Personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure

Abstract

The invention discloses a personalized drug recommendation system for rheumatoid arthritis based on a graph embedded hierarchical structure, and the system obtains the doctor-seeing record of a patient diagnosed with rheumatoid arthritis, and carries out the information representation of the patient through the category probability distribution of drugs. Taking drugs as nodes, constructing EHR and DDI matrixes based on the occurrence frequency of drug pairs and the interaction between the drug pairs, processing EHR and DDI graph features through a graph convolutional network, and updating node representation in combination with an adjacent matrix and a feature matrix; and in combination with patient features and graph embedding information, feature enhancement is carried out on drug category probability distribution information and weighted information obtained by DDI and EHR matrixes, feature fusion of multi-level weighting based on attention is realized, and an output result is obtained. According to the invention, the design of a drug map enhancement encoder module and a layered structure decoder is combined, and a personalized drug scheme with an increased remission rate after treatment of a patient with rheumatoid arthritis single diagnosis is recommended.

Description

Technical Field

[0001] The present invention relates to the field of medical health information technology, and in particular to a personalized drug recommendation system for rheumatoid arthritis based on a graph embedding hierarchical structure. Background Art

[0002] Rheumatoid arthritis (RA) is a systemic autoimmune disease that usually recurs and continues to progress, eventually causing disability or even death due to major complications. Among them, disease activity (DA), as one of the important indicators for evaluating RA, decisively affects the doctor's treatment strategy. At present, the treatment of rheumatoid arthritis (RA) is still low in terms of achieving remission and response rates. This is partly because of the current lack of understanding of rheumatoid treatment target strategies, which has led to rheumatologists being unable to fully support treatment based on treatment guidelines. In addition, some treatment methods themselves have failed to effectively improve the remission rate. At the same time, most clinicians and patients have not yet fully accepted the concept of target treatment, resulting in the failure to fully implement the target treatment principle in clinical practice.

[0003] Therefore, it is particularly important to develop personalized and effective clinical treatment plans based on the specific conditions of different patients to ensure the optimal treatment results and improve the patient's quality of life.

[0004] Early work on drug recommendations focused on learning a set of rules from EHRs. Subsequent work used recurrent neural networks to model the sequential dependencies of patients' past visits to improve the effectiveness of drug recommendations. However, these methods did not consider drug interactions, which are critical to minimizing adverse drug reactions.

[0005] At present, the research on drug recommendation models mainly focuses on multi-diseases and diseases with relatively mature pathogenesis research. Among the common multi-disease drug recommendation models, LEAP combines an external drug-drug interaction (DDI) database and uses a recurrent decoder to model the drug-disease relationship in the current visit to recommend a set of drugs. G-Bert uses all data to pre-train diagnosis and drug encoders, but historical drug records are required in the fine-tuning stage. COGNet copies past medication records to current visit information. REFINE recommends modeling the severity of drug interactions and fine-grained drug dosage. However, these networks all rely on historical medication records as input to improve performance. This model based on historical data lacks effective support for patients who have visited a single doctor.

[0006] In summary, existing studies emphasize the feasibility and clinical utility of using machine learning to stratify patients and predict treatment responses in rheumatology. Current drug recommendation models mainly focus on general practice and well-researched diseases, while for chronic diseases such as RA, rational drug selection based on drug interactions remains a key unmet need. Specifically, existing technologies still have the following deficiencies: 1) Insufficient accurate prediction of treatment response rate and remission, over-reliance on multiple diagnostic records Existing treatment options fail to accurately predict treatment response and remission in patients with rheumatoid arthritis and rely on multiple diagnostic records, ignoring the need for recommendations for patients who have only seen a single visit. In real-world healthcare systems, patients who have only seen a single visit cannot be excluded, indicating that this is a key area for improvement.

