Personalized hypertensive nephropathy risk assessment method and system based on graph convolutional network

Abstract

The invention discloses a personalized hypertensive nephropathy risk assessment method and system based on a graph convolutional network, and the method comprises the steps: collecting modal data related to hypertensive nephropathy, and carrying out the preprocessing; constructing a patient similarity graph; performing multi-source information fusion and domain knowledge integration; sparse processing is carried out on the graph structure, and standardization processing is carried out on an adjacent matrix; modeling through a time sequence diagram convolutional network and a long-short term memory network, and capturing the evolution trend of patient characteristics along with time; through a self-attention mechanism and a graph regularization item, the transmission of important features is enhanced; a multi-head attention mechanism is adopted, and different types of relation modes are captured; optimizing model parameters by using an Adam optimization algorithm and a regularization method; and a risk assessment result is optimized by adopting weighted stacking and a self-adaptive attention mechanism. According to the risk assessment method, accurate assessment of the risk of hypertensive nephropathy is realized by constructing the multi-dimensional and dynamically changing patient similarity graph.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a personalized hypertensive nephropathy risk assessment method and system based on a graph convolutional network. Background Art

[0002] Hypertensive nephropathy (HN) is kidney damage caused by long-term uncontrolled hypertension, and has become one of the main causes of chronic kidney disease (CKD) and end-stage renal disease (ESRD) worldwide. In pathophysiology, kidney damage caused by hypertension is mainly manifested as renal arteriosclerosis, glomerular sclerosis, and renal interstitial fibrosis, which can lead to a gradual decrease in glomerular filtration rate (GFR) and eventually lead to renal failure. Since the early symptoms of hypertensive nephropathy are usually not obvious, many patients are diagnosed only when the disease progresses to the late stage, missing the best time for intervention and treatment. Therefore, early identification of high-risk groups and risk assessment are crucial to delay disease progression and improve patient prognosis. At present, the risk assessment of hypertensive nephropathy mainly relies on traditional statistical methods and scoring systems based on clinical experience, such as the Framingham risk score and the epidemiological collaboration (Equation). However, these methods have limitations in dealing with high-dimensional, multimodal and nonlinear data, and it is difficult to accurately capture the complex similarity relationships between patients and the dynamic characteristics of disease progression. In addition, heterogeneous factors such as genetic background, environmental factors and lifestyle of individual patients also affect the accuracy of risk assessment. With the rapid development of artificial intelligence and deep learning technology, advanced models such as deep feedforward neural network (DFNN) and graph neural network (GNN) have been introduced into the medical field, providing new ideas for risk assessment and disease prediction. In particular, graph convolutional network (GCN) performs well in processing non-Euclidean spatial data and capturing high-order neighborhood relationships, and has achieved success in bioinformatics, drug discovery and social network analysis. However, a single model still has shortcomings in dealing with multi-source heterogeneous data, dynamically changing disease characteristics and individualized risk prediction, and it is difficult to meet the needs of clinical practice. Summary of the invention

[0003] The purpose of the present invention is to provide a personalized hypertensive nephropathy risk assessment method and system based on graph convolutional network to solve the problems mentioned in the background technology.

[0004] To achieve the above object, the present invention provides a personalized hypertensive nephropathy risk assessment method based on graph convolutional network, comprising the following steps:

[0005] Step 100: collecting modal data related to hypertensive nephropathy, preprocessing the modal data, and then establishing a patient feature matrix;

[0006] Step 200: construct a patient similarity graph by measuring and encoding similarities between patients and enhancing patient feature representation using a graph convolutional network;

[0007] Step 300: performing multi-source information fusion and domain knowledge integration, wherein the multi-source information fusion includes structured data fusion, unstructured data fusion and image feature fusion;

[0008] Step 400: Sparse the graph structure and then normalize the adjacency matrix;

[0009] Step 500: Processing time series data and modeling through time series graph convolutional networks and long short-term memory networks to capture the evolution trend of patient characteristics over time;

[0010] Step 600: Strengthen the transfer of important features through the self-attention mechanism and graph regularization term;

[0011] Step 700: In each layer of the temporal graph convolutional network, a multi-head attention mechanism is used to capture different types of relationship patterns;

[0012] Step 800: Optimize GCN parameters, base model parameters, integration weights and personalized attention parameters using the Adam optimization algorithm and regularization method;

[0013] Step 900: Adopting weighted stacking and adaptive attention mechanism and optimizing the final risk assessment result by learning personalized loss function;

[0014] Step 800 is specifically as follows:

[0015] The Adam optimization algorithm is used to optimize all model parameters through the gradient descent method. The formula is as follows:

[0016]

[0017] Where η is the learning rate, and are the first-order and second-order moment estimates, ∈ is a small constant to prevent division by zero, and θ t+1 and θ t Represent the model parameters in t+1 and t iterations respectively;

[0018] At the same time, a variety of regularization techniques are used, including L2 regularization and Dropout method. Dropout is introduced in each hidden layer to randomly discard a part of neurons to enhance the generalization ability of the model:

[0019] h ′ =h⊙d,d~Bernoulli(p);

[0020] Among them, h ′ represents the feature vector after Dropout, h represents the original feature vector, and d represents the Dropout mask vector, whose elements conform to the Bernoulli distribution.

[0021] Preferably, in step 100:

[0022] The modal data includes biomarker data, electronic health records, lifestyle data and imaging data; the biomarker data includes serum creatinine, urine protein, blood pressure, blood sugar level, and blood lipid index; the electronic health records include medical history, diagnosis information, medication records, and complication information; the lifestyle data includes eating habits, exercise frequency, smoking and drinking, and sleep patterns; the imaging data includes renal ultrasound images and magnetic resonance imaging;

[0023] The pre-processing step comprises:

[0024] For different modal data, missing data are filled by using missing value filling and multiple interpolation methods;

[0025] Identify and handle outliers through statistical method box plot analysis;

[0026] Perform Z-score standardization on numerical features to ensure that all features are on the same scale. The formula is as follows:

[0027]

[0028] Among them, f′ i is the standardized i-th modal data, f i represents the original i-th modal data, μ i and σ i are the mean and standard deviation of the i-th mode data respectively;

[0029] The patient feature matrix is Where N is the number of patients and D is the feature dimension.

