Method, system and device for realizing AFD phase inversion risk prediction based on function connection and graph neural network, processor and medium

Abstract

The invention relates to a method for realizing AFD phase inversion risk prediction based on function connection and a graph neural network, and the method comprises the following steps: preprocessing an image, extracting a time sequence of each brain region based on a predefined brain map, and calculating a whole brain FC matrix; constructing the FC matrix into graph structure data, and inputting the graph structure data into a graph neural network; the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient population is calculated. The method, the system, the device, the processor and the medium for realizing AFD phase inversion risk prediction based on the function connection and the graph neural network are high in prediction precision, combine GNN with an edge weight attention mechanism, can capture high-order topological characteristics of a brain network, have early warning capability, and can predict and output FC and brain regions which are most critical to decision. The technology fusion innovativeness is high, the brain connection omics, the graph neural network and representation learning are seamlessly fused, and the clinical transformation potential is large.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical artificial intelligence, in particular to the field of cross technology of neuroimaging analysis, and more particularly to a method, system, device, processor and medium for realizing AFD conversion risk prediction based on functional connectivity and graph neural network. BACKGROUND

[0002] It is a key and difficult point in clinical diagnosis and treatment to identify unipolar and bipolar depression in early stage. AFD is a subtype of MDD, and there is a high overlap in clinical symptoms with BD, especially bipolar II disorder. The risk of AFD converting to BD is significantly higher than that of ordinary MDD patients. It is extremely challenging to accurately identify patients who will soon convert to BD from AFD in clinical practice. AFD refers to atypical depression.

[0003] The prior art mainly relies on clinical scales (such as hypomanic symptom checklist) and clinical features (such as early onset and family history) for risk assessment. These methods are highly subjective and lack specificity, and lack objective biological markers. Some neuroimaging studies attempt to use traditional machine learning methods to analyze brain functional connectivity (FC) to distinguish MDD and BD, but these methods usually have the following limitations: feature extraction relies on artificial prior knowledge, and cannot capture high-order, complex nonlinear connection patterns; the sample is highly heterogeneous and weakly specific, and is not anchored to patients with high conversion risk with AFD; unable to effectively quantify the "similarity" of individual patients to the target disease at the neuroimaging level, thus failing to achieve personalized conversion risk and time course prediction; lack of dynamic risk prediction, existing methods mostly focus on disease classification, and cannot quantitatively predict the future disease evolution trajectory of a specific patient group. SUMMARY

[0004] The purpose of the present application is to overcome the above-mentioned shortcomings of the prior art, and to provide a method, system, device, processor and medium for realizing AFD conversion risk prediction based on functional connectivity and graph neural network, which meets the requirements of objectivity, accuracy and individualization.

[0005] In order to achieve the above-mentioned purpose, the method, system, device, processor and medium for realizing AFD conversion risk prediction based on functional connectivity and graph neural network of the present application are as follows: The method for realizing AFD conversion risk prediction based on functional connectivity and graph neural network, which mainly comprises the following steps: (1) Preprocess the image, extract the time series of each brain region based on the pre-defined brain atlas, and calculate the whole brain FC matrix; (2) Construct the FC matrix into a graph structure data and input it into a graph neural network; (3) Calculate the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient population.

[0006] Preferably, the step (2) specifically comprises the following steps: (2.1) Construct a functional connection graph; (2.2) Construct and train a graph neural network model; (2.3) Extract features and make risk prediction; (2.4) Identify key biomarkers.

[0007] Preferably, the step (2.1) specifically comprises the following steps: Define the whole brain FC matrix as a graph G = (V, E, X); Where the node set V corresponds to 90 brain regions, the weight e ij is obtained by converting the FC value r ij , the FC value r ij is the Pearson correlation coefficient of the time series of brain region i and brain region j, and the feature vector X of each node is the FC value of the node and all other brain regions, indicating the connection strength of the node and all other 89 nodes.

