Readmission risk prediction method and system based on patient diagnosis and treatment disease information

Abstract

The present invention belongs to the field of smart medical technology and provides a method and system for predicting hospital readmission risk based on patient diagnosis and treatment disease information. The method comprises constructing a patient representation graph and performing feature extraction to obtain a representation of the patient representation graph; operating on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, and calculating the contrast loss between the two; constructing a representation of a patient's single hospitalization based on the representation of the patient representation graph, the node sampling view embedding, and the edge sampling view embedding; inputting the representation of the patient's single hospitalization into a readmission prediction model to obtain a predicted label, and calculating the supervised loss between the predicted label and the true label; optimizing the parameters of the readmission prediction model by combining the contrast loss and the supervised loss to obtain a trained readmission prediction model; and performing prediction using the trained model. The present invention can improve data richness, reduce the impact of data differences on the model, and improve model generalization performance.

Description

Technical Field

[0001] The present disclosure relates to the field of smart medical technology, and in particular to a method and system for predicting hospital readmission risk based on patient diagnosis and treatment disease information. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Readmission refers to a patient's need for re-admission to the hospital sometime after completing a hospital treatment due to a recurrence of a disease or complications. This is a significant concern within the healthcare system, as the high rate of readmissions not only challenges the utilization of medical resources but also negatively impacts healthcare efficiency and patients' quality of life. In recent years, with the widespread adoption of electronic health records (EHRs), a vast amount of patient information, including demographics, clinical information, diagnostic results, and treatment history, has been captured. This provides a valuable data foundation for predicting readmissions. Artificial intelligence models can be used to deeply analyze and mine this data, effectively identifying key insights and enabling accurate and reliable readmission prediction. Using readmission prediction models, healthcare professionals can promptly identify patients at risk of readmission within 30 days and implement targeted preventive interventions. These measures can include adjusting medication regimens, strengthening rehabilitation care, and providing health education, all aimed at reducing the likelihood of readmission. This approach not only helps reduce the waste of medical resources and improve healthcare efficiency, but also improves patients' quality of life and mitigates the negative impact of illness on their health and well-being. Therefore, it is very necessary to carry out readmission prediction to provide medical institutions with scientific decision-making support and personalized medical management plans, so as to achieve effective utilization of medical resources and comprehensive management of patient health.

[0004] Existing models for predicting readmissions generally overlook the impact of inter-coding interactions on readmission risk and suffer from poor generalization. First, there are inherent connections between diseases, but most models overlook these connections. For example, some models may find a correlation between cardiovascular disease and readmission, but fail to further explore the combined impact of cardiovascular disease and gastric ulcers on readmission risk. However, these inherent connections are crucial for predicting readmission risk. Second, there are also inherent connections between diagnoses and treatments, and between diseases and treatments, which can have an invaluable impact on readmission risk prediction. A limitation of existing readmission prediction models is that they fail to adequately account for individual patient differences, including physiological conditions, lifestyles, locations, and medical histories. Due to these differences, even if a prediction model performs well on training data, it may not be able to effectively adapt and predict readmission risk in new patient populations. This presents challenges in practical application, especially in environments with high uncertainty, such as hospitals, where the model's generalization is limited. In response to the two problems mentioned above, how to quickly and effectively explore the intrinsic connections between codes, consider the impact of their coexistence on the readmission prediction model, and how to effectively process and utilize the diversity of data between patients to improve the model's generalization ability and predictive performance for different patients have become important tasks in current readmission prediction research. Summary of the Invention

[0005] In response to the above-mentioned defects, the present disclosure proposes a method and system for predicting the risk of readmission based on patient diagnosis and treatment disease information. The co-occurrence information between disease codes and diagnosis and treatment codes is used to mine potential correlation information, and a patient representation graph is constructed based on the potential correlation. A data enhancement method is adopted, and graph comparative learning is used to adaptively capture task-related topological structures. New views are constructed through node and edge sampling to improve data richness, reduce the impact of data differences on the model, and improve the generalization performance of the model.