[0030] Preferably, in step 200:

[0031] The construction of the patient similarity graph includes graph structure definition and similarity measurement;

[0032] The graph structure definition is: the nodes of the patient similarity graph represent each patient, and the feature vector of the node is the preprocessed patient feature, expressed as Among them, x i represents the patient characteristics of the i-th modality data, and the edges of the patient similarity graph represent the similarities between patients;

[0033] Similarity measures include:

[0034] Biomarker similarity measurement: Based on the similarity measurement of patient biomarkers, the Gaussian similarity function is used to obtain the following formula:

[0035]

[0036] Among them, S biomarker (i,j) represents the Gaussian similarity measure of the patient’s biomarkers, x i,biomarker 、x j,biomarker Represents the feature vectors of the patient’s biomarkers at nodes i and j, σ biomarker represents the Gaussian kernel width of the patient’s biomarker;

[0037] Lifestyle similarity measurement: Based on the similarity measurement of lifestyle data, cosine similarity is used, and the formula is as follows:

[0038]

[0039] Among them, S lifestyle (i,j) represents the similarity measure of the patient’s lifestyle, x i,lifestyle 、x j,lifestyle Represent the feature vectors of the patient’s lifestyle at nodes i and j respectively;

[0040] Medical history similarity measurement: Based on the similarity measurement of the patient's medical history information, Jaccard similarity is used, and the formula is as follows:

[0041]

[0042] Among them, S history (i,j) The feature vector of the patient’s medical history information at nodes i and j, Represent the medical history set of the patient at nodes i and j respectively;

[0043] Dynamic similarity measurement: Using the dynamic weight adjustment mechanism, the weights of each similarity measurement are dynamically adjusted by learning the weight parameters, and the formula is as follows:

[0044]

[0045] Among them, A ijrepresents the variable weight between node i and node j in the patient similarity graph, x ik represents the kth eigenvalue of node i, x jk represents the kth eigenvalue of node j, K is the dimension of similarity measurement, w k is the weight of the k-th similarity dimension.

[0046] Preferably, in step 300:

[0047] The structured data fusion integrates biomarkers, electronic health records, etc., and calculates the similarity between patients through the similarity measurement method; the unstructured data fusion uses natural language processing to encode text data and incorporates the extracted high-dimensional text features into the similarity measurement; the image feature fusion extracts deep features of image data through pre-trained convolutional neural networks, and fuses the deep features with other modal data to enhance the depth and diversity of similarity measurement;

[0048] The domain knowledge integration includes defining similarities between patients based on clinical pathways and progression trajectories, adjusting similarity measurement weights through expert knowledge guidance, and dividing similarities into different levels, calculating them separately and integrating them into the overall similarity score.

[0049] Preferably, in step 400:

[0050] The thinning process includes:

[0051] By setting a similarity threshold θ, only edges with a similarity higher than θ are retained. The formula is as follows:

[0052]

[0053] The K nearest neighbor retention strategy is adopted. For each node, only its most similar K neighbors are retained. The formula is as follows:

[0054]

[0055] Among them, Top-K(A i,: ) represents the K nodes with the highest similarity between node i and other nodes;

[0056] The steps of the normalization process are as follows:

[0057]

[0058] Where I is the identity matrix, is the degree matrix, is the normalized adjacency matrix, represents the normalized adjacency matrix, Represents the normalized adjacency matrix The element at row i and column j in .

[0059] Preferably, step 500 is specifically as follows:

[0060] Step 510: Divide the patient data into multiple time windows according to time. Each time window t corresponds to a graph structure at a moment. In each time window t, recalculate the similarity measure based on the latest patient feature data and update the adjacency matrix A. t ,Through dynamic weight adjustment and multi-source information fusion, it reflects the latest patient characteristics and similarity relationships;

[0061] Step 520: Combined with the time series information, the graph structure data in the time window sequence is processed by the time series graph convolution network to capture the time series evolution of the patient characteristics. The formula is as follows:

[0062]

[0063] in, is the node feature matrix of the l+1th layer, is the normalized adjacency matrix of the t-th time window, is the node feature matrix of the lth layer, W (l) represents the weight matrix of the lth layer;

[0064] Step 530: Introduce a memory mechanism to serialize the graph convolution outputs of different time windows to capture long-term dependencies and feature evolution trends.

[0065] Preferably, step 600 is specifically as follows:

[0066] Step 610: Use the self-attention mechanism to enhance the node features and dynamically adjust the importance of different feature dimensions. The formula is as follows:

[0067]

[0068] Among them, α ij is the self-attention weight, h j is the feature vector of node j, W represents the weight matrix, h′ i Represents the new feature vector of node i after being enhanced by the self-attention mechanism;

[0069] Step 620: Add a graph regularization term to the loss function to make the features of similar nodes closer:

[0070]

[0071] Among them, ε is the edge set in the graph, h i The feature vector of node i, represents the graph regularization loss term;

[0072] Step 630: Randomly delete or add edge and node noise to enhance patient similarity graph data, the method includes edge perturbation, node perturbation and feature perturbation.

