Method and device for predicting comorbidity risk of chronic diseases based on patient network

Abstract

This invention discloses a method and apparatus for predicting the risk of chronic disease comorbidity based on a patient network. The method includes: acquiring patient personal characteristic information and medical record information, integrating them to obtain a patient disease set; constructing a patient-disease bipartite graph based on the patient disease set, calculating the comprehensive similarity between patients, connecting edges according to a Top-K strategy, and finally constructing a patient network to reflect the correlation between patients; constructing a chronic disease comorbidity prediction model and inputting the patient network and initial feature representations, updating node features through graph convolution fusion attention mechanism, and outputting the target chronic disease probability. This invention is the first to combine a clinically adaptive patient network with a graph neural network to train a model for chronic disease prediction. By modeling patient similarity and disease interaction relationships, it overcomes the shortcomings of traditional methods that ignore the similarity between patients, improves prediction accuracy, and is suitable for early clinical intervention.

Description

Technical Field

[0001] This invention belongs to the field of medical testing and artificial intelligence, and relates to a method and device for predicting the risk of chronic disease comorbidities based on patient networks. Background Technology

[0002] Chronic diseases have a high incidence rate and a high risk of serious complications, posing a continuous challenge to public health systems. In clinical practice, chronic diseases rarely exist in isolation. For example, patients with hypertension often have multiple organ damage, such as damage to the heart, brain, and kidneys, as well as related metabolic diseases. This "comorbidity" phenomenon significantly increases the complexity of health management and the risk to patient prognosis.

[0003] Currently, predictive models for various chronic diseases and their complications mostly rely on traditional machine learning methods (such as logistic regression, support vector machines, etc.) or basic convolutional neural networks (CNNs). While these methods can process structured medical record data to some extent, they generally have several limitations: First, patients are not isolated individuals. Patient groups that share similar medical treatment patterns, demographic characteristics, or disease spectra may imply stronger comorbidity risk associations. Traditional models treat patients as isolated individuals and cannot utilize the group-based, networked association information among patients, resulting in insufficient accuracy of prediction results. Second, medical data is characterized by high dimensionality and sparsity. Directly using raw diagnostic codes (such as ICDs) to construct feature vectors will face the curse of dimensionality and sample sparsity problems, affecting model performance. Third, in the real world, patients with specific comorbidities (case group) are usually far fewer than those without (control group). This severe class imbalance will cause the model to predict biased towards the majority class, reducing its ability to identify high-risk minority classes.

[0004] Graph Neural Networks (GNNs) can effectively process relational data, providing a new approach to modeling complex relationships between patients. However, how to scientifically and rationally construct networks reflecting disease co-occurrence and patient similarity from raw administrative and diagnostic data, and how to design targeted GNN models to address the specific challenges in the medical field, remain current technical challenges. Summary of the Invention

[0005] To overcome the above-mentioned shortcomings, this invention proposes a method and device for predicting the risk of chronic disease comorbidity based on patient networks. By constructing a relationship graph with patients as nodes, the method integrates patient similarity and correlation between patients into a unified learning framework, and uses a graph neural network that integrates attention mechanism and multi-task learning for training, thereby achieving more accurate modeling and chronic disease prediction.

[0006] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0007] This invention provides a method for predicting the risk of chronic disease comorbidities based on patient networks, comprising the following steps:

[0008] S1. Obtain patient characteristic information, construct patient samples, and use whether or not the patient has the target chronic disease as the target label; patient information includes personal characteristic information and medical record information. Integrate the patient's medical record information to form a disease set for each patient.

[0009] S2, construct a patient-disease bipartite graph based on the disease set of each patient, including patient nodes and disease nodes, with edges representing patients suffering from corresponding chronic diseases; calculate the multi-dimensional similarity between patients, and then calculate the comprehensive similarity between patients; construct a patient network, with patients as nodes, and use the TOP-K strategy to connect patient nodes with a comprehensive similarity greater than the set similarity threshold;

[0010] S3, construct a chronic disease comorbidity prediction model, using the patient network and initial patient characteristics as input, to predict the probability of having the target chronic disease; wherein, the patient characteristic information and lifestyle characteristic information obtained in S1 are used as the initial patient characteristic representation;

[0011] S4. Construct training samples, construct a binary cross-entropy loss function, and optimize model parameters; collect feature information of patients to be predicted, construct their patient-disease bipartite graph, calculate the comprehensive similarity between the patients to be predicted and existing patient nodes, integrate the patients to be predicted into the existing patient network, input them into the trained chronic disease comorbidity prediction model, and predict the probability of the target chronic disease.