[0007] 2) Ignoring the deep integration of patient characteristics and the complexity of drug relationships, and insufficient drug relationship modeling capabilities The differences in drug relationships in single disease scenarios have not been fully considered, and the combination of patients' individualized clinical information and drug characteristics has not been deeply explored. Traditional drug recommendation methods often process patient information and drug characteristics separately, simply relying on basic information (such as age, gender, etc.), while ignoring the complex co-occurrence and interaction relationships between patients' clinical manifestations and drugs. Some methods try to combine patient information and drug characteristics through graph embedding encoders, but there are still deficiencies in the deep modeling of drug relationships, especially the inability to handle complex dependencies between drugs.

[0008] 3) Lack of hierarchical selection of drug relationships, confusion and loss in the information aggregation process In the processing of drug-chemical relationships, traditional models fail to distinguish the hierarchical structure of drug categories and their subcategories, resulting in confusion and loss in the process of information aggregation. In the scenario of multi-label drug recommendation, the model fails to fully consider the relationship between drug categories and subcategories, which makes it impossible for the model to make accurate hierarchical decisions in complex drug selection. Therefore, when faced with multiple drug choices, drug recommendation systems often cannot provide personalized and refined suggestions. Summary of the invention

[0009] The present invention aims to address the deficiencies of the prior art and propose a personalized drug recommendation system for rheumatoid arthritis based on a graph embedded hierarchical structure. Combining the drug graph enhanced encoder module and the hierarchical structure decoder design, a personalized drug regimen with an increased remission rate after treatment is recommended for patients with a single diagnosis of rheumatoid arthritis.

[0010] The objective of the present invention is achieved through the following technical solution: a personalized drug recommendation system for rheumatoid arthritis based on a graph embedding hierarchical structure, the system comprising: The feature processing and patient information representation module is used to obtain the medical records of patients diagnosed with rheumatoid arthritis, perform data feature processing, and obtain the probability distribution of the drug categories used by the patients based on the electronic medical record information of the first diagnosis of the disease to represent the patient information; The EHR and DDI matrix module is used to construct a top-level EHR matrix, a bottom-level EHR matrix, and a bottom-level DDI matrix using drugs as nodes and the number of occurrences of drug pairs in the prescription drug set and the interactions between drug pairs as weighted edges; The drug graph embedding encoder module is used to process the EHR and DDI graph features through a graph convolutional network, combine the adjacency matrix and feature matrix, and update the node representation; The hierarchical drug recommendation decoder first fuses the graph embedding representation of the patient's electronic health record (EHR) matrix with the probability distribution of its current drug category to generate a preliminary medication classification result. Then, at the bottom stage, the drug probability distribution information output by the top layer is secondary fused with the graph embedding representation of the underlying EHR data and drug-drug interaction (DDI) network, and finally outputs the optimized drug recommendation result through dual knowledge integration.

[0011] Furthermore, the medical records of patients diagnosed with rheumatoid arthritis were obtained as follows: patients with two consecutive clinical records since the first admission were selected, and the treatment effects before and after medication were compared through disease assessment, and the samples of patients who met the standards were selected.

[0012] Furthermore, the data feature processing is specifically as follows: using variance filtering, recursive feature elimination (RFE) in the wrapping method, recursive feature elimination based on weight sorting, and feature selection method with L2 regularization in the embedding method to filter out the union.

[0013] Furthermore, the patient information is specifically represented as follows: a random forest model is used, with the patient's first diagnosed electronic medical record information as the patient's input feature data, to obtain the drug category probability distribution output corresponding to the patient.

[0014] Furthermore, the number of occurrences of drug pairs in the prescription drug set is calculated in the EHR matrix to calculate the co-occurrence frequency as a weighted edge, and the interactions between drug pairs in the DDI matrix are used as weighted edges in the drug interaction graph.

[0015] Furthermore, the weights in the drug interaction graph represent the severity of drug-drug interactions, and the severity information is classified as major, moderate, minor, or unknown / no reaction.