[0073] Preferably, step 700 is specifically as follows:

[0074] Step 710: In each layer of the temporal graph convolutional network GCN, a multi-head attention mechanism is used to calculate different attention weights in parallel through multiple attention heads to capture different types of relationship patterns. The formula is as follows:

[0075] MultiHead(Q,K,V)=Concat(head1,…,head h )W O ;

[0076] Among them, head i =Attention(Q i ,K i ,V i ), h is the number of attention heads, W O is the output weight matrix, MultiHead(Q,K,V) represents the multi-head attention mechanism, Q is the query vector, K is the key vector, and V is the value vector;

[0077] Step 720: Generate adversarial samples to improve the robustness of the model to input perturbations. Specific methods include fast gradient sign method and projected gradient descent:

[0078] Fast Gradient Sign Method: Generate adversarial samples using gradient information:

[0079]

[0080] Among them, x represents the original input sample, x ′ represents an adversarial sample, generated by adding a small perturbation to the original input sample. represents the loss function measuring the difference between the model prediction and the true label, ∈ represents the perturbation amplitude, Represents the loss function Partial derivative with respect to the original input sample x;

[0081] Projected Gradient Descent: Iteratively generate adversarial samples, the formula is as follows:

[0082]

[0083] in, is the set of constraints for adversarial samples, α is the step size, x t ′ represents the adversarial sample generated by t iterations, x t′ +1 represents the adversarial sample generated by t+1 iterations;

[0084] Step 730: Performing deep learning integrated and personalized risk assessment to ensure collaborative optimization of the model at the integrated and personalized levels;

[0085] Step 730 includes the following steps:

[0086] Step 731: Randomly initialize all model parameters, including GCN parameters, base model parameters, integrated weight coefficient α and personalized attention parameters;

[0087] Step 732: Optimize the integrated prediction loss and personalized loss to obtain the total loss function for:

[0088]

[0089] in, is the loss function for ensemble prediction, is the loss function of personalized prediction, γ is the weight of personalized loss, and λ is the regularization parameter.

[0090] Preferably, the specific steps of step 900 are as follows:

[0091] Step 910: Design a secondary model as the weight coefficient of the learning base model. The final prediction result is the weighted sum of the prediction results of each base model:

[0092]

[0093] in, Represents the final prediction result, α i represents the weight coefficient of the i-th base model, y i represents the prediction result of the i-th base model, m represents the total number of base models in the ensemble method,

[0094] Through back-propagation and optimization algorithms, the weight coefficient α is learned to minimize the loss function of the integrated prediction;

[0095] Combined with the individual characteristics of the patient, the adaptive attention mechanism is used to predict the risk of each patient. The specific steps include feature extraction, attention weight calculation and personalized weighting. The formula is as follows:

[0096] e = tanh(W1p+b1);

[0097] α=softmax(W2e+b2);

[0098]

[0099] Among them, e represents the intermediate representation vector obtained after linear transformation and nonlinear activation of the patient's individual features, p is the patient's individual feature vector, W1 and W2 are learnable weight matrices, b1 and b2 are bias vectors, and α is the attention weight;

[0100] Step 920: Design a personalized loss function, combining the error between the personalized prediction result and the true label, the formula is as follows:

[0101]

[0102] Where N is the number of samples, y k is the true label of the kth sample, Predict results for personalization.

[0103] The present invention also provides a personalized hypertensive nephropathy risk assessment system based on a graph convolutional network, comprising:

[0104] Data acquisition and preprocessing module: used to collect modal data related to hypertensive nephropathy, preprocess the modal data, and then establish a patient feature matrix;

[0105] Patient Similarity Graph Construction Module: It is used to construct a patient similarity graph by measuring and encoding the similarities between patients and enhancing the patient feature representation using a graph convolutional network;

[0106] Multi-source information fusion module: used for multi-source information fusion and domain knowledge integration;

[0107] Sparseness and normalization processing module: used to perform sparseness processing on the graph structure and then normalize the adjacency matrix;

[0108] Time series processing module: used to process time series data and capture the evolution trend of patient characteristics over time through time series graph convolutional network and long short-term memory network modeling;

[0109] Feature transfer enhancement module: used to enhance the transfer of important features through self-attention mechanism and graph regularization term;

[0110] Relationship pattern capture module: It is used to capture different types of relationship patterns using a multi-head attention mechanism in each layer of the temporal graph convolutional network;

[0111] Parameter optimization module: used to optimize GCN parameters, base model parameters, integration weights, and personalized attention parameters through the Adam optimization algorithm and regularization method;

[0112] Personalized risk assessment module: It is used to optimize the final risk assessment results by adopting weighted stacking and adaptive attention mechanism and learning personalized loss function.

[0113] Therefore, the present invention adopts the above-mentioned personalized hypertensive nephropathy risk assessment method and system based on graph convolutional network, which has the following beneficial effects:

[0114] (1) By constructing a multi-dimensional, dynamically updated patient similarity graph, fusing multi-source heterogeneous data, and adopting a weighted stacking integration strategy and adaptive attention mechanism, we can achieve accurate and personalized assessment of the risk of hypertensive nephropathy.

[0115] (2) It not only fully utilizes the advantages of GCN in capturing high-order relationships and complex similarities, but also improves the model's sensitivity and adaptability to changes in patient characteristics through multimodal data fusion and dynamic graph structure updates.

[0116] (3) This method will provide clinicians with an efficient and reliable risk prediction tool for hypertensive nephropathy, assist in early diagnosis and intervention, and ultimately improve patient prognosis and reduce the burden on the medical system.

[0117] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1 This is a flow chart of a personalized hypertensive nephropathy risk assessment method based on a graph convolutional network according to an embodiment of the present invention;

[0119] Figure 2 It is a model framework diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0120] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0121] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0122] Example

[0123] like Figure 1-2 As shown, the present invention provides a personalized hypertensive nephropathy risk assessment method based on graph convolutional network, aiming to achieve accurate assessment of hypertensive nephropathy (HN) risk. The core innovation lies in constructing a patient similarity graph (PSG) and using graph convolutional network to enhance patient feature representation, thereby achieving personalized risk prediction. The entire process runs through data collection and preprocessing, patient similarity graph construction, graph convolutional network feature enhancement, integrated deep learning prediction module and personalized risk assessment, and finally achieves efficient collaboration of the model through mathematical optimization and training strategy.

