A Drug Combination Recommendation Method and System Based on Hierarchical Hypergraph Convolutional Networks

Abstract

This invention discloses a drug combination recommendation method and system based on a hierarchical hypergraph convolutional network, belonging to the field of machine learning technology. The method includes: inputting a test instance and a historical medical visit sequence into a drug combination recommendation model constructed from a hierarchical hypergraph convolutional network to obtain recommendation scores for different drugs for the test instance; and obtaining a drug combination recommendation result for the test instance based on the recommendation scores of different drugs for the test instance. This invention can generate drug combination recommendation results more accurately.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and in particular to a method and system for recommending drug combinations based on hierarchical hypergraph convolutional networks. Technical Background

[0002] In the field of personalized medicine, drug combination recommendations aim to provide patients with individualized treatment plans, selecting appropriate drug combinations based on their medical history and current condition. Compared to traditional standardized treatments, this approach can provide more effective treatment plans tailored to each patient's unique medical history, thereby significantly improving treatment outcomes.

[0003] In recent years, the vast accumulation of Electronic Health Records (EHRs) has provided rich data support for personalized drug combination recommendations. These EHR data typically include detailed records of patients' demographic and clinical information, covering multiple patient visits, diagnoses, surgical procedures, and medication use. Researchers model the complete patient medical record sequence by mining temporal patterns and disease progression trends. While these methods have improved the personalization of drug combination recommendations to some extent, most methods are limited to capturing simple sequential relationships of medical events, failing to fully consider the complex interactions and dependencies between diagnoses, surgeries, and medications. Furthermore, patient medical records contain various types of diagnoses, surgeries, and potential complications, but existing methods often cannot adequately model the multidimensional attributes of medical events.

[0004] Furthermore, Graph Neural Networks (GNNs) are widely used in drug combination recommendation due to their ability to model structured data and capture complex dependencies. However, most GNN-based research focuses on traditional graph structures and pairwise relationships, making it difficult to effectively capture higher-order connections between medical entities. This limitation leads to the omission of crucial information about complex patterns and potential dependencies, further restricting recommendation performance and failing to comprehensively model the complex conditions of patients.

[0005] The existing patent solution, "A Drug Recommendation Method and System Based on Heterogeneous EHR Network Representation Learning," models the relationship between patients and drugs using ordinary graphs and aggregates information using graph neural networks. However, ordinary graphs can only represent binary relationships between entities and cannot effectively represent complex higher-order connections and combinatorial effects. This invention employs an innovative hierarchical hypergraph convolutional network to model higher-order relationships between patients and drugs through a hierarchical hypergraph structure. Hypergraphs can not only represent binary relationships but also capture multi-faceted relationships and more complex interactions between entities, thereby improving the accuracy and personalization of drug combination recommendations.

[0006] In summary, existing methods have significant shortcomings in multidimensional feature representation, complex interaction capture, and high-order relationship modeling. In particular, their ability to comprehensively uncover potential dependencies between medical entities and handle the complexity of patient medical records still needs improvement, making it difficult to fully meet the actual needs of personalized drug combination recommendations. Summary of the Invention

[0007] To overcome the limitations of existing technologies, this invention proposes a drug combination recommendation method and system based on hierarchical hypergraph convolutional networks, which can generate drug combination recommendation results more accurately.

[0008] To achieve the above objectives, the technical solution of the present invention includes the following:

[0009] A drug combination recommendation method based on hierarchical hypergraph convolutional networks, the method comprising:

[0010] The diagnostic and surgical sets of the test instances are input into the drug combination recommendation model to obtain a recommendation score for the drug combination; wherein, the drug combination recommendation model is used for:

[0011] The embedding representations of patient hyperedges and visit hyperedges in a hierarchical hypergraph are computed; wherein, the hierarchical hypergraph includes: entity hyperedges, visit hyperedges and patient hyperedges, the entity hyperedges are constructed based on the diagnosis set, surgery set and drug set in the historical medical records of patient instances, the visit hyperedges are entity hyperedges belonging to the same visit, and the patient hyperedges are all visit hyperedges connecting the same patient;

[0012] Based on the diagnostic set and surgical set of the test instance, an embedded representation of the test instance is generated;

[0013] The recommendation score for the drug combination is obtained based on the similarity between the embedding representation of the test instance and the embedding representation of the patient hyperedge and the access hyperedge.

[0014] Based on the recommended score of the drug combination, the recommended drug combination for this test instance is obtained.