[0012] Furthermore, in S1, a case group and a control group are constructed based on the target label, and the SMOTE algorithm is used to oversample the group with the smaller sample size in the case group and the control group.

[0013] Furthermore, in S2, multi-dimensional similarity between patients is calculated, and then the comprehensive similarity between patients is calculated. The process is as follows:

[0014] Demographic similarity is calculated based on age difference and gender matching degree. :

[0015] (1)

[0016] in, For indicator functions, As weight, The scaling parameter for age difference. and Let i and j represent the genders of patients respectively. and Let i and j represent the ages of patient i and patient j, respectively.

[0017] Patient and The connection relationships with all disease nodes are represented as vectors. and And calculate their disease cosine similarity. :

[0018] (2)

[0019] in, and Let these represent the disease vectors of patient i and patient j, respectively. Represents the magnitude of a vector.

[0020] Based on patients and patients Each of their own disease sets and Calculate disease Jaccard similarity The formula is as follows:

[0021] (3)

[0022] in, This represents the union of diseases among patients. This indicates the overlap in the diseases among patients.

[0023] Construct and calculate the comprehensive similarity :

[0024] (4)

[0025] in, These are weighting coefficients, and they have... .

[0026] Furthermore, for patients over 60 years of age, it is necessary to increase the dosage. The weights in the overall similarity score are used to reflect the characteristics of multiple comorbidities in the elderly population, therefore... The following age-adaptive adjustment is introduced:

[0027] (5)

[0028] in, It is the basic weight. It is an indicator function. It is the enhancement coefficient. This represents the OR operation in statistics. When... When making adaptive adjustments, remain unchanged. As it changes, .

[0029] Furthermore, in S2, the patient network Includes the patient node set U and the edge set Set a similarity threshold For patients and patients If both In the patient network The middle is the node and Establish edges Among them, the similarity threshold The value ranges from 0 to 1.

[0030] Furthermore, in S3, the chronic disease comorbidity prediction model includes an information propagation module and a decoding prediction module. The information propagation module uses a multi-layer graph convolutional network as its skeleton and adds a hierarchical attention mechanism to each layer to aggregate and update patient node features and obtain the final embedded representation of the patient node. The decoding prediction module is used to decode the final embedded representation and predict the probability that the patient has the target chronic disease.

[0031] Furthermore, the information propagation module includes an input layer, a multi-layer graph convolution and attention aggregation layer, and an output layer;

[0032] The input layer receives the patient network and the initial feature representation of each patient node, and transforms the patient network into an adjacency matrix;

[0033] The aforementioned multi-layer graph convolution and attention aggregation layer is composed of K layers of graph convolution-attention layers stacked together, and the processing flow of each layer includes:

[0034] 1) Through the parameter matrix Target node feature representation and the feature representation of its neighboring nodes Perform a linear transformation to obtain the eigenvectors respectively. and Where k represents the k-th graph convolutional-attention layer; the feature vector after linear transformation and After concatenation, it is combined with the attention vector. Perform a dot product to achieve concatenation and mapping; then apply the result to the LeakyReLU activation function for non-linear activation to obtain the unnormalized attention coefficients. :

[0035] (6)

[0036] In the formula, This indicates a splicing operation.

[0037] The attention coefficients are normalized using the softmax function to obtain the attention weights. :

[0038] (7)

[0039] in, yes The set of neighboring nodes.

[0040] 2) Transfer the feature vectors of neighboring nodes By attention weight After weighted summation, a nonlinear activation function is applied. Perform nonlinear transformation to generate target nodes New feature representation at layer k :

[0041] (8)

[0042] After information propagation through K layers of graph convolution and attention layers, the final embedding representation is obtained. The information propagation module outputs information through the output layer.