[0124] The risk assessment methodology includes the following steps:

[0125] Step 100: collecting modal data related to hypertensive nephropathy, preprocessing the modal data, and then establishing a patient feature matrix;

[0126] Modal data includes biomarker data, electronic health records, lifestyle data, and imaging data. Biomarker data include serum creatinine, urine protein, blood pressure, blood sugar levels, blood lipid indicators, etc. Electronic Health Records (EHR) include medical history, diagnostic information, medication records, complication information, etc. Lifestyle data include dietary habits (e.g., whether high-salt diet intake), exercise frequency, smoking and drinking, sleep patterns, etc., and dietary habits. Imaging data includes renal ultrasound images, magnetic resonance imaging (MRI), etc.

[0127] The preprocessing steps include filling missing values, outlier processing and standardization. The specific process is as follows:

[0128] For different modal data, missing data are filled by missing value filling and multiple interpolation methods to ensure data integrity; outliers are identified and processed by statistical box plot analysis to prevent them from having an adverse effect on model training; numerical features are Z-score standardized to ensure that each feature is on the same scale. The formula is as follows:

[0129]

[0130] Among them, f′ i is the standardized i-th modal data, f irepresents the original i-th modal data, μ i and σ i are the mean and standard deviation of the i-th mode data respectively.

[0131] After preprocessing, the patient feature matrix is ​​generated Where N is the number of patients and D is the feature dimension.

[0132] Step 200: By systematically measuring and encoding the similarities between patients and enhancing the patient feature representation using a graph convolutional network (GCN), a patient similarity graph (PSG) is constructed to achieve personalized risk assessment. The construction of PSG is one of the core of this method.

[0133] The construction of PSG involves several key steps, including graph structure definition and similarity measurement.

[0134] The graph structure definition refers to the definition of the graph structure of PSG. The nodes of PSG represent each patient, and the feature vector of the node is the preprocessed patient feature, expressed as Among them, x i Represents the patient characteristics of the i-th modality data. The edges of PSG represent the similarities between patients, and the weight of the edge reflects the degree of similarity between patients.

[0135] Similarity measurement is the key to PSG construction, which determines the closeness of the connection between patients and their representation in the graph. Although traditional similarity measurement methods such as cosine similarity and Euclidean distance are effective, they may not fully reflect the true relationship between patients in multidimensional complex data. To this end, this method adopts an innovative method of multidimensional similarity measurement and dynamic weight adjustment, comprehensively considering multiple dimensions such as biomarkers, lifestyle, medical history, etc., to ensure the comprehensiveness and accuracy of similarity assessment, including the following aspects:

[0136] Biomarker similarity: Based on the similarity measure of patient-based biomarkers (such as serum creatinine, urine protein, etc.), a Gaussian similarity function is used:

[0137]

[0138] Among them, S biomarker (i,j) represents the Gaussian similarity measure of the patient’s biomarkers, x i,biomarker 、x j,biomarker Represents the feature vectors of the patient’s biomarkers at nodes i and j, σ biomarker Represents the Gaussian kernel width of the patient biomarker.

[0139] Lifestyle similarity: A similarity measure based on lifestyle data (diet, exercise habits), using cosine similarity:

[0140]

[0141] Among them, S lifestyle (i,j) represents the similarity measure of the patient’s lifestyle, x i,lifestyle 、x j,lifestyle Represent the feature vectors of the patient’s lifestyle at nodes i and j respectively.

[0142] Medical history similarity: Based on the similarity measurement of the patient's medical history information (such as previous illness, family history), the Jaccard similarity formula is as follows:

[0143]

[0144] Among them, S history (i,j) The feature vector of the patient’s medical history information at nodes i and j, Represent the medical history sets of the patient at nodes i and j respectively.

[0145] Dynamic similarity measurement: Taking into account the different importance of different feature dimensions, the dynamic weight adjustment mechanism is used to dynamically adjust the weights of each similarity measurement by learning weight parameters. The formula is as follows:

[0146]

[0147] Among them, A ij represents the variable weight between node i and node j in the patient similarity graph, which is used to quantify the similarity between patients. ik represents the kth eigenvalue of node i, x jk represents the kth eigenvalue of node j, K is the dimension of similarity measurement, w k is the weight of the k-th similarity dimension, which is adaptively learned through the training process to reflect the contribution of different features to the similarity.

[0148] Step 300: Multi-source information fusion and domain knowledge integration are performed, wherein multi-source information fusion includes structured data fusion, unstructured data fusion and image feature fusion. In the process of PSG construction, a single data source often cannot fully reflect the similarity relationship between patients. Therefore, this method further improves the comprehensiveness and depth of similarity evaluation through multi-source information fusion and domain knowledge integration.

[0149] Structured data fusion integrates biomarkers, electronic health records, etc., and calculates the similarity between patients through the above-mentioned multidimensional similarity measurement method; unstructured data fusion uses natural language processing technology (such as BERT) to encode text data (medical record descriptions) and incorporates the extracted high-dimensional text features into the similarity measurement; image feature fusion extracts deep features of image data through pre-trained convolutional neural networks, and fuses these features with other modal data to enhance the depth and diversity of similarity measurement.

[0150] In addition, domain knowledge integration includes defining the similarity between patients based on clinical pathways and progression trajectories, adjusting the similarity measurement weights guided by expert knowledge, and hierarchical similarity, dividing the similarity into different levels (such as basic feature similarity, pathological progression similarity), calculating them separately and integrating them into the overall similarity score to ensure that the model fully captures the similarity between patients at different levels.

[0151] Step 400: perform sparse processing on the graph structure, and then perform normalization processing on the adjacency matrix.