[0015] Further, the generation of an embedded representation of the test instance based on the diagnostic set and surgical set of the test instance includes:

[0016] Based on the diagnosis number and surgery number of the test instance, the diagnosis set and surgery set are one-hot encoded respectively to obtain the diagnosis one-hot vector and the encoding one-hot vector;

[0017] Obtain the learned embedding matrix;

[0018] The learned embedding matrix is ​​multiplied by the diagnostic one-hot vector and the encoded one-hot vector to obtain the diagnostic embedding representation and the surgical embedding representation;

[0019] The diagnostic embedding representation and the surgical embedding representation are respectively input into the capsule network to obtain the output value of each capsule, and the output values ​​of the corresponding capsules are concatenated to obtain the capsule network diagnostic output and capsule network surgical output.

[0020] By concatenating the capsule network diagnostic output and the capsule network surgical output, an embedded representation of the test instance is obtained.

[0021] Furthermore, the embedded representation of access hyperedges in the computational hierarchical hypergraph includes:

[0022] Initialize the feature matrix of entity nodes in the drug combination recommendation hierarchy hypergraph; wherein, the entity nodes include: diagnostic nodes, surgical nodes, and drug nodes;

[0023] A self-attention mechanism is used to convolve different entity nodes inside each entity hyperedge to obtain an updated entity node feature matrix; wherein, the entity hyperedges include: diagnostic hyperedges, surgical hyperedges and drug hyperedges;

[0024] The updated entity node feature matrix is ​​subjected to average pooling to obtain the embedding representation of each entity hyperedge;

[0025] We weight the embedding representations of the diagnostic, surgical, and drug hyperedges for the same visit to obtain the embedding representation of the visit hyperedge.

[0026] Furthermore, the embedded representation of the patient hyperedge in the computational hierarchical hypergraph includes:

[0027] Introduce a time decay factor to all access superedges belonging to the same patient instance. Where t j λ is the time interval between the j-th access and the current time, and λ is a hyperparameter that controls the decay rate.

[0028] According to the time decay factor γ j Calculate the weight of each visited hyperedge;

[0029] The embedding representation of the patient's hyperedge is obtained by weighting the access hyperedges by combining their weights.

[0030] Further, the step of obtaining a recommendation score for the drug combination based on the similarity between the embedding representation of the test instance and the embedding representations of the patient hyperedge and the access hyperedge includes:

[0031] Calculate the similarity score Q between the embedding representation of the test instance and the embedding representations of all visited superedges. v And based on the first set threshold, the first preliminary recommended drug set for this test instance is obtained;

[0032] Calculate the similarity score Q between the embedded representation of the test instance and the embedded representation of all patient hyperedges. u ;

[0033] At any similarity score Q u If all similarity scores are less than the second set threshold, the similarity scores Q corresponding to each drug combination in the first preliminary recommended drug set will be... v As a recommended score for the corresponding drug combination;

[0034] At least one similarity score Q u If the value exceeds the second preset threshold, a second preliminary recommended drug set is obtained for the test instance based on the second preset threshold, and the similarity score Q corresponding to the drug combinations in the first preliminary recommended drug set is combined. v The similarity score Q corresponding to the second preliminary recommended drug set u The recommended score for the corresponding drug combination is obtained.

[0035] Furthermore, in each training cycle of the drug combination recommendation model, the training loss function is L = π[0].L bce +π[1]·L multi +L DDI L bce L represents the binary cross-entropy loss function used to measure the difference between the predicted probability and the true label. multi L represents the multi-label margin loss function used to ensure correct sorting. DDI Let represent the DDI loss function used to prevent adverse drug interactions, and π[0] and π[1] represent the weight coefficients of the binary cross-entropy and multi-label margin loss in the multi-label prediction loss function, respectively, and π[0]+π[1]=1.

[0036] A drug combination recommendation system based on hierarchical hypergraph convolutional networks, the system comprising:

[0037] The recommendation score calculation module is used to input the diagnosis set and surgery set of the test instances into the drug combination recommendation model to obtain the recommendation score of the drug combination; wherein, the drug combination recommendation model is used for:

[0038] The embedding representations of patient hyperedges and visit hyperedges in a hierarchical hypergraph are computed; wherein, the hierarchical hypergraph includes: entity hyperedges, visit hyperedges and patient hyperedges, the entity hyperedges are constructed based on the diagnosis set, surgery set and drug set in the historical medical records of patient instances, the visit hyperedges are entity hyperedges belonging to the same visit, and the patient hyperedges are all visit hyperedges connecting the same patient;

[0039] Based on the diagnostic set and surgical set of the test instance, an embedded representation of the test instance is generated;

[0040] The recommendation score for the drug combination is obtained based on the similarity between the embedding representation of the test instance and the embedding representation of the patient hyperedge and the access hyperedge.

