Cardiac stroke prediction method and device based on knowledge enhancement graph neural network

Abstract

The invention discloses a cardiac stroke prediction method and device based on a knowledge enhancement graph neural network, and belongs to the technical field of medical health information.The method comprises the steps that key medical information is extracted from an electronic medical record EMR, and a patient-symptom-knowledge diagnosis graph is constructed; wherein the categories of the nodes comprise patients, symptoms and knowledge; the edges between the nodes comprise patient-symptom and symptom-knowledge; representing an interaction relationship among the patient, the symptom and the medical knowledge through the patient-symptom-knowledge diagnosis graph; designing a knowledge-perceived graph-enhanced diagnosis network model and performing training; according to the method, expert knowledge and electronic medical record data in the knowledge graph can be effectively integrated, EMR and information in the medical knowledge graph are fully utilized, and the accuracy and robustness of cardiac stroke prediction and diagnosis are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical health information technology, and relates to a medical diagnosis technology assisted by a neural network model, and in particular to an auxiliary disease diagnosis prediction device and method based on a knowledge-enhanced graph neural network. Background Art

[0002] Cardiogenic stroke is a type of stroke caused by blood clots or other substances entering the cerebral blood circulation due to heart problems, causing brain tissue damage. With the widespread use of electronic medical records (EMR) in medical institutions, disease prediction technology based on deep learning has developed rapidly. Existing research work has achieved certain results by combining EMR and medical knowledge to assist diagnosis. However, these methods often ignore the joint structural information in EMR and knowledge graphs (KG) and fail to effectively deal with data sparsity problems. In addition, existing methods often find it difficult to capture the complex relationship between symptoms and diseases when dealing with multidimensional and heterogeneous medical data, which limits the accuracy and depth of diagnosis.

[0003] In existing work, Cheng et al. [1] proposed a deep learning method for phenotyping from patient electronic medical records (EHR). Chu et al. [2] proposed a new context-aware attention mechanism for detecting Adverse Medical Events (AME) of cardiovascular diseases to learn the local context information of words in medical text. Reddy and Delen [3] used the RNN-LSTM method to extract temporal relationships from longitudinal EHR clinical data to predict the probability of readmission within 30 days in lupus patients. Luo et al. [4] used a self-attention-based method to encode temporal information and detect key time steps in the patient's medical history.

[0004] In terms of using graph neural networks to model EHR, GRAM [4] and G-BERT

[20] both build medical ontology graphs based on disease domain knowledge. Choi et al. proposed GCT [5], a graph convolutional transformer, by building graphs of diagnosis, treatment, and laboratory results. Lu et al.

[13] considered the horizontal links in the medical ontology graph and constructed a patient-disease graph to learn hidden disease relationships. Although these methods utilize different EHR data graph structures, they do not explicitly model EHR and knowledge graphs in the same graph, which makes it difficult to establish direct or high-level connections between them.

[0005] There are also some models that focus on combining external knowledge and using external medical knowledge graphs (KGs) to improve representation learning and thus improve prediction performance. For example, some works use the ancestral information of nodes in KGs to improve the quality of medical representation learning, such as GRAM [5], KAME [6]. Or use graph neural networks to capture high-order connection information in KGs and fuse it into the input representation, such as MedPath [7]. However, these works do not consider the data sparsity problem when the amount of data is small, which seriously affects the accuracy of model training and disease diagnosis prediction.