[0043] Furthermore, the decoding and prediction module decodes the final embedded representation of each patient node and predicts the target patient. Probability of having a target chronic disease :

[0044] (9)

[0045] in, It is a bilinear weight matrix. It is a bias term.

[0046] Furthermore, in S4, the process of constructing training samples is as follows: for the target chronic disease, the patient samples in the training dataset are divided into multiple subgroups according to different baseline features, and the probability of patients in each subgroup having the target chronic disease is predicted as a subtask for model training; wherein, the baseline features include single predictor chronic diseases, and patient samples with only the baseline features and patient samples with comorbidities including the baseline features are divided into a subgroup.

[0047] Furthermore, in S4, the binary cross-entropy loss function is as follows:

[0048] (10)

[0049] In the formula, N represents the total number of samples. Indicates the number of samples in the subgroup. Indicates target patients The probability of having the target chronic disease. Indicates the patient The target label of the sample.

[0050] Another aspect of the present invention provides a chronic disease comorbidity risk prediction device based on a patient network, comprising:

[0051] The data acquisition and preprocessing module is used to acquire patient characteristic information, construct patient samples, and label them with whether they have the target chronic disease. Patient information includes personal characteristic information and medical record information. The medical record information of patients is integrated to form a disease set for each patient.

[0052] The similarity calculation module is used to construct a patient-disease bipartite graph based on the disease set of each patient, including patient nodes and disease nodes, with edges representing that the patient suffers from the corresponding chronic disease; calculate the multi-dimensional similarity between patients, and then calculate the comprehensive similarity between patients; construct a patient network, with patients as nodes, and connect patient nodes with a comprehensive similarity greater than a set similarity threshold.

[0053] The network construction module is used to build a chronic disease comorbidity prediction model. It uses the patient network and the patient's initial characteristics as input to predict the probability of having the target chronic disease.

[0054] The model training and prediction module is used to construct training samples, build a binary cross-entropy loss function, optimize model parameters, collect feature information of patients to be predicted, construct their patient-disease bipartite graph, calculate the comprehensive similarity between the patients to be predicted and existing patient nodes, integrate the patients to be predicted into the existing patient network, input them into the trained chronic disease comorbidity prediction model, and predict the probability of the target chronic disease.

[0055] Compared with the prior art, the present invention has at least the following beneficial effects:

[0056] This invention scientifically derives a patient-disease bipartite graph to deduce a patient association network and uses a Generative Neural Network (GNN) with an attention mechanism for modeling, fully mining group risk information and significantly improving prediction accuracy. It alleviates sample imbalance through SMOTE, mitigates data sparsity through bipartite graph embedding, and makes the model more suitable for geriatric medicine through age-adaptive weighting. Attention weights can reveal similar patients who contribute most to the prediction, and the bilinear decoder can also provide insights into feature interactions, enhancing the clinical credibility of the model. The model can output the risk probabilities of various chronic diseases through multiple comorbidities, and the network construction process mainly relies on structured administrative and diagnostic data, facilitating deployment and application in medical information systems. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram illustrating the construction of the patient-disease bipartite graph and patient network in an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram illustrating the use of a hierarchical attention mechanism to aggregate neighbor information in the information propagation module of the chronic disease comorbidity prediction model in this embodiment of the invention.

[0060] Figure 4 This is a structural block diagram of the device provided in the embodiments of the present invention. Detailed Implementation

[0061] To further understand the implementation method of the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described. However, the embodiments described below are merely a part of the present invention, and not all of it. The following description is only for further illustrating the advantages and features of the present invention, and not for claiming limitations on the present invention. Other embodiments obtained by those skilled in the art without inventive effort are all within the scope of protection of the present invention.

[0062] The core innovation of this invention lies in calculating multi-dimensional similarity, including personal characteristics and comorbidities, between patient nodes based on a constructed patient-disease bipartite graph to quantify the correlation of patient groups. Based on this correlation, a patient network is scientifically derived, and then trained and used with a graph neural network integrating attention mechanisms and multi-task learning for chronic disease risk prediction. This method utilizes other chronic diseases already present in patients (such as diabetes and coronary heart disease) and the similarities between patients to predict their future risk of developing hypertension, thereby achieving early warning.