[0152] In order to reduce the complexity of the graph and highlight the key similarity relationships, a graph structure sparseness strategy is introduced. Specifically, first, by setting a similarity threshold θ, only edges with similarity higher than θ are retained. The formula is as follows:

[0153]

[0154] At the same time, the K nearest neighbor retention strategy is adopted. For each node, only its most similar K neighbors are retained. The formula is as follows:

[0155]

[0156] Among them, Top-K(A i,: ) represents the K nodes with the highest similarity between node i and other nodes.

[0157] By setting the similarity threshold and retaining the K-nearest neighbor strategy, the edge structure of PSG is effectively sparse, which not only reduces the computational complexity but also ensures that only key similarity relationships are retained in the graph, thereby improving the efficiency and performance of the model.

[0158] Next, the adjacency matrix is ​​normalized to meet the input requirements of the graph convolutional network. The normalization steps are as follows:

[0159]

[0160]

[0161] Where I is the identity matrix, is the degree matrix, is the normalized adjacency matrix. Represents the adjacency matrix after normalization. Normalization aims to balance the degree of nodes, ensure the effective transmission of information in graph convolution, and avoid excessive influence of high-connectivity nodes on feature updates. Represents the normalized adjacency matrix This normalization method can balance the degree of nodes, ensure the effective transmission of information in graph convolution, and avoid excessive influence of high-degree nodes on feature updates.

[0162] Step 500: Process the time series data and capture the evolution trend of patient characteristics over time through time series graph convolutional network and long short-term memory network modeling. This step can reflect the dynamic characteristics of the patient's health status and lifestyle over time, and specifically includes the following steps:

[0163] Step 510: Divide the patient data into multiple time windows (such as monthly, quarterly), each time window corresponds to a graph structure at a certain moment, there are T time windows, each time window corresponds to an adjacency matrix A t In each time window t, the similarity measure is recalculated based on the latest patient characteristic data and the adjacency matrix A is updated t ,Through dynamic weight adjustment and multi-source information fusion, it ensures that the graph structure can reflect the latest patient characteristics and similarity relationships in real time.

[0164] Step 520: Combined with the time series information, the graph structure data in the time window sequence is processed by a time series graph convolutional network (such as Dynamic GCN, Temporal GCN) to capture the time series evolution of the patient's characteristics. The formula is as follows:

[0165]

[0166] in, is the node feature matrix of the l+1th layer, is the normalized adjacency matrix of the t-th time window, is the node feature matrix of the lth layer, W (l) Represents the weight matrix of the lth layer. Through the temporal graph convolutional network, the model can capture the evolution trend of patient characteristics over time and improve the timeliness and accuracy of risk assessment.

[0167] Step 530: Introduce a memory mechanism (such as a long short-term memory network, LSTM) to serialize the graph convolution outputs of different time windows, capture long-term dependencies and feature evolution trends, and further enhance the model's time series prediction capabilities.

[0168] Step 600: Through the self-attention mechanism and graph regularization term, the transmission of important features is strengthened, the expressiveness of PSG and the robustness of the model are further improved, and the model's adaptability to changes in graph results is improved.

[0169] Step 610: Use the self-attention mechanism to enhance node features and dynamically adjust the importance of different feature dimensions:

[0170]

[0171] Among them, α ij is the self-attention weight, h j is the feature vector of node j, W represents the weight matrix, h ′ i Represents the new feature vector of node i after being enhanced by the self-attention mechanism. This new feature vector combines the important feature information of the node itself and its neighboring nodes. Through the calculation of the self-attention mechanism, the weight can be dynamically adjusted according to the feature similarity between nodes to strengthen the information transmission of important neighboring nodes.

[0172] Step 620: Add a graph regularization term to the loss function to encourage similar nodes to have closer features:

[0173]

[0174] Among them, ε is the edge set in the graph, h i The feature vector of node i, Represents the graph regularization loss term, which is used to constrain the model so that the feature vectors of connected nodes (i.e. similar patients) in the graph are as close as possible. Through graph regularization, the model can better capture the similarity relationship between nodes and improve the consistency and robustness of feature representation.

[0175] Step 630: By randomly deleting or adding edges, node noise and other graph data enhancement methods, the model's adaptability to graph structure changes and the model's generalization ability are improved. Specific methods include edge perturbation, node perturbation and feature perturbation.

[0176] Step 700: In each layer of the temporal graph convolutional network, a multi-head attention mechanism is used to capture different types of relationship patterns and enhance the diversity of feature representation.

[0177] Step 710: In order to further improve the expressiveness and flexibility of the graph convolutional network, a multi-head attention mechanism is used in each layer of the temporal graph convolutional network GCN. Different attention weights are calculated in parallel by multiple attention heads to capture different types of relationship patterns:

[0178] MultiHead(Q,K,V)=Concat(head1,…,head h )W O ;

[0179] Among them, head i =Attention(Q i ,K i ,V i ), h is the number of attention heads, W O is the output weight matrix. MultiHead(Q,K,V) represents a multi-head attention mechanism, where Q is the query vector, which is used to calculate the attention weight with the key vector, K is the key vector, which determines the attention weight together with the query vector, and V is the value vector, which is weighted and summed according to the attention weight to generate the final attention output.

[0180] MultiHead(Q,K,V) can capture diverse relational patterns from different representation subspaces through multiple parallel attention heads. That is, through the multi-head attention mechanism, the model can capture different relational patterns in parallel and improve the diversity and richness of feature representation.

[0181] Step 720: Improve the robustness of the model to input perturbations by generating adversarial examples, including the fast gradient sign method (FGSM) and projected gradient descent (PGD).