[0006] References: [1] Yu Cheng, Fei Wang, Ping Zhang, and Jianying Hu. 2016. Riskprediction with electronic health records: A deep learning approach. InProceedings of the 2016 SIAM international conference on data mining. SIAM,432–440. [2] Jiebin Chu, Wei Dong, Kunlun He, HuilongDuan, and ZhengxingHuang. 2018. Using neural attention networks to detect adverse medical events from electronic health records. Journal of biomedical informatics 87 (2018), 118–130. [3] Bhargava K Reddy and DursunDelen. 2018. Predictinghospitalreadmission for lupus patients: an RNN-LSTM-based deep-learningmethodology. Computers in biology and medicine 101 (2018), 199–209. [4] Junyu Luo, Muchao Ye, Cao Xiao, and Fenglong Ma. 2020. Hitanet:Hierarchical time-aware attention networks for risk prediction on electronichealth records. In Proceedings of the 26th ACM SIGKDD internationalconferenceon knowledge discovery&data mining. 647–656. [5] Edward Choi, Mohammad Taha Bahadori, Le Song, Walter F Stewart,andJimengSun. 2017. GRAM: graph-based attention model for healthcarerepresentationlearning. In Proceedings of the 23rd ACM SIGKDDinternationalconference onknowledge discovery and data mining. 787–795. [6] Fenglong Ma, Quanzeng You, Houping Xiao, Radha Chitta, Jing Zhou,andJing Gao. 2018. Kame: Knowledge-based attention model for diagnosisprediction in healthcare. In Proceedings of the 27th ACM internationalconference on informationand knowledge management. 743–752. [7] Muchao Ye, Suhan Cui, Yaqing Wang, Junyu Luo, Cao Xiao, andFenglongMa. 2021. Medpath: Augmenting health risk prediction via medicalknowledge paths. In Proceedings of the Web Conference 2021. 1397–1409. Summary of the invention

[0007] In view of the complex causes of cardiogenic stroke, the difficulty in diagnosis and the shortcomings of the above-mentioned prior arts, the present invention provides a cardiogenic stroke diagnosis device and method based on knowledge-enhanced graph neural network, which can make full use of the information in EMR and medical knowledge graphs to improve the accuracy and robustness of diagnosis.

[0008] In order to achieve this purpose, the present invention adopts the following technical solutions: A cardiogenic stroke prediction method based on knowledge-enhanced graph neural network, by constructing a knowledge-aware graph-enhanced diagnosis network model, realizes the prediction and diagnosis of cardiogenic stroke diseases for cardiogenic stroke medical information, and improves the accuracy of diagnosis; the method comprises the following steps: 1) Extract key medical information from cardioembolic stroke-related electronic medical records and construct a patient-symptom-knowledge diagnostic map; The categories of nodes in the patient-symptom-knowledge diagnostic graph include: patient, symptom, knowledge; the edges between nodes include: patient-symptom, symptom-knowledge; the interactive relationship between patients, symptoms and medical knowledge is represented by the patient-symptom-knowledge diagnostic graph; 2) Based on the patient-symptom-knowledge diagnosis graph, a knowledge-aware graph-enhanced diagnosis network model is constructed, including a feature extraction module based on a graph attention network, a feature-aware knowledge extraction module, and a feature perturbation data enhancement module; the input of the model is the patient-symptom-knowledge diagnosis graph; the output is the result of whether the patient has cardioembolic stroke; 21) Feature extraction based on graph attention network includes learning of patient representation features and extraction of symptom attributes; 22) Feature-aware knowledge extraction includes encoding feature-aware guidance signals and performing knowledge extraction and aggregation: a. Encoding feature-aware guidance signals, i.e., encoding the embedded feature representations of target patients and target symptoms as guidance signals; b. Knowledge extraction and aggregation include: Filtering noise on medical knowledge graphs using feature-aware guidance signals; Initialize the relation embedding representation, and then calculate the custom embedding representation; The guidance signal encoding of feature perception is obtained by fusing patient and symptom characteristics; Calculate the attention weight of the symptom to the entity and get the attention score; Normalize the attention scores; Compute latent feature representations of symptom neighbor entities; That is, the information aggregation of the first-order neighbors of the symptom is obtained; 23) Perform feature perturbations on node embeddings in the patient-symptom-knowledge diagnostic graph and perform data augmentation on the feature perturbations; 3) Optimize the knowledge-aware graph-enhanced diagnosis network model to improve the model’s diagnosis accuracy; The loss function L of the knowledge-aware graph-enhanced diagnosis network model is designed as: Among them, CrossEntropy is the cross entropy loss function, It is to calculate L2 regularization for all trainable parameters. It is a hyperparameter used to adjust L2 regularization; is the true label; Predict results for the model; The electronic medical records of the cardioembolic stroke patients to be predicted are preprocessed into a patient-symptom-knowledge diagnostic graph, which is then input into a trained knowledge-aware graph-enhanced diagnostic network model for prediction. The prediction result is whether the patient suffers from cardioembolic stroke.