[0063] This embodiment provides a method for predicting the risk of chronic disease comorbidities based on patient networks; the process can be found in [reference needed]. Figure 1 It includes the following steps:

[0064] S101, Data Acquisition and Preprocessing.

[0065] Patient data was obtained from the China Health and Retirement Longitudinal Study (CHARLS 2020), specifically from anonymized publicly available patient data. The required patient data included:

[0066] 1) Patient's personal characteristics, including at least age and gender;

[0067] 2) Case record information, including diagnostic records coded according to the International Classification of Diseases (ICD-10) standard; the focus is on extracting diagnostic records of other predictor chronic diseases used to predict the target chronic disease, and recording whether the patient was diagnosed with the target chronic disease in subsequent follow-up, which is used as the target label for model prediction;

[0068] 3) Lifestyle characteristics, including behaviors such as drinking and smoking, as features of the patient.

[0069] By integrating multiple medical records of the same patient and summarizing all their predictive factors for chronic disease diagnosis, a set of characteristic diseases for that patient is formed.

[0070] Patient samples were divided into case and control groups based on the target label, and a training dataset was constructed. The case group included patient samples with the target chronic disease comorbidity, and the control group included patient samples without the target chronic disease comorbidity.

[0071] For example, using hypertension as the target chronic disease, its predictor chronic diseases include diabetes (E10-E14), coronary heart disease (I20-I25), chronic kidney disease (N18), etc.; patient data is collected, and the diagnostic records of predictor chronic diseases are summarized to form a characteristic disease set for the patient; if the patient is diagnosed with hypertension after a certain visit, the target label for that patient sample is marked as... If a case is identified, it should be classified as a case group; otherwise, it should be marked as a case group. They were then grouped into the control group, among which, This represents the target label for the i-th patient sample.

[0072] Normalize the continuous features in the patient data; for example, perform min-max normalization on the age feature.

[0073] To address the issue of imbalanced sample sizes between the case and control groups, the SMOTE algorithm is used to oversample the group with fewer samples, increasing the number of composite samples to balance the two classes. Typically, the control group has significantly fewer samples than the case group; in such cases, oversampling is performed on the control group to increase its sample size.

[0074] S102. Construct a patient-disease bipartite graph and generate a patient network.

[0075] like Figure 2 As shown, a patient-disease bipartite graph is first constructed based on the integrated patient medical record information. This transforms high-dimensional feature relationships into intuitive graph connections; among them, For the set of patient nodes, This is a set of chronic disease nodes, including other predictors of the target chronic disease. Let the set be an edge set, if the patient Diagnosed with a disease Then at the patient node and disease nodes Establish an undirected edge between them ,in, , Indicates the total number of patients. , Represents the total number of predictor chronic diseases, and the set This excludes target chronic diseases, which are used as prediction targets.

[0076] Calculate multidimensional similarity among patients based on the patient-disease bipartite graph of each patient sample:

[0077] Calculate demographic similarity using formula (1) ;Calculate the cosine similarity of diseases using formula (2) ;Calculate disease Jaccard similarity using formula (3) .

[0078] Construct and calculate the comprehensive similarity As shown in formula (4); for patients older than 60 years, the dosage needs to be increased. The weights in the overall similarity are adjusted to better capture the similarity of comorbidity patterns in elderly patients, therefore the weight coefficients in formula (4) are adjusted accordingly. An age-adaptive adjustment is introduced, as shown in formula (5); where the weight coefficients are... and basic weights Within its value range, optimization is performed through traversal, and the optimal value is determined by combining the training and debugging results of the constructed chronic disease comorbidity prediction model. For models requiring further optimization... In the case of adaptive adjustment, and To ensure that the dominant role of disease vectors in similarity calculation is not shaken by age adjustment, only for... Make adjustments to make The enhancement coefficient in formula (5) The optimal value is also determined through parameter tuning during model training.