[0041] The drug combination recommendation module is used to obtain recommended drug combinations for this test instance based on the recommendation scores of the drug combinations.

[0042] An electronic device, characterized in that the electronic device comprises: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the drug combination recommendation method based on hierarchical hypergraph convolutional networks as described above.

[0043] A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the drug combination recommendation method based on hierarchical hypergraph convolutional networks as described above.

[0044] A computer program product, characterized in that, when the computer program product is run on a computer device, it causes the computer device to execute the drug combination recommendation method based on hierarchical hypergraph convolutional networks as described above.

[0045] Compared with existing technologies, this invention fully combines the representational capabilities of capsule networks with the structural modeling advantages of hierarchical hypergraphs to achieve efficient modeling of personalized patient representations and complex medical data relationships. In this method, the invention captures multifaceted features of patient instances in diagnostic and surgical data through a capsule-based patient representation learning module, enhancing the expressive power of patient representations. A hierarchical hypergraph structure is designed and combined with hypergraph convolution technology to model medical entities and their complex interactions in electronic medical records and historical visit sequences, obtaining embedded representations of global and local information. Furthermore, this invention integrates the outputs of multiple modules to construct a drug combination prediction mechanism, accurately recommending personalized drug combinations for test patient instances, thereby effectively improving the application performance of recommendation systems in complex medical scenarios. Attached Figure Description

[0046] Figure 1 This is a flowchart of a drug combination recommendation method based on hierarchical hypergraph convolutional networks.

[0047] Figure 2 This is a flowchart of the model training process.

[0048] Figure 3 This is a flowchart of the model during the prediction phase.

[0049] Figure 4This is a first-result graph showing the model's performance compared to existing models on common public datasets.

[0050] Figure 5 This is a second result graph showing the performance of the model and existing models on common public datasets. Detailed Implementation

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0052] The drug combination recommendation method based on hierarchical hypergraph convolutional networks of the present invention, such as... Figure 1 As shown, it includes the following steps.

[0053] Step 1: Input the test instance and historical medical visit sequence into a drug combination recommendation model composed of a hierarchical hypergraph convolutional network to obtain the recommendation scores of different drugs for the test instance.

[0054] The drug combination recommendation method based on hierarchical hypergraph convolutional networks of this invention is implemented based on a drug combination recommendation model. This model first embeds the patient query representation into a latent space through a capsule network. Then, it maps the embedded representations of different medical entities (such as diagnosis, surgery, and drugs) to multi-level specific spaces using hierarchical hypergraph convolutional modules. By combining the weights of each hyperedge in these specific spaces, it captures the complex interaction relationships between medical entities and the influence of historical records on the current recommendation, thereby generating a drug combination recommendation score for the patient query.

[0055] Figure 2 The training flowchart for this drug combination recommendation model includes: inputting test instances into a patient representation learning module composed of capsule networks to obtain the embedded representation of the test instances; inputting electronic medical records into a hierarchical hypergraph convolutional model to construct entity hyperedges, access hyperedges, and patient hyperedges, and obtaining the embedded representations of each access hyperedge and patient hyperedge; calculating the similarity between the embedded representation of the test instance and the patient and access hyperedges, and obtaining the drug set corresponding to each access through the association of hyperedges; calculating the binary cross-entropy loss, multi-label margin loss, and DDI loss used to prevent adverse drug interactions, and calculating the final loss function to optimize the model parameters. For a detailed training process of this drug combination recommendation model, please refer to the following steps.

[0056] Step 1.1: Input the test instance into the patient representation learning module composed of capsule networks to obtain the embedded representation of the test instance.

[0057] This invention obtains the embedded representation of an instance by embedding the diagnostic and surgical sets of an instance into a latent space and inputting them into a patient representation learning module composed of a capsule network. In other words, the training device in this embodiment acquires a sample set, including sample data and sample annotations, data acquisition, and preprocessing. First, the instance is embedded into the latent space. Specifically, the diagnostic and surgical sets of the instance are converted into one-hot codes to obtain an encoded representation for each instance, ensuring that all features are appropriately scaled in real-world applications. After obtaining the encoded representation, the training device inputs the diagnostic and surgical embedded representations of the instance into the patient representation learning module composed of a capsule network to obtain the embedded representation of each instance.

[0058] Specifically, the training device first embeds the diagnostic procedures of the instances, mapping them into the latent space. This requires encoding the instance's diagnosis and procedures first. Ultimately, instance u... i The diagnosis and surgery are represented as two vectors in the feature space. and Where t represents the last access sequence time of the instance to be recommended for drug testing. One-hot vector. and It is generated based on the diagnosis and surgery codes. For example, if there are 6 diagnosis codes, and the diagnosis set of this patient instance includes the first diagnosis and the fifth diagnosis, then the patient's diagnosis one-hot vector is [1, 0, 0, 0, 1, 0]. The surgery codes are generated in the same way.

