Cross-domain disease prediction method based on graph self-supervised transfer learning

Abstract

The present invention discloses a cross-domain disease prediction method based on graph self-supervised transfer learning, the method is: processing functional magnetic resonance imaging data and clinical phenotype data; modeling a brain feature structure topology map from the functional magnetic resonance imaging data of the subject; sampling the nodes of the brain feature structure topology map to obtain a subgraph and selecting the subgraph with the highest intimacy score; mapping the node features of the target subgraph to a low-dimensional vector space to obtain a node embedding matrix; converting the node embedding matrix into a reconstruction matrix, calculating the Euclidean distance between the original feature vector and the reconstructed feature vector to obtain a generative learning score; constructing a comparative relationship between the target node and the adjacent subgraph and calculating the comparative learning score; generating a learning score and a weighted sum of the comparative learning score to obtain a predicted probability value of the disease category; inputting the data to be migrated into the prediction model to obtain a prediction result of the corresponding disease category. The present invention improves the accuracy and efficiency of cross-domain disease prediction.

Description

Technical Field

[0001] The present invention relates to artificial intelligence and biomedicine, and in particular to a cross-domain disease prediction method based on graph self-supervised transfer learning. Background Art

[0002] With the advancement of biomedical imaging technology, functional magnetic resonance imaging (fMRI) has become an important tool for studying brain function and diagnosing neurological diseases. fMRI captures changes in functional connectivity between brain regions by measuring blood oxygen level-dependent (BOLD) signals in different brain regions, providing a non-invasive means to study the pathological characteristics of neurological diseases. However, current disease prediction models often face challenges when applied across domains, mainly manifested in the problems of domain shift and dataset heterogeneity. These problems are particularly evident in data across institutions and devices, because different population characteristics, imaging devices, and scanning protocols may lead to significant differences in data, and traditional machine learning methods often experience performance degradation when processing such heterogeneous data.

[0003] The domain shift problem is particularly serious in the field of biomedical and health informatics (BHI), affecting the generalization ability and clinical applicability of models. For example, the diagnosis of neurodevelopmental diseases such as autism spectrum disorder (ASD) often relies on behavioral assessments, which are often subject to subjective differences between assessors and lack objective and reliable indicators. Functional magnetic resonance imaging provides a possibility for accurate diagnosis of these diseases, but due to the different imaging equipment and protocols used by various hospitals and research institutions, it is difficult for a single model to maintain high performance in all clinical settings.

[0004] In order to address these challenges, transfer learning and self-supervised learning techniques have received widespread attention in recent years. Transfer learning significantly improves the generalization performance of the model by applying the knowledge learned in the source domain to the target domain, which is especially suitable for the common data heterogeneity problem in the field of biomedical imaging. Self-supervised learning can automatically extract useful feature representations from unlabeled data, reducing the dependence on large amounts of labeled data. Data labeling in the biomedical field is costly and time-consuming, and this problem can be effectively solved through self-supervised learning technology. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of data heterogeneity and domain shift caused by different population characteristics, imaging devices and scanning protocols in biomedical image analysis; to provide a cross-domain disease prediction method based on graph self-supervised transfer learning, to improve the analysis of functional magnetic resonance imaging (fMRI) data, so as to perform accurate disease diagnosis in different populations and devices.

[0006] The technical solution adopted by the present invention is:

[0007] The cross-domain disease prediction method based on graph self-supervised transfer learning includes the following steps:

[0008] Step 1: Processing functional magnetic resonance imaging (fMRI) data and clinical phenotype data from different institutions;

[0009] Step 2: Model the preprocessed fMRI data and phenotypic data to define a topological map of the connected brain features constructed from the fMRI data of the subject group. ;in, A collection of nodes representing the topological map of brain feature structures, specifically corresponding to the subject groups; The edge set representing the topological graph of brain feature structure specifically corresponds to the edge set consisting of similar fMRI images; The set of edge weights representing the topological graph of brain feature structures, specifically corresponding to the degree of similarity;

[0010] Step 3: Use the restarted random walk algorithm to sample the nodes in the brain feature structure topology map to obtain multiple subgraphs. The size of each subgraph obtained by sampling is a fixed value. (i.e., each subgraph contains nodes), and then evaluate the intimacy scores of the neighboring nodes of the target node in multiple subgraphs; select the node with the highest intimacy score subgraph as the target node Submap View ,in, Indicates the number of sub-graphs selected; Indicates the selected sub-graphs;

[0011] Step 4: Map the node features of the subgraph view of the target node into the low-dimensional vector space to obtain the node embedding matrix. The mapping formula is:

[0012] ,

[0013] in, Representation Node The embedding matrix of Is a node The characteristic matrix is ​​obtained from the characteristic matrix Extracted from the node and its neighbor nodes as rows, a matrix containing the eigenvectors corresponding to these nodes; Representation Node The weight matrix is ​​obtained from the weight matrix Extracted from the node A matrix centered on , containing the weight relationships between all pairs of nodes in the subgraph; is the activation function; ; is the weight matrix encoding the parameters used in the training process; is the index into the subgraph; is the identity matrix.