[0006] In order to achieve the above objectives, the present disclosure adopts the following technical solutions:

[0007] A first aspect of the present disclosure provides a method for predicting hospital readmission risk based on patient diagnosis and treatment disease information, comprising the following steps:

[0008] Construct a patient representation graph based on disease codes and diagnosis and treatment codes, extract features from the patient representation graph, and obtain a representation of the patient representation graph;

[0009] Perform multiple operations on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, and calculate the contrast loss between the two;

[0010] Construct a representation of a single patient admission based on the patient representation graph, node sampling view embedding, and edge sampling view embedding;

[0011] Build a readmission prediction model, input the representation of a patient's single hospitalization into the readmission prediction model, obtain the predicted label, and calculate the supervised loss between the predicted label and the true label;

[0012] The parameters of the readmission prediction model are optimized by combining contrastive loss and supervised loss to obtain a trained readmission prediction model.

[0013] Use the trained readmission prediction model to predict patient readmission.

[0014] As a further implementation, multiple operations are performed on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, specifically:

[0015] performing a data augmentation operation on the representation of the patient representation graph to obtain an enhanced patient representation graph;

[0016] The category-aware graph attention mechanism is used to propagate and aggregate information on the enhanced patient representation graph, and the extracted features are pooled and projected to obtain node sampling view embedding and edge sampling view embedding.

[0017] As a further implementation method, before performing data enhancement, it is also necessary to use the node embedding layer to encode the nodes in the patient representation graph to obtain the embedded representation of the nodes, and use the edge embedding layer to encode the edges in the patient representation graph to obtain the embedded representation of the edges.

[0018] As a further implementation, data augmentation is performed on the representation of the patient representation graph. Specifically, the node embedding representation and the patient representation graph are input into the node sampler to generate a new node embedding. The new node embedding is input into the linear layer to obtain the probability of the node sampling operation on the topological structure, and the Gumbel-Softmax is used for reparameterization to obtain the mask vector indicating whether the node is retained, thereby obtaining a new view of node-level sampling.

[0019] The embedded representations of the nodes at both ends of the edge, the representation of the patient representation graph, and the embedded representation of the edge are input into the edge sampler to generate a new edge embedding. The new edge embedding is input into the linear layer to obtain the probability of the edge sampling operation on the topological structure, and Gumbel-Softmax is used for reparameterization to obtain the mask vector of whether the node is retained, thereby obtaining a new view of edge-level sampling.

[0020] As a further implementation, the enhanced patient representation graph includes node-level sampling new views and edge-level sampling new views.

[0021] As a further implementation, the category-aware graph attention mechanism is used to propagate and aggregate information on the enhanced patient representation graph, specifically:

[0022] Calculate the attention weights between each node and its neighboring nodes in the node-level sampling new view and the edge-level sampling new view respectively;

[0023] Normalize the calculated attention weights across all nodes’ neighbors;

[0024] The information of neighbor entities is summarized in a linear combination using attention weights;

[0025] After performing n layers of information propagation and aggregation, a multi-layer representation of each node is obtained, and the representations of each layer are concatenated into a single vector as the final feature representation of each node in the view.

[0026] As a further implementation method, a patient representation diagram is constructed based on disease codes and diagnosis and treatment codes, specifically:

[0027] Different disease codes and different diagnosis and treatment codes during a single visit are regarded as nodes, and the meanings of the disease codes and diagnosis and treatment codes themselves are used as the characteristics of the nodes themselves to construct a node feature matrix.

[0028] Obtain the co-occurrence information between diseases, diagnoses and treatments, and between diseases and treatments, calculate the mutual information between nodes, and construct an adjacency matrix;

[0029] Construct a patient representation graph based on the node feature matrix and adjacency matrix.