[0059] Next, this invention first randomly initializes an embedding matrix of dimension |D|×d and an embedding matrix of dimension |P|×d, where |D| represents the number of all diagnostic codes, |P| represents the number of all surgical codes, and d represents the dimension of the embedding representation. Then, it is multiplied with the one-hot vector corresponding to the patient, so that the embedding matrix only contains the initial embedding corresponding to that patient's code. As the model trains, the embedding representation of each code (corresponding to a column in the embedding matrix) is learned and updated, and then input into the capsule network module.

[0060] Finally, the training device first performs a nonlinear transformation on the embedding matrices of diagnosis and surgery using a dynamic routing mechanism to obtain the output vector of each capsule network. b ij W is the coupling coefficient calculated through a dynamic routing mechanism. j The weight matrix is ​​a linear transformation matrix, and squash(·) is a non-linear activation function that restricts the magnitude of the output vector to the range (0, 1). and This indicates that the j-th capsule is effective for patient u. i The diagnostic and surgical outputs are then concatenated; the final embedded representation of the diagnosis and surgery is obtained by splicing the outputs of all capsules. and Finally, the diagnostic and surgical embedding representations of this instance are concatenated to obtain the embedding representation of this instance. in For the patient's embedding representation, CONCAT represents the concatenation of the embedding representations of the diagnosis and surgery. As the model trains, the embedding table for each patient instance is learned and updated, resulting in the trained embeddings.

[0061] Through the above steps, the features of the instances are effectively embedded into the latent space, providing a more accurate and robust foundation for subsequent recommendations and predictions.

[0062] Step 1.2: Input the historical records of patient instances into the hierarchical hypergraph convolution model to obtain the embedding representations of each patient hyperedge and access hyperedge.

[0063] This invention inputs historical medical records into a designed hierarchical hypergraph construction and convolution module to obtain embedded representations of each node and hyperedge for subsequent similarity calculation. To this end, this technical solution constructs hyperedges for the diagnosis set, surgery set, and medication set in each medical record to obtain entity hyperedges. and in, Indicates access to v i Diagnostic entity hyperedge, Indicates access to v i The surgical entity is beyond the edge. Indicates access to v i The drug entity hyperedge; then, the connection belongs to the same access v i The three types of entity hyperedges are: diagnostic entity hyperedge, surgical entity hyperedge, and drug entity hyperedge. The visit hyperedge for this visit is obtained. Finally, the connection belongs to the same patient u j All access hyperedges Obtain the patient's super-border Where m and n represent patient u j The beginning and end access sequences are defined. Based on the above, this invention constructs a hierarchical hypergraph structure. To mine the local and global interaction features of this hierarchical hypergraph structure, this invention designs a corresponding hierarchical hypergraph convolutional network.

[0064] Specifically, to compute the representations of each node, including diagnostic nodes, surgical nodes, and drug nodes, as well as each hyperedge, including entity hyperedges, access hyperedges, and patient hyperedges, a hierarchical hypergraph convolutional network was designed for the training device. Specifically, the feature matrices of three medical entity nodes are first initialized. These represent the feature matrices of the diagnostic node, the surgical node, and the drug node, respectively. To obtain the node representation after convolution within the corresponding entity hyperedge, a self-attention mechanism is used to perform convolution within each entity hyperedge.

[0065]

[0066] The corresponding correlation matrix is: |ε d | represents the number of diagnostic superedges, |ε p | represents the number of surgical hyperedges, |ε m | represents the number of drug hyperedges. Att(X) represents the self-attention mechanism acting on the feature matrix X. and Z represents the weight matrix of the l-th layer, and σ represents the sigmoid activation function. Subsequently, to obtain the embedding representation of each entity hyperedge, we perform mean pooling on the updated node representation. Thus, the updated embedding representation of the diagnostic entity hyperedge is Z. d The embedding of the surgical entity hyperedge is represented by Z. p The embedding of the drug entity hyperedge is represented as Z. m .

[0067] To aggregate the hyperedges of diagnostic, surgical, and drug entities in each visit into the visit hyperedge of that visit, this technical solution introduces a learnable weight α. d α p and α m The relative importance of the three types of entity hyperedges is determined by normalizing them using the softmax function.

[0068]

[0069] Where, ω i These are learnable scalar parameters.