[0014] Step 5: Use the graph neural network decoder to convert the node embedding matrix into a reconstruction matrix. The decoding process is:

[0015] ,

[0016] in, Represents the parameter matrix used in the decoding training process; Representation Node The embedding matrix of Is a node The reconstructed feature matrix of Representation Node The weight matrix of is the index into the subgraph; is the activation function. Calculate the target node The original feature vector of and the feature vector reconstructed by the decoder The Euclidean distance between them is used to generate the learning score as follows:

[0017] ,

[0018] in, is the parameter that adjusts the final score to fall into [0,1]; Indicates the number of selected sub-graphs; is the index value of the subgraph; Indicates that the last row of the matrix is ​​selected as the reconstructed eigenvector.

[0019] Step 6: introduce a multi-view comparative learning module to build a comparative relationship between the target node and its adjacent context subgraphs, and make full use of the rich structural information in the graph to calculate the comparative learning score;

[0020] Step 7: The learning score and the comparative learning score are weighted and calculated to obtain the predicted probability value of the subject's corresponding disease category to form a prediction model, and the calculation formula is as follows:

[0021] : ,

[0022] in, and is an adjustable coefficient used to adjust the impact of the generation and comparison scoring functions;

[0023] Step 8: Input the functional magnetic resonance imaging data and clinical phenotype data of the target domain to be migrated into the prediction model, and output the prediction results of the corresponding disease categories of the patients in the target domain to complete the cross-domain auxiliary diagnosis of medicine.

[0024] Furthermore, the specific steps of step 1 are:

[0025] Step 1-1: Determine the specific types of fMRI data and clinical phenotype data required based on the characteristics of the disease to be predicted, including information such as gender, age, medical history, etc. Depending on the disease, specific clinical parameters may also be required, such as the stage of disease onset, medication history, family history, etc., as well as other biomarker data related to the disease.

[0026] Step 1-2: In the existing data set, select the corresponding image data (such as fMRI images) and clinical phenotype data (such as basic information of the patient, medical history, examination results, etc.) according to the determined demand type. Ensure that the selected data format meets the predetermined requirements and that the coding of the clinical data is consistent with the research requirements to avoid analysis problems caused by data format mismatch.

[0027] Steps 1-3: Clean and process the data to ensure data quality. First, remove duplicate data and unify the data format to ensure consistency between clinical phenotype data and fMRI image data. Then, handle missing values ​​and fill missing data by median imputation. Next, detect and process outliers, use visualization and statistical methods to identify outliers, and delete these outliers. Finally, integrate data from different sources and perform quality verification to ensure that the clinical and imaging data of each patient are correctly matched and there are no omissions or errors.

[0028] Furthermore, the specific steps of step 2 are:

[0029] Step 2-1, define the brain feature structure topology A collection of nodes , each subject is defined as a brain feature structure topology map , then the subject group is constructed into a brain feature structure topology map A collection of nodes ;

[0030] Step 2-2: Use an atlas based on anatomical landmarks to divide the cerebral cortex and subcortical regions into Areas of interest ,in Indicates The resting state functional magnetic resonance imaging data were registered to the standard anatomical space, and each region of interest was calculated Average time series of resting-state activity in the middle;

[0031] Step 2-3, use the following Pearson correlation coefficient Calculate different regions of interest and The functional connectivity between them, the correlation coefficient is calculated as follows:

[0032]

[0033] in, represents the Pearson correlation coefficient, and Respectively represent and Region of interest; Indicates Areas of interest At the point in time Time series values ​​of; Indicates Areas of interest At the point in time The time series value of Indicates Areas of interest The average of the time series values ​​of ; Indicates Areas of interest The average of the time series values ​​of ; The time point indicating the presence of the region of interest;

[0034] In step 2-4, the following Fisher transformation is applied to improve the normality of the correlation coefficient:

[0035] ,

[0036] in, Represents the Fisher transformed and Pearson correlation coefficient between them; is the inverse tangent function;

[0037] The feature matrix of all nodes It is expressed as: ,in Each row represents a node The eigenvector of ; is the number of regions of interest;