[0008] Preferably, the step (2.2) specifically comprises the following steps: (2.2.1) Input the constructed graph G; (2.2.2) Construct a graph neural network model; (2.2.3) Model training.

[0009] Preferably, the step (2.2.2) specifically comprises the following steps: (2.2.2.1) Based on the edge-weighted graph attention layer, the original FC weight e ij is integrated into the attention mechanism; (2.2.2.2) Based on the differentiable graph pooling layer, 90 nodes are aggregated into 32 nodes through a learnable allocation matrix S, and the graph structure is coarsened; (2.2.2.3) Repeat steps (2.2.2.1) and (2.2.2.2) to aggregate 32 nodes into 4 nodes; (2.2.2.4) Concatenate the features of the 4 nodes to form a graph-level feature vector; pass through two fully connected layers to output the final classification result.

[0010] Preferably, step (2.2.2.1) specifically includes the following steps: (2.2.2.1.1) Calculate the original attention coefficients between nodes; (2.2.2.1.2) Through edge weight e ij Modulate the attention coefficient; (2.2.2.1.3) The features of neighboring nodes are weighted and summed using the modulated attention coefficients and then passed through a nonlinear activation function σ to obtain new features of the nodes.

[0011] Preferably, step (2.2.3) specifically includes: Using cross-entropy loss L ent As the main loss function, an auxiliary link prediction loss L is introduced. aux Optimize the pooling process.

[0012] Preferably, step (2.3) specifically includes the following steps: (2.3.1) The trained GNN model is used as a feature extractor. The FC graph of the subject is input to extract a high-dimensional feature vector. (2.3.2) Calculate the Euclidean distance between the feature vector of each AFD patient and the average feature vector of all patients in the BD group, and establish a risk prediction model.

[0013] Preferably, step (2.3.2) further includes calculating the cosine similarity or Mahalanobis distance between the feature vector of each AFD patient and the average feature vector of all patients in the BD group.

[0014] Preferably, step (2.4) specifically includes the following steps: The attention matrix is ​​extracted from the trained EGAT layer, all weights are sorted, the top 100 FCs with the highest weights are selected, and the top 5 key brain regions are identified by accumulating the attention weights of the functional connections of each brain region.

[0015] The system for predicting AFD phase transition risk based on functional connectivity and graph neural networks is characterized by the following: The data processing module receives resting-state functional magnetic resonance imaging (fMRI) images of an individual at its input end. It is used to preprocess the images, extract time series of each brain region based on a predefined brain atlas, and calculate the whole-brain FC matrix. The graph neural network feature extraction module is connected to the data processing module and is used to construct the FC matrix into graph structure data and input it into a graph neural network. The similarity calculation and risk prediction module is connected with the graph neural network feature extraction module, and is used for calculating the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient population.

[0016] The device for realizing AFD conversion risk prediction based on functional connection and graph neural network mainly comprises: The processor is configured to execute computer executable instructions; The memory stores one or more computer executable instructions, and when the computer executable instructions are executed by the processor, the steps of the method for realizing AFD conversion risk prediction based on functional connection and graph neural network are realized.

[0017] The processor for realizing AFD conversion risk prediction based on functional connection and graph neural network mainly comprises:

[0018] The computer readable storage medium mainly comprises:

[0019] The method, system, device, processor and medium for realizing AFD conversion risk prediction based on functional connection and graph neural network have high prediction accuracy, and for the first time, GNN and edge weight attention mechanism are combined for this field, can capture high-order topological features of brain networks, and 75% accuracy is achieved in MDD and BD classification tasks, which is significantly better than traditional machine learning methods; have early warning ability, propose a quantitative index of "neural image similarity", and confirm that it is related to conversion time. This enables doctors to evaluate the BD conversion risk of AFD patients individually and quantitatively before the appearance of clinical symptoms, realizing the leap from "static diagnosis" to "dynamic prediction". The model of the application can not only predict, but also output the most critical FC and brain area for decision making. The application finds that the connections within the default mode network, the dorsal / ventral attention network and the visual network are key features, which have important scientific research value. The technical fusion of the application has strong innovation, seamlessly integrates brain connectomics, graph neural networks and representation learning, and provides a new and powerful technical framework for solving the heterogeneity and early prediction of mental illness. The clinical conversion potential of the application is great, and ultimately can be deployed as an auxiliary diagnosis software to provide objective decision support for psychiatrists, which helps to avoid misdiagnosis and mistreatment and improve long-term prognosis of patients. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A schematic diagram for constructing a graph neural network model of the method for realizing AFD phase conversion risk prediction based on functional connectivity and graph neural network of the application. DETAILED DESCRIPTION

[0021] In order to more clearly describe the technical content of the application, further description will be made below in combination with specific embodiments.

[0022] The method for realizing AFD phase conversion risk prediction based on functional connectivity and graph neural network of the application, wherein the method comprises the following steps: (1) pre-processing the image, extracting the time series of each brain region based on the pre-defined brain atlas, and calculating the whole brain FC matrix; (2) constructing the FC matrix into graph structure data and inputting into a graph neural network; (3) calculating the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient population.

[0023] As a preferred embodiment of the application, the step (2) specifically comprises the following steps: (2.1) constructing a functional connectivity graph; (2.2) constructing and training a graph neural network model; (2.3) extracting features and performing risk prediction; (2.4) identifying key biomarkers.

[0024] As a preferred embodiment of the application, the step (2.1) specifically comprises the following steps: The whole brain FC matrix is defined as a graph G=(V, E, X); Wherein, the node set V corresponds to 90 brain regions, the weight e ij The FC value r ij is converted to obtain the FC value r ij is the Pearson correlation coefficient of the time series of the brain region i and the brain region j, and the feature vector X of each node is the FC value of the node and all other brain regions, which indicates the connection strength of the node and all other 89 nodes.

[0025] As a preferred embodiment of the application, the step (2.2) specifically comprises the following steps: (2.2.1) inputting the constructed graph G; (2.2.2) constructing a graph neural network model; (2.2.3) model training.

[0026] In a preferred embodiment of the present invention, step (2.2.2) specifically includes the following steps: (2.2.2.1) Based on the edge-weighted graph attention layer, the original FC weights e ij Integrate into the attention mechanism; (2.2.2.2) Based on the differentiable graph pooling layer, 90 nodes are aggregated into 32 nodes through a learnable allocation matrix S to coarsen the graph structure; (2.2.2.3) Repeat steps (2.2.2.1) and (2.2.2.2) to aggregate the 32 nodes into 4 nodes; (2.2.2.4) The features of the four nodes are concatenated to form a graph-level feature vector; after passing through two fully connected layers, the final classification result is output.

[0027] In a preferred embodiment of the present invention, step (2.2.2.1) specifically includes the following steps: (2.2.2.1.1) Calculate the original attention coefficients between nodes; (2.2.2.1.2) Through edge weight e ij Modulate the attention coefficient; (2.2.2.1.3) The features of neighboring nodes are weighted and summed using the modulated attention coefficients and then passed through a nonlinear activation function σ to obtain new features of the nodes.

[0028] In a preferred embodiment of the present invention, step (2.2.3) specifically comprises: Using cross-entropy loss L ent As the main loss function, an auxiliary link prediction loss L is introduced. aux Optimize the pooling process.

[0029] In a preferred embodiment of the present invention, step (2.3) specifically includes the following steps: (2.3.1) The trained GNN model is used as a feature extractor. The FC graph of the subject is input to extract a high-dimensional feature vector. (2.3.2) Calculate the Euclidean distance between the feature vector of each AFD patient and the average feature vector of all patients in the BD group, and establish a risk prediction model.