[0030] A second aspect of the present disclosure provides a system for predicting hospital readmission risk based on patient diagnosis and treatment disease information, comprising:

[0031] The representation graph construction module is configured to: construct a patient representation graph according to the disease code and the diagnosis and treatment code, perform feature extraction on the patient representation graph, and obtain a representation of the patient representation graph;

[0032] A contrast loss calculation module is configured to: perform multiple operations on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, and calculate the contrast loss between the two;

[0033] The representation construction module for a single patient admission is configured to: construct a representation of a single patient admission based on the representation of the patient representation graph, the node sampling view embedding, and the edge sampling view embedding;

[0034] A supervised loss calculation module is configured to: build a readmission prediction model, input the representation of a patient's single hospitalization into the readmission prediction model, obtain a predicted label, and calculate a supervised loss between the predicted label and the true label;

[0035] The model training module is configured to: optimize the parameters of the readmission prediction model by combining contrastive loss and supervised loss to obtain a trained readmission prediction model;

[0036] The model prediction module is configured to use the trained readmission prediction model to predict patient readmission.

[0037] A third aspect of the present disclosure provides a medium having a program stored thereon, which, when executed by a processor, implements the steps of the method for predicting the risk of readmission based on patient diagnosis and treatment disease information described in the first aspect of the present disclosure.

[0038] The fourth aspect of the present disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for predicting the risk of readmission based on patient diagnosis and treatment disease information described in the first aspect of the present disclosure are implemented.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention proposes a method for predicting the risk of hospital readmission based on patient diagnosis and treatment information. Since there are inherent connections between diseases, between diagnoses and treatments, and between diseases and diagnoses and treatments that cannot be ignored, it plays an important role in the task of predicting hospital readmission. The present invention uses the co-occurrence information between disease codes and diagnosis and treatment codes in the training data set to calculate the point mutual information value matrix (PMI), realize the mining of potential correlation information between diseases, diagnoses and treatments, and diseases and treatments, and constructs a patient representation map based on the potential correlation information for model training, which increases the cognitive scope of the model and improves the prediction performance of the model.

[0041] (2) The method for predicting readmission risk based on patient diagnosis and treatment disease information of the present invention addresses the low generalization ability of traditional readmission models. Using inherent training data for model training often leads to model failure when inputting unseen patient samples. The method disclosed in the present invention applies a data augmentation operation based on graph contrast learning to the patient representation graph, adaptively capturing the topological structure related to the task, and simultaneously applying node sampling and edge sampling to generate two enhanced views, thereby increasing the richness of the samples, helping the model to cope with unknown patient sample data, mining the commonalities in the data, and improving the generalization ability of the model.

[0042] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which constitute a part of the present disclosure, are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure.

[0044] Figure 1 is a flow chart of a method for constructing a patient representation map in an embodiment of the present disclosure;

[0045] Figure 2 is a flow chart of a data enhancement method based on contrastive learning in an embodiment of the present disclosure;

[0046] Figure 3 4 is a flow chart of a method for predicting hospital readmission in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0047] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0049] In the absence of conflict, the embodiments of the present disclosure and the features thereof may be combined with each other.

[0050] Example 1

[0051] like Figure 1-3 As shown, the first embodiment of the present disclosure provides a method for predicting the risk of hospital readmission based on patient diagnosis and treatment disease information, including the following steps:

[0052] S1 constructs a patient representation graph based on the disease code and diagnosis and treatment code, extracts features from the patient representation graph, and obtains a representation of the patient representation graph;

[0053] S2 performs multiple operations on the representation of the patient representation graph to obtain the node sampling view embedding and the edge sampling view embedding, and calculates the contrast loss between the two;

[0054] S3 constructs the representation of a patient's single hospital admission based on the representation of the patient representation graph, node sampling view embedding, and edge sampling view embedding;

[0055] S4 builds a readmission prediction model, inputs the representation of a patient's single hospitalization into the readmission prediction model, obtains the predicted label, and calculates the supervised loss between the predicted label and the true label;

[0056] S5 combines contrast loss and supervised loss to optimize the parameters of the readmission prediction model to obtain a trained readmission prediction model;

[0057] S6 uses the trained readmission prediction model to predict patient readmission.

[0058] The following describes the technical solution of the present invention using a patient sample in a training set as an example, including the training process of the model.