[0038] Step 2-5, define the brain feature structure topology The edge set , the similarity threshold is pre-set as . Through different nodes and Calculate the cosine similarity between the feature vectors , the calculation formula is as follows:

[0039] ,

[0040] in, For Node The eigenvector of For Node feature vector; determine whether the similarity between feature vectors is not less than a pre-given similarity threshold ; If so, then the two nodes and There is an edge between , and the corresponding adjacency matrix Elements in is 1. Otherwise, the two nodes and There is no edge between them, and the corresponding adjacency matrix Elements in is 0. Then all the edges form the edge set of the brain feature structure topology graph , and calculate the edge set as above The corresponding adjacency matrix ;

[0041] Step 2-6, define the brain feature structure topology The edge weight matrix , selected from clinical phenotype data Non-image indicators, each non-image indicator calculation formula is as follows:

[0042] ,

[0043] in, Representation Node No. Non-imaging indicators; Indicates the total number of selected non-image indicators; Representation Node A collection of non-image indicator data; Indicates the specific The index value of non-image indicator data;

[0044] side The weight value is calculated as follows:

[0045] ,

[0046] in, It is a measure of the distance between objects; Represents edge The weighted relationship of The weight value of .

[0047] Furthermore, the specific steps of step 3 are:

[0048] Step 3-1, the specific implementation of the restart walk algorithm is as follows: Preset the walk probability value and the number of nodes in the subgraph Starting from the current node, simulate the movement of random walkers in the network. At each step, the walker moves with probability Move to the adjacent node with probability ( Return to the starting node and iterate this process until the number of nodes in the subgraph reaches until.

[0049] In step 3-2, the calculation method of the intimacy score for each selected subgraph is as follows:

[0050] ,

[0051] in, Represents a node Subgraph of the destination node The intimacy score of For subgraph Medium The remaining nodes outside is a constant parameter, usually set to 0.15; represents the column-normalized adjacency matrix; is a diagonal matrix; On the diagonal of is the identity matrix.

[0052] Furthermore, the specific steps of step 6 are:

[0053] Step 6-1: Use the output node embedding matrix of the graph neural network encoder as the input of the comparative learning module, and transform the original feature vector of each node to obtain the node embedding vector The feature conversion method for a single node is ,in, For Node The eigenvector of is the weight matrix encoding the parameters used in the training process; is the activation function;

[0054] Step 6-2, use the average pooling function to get the sub-image The node embedding matrix Aggregate into a fixed-dimensional vector representation , the function is as follows:

[0055] ,

[0056] in, is the number of nodes in the subgraph; For all The index of the subgraph; Indicates the selection matrix OK;

[0057] Step 6-3, select As a positive example of node, the corresponding negative example is , is a feature matrix with a shuffled order; the scores of positive and negative sample pairs are as follows:

[0058] ,

[0059] ,

[0060] in, To represent the learnable parameter matrix, the same weight matrix is ​​used for the positive and negative examples. The scores of the positive and negative samples can be calculated using the above two formulas; the scores of the positive and negative samples are , the score corresponding to the negative sample pair is , Represents a node, Represents the node No. Subgraph , is the activation function.

[0061] Step 6-4, calculate the final contrastive learning score , by computing nodes of The average of the positive and negative sample scores of the selected subgraphs is used as the contrastive learning score of the current node.

[0062] The present invention adopts the above technical scheme, combines functional magnetic resonance imaging data with other clinical phenotype data, and constructs a brain feature topology map that accurately reflects the patient's pathological characteristics. On this basis, the present invention uses self-supervised learning to identify diseased subjects in the generated brain feature topology map. The present invention reveals any inconsistency in the attributes between the target node and its neighborhood by calculating the regression loss between the reconstructed feature vector and the original feature vector. At the same time, two scoring functions are designed to assign higher disease scores to attribute abnormalities and structural abnormalities in the graph. Nodes with scores exceeding the threshold are regarded as diseased subjects, while nodes with scores below the threshold are regarded as normal subjects. This method effectively converts the classification problem into a score comparison process, which is convenient for rapid screening of possible disease patients and provides auxiliary help information for subsequent diagnosis, thereby improving the accuracy and efficiency of disease diagnosis.