[0070] Then, to propagate information to the patient level, this invention aggregates visit hyperedges into patient hyperedges. Each patient hyperedge consists of multiple visit hyperedges. Considering that different visits may contribute differently to the patient-level representation, this invention introduces a learnable weight parameter β for each visit hyperedge. j The weights are normalized using the softmax function and further modulated by a time decay factor to reflect the relative importance of each patient visit.

[0071]

[0072] Among them, b j As a learnable parameter, N represents the total number of patient visits, and γ j This is the introduced time decay factor.

[0073] Finally, the patient hyperedge embedding representation It is obtained by weighted summation of all associated access superedges.

[0074]

[0075] in, Let β represent the embedding representation of the superedge visited for the j-th time. j To access v j Time decay weight, The learnable weight matrix and correlation matrix represent the patient-layer feature transformation. This indicates the relationship between a patient and related medical visits, where |ε u | represents the patient's excess edge number, |ε v | represents the number of accessed hyperedges, and σ represents the activation function.

[0076] In this way, a holistic view of the patient's medical history can be captured, effectively integrating information from all visits, while allowing the model to learn the relative importance of each visit. Furthermore, the structural relationships between visits are considered, and a time decay factor is designed to model the temporal sequence.

[0077] Step 1.3: Calculate the similarity between the embedded representation of the test instance and the patient hyperedge and the access hyperedge, and obtain the drug set corresponding to each access through the association of hyperedges.

[0078] This invention obtains a set of recommended drugs by calculating the similarity between the embedded representation of a test instance and the patient and access hyperedges. Specifically, firstly, the relevance of the patient to each access hyperedge is assessed based on the similarity between the patient's embedded representation and the embedded representation of the access hyperedge. This identifies which access hyperedges are more similar to the current patient's history and needs. Secondly, based on the similarity between the patient instance and the patient hyperedge, the degree of association between the patient and their own historical treatment records is further analyzed. This calculation helps capture the patient's personalized characteristics and avoids excessive interference from the patient's own hyperedges on the final recommendation results. Combining the similarity between the patient instance and the access hyperedge generates a comprehensive similarity score. This score not only considers the matching degree between the patient and the access hyperedge but also fully reflects the correlation between the patient's historical data and access data. Flexible optimization of the similarity score is achieved by adjusting the learnable parameters. Finally, based on the calculated similarity score, the access hyperedges most relevant to the patient are selected, and the top-k most matching drug sets are chosen from them. By weighting the similarity scores, the final recommended drug set is calculated, thus providing the patient with an optimal drug combination recommendation scheme.

[0079] Step 1.4: Calculate the binary cross-entropy loss, multi-label margin loss, and DDI loss used to prevent adverse drug interactions, and calculate the final loss function to optimize the model parameters.

[0080] This invention aims to better balance safety and accuracy in drug combination recommendation by introducing a combined loss. Accuracy and DDI (Drug-to-Drug Inference) values ​​often improve simultaneously during training because drug-drug interactions exist in real-world EHR (Experienced Health Rate) data. Therefore, both incorrect and correct drug predictions can increase the DDI rate. To achieve an accurate model with a low DDI rate, this invention needs to find a balance. To this end, this invention introduces a combined loss. Specifically, this technical solution first uses two commonly used loss functions: binary cross-entropy loss and multi-label marginal loss. Binary cross-entropy loss measures the difference between the predicted probability and the actual binary label for each drug, while multi-label marginal loss measures the difference between the scores of relevant and irrelevant drugs to ensure appropriate ranking. Simultaneously, this invention also designs a DDI loss to reduce the DDI rate.

[0081] Specifically, the advantage of using a combined loss function design is that it can simultaneously consider the accuracy of drug recommendations and the risk of drug-drug interactions (DDIs), thus effectively balancing the contradiction between the two. Traditional drug combination recommendation models typically focus on optimizing prediction accuracy but often ignore the potential risks of drug-drug interactions, which may lead to patients facing unsafe drug combinations in actual use. Focusing solely on DDI rates may result in overly conservative recommendations, affecting the accuracy and effectiveness of the recommendations. By introducing binary cross-entropy loss and multi-label marginal loss, the model can accurately predict the recommendation probability of each drug while ensuring that the ranking order of drugs aligns with actual medication logic. Binary cross-entropy loss accurately measures whether a drug should be recommended to a patient, while multi-label marginal loss, by incorporating the relative relationships between drugs, avoids bias in the recommendation model and improves the quality of recommendations.

[0082]

[0083] L p =π[0]L bce +π[1]L multi

[0084] in, This represents the true label of the i-th drug. Let represent the predicted probability of the i-th drug at time t, where π[0] and π[1] are mixed weights, and π[0]≥0, π[1]≥0 and π[0]+π[1]=1 are guaranteed.