[0016] Furthermore, the drug graph is embedded in the encoder module, and a two-layer graph convolutional network is used to update the node representation, so that the features of each node can be combined with the information of neighboring nodes, thereby generating a node embedding vector that integrates the graph structure information; and the node embedding vector is fused with the original input features.

[0017] Furthermore, the probability distribution information of the drug category characteristics obtained from the patients is used, and the graph embedding information after the top-level EHR matrix is ​​processed by the graph convolutional network is used for feature enhancement. The top-level classification result combining the patient characteristics and the graph embedding information is obtained through feature weighted fusion based on the attention mechanism. The top-level classification result is input into the bottom-level classifier, and the probability distribution information of the drug category characteristics obtained at the bottom level is enhanced with the weighted information obtained from the bottom-level EHR matrix and the bottom-level DDI matrix to achieve multi-level weighted feature fusion based on attention, and obtain the final output result.

[0018] Furthermore, the predicted value of each drug category label is converted into a probability value and the loss is calculated. The Sigmoid activation function is used to map the value of each drug category label to the interval [0, 1] to obtain the predicted probability of each label.

[0019] Furthermore, the model’s ability to distinguish each label is improved by maximizing the difference between the predicted scores of the correct label and other labels.

[0020] Beneficial effects of the present invention: 1) Improve the accuracy of predicting treatment remission rate for patients with a single diagnosis record Existing drug recommendation models mainly rely on multiple medical records, but for patients with only a single medical record, existing technologies have difficulty providing accurate personalized drug recommendations. The treatment of patients with rheumatoid arthritis usually needs to be adjusted based on multiple treatment responses, but for patients with a single visit, the existing model cannot effectively capture the patient's treatment response, resulting in the recommended drug regimen being insufficient to improve the patient's remission rate. The present invention aims to solve the problem of personalized drug recommendations to improve the remission rate of patients with a single visit after treatment. A representation method based on the data of patients who meet the remission standard after treatment is proposed, and effective features are screened through and adopt a variety of feature selection methods to improve the accuracy of recommendations for patients with a single visit.

[0021] 2) Deeply integrate patient personalized characteristics with drug relationships and enhance drug relationship modeling The drug co-occurrence matrix is ​​constructed using the frequency of a single drug in a drug combination in the treatment regimen of the entire patient in the training set. The drug-drug interaction (DDI) matrix is ​​used to represent the interactions between different drugs, avoiding adverse drug reactions while enhancing the safety of recommendations and enhancing treatment effects.

[0022] In this task, a two-layer graph convolutional network is used to process the graph features of the input EHR and DDI. The representation of the nodes is updated through the adjacency matrix and feature matrix, so that the features of each node can be combined with the features of its neighboring nodes, thereby generating a node embedding vector that integrates the graph structure information. It is then fused with the feature representation based on the model prediction probability to obtain the enhanced information of the graph.

[0023] 3) Constructing a hierarchical selection mechanism for drug relationships Through the hierarchical drug recommendation decoding method, the drug recommendation problem is transformed into a hierarchical classification task, and the top-down use of the category hierarchy relationship is used to improve the model classification performance. The major categories of common first-line RA treatment drugs and their corresponding subcategories are used as the final recommendation labels. The model effectively handles complex multi-label recommendation tasks by refining the classification granularity, solving the problem of information aggregation decision-making in drug recommendation that fails to effectively capture the hierarchical structure and interdependence between major drug categories and subcategories. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 This is a schematic diagram of a personalized drug recommendation system for rheumatoid arthritis based on a graph embedding hierarchical structure provided by the present invention.