[0059] First, we collected the co-occurrence statistics of diseases in the entire training dataset and calculated the first point-to-point mutual information (PMI) as an indicator to measure the correlation between diseases. The specific calculation formula of PMI is as follows:

[0060]

[0061] in, represents the probability that diseases i and j appear simultaneously in a single admission record of the same patient, represents the probability of disease i appearing in a patient's single admission record, represents the probability of disease j appearing in a patient's single admission record, N represents the total number of records, N i 、N j Represent the total number of occurrences of disease i and disease j respectively.

[0062] Secondly, the co-occurrence statistics of diagnosis and treatment are collected in the entire training data set, and the second point mutual information (PMI) is calculated as an indicator to measure the correlation between diagnosis and treatment; similarly, the co-occurrence statistics of disease and diagnosis and treatment are collected in the entire training data set, and the third point mutual information (PMI) is calculated as an indicator to measure the correlation between disease diagnosis and treatment. The calculation method is similar to the above. Pay attention to which type of coding is used in the calculation process. After calculating the co-occurrence statistics, for a specific patient, a single visit has a disease coding sequence, expressed as C = {c1,…,c D} and diagnosis and treatment coding sequence M={m1,…,m M}, where D represents the number of disease codes and M represents the number of diagnosis and treatment codes. The patient representation graph is composed of disease codes and diagnosis and treatment codes. The number of disease code nodes |V1| depends on the number of different disease codes during the patient's visit, and the number of diagnosis and treatment code nodes |V2| depends on the number of different diagnosis and treatment codes during the patient's visit. The total number of nodes |V| is the sum of the two. The meaning of the code itself is used as the feature of the node itself, and the node feature matrix X = {x1, x2,…, x |V|}. Determine whether there is a correlation between two nodes based on the PMI value calculated at the beginning. A positive PMI value indicates that a high correlation between codes is shown in the training dataset, while a negative PMI value indicates that the combination of codes has a low correlation. A PMI value of 0 means that the correlation between the pair of codes cannot be reflected in the training dataset. This is more common due to the limited sample size of the training data. In addition, the co-occurrence information between diseases, diagnoses and treatments, and diseases and treatments is taken into account here, so it can be considered that there are three types of edges, represented by 1, 2, and 3. Specifically, the edge between diseases can be expressed as:

[0063]

[0064] PMI is the point mutual information value between disease i and disease j. Nodes i and j here represent disease codes, and edges of type 1 are used to represent edges between diseases.

[0065] The edge between diagnosis and treatment can be expressed as:

[0066]

[0067] Here, nodes i and j both represent diagnosis and treatment codes. The PMI calculation formula is similar, but note that both nodes must be diagnosis and treatment codes, and type 2 edges are used to represent edges between diagnoses and treatments.

[0068] The edge between disease and treatment can be expressed as:

[0069]

[0070] Here, one of the nodes i and j must be a disease code and the other must be a diagnosis and treatment code. The calculation formula for PMI is similar, but note that both the disease code and the diagnosis and treatment code must be present. The edge of type 3 is used to represent the edge between the disease and the diagnosis and treatment.

[0071] In this way, we can determine which coding nodes have edges and obtain the adjacency matrix A. Finally, we can construct a patient representation graph for a single hospitalization based on the adjacency matrix A and the node feature matrix X, that is, G = {X, A}.

[0072] Subsequently, the constructed patient representation graph of the patient is passed through an algorithm in the graph neural network (such as GraphSAGE, GCN, etc.) for information transmission and aggregation to obtain the representation of the patient representation graph Used to guide subsequent graph data augmentation operations. At the beginning of the data augmentation operation, the node embedding layer is used to encode the nodes in the patient representation graph, obtain the embedded representation of the node, and represent it as s v. The edge embedding layer is then used to encode the three types of edges in the patient representation graph to obtain the embedded representation of the edge, which is represented as s e .