[0085] Furthermore, the introduction of DDI loss allows the model to focus not only on the accuracy of drug prediction but also to actively consider potential adverse interactions between drugs during training, thereby suppressing the generation of high-risk drug combinations. By reducing the DDI rate, the model can recommend drug combinations that are both effective and safe, thus improving the feasibility and safety of clinical recommendations.

[0086]

[0087] Where · represents the dot product between scalars. DDI adjacency matrix A ddi [i, j] = 1 indicates that the i-th drug interacts with the j-th drug. This loss function is defined for a single visit. During model training, backpropagation of the loss will be performed at the patient level using the average loss from all visits.

[0088] Ultimately, the design of the combined loss function achieves multi-dimensional optimization of drug recommendations, enabling the model to improve recommendation accuracy while effectively controlling adverse interactions between drugs. This design not only meets the clinical practice's demand for high accuracy and low risk in drug recommendation systems but also provides strong support for personalized medicine and precision treatment.

[0089]

[0090] Where s′ is the current DDI value and s is the target value. If s′ is greater than s, then L is applied. DDI The chances are higher; otherwise, use L. p As training progresses, the temperature (Temp) decreases, reducing the need for L... DDI The possibility of [something], and shift towards accuracy.

[0091] In one embodiment, after training the prototype network using the objective function, the training device can also output the trained prototype network to the execution device (e.g., a terminal device, an edge device, or other target device). In the above embodiments of this application, it is specifically described that the prototype network trained by the training device can be sent to the execution device, where the execution device can specifically apply the trained prototype network. This separates the training and application processes of the prototype network, reducing the computational load on the execution device and saving its computing power.

[0092] Step 2: Based on the recommendation scores of different drugs for this test instance, obtain the drug combination recommendation results for the test instance.

[0093] Relevant devices (such as execution devices or training devices) can make predictions using this trained drug combination recommendation model. The execution device can utilize a novel model application approach. In the application phase, the execution device first calculates the similarity between the embedded representation of the test instance and the access hyperedges. This process identifies the access hyperedges most relevant to the patient's needs by evaluating the similarity between the patient embedding and each access hyperedge embedding. This step helps the execution device assess the relevance of the patient to each access hyperedge from a local perspective, thereby providing a basis for subsequent drug recommendations.

[0094] However, relying solely on local similarity can lead to uncertainty in recommendations for some instances, especially when the similarity values ​​of certain access hyperedges are close. To address this issue, the execution device also considers the similarity between the patient instance and the patient hyperedge, further analyzing the relationship between the patient and their historical treatment records. In this process, the execution device can effectively capture the patient's personalized characteristics and reduce the impact of the patient's own hyperedge on the final recommendation results.

[0095] By combining global and local similarity, the execution device generates a comprehensive similarity score. This score considers not only the match between the patient and the access hyperedge but also fully reflects the correlation between the patient's historical and current data. After obtaining the comprehensive similarity score, the execution device filters out the access hyperedges most relevant to the patient and selects the top-k best-matching drug sets from them. Using a weighted similarity score, the execution device finally calculates the recommended drug set, providing accurate drug combination recommendations. This method ensures that the drug recommendation process is both highly accurate and can provide tailored recommendations based on the patient's individual needs.

[0096] Specifically Figure 3 This is a flowchart of the model during the prediction phase. During prediction, the execution device first computes the patient embedding from a local perspective. and each access superedge embedding Z u similarity in, The first column stores the access superedge ID, the second column stores the similarity between each access superedge embedding and the patient superedge embedding, and then the execution device retains access superedges with high similarity scores according to a first set threshold.

[0097] Then, from a global perspective, the similarity between the patient's hyperedge and the hyperedges of the test instance is calculated. in, The first column stores the patient ID, and the second column stores the similarity between the patient's hyperedge and the patient instance embedding. Let the embedding representation of the superedge of the i-th patient be represented. The embedding representation of the j-th patient's superedge is given. To ensure that the patient's own superedge does not dominate the final result, a learnable weight ω is introduced when i = j. Then, the execution device retains the patient superedge with high similarity score according to the second set threshold.

[0098] If the similarity of all patient hyperedges is less than the second set threshold, then the score of the drug set is obtained only based on the similarity of the accessed hyperedges.

[0099] If the similarity of a patient's hyperedge exceeds a second predefined threshold, then a similarity score S = α is calculated for the patient's hyperedge and all access hyperedges of the same patient. u ·Q u +β v ·Q v , where α u and β v The parameters are learnable and satisfy α u +β v =1. The execution device selects the set of drugs with the highest scores as the recommended drugs for the instance.