[0063] The beneficial effects of the present invention are: 1. A graph-based domain adaptation method is proposed to combine functional magnetic resonance imaging data with other clinical phenotype data to construct a topological map reflecting the patient's pathological characteristics. This method can adapt to the data heterogeneity generated by different imaging devices and protocols, and significantly enhance the generalization ability of the model in different clinical environments. 2. A graph self-supervised learning method is introduced. Through this method, the model can automatically learn useful feature representations on unlabeled medical datasets, thereby greatly reducing dependence on large amounts of annotated data. 3. The use of transfer learning methods ensures that the model can be effectively applied in new fields outside the source field and adapt to different patient groups and clinical environments. This is particularly important for medical research across institutions and populations. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments;

[0065] Figure 1 It is a flow chart of the cross-domain disease prediction method based on graph self-supervised transfer learning of the present invention;

[0066] Figure 2 It is a structural schematic diagram of the cross-domain disease prediction method based on graph self-supervised transfer learning of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0068] With the continuous advancement of biomedical imaging technology, functional magnetic resonance imaging (fMRI) has become a key tool for studying brain function and disease diagnosis. However, due to the diversity of data sources, including different population characteristics, imaging equipment, and scanning protocols, traditional machine learning models often encounter performance degradation in new scenarios, which is called the domain shift or dataset shift problem.

[0069] The present invention represents the patient's pathological information as a topological graph. Figure 2 As shown in the upper left, each patient is regarded as a node, and its feature vector is generated by imaging data, for example, by calculating functional connectivity of functional magnetic resonance imaging (fMRI) data and then normalizing the Pearson correlation coefficient using Fisher transformation to generate features. The edges in the topological graph symbolize the similarity between nodes (patients), and their weights are established by calculating the correlation between node feature vectors and converting them into edge weights. On this basis, the present invention uses a self-supervised learning algorithm to identify anomalies in the generated overall graph. The specific process includes selecting a subset of the subgraph corresponding to the target node using a restarted random walk algorithm. Subsequently, the proximity of the nodes in the subgraph to the target node is calculated according to the PersonalizedPageRank algorithm, and only the subgraph with higher proximity is retained for further detection of abnormalities in the target node. The present invention designs two self-supervised learning objectives to detect diseased subjects: one is attribute-based generative learning, and the other is structure-based comparative learning. Generative learning uses a graph neural network encoder and decoder to reconstruct the feature vector of the target node using the attribute information of the target node neighborhood, and identifies attribute abnormalities by minimizing the mean square error (MSE) between the original feature vector and the reconstructed feature vector. In contrast, structure-based comparative learning directly compares the differences between the target node embedding and its neighborhood context embedding through Jensen-Shannon divergence to identify structural anomalies. In the inference stage, the present invention calculates the final anomaly score of the target node based on the weighted sum of the attribute anomaly score and the structural anomaly score to assist medical diagnosis. This method effectively transforms the classification problem into a score comparison process, which not only reduces the dependence on a large amount of labeled data, but also significantly improves the accuracy and efficiency of disease diagnosis, while demonstrating strong capabilities in multimodal data fusion and cross-domain adaptation.

[0070] like Figure 1 As shown in or 2, the present invention discloses a cross-domain disease prediction method based on graph self-supervised transfer learning, which includes the following steps:

[0071] Step 1: Processing functional magnetic resonance imaging (fMRI) data and clinical phenotype data from different institutions;

[0072] Step 2: Model the preprocessed fMRI data and phenotypic data to define a topological map of the connected brain features constructed from the fMRI data of the subject group. ;in, A collection of nodes representing the topological map of brain feature structures, specifically corresponding to the subject groups; The edge set representing the topological graph of brain feature structure specifically corresponds to the edge set consisting of similar fMRI images; The set of edge weights representing the topological graph of brain feature structures, specifically corresponding to the degree of similarity;

[0073] Step 3: Use the restarted random walk algorithm to sample the nodes in the brain feature structure topology map to obtain multiple subgraphs. The size of each subgraph obtained by sampling is a fixed value. (i.e., each subgraph contains nodes), and then evaluate the intimacy scores of the neighboring nodes of the target node in multiple subgraphs; select the node with the highest intimacy score subgraph as the target node Submap View ,in, Indicates the number of sub-graphs selected; Indicates the selected sub-graphs;

[0074] Step 4: Map the node features of the subgraph view of the target node into the low-dimensional vector space to obtain the node embedding matrix. The mapping formula is:

[0075] ,

[0076] in, Representation Node The embedding matrix of Is a node The characteristic matrix is ​​obtained from the characteristic matrix Extracted from the node and its neighbor nodes as rows, a matrix containing the eigenvectors corresponding to these nodes; Representation Node The weight matrix is ​​obtained from the weight matrix Extracted from the node A matrix centered on , containing the weight relationships between all pairs of nodes in the subgraph; is the activation function; ; is the weight matrix encoding the parameters used in the training process; is the index into the subgraph; is the identity matrix.