[0009] In specific implementation, the present invention realizes a cardiogenic stroke diagnosis device based on knowledge-enhanced graph neural network, including: a patient-symptom-knowledge diagnosis graph construction module, a knowledge-aware graph-enhanced diagnosis network model construction module, and a model optimization module; Patient-symptom-knowledge diagnostic graph construction module, the constructed patient-symptom-knowledge diagnostic graph is used to represent the complex interactive relationship between patients, symptoms and medical knowledge; the patient-symptom-knowledge diagnostic graph includes patients and symptoms in the electronic medical record and knowledge entities in the medical knowledge graph (cardioembolic stroke disease, drugs, symptoms, risk factors, examination items, etc.); there are three types of nodes in the patient-symptom-knowledge diagnostic graph: patient, symptom, knowledge; there are two types of edges: patient-symptom, symptom-knowledge; The knowledge-aware graph-enhanced diagnostic network model building module is used to extract the features of the patient-symptom-knowledge diagnostic graph; the knowledge-aware graph-enhanced diagnostic network model building module includes a multi-layer graph convolutional network and an attention mechanism module, wherein the multi-layer graph convolutional network captures the high-order relationship between patients and symptoms through graph convolution operations; the attention mechanism module is used to dynamically assign attention weights to different neighbor nodes according to the representation of patients and symptoms to extract key medical features.

[0010] The model optimization module is used to optimize the knowledge-aware graph-enhanced diagnostic network model to obtain the prediction results of whether a patient suffers from cardiogenic stroke and improve the diagnostic accuracy of the model.

[0011] First, a patient-symptom-knowledge diagnostic graph is constructed. This graph structure integrates patient information, symptoms and knowledge entities in the electronic medical record (EMR) and medical knowledge graph to capture the structural information of the data and the subtle interactions between features. A knowledge-aware graph-enhanced diagnostic network model construction module is designed, including a feature extraction module based on a graph attention network, a feature-aware knowledge extraction module and a feature perturbation data enhancement module. The network is built based on the patient-symptom-knowledge diagnostic graph, and captures the high-order relationship between patients and symptoms through multi-layer graph convolution. The representation of patients and symptoms is used to guide the extraction of knowledge to enhance the accuracy and interpretability of the diagnosis. Subsequently, feature perturbation is used for data enhancement to solve the sparsity problem in medical data and improve the generalization performance of the model. Finally, experiments are conducted on a self-constructed cardioembolic stroke dataset to verify the effectiveness of the method.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention effectively integrates expert knowledge and electronic medical record data in the knowledge graph to construct a patient-symptom-knowledge diagnostic graph. On the patient-symptom-knowledge diagnostic graph, the graph convolutional neural network method is used to learn local information and global information. The knowledge-aware graph-enhanced diagnostic model can perform end-to-end training on knowledge and data at the same time. In the model optimization goal, in addition to optimizing the disease prediction task, supervision information on knowledge relationships (disease comparison learning part and disease-symptom relationship learning part) is also added to ensure that the disease prediction task can effectively utilize knowledge and that knowledge representation is not affected by data noise. In view of the problem that there are many predicted diseases and a limited number of patients corresponding to some diseases, a multi-label hierarchical classification is designed to improve the prediction effect of diseases in small sample categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flowchart of the cardiogenic stroke diagnosis method based on knowledge-enhanced graph neural network provided by the present invention.

[0014] Figure 2A patient-symptom-knowledge graph constructed for the present invention; wherein p represents the patient, f represents the symptom, and v represents the knowledge entity.

[0015] Figure 3 A schematic diagram of the structure of a knowledge-aware graph-enhanced diagnostic network model building module constructed by the present invention; Where p represents the patient, f represents the symptom, and v represents the knowledge entity. Taking patient p1 and symptom f1 as an example, the patient (p * 1) and symptoms (f * 1) The process of final representation. DETAILED DESCRIPTION

[0016] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0017] The present invention provides a cardiogenic stroke diagnosis technology based on a knowledge-enhanced graph neural network, including the steps of constructing a patient-symptom-knowledge diagnosis graph, designing a knowledge-aware graph-enhanced diagnosis network model construction module, a data enhancement module based on feature perturbation, and a model optimization module.