[0079] Building a patient network It includes the patient node set U and the edge set. ; Set a similarity threshold using the Top-K strategy For example, if K is 50, then the threshold value is... The bit is set to 0.7. For the patient... and patients If both In the patient network Patient nodes in and Establish an edge between them Among them, the similarity threshold The value range is from 0 to 1, for example = 0.7.

[0080] The patient network reflects patient groups similar in terms of predictor disease spectrum and demographic characteristics. To control network density, reduce noisy connections, and highlight strong associations, a Top-K strategy is employed by calculating multi-dimensional comprehensive similarity among patients, retaining only nodes with similarity greater than a similarity threshold. The edges between neighbors are used to construct a sparse weighted undirected graph. This sparse bipartite graph is further compressed into a patient network, transforming the model input from thousands of dimensions of ICD disease codes into a patient network consisting only of patient nodes and similarity edges. The dimensionality of the patient network depends on the number of patients, but is far lower than the total number of disease codes, fundamentally avoiding the curse of dimensionality. This embodiment uses a "patient-disease bipartite graph" and a "patient network" for dimensionality reduction and information condensation, avoiding the direct use of a massive "patient-disease" matrix. Furthermore, through... Metrics such as Jaccard similarity can calculate effective similarity even when the disease sets of two patients have very little overlap, thereby establishing connections. This ensures that valuable associations can be mined from sparse data, alleviating the problem of sample sparsity.

[0081] S103. Construct a predictive model for chronic disease comorbidity.

[0082] The initial features of each patient node are obtained as its initial feature representation. The initial features include the patient's normalized age, gender, association vector with the predictor disease, and lifestyle characteristics. The patient network constructed in step S102 is then used to represent these features. The initial feature representation of each patient node is input into the constructed chronic disease comorbidity prediction model. The chronic disease comorbidity prediction model includes an information propagation module and a decoding prediction module. The information propagation module is used to aggregate and update the features of patient nodes and obtain the final embedded representation of each patient node. The decoding prediction module is used to decode the final embedded representation and predict the probability that the patient has the target chronic disease.

[0083] For example, the information propagation module employs a multi-layer graph convolutional network combined with a hierarchical attention mechanism. This module is an end-to-end neural network structure, primarily composed of a multi-layer graph convolutional network (GCN) as its skeleton, with a hierarchical attention mechanism embedded in each layer. The information propagation module is not a single network layer, but a multi-layered, stackable information processing pipeline including an input layer, multi-layer graph convolutional and attention aggregation layers, and an output layer.

[0084] The input layer receives the patient network and the initial feature representation of each patient node, and transforms the patient network into an adjacency matrix.

[0085] The aforementioned multi-layer graph convolutional and attention aggregation layers, serving as the main body of the information propagation module, are typically composed of K layers (e.g., 2-3 layers) of identical graph convolutional-attention layers stacked together. The processing flow of each layer is as follows: Figure 3 As shown, it includes:

[0086] 1) Calculate the attention coefficient to dynamically evaluate the importance of each neighbor node to the corresponding target node, rather than treating all neighbor nodes equally.

[0087] First, through the trainable parameter matrix Target node feature representation and the feature representation of its neighboring nodes Perform a linear transformation to obtain the eigenvectors respectively. and Where k represents the k-th graph convolutional-attention layer; secondly, the linearly transformed feature vector... and After concatenation, it is combined with the trainable attention vector. Perform a dot product to achieve concatenation and mapping; then apply the result to the LeakyReLU activation function for non-linear activation to obtain the unnormalized attention coefficients. :

[0088] (6)

[0089] In the formula, This indicates a splicing operation.

[0090] The attention coefficients of all neighboring nodes of the target node are normalized using the softmax function to obtain the final, interpretable attention weights. And the sum of the attention weights of all neighboring nodes of the target node is 1:

[0091] (7)

[0092] in, yes The set of neighboring nodes.

[0093] 2) Weighted aggregation of neighbor information: selectively aggregate features of neighbor nodes based on attention weights.