[0026] Figure 2 It is a schematic diagram of the model of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with the accompanying drawings and implementation examples. It should be understood that the specific implementation examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0028] like Figure 1 As shown, the present invention provides a personalized drug recommendation system for rheumatoid arthritis based on a graph embedding hierarchical structure, which designs modules such as data preprocessing and feature extraction, patient information representation, graph embedding enhancement (including EHR and DDI matrix modules and drug graph embedding encoder modules), a hierarchical drug recommendation decoder, and loss function design to complete the personalized drug recommendation process; specifically, it includes the following modules: Feature processing and patient information representation module: First, all medical records of patients with confirmed rheumatoid arthritis were extracted, and all outpatient records and hospitalization records with missing admission and discharge diagnoses were excluded. Patients with two consecutive clinical records since the first admission were selected, and the treatment effects before and after medication were compared by condition assessment. The existing results of four indicators such as DAS28CRP, DAS28ESR, CDAI, and SDAI were used as references. According to the evaluation criteria proposed by the European League Against Rheumatism, when the DAS28CRP score is less than 2.6, the DAS28ESR score is less than 2.6, the CDAI score is less than 2.8, and the SDAI score is less than 2.6, it is regarded as disease remission under the definition of each indicator. In order to reflect the average situation under different indicators, the present invention defines any one of the four indicators detected during the second visit as disease remission (target). When the DAS28CRP or DAS28ESR score is less than 5.1, and the difference between the two DAS28CRP or DAS28ESR scores is greater than 0.6, the treatment response is considered. If the difference between any of the two indicators detected during the second visit reaches the standard, it is defined as a response. In order to propose a medication recommendation model with a better prognosis, thereby improving the disease treatment target rate, the present invention selects patients who meet the target as research subjects.

[0029] Taking into account the difficulty of data collection, feature selection methods such as variance filtering, recursive feature elimination (RFE) in the wrapping method, recursive feature elimination based on weight sorting, and L2 regularization in the embedding method are used to screen out the union. The specific process is as follows: Suppose there is a feature matrix , where n is the number of samples and m is the number of features, represents the corresponding j-th feature, represents the specific value of the i-th sample on the j-th feature, represents the mean of the jth feature. The process of variance filtering can be expressed as the following formula. The feature selection rule of variance filtering is to remove features with variance less than a certain threshold. Features, is the variance-preserving feature set.

[0030]

[0031]

[0032] Recursive feature elimination constructs a feature representation learning method based on model prediction probability, and uses the probability information obtained by ensemble learning methods such as random forest as a new feature to enhance the expression and prediction capabilities of downstream models when processing complex medical data. Each time recursively, the selection of features can be expressed as the following formula. represents the k features finally selected, S is the set of features, and L is the loss function. Represents the use of parameters The trained model refers to an integrated model such as random forest. Represents, for example, the number and depth of trees in a random forest.

[0033]

[0034] Weighted RFE is similar to standard RFE, but it takes into account the weighted ranking of features for further optimization. Assume that the model used can assign weights to each feature. .

[0035]

[0036] In the embedding method, L2 regularization (also known as ridge regression) is a feature selection method that optimizes the model by adding a penalty term. The goal of L2 regularization is to minimize the objective function while penalizing the weight of the feature so that the weight of unimportant features is close to zero. Assume that a linear regression model is used for training, n and m represent the total number of samples and the number of features, respectively. is the true label of the i-th sample, is the feature vector of the i-th sample, is the weight of the jth feature. And add the L2 regularization penalty term, is the regularization strength hyperparameter, and the objective function The expression is as follows. Filter out features with non-zero weights for use in the final model.

[0037]

[0038]

[0039] Comprehensive features The selection process is as follows:

[0040] First, the present invention predicts the prediction probability of a given sample X by training a random forest model. The random forest consists of N trees, and each tree t gives a probability output about the drug category Represents the predicted probability of drug category k for the nth patient on tree t:

[0041] The present invention uses the electronic medical record information of the first diagnosis of the disease as input. is the input feature data of the nth patient, is the drug probability distribution output for the nth patient, the probability of each drug category , where k is the number of drug categories. Represents the parameters of the tree t.