[0018] (1) Constructing the patient-symptom-knowledge diagnostic graph module: Use natural language processing technology to extract key medical information from electronic medical records (EMRs), such as symptoms, diagnosis, treatment, etc., as node features. Construct a patient-symptom interaction graph G_{PF}, where nodes represent patients or symptoms, and edges represent the interaction between patients and symptoms. Construct a medical knowledge graph G_{KG}, which contains medical concepts, relationships, and rules, such as the association between diseases, symptoms, and drugs. Integrate G_{PF} and G_{KG} into a unified graph (patient-symptom-knowledge diagnostic graph) G to form a patient-symptom-knowledge diagnostic graph, which has three types of nodes: patients (p), symptoms (f), and knowledge (v); there are two types of edges: patient-symptom and symptom-knowledge, to capture the complex interaction between patients, symptoms, and medical knowledge.

[0019] (2) Design a knowledge-aware graph-enhanced diagnostic network model building module Based on constructing a patient-symptom-knowledge diagnostic graph, a knowledge-aware graph-enhanced diagnostic network model building module is constructed, including designing a multi-layer graph convolutional network to capture the high-order relationship between patients and symptoms through graph convolution operations. The attention mechanism is introduced to dynamically assign attention weights to different neighbor nodes according to the representation of patients and symptoms to extract key medical features. The representation of patients and symptoms is used to guide the extraction of knowledge in the knowledge graph, enhancing the accuracy and interpretability of the model for disease diagnosis.

[0020] (3) Data enhancement based on feature perturbation The node embeddings (node ​​representations) in the patient-symptom-knowledge diagnostic graph G are perturbed to increase the amount of information and improve the generalization ability of the model. The node embeddings are perturbed using a zero-mean Gaussian distributed noise vector to simulate individual differences and provide more personalized diagnosis. Through feature perturbation, the model can learn more robust feature representations, thereby maintaining high diagnostic performance when faced with sparse or noisy medical data.

[0021] (4) Model optimization In the knowledge-aware graph-enhanced diagnostic network model building module, after L layers of graph convolution, the patient representation features of each layer are added together to obtain the final patient representation features, and then input into a multi-layer perceptron (MLP) for binary classification (whether it is cardioembolic stroke). The cross entropy loss function and L2 regularization are used to optimize the model to avoid overfitting and improve the generalization ability of the model. The model performance is further optimized by adjusting the model parameters and hyperparameters, such as learning rate, number of layers, number of hidden units, etc.

[0022] In specific implementation, the electronic medical record system records the basic information and symptoms of patients with symptoms such as headache and dyspnea. This information is input into the "patient-symptom graph construction module", and the patient's electronic medical record data is integrated to form a patient node, which is connected to the symptom node through edges to build a preliminary patient-symptom graph.

[0023] Next, the preliminary patient-symptom graph is sent to the "cardioembolic stroke knowledge graph construction module" to form a richer patient-symptom-knowledge diagnostic graph by fusing the knowledge entities related to cardioembolic stroke in the medical knowledge graph, such as possible causes, related symptoms, diagnostic criteria, etc., with the patient-symptom graph. The diagnostic graph contains three types of nodes: patients, symptoms, and knowledge, and two types of edges: patients-symptoms and symptoms-knowledge. Subsequently, the patient-symptom-knowledge diagnostic graph is passed to the "knowledge fusion module" to further integrate and process the information in preparation for subsequent diagnostic analysis. The fused information is sent to the "patient-symptom-knowledge diagnostic graph construction module" to convert the fused information into a format that can be processed by the diagnostic network.

[0024] Furthermore, the knowledge-aware graph-enhanced diagnostic network model building module is used for in-depth feature extraction and analysis. The model includes a multi-layer graph convolutional network and an attention mechanism module. The multi-layer graph convolutional network captures the high-order relationship between patients and symptoms through graph convolution operations, while the attention mechanism module dynamically assigns attention weights to different neighbor nodes according to the representation of patients and symptoms to extract key medical features.

[0025] In order to improve the generalization performance of the model, the system perturbs the extracted features in the feature perturbation-based data enhancement module to simulate different patient conditions. This can help the model better learn the complex relationship between patients and symptoms and improve the accuracy of cardiogenic stroke diagnosis.

[0026] Finally, the model system is trained and optimized to obtain the model's predictive analysis results, and then a diagnostic report is generated, including the probability of the patient suffering from cardioembolic stroke, and further examination recommendations and treatment plans are provided. The system's diagnostic results and recommendations can be used as an auxiliary basis, on which the doctor can make the final diagnosis and treatment decision based on his or her own clinical experience.