[0094] The neighbor node feature vector after linear transformation By attention weight After weighted summation, a nonlinear activation function is applied. (e.g., ReLU) performs a nonlinear transformation to generate the target node. New feature representation at layer k :

[0095] (8)

[0096] After information propagation through the K-layer graph convolution-attention layer, each patient node obtains a final, high-order embedding representation. The information propagation module outputs information through the output layer. It not only includes the original feature information of the node itself, but also aggregates the feature information of its neighbors, i.e., similar patients, through attention weighting, thus comprehensively reflecting the structural position and risk pattern of the patient in the patient network.

[0097] For example, the decoding and prediction module employs a bilinear decoder to decode the final embedded representation of each patient node and predict the target patient. Probability of having a target chronic disease :

[0098] (9)

[0099] in, It is a trainable bilinear weight matrix. This is the bias term. Bilinear interaction can effectively capture the complex relationship between patient characteristics (such as pre-existing chronic diseases, age, gender, etc.) and the risk of the target chronic disease.

[0100] S104. Optimization and prediction of chronic disease comorbidity prediction model.

[0101] For the target chronic disease, patient samples in the training dataset are divided into multiple subgroups based on different baseline features. Predicting the probability of patients in each subgroup having the target chronic disease is used as a subtask for model training. Specifically, baseline features include single predictor chronic diseases. Patient samples with only the baseline feature and patients with comorbidities including the baseline feature are grouped into a subgroup. For example, for predicting the probability of hypertension comorbidity risk, based on the main predictor chronic disease type, patient samples are divided into a "diabetes baseline group" and a "coronary artery disease baseline group." All patients in the "diabetes baseline group" have diabetes, and all patients in the "coronary artery disease baseline group" have coronary artery disease. Predicting the probability of hypertension risk in different subgroups is used as a subtask. Correspondingly, for different target chronic diseases, subtasks need to be constructed separately to train the model. The trained model is then used only to predict the corresponding target chronic disease.

[0102] Patient networks were constructed from patient samples in each subgroup according to S102 and input into the chronic disease comorbidity prediction model for model training. These subtasks shared the parameters of the information propagation module, but each had its own independent decoder parameters (M,b), so that the model could focus more on the risk patterns of specific groups.

[0103] Construct the following binary cross-entropy loss function L:

[0104] (10)

[0105] In the formula, N represents the total number of samples in the training set, i.e., the total number of patients. This indicates the number of samples in the subgroup.

[0106] A hard example mining optimization strategy is adopted to dynamically adjust the weights of misclassified samples, and the Adam optimizer is used with learning rate decay to optimize model parameters.

[0107] The specific training process is as follows: sample a batch of training data, input it into the model for forward propagation to obtain the predicted probability; calculate the loss for each sample, and apply a hard example mining strategy to dynamically calculate the sample weights based on the loss magnitude. The weighted total batch loss Loss calculated via backpropagation Regarding the gradients of all model parameters, the Adam optimizer uses these gradients, along with its internally maintained momentum and variance, to calculate the specific update amount for each parameter. Based on the current learning rate decay schedule, the actual learning rate for this iteration is obtained; the Adam optimizer uses this learning rate to perform parameter update operations; this process is repeated until the model's performance on the validation set converges or the preset number of training epochs is reached.

[0108] For patients to be predicted, their personal characteristics, medical records, and lifestyle information are collected to construct a patient-disease bipartite graph. The comprehensive similarity between the patient and each patient node in the existing target chronic disease patient network is calculated, and edges are established to integrate the patient into the existing patient network. The updated patient network is then input into the trained chronic disease comorbidity prediction model to obtain the risk probability value of the patient developing the target chronic disease, thereby achieving early risk stratification and clinical intervention.