[0073] Now we start the formal data enhancement operation. Under the guidance of , two differentiable graph samplers are designed, namely node sampler and edge sampler. For the node sampler, the node is embedded in the representation s v Graph embedding representation of the original patient representation graph Combined, a new node embedding is generated, which is then fed into a linear layer to obtain the probability of node sampling operations on the topological structure. The Gumbel-Softmax reparameterization technique is then used to obtain a mask vector indicating whether node v is retained in the first augmented view, with 1 being retained and 0 being deleted. The first augmented view is generated using the following calculation formula:

[0074]

[0075] Among them, W θ and b θ is a parameter, ⊕ represents the connection operation, m v This is the generated node-level mask vector. At the same time, the edges connected to any deleted nodes will be removed, and the final node sampling will obtain a new node-level view G based on the original sample. v .

[0076] For the edge sampler, the representation after the input embedding layer is obtained by splicing the nodes at both ends of the edge and Graph embedding representation of the original patient representation graph and its own edge embedding representation s e , and get the edge embedding. Then pass it through the linear layer and apply the Gumbel-Softmax reparameterization technique to obtain whether each edge in the graph is sampled, and then generate the second enhanced view. The specific calculation formula is as follows:

[0077]

[0078] in and They are edge e ij The node embedding vectors at both ends, and is a trainable parameter, m ij This is the generated edge-level mask vector. By cropping the edges, we can obtain a new edge-level view G of the original sample. e .

[0079] After acquiring two enhanced views of the patient G v and Ge Afterwards, the Category-Aware Graph Attention Mechanism (CAGAT) is used to control the information propagation process between the two enhanced views. The specific propagation process of CAGAT is as follows:

[0080] In each augmented view, for each node v and its neighbor u∈N(v), where N(v) is the set of neighbors of node v, the attention weight α(v,r,u) between the two nodes is calculated:

[0081]

[0082] where s u and s v are the embeddings of nodes u and v in the CAGAT layer, r is the type embedding of the edge connecting these two nodes, and W v is the trainable projection matrix, W α is a training parameter. To ensure that the gap is not too large, the calculated attention score is normalized among all neighbors of node v using the softmax function, as follows:

[0083]

[0084] Next, the information of the neighbor entity N(v) is summarized in a linear combination using the attention score to obtain the representation of the node v after the message propagation:

[0085]

[0086] After performing n layers of information propagation and aggregation, a multi-layer representation of each node is obtained The representation of each layer is concatenated into a single vector as the final representation of each node in the view. It is then passed through the pooling function and the projection function to obtain the graph-level representation, which is represented as the node sampling view embedding g v and edge sampling view embedding g e .

[0087] After the above operations, we adopt the idea of ​​graph contrastive learning. By designing a contrastive loss function to guide the optimization of data augmentation operations, we aim to maximize the similarity between different sampled views of the patient and reduce the similarity between sampled views of different patients.

[0088] Specifically, the patient representation graph is taken as the i-th graph in its training batch, and g v and g e Re-annotate as g v,i and g e,i , we regard the enhanced views generated from the other i-1 images in the same batch as negative samples, and regard the positive sample pairs of the i-th image (G v,G e ) is expressed as

[0089]

[0090] Where τ is the temperature parameter. Since the order of the embedding representation pairs is changed, the cosine similarity value will also change. Therefore, the final contrast loss is calculated as follows:

[0091]

[0092] After the above steps, by using contrastive learning to guide the graph data enhancement operation, the contrastive loss is obtained, and node sampling and edge sampling are effectively performed.

[0093] By fusing the original graph representation of this patient's admission And the graph representation g of the generated new view v and g e , generate the embedding vector s of the patient's admission c , as follows:

[0094]

[0095] Will s c Send it to MLP for readmission prediction task and get the predicted label y ′ , which is the predicted probability of whether the patient will be readmitted within 30 days. It is compared with the true label y and the cross entropy loss function is used to improve the prediction accuracy in a supervised learning manner:

[0096]

[0097] The supervised loss is obtained, and the final loss function combines the contrast loss and the supervised loss to form the training process of the entire model:

[0098] L t =λ*L sup +(1-λ)*L cl

[0099] Here, λ is a hyperparameter that balances the two loss functions. This loss is continuously optimized to train the model. The trained model is then used to predict patient readmissions.