[0030] As a preferred embodiment of the present invention, step (2.3.2) further includes calculating the cosine similarity or Mahalanobis distance between the feature vector of each AFD patient and the average feature vector of all patients in the BD group.

[0031] In a preferred embodiment of the present invention, step (2.4) specifically includes the following steps: The attention matrix is extracted from the trained EGAT layer, all weights are sorted, the top 100 FCs with the highest weights are selected, the attention weight of the functional connection where each brain region is located is accumulated, and the top 5 key brain regions are identified.

[0032] The system for realizing AFD phase conversion risk prediction based on functional connection and graph neural network, wherein the system comprises: A data processing module, an input end of which receives a resting state functional magnetic resonance image of an individual, is used for pre-processing the image, extracting the time series of each brain region based on a pre-defined brain atlas, and calculating a whole brain FC matrix; A graph neural network feature extraction module, connected with the data processing module, is used for constructing the FC matrix into graph structure data and inputting into a graph neural network; A similarity calculation and risk prediction module, connected with the graph neural network feature extraction module, is used for calculating the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient population.

[0033] The device for realizing AFD phase conversion risk prediction based on functional connection and graph neural network, wherein the device comprises: A processor configured to execute computer executable instructions; A memory storing one or more computer executable instructions, wherein the computer executable instructions are executed by the processor to realize the steps of the method for realizing AFD phase conversion risk prediction based on functional connection and graph neural network.

[0034] The processor for realizing AFD phase conversion risk prediction based on functional connection and graph neural network, wherein the processor is configured to execute computer executable instructions, and the computer executable instructions are executed by the processor to realize the steps of the method for realizing AFD phase conversion risk prediction based on functional connection and graph neural network.

[0035] The computer readable storage medium of the application, wherein the computer readable storage medium stores a computer program, and the computer program can be executed by a processor to realize the steps of the method for realizing AFD phase conversion risk prediction based on functional connection and graph neural network.

[0036] The application belongs to the technical field of medical artificial intelligence and neuroimaging analysis, and particularly relates to a method and system for constructing an objective biomarker model for early identification and prediction of the risk of conversion of major depressive disorder with atypical features (AFD) to bipolar disorder (BD) based on resting-state functional magnetic resonance imaging, graph neural networks and representation learning technology.

[0037] The application aims to solve the above-mentioned deficiencies in the prior art, and specifically aims to solve the technical problem of how to provide an objective and accurate personalized prediction method to early identify a subgroup of AFD patients with high conversion risk and quantify the risk and time course of conversion to BD.

[0038] 1. System composition and process The system of the application comprises a data processing module, a graph neural network feature extraction module and a similarity calculation and risk prediction module.

[0039] 1.1 Data processing module: the input is the resting-state functional magnetic resonance image of an individual. The module pre-processes the image and extracts the time series of each brain region based on a pre-defined brain atlas (such as AAL-90), and then calculates the whole brain FC matrix.

[0040] 1.2 Graph neural network feature extraction module: the FC matrix is constructed as graph structure data (brain regions as nodes and FC values as edge weights) and input into a graph neural network GNN. The graph neural network GNN is trained to effectively distinguish MDD and BD and extract high-dimensional and abstract feature representations from them. The feature condenses the information of the individual brain FC to reflect the individual brain disease state.

[0041] 1.3 Similarity calculation and risk prediction module: the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient population is calculated. The smaller the distance, the more similar the neural image features, the higher the risk of conversion of AFD to BD and the shorter the conversion time.

[0042] 2. Construction of graph neural network model (key technical point) This is the core innovation point of the application, and its structure is shown in Figure 1 .

[0043] Figure 1 A in the above formula is a disease diagnosis model based on a graph neural network (GNN), which can distinguish between depression (MDD) and bipolar disorder (BD); Figure 1 B in the above formula is a trained model that can be used for feature extraction to capture whether a subject has MDD or BD characteristics; Figure 1C in the figure shows the detailed structure of the GNN model, including attention coefficient calculation, information aggregation, node feature update, etc.