[0109] The method described in this embodiment was compared with traditional comorbidity-based patient network models (such as the GCN model) using accuracy, F1 score, and AUC as the core evaluation metrics. The tests were conducted on the same dataset using the same training and prediction parameters. Traditional comorbidity-based patient network models only construct connections based on comorbidity relationships between patients, while the method described in this embodiment effectively integrates multi-source clinical information and constructs a reasonable patient network through comprehensive similarity. Test results show that the accuracy of the method described in this embodiment is 69.36%, the F1 score is 65.75%, and the AUC is 0.748; the accuracy of the GCN model is 65.58%, the F1 score is 64.26%, and the AUC is 0.726. The method described in this embodiment improves accuracy by approximately 4%, and both the F1 score and AUC are improved, confirming its superiority. In summary, this invention demonstrates its performance advantages through empirical data. Its scientifically refined network construction method overcomes the limitations of traditional methods, providing reliable support for the application of graph neural networks in disease risk prediction and highlighting its value in clinical decision support.

[0110] Another aspect of this embodiment provides a chronic disease comorbidity risk prediction device based on a patient network, such as... Figure 4 As shown, it includes:

[0111] The data acquisition and preprocessing module is used to acquire patient characteristic information, construct patient samples, and label them with whether they have the target chronic disease. Patient information includes personal characteristic information and medical record information. The medical record information of patients is integrated to form a disease set for each patient.

[0112] The similarity calculation module is used to construct a patient-disease bipartite graph based on the disease set of each patient, including patient nodes and disease nodes, with edges representing that the patient suffers from the corresponding chronic disease; calculate the multi-dimensional similarity between patients, and then calculate the comprehensive similarity between patients; construct a patient network, with patients as nodes, and connect patient nodes with a comprehensive similarity greater than a set similarity threshold.

[0113] The network construction module is used to build a chronic disease comorbidity prediction model. It uses the patient network and the patient's initial characteristics as input to predict the probability of having the target chronic disease.

[0114] The model training and prediction module is used to construct training samples, build a binary cross-entropy loss function, optimize model parameters, collect feature information of patients to be predicted, construct their patient-disease bipartite graph, calculate the comprehensive similarity between the patients to be predicted and existing patient nodes, integrate the patients to be predicted into the existing patient network, input them into the trained chronic disease comorbidity prediction model, and predict the probability of the target chronic disease.

[0115] The embodiments described above are merely illustrative of implementation methods of the present invention and should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for predicting the risk of chronic disease comorbidities based on patient networks, characterized in that, include: S1. Obtain patient characteristic information, construct patient samples, and use whether or not the patient has the target chronic disease as the target label; Patient information includes personal characteristics and medical records. The medical records of each patient are integrated to form a set of diseases for each patient. S2, construct a patient-disease bipartite graph based on the disease set of each patient, including patient nodes and disease nodes, with edges representing that the patient suffers from the corresponding chronic disease; Calculate the multi-dimensional similarity between patients, and then calculate the comprehensive similarity between patients; construct a patient network, with patients as nodes, and connect patient nodes whose comprehensive similarity is greater than the set similarity threshold; S3, construct a chronic disease comorbidity prediction model, using the patient network and initial patient characteristics as input, to predict the probability of having the target chronic disease; wherein, the patient characteristic information and lifestyle characteristic information obtained in S1 are used as the initial patient characteristic representation; S4. Construct training samples, construct a binary cross-entropy loss function, and optimize model parameters; collect feature information of patients to be predicted, construct their patient-disease bipartite graph, calculate the comprehensive similarity between the patients to be predicted and existing patient nodes, integrate the patients to be predicted into the existing patient network, input them into the trained chronic disease comorbidity prediction model, and predict the probability of the target chronic disease.

2. The method according to claim 1, characterized in that, In S1, a case group and a control group are constructed based on the target label, and the SMOTE algorithm is used to oversample the case group and the control group with a smaller sample size.

3. The method according to claim 1, characterized in that, In S2, the calculation process for multi-dimensional similarity and comprehensive similarity between patients is as follows: Calculate demographic similarity : （1） in, For indicator functions, As weight, The scaling parameter for age difference. and Let i and j represent the genders of patients respectively. and Let i and j represent the ages of patient i and patient j, respectively. Patient and The connection relationships with all disease nodes are represented as vectors. and And calculate their disease cosine similarity. : （2） in, Represents the magnitude of a vector; Based on patients and patients Each of their own disease sets and Calculating disease Jacard similarity The formula is as follows: （3） in, This represents the union of diseases among patients. Indicates the overlap of diseases among patients; Construct and calculate the comprehensive similarity : （4） in, , and These are weighting coefficients, and they have... .