[0100] Example 2

[0101] A second embodiment of the present disclosure provides a system for predicting hospital readmission risk based on patient diagnosis and treatment disease information, including:

[0102] The representation graph construction module is configured to: construct a patient representation graph according to the disease code and the diagnosis and treatment code, perform feature extraction on the patient representation graph, and obtain a representation of the patient representation graph;

[0103] A contrast loss calculation module is configured to: perform multiple operations on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, and calculate the contrast loss between the two;

[0104] The representation construction module for a single patient admission is configured to: construct a representation of a single patient admission based on the representation of the patient representation graph, the node sampling view embedding, and the edge sampling view embedding;

[0105] A supervised loss calculation module is configured to: build a readmission prediction model, input the representation of a patient's single hospitalization into the readmission prediction model, obtain a predicted label, and calculate a supervised loss between the predicted label and the true label;

[0106] The model training module is configured to: optimize the parameters of the readmission prediction model by combining contrastive loss and supervised loss to obtain a trained readmission prediction model;

[0107] The model prediction module is configured to use the trained readmission prediction model to predict patient readmission.

[0108] The more detailed steps are the same as those in Example 1 and will not be repeated here.

[0109] Example 3

[0110] A third embodiment of the present disclosure provides a medium having a program stored thereon, which, when executed by a processor, implements the steps of the method for predicting the risk of readmission based on patient diagnosis and treatment disease information as described in the first embodiment of the present disclosure.

[0111] The more detailed steps are the same as those in Example 1 and will not be repeated here.

[0112] Example 4

[0113] Embodiment 4 of the present disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for predicting the risk of readmission based on patient diagnosis and treatment disease information as described in Embodiment 1 of the present disclosure are implemented.

[0114] The more detailed steps are the same as those in Example 1 and will not be repeated here.

[0115] The foregoing description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

Claims

1. A method for predicting hospital readmission risk based on patient diagnosis and treatment disease information, characterized in that: The following steps are involved: Construct a patient representation graph based on disease codes and diagnosis and treatment codes, extract features from the patient representation graph, and obtain a representation of the patient representation graph; Construct a patient representation diagram based on disease codes and diagnosis and treatment codes, specifically: Different disease codes and different diagnosis and treatment codes during a single visit are regarded as nodes, and the meanings of the disease codes and diagnosis and treatment codes themselves are used as the characteristics of the nodes themselves to construct a node feature matrix. Obtain the co-occurrence information between diseases, diagnoses and treatments, and between diseases and treatments, calculate the mutual information between nodes, and construct an adjacency matrix; Construct a patient representation graph based on the node feature matrix and adjacency matrix; Perform multiple operations on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, and calculate the contrast loss between the two; Multiple operations are performed on the representation of the patient representation graph to obtain the node sampling view embedding and edge sampling view embedding, specifically: performing a data augmentation operation on the representation of the patient representation graph to obtain an enhanced patient representation graph; The category-aware graph attention mechanism is used to propagate and aggregate information on the enhanced patient representation graph, and the extracted features are pooled and projected to obtain node sampling view embedding and edge sampling view embedding. Construct a representation of a single patient admission based on the patient representation graph, node sampling view embedding, and edge sampling view embedding; Build a readmission prediction model, input the representation of a patient's single hospitalization into the readmission prediction model, obtain the predicted label, and calculate the supervised loss between the predicted label and the true label; The parameters of the readmission prediction model are optimized by combining contrastive loss and supervised loss to obtain a trained readmission prediction model. Use the trained readmission prediction model to predict patient readmission.

2. The method for predicting hospital readmission risk based on patient diagnosis and treatment disease information as claimed in claim 1, characterized in that: Before data enhancement, it is necessary to use the node embedding layer to encode the nodes in the patient representation graph to obtain the embedded representation of the nodes, and use the edge embedding layer to encode the edges in the patient representation graph to obtain the embedded representation of the edges.