[0077] Step 5: Use the graph neural network decoder to convert the node embedding matrix into a reconstruction matrix. The decoding process is:

[0078] ,

[0079] in, Represents the parameter matrix used in the decoding training process; Representation Node The embedding matrix of Is a node The reconstructed feature matrix of Representation Node The weight matrix of is the index into the subgraph; is the activation function. Calculate the target node The original feature vector of and the feature vector reconstructed by the decoder The Euclidean distance between them is used to generate the learning score as follows:

[0080] ,

[0081] in, is the parameter that adjusts the final score to fall into [0,1]; Indicates the number of sub-graphs selected; is the index value of the subgraph; Indicates that the last row of the matrix is ​​selected as the reconstructed eigenvector.

[0082] Step 6: introduce a multi-view comparative learning module to build a comparative relationship between the target node and its adjacent context subgraphs, and make full use of the rich structural information in the graph to calculate the comparative learning score;

[0083] Step 7: Weighted sum calculation of the learning score and the comparative learning score is performed to obtain the predicted probability value of the corresponding disease category to form a prediction model, and the calculation formula is as follows:

[0084] : ,

[0085] in, and is an adjustable coefficient used to adjust the impact of the generation and comparison scoring functions;

[0086] Step 8: Input the functional magnetic resonance imaging data and clinical phenotype data of the target domain to be migrated into the prediction model, and output the prediction results of the corresponding disease categories of the patients in the target domain to complete the cross-domain auxiliary diagnosis of medicine.

[0087] Furthermore, the specific steps of step 1 are:

[0088] Step 1-1: Determine the specific types of fMRI data and clinical phenotype data required based on the characteristics of the disease to be predicted, including information such as gender, age, medical history, etc. Depending on the disease, specific clinical parameters may also be required, such as the stage of disease onset, medication history, family history, etc., as well as other biomarker data related to the disease.

[0089] Step 1-2: In the existing data set, select the corresponding image data (such as fMRI images) and clinical phenotype data (such as basic information of the patient, medical history, examination results, etc.) according to the determined demand type. Ensure that the selected data format meets the predetermined requirements and that the coding of the clinical data is consistent with the research requirements to avoid analysis problems caused by data format mismatch.

[0090] Steps 1-3: Clean and process the data to ensure data quality. First, remove duplicate data and unify the data format to ensure consistency between clinical phenotype data and fMRI image data. Then, handle missing values ​​and fill missing data by median imputation. Next, detect and process outliers, use visualization and statistical methods to identify outliers, and delete these outliers. Finally, integrate data from different sources and perform quality verification to ensure that the clinical and imaging data of each patient are correctly matched and there are no omissions or errors.

[0091] Furthermore, the specific steps of step 2 are:

[0092] Step 2-1, define the brain feature structure topology A collection of nodes , each subject is defined as a brain feature structure topology map , then the subject group is constructed into a brain feature structure topology map A collection of nodes ;

[0093] Step 2-2: Use an atlas based on anatomical landmarks to divide the cerebral cortex and subcortical regions into Areas of interest ,in Indicates The resting state functional magnetic resonance imaging data were registered to the standard anatomical space, and each region of interest was calculated Average time series of resting-state activity in the middle;

[0094] Step 2-3, use the following Pearson correlation coefficient Calculate different regions of interest and The functional connectivity between them, the correlation coefficient is calculated as follows:

[0095]

[0096] in, represents the Pearson correlation coefficient; and Respectively represent and Region of interest; Indicates Areas of interest At the point in time The time series value of Indicates Areas of interest At the point in time The time series value of Indicates Areas of interest The average of the time series values ​​of ; Indicates Areas of interest The average of the time series values ​​of ; The time point indicating the presence of the region of interest;

[0097] In step 2-4, the following Fisher transformation is applied to improve the normality of the correlation coefficient:

[0098] ,

[0099] in, Represents the Fisher transformed and Pearson correlation coefficient between them; is the inverse tangent function; the characteristic matrix of all nodes It is expressed as: ,in, Each row represents a node The eigenvector of ; is the number of regions of interest;

[0100] Step 2-5, define the brain feature structure topology The edge set , the similarity threshold is pre-set as . Through different nodes and Calculate the cosine similarity between the feature vectors , the calculation formula is as follows:

[0101] ,

[0102] in, For Node The eigenvector of For Node feature vector; determine whether the similarity between feature vectors is not less than a pre-given similarity threshold ; If so, then the two nodes and There is an edge between , and the corresponding adjacency matrix Elements in is 1. Otherwise, the two nodes and There is no edge between them, and the corresponding adjacency matrix Elements in is 0. Then all the edges form the edge set of the brain feature structure topology graph , and calculate the edge set as above The corresponding adjacency matrix ;

[0103] Step 2-6, define the brain feature structure topology The edge weight matrix , selected from clinical phenotype data Non-image indicators, each non-image indicator calculation formula is as follows:

[0104] ,

[0105] in Representation Node No. Non-imaging indicators; Indicates the total number of selected non-image indicators; Representation Node A collection of non-image indicator data; Indicates the specific The index value of non-image indicator data; The weight value is calculated as follows:

[0106] ,

[0107] in, It is a measure of the distance between objects; Represents edge The weighted relationship of The weight value of .

