter.

Images