4. The method according to claim 3, characterized in that, For patients over 60 years of age, yes, yes Introducing age-adaptive adjustment to increase Weight: （5） in, It is the basic weight. It is an indicator function. It is the enhancement coefficient. This represents the OR operation in statistics; when When making adaptive adjustments, remain unchanged. As it changes, .

5. The method according to claim 1, characterized in that, In S3, the chronic disease comorbidity prediction model includes an information propagation module and a decoding prediction module. The information propagation module uses a multi-layer graph convolutional network as the skeleton and adds a hierarchical attention mechanism to each layer to aggregate and update the patient node features and obtain the final embedded representation of the patient node. The decoding prediction module is used to decode the final embedded representation and predict the probability that the patient has the target chronic disease.

6. The method according to claim 5, characterized in that, The information propagation module includes an input layer, a multi-layer graph convolution and attention aggregation layer, and an output layer; The input layer receives the patient network and the initial feature representation of each patient node, and transforms the patient network into an adjacency matrix; The aforementioned multi-layer graph convolution and attention aggregation layer is composed of K layers of graph convolution-attention layers stacked together, and the processing flow of each layer includes: 1) Through the parameter matrix Target node feature representation and the feature representation of its neighboring nodes Perform a linear transformation to obtain the eigenvectors respectively. and Where k represents the k-th graph convolutional-attention layer; and After concatenation, it is combined with the attention vector. Perform a dot product; then apply a non-linear activation to the result to obtain the attention coefficient. The attention weights are obtained after normalization. ; 2) Transfer the feature vectors of neighboring nodes By attention weight After weighted summation, a nonlinear transformation is performed to generate the target node. New feature representation at layer k ; After information propagation through K layers of graph convolution and attention layers, the final embedding representation is obtained. It is output through the output layer.

7. The method according to claim 5 or 6, characterized in that, The decoding prediction module provides the final embedded representation of the patient node. Decode and predict the target patient Probability of having a target chronic disease : （9） in, It is a bilinear weight matrix. It is a bias term.

8. The method according to claim 1, characterized in that, In S4, the training sample construction process is as follows: for the target chronic disease, the patient samples in the training dataset are divided into multiple subgroups according to different baseline features. Predicting the probability of each subgroup of patients having the target chronic disease is used as a subtask for model training. The baseline features include single predictor chronic diseases. Patient samples with only the baseline features and patients with comorbidities including the baseline features are divided into a subgroup.

9. The method according to claim 1, characterized in that, In S4, the binary cross-entropy loss function is as follows: （10） In the formula, N represents the total number of samples. Indicates the number of samples in the subgroup. Indicates target patients The probability of having the target chronic disease. Indicates the patient The target label of the sample.

10. A chronic disease comorbidity risk prediction device based on a patient network, used to implement the method as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to acquire patient characteristic information, construct patient samples, and label them with whether or not they have the target chronic disease. Patient information includes personal characteristics and medical records. The medical records of each patient are integrated to form a set of diseases for each patient. The similarity calculation module is used to construct a patient-disease bipartite graph based on the disease set of each patient, including patient nodes and disease nodes, with edges representing that the patient suffers from the corresponding chronic disease; Calculate the multi-dimensional similarity between patients, and then calculate the comprehensive similarity between patients; construct a patient network, with patients as nodes, and connect patient nodes whose comprehensive similarity is greater than the set similarity threshold; The network construction module is used to build a chronic disease comorbidity prediction model. It uses the patient network and the patient's initial characteristics as input to predict the probability of having the target chronic disease. The model training and prediction module is used to construct training samples, build a binary cross-entropy loss function, optimize model parameters, collect feature information of patients to be predicted, construct their patient-disease bipartite graph, calculate the comprehensive similarity between the patients to be predicted and existing patient nodes, integrate the patients to be predicted into the existing patient network, input them into the trained chronic disease comorbidity prediction model, and predict the probability of the target chronic disease.

Images