[0044] 2.1 Functional connectivity graph construction: Define the whole brain FC as a graph G = (V, E, X). Where the node set V corresponds to 90 brain regions; the weight e ij of the edge set E is derived from the FC value r ij . ij = 1- r ij is the Pearson correlation coefficient of the time series of brain regions i and j. This transformation maps the correlation range [-1, 1] to the numerical range [0, 1] more suitable for graph neural network processing; the feature vector X of each node is the FC value of all other brain regions, representing the connection strength between this node and all other 89 nodes.

[0045] 2.2 Graph neural network model construction and training (core technology) 2.21 Model input: the graph G constructed above.

[0046] 2.22 Model architecture: First layer: Edge-Weighted Graph Attention layers (EGAT). This layer is the key improvement point of the invention, which integrates the original FC weight e ij into the attention mechanism. First, calculate the original attention coefficient between nodes: att ij = a (Wh i , Wh j ), where W is a learnable weight matrix, and a is a single-layer feedforward neural network. Then, modulate the attention coefficient with the edge weight e ij : eatt ij = softmax j (att ij e ij ) = att . Finally, use the modulated attention coefficient to weight the sum of neighbor node features, and pass through a nonlinear activation function σ to get the new features of the node: .

[0047] In the graph neural network model of the invention, the nonlinear activation function σ used in the first layer is the ELU (Exponential Linear Unit) activation function. This function is one of the commonly used deep learning activation functions, which can provide nonlinear transformation in different input intervals. ELU has advantages such as stable convergence, smooth gradient change in dealing with frequent noise and large differences in connection strength commonly found in brain network data, and is very suitable as the first layer activation method of the graph neural network.

[0048] Second layer: Differentiable Graph Pooling (DGP). This layer aggregates 90 nodes into 32 nodes through a learnable assignment matrix S, to coarsen the graph structure and learn hierarchical features. Node features and edge weights are pooled as follows: out = S T H in , E out = S T E in S.

[0049] Third layer: Edge-weighted graph attention layer and differentiable graph pooling layer are repeated once, further aggregating 32 nodes into 4 nodes.

[0050] Fourth layer: graph-level feature extraction and classification. The features of the last 4 nodes are concatenated to form a graph-level feature vector. Then, two fully connected layers (containing 8 and 2 neurons, respectively) are applied to output the final classification result (MDD or BD).

[0051] 2.23 Model training: Cross-entropy loss L ent is used as the main loss function, and an auxiliary link prediction loss L aux = ‖E in , SS T ‖ Fr is introduced to optimize the pooling process. The total loss function is L = L ent + λL aux (λ = 1).

[0052] 2.3 Feature extraction and risk prediction: 2.31 Feature extraction: The trained GNN model (with the last classification layer removed) is used as a feature extractor. For any subject (including AFD), its FC graph is input, and a high-dimensional feature vector is extracted from before the last fully connected layer as its "fingerprint" of brain FC.

[0053] 2.32 Similarity calculation and risk prediction: The Euclidean distance between the feature vector of each AFD patient and the average of all patients in the BD group is calculated. The smaller the distance, the more similar the neuroimaging features to BD. Establish a risk prediction model: the study found that the Euclidean distance is positively correlated with the time of AFD transforming into BD. That is, the more similar the features of the patient to BD, the shorter the transformation time. Therefore, the Euclidean distance can be used as a quantitative risk indicator to predict the transformation risk and time trend of individual patients.

[0054] 2.4 Key biomarker identification (model interpretability) The attention matrix is extracted from the trained EGAT layer, which contains the contribution weight of each FC to the model decision. All weights are sorted, and the top 100 FCs with the highest weights are selected as the "key FCs" that distinguish MDD from BD. By accumulating the attention weights of the functional connections in which each brain region is located, the top 5 key brain regions (left superior medial frontal gyrus, left middle temporal gyrus, left inferior temporal gyrus, left fusiform gyrus, and right occipital gyrus) are identified.