[0182] FGSM (Fast Gradient Sign Method): Generate adversarial samples using gradient information:

[0183]

[0184] Here, x represents the original input sample, which is the unperturbed input data received by the model during training or prediction. ′ Denotes adversarial examples, which are generated by adding small perturbations to the original input samples, aiming to fool the model into making incorrect predictions. Denotes the loss function measuring the difference between the model prediction and the true label, ∈ denotes the perturbation amplitude, which is a small constant used to control the size of the perturbation added to the original input sample. Represents the loss function The partial derivative with respect to the original input sample x indicates how to adjust x to increase or decrease the loss.

[0185] PGD ​​(Projected Gradient Descent): Iteratively generates adversarial samples. The formula is as follows:

[0186]

[0187] in, is the set of constraints for adversarial samples, α is the step size, x t ′ represents the adversarial sample generated by t iterations, x t ′ +1 represents the adversarial sample generated by t+1 iterations. Through adversarial training, the model can better adapt to the diversity and uncertainty of input data.

[0188] Step 730: Performing deep learning integrated and personalized risk assessment to ensure collaborative optimization of the model at the integrated and personalized levels, including the following steps:

[0189] Step 731: Randomly initialize all model parameters, including GCN parameters, base model parameters, integrated weight coefficient α and personalized attention parameters;

[0190] Step 732: Optimize the integrated prediction loss and personalized loss to obtain the total loss function for:

[0191]

[0192] in, is the loss function for ensemble prediction, is the loss function of personalized prediction, γ is the weight of personalized loss, and λ is the regularization parameter.

[0193] Step 800: Adopt the Adam optimization algorithm to optimize all model parameters by gradient descent method. The formula is as follows:

[0194]

[0195] Where η is the learning rate, and are the first-order and second-order moment estimates, ∈ is a small constant to prevent division by zero, and θ t+1 and θ t denote the model parameters in the t+1 and t iterations respectively.

[0196] At the same time, in order to prevent the model from overfitting, a variety of regularization techniques are used, including L2 regularization and Dropout method. Dropout is introduced in each hidden layer to randomly discard a part of neurons to enhance the generalization ability of the model:

[0197] h ′ =h⊙d,d~Bernoulli(p);

[0198] Among them, h ′ represents the feature vector after Dropout, h represents the original feature vector, and d represents the Dropout mask vector, whose elements conform to the Bernoulli distribution.

[0199] Through joint optimization and collaborative training strategies, the model can simultaneously optimize GCN parameters, base model parameters, integration weights and personalized attention parameters, improving the performance and generalization ability of the overall model.

[0200] Step 900: Perform integrated and personalized risk assessment, and ensure the accuracy of personalized prediction through weighted stacking and adaptive attention mechanism. Optimize the final risk assessment result by learning personalized loss function. The specific steps are as follows:

[0201] Step 910: Use a weighted stacking strategy to weight the prediction results of multiple base models to improve the overall prediction performance. Specifically, a secondary model (Meta-Learner) is designed to learn the weight coefficients of the base models, and the final prediction result is the weighted sum of the prediction results of each base model:

[0202]

[0203] in, Represents the final prediction result, α i represents the weight coefficient of the i-th base model, y i represents the prediction result of the i-th base model, m represents the total number of base models in the ensemble method,

[0204] Through back-propagation and optimization algorithms, the weight coefficient α is learned to minimize the loss function of the ensemble prediction.

[0205] Combined with the individual characteristics of patients, customized risk prediction for each patient is achieved through an adaptive attention mechanism. The specific steps include feature extraction, attention weight calculation and personalized weighting. The formula is as follows:

[0206] e = tanh(W1p+b1);

[0207] α=softmax(W2e+b2);

[0208]

[0209] Among them, e represents the intermediate representation vector obtained after linear transformation and nonlinear activation of the patient's individual characteristics, which is used as the basis for calculating the attention weight, p is the patient's individual feature vector, W1 and W2 are learnable weight matrices, b1 and b2 are bias vectors, and α is the attention weight.

[0210] Step 920: In order to optimize the personalized prediction results, a personalized loss function is designed, combining the error between the personalized prediction results and the true label:

[0211]

[0212] Where N is the number of samples, y k is the true label of the kth sample, The above method can provide a personalized prediction result. Finally, an accurate assessment result of the risk of hypertensive nephropathy is obtained.

[0213] The present invention also provides a personalized hypertensive nephropathy risk assessment system based on a graph convolutional network, comprising:

[0214] Data acquisition and preprocessing module: used to collect modal data related to hypertensive nephropathy, preprocess the modal data, and then establish a patient feature matrix;

[0215] Patient Similarity Graph Construction Module: It is used to construct a patient similarity graph by measuring and encoding the similarities between patients and enhancing the patient feature representation using a graph convolutional network;

[0216] Multi-source information fusion module: used for multi-source information fusion and domain knowledge integration;

[0217] Sparseness and normalization processing module: used to perform sparseness processing on the graph structure and then normalize the adjacency matrix;

[0218] Time series processing module: used to process time series data and capture the evolution trend of patient characteristics over time through time series graph convolutional network and long short-term memory network modeling;

[0219] Feature transfer enhancement module: used to enhance the transfer of important features through self-attention mechanism and graph regularization term;

[0220] Relationship pattern capture module: It is used to capture different types of relationship patterns using a multi-head attention mechanism in each layer of the temporal graph convolutional network;

[0221] Parameter optimization module: used to optimize GCN parameters, base model parameters, integration weights, and personalized attention parameters through the Adam optimization algorithm and regularization method;

[0222] Personalized risk assessment module: It is used to optimize the final risk assessment results by adopting weighted stacking and adaptive attention mechanism and learning personalized loss function.

[0223] Therefore, the present invention adopts the above-mentioned personalized hypertensive nephropathy risk assessment method and system based on graph convolutional network, adopts multi-source information fusion and complex similarity modeling, combines time series information and personalized adjustment, and can provide accurate hypertensive nephropathy risk assessment.