[0042]

[0043] EHR and DDI Matrix Module: The drug EHR and DDI matrices are represented as Ge = { }e and Gd = { }, each drug node . Represents the co-occurrence relationship of drugs in drug use plans, with weighted edges The co-occurrence frequency is calculated by counting the number of occurrences of drug pairs in the prescription drug set in the EHR, where the weight Indicates the frequency of co-occurrence between drugs. represents a drug interaction graph, where the weighted edges represents the interaction between any pair of drugs i and j. The weight Indicates the severity of drug interactions. In the DDInter database, severity information is divided into major, moderate, minor, or unknown / no reaction. The present invention uses 3, 2, 1, and 0 to represent these severity levels, and the severity levels are standardized to a range between 0 and 1.

[0044] The present invention involves a top-level EHR matrix , an underlying EHR matrix and an underlying DDI matrix .

[0045] Drug Graph Embedding Encoder Module: Graph convolutional network (GCN) is an effective node representation learning method, which is widely used in graph structured data and can capture local structural information between nodes (such as feature dependencies in clinical data). GCN combines the features of each node with the features of its neighboring nodes through the convolution operation of the adjacency matrix to generate a deep node representation. GCN has been successfully applied to biomedical networks, especially drug-drug interaction (DDI) graphs, modeling drugs as nodes and DDI as the relationship between nodes, thereby enhancing the effectiveness of drug recommendation systems.

[0046] The calculation of GCN usually includes the following steps. The first is the normalization of the adjacency matrix. In order to avoid the dimension problem during feature propagation, GCN usually normalizes the adjacency matrix to eliminate the influence of node degree differences and make the graph convolution operation more balanced. As shown in the following formula, is the original adjacency matrix, which represents the connection relationship between nodes. is the identity matrix, is the degree matrix.

[0047]

[0048] In the task of the present invention, the EHR and DDI graph features are processed by a double-layer GCN. GCN combines the adjacency matrix and the feature matrix to update the node representation so that the features of each node can be combined with the information of the neighboring nodes, thereby generating a node embedding vector that integrates the graph structure information. The node embedding vector processed by GCN will be fused with the original input features. The fusion method can use methods such as splicing or weighted summation to combine the structural information extracted by the graph convolutional network with the original features to enhance the expression ability of the model.

[0049] First, the graph convolution is performed through the normalized adjacency matrix Propagate information about node characteristics. is the trainable parameter matrix in the GCN layer, which is used for feature transformation and dimension mapping. Secondly, through matrix multiplication Update the node features, then introduce nonlinear interactions through the ReLU activation function, and perform a second matrix multiplication with the weights Combined, we finally get the final embedding representation of each node .

[0050]

[0051] The core goal of this invention is to solve the hierarchical modeling problem of drug relationships. By introducing GCN and multi-level classification structures, the performance of the drug recommendation system in complex multi-label scenarios is improved, especially when faced with multiple drug options, accurate decisions can be made and information loss can be effectively avoided.

[0052] Hierarchical drug recommendation decoder: Directly classifying all text categories may face the problem of uneven distribution of category samples and it is difficult to learn the relationship between different types of drugs. Therefore, the present invention transforms the problem into a hierarchical classification problem and improves the classification performance of the model by utilizing the hierarchical relationship between categories, thereby proposing an RA double-layer drug recommendation model.

[0053] Directly classifying or encoding patient information will face problems such as unbalanced distribution of category samples and difficulty in learning the relationship between drug types. Therefore, this study takes drug categories as a benchmark, transforms it into a classification problem, and improves the model classification performance by utilizing the category hierarchy relationship from top to bottom.

[0054] The specific steps are to use the patient to obtain characteristic probability distribution information , using The final embedding representation of the node after GCN double-layer processing , as a graph embedding enhancement module to enhance features . Patient features and graph embedding information are combined through feature weighted fusion based on attention mechanism. Feature enhancement By summing up the graph matrix, It is a custom scale parameter used to adjust the weight ratio of the matrix.