[0108] Furthermore, the specific steps of step 3 are:

[0109] Step 3-1, the specific implementation of the restart walk algorithm is as follows: Preset the walk probability value and the number of nodes in the subgraph Starting from the current node, simulate the movement of random walkers in the network. At each step, the walker moves with probability Move to the adjacent node with probability ( Return to the starting node and iterate this process until the number of nodes in the subgraph reaches until.

[0110] In step 3-2, the calculation method of the intimacy score for each selected subgraph is as follows:

[0111] ,

[0112] in, Represents a node Subgraph of the destination node Intimacy rating; For subgraph Medium The remaining nodes outside is a constant parameter, usually set to 0.15; represents a column-normalized adjacency matrix; is a diagonal matrix; On the diagonal of is the identity matrix.

[0113] Furthermore, the specific steps of step 6 are:

[0114] Step 6-1: Use the output node embedding matrix of the graph neural network encoder as the input of the comparative learning module, and transform the original feature vector of each node to obtain the node embedding vector The feature conversion method for a single node is ,in, For Node The eigenvector of is the weight matrix encoding the parameters used in the training process; is the activation function;

[0115] Step 6-2, use the average pooling function to get the sub-image The node embedding matrix Aggregate into a fixed-dimensional vector representation , the function is as follows:

[0116] ,

[0117] in, is the number of nodes in the subgraph; For all The index of the subgraph; Indicates the selection matrix OK;

[0118] Step 6-3, select As a positive example of node, the corresponding negative example is , is a feature matrix with a shuffled order; the scores of positive and negative sample pairs are as follows:

[0119] ,

[0120] ,

[0121] in, To represent the learnable parameter matrix, the same weight matrix is ​​used for the positive and negative examples. The scores of the positive and negative samples can be calculated using the above two formulas; the scores of the positive and negative samples are , the score corresponding to the negative sample pair is , Represents a node, Represents the node No. Subgraph , is the activation function;

[0122] Step 6-4, calculate the final contrastive learning score , by computing nodes of The average of the positive and negative sample scores of the selected subgraphs is used as the contrastive learning score of the current node.

[0123] The present invention successfully proposed a cross-domain disease prediction method based on graph self-supervised transfer learning, which effectively solved the increasingly prominent domain transfer problem in the field of biomedicine and health informatics (BHI). Extensive test results on the ABIDE dataset showed that the method of the present invention performed well in detecting ASD and achieved excellent results on multiple key indicators. These results demonstrated strong generalization capabilities across different patient groups and clinical environments, proving that the model of the present invention can effectively transfer knowledge from the source domain to improve disease prediction performance in the target domain. In addition, by combining self-supervised learning and transfer learning, the model of the present invention can automatically learn useful feature representations.

[0124] The present invention adopts the above technical scheme, combines functional magnetic resonance imaging data with other clinical phenotype data, and constructs a brain feature topology map that accurately reflects the patient's pathological characteristics. On this basis, the present invention uses self-supervised learning to identify diseased subjects in the generated brain feature topology map. The present invention reveals any inconsistency in the attributes between the target node and its neighborhood by calculating the regression loss between the reconstructed feature vector and the original feature vector. At the same time, two scoring functions are designed to assign higher disease prediction probability values ​​to attribute abnormalities and structural abnormalities in the graph to assist medical diagnosis. This method effectively converts the classification problem into a score comparison process, facilitates the rapid screening of possible disease patients, and provides auxiliary help information for subsequent diagnosis, thereby improving the accuracy and efficiency of disease diagnosis.

[0125] The beneficial effects of the present invention are: 1. A graph-based domain adaptation method is proposed to combine functional magnetic resonance imaging data with other clinical phenotype data to construct a topological map reflecting the patient's pathological characteristics. This method can adapt to the data heterogeneity generated by different imaging devices and protocols, and significantly enhance the generalization ability of the model in different clinical environments. 2. A graph self-supervised learning method is introduced. Through this method, the model can automatically learn useful feature representations on unlabeled medical datasets, thereby greatly reducing dependence on large amounts of annotated data. 3. The use of transfer learning methods ensures that the model can be effectively applied in new fields outside the source field and adapt to different patient groups and clinical environments. This is particularly important for medical research across institutions and populations.