[0027] Through these modules and processes, the cardiogenic stroke diagnosis system can make full use of the information in electronic medical records and medical knowledge graphs, improve the accuracy and robustness of diagnosis, and provide patients with more effective medical services.

[0028] Experimental results show that the proposed method outperforms the current state-of-the-art technology on cardioembolic stroke datasets and has significant diagnostic accuracy and robustness. Through knowledge graph enhancement and feature perturbation data enhancement, the proposed method can effectively handle data sparsity problems and has made breakthrough progress in the fine-grained diagnosis of cardioembolic stroke. In addition, the interpretability of the method has also been improved, which helps doctors understand the diagnostic basis of the model and enhances the credibility and practicality of the model in clinical applications.

[0029] The specific implementation includes the following steps: (1) Patient-symptom-knowledge diagnostic graph construction module First, a large language model tool (such as ChatGLM) is used to extract keywords (symptoms or diseases that reflect key information in EMR, such as hypertension, obesity, etc.) from EMR as the patient's symptoms, and the patient-symptom interaction graph G PF Defined as: Among them, p and f represent patients and symptoms respectively, P and F represent patient sets and symptom sets respectively, 1 indicates that an interaction between patient p and symptom f is observed; otherwise is 0.

[0030] For the knowledge graph G KG , the calculation formula is as follows: Each triple Describes the relationship r between the head entity h and the tail entity t, Represents a set of nodes in the knowledge graph, Represents the edge set, and finally obtains the unified graph G: in ; w and v represent entities, patients, and symptom nodes in the knowledge graph.

[0031] (2) Design a knowledge-aware graph-enhanced diagnostic network model building module, including a feature extraction module based on a graph attention network, a feature-aware knowledge extraction module, and a feature perturbation data enhancement module; Based on the patient-symptom-knowledge diagnostic graph, the graph attention network (GAT) is used to calculate the representation of patients, symptoms, and knowledge. Patients with similar diseases are connected through common neighbor nodes, thereby capturing the similarities between patients and diseases. For example, two patients diagnosed with the same condition may have different symptoms and medical histories, but GAT can identify the commonalities of their diseases by analyzing the connectivity within the patient-symptom-knowledge diagnostic graph. In addition, this connectivity is not limited to direct symptoms or treatments, but extends to a wider range of disease characteristics, including complications and risk factors. In this way, GAT is able to understand the disease state more carefully, allowing for more accurate prediction and diagnosis.

[0032] ) Feature extraction based on graph attention network, including patient representation feature learning and symptom attribute extraction: 1) Patient representation feature learning: First, structural information is extracted from the patient-symptom interaction graph G_{PF}. Consider patient p, using To represent the neighborhood of patient p on G_{pf}, that is, the symptom set adjacent to patient p. The patient representation feature of patient p in the first layer of the graph attention network can be calculated by the following formula: in, Represents features for patients from layer 1. Considering that different symptoms react to different degrees to the patient's condition, we introduce the attention score To distinguish the importance of each symptom, expressed as: Through the above information propagation process, we can get the patient's characteristic representation .

[0033] 2) Symptom attribute extraction: In order to learn similar information between patients, enhance the generalization ability of the model, and reflect more common medical phenomena. The embedding of symptom f can be calculated by the following formula: in, is the first layer representation of the symptom f in the graph attention network (GAT), In order to distinguish the attention weights of different symptoms in reflecting the importance of disease characteristics, the calculation process is as follows: Through the attention mechanism, the model can identify the most critical symptoms for disease diagnosis, thereby improving the accuracy of diagnosis. For patients with chronic diseases, the attention mechanism can help the model focus on key indicators of disease changes and achieve dynamic monitoring of disease progression.

[0034] ) Feature-aware knowledge extraction, including encoding feature-aware guidance signals and performing knowledge extraction and aggregation: Medical knowledge graphs contain a large number of medical concepts, medical relationships, and medical rules, which can provide rich knowledge for the model. By combining medical knowledge graphs, the model can use existing medical knowledge to assist diagnosis, such as improving the accuracy of diagnosis by identifying standard associations between symptoms and diseases. In addition, the standardized information in the medical knowledge graph helps the model process and unify medical data from different sources and formats, reducing data inconsistency and ambiguity.