[0055]

[0056]

[0057] Through matrix multiplication Calculate the similarity matrix between patient features and graph embeddings. Apply Softmax to normalize the similarity and use hyperparameters Control the aggregation strength and Add together, retain individual feature details, and get the top-level classification result . Top-level output As input, the features are further extracted by the underlying classifier (such as a linear layer or a shallow GCN).

[0058] The feature probability distribution obtained at the bottom layer The feature enhancement is performed with the weighted information obtained from the DDI and EHR matrices to achieve multi-level weighted feature fusion based on attention, and finally the output result is obtained. .

[0059]

[0060]

[0061] Loss function design module: The loss function is divided into two parts: the first part uses binary cross entropy loss, a common standard loss function in multi-label tasks. In this task, the output of the model is usually unactivated logits, which represent the predicted value of each label. In order to convert these logits into probability values ​​and calculate the loss, the Sigmoid activation function is used to map each logit value to the interval [0, 1] to obtain the predicted probability of each label.

[0062] The second part of the loss function uses multi-label marginal loss. Unlike the traditional multi-category cross entropy loss, multi-label marginal loss does not require mutual exclusivity between labels, but optimizes the performance of the model by focusing on the marginal differences between labels. The design concept of marginal loss is to improve the model's discrimination of each label by maximizing the difference between the predicted score of the correct label and other labels.

[0063] In order to comprehensively consider the advantages of binary cross entropy loss and multi-label margin loss, the weighted sum of the two loss functions is used as the final total loss function. The weighting coefficient is controlled by hyperparameters, and the values ​​of these hyperparameters need to be tuned according to the actual task. The specific loss function is as follows. is the true label of sample i, Represents the predicted output of sample i, which is converted into probability through the Sigmoid function

[0064] Comparison of data processing and model results: The present invention collected data on rheumatoid arthritis patients from 352 clinical hospitals in multiple provinces between 2000 and 2023, excluding patients without drug treatment plans, patients without re-evaluation records after drug treatment, and patients who did not meet the treatment targets after the first treatment, and obtained 4,975 patients who met the targets after a single treatment.

[0065] Taking into account the difficulty of data collection, a total of 44 clinical features were selected from the statistical data of the patients in the cohort, including basic information, diagnosis confirmation, medical history, joint assessment, organ involvement, condition assessment, laboratory tests, etc. Variance filtering, recursive feature elimination (RFE) in the wrapping method, recursive feature elimination based on weight sorting, and L2 regularization in the embedding method were used to select features, and 37 features were left after the union was screened.

[0066] Four major categories of drugs were selected from the common first-line RA treatment drugs, including glucocorticoids, immunosuppressants, special immunosuppressants and nonsteroidal anti-inflammatory drugs (NSAIDs), and the two most frequently used specific drug chemical names under each drug category were selected as the final recommended labels, totaling 8 subcategories.

[0067] Table 1 Comparison of indicators of drug recommendation models

[0068] The model proposed in the present invention is compared with the classic multi-label model CC chain classifier, common machine learning SVM, LightGBM, RF, classic neural network MLP and CNN, and the multi-disease drug recommendation model LEAP that can be based on a single visit. At the same time, the present invention also includes a non-hierarchical model (Our model (-) in Table 1) with a graph embedding enhancement module for comparison. The comparison indicators include three basic indicators, Micro F1, AUC and Micro Jaccard, to evaluate the performance of the classification model, as well as the number of drug recommendations. The model of the present invention (Our model in Table 1) has the best overall performance. The evaluation adopted a five-fold cross validation. The t-test was used to evaluate LEAP and the model proposed in the present invention, and the statistical difference between AUC p<0.05 was significant.