[0126] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians of the art without making creative work are within the scope of protection of the present application.

Claims

1. A cross-domain disease prediction method based on graph self-supervised transfer learning, characterized by: It includes the following steps: Step 1: Processing functional magnetic resonance imaging data and clinical phenotype data from different institutions; Step 2: Model the preprocessed fMRI data and phenotypic data to define a topological map of the connected brain features constructed from the fMRI data of the subject group. ;in, A collection of nodes representing the topological map of brain feature structures, specifically corresponding to the subject groups; The edge set representing the topological graph of the brain characteristic structure specifically corresponds to the edge set consisting of similar functional magnetic resonance imaging images; The set of edge weights representing the topological graph of brain feature structures, specifically corresponding to the degree of similarity; Step 3: Use the restarted random walk algorithm to sample the nodes in the brain feature structure topology map to obtain multiple subgraphs. The size of each subgraph obtained by sampling is a fixed value. , that is, each subgraph contains nodes, and then evaluate the intimacy scores of the neighboring nodes of the target node in multiple subgraphs; select the node with the highest intimacy score subgraph as the target node Submap View ,in, Indicates the number of sub-graphs selected; Indicates the selected sub-graphs, ; Each selected subgraph The intimacy score is calculated as follows: ， in, Represents a node Subgraph of the destination node The intimacy score of ; For subgraph Medium The remaining nodes outside is a constant parameter; represents the column-normalized adjacency matrix, Represents edge The weighted relationship of The weight value of For two nodes and There is an edge between is a diagonal matrix; On the diagonal of is the identity matrix; Step 4, mapping the node features of the subgraph view of the target node into a low-dimensional vector space to obtain a node embedding matrix; Step 5, convert the node embedding matrix into a reconstruction matrix, and calculate the Euclidean distance between the original feature vector of the target node and the feature vector reconstructed by the decoder to obtain the generative learning score; Step 6: introduce a multi-view comparative learning module to build a comparative relationship between the target node and its adjacent context subgraphs, and use the structural information in the graph to calculate the comparative learning score. The specific steps of step 6 are: Step 6-1: Use the output node embedding matrix of the graph neural network encoder as the input of the comparative learning module, and transform the original feature vector of each node to obtain the node embedding vector ; The feature conversion method for a single node is ,in, For Node The eigenvector of is the weight matrix encoding the parameters used in the training process; is the activation function; Step 6-2, use the average pooling function to get the sub-image Node The embedding matrix Aggregate into a fixed-dimensional vector representation , the function is as follows: ， in, is the number of nodes in the subgraph; For all The index of the subgraph; Indicates the selection matrix OK; Step 6-3, select As a positive example of node, the corresponding negative example is , is a feature matrix with a shuffled order; the scores of positive and negative sample pairs are as follows: ， ， in, To represent the learnable parameter matrix, the positive and negative examples use the same weight matrix in the calculation to calculate the scores of positive and negative samples; the corresponding scores of positive sample pairs are , the score corresponding to the negative sample pair is , Represents a node, Represents the node No. subgraphs and is the activation function; Step 6-4, calculate the final contrastive learning score , by computing nodes of The average of the positive and negative sample scores of the selected subgraphs is used as the contrastive learning score of the current node; Step 7, weighting and calculating the generated learning score and the comparative learning score to obtain the predicted probability value of the subject's corresponding disease category to form a prediction model; Step 8: Input the functional magnetic resonance imaging data and clinical phenotype data of the target domain to be migrated into the prediction model, and output the prediction results of the corresponding disease categories of the patients in the target domain to complete the cross-domain auxiliary diagnosis of medicine.

2. The cross-domain disease prediction method based on graph self-supervised transfer learning according to claim 1 is characterized in that: The specific steps of step 1 are: Step 1-1, determine the specific types of fMRI data and clinical phenotype data required according to the actual situation of the disease to be predicted, including the patient's gender, age, and medical history; Step 1-2: Search for graph data with consistent data format and clinical phenotype data with consistent coding in research institutions and data warehouses according to the type of demand; Steps 1-3, clean the data to be found and handle missing values ​​and diseased values.