[0035] 1) Encoding feature-aware guidance signals: Target patients and target symptoms We can encode them into a guidance signal by the following formula: in is a learnable weight; h is the guidance signal calculation function.

[0036] 2) Perform knowledge extraction and aggregation: We use feature-aware guidance signals to filter noise on the medical knowledge graph G_{KG}. Given the first-order neighbors of f on G_{KG} , we define To express the relationship between f and p By target triple First, initialize the relation embedding representation, and then calculate the custom embedding representation. The calculation method is as follows: in The original relationship is expressed as, It is the relational representation under the guidance of f and p; Represents the element product. Based on this, we calculate the attention weight of symptom f on entity v, expressed as: in, Score for attention; is a trainable transformation matrix, and Transform to the latent space of relation r'; tanh is a nonlinear activation function. Next, we use the softmax function to convert the attention score Normalized, expressed as: in, is the normalized attention score; We then proceed to compute the latent feature representations of the neighboring entities of symptom f as follows: Then, we can get the information aggregation of the symptom's first-order neighbors: 23) Perform data augmentation on feature perturbations to enhance the model’s robustness to outliers: Design a perturbation strategy to perturb the node embeddings in the graph. The perturbation can be random or based on a distribution (such as a Gaussian distribution). The purpose of the perturbation is to simulate the noise and uncertainty in the data and enhance the robustness of the model to outliers. Through feature perturbations, the model can learn to make accurate diagnoses in the face of incomplete or inaccurate data. This is particularly important for processing real medical data, which often contains missing values ​​and noise.

[0037] Assume that any node in the graph is o, and its embedding representation is recorded as , then the characteristic perturbation can be expressed as: in The standard deviation is The noise vector is extracted from the zero-mean Gaussian distribution of , and I is the identity matrix.

[0038] (3) Model optimization to improve the diagnostic accuracy of the model; Design the loss function of the model: Use the cross entropy loss function to train the model to classify cardioembolic stroke. The cross entropy loss function measures the difference between the probability distribution predicted by the model and the probability distribution of the true label. Use L2 regularization to avoid overfitting. L2 regularization limits the complexity of the model by penalizing the sum of squares of the model parameters, thereby improving the generalization ability of the model. Adjust the model's hyperparameters, such as learning rate, number of layers, number of hidden units, etc., through grid search, random search, or Bayesian optimization to find the best model configuration, calculated as follows: Among them, CrossEntropy is the cross entropy loss function, It is to calculate L2 regularization for all trainable parameters. It is a hyperparameter used to adjust L2 regularization; is the true label; Predict the results for the model.

[0039] Through the specific implementation of the above-mentioned cardiogenic stroke diagnosis method, the method of the present invention can make full use of the structured and unstructured information in the medical data to improve the accuracy and robustness of the diagnosis. In addition, the method also has good interpretability, which helps doctors understand the diagnostic basis of the model and enhances the credibility and practicality of the model in clinical applications.

[0040] It should be noted that the purpose of publishing the embodiments is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A cardiogenic stroke prediction method based on knowledge-enhanced graph neural network, characterized in that: The steps include: 1) Extract key medical information from cardioembolic stroke-related electronic medical records and construct a patient-symptom-knowledge diagnostic map; The categories of nodes in the patient-symptom-knowledge diagnostic graph include: patient, symptom, knowledge; the edges between nodes include: patient-symptom, symptom-knowledge; the interactive relationship between patients, symptoms and medical knowledge is represented by the patient-symptom-knowledge diagnostic graph; 2) Based on the patient-symptom-knowledge diagnostic graph, a knowledge-aware graph-enhanced diagnostic network model is constructed, including a feature extraction module based on a graph attention network, a feature-aware knowledge extraction module, and a feature perturbation data enhancement module; 21) Feature extraction based on graph attention network includes learning of patient representation features and extraction of symptom attributes; 22) Feature-aware knowledge extraction includes encoding feature-aware guidance signals and performing knowledge extraction and aggregation: a. Encoding feature-aware guidance signals, i.e., encoding the embedded feature representations of target patients and target symptoms as guidance signals; b. Knowledge extraction and aggregation include: Filtering noise on medical knowledge graphs using feature-aware guidance signals; Initialize the relation embedding representation, and then calculate the custom embedding representation; The guidance signal encoding of feature perception is obtained by fusing patient and symptom characteristics; Calculate the attention weight of the symptom to the entity and get the attention score; Normalize the attention scores; Compute latent feature representations of symptom neighbor entities; That is, the information aggregation of the first-order neighbors of the symptom is obtained; 23) Perform feature perturbations on node embeddings in the patient-symptom-knowledge diagnostic graph and perform data augmentation on the feature perturbations; 3) Optimize the knowledge-aware graph-enhanced diagnostic network model to obtain the prediction results of whether a patient has cardioembolic stroke and improve the diagnostic accuracy of the model; The loss function L of the knowledge-aware graph-enhanced diagnosis network model is designed as: Among them, CrossEntropy is the cross entropy loss function, It is to calculate L2 regularization for all trainable parameters. It is a hyperparameter used to adjust L2 regularization; is the true label; Predict results for the model; The patient-symptom-knowledge diagnostic graph obtained by preprocessing the electronic medical records of the cardioembolic stroke patients to be predicted is input into the trained knowledge-aware graph-enhanced diagnostic network model for prediction, and the prediction result is whether the patient suffers from cardioembolic stroke.