[0055] 3. Alternative 3.1 The GNN (graph neural networks) can be replaced by other graph neural network models such as GCN (graph convolutional network), GraphSAGE (graph sample and aggreGatE), but EGAT is preferred due to its fusion of edge weights.

[0056] 3.2 In addition to Euclidean distance, cosine similarity, Mahalanobis distance, etc. can also be used.

[0057] 3.3 Clinical variables (such as depression and hypomania scale scores, family history) and image features can be fused and input into a comprehensive prediction model.

[0058] 4. Key points of technology and points to be protected Key point 1: Construct the FC matrix as a graph representation, and use e ij =1- as the calculation method of edge weight.

[0059] Key point 2: Use edge-weighted graph attention mechanism (EGAT) in GNN, and integrate the original FC strength as prior knowledge into attention calculation.

[0060] Key point 3: Use the trained GNN model to extract high-dimensional feature representation, and use its Euclidean distance with BD as a biomarker for quantifying conversion risk.

[0061] Key point 4: Based on the attention matrix of GNN, automatically identify the key FCs and brain regions most relevant to disease differentiation and conversion.

[0062] In the specific embodiments of the present application, the following specific examples are provided: Example 1: Model construction and training Data: 194 BD patients and 229 non-AFD (NAFD) patients with depression were recruited as the training set.

[0063] Steps: Follow the a, b, c parts of the above "technical solution" strictly. Use five-fold cross-validation to optimize the model hyperparameters, and the final model achieves a classification accuracy of 75% on the independent test set.

[0064] Results: A GNN model and its parameters that can robustly distinguish MDD from BD are successfully obtained.

[0065] Example 2: Clinical application of AFD conversion risk prediction Data: 85 patients diagnosed with AFD at baseline, of which 5 converted to BD within a one-year follow-up period and had baseline MRI data.

[0066] Steps: Using the model trained in Example 1, extract the baseline high-dimensional feature vectors of the 5 AFD patients who converted and the remaining AFD patients who did not convert. Calculate the Euclidean distance between each AFD patient's features and the average features of the BD group. Plot the scatter plot of Euclidean distance vs. conversion time and perform correlation analysis.

[0067] Results: The Euclidean distance is positively correlated with the conversion time (R>0, P<0.05), confirming the effectiveness of the index. For a new AFD patient, only the distance needs to be calculated to assess the short-term conversion risk.

[0068] Example 3: Verification of key biomarkers and symptom association Steps: Based on the attention weights of the trained model, extract the top 100 key FCs. In the AFD patient population, analyze the correlation between the intensity of these key FCs and the scores of each atypical symptom (such as increased appetite, sleepiness, and sensitivity to rejection) in the 30-item depressive symptom questionnaire (the Inventory of Depressive Symptomatology, IDS-30).

[0069] Results: Multiple key FCs were found to be significantly correlated with specific symptoms (e.g., the connection between the left middle temporal gyrus and the left occipital lobe was positively correlated with appetite). This not only verifies the reliability of the model in discovering biomarkers, but also links microscopic brain connectivity changes to macroscopic clinical symptoms, greatly enhancing the biological and clinical significance of the invention.

[0070] The invention introduces the weight of each functional connection into the attention calculation process, enabling the model to explicitly utilize the strength of neural activity coupling and improving sensitivity to key brain network features such as weak connections and negative connections.

[0071] The invention provides risk prediction for the conversion of atypical depression to bipolar disorder, which is a completely different clinical task scenario involving longitudinal follow-up data and disease evolution rules, requiring completely different model structures, loss functions, and feature interpretation methods.

[0072] The application uses a trained GNN to extract high-dimensional functional connectivity features; calculates the distance between the individual and the bipolar disorder feature center; uses the distance as a risk indicator to predict whether it is more likely to be converted into bipolar disorder in the future.