[0100] Finally, based on the similarity score S, the top-k most similar drug sets M corresponding to the visits are selected. vv The final drug recommendation result is obtained by weighting the results. Among them, S[v i ] indicates accessing v i The similarity score, δ(m) k ∈M vv [v i ]) is an indicator function, indicating whether the drug exists in the drug set M. vv [v i If the result is 1, then it equals 1; otherwise, it equals 0.

[0101] In summary, this invention provides a drug combination recommendation model based on patient embedding and hyperedge similarity calculation. By combining local similarity and global information, the training device can more accurately capture the correlation features between patients' historical treatment records and candidate drug sets, thereby improving the accuracy and safety of drug recommendations. By introducing a comparison mechanism between patient instances and patient hyperedges, the training device can effectively distinguish the individual characteristics of different patients, reduce the interference of patient hyperedges on recommendation results, and improve the model's performance in complex medication scenarios. Because a learnable parameter optimization strategy is used during training, the execution device can flexibly balance global and local similarity information during the inference phase, thereby selecting the optimal drug combination and ensuring the accuracy and safety of the recommendation results. Furthermore, the method of this invention, by dynamically adjusting the similarity scoring mechanism, effectively adapts to the complex needs of different patient instances, further enhancing the model's generalization ability and robustness in diverse medical scenarios.

[0102] When using the training method provided by this invention, the obtained recommendation model significantly outperforms traditional methods in drug recommendation tasks. Experimental results fully demonstrate the effectiveness and reliability of the method in terms of accuracy, safety, and generalization ability. The method of this invention is highly adaptable and can flexibly handle different patient instances and diverse medical scenarios. The method of this invention can effectively achieve accurate recommendation and optimization of drug combinations; its design is reasonable, computationally efficient, and easy to deploy, providing strong technical support for clinical drug recommendation.

[0103] Figure 4 and Figure 5 This figure compares the experimental performance of the method proposed in this invention with existing drug combination recommendation methods. The dataset selected in this invention is a standard dataset commonly used in existing methods. This dataset was obtained from a real hospital setting, and patient information was anonymized. The method proposed in this invention demonstrates superior recommendation accuracy and a lower DDI value on this dataset.

[0104] Finally, the method of this invention is merely a preferred embodiment and is not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A drug combination recommendation method based on hierarchical hypergraph convolutional networks, characterized in that, The method includes: The diagnostic and surgical sets of the test instances are input into the drug combination recommendation model to obtain a recommendation score for the drug combination; wherein, the drug combination recommendation model is used for: The embedding representations of patient hyperedges and visit hyperedges in a hierarchical hypergraph are computed; wherein, the hierarchical hypergraph includes: entity hyperedges, visit hyperedges and patient hyperedges, the entity hyperedges are constructed based on the diagnosis set, surgery set and drug set in the historical medical records of patient instances, the visit hyperedges are entity hyperedges belonging to the same visit, and the patient hyperedges are all visit hyperedges connecting the same patient; Based on the diagnostic set and surgical set of the test instance, an embedded representation of the test instance is generated; The recommendation score for the drug combination is obtained based on the similarity between the embedding representation of the test instance and the embedding representation of the patient hyperedge and the access hyperedge. Based on the recommended score of the drug combination, the recommended drug combination for this test instance is obtained; The step of generating an embedded representation of a test instance based on the diagnostic set and surgical set of the test instance includes: Based on the diagnosis number and surgery number of the test instance, the diagnosis set and surgery set are one-hot encoded respectively to obtain the diagnosis one-hot vector and the encoding one-hot vector; Obtain the learned embedding matrix; The learned embedding matrix is ​​multiplied by the diagnostic one-hot vector and the encoded one-hot vector to obtain the diagnostic embedding representation and the surgical embedding representation; The diagnostic embedding representation and the surgical embedding representation are respectively input into the capsule network to obtain the output value of each capsule, and the output values ​​of the corresponding capsules are concatenated to obtain the capsule network diagnostic output and capsule network surgical output. By concatenating the capsule network diagnostic output and the capsule network surgical output, an embedded representation of the test instance is obtained.

2. The method according to claim 1, characterized in that, The embedded representation of access hyperedges in the computational hierarchy hypergraph includes: Initialize the feature matrix of entity nodes in the drug combination recommendation hierarchy hypergraph; wherein, the entity nodes include: diagnostic nodes, surgical nodes, and drug nodes; A self-attention mechanism is used to convolve different entity nodes inside each entity hyperedge to obtain an updated entity node feature matrix; wherein, the entity hyperedges include: diagnostic hyperedges, surgical hyperedges and drug hyperedges; The updated entity node feature matrix is ​​subjected to average pooling to obtain the embedding representation of each entity hyperedge; We weight the embedding representations of the diagnostic, surgical, and drug hyperedges for the same visit to obtain the embedding representation of the visit hyperedge.