3. The method for predicting hospital readmission risk based on patient diagnosis and treatment disease information as claimed in claim 2, characterized in that: Data augmentation is performed on the representation of the patient representation graph. Specifically, the node embedding representation and the patient representation graph representation are input into the node sampler to generate a new node embedding. The new node embedding is input into the linear layer to obtain the probability of the node sampling operation on the topological structure. The mask vector indicating whether the node is retained is obtained by using Gumbel-Softmax reparameterization, thereby obtaining a new view of node-level sampling. The embedded representations of the nodes at both ends of the edge, the representation of the patient representation graph, and the embedded representation of the edge are input into the edge sampler to generate a new edge embedding. The new edge embedding is input into the linear layer to obtain the probability of the edge sampling operation on the topological structure, and Gumbel-Softmax is used for reparameterization to obtain the mask vector of whether the node is retained, thereby obtaining a new view of edge-level sampling.

4. The method for predicting hospital readmission risk based on patient diagnosis and treatment disease information as claimed in claim 3, characterized in that: The enhanced patient representation graph includes new views of node-level sampling and new views of edge-level sampling.

5. The method for predicting hospital readmission risk based on patient diagnosis and treatment disease information as claimed in claim 1, characterized in that: The category-aware graph attention mechanism is used to propagate and aggregate information on the enhanced patient representation graph, specifically: Calculate the attention weights between each node and its neighboring nodes in the node-level sampling new view and the edge-level sampling new view respectively; Normalize the calculated attention weights across all nodes’ neighbors; The information of neighbor entities is summarized in a linear combination using attention weights; After performing n layers of information propagation and aggregation, a multi-layer representation of each node is obtained, and the representations of each layer are concatenated into a single vector as the final feature representation of each node in the view.

6. A readmission risk prediction system based on patient diagnosis and treatment disease information, characterized by: include: The representation graph construction module is configured to: construct a patient representation graph based on the disease code and the diagnosis and treatment code, extract features from the patient representation graph, and obtain a representation of the patient representation graph; construct the patient representation graph based on the disease code and the diagnosis and treatment code, specifically: Different disease codes and different diagnosis and treatment codes during a single visit are regarded as nodes, and the meanings of the disease codes and diagnosis and treatment codes themselves are used as the characteristics of the nodes themselves to construct a node feature matrix. Obtain the co-occurrence information between diseases, diagnoses and treatments, and between diseases and treatments, calculate the mutual information between nodes, and construct an adjacency matrix; Construct a patient representation graph based on the node feature matrix and adjacency matrix; A contrast loss calculation module is configured to: perform multiple operations on the representation of the patient representation graph to obtain node sampling view embedding and edge sampling view embedding, and calculate the contrast loss between the two; Multiple operations are performed on the representation of the patient representation graph to obtain the node sampling view embedding and edge sampling view embedding, specifically: performing a data augmentation operation on the representation of the patient representation graph to obtain an enhanced patient representation graph; The category-aware graph attention mechanism is used to propagate and aggregate information on the enhanced patient representation graph, and the extracted features are pooled and projected to obtain node sampling view embedding and edge sampling view embedding. The representation construction module for a single patient admission is configured to: construct a representation of a single patient admission based on the representation of the patient representation graph, the node sampling view embedding, and the edge sampling view embedding; A supervised loss calculation module is configured to: build a readmission prediction model, input the representation of a patient's single hospitalization into the readmission prediction model, obtain a predicted label, and calculate a supervised loss between the predicted label and the true label; The model training module is configured to: optimize the parameters of the readmission prediction model by combining contrastive loss and supervised loss to obtain a trained readmission prediction model; The model prediction module is configured to use the trained readmission prediction model to predict patient readmission.

7. A medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for predicting the risk of hospital readmission based on patient diagnosis and treatment disease information as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method for predicting the risk of readmission based on patient diagnosis and treatment disease information as described in any one of claims 1 to 5 are implemented.

Images