[0073] The specific implementation schemes of the embodiments can be understood with reference to the related descriptions in the above embodiments, which will not be repeated here.

[0074] It can be understood that the same or similar parts in the above embodiments can be mutually referred to, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0075] It should be noted that in the description of the present application, the terms "first", "second", etc. are only for the purpose of description, and cannot be understood as indicating or implying relative importance. In addition, in the description of the present application, unless otherwise specified, the meaning of "a plurality of" is at least two.

[0076] Any process or method descriptions in flow charts or described elsewhere herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions or other processes, and the various embodiments of the present application can include additional or fewer steps performing the described functions in the illustrated order, including using portions of code performing the functions in substantially the same way, or in the reverse order, or in substantially simultaneous operations with other functions performing in the background, as will be understood by those skilled in the art of the present application.

[0077] It should be understood that each part of the present application can be realized by hardware, software, firmware or their combination. In the above embodiments, a plurality of steps or methods can be realized by software or firmware stored in a memory and executed by a suitable instruction execution device. For example, if realized by hardware, and as in another embodiment, it can be realized by any one or their combination of the following technologies known in the art: discrete logic circuit with logic gate circuit for implementing logic function on data signal, application specific integrated circuit with suitable combination logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA) and the like.

[0078] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment method can be instructed by a program to complete the relevant hardware, and the corresponding program can be stored in a computer readable storage medium. The program, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0079] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing module, or each unit can be physically present separately, or two or more units can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. If the integrated module is realized in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0080] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0081] In the description of the present specification, the description referring to the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0082] The method, system, device, processor and medium for realizing AFD conversion risk prediction based on functional connection and graph neural network of the present application have high prediction accuracy, for the first time combine GNN and edge weight attention mechanism for this field, can capture the high-order topological features of brain network, and have achieved 75% accuracy on MDD and BD classification tasks, which is significantly better than traditional machine learning methods; have early warning capability, propose a quantitative index "neuroimage similarity", and prove its correlation with conversion time. This enables doctors to conduct individualized and quantitative assessment of the BD conversion risk of AFD patients before the appearance of clinical symptoms, realizing the leap from "static diagnosis" to "dynamic prediction". The model of the present application not only can predict, but also can output the FC and brain region most critical to decision making. The present application finds that the connections within the default mode network, the dorsal / ventral attention network and the visual network are key features, which have important scientific research value. The technical fusion of the present application is highly innovative, seamlessly integrates brain connectomics, graph neural network and representation learning, and provides a new and powerful technical framework for solving the heterogeneity of mental illness and the early prediction problem. The clinical conversion potential of the present application is great, and ultimately can be deployed as an auxiliary diagnosis software to provide objective decision support for psychiatrists, which helps to avoid misdiagnosis and mistreatment and improve the long-term prognosis of patients.

[0083] In this specification, the present application has been described with reference to its specific embodiments. However, it is obvious that various modifications and changes can be made without departing from the spirit and scope of the present application. Therefore, the specification and drawings should be considered as illustrative rather than limiting.

Claims

1. A method for predicting AFD phase transition risk based on functional connectivity and graph neural networks, characterized in that, The method includes the following steps: (1) Preprocess the images, extract the time series of each brain region based on the predefined brain atlas, and calculate the whole brain FC matrix; (2) Construct the FC matrix into graph structure data and input it into a graph neural network; (3) Calculate the Euclidean distance between the characteristic representation of the AFD patient to be evaluated and the average characteristic representation of the BD patient population.

2. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 1, characterized in that, Step (2) specifically includes the following steps: (2.1) Construct a functional connection diagram; (2.2) Construct and train the graph neural network model; (2.3) Extract features and perform risk prediction; (2.4) Identify key biomarkers.

3. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 2, characterized in that, Step (2.1) specifically includes the following steps: The whole brain FC matrix is ​​defined as graph G = (V, E, X); Here, the node set V corresponds to 90 brain regions, and the weight e of the edge set E is... ij From FC value r ij The converted FC value r ij is the Pearson correlation coefficient of the time series of brain region i and brain region j. The feature vector X of each node is the FC value of that node with all other brain regions, indicating the connection strength of that node with all other 89 nodes.

4. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 2, characterized in that, Step (2.2) specifically includes the following steps: (2.2.1) Input the constructed graph G; (2.2.2) Construct a graph neural network model; (2.2.3) Train the model.

5. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 4, characterized in that, Step (2.2.2) specifically includes the following steps: (2.2.2.1) Based on the edge-weighted graph attention layer, the original FC weights e ij Integrate into the attention mechanism; (2.2.2.2) Based on the differentiable graph pooling layer, 90 nodes are aggregated into 32 nodes through a learnable allocation matrix S to coarsen the graph structure; (2.2.2.3) Repeat steps (2.2.2.1) and (2.2.2.2) to aggregate the 32 nodes into 4 nodes; (2.2.2.4) The features of the four nodes are concatenated to form a graph-level feature vector; after passing through two fully connected layers, the final classification result is output.

6. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 5, characterized in that, The aforementioned step (2.2.2.1) specifically includes the following steps: (2.2.2.1.1) Calculate the original attention coefficients between nodes; (2.2.2.1.2) Through edge weight e ij Modulate the attention coefficient; (2.2.2.1.3) The features of neighboring nodes are weighted and summed using the modulated attention coefficients and then passed through a nonlinear activation function σ to obtain new features of the nodes.

7. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 4, characterized in that, The specific steps (2.2.3) are as follows: Using cross-entropy loss L ent As the main loss function, an auxiliary link prediction loss L is introduced. aux Optimize the pooling process.

8. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 2, characterized in that, Step (2.3) specifically includes the following steps: (2.3.1) The trained graph neural network (GNN) model is used as a feature extractor. The subject's full-factor graph is input to extract a high-dimensional feature vector. (2.3.2) Calculate the Euclidean distance between the feature vector of each AFD patient and the average feature vector of all patients in the BD group, and establish a risk prediction model.

9. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 8, characterized in that, The step (2.3.2) also includes calculating the cosine similarity or Mahalanobis distance between the feature vector of each AFD patient and the average feature vector of all patients in the BD group.

10. The method for predicting AFD phase transition risk based on functional connectivity and graph neural networks according to claim 2, characterized in that, Step (2.4) specifically includes the following steps: The attention matrix is ​​extracted from the trained EGAT layer, all weights are sorted, the top 100 FCs with the highest weights are selected, and the top 5 key brain regions are identified by accumulating the attention weights of the functional connections of each brain region.

11. A system for predicting AFD phase transition risk based on functional connectivity and graph neural networks, characterized in that, The system includes: The data processing module receives resting-state functional magnetic resonance imaging (fMRI) images of an individual at its input end. It is used to preprocess the images, extract time series of each brain region based on a predefined brain atlas, and calculate the whole-brain FC matrix. The graph neural network feature extraction module is connected to the data processing module and is used to construct the FC matrix into graph structure data and input it into a graph neural network. The similarity calculation and risk prediction module, connected to the graph neural network feature extraction module, is used to calculate the Euclidean distance between the feature representation of the AFD patient to be evaluated and the average feature representation of the BD patient group.

12. A device for predicting AFD phase transition risk based on functional connectivity and graph neural networks, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for predicting AFD phase transition risk based on functional connectivity and graph neural networks as described in any one of claims 1 to 10.

13. A processor for predicting AFD phase transition risk based on functional connectivity and graph neural networks, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the method for predicting AFD phase transition risk based on functional connectivity and graph neural networks as described in any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method for predicting AFD phase transition risk based on functional connectivity and graph neural networks as described in any one of claims 1 to 10.

Images