3. The cross-domain disease prediction method based on graph self-supervised transfer learning according to claim 1, characterized in that: The specific steps of step 2 are: Step 2-1, define the brain feature structure topology A collection of nodes , each subject is defined as a brain feature structure topology map , then the subject group is constructed into a brain feature structure topology map A collection of nodes ; Step 2-2: Use an atlas based on anatomical landmarks to divide the cerebral cortex and subcortical regions into Areas of interest ,in Indicates The resting state functional magnetic resonance imaging data were registered to the standard anatomical space, and each region of interest was calculated Average time series of resting-state activity in the middle; Step 2-3, use the following Pearson correlation coefficient Calculate different regions of interest and The functional connectivity between them, the correlation coefficient is calculated as follows: ; in, represents the Pearson correlation coefficient, and Respectively represent and Region of interest; Indicates Areas of interest At the point in time The time series value of; Indicates Areas of interest At the point in time The time series value of Indicates Areas of interest The average of the time series values ​​of ; Indicates Areas of interest The average of the time series values ​​of ; The time point indicating the existence of the region of interest; In step 2-4, the following Fisher transformation is applied to improve the normality of the correlation coefficient: ， in, Represents the Fisher transformed and Pearson correlation coefficient between them; is the inverse tangent function; The feature matrix of all nodes It is expressed as: ,in Each row represents a node The eigenvector of ; is the number of regions of interest; Step 2-5, define the brain feature structure topology The edge set , the similarity threshold is pre-set as ; Through different nodes and Calculate the cosine similarity between the feature vectors , the calculation formula is as follows: ， in, For Node The eigenvector of For Node feature vector; determine whether the similarity between feature vectors is not less than a pre-given similarity threshold ; If so, then the two nodes and There is an edge between , and the corresponding adjacency matrix Elements in is 1; otherwise, the two nodes and There is no edge between them, and the corresponding adjacency matrix Elements in is 0; then all the edges form the edge set of the brain feature structure topology graph , and calculate the edge set The corresponding adjacency matrix ; Step 2-6, define the brain feature structure topology The edge weight set , selected from clinical phenotype data Non-image indicators, each non-image indicator calculation formula is as follows: ， in, Representation Node No. Non-imaging indicators, ; Indicates the total number of selected non-image indicators; Representation Node A collection of non-image indicator data; Indicates the specific The index value of non-image indicator data; The weight value is calculated as follows: ， in, It is a measure of the distance between objects; Represents edge The weighted relationship of The weight value of .

4. The cross-domain disease prediction method based on graph self-supervised transfer learning according to claim 1, characterized in that: The specific steps in step 3 are: Step 3-1, use the restart random walk algorithm to sample the destination node in the brain feature structure topology map to obtain multiple sub-graphs. The specific restart walk algorithm is as follows: Preset the walk probability value and the number of nodes in the subgraph ; Starting from the current node, simulate the movement of random walkers in the network; At each step, the walker Move to the adjacent node with probability ( Return to the starting node and iterate until the number of nodes in the subgraph reaches until; Step 3-2, calculate each selected subgraph Intimacy rating.

5. The cross-domain disease prediction method based on graph self-supervised transfer learning according to claim 1, characterized in that: In step 4, the subgraph of the target node The mapping formula when mapping the node features to the low-dimensional feature vector space is as follows: ， in, Representation Node The embedding matrix of Is a node The characteristic matrix is ​​obtained from the characteristic matrix Extracted from the node and Node The neighbor nodes of are rows, and the matrix contains the eigenvectors corresponding to the nodes; Representation Node The weight matrix is ​​obtained from the weight matrix Extracted from the node A matrix centered on , containing the weight relationships between all pairs of nodes in the subgraph; is the activation function; , is a diagonal matrix, On the diagonal of ; is the weight matrix encoding the parameters used in the training process; is the index into the subgraph; is the identity matrix.

6. The cross-domain disease prediction method based on graph self-supervised transfer learning according to claim 1, characterized in that: The decoding process of converting the node embedding matrix into the reconstruction matrix in step 5 is as follows: ， in, Represents the parameter matrix used in the decoding training process; Representation Node The embedding matrix of Is a node The reconstructed feature matrix of , is a diagonal matrix, On the diagonal of ; Representation Node The weight matrix of is the index into the subgraph; is the activation function; Calculate target node The original feature vector of and the feature vector reconstructed by the decoder The Euclidean distance between them is used to generate the learning score as follows: ， in, is the parameter that adjusts the final score to fall into [0,1]; Indicates the number of selected sub-graphs; is the index value of the subgraph; Indicates that the last row of the matrix is ​​selected as the reconstructed eigenvector.

7. The cross-domain disease prediction method based on graph self-supervised transfer learning according to claim 1, characterized in that: The specific steps of step 7 are: The generated learning score and the comparative learning score are weighted and calculated to obtain the predicted probability value of the corresponding disease category of the subject as follows: ， in, represents the generated learning score; represents contrastive learning score; and is an adjustable coefficient used to adjust the impact of the generation and comparison scoring functions.

Images