[0069] In order to more intuitively and comprehensively display the results of drug recommendation, the average number of recommended drugs, the average number of additional drugs, and the average number of missed drugs of the present invention and the comparative experiment are compared. It can be seen that although the chain classifier recommends more drugs, it also recommends wrong drugs, resulting in more missed drugs. The model of the present invention misses the least number of drugs, recommends more correct drugs, and achieves better results.

[0070] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A personalized drug recommendation system for rheumatoid arthritis based on a graph embedding hierarchical structure, characterized in that: The system includes: The feature processing and patient information representation module is used to obtain the medical records of patients diagnosed with rheumatoid arthritis, perform data feature processing, and obtain the probability distribution of the drug categories used by the patients based on the electronic medical record information of the first diagnosis of the disease to represent the patient information; The EHR and DDI matrix module is used to construct a top-level EHR matrix, a bottom-level EHR matrix, and a bottom-level DDI matrix using drugs as nodes and the number of occurrences of drug pairs in the prescription drug set and the interactions between drug pairs as weighted edges; The drug graph embedding encoder module is used to process the EHR and DDI feature matrix graphs through a graph convolutional network, combine the adjacency matrix and feature matrix, and update the node representation; At the top-level stage, the hierarchical drug recommendation decoder first fuses the graph embedding representation of the patient's electronic health record (EHR) matrix with the probability distribution of its current drug category to generate a preliminary medication classification result. Then, at the bottom-level stage, the drug probability distribution information output by the top-level is secondary fused with the graph embedding representation of the underlying EHR data and drug-drug interaction (DDI) network, and finally outputs the optimized drug recommendation result through dual knowledge integration.

2. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: The medical records of patients diagnosed with rheumatoid arthritis were obtained as follows: patients with two consecutive clinical records since their first admission were selected, and the treatment effects before and after medication were compared through disease assessment, and the samples of patients who met the standards were selected.

3. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: The data feature processing is specifically as follows: using variance filtering, recursive feature elimination (RFE) in the wrapping method, recursive feature elimination based on weight sorting, and feature selection method with L2 regularization in the embedding method to filter out the union.

4. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: The patient information is specifically represented as follows: a random forest model is used, with the patient's first diagnosed electronic medical record information as the patient's input feature data, to obtain the corresponding patient's drug category probability distribution output.

5. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: In the EHR matrix, the number of occurrences of drug pairs in the prescription drug set is calculated to calculate the co-occurrence frequency as a weighted edge, and in the DDI matrix, the interactions between drug pairs are used as weighted edges in the drug interaction graph.

6. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 5, characterized in that: The weights in the drug interaction diagram represent the severity of the drug-drug interaction, and the severity information is divided into major, moderate, minor, or unknown / no reaction.

7. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: In the drug graph embedding encoder module, a two-layer graph convolutional network is used to update the node representation, so that the features of each node can be combined with the information of neighboring nodes, thereby generating a node embedding vector that integrates the graph structure information; And the node embedding vector will be fused with the original input features.

8. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: The probability distribution information of the drug category characteristics obtained from the patients is used, and the graph embedding information after the top-level EHR matrix is ​​processed by the graph convolutional network is used for feature enhancement. The top-level classification result combining the patient characteristics and the graph embedding information is obtained through the feature weighted fusion based on the attention mechanism. The top-level classification result is input into the bottom-level classifier, and the probability distribution information of the drug category characteristics obtained at the bottom level is enhanced with the weighted information obtained from the bottom-level EHR matrix and the bottom-level DDI matrix to achieve multi-level weighted feature fusion based on attention and obtain the final output result.

9. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: The predicted value of each drug category label is converted into a probability value and the loss is calculated. The Sigmoid activation function is used to map the value of each drug category label to the interval [0, 1] to obtain the predicted probability of each label.

10. The personalized drug recommendation system for rheumatoid arthritis based on graph embedding hierarchical structure according to claim 1, characterized in that: By maximizing the difference between the predicted scores of the correct label and other labels, the model can improve the discrimination of each label.

Images