[0224] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A personalized hypertensive nephropathy risk assessment method based on graph convolutional network, characterized in that: The following steps are involved: Step 100: collecting modal data related to hypertensive nephropathy, preprocessing the modal data, and then establishing a patient feature matrix; Step 200: construct a patient similarity graph by measuring and encoding similarities between patients and enhancing patient feature representation using a graph convolutional network; Step 300: performing multi-source information fusion and domain knowledge integration, wherein the multi-source information fusion includes structured data fusion, unstructured data fusion and image feature fusion; Step 400: Sparse the graph structure and then normalize the adjacency matrix; Step 500: Processing time series data and modeling through time series graph convolutional networks and long short-term memory networks to capture the evolution trend of patient characteristics over time; Step 600: Strengthen the transfer of important features through the self-attention mechanism and graph regularization term; Step 700: In each layer of the temporal graph convolutional network, a multi-head attention mechanism is used to capture different types of relationship patterns; Step 800: Optimize GCN parameters, base model parameters, integration weights and personalized attention parameters using the Adam optimization algorithm and regularization method; Step 900: Adopting weighted stacking and adaptive attention mechanism and optimizing the final risk assessment result by learning personalized loss function; Step 800 is specifically as follows: The Adam optimization algorithm is used to optimize all model parameters through the gradient descent method. The formula is: Where η is the learning rate, and are the first-order and second-order moment estimates, ∈ is a small constant to prevent division by zero, and θ t+1 and θ t Represent the model parameters in t+1 and t iterations respectively; At the same time, a variety of regularization techniques are used, including L2 regularization and Dropout method. Dropout is introduced in each hidden layer to randomly discard a part of neurons to enhance the generalization ability of the model, and the formula is obtained: h′=h⊙d,d~Bernoulli(p); Among them, h′ represents the feature vector after Dropout, h represents the original feature vector, and d represents the Dropout mask vector, whose elements conform to the Bernoulli distribution.

2. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: In step 100: The modal data includes biomarker data, electronic health records, lifestyle data and imaging data; the biomarker data includes serum creatinine, urine protein, blood pressure, blood sugar level, and blood lipid index; the electronic health records include medical history, diagnosis information, medication records, and complication information; the lifestyle data includes eating habits, exercise frequency, smoking and drinking, and sleep patterns; the imaging data includes renal ultrasound images and magnetic resonance imaging; The pre-processing step comprises: For different modal data, missing data are filled by using missing value filling and multiple interpolation methods; Identify and handle outliers through statistical method box plot analysis; Perform Z-score standardization on numerical features to ensure that all features are on the same scale. The formula is as follows: Among them, f′ i is the standardized i-th modal data, f i represents the original i-th modal data, μ i and σ i are the mean and standard deviation of the i-th mode data respectively; The patient feature matrix is Where N is the number of patients and D is the feature dimension.

3. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: In step 200: The construction of the patient similarity graph includes graph structure definition and similarity measurement; The graph structure definition is: the nodes of the patient similarity graph represent each patient, and the feature vector of the node is the preprocessed patient feature, expressed as Among them, x i represents the patient characteristics of the i-th modality data, and the edges of the patient similarity graph represent the similarities between patients; The similarity measure includes: Biomarker similarity metric: Based on the similarity metric of patient biomarkers, the Gaussian similarity function is used to obtain the formula: Among them, S biomarker (i,j) represents the Gaussian similarity measure of the patient’s biomarkers, x i,biomarker 、x j,biomarker Represents the feature vectors of the patient’s biomarkers at nodes i and j, σ biomarker represents the Gaussian kernel width of the patient’s biomarker; Lifestyle similarity measurement: Based on the similarity measurement of lifestyle data, cosine similarity is used to obtain the formula: Among them, S lifestyle (i,j) represents the similarity measure of the patient’s lifestyle, x i,lifestyle 、x j,lifestyle Represent the feature vectors of the patient’s lifestyle at nodes i and j respectively; Medical history similarity measurement: Based on the similarity measurement of the patient's medical history information, Jaccard similarity is used to obtain the formula: Among them, S history (i,j) The feature vector of the patient’s medical history information at nodes i and j, Represent the medical history set of the patient at nodes i and j respectively; Dynamic similarity measurement: Using the dynamic weight adjustment mechanism, the weights of each similarity measurement are dynamically adjusted by learning the weight parameters, and the formula is obtained: Among them, A ij represents the variable weight between node i and node j in the patient similarity graph, x ik represents the kth eigenvalue of node i, x jk represents the kth eigenvalue of node j, K is the dimension of similarity measurement, w k is the weight of the k-th similarity dimension.

4. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: In step 300: The structured data fusion integrates biomarkers, electronic health records, etc., and calculates the similarity between patients through the similarity measurement method; the unstructured data fusion uses natural language processing to encode text data and incorporates the extracted high-dimensional text features into the similarity measurement; the image feature fusion extracts deep features of image data through pre-trained convolutional neural networks, and fuses the deep features with other modal data to enhance the depth and diversity of similarity measurement; The domain knowledge integration includes defining similarities between patients based on clinical pathways and progression trajectories, adjusting similarity measurement weights through expert knowledge guidance, and dividing similarities into different levels, calculating them separately and integrating them into the overall similarity score.

5. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: In step 400: The thinning process includes: By setting a similarity threshold θ, only edges with a similarity higher than θ are retained. The formula is as follows: The K nearest neighbor retention strategy is adopted. For each node, only its most similar K neighbors are retained. The formula is as follows: Among them, Top-K(A i,: ) represents the K nodes with the highest similarity between node i and other nodes; The steps of the normalization process are as follows: Where I is the identity matrix, is the degree matrix, is the normalized adjacency matrix, represents the normalized adjacency matrix, Represents the normalized adjacency matrix The element at row i and column j in .

6. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: Step 500 includes: Step 510: Divide the patient data into multiple time windows according to time. Each time window t corresponds to a graph structure at a moment. In each time window t, recalculate the similarity measure based on the latest patient feature data and update the adjacency matrix A. t ,Through dynamic weight adjustment and multi-source information fusion, it reflects the latest patient characteristics and similarity relationships; Step 520: Combined with the time series information, the graph structure data in the time window sequence is processed by the time series graph convolution network to capture the time series evolution of the patient characteristics. The formula is as follows: in, is the node feature matrix of the l+1th layer, is the normalized adjacency matrix of the t-th time window, is the node feature matrix of the lth layer, W (l) represents the weight matrix of the lth layer; Step 530: Introduce a memory mechanism to serialize the graph convolution outputs of different time windows to capture long-term dependencies and feature evolution trends.

7. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: Step 600 includes: Step 610: Use the self-attention mechanism to enhance the node features and dynamically adjust the importance of different feature dimensions. The formula is as follows: Among them, α ij is the self-attention weight, h j is the feature vector of node j, W represents the weight matrix, h′ i Represents the new feature vector of node i after being enhanced by the self-attention mechanism; Step 620: Add a graph regularization term to the loss function to make the features of similar nodes closer: Among them, ε is the edge set in the graph, h i The feature vector of node i, represents the graph regularization loss term; Step 630: Randomly delete or add edge and node noise to enhance patient similarity graph data, the method includes edge perturbation, node perturbation and feature perturbation.

8. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: Step 700 includes: Step 710: In each layer of the temporal graph convolutional network GCN, a multi-head attention mechanism is used to calculate different attention weights in parallel through multiple attention heads to capture different types of relationship patterns. The formula is as follows: MultiHead(Q,K,V)=Concat(head1,…,head h )W O ; Among them, head i =Attention(Q i ,K i ,V i ), h is the number of attention heads, W O is the output weight matrix, MultiHead(Q,K,V) represents the multi-head attention mechanism, Q is the query vector, K is the key vector, and V is the value vector; Step 720: Generate adversarial samples to improve the robustness of the model to input perturbations. Specific methods include fast gradient sign method and projected gradient descent: Fast Gradient Sign Method: Generate adversarial samples using gradient information: Where x represents the original input sample, x′ represents the adversarial sample, which is generated by adding a small perturbation to the original input sample. represents the loss function measuring the difference between the model prediction and the true label, ∈ represents the perturbation amplitude, Represents the loss function Partial derivative with respect to the original input sample x; Projected Gradient Descent: Iteratively generate adversarial samples, the formula is as follows: in, is the set of constraints for adversarial samples, α is the step size, x t ′ represents the adversarial sample generated by t iterations, x t ' +1 represents the adversarial sample generated by t+1 iterations; Step 730: Performing deep learning integrated and personalized risk assessment to ensure collaborative optimization of the model at the integrated and personalized levels; Step 730 includes the following steps: Step 731: Randomly initialize all model parameters, including GCN parameters, base model parameters, integrated weight coefficient α and personalized attention parameters; Step 732: Optimize the integrated prediction loss and personalized loss to obtain the total loss function for: in, is the loss function for ensemble prediction, is the loss function of personalized prediction, γ is the weight of personalized loss, and λ is the regularization parameter.

9. The personalized hypertensive nephropathy risk assessment method based on graph convolutional network according to claim 1, characterized in that: The specific steps of step 900 are as follows: Step 910: Design a secondary model as the weight coefficient of the learning base model. The final prediction result is the weighted sum of the prediction results of each base model: in, represents the final prediction result, α i represents the weight coefficient of the i-th base model, y i represents the prediction result of the i-th base model, m represents the total number of base models in the ensemble method, α i ≥0; Through back-propagation and optimization algorithms, the weight coefficient α is learned to minimize the loss function of the integrated prediction; Combined with the individual characteristics of the patient, the adaptive attention mechanism is used to predict the risk of each patient. The specific steps include feature extraction, attention weight calculation and personalized weighting. The formula is as follows: e = tanh(W1p+b1); α=softmax(W2e+b2); Among them, e represents the intermediate representation vector obtained after linear transformation and nonlinear activation of the patient's individual features, p is the patient's individual feature vector, W1 and W2 are learnable weight matrices, b1 and b2 are bias vectors, and α is the attention weight; Step 920: Design a personalized loss function, combining the error between the personalized prediction result and the true label, the formula is as follows: Where N is the number of samples, y k is the true label of the kth sample, Predict results for personalization.

10. A personalized hypertensive nephropathy risk assessment system based on graph convolutional networks, applied to the personalized hypertensive nephropathy risk assessment method based on graph convolutional networks as described in claims 1-9, characterized in that: include: Data acquisition and preprocessing module: used to collect modal data related to hypertensive nephropathy, preprocess the modal data, and then establish a patient feature matrix; Patient Similarity Graph Construction Module: It is used to construct a patient similarity graph by measuring and encoding the similarities between patients and enhancing the patient feature representation using a graph convolutional network; Multi-source information fusion module: used for multi-source information fusion and domain knowledge integration; Sparse and normalized processing module: used to perform sparse processing on the graph structure and then normalize the adjacency matrix; Time series processing module: used to process time series data and capture the evolution trend of patient characteristics over time through time series graph convolutional network and long short-term memory network modeling; Feature transfer enhancement module: used to enhance the transfer of important features through self-attention mechanism and graph regularization term; Relationship pattern capture module: It is used to capture different types of relationship patterns using a multi-head attention mechanism in each layer of the temporal graph convolutional network; Parameter optimization module: used to optimize GCN parameters, base model parameters, integration weights, and personalized attention parameters through the Adam optimization algorithm and regularization method; Personalized risk assessment module: It is used to optimize the final risk assessment results by adopting weighted stacking and adaptive attention mechanism and learning personalized loss function.

Images