3. The method according to claim 1, characterized in that, The embedded representation of the patient hyperedge in the computational hierarchical hypergraph includes: Introduce a time decay factor to all access superedges belonging to the same patient instance. Where t j λ is the time interval between the j-th access and the current time, and λ is a hyperparameter that controls the decay rate. According to the time decay factor γ j Calculate the weight of each visited hyperedge; The embedding representation of the patient's hyperedge is obtained by weighting the access hyperedges by combining their weights.

4. The method according to claim 1, characterized in that, The recommendation score for the drug combination is obtained based on the similarity between the embedding representation of the test instance and the embedding representation of the patient hyperedge and the access hyperedge, including: Calculate the similarity score Q between the embedding representation of the test instance and the embedding representations of all visited superedges. v And based on the first set threshold, the first preliminary recommended drug set for the test instance is obtained; Calculate the similarity score Q between the embedded representation of the test instance and the embedded representation of all patient hyperedges. u ; At any similarity score Q u If all similarity scores are less than the second set threshold, the similarity scores Q corresponding to each drug combination in the first preliminary recommended drug set will be... v As a recommended score for the corresponding drug combination; At least one similarity score Q u If the value exceeds the second preset threshold, a second preliminary recommended drug set is obtained for the test instance based on the second preset threshold, and the similarity score Q corresponding to the drug combinations in the first preliminary recommended drug set is combined. v The similarity score Q corresponding to the second preliminary recommended drug set u The recommended score for the corresponding drug combination is obtained.

5. The method according to any one of claims 1 to 4, characterized in that, In each training cycle of the drug combination recommendation model, the training loss function is L = π[0]·L bce +π[1]·L multi +L DDI L bce L represents the binary cross-entropy loss function used to measure the difference between the predicted probability and the true label. multi L represents the multi-label margin loss function used to ensure correct sorting. DDI Let represent the DDI loss function used to prevent adverse drug interactions, and π[0] and π[1] represent the weight coefficients of the binary cross-entropy and multi-label margin loss in the multi-label prediction loss function, respectively, and π[0]+π[1]=1.

6. A drug combination recommendation system based on hierarchical hypergraph convolutional networks, characterized in that, The system includes: The recommendation score calculation module is used to input the diagnosis set and surgery set of the test instances into the drug combination recommendation model to obtain the recommendation score of the drug combination; wherein, the drug combination recommendation model is used for: The embedding representations of patient hyperedges and visit hyperedges in a hierarchical hypergraph are computed; wherein, the hierarchical hypergraph includes: entity hyperedges, visit hyperedges and patient hyperedges, the entity hyperedges are constructed based on the diagnosis set, surgery set and drug set in the historical medical records of patient instances, the visit hyperedges are entity hyperedges belonging to the same visit, and the patient hyperedges are all visit hyperedges connecting the same patient; Based on the diagnostic set and surgical set of the test instance, an embedded representation of the test instance is generated; The recommendation score for the drug combination is obtained based on the similarity between the embedding representation of the test instance and the embedding representation of the patient hyperedge and the access hyperedge. The drug combination recommendation module is used to obtain recommended drug combinations for the test instance based on the recommendation score of the drug combinations; wherein, the step of generating an embedded representation of the test instance based on the diagnostic set and surgical set of the test instance includes: Based on the diagnosis number and surgery number of the test instance, the diagnosis set and surgery set are one-hot encoded respectively to obtain the diagnosis one-hot vector and the encoding one-hot vector; Obtain the learned embedding matrix; The learned embedding matrix is ​​multiplied by the diagnostic one-hot vector and the encoded one-hot vector to obtain the diagnostic embedding representation and the surgical embedding representation; The diagnostic embedding representation and the surgical embedding representation are respectively input into the capsule network to obtain the output value of each capsule, and the output values ​​of the corresponding capsules are concatenated to obtain the capsule network diagnostic output and capsule network surgical output. By concatenating the capsule network diagnostic output and the capsule network surgical output, an embedded representation of the test instance is obtained.

7. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the drug combination recommendation method based on hierarchical hypergraph convolutional networks as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the drug combination recommendation method based on a hierarchical hypergraph convolutional network as described in any one of claims 1-5.

9. A computer program product, characterized in that, When the computer program product is run on a computer device, the computer device performs the drug combination recommendation method based on hierarchical hypergraph convolutional networks as described in any one of claims 1-5.

Images