2. The cardioembolic stroke prediction method based on knowledge-enhanced graph neural network as claimed in claim 1, characterized in that: In step 22), target patients and target symptoms The embedded feature representation of is encoded as a guidance signal, which is specifically expressed as: in, is a learnable weight; h is the guidance signal calculation function.

3. The cardiogenic stroke prediction method based on knowledge-enhanced graph neural network as claimed in claim 2, characterized in that: Perform knowledge extraction and aggregation, including: definition express By target triple Guidance; the first-order neighbors of f on G_{KG} are denoted as ; Calculate the custom embedding representation. The calculation method is expressed as: in, represents the element-wise product; the feature-aware guidance signal encoding obtained by fusing patient and symptom features, A relational representation that represents the relation r', the symptom f, and the synergistic information between p and f; Calculate the attention weight of symptom f on entity v, expressed as: in, Score for attention; is a trainable transformation matrix, and Transform to the latent space of relation r'; tanh is a nonlinear activation function; The softmax function is used to normalize the attention score, expressed as: in is the normalized attention score; Calculate the latent feature representation of the neighboring entities of symptom f: That is, the information aggregation of the first-order neighbors of the symptom is obtained, which is expressed as: 。 4. The cardiogenic stroke prediction method based on knowledge-enhanced graph neural network as claimed in claim 3, characterized in that: The characteristic perturbation is expressed as: Where o is a node in the graph; is the embedding representation of the node; The standard deviation is The noise vector is extracted from the zero-mean Gaussian distribution of , and I is the identity matrix.

5. A cardiogenic stroke prediction device based on knowledge-enhanced graph neural network implemented by the method of claim 1, characterized in that: include: Patient-symptom-knowledge diagnostic graph construction module, knowledge-aware graph-enhanced diagnostic network model construction module, and model optimization module; among which: The patient-symptom-knowledge diagnostic graph module is used to represent the interactive relationship between patients, symptoms and medical knowledge; the patient-symptom-knowledge diagnostic graph includes patients, symptoms in the electronic medical record and knowledge entities in the medical knowledge graph; The knowledge-aware graph-enhanced diagnosis network module is used to extract the features of the patient-symptom-knowledge diagnosis graph; the knowledge-aware graph-enhanced diagnosis network module includes a multi-layer graph convolutional network and an attention mechanism module, wherein the multi-layer graph convolutional network captures the high-order relationship between patients and symptoms through graph convolution operations; the attention mechanism module is used to dynamically assign attention weights to different neighbor nodes according to the representation of patients and symptoms, and extract key medical features of cardiogenic stroke diseases; The model optimization module is used to optimize the knowledge-aware graph-enhanced diagnostic network model to obtain the prediction results of whether a patient suffers from cardiogenic stroke and improve the diagnostic accuracy of the model.

6. The cardiogenic stroke prediction device based on knowledge-enhanced graph neural network as claimed in claim 5, characterized in that: The node categories in the patient-symptom-knowledge diagnostic graph are: patient, symptom, knowledge; the edges in the graph are: patient-symptom, symptom-knowledge edges.

Images