Disease prediction system based on heterogeneous information network

Abstract

The application discloses a disease prediction system based on a heterogeneous information network, comprising: an acquisition module configured to acquire a current electronic medical record of a patient to be predicted; a disease prediction module configured to input the current electronic medical record of the patient to be predicted into a trained disease prediction model and output a disease prediction result; wherein the working principle of the disease prediction model is that the current electronic medical record of the patient to be predicted is data augmented, an electronic medical record heterogeneous graph is constructed based on the augmented data; on the electronic medical record heterogeneous graph, an embedding representation of the electronic medical record is performed based on a meta-path learning representation, and aggregation of a meta-path neighbor node is performed through an attention mechanism; and finally, disease prediction of the patient to be predicted is realized. In order to accurately predict whether the patient has a potential disease and make an accurate prediction on the patient.

Description

Technical Field

[0001] This invention relates to the fields of disease diagnosis and heterogeneous information network technology, and in particular to a disease prediction system based on heterogeneous information networks. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] With the widespread adoption of electronic medical records (EMRs), disease prediction has recently garnered significant research attention. Disease prediction requires training an accurate classifier to map input prediction signals (such as symptoms, patient demographics, etc.) to the estimated disease for each patient. However, existing machine learning-based solutions heavily rely on large amounts of manually labeled EMR training data, leading to data insufficiency. Most existing disease prediction methods are based on continuous EMRs collected from individual patients, failing to address patients without historical EMR information, thus reducing their practicality in real-world applications.

[0004] Most existing methods are based on surface phenomena for data analysis to make diagnoses for patients, without delving into the patient's information, integrating and analyzing all patient information, or obtaining more accurate predictions by mining all the information. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a disease prediction system based on heterogeneous information networks. It utilizes the data mining capabilities of heterogeneous information networks to indirectly expand insufficient electronic medical record data and learns highly representative node embedding representations for patients. To address the data deficiency problem, it supplements disease and symptom data by establishing a medical concept graph and a medical record graph based on a medical knowledge base and all electronic medical records. By aggregating information from neighboring nodes and based on symptom information from new patients' electronic medical records, it performs patient embedding representation learning to introduce an attention mechanism to integrate information from different aspects, thereby accurately predicting whether a patient has a potential disease and making accurate predictions for the patient.

[0006] This invention provides a disease prediction system based on heterogeneous information networks;

[0007] Disease prediction systems based on heterogeneous information networks include:

[0008] The acquisition module is configured to acquire the current electronic medical record of the patient to be predicted.

[0009] The disease prediction module is configured to: input the current electronic medical record of the patient to be predicted into the trained disease prediction model and output the disease prediction result;

[0010] The disease prediction model works as follows: it augments the current electronic medical record of the patient to be predicted with data, and constructs an electronic medical record heterogeneous graph based on the augmented data; on the electronic medical record heterogeneous graph, it performs embedded representation of the electronic medical record based on the learning representation of meta-paths, and aggregates neighbor nodes through an attention mechanism; and finally achieves disease prediction for the patient to be predicted.

[0011] Compared with the prior art, the beneficial effects of the present invention are:

[0012] This invention applies heterogeneous information networks to the medical field, proposing a disease prediction method based on heterogeneous information networks. Using patients' electronic medical record (EMR) data, the data is expanded through an external medical knowledge graph, and all EMR data are modeled as a heterogeneous EMR information network. On this heterogeneous EMR graph, meta-path-based representation learning is used to embed patients, and an attention mechanism is used to aggregate neighbor nodes, ultimately leading to a disease prediction result for the patient. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is the overall algorithm framework for Example 1;

[0015] Figure 2 This completes the external knowledge required for Example 1.

[0016] Figures 3(a) to 3(d) This is an example of a heterogeneous electronic medical record information network. Detailed Implementation

[0017] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0020] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0021] Heterogeneous information networks have demonstrated powerful data analysis capabilities in fields such as data mining, link prediction, clustering, and recommendation. However, in-depth research on heterogeneous information networks in the medical and disease fields has not yet been carried out. Since the structure of electronic medical records is very similar to the network pattern of heterogeneous information networks, it is possible to transfer heterogeneous information networks to the field of disease prediction.

[0022] Example 1

[0023] This embodiment provides a disease prediction system based on heterogeneous information networks;

[0024] like Figure 1 and Figure 2 As shown, a disease prediction system based on heterogeneous information networks includes:

[0025] The acquisition module is configured to acquire the current electronic medical record of the patient to be predicted.

[0026] The disease prediction module is configured to: input the current electronic medical record of the patient to be predicted into the trained disease prediction model and output the disease prediction result;

[0027] The disease prediction model works as follows: data is augmented to the current electronic medical record of the patient to be predicted, and a heterogeneous graph of electronic medical records is constructed based on the augmented data; on the heterogeneous graph of electronic medical records, the embedded representation of electronic medical records is performed based on the learning representation of meta-paths, and the neighbor nodes of meta-paths are aggregated through an attention mechanism, so as to finally achieve disease prediction for the patient to be predicted.

[0028] Furthermore, the current electronic medical record of the patient to be predicted is obtained; the current electronic medical record includes the patient's basic information and symptom information. The basic information includes the patient's name, gender, age, etc., and the symptom information includes the detected physiological indicators, such as blood pressure, pulse, etc.

[0029] Furthermore, the disease prediction model includes an encoder; wherein the encoder is implemented using a Heterogeneous Graph Attention Network (HAN).

[0030] Furthermore, the current electronic medical records of the patients to be predicted are augmented with data, and an electronic medical record heterogeneity graph is constructed based on the augmented data; specifically including:

[0031] Construct a medical concept graph based on a medical knowledge base; the medical concept graph contains two types of nodes: symptom and disease name, and each disease node is connected to its corresponding symptom node by an edge;

[0032] Construct a full-staff electronic medical record graph; the electronic medical record graph contains three types of nodes: patient, symptom, and disease. Patients are connected to their corresponding symptoms by edges, and symptoms are connected to their corresponding diseases by edges.

[0033] Based on the medical concept map and the electronic medical record map of all patients, the electronic medical record heterogeneity map of the patients to be predicted is obtained by data augmentation of the electronic medical records.

[0034] Furthermore, after augmenting the current electronic medical record of the patient to be predicted with data and constructing an electronic medical record heterogeneous graph based on the augmented data; and before embedding the electronic medical record into the electronic medical record based on the meta-path learning representation on the electronic medical record heterogeneous graph; it also includes:

[0035] The heterogeneous graph of electronic medical records is decomposed to obtain several bipartite graphs; the neighboring nodes of each node in the bipartite graph are randomly sampled; the sampled information of the neighboring nodes is aggregated to complete the data update.

[0036] Furthermore, the aggregation operation of the sampled information of neighboring nodes specifically includes:

[0037] Based on the first-order neighbor characteristics of each node Calculate weight information, perform information aggregation on the first-order neighbor features of each node, and perform the same operation on all nodes in the graph:

[0038]

[0039] Where sigmoid is the activation function, α i,j Let W be a node and W be a weight coefficient.

[0040] After aggregating the first-order neighbor information of each node, the information from each order is then concatenated to achieve the purpose of information transmission across multiple order nodes. The transmission formula is as follows:

[0041]

[0042] in, This represents the information of the l-th layer of the updated node, where W represents the weight coefficient. This represents the information of the (l-1)th level of the node. This represents the node information at level l before it is updated, and || represents the transmission of information.

[0043] By connecting defined nodes and edges, information is propagated at multiple levels, and the relationships between nodes are inferred through the transmission of information.

[0044] It should be understood that data augmentation is performed using external knowledge. This is done by decomposing the heterogeneous graph into multiple bipartite graphs, sampling the neighbors through the bipartite graphs, and then embedding the sampled information into different structural graphs to complete the data augmentation.

[0045] Furthermore, the embedding representation of electronic medical records on the heterogeneous graph of electronic medical records, based on the learned representation of meta-paths, specifically includes:

[0046] Perform a non-linear activation operation on each bipartite graph;

[0047] Then, the context node entity of each node is embedded into the result of the nonlinear activation operation to obtain the final embedded representation.

[0048] Furthermore, each bipartite graph undergoes a nonlinear activation operation; specifically, the nonlinear activation function used is sigmoid.

[0049] Furthermore, the context node entity of each node is embedded into the result of the nonlinear activation operation, wherein the embedding expression formula is as follows:

[0050]

[0051] Where W represents the projection weight matrix, which is randomly initialized and then iteratively updated and adjusted. Z v h is the final embedding representation of node v. v Let σ represent node information, σ represent the activation function, and G represent the set of all nodes.

[0052] It should be understood that different projection operations are performed on different bipartite graphs. When performing embedding operations in the vector space, a nonlinear activation operation is performed first, and then context entities are used to embed them.

[0053] Furthermore, the aggregation of meta-path neighbor nodes through the attention mechanism specifically includes:

[0054] Through the attention mechanism, the weights of different nodes based on different meta-paths are learned, and the features of different types of nodes are projected onto a unified feature space based on the weights.

[0055] For example, the aggregation of neighbor nodes through the attention mechanism specifically includes:

[0056] The features of different types of nodes are projected onto the same feature space, and an attention mechanism is used to learn the weights between different nodes based on different meta-paths.

[0057] Given a pair of nodes, calculate the probability of the node pair being connected under different meta-paths, and obtain the weight probability information of the first node of the pair under different meta-paths;

[0058]

[0059] Among them, C i,j To calculate the probability, Z i With Z j This represents the patient node and the target node.

[0060] Furthermore, the final goal of predicting the disease of the patient to be predicted includes:

[0061] Learn the neighbor information of nodes under each metapath to obtain the information representation of nodes under the metapath.

[0062] The process of learning the neighbor information of nodes under each meta-path to obtain the information representation of nodes under the meta-path is specifically as follows:

[0063] α T =σ(∑ j∈N α j ·Z v (5)

[0064] Where σ represents a nonlinear activation function, α T This represents the learned meta-path information. Z v For the final embedding representation of node v, α j This represents the weights under different metapaths.

[0065] The learned information is represented as α T Normalization is performed to obtain a standard information representation α, which is then used as the input value for formula (6) to predict the final result:

[0066]

[0067] The information contained in the symptom, disease, and patient nodes is expressed to output a predicted disease.

[0068] Furthermore, the trained disease prediction model; the training process includes:

[0069] Construct a training set; wherein the training set consists of electronic medical records with known disease type labels;

[0070] The training set is input into the disease prediction model to train the model. Training is stopped when the loss function value of the model no longer decreases or the number of iterations reaches the minimum value, and the trained disease prediction model is obtained.

[0071] The loss function used is the negative log-likelihood loss function.

[0072] The negative log-likelihood loss function is expressed as follows:

[0073]

[0074] A ranking function is obtained by using a loss function to determine the prediction result. This ranking function uses the top few results as the prediction result, α. v,v′ The learned node interaction probabilities, The patient's diagnosis is performed based on the updated learned node interaction probabilities. Figures 3(a) to 3(d) This is an example of a heterogeneous electronic medical record information network.

[0075] In this embodiment, disease information is first expanded through an external medical knowledge base to form a heterogeneous electronic medical record graph. In the heterogeneous graph, different neighbor nodes are sampled, and information is aggregated through an attention mechanism. Then, a new vector representation is learned through representation learning, and finally, the patient's disease is predicted.

[0076] In this embodiment, a Heterogeneous Information Attention Network (HAN) encoder is selected as the encoder for the final disease prediction task. It follows neighborhood-based information aggregation rules and emphasizes graph structure information, distinguishing between node-level and semantic-level information during aggregation. The HAN uses both semantic-level and node-level attention mechanisms to simultaneously learn the importance of meta-paths and node neighbors, and obtains the final node information through corresponding aggregation operations. This invention uses information mining from heterogeneous information networks to predict patient diseases.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A disease prediction system based on heterogeneous information networks, characterized by: include: The acquisition module is configured to acquire the current electronic medical record of the patient to be predicted. The disease prediction module is configured to: input the current electronic medical record of the patient to be predicted into the trained disease prediction model and output the disease prediction result; The working principle of the disease prediction model is as follows: the current electronic medical record of the patient to be predicted is expanded with data, and a heterogeneous graph of electronic medical record is constructed based on the expanded data. Specifically, this includes: constructing a medical concept graph based on a medical knowledge base; the medical concept graph contains two types of nodes: symptom and disease name, and each disease node is connected to its corresponding symptom node through an edge. Construct a full-staff electronic medical record graph; the electronic medical record graph contains three types of nodes: patient, symptom, and disease. Patients are connected to their corresponding symptoms by edges, and symptoms are connected to their corresponding diseases by edges. Based on the medical concept map and the electronic medical record map of all staff, the electronic medical records of the patients to be predicted are expanded to obtain the electronic medical record heterogeneity map. The heterogeneous graph of electronic medical records is decomposed to obtain several bipartite graphs; the neighboring nodes of each node in the bipartite graph are randomly sampled; the sampled information of the neighboring nodes is aggregated to complete the data update. On the heterogeneous graph of electronic medical records, the embedded representation of electronic medical records is performed based on the learning representation of meta-paths, and the neighbor nodes of meta-paths are aggregated through the attention mechanism; ultimately, the disease prediction of the patient to be predicted is realized. The aggregation of meta-path neighbor nodes through an attention mechanism specifically includes: The features of different types of nodes are projected onto the same feature space, and the attention mechanism is used to learn the weights between different nodes based on different meta-paths; Given a pair of nodes, calculate the probability of the node pair being connected under different meta-paths, and obtain the weight probability information of the first node of the pair under different meta-paths; （4） in, To calculate the probability, and Represent the patient node and the target node; The ultimate goal is to predict the disease of the patient to be predicted; specifically including: Learn the neighbor information of nodes under each metapath to obtain the information representation of nodes under the metapath; The process of learning the neighbor information of nodes under each meta-path to obtain the information representation of nodes under the meta-path is specifically as follows: （5） in, Represents a non-linear activation function. This represents the learned meta-path information; For nodes The final embedding representation, Indicates the weights under different metapaths; Representation of learned information Perform normalization to obtain a standard information representation. And use it as the input value of formula (6) to predict the final result: （6） The information contained in the symptom, disease, and patient nodes is expressed to output a predicted disease.

2. The disease prediction system based on heterogeneous information networks as described in claim 1, characterized in that, The aggregation operation of the sampled information of neighboring nodes specifically includes: Based on the first-order neighbor characteristics of each node Calculate weight information, perform information aggregation on the first-order neighbor features of each node, and perform the same operation on all nodes in the graph: (1) in, For activation function, These are the weighting coefficients; After aggregating the first-order neighbor information of each node, the information from each order is then concatenated to achieve the purpose of information transmission across multiple order nodes. The transmission formula is as follows: (2) in, This represents the information of the l-th level of the updated node. This represents the information of the (l-1)th layer of the node. This represents the node information of layer l before it is updated. It indicates the transmission of information.

3. The disease prediction system based on heterogeneous information networks as described in claim 1, characterized in that, The embedding representation of electronic medical records on the heterogeneous graph of electronic medical records, based on the learning representation of meta-paths, specifically includes: Each bipartite graph is subjected to a nonlinear activation operation; then, the context node entity of each node is embedded into the result of the nonlinear activation operation to obtain the final embedded representation.

4. The disease prediction system based on heterogeneous information networks as described in claim 3, characterized in that, The context node entity of each node is embedded into the result of the nonlinear activation operation, where the embedding expression formula is as follows: (3) in, This represents the projection weight matrix, which is randomly initialized and then iteratively updated and adjusted. For nodes The final embedding representation, Represents node information, This represents the activation function. This represents the set of all nodes.

5. The disease prediction system based on heterogeneous information networks as described in claim 1, characterized in that, The aggregation of meta-path neighbor nodes through an attention mechanism specifically includes: Through the attention mechanism, the weights of different nodes based on different meta-paths are learned, and the features of different types of nodes are projected onto a unified feature space based on the weights.

6. The disease prediction system based on heterogeneous information networks as described in claim 4, characterized in that, The trained disease prediction model; The training process includes: Construct a training set; wherein the training set consists of electronic medical records with known disease type labels; The training set is input into the disease prediction model to train the model. Training is stopped when the loss function value of the model no longer decreases or the number of iterations reaches the minimum value, and the trained disease prediction model is obtained. The loss function is the negative log-likelihood loss function. The negative log-likelihood loss function is expressed as follows: (7) A ranking function is obtained by using a loss function to determine the prediction results. This ranking function then uses the top few results as the final prediction. The learned node interaction probabilities, The patient's diagnosis is performed based on the updated learned node interaction